Split (1)

train · 7.63k rows

title string	paper_decision string	review_1 string	rebuttals_1 string	review_2 string	rebuttals_2 string	review_3 string	rebuttals_3 string	review_4 string	rebuttals_4 string	global_rebuttals string	dataset_source string	conference_year int64	review_5 string	rebuttals_5 string	review_6 string	rebuttals_6 string	review_7 string	rebuttals_7 string	review_8 string	rebuttals_8 string
Neglected Hessian component explains mysteries in sharpness regularization	Accept (spotlight)	Summary: It is known that SAM can improve generalization, while weight noises and gradient penalties often fail. This work reported that the structure of the Hessian can explain the inconsistency by identifying the key role of NME. This work first studied gradient penalties and show that methods using second order information are sensitive to activation functions due to NME. Moreover, as NME matters to Hessian penalties, weight noises which minimizes NME often behave poorly. Strengths: - This work reports a novel and interesting conclusion that the Hessian structure, particularly the overlooked NME, can explain the poor performance of gradient penalties and Hessian penalties. This is significant for understanding the second order information in optimization and generalization of deep learning. - The theoretical analysis seems clear and reasonable. - This work empirically verified the theoretical results and identified the key role of NME. Weaknesses: -The empirical results only include ResNets. May the empirical conclusion also depend on model architectures? How about the results of very simple models (FCN/LR) and very complex models (Transformers)? Note that they have very different loss landscape. -Training with weight noises only implicitly penalize the Hessian. This work often abuse Hessian penalties and weight noises. -This work provides insights but failed to show how to improve or design better second-order-based methods. This can further support the conclusion of this work. Technical Quality: 2 Clarity: 3 Questions for Authors: Please see the weaknesses. Confidence: 4 Soundness: 2 Presentation: 3 Contribution: 3 Limitations: This work did not discuss the limitations. Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We thank the reviewer for their questions, and answer some here. _The empirical results only include ResNets. May the empirical conclusion also depend on model architectures? How about the results of very simple models (FCN/LR) and very complex models (Transformers)? Note that they have very different loss landscape._ We agree that studying more architectures and datasets would be interesting; for this initial work we focused on very detailed experiments in our chosen settings rather than shallow experiments across a broad range of settings. _Training with weight noises only implicitly penalize the Hessian. This work often abuse Hessian penalties and weight noises._ In Sections 5.1 and 5.2, we review the links made _in the literature_ between weight noise and Hessian penalties in order to frame our work. However, in all our experiments we make the procedure clear, and mainly focus on explicit regularization of the Hessian/Gauss Newton trace. _This work provides insights but failed to show how to improve or design better second-order-based methods. This can further support the conclusion of this work._ We agree that additional research is needed to further improve second order methods/regularizers; however our work showed that the Gauss Newton trace is another useful regularizer. In addition, our synthetic second derivative approach in Section 4.5 shows how we can overcome the poor second derivatives of ReLU in a practical, efficient way. --- Rebuttal Comment 1.1: Title: Thanks for the rebuttal Comment: Thanks for the rebuttal. It addressed some of my concerns. I tend to keep the rating of this work as Weak Accept.	Summary: The authors examine the performance of training methods for neural networks, which utilize (approximate) second order information. They note that often only the curvature of the loss function is taken into account (rather than that of the loss functions) and demonstrate that this approach leads to drawback in training in particular regrading sharpness regularization such as gradient penalties / weight noise. They examine the effect of including / excluding second-order information in the form of regularizations on generalizability based on numerical experiments. Strengths: The authors demonstrate a problem in learning regularizations that has so far no been addressed, namely the varying performance of the regularizations and the mixed success of for example gradient norm penalties. The observations made have the potential to aid in the development of learning algorithms. Weaknesses: In terms of the use of second derivatives the manuscript sends a bit of a mixed message. The initial experiment in Section 4 seems to show that accurate second-order information of the model is useful when using gradient penalties in learning algorithms, the Hessian penalties in Section 5 seem to pain a different picture, where the entire Hessian trace penalty performs the worst, so neglecting Hessian components may in fact not be such a bad idea in general. Technical Quality: 3 Clarity: 3 Questions for Authors: - 83 - 84: It would be helpful to reformulate the defintion of the trace to make clear that the trace being the sum of eigenvalues is an established result rather than a definition of the trace operator to avoid confusion. - 123: The function L changes its meaning halfway through the manuscript. Initially, it depends on z and y, i.e., on model output and labels. Starting with 123 it becomes a function with the sole input being a vector of weights. I think this is supposed to be L(z(theta, x), y) for fixed training data (x, y). Please introduce another function to make this clearer to the reader. - Some abbreviations could be given once in full in order to jog the memory of readers who are not very familiar with the precise context (NTK, ReLU / GELU) - 143: Should this be SGD (there are not batches to be seen) or a normal gradient descent? - 226: "Minimizing the NME is detrimental" Shouldn't this state that "ignoring" the NME is a problem? Confidence: 4 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: NA Flag For Ethics Review: ['No ethics review needed.'] Rating: 7 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We thank the reviewer for their comments, and will fix the errors noted. We address some selected concerns below. _In terms of the use of second derivatives the manuscript sends a bit of a mixed message._ The overall point about second derivatives is a bit subtle; our work suggests that second derivatives in the _update rule_ are helpful (get full NME information in HVP). In contrast, the full Hessian trace penalty has second derivative information _in the regularizer_. Minimizing this information during optimization reduces the effect of second derivatives in the _update rule_. It also introduces _third derivatives_ into the update rule; these higher order derivatives in the updates seem to hurt training. It remains an open question whether or not other forms of higher order derivatives are useful. _143: Should this be SGD (there are not batches to be seen) or a normal gradient descent?_ We wrote the rule for a single batch; in practice this update rule would be combined with batching/SGD. _226: "Minimizing the NME is detrimental" Shouldn't this state that "ignoring" the NME is a problem?_ This should read “Minimizing the NME trace” is detrimental, as the GN trace penalty (no NME in the regularizer) works well. --- Rebuttal Comment 1.1: Comment: Thank you for the response. Barring these small concerns, I think the paper provides interesting insights into regularization techniques, and I recommend its acceptance.	Summary: This paper investigates the importance of considering second order information, specifically the structure of the Hessian of the loss, in deep learning. It decomposes the Hessian into the Gauss-Newton matrix and the Nonlinear Modeling Error (NME) matrix, with focus on the often-overlooked NME. Through empirical and theoretical evidence, the study demonstrates the significance of the NME in the performance of gradient penalties and their sensitivity to activation functions. The difference in regularization performance between gradient penalties and weight noise is also attributed to the NME. The findings underscore the need to consider the NME in experimental design and theoretical analysis for sharpness regularization, potentially leading to new classes of second order algorithms that utilize the loss landscape geometry differently. Strengths: Strengths 1. The paper is clearly written and easy to follow. 2. The paper focuses on an important topic in the area of sharpness learning, which is worthy of investigation. 3. The paper provides a novel understanding of the connections between SAM and gradient norm penalty through the perspective of NME, offering interesting insights. 4. The paper highlights the pitfalls of using different activations with the Hessian, providing valuable guidance for practical training. Weaknesses: Weakness 1. How do you solve SAM? Have you also neglected the second-order term in your empirical analysis? What would be the effect if this term were not neglected in your situation? 2. Could you provide some demonstrations regarding the off-diagonal elements in the NME? It is suggested that these elements may also play an important role. 3. I recommend that the theoretical analysis be expressed in a more formal and clear style. 4. There are a couple of typos: Line 133, "$p = 1$" -> "$\rho = 1$"; Line 208, "if this is link" -> "if this link". Technical Quality: 2 Clarity: 3 Questions for Authors: See Weakness. Confidence: 4 Soundness: 2 Presentation: 3 Contribution: 3 Limitations: I have not found any discussions about the limitations and potential negative societal impact. But in my opinion, this may not be a problem, since the work only focuses on understanding the sharpness learning. Still, it is highly encouraged to add corresponding discussions. Flag For Ethics Review: ['No ethics review needed.'] Rating: 7 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We would like to thank the reviewer for their insightful review. We address the reviewer's questions and suggestions below. _How do you solve SAM? Have you also neglected the second-order term in your empirical analysis? What would be the effect if this term were not neglected in your situation?_ Our experimental results for SAM are based on the SAM algorithm proposed by Foret et al. Specifically, the SAM update does not utilize any Hessian (but does evaluate an extra gradient). The reason the Hessian is absent from the SAM update is due to the approximation employed by Foret et al (using stop_gradient on the adversarial perturbation). Therefore, our implementation of SAM, which aligns with that of Foret et al, indeed neglects the Hessian. We have not investigated the implications of retaining the Hessian in the SAM formulation because this was already explored by Foret et al. They conducted this experiment and observed that by keeping the Hessian, SAM's performance slightly decreases compared to when the Hessian is neglected (which is consistent with our story). This is discussed in Figure 4 of their paper. To derive PSAM from SAM Equation 6, we employ a first-order approximation as shown in Equations 9-11. This differs from the approximation used in the original SAM algorithm [Foret et al., 2021]. While exploring a second-order approximation could be a promising avenue, it falls outside the scope of this current work. _Could you provide some demonstrations regarding the off-diagonal elements in the NME? It is suggested that these elements may also play an important role._ The experiments in Section 4.4 show that the off-diagonal elements of the NME by themselves are not sufficient to obtain good results with PSAM. In these experiments we artificially zero out the diagonal elements of the NME of GELU. It would be interesting to see how well PSAM with ONLY the diagonal elements performs, but we did not have time in the rebuttal period to design and run this experiment. _I recommend that the theoretical analysis be expressed in a more formal and clear style._ Thank you for your suggestion. We will improve the style of the theoretical analysis and address the typos you identified. --- Rebuttal Comment 1.1: Comment: Thanks for the authors' kind response. Having no further concerns, I believe that this paper serves as an excellent resource for exploring gradient regularization and SAM, and I strongly recommend its acceptance.	Summary: This paper studies the influence of the second-order component of the Hessian in sharpness-aware minimization and optimization methods that involve gradient penalties. First, they show that the Hessian decomposes into the component of the Hessian that people usually consider [GN] and a term that includes the second derivative with respect to the parameters [NME]. From there, they demonstrate that a gap between penalty sharpness-aware minimization and general sharpness-aware minimization exists when activations are ReLU but not when activations are GeLU. They study how the NME explains this gap by showing that ablating the NME component in GeLU recovers the ReLU generalization. Further, they also show that including a term related to the NME in the ReLU case removes this gap. Finally, they ask the question of whether explicitly accounting for the NME in sharpness-aware minimization / gradient-penalty via weight noise can improve generalization of solutions, finding that it cannot in general. In the end, this work, to my understanding, deepens our understanding of the role of the usually-ignored NME term in various settings. Strengths: - identifies a part of the hessian that is not usually considered in sharpness-aware minimization and isolates its effect in various settings, showing that in some cases it is relevant for understanding the behavior, while in others there are reasons to not include it in algorithms. - theoretical analysis as well as experiments on real-scale datasets - appropriate ablations to study the effects of terms, mainly on outcome. I would be interested in future work looking at dynamics, as well! - I like that the work studies the nuances of the effects of the NME, not just tells a single story and leaves it at that. Weaknesses: overall, I feel the work studies the questions asked in detail. See some questions / writing suggestions below. Technical Quality: 4 Clarity: 4 Questions for Authors: - In section 2 when deriving the NME, can you include an explicit example where the NME is large? Either analytically or a plot demonstrating the evolution of, e.g., its Frobenius norm over the course of training. - small thing: perhaps when you mention the 3 datasets, reverse the order to have them in increasing order of difficulty - one thing you do not discuss but I would want to know more about is — PSAM method only should work when rho is small. And indeed, when rho is small, even for ReLU the methods match. Of course, since the GeLU has the match even at the values of rho where there is a gap for ReLU, it seems that the size of rho is not the determining factor. But could you still comment on how to know whether the issue is just the size of rho? - any comments on dynamics of training with and without the additional penalty in section 5.2? Seems like that might also be interesting. Confidence: 2 Soundness: 4 Presentation: 4 Contribution: 4 Limitations: the work analyzes the effect of the NME in various settings and is clear about the particular settings which it studies. Flag For Ethics Review: ['No ethics review needed.'] Rating: 8 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We thank the reviewer for their thorough review. We address the reviewer's questions and suggestions below. _In section 2 when deriving the NME, can you include an explicit example where the NME is large? Either analytically or a plot demonstrating the evolution of, e.g., its Frobenius norm over the course of training._ The suggestion to identify specific examples where the NME is large is valuable. We are actively investigating this, but due to time constraints in preparing the rebuttal we don’t have definite results yet. _small thing: perhaps when you mention the 3 datasets, reverse the order to have them in increasing order of difficulty_ We will change the order as you suggested. _one thing you do not discuss but I would want to know more about is — PSAM method only should work when rho is small. And indeed, when rho is small, even for ReLU the methods match. Of course, since the GeLU has the match even at the values of rho where there is a gap for ReLU, it seems that the size of rho is not the determining factor. But could you still comment on how to know whether the issue is just the size of rho?_ Further evidence that the size of rho isn’t the sole factor can be found in Figure 3, which shows the gap for ReLU can be narrowed even at larger rho values by using a synthetic activation NME. --- Rebuttal Comment 1.1: Title: Thanks! Comment: Thanks for your response! And Figure 3 is noted, thanks!	null	NeurIPS_2024_submissions_huggingface	2,024	null	null	null	null	null	null	null	null
How many classifiers do we need?	Accept (poster)	Summary: The authors address the setting of ensembling, where the predictions of multiple models are combined to improve accuracy and form more robust conclusions. The authors define a quantity η (polarity among agents within a dataset), and show both empirically and theoretically that this quantity is nearly constant regardless of hyperparameters or architectures of classifiers. The authors present a tight upper bound for the error of majority vote under restricted entropy conditions. This bound indicates that the disagreement is linearly correlated with the target, and the slope is linear in polarity, η. Finally, the authors prove asymptotic behavior of disagreement in terms of the number of agents, which can help predicting the performance for a larger number of agents from that of a smaller number. Strengths: 1. Section 2 is, by in large, excellent, and does a very good job of summarizing the existing work in this area. 2. The concept of polarity is clearly defined and well supported via the helpful example in Fig. 1. 3. The tighter bounds in Sec. 4 are well-described and scoped. Weaknesses: 1. The claim that the authors have shown empirically that η is nearly constant regardless of hparams or architectures is overbroad for the provided evidence; the authors should consider scoping their empirical claims more modestly, in keeping with the small scale experiments in this paper. 2. The phrase "a distribution over a parametric family hθ , e.g., a distribution of classifiers that is trained from a neural network or a random forest", could be misinterpreted as a claim that random forests are parametric models 3. Equation (6) appears to have a formatting error, there's a long underscore under one half of the equation 4. Conjecture 1 uses the term interpolating neural networks; however, this term is not defined in Sec. 2 and is overloaded in the literature. Please provide a definition in Sec. 2 or a relevant citation. 5. "Since the denominator ... is invariant to the number of classifiers and the numerator resembles the disagreement between classifiers, δ is expected to follow a similar pattern": this point could use more clarification Technical Quality: 4 Clarity: 3 Questions for Authors: QUESTIONS * How do the authors define interpolating neural networks in this work? * Can the authors please expand on the point raised in W5? SUMMARY OF THE REVIEW This paper presents significant new theoretical contributions in its area, which is important and of interest to the larger community, but has a few overclaims and oversights which currently limit it. As the paper stands, I will give a weak accept, but if the weaknesses I list are addressed, I will raise my score. Confidence: 3 Soundness: 4 Presentation: 3 Contribution: 3 Limitations: The authors have adequately addressed the limitations and potential negative societal impact of their work. Flag For Ethics Review: ['No ethics review needed.'] Rating: 7 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thank you for reviewing this paper and we really appreciate your comments and questions. [Weakness 1] Agreed. That was the reason why we narrowed down the scope to interpolating models in Conjecture 1 although Theorem 1 holds for any ensemble. We think it would be an interesting direction to see how $\eta$ differs in non-interpolating models in different fields (e.g. language models). [Weakness 2] Thank you for catching this. We will fix this in the final version. [Weakness 3] The long underscore is due to the footnote below the formula. [Weakness 4, Question 1] Please see the general response: "interpolator" is a model that is trained to perfectly fit the train data. In Conjecture 1, we wanted to draw a statement analogous to so-called 'neural scaling laws', which states that the out-of-sample performance of an interpolating neural network scales with the width of the neural net. We will add a paragraph about this on the final version. [Weakness 5, Question 2] In Theorem 5, we prove that the disagreement for $N$ classifiers has a representation as $(1-1/N)(D_\infty + N^{-1/2}Z_N)$, and through Donsker's invariance principle, it can be approximated through the disagreement of a smaller number of classifiers (say $M$) as $(1-1/N) / (1-1/M) * (\text{Disg. of }M\text{ classifiers})$. The justification for this argument is that the disagreement is a $V$-statistic, and so its bias can be accounted for. The same is true for the numerator of $\delta$ (the denominator is unbiased): it is also a $V$-statistic, and should behave similarly to the disagreement, as its form is the same with only one less label. We will add this clarification in the final version. --- Rebuttal 2: Comment: I would like to thank the authors for their detailed response. To acknowledge that my concerns have been addressed, I will update my score. --- Rebuttal Comment 2.1: Comment: We are glad your concerns were addressed. Please let us know if you have any additional questions.	Summary: This paper analyzes the majority vote error of an ensemble of classifiers. A new quantity called polarity is introduced. The polarity of neural networks is analyzed empirically and theoretically, and stronger bounds on the majority vote error are derived based on the polarity. Finally, the previously derived bounds are used to predict the majority vote error rate of a large ensemble from that of a smaller ensemble. Strengths: As far as I am aware, the notion of an ensemble's polarity around which the paper is developed is a novel concept. The derived bounds based on polarity are stronger than previously known bounds and even subsume some of them, for example those based on the notion of competent ensembles. Weaknesses: The conjecture of the neural polarity is somewhat lacking in evidence given the generality of the statement. While Figure 1 provides some empirical evidence for a variety of models and hyperparameters, only the relatively simple CIFAR-10 dataset is considered. It would be interesting to see if the neural polarity law still holds for more challenging tasks, where the performance of each individual classifier would be lower. I also feel that some parts of the paper could be improved for clarity. In particular, in Section 5 it is somewhat unclear how all the results fit together, which can detract from their potential significance. See some of the questions below. Technical Quality: 3 Clarity: 2 Questions for Authors: 1. The description of Theorem 1 as a lower bound in line 127 is somewhat misleading. As mentioned later in the paper, models can be polarized for $\eta$ lower than the derived bound. I wonder if the wording here can be improved. 2. I do not quite see the connection to neural collapse in line 159. Neural collapse describes the phenomenon that representations of examples from the same class collapse to a single point and the class means form a simplex ETF. Still, it does not state how the probabilities of any given example are distributed among the labels. 3. In theorem 5, what is $\sigma^2$? And what is the significance of the scaled random walk converging to Brownian motion? As far as I can tell, the convergence to Brownian motion is a restatement of a standard result from stochastic calculus and is not used anywhere else in the paper. 4. In the definition of $\mathcal{L}(h)$ in line 215, it appears that the symbol $h$ is overloaded. Confidence: 4 Soundness: 3 Presentation: 2 Contribution: 3 Limitations: Limitations are addressed. Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thank you for reviewing this paper; we really appreciate your comments and questions. [Weakness 1] We agree that further empirical evidence would strengthen our claim. It is difficult to obtain an ensemble of well-trained interpolating models in practice, so we are limited in our selection of models and datasets. For now, we provide further evidence of the universality of $\eta<4/3$ on two additional datasets in the figure document: KMNIST and Fashion-MNIST. Please refer to Figure A on the figure document attached to the rebuttal for all reviewers above. On KMNIST, which is an easier task than CIFAR-10, ResNet18 with various width (48, 64) were trained with various batch size (32,64,128) to interpolate the train data. On Fashion-MNIST, ResNet18 with width 48 was trained with batch size 128 to interpolate the train data. [Weakness 2, Question 3] The main role of Theorem 5 is to justify our approximation of the entire disagreement curve by neglecting $N^{-1/2}Z_N$. If we used a ordinary central limit theorem, we could only justify approximations of $D_n$ for large $n$. Appealing to Donsker's invariance principle, we can justify our procedure of extrapolating the disagreement as $D_n = D_3 * (1-1/n)/(1-1/3)$. In this case, $\sigma^2$ is equal to $Var_{h\sim\rho}(L(h))$. We will update this in the final version along with the typo on the $U_N$ Hoeffding decomposition in Line 214 where $E_{\rho_N^2}$ should be $E_{\rho^2}$ instead. [Question 1] We rephrased the theorem and conjecture and posted as a rebuttal for all reviewers above. We hope this clarifies the statement. [Question 2] Perhaps it is only at an intuitive level, but we see neural collapse and reduced entropy on the probabilities as symptoms of the same general phenomenon. If the representations within each class are reduced to a single point, then a classifier ensemble is going to become increasingly confident within each class, as there is less variability in the outputs. [Question 3] Answered with Weakness 2 above. [Question 4] Thank you for pointing this out. We will update this to $\mathcal{L}(h)=\mathbb{E}\_{h'\sim\rho}\mathbb{P}\_{\mathcal{D}}(h(X)\neq h'(X))-\mathbb{E}\_{(h',h'')\sim\rho^2}\mathbb{P}\_{\mathcal{D}}(h'(X)\neq h''(X))$ in the final version. --- Rebuttal Comment 1.1: Comment: Thank you for the additional results and clarifications. I would be glad to increase my score. --- Reply to Comment 1.1.1: Comment: We are glad the additional results and clarification were helpful. Please let us know if you have any additional questions or need further clarification.	Summary: This paper focuses on quantifying the impact of number of classifiers on the error rate of majority vote decision strategy for ensemble classifiers. They define the notion of “polarity” of an ensemble and use this notion to characterize the relationship between majority vote error rate and disagreement between classifiers in any relevant ensemble. Empirical analysis over CIFAR and MNIST datasets shows that the paper’s method provides a tighter bound for majority error vote than other prior methods. Strengths: 1. Overall, the paper is well-written, concise, and would be valuable to the wide audience working on predictive optimization and ensemble models. 2. The use of the notion of polarity is novel and seems to capture two relevant properties simultaneously: (a) whether a large fraction of classifiers are making an incorrect prediction, and (b) whether the predictions of large groups of classifiers differ from each other. This characterization specially seems to assist in handling the presence of multiple classes as well. The conjectured upper bound on polarity is also independently interesting.   3. I liked the focus on entropy-restricted ensembles. The focus on ensembles with quantifiable “disagreement properties“ (as in Thms 3, 4) allows for stronger bounds on majority vote error rate, although it does seem difficult to then derive bounds for these “disagreement properties” (like the parameter $\delta$). Nevertheless, the approach of bounding majority vote error rate using disagreement notions is intuitively appealing and, as the paper shows, it does provide reasonable theoretical bounds. Weaknesses: 1. Some of the terms used in the paper could use additional descriptions. For example, I don’t see the description of “interpolation” or “interpolating models” anywhere. Considering that it is important in interpreting Figure 1 and also determining when the conjectured polarity upper bound is likely to hold, I would suggest spending a paragraph on prior work around interpolation.  2. Related to the above point, some more details around the Remark in Lines 139-145 would be helpful. It looks like for the non-constant term in the maximum of Thm 1, the numerator and denominator are empirical approximations of the numerator and denominator of $\eta$. Given that, could these converge to $\eta$ itself as $m \rightarrow \infty$? And if so, isn’t it possible for the Conjecture 1 to be violated as even values of $\eta > 4/3$ could satisfy the condition in Theorem 1? Let me know if I am misinterpreting something here.   3. More details of empirical analysis would be good to include in the main body. For instance, details about number of classifiers in each ensemble and whether the classifiers only differed in starting points (or other parameters as well) would be good to know while reading the figures themselves Technical Quality: 3 Clarity: 3 Questions for Authors: 1. On the point of polarity, it would be good to get more details around the discussion in the Remark on Lines 139-145. The specific questions I have on this are noted in the point above.  2. Regarding my point on empirical analysis, did the trained classifier ensemble consist of classifiers with just different starting points or were other training details/parameters changed? 3. Minor point, but $\delta$ notation is used in both Theorem 1 (for probability) and Theorems 3, 4 (for disagreement). And it looks like this notation is used in different contexts in these different results. Might be good to use different notations at different places if they are not related. Confidence: 2 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: Some limitations discussed in the conclusion section. Flag For Ethics Review: ['No ethics review needed.'] Rating: 7 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thank you for your positive appraisal of our work! [Weakness 1] Please see the general response. Interpolators are models that are trained to perfectly fit the train data, and appear prominently in the double descent literature, as well as prior explorations into the benefits of ensembling. [Weakness 2, Question 1] Thank you for pointing this out. It is true that the empirical approximation converges to $\eta$ itself. Here the bound becomes trivial with large $m$, as $\eta$ is clearly bounded by the maximum of $4/3$ and itself. We rephrased the theorem and conjecture and posted as a rebuttal for all reviewers above. We hope this clarifies the statement. Our conjecture follows from the observation that $P/S$ is less than $4/3$ (and hence $\max\{4/3, P/S\}$) for interpolators. To prove this in the theory is a fascinating open problem for us. [Weakness 3, Question 2] We will add more implementation details to Appendix C in the final version. The constituent classifiers differ in weight initialization and vary due to the randomized batches used during training. We see that the number of classifiers we used (mostly 5) is specified in Appendix C.1, but please kindly let us know if you feel we missed something. [Question 3] Thank you for pointing this out. We will correct this in the final version. --- Rebuttal Comment 1.1: Comment: Thanks to the authors for the clarifications. I will keep my score as it is but I would suggest moving some of the implementation details from the appendix to the main body if possible. --- Reply to Comment 1.1.1: Comment: We are glad that we could address your concerns and questions regarding our work. We will move some key implementation details to the main body in the final version.	Summary: The authors introduce polarity $\eta$ to characterize an ensemble of classifiers. Polarity is the probability that over half the models in an ensemble make an incorrect prediction divided by the expected square fraction of models that make an incorrect prediction. This quantity is bound by a concentration inequality, and the authors empirically demonstrate that this value tends to hover around 4/3. The authors then bound the ensemble error probability via a relation that involves polarity, and refine this bound based on further assumptions on the behavior of the ensemble. Finally, the authors show how to extrapolate small ensemble statistics to estimate error rates of larger ensembles. Strengths: The work is well motivated and clearly presented. The authors show that the theorems they derive can be used to estimate the performance of large ensembles (although further validation of this would strengthen the paper, see weaknesses). Weaknesses: It seems that Theorem 1 would hold if we replace 4/3 with any quantity that is larger than 4/3 as well, so theoretically it is unclear where this quantity really arises from and what prevents this quantity from being made smaller. Perhaps the authors can sketch out some intuition on this. Some additional experimental validation could help strengthen paper, since some of the results are empirical. For example, the experiments shown in Figure 3 could be extended to the other models discussed in the paper, and ideally should be performed multiple times to obtain a good idea of how reliable the estimates will be. Furthermore, some statistics could be reported (e.g. mean and variances for the distributions in Figure 1). However this is primarily theoretical work, so this is not a major concern. Technical Quality: 3 Clarity: 3 Questions for Authors: Figure 1 appears to show a positive correlation between error rate and polarity. I imagine this is driven by the numerator term in polarity. Could the authors comment on this, and whether they would expect the 4/3 rule to hold in a small or large error regime (e.g. $\epsilon$ or $0.5-\epsilon$?). For the experimental results shown in Figure 3, is it possible to plot some of the other bounds described in Section 2.2 for reference? The ``#’’ sign shows up in an equation on line 131, which was not defined. Is this an error? Confidence: 3 Soundness: 3 Presentation: 3 Contribution: 2 Limitations: Yes. Flag For Ethics Review: ['No ethics review needed.'] Rating: 7 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thank you for your positive review of our paper; we greatly appreciate your comments and questions. [Weakness 1] Please see the general response: we hope that the newly presented form of Theorem 1 addresses your concerns. In the original wording, we state that an $\eta$ greater than the stated lower bound is guaranteed. This is true, but one should pick the best possible $\eta$ for the task, which is what the newly worded theorem now states. $4/3$ in the theorem mainly arises from Lemma 1, equation (9). The notable point is that the value $4/3$ aligns well with the empirical result as shown in Figure 1. [Weakness 2] Thank you for your comments. Please refer to Figure C on the figure document attached to the rebuttal for all reviewers above. On KMNIST, which is an easier task than CIFAR-10, ResNet18 with various width (48, 64) were trained with batch size (32,64,128) to interpolate the train data. On Fashion-MNIST, ResNet18 with width 48 was trained with batch size 128 to interpolate the train data. [Question 1] Lemma 2 states that the majority vote error rate is smaller than $P(W_\rho > 1/2)$, the numerator term in the definition of polarity. If the error W_\rho(X) is smaller than 0.5 for all datapoints in the test set, the equality holds and the majority vote error rate $= P(W_\rho > 1/2) = 0$. This ensemble is $0$-polarized and thus $4/3$-polarized. We may have misunderstood your question. Could you explain a bit more about what you meant by 'small or large error regime (e.g. $\epsilon$ or $0.5-\epsilon$)'? [Question 2] We avoided overlaying those bounds (described in Section 2.2) as they are so far above the current set of axes that one cannot deduce the accuracy of our novel bounds. Please refer to Figure B in the figure document attached to the rebuttal for all reviewers above [Question 3] We used # as an indicator function 'f(A)=1 if A is true, otherwise 0'. For clarity, we will change this in the final version to a more common notation $1_A$. --- Rebuttal Comment 1.1: Comment: Thank you for your response. I remain unsure whether 4/3 is an intrinsic property of polarity of whether it is an artifact of the proof - I would need to spend more time with the paper, which unfortunately we are short on. I am satisfied with the new figures the authors have provided to address [Weakness 2] and I have updated my score to reflect this. On question 1: If you look at Figure 1 it seems as though there is a positive correlation. I was asking what you would expect to happen if the x-axis is expanded to 0-0.5 and populated with additional points in those unseen regimes. --- Rebuttal 2: Comment: Our current theory doesn’t suggest a _general_ positive correlation. This is reflected in our experiments: when comparing the CIFAR10 and CIFAR10-1 examples in Figure A ([figure_document](https://openreview.net/attachment?id=RvSCTfKipG&name=pdf)), the polarity of the interpolators hasn’t changed significantly, even though the majority vote error rate has increased by ~0.1 on average. However, we do not reject the possibility that a correlation might exist on a case-by-case basis. Our listed examples, Example 1 and Example 2 on page 4, highlight two extreme cases where this correlation is present. We are interested in conducting further investigations in future work. Regarding polarity, as both Figure 1 and Figure A ([figure_document](https://openreview.net/attachment?id=RvSCTfKipG&name=pdf)) suggest, and based on our efforts to make Lemma 1 reasonably sharp, we conjecture that 4/3 is not purely an artifact of the proof.	Rebuttal 1: Rebuttal: As there are overlapping comments regarding Theorem 1 and Conjecture 1, we provide clarification and a rephrased version of the theorem and conjecture here. Firstly, for Theorem 1, we recognize that the current wording may be misleading, so we reword the theorem in the following way for clarity: Theorem 1: Letting $$\eta = \max\left\\{ \frac{4}{3},\left(\frac{\sqrt{\frac{3}{8m}\log\frac{1}{\delta}}+\sqrt{\frac{3}{8m}\log\frac{1}{\delta}+4SP}}{2S}\right)^{2}\right\\}$$, the ensemble $\rho$ is $\eta$-polarized with probability at least $1-\delta$. No modifications to the proof are necessary, as this is simply a restatement of the finding obtained at the end of the proof. Definition of Interpolator, Interpolating: We define a neural net (or any other type of model) as an interpolator and say it is interpolating if it is trained to exactly fit the training data, i.e. $L_{train} = 0$. This term was recently popularized in Belkin et al.'s paper, "Reconciling Modern Machine-Learning Practice and the Classical Bias–Variance Trade-Off", which we now cite. Conjecture 1: An ensemble $\rho$ comprised of independently trained high-performing interpolating neural networks is $\eta$-polarized for $\eta \leq 4/3$. As some reviewers have pointed out, the previous wording suggested that interpolating neural network ensembles were $\eta$-polarized for $\eta$ precisely equal to $4/3$ and no lower. To clarify, we are asserting that $4/3$ suffices, as evidenced by Figure 1, which highlights that estimates for polarity (that is, $P/S$) are consistently below $4/3$ for interpolating models. We also wish to add that the "high-performing" moniker should be necessary here to ensure that $P$ is not too large in general. Pdf: /pdf/4cadd7631d6124579aaecc6083bd5bcbe412c80c.pdf	NeurIPS_2024_submissions_huggingface	2,024	null	null	null	null	null	null	null	null
Label Delay in Online Continual Learning	Accept (poster)	Summary: This paper outlines a new continual learning framework to handle label delays in data streams, a situation where labels lag behind data collection. Extensive testing revealed that neither increased computational resources nor advanced techniques like Self-Supervised Learning significantly overcome the challenges posed by label delays. To address this, the authors introduce Importance Weighted Memory Sampling, a method that effectively uses memory samples resembling new unlabeled data to bridge accuracy gaps caused by delays, achieving performance on par with non-delayed scenarios. Strengths: 1. I fully agree with the idea that the assumption of instantaneous annotation in Online Continual Learning (OCL) rarely holds in real-world applications. This makes the following discussions particularly interesting and relevant. 2. The proposed method is simple and easy to understand. Weaknesses: 1. While the discussion on the instantaneous nature of the annotation process is intriguing, I remain curious about the practicality of the proposed Online Continual Learning (OCL) model that incorporates delayed labels. In real-world scenarios, the challenge of matching input data at time step $t$ with labels at time step $t+d$ perfectly may prove more difficult than acquiring a data stream with instantaneous annotations. 2. The proposed setting is not clearly formulized. If the annotator can provide both the image and the corresponding label which makes the memory constraints and proposed label delay quite meaningless, simply train on data stream given by the annotator will be a better choice. 3. Unclear notations. The notation and evaluation criteria in the paper are somewhat unclear. Unless I am mistaken, it seems the model is assessed based on its performance on unlabeled training data. This approach may overlook inherent problems such as forgetting or performance degradation in Online Continual Learning (OCL), since ideally, the model should be evaluated across all previously encountered data distributions. A clearer definition of the OCL model mentioned in the paper is necessary. Additionally, there is notable existing research on online learning methods that handle delayed labels. A comparison with these methods would enrich the discussion and provide a more comprehensive evaluation of the proposed model's effectiveness in the context of OCL. 4. More comparisons with existing OCL methods should be included. 5. The paper does not provide a convincing comparison of its contribution with respect to existing work like[A,B]. [A]: Mesterharm, Chris. "On-line learning with delayed label feedback." International Conference on Algorithmic Learning Theory. Berlin, Heidelberg: Springer Berlin Heidelberg, 2005. [B]: Gomes, Heitor Murilo, et al. "A survey on semi-supervised learning for delayed partially labelled data streams." ACM Computing Surveys 55.4 (2022): 1-42. Technical Quality: 3 Clarity: 3 Questions for Authors: See the weakness. Confidence: 4 Soundness: 3 Presentation: 3 Contribution: 2 Limitations: N/A Flag For Ethics Review: ['No ethics review needed.'] Rating: 5 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We thank the reviewer for their efforts reviewing our paper and that the reviewer finds our method to be simple and that they agree on the setup of delayed labels. Q1: the challenge of matching input data at time step 𝑡 with labels at time step 𝑡+𝑑 perfectly may prove more difficult than acquiring a data stream with instantaneous annotations. Throughout our work we argue that modeling the delay is essential in OCL settings, and we found that a straightforward experimental model for demonstrating how such delays impact the model performance is the ($x_t, y_{t-d}$) formulation. Throughout our work, we argue that modeling delay is crucial in OCL settings. We found that a straightforward experimental model, using the (x_t, y_{t-d}) formulation, effectively demonstrates how delays impact model performance. Decoupling the input and label streams could be an interesting generalization, where data and labels arrive asynchronously. We excluded this to maintain focus on our main message. Including this discussion in future work would enhance the manuscript. However, we highlight the numerous experiments varying delay times (as in Figures 2, 3, 4), demonstrating IWMS's superiority over other methods. We do not expect performance rankings to differ significantly with variable delays. Q2: The proposed setting is not clearly formalized We will review our manuscript to ensure the notations and explanations are as clear as possible and will add clarifications where necessary to prevent misunderstandings. Feedback from other reviewers (bdVn, dj2M, and HwKE) indicated that our presentation was clear, but we understand that improvements can always be made. Please let us know any specific comments the reviewer has in mind. The Annotator does not provide the samples, only the labels. The scenario, to which the reviewer refers to, in which we simply wait for the annotator to reveal the labels and ignore the delays, is described as the Naïve case as this is the simplest and in practice the most popular approach. Q3: part 1: Unclear notations. The notation and evaluation criteria in the paper are somewhat unclear. We will review our manuscript to ensure the notations and explanations are as clear as possible and will add clarifications where necessary to prevent misunderstandings. Please let us know any specific comments you have in mind. We appreciate the reviewer's feedback on the evaluation criteria and the suggestion to include more comparisons with existing Online Continual Learning (OCL) methods. However, we would like to clarify both the original OCL formulation ([Cai] in Figure 3c and our manuscript does indeed address the evaluation of forgetting and performance degradation. Specifically, these aspects are thoroughly considered and illustrated in Figure 4 of the main paper, which evaluates the model's performance across all previously encountered data distributions. We apologize if this was not immediately clear and will ensure that the manuscript highlights these evaluations more prominently. OCL papers often focus less on forgetting and more on online accuracy, i.e., next batch accuracy, because the data stream is chronologically ordered. In this setup, older data (e.g., from 2010) always precedes newer data (e.g., from 2024) and might never reappear. For example, images of Nokia phones are far less common now than those of iPhones. Thus, the goal is to predict future data accurately. Despite this, we still report forgetting and past data accuracy, detailed in the supplementary material. Q3: part 2: A comparison with these online learning with label delay We would like to point the reviewer to the supplementary material, which explains the primary reasons for the incompatibility of traditional online learning approaches with modern OCL settings. Briefly, the main issue is the general lack of information retention mechanisms in traditional methods, which are crucial for addressing the complexities of continual learning tasks. Traditional online learning methods often assume immediate access to labels and do not account for the challenges of learning from data streams with delays. In contrast, modern OCL approaches, including our proposed framework, focus on retaining and utilizing past information effectively to mitigate forgetting and adapt to new data distributions. We have highlighted these distinctions in our paper and supplementary material to clarify why direct comparisons with traditional methods may not fully capture the unique challenges and contributions of our work. However, we will ensure that these discussions are more clearly articulated in the manuscript. Q4: More comparisons with existing OCL Could the reviewer point out which exact method should be included in our comparisons? Please note we used the best performing (under computationally budgeted) OCL as shown recently in [Ghunaim et al.]. Q5: No convincing comparison to [A, B] We appreciate the reviewer's suggestion to improve our comparison with prior work. Figures 16 and 17 in the Supplementary Material visually explain the core difference with the framework considered by [B]. In our framework, all labels become available after a fixed duration, making it fully supervised. Gomez et al. do not consider unsupervised and semi-supervised methods for cases where newer data distribution states are accessed using unsupervised techniques, which is our core contribution. We'll emphasize this distinction in the main paper. Regarding [A], we address the incompatibility issues of such online learning work from 2005 in the related work section, L66. The methods proposed by [A] describe how to wait for labels that may come in a different order than the data points, involving sorting labeled instances or updating on the most recent labels. Our new experiments show that simply rehearsing on the newest supervised data, as [A] suggests, leads to performance collapse (i.e., per chance accuracy). --- Rebuttal Comment 1.1: Title: Experimental results on [A] Comment: Briefly, the main issue is the general lack of information retention mechanisms in traditional methods, which are crucial for addressing the complexities of real world continual learning tasks, such as training a feature extractor that both learns new concepts faster (forward transfer) without losing the capability to perform well on already seen problems (backward transfer). To highlight that without rehearsing on memory samples the methods suffer significant performance degradation, we implemented the OL algorithm that is analogous to [A] in the special case in which all the labels (or feedback) arrives in order with a fixed constant delay. We ran new experiments (with identical experimental environment described in the main experimental section, Section 6) on the two largest datasets, CLOC (39M) and CGLM (580K), with computational budget $\mathcal{C}=2, 8$ respectively, for delay parameters $d=10, 50$ and $d=10, 50, 100$ respectively. The results show extreme underperformance: ### Online Accuracy of Online-Learning (no memory rehearsal) on CLOC \| Time Steps \| delay=10 \| delay=50 \| \| --- \| --- \| --- \| \| 5000 \| 0.195 \| 0.163 \| \| 15000 \| 2.142 \| 1.354 \| \| 25000 \| 2.960 \| 1.793 \| \| 40000 \| 3.467 \| 2.157 \| \| 50000 \| 4.202 \| 2.451 \| \| 60000 \| 4.838 \| 2.699 \| \| 75000 \| 5.238 \| 2.898 \| \| 85000 \| 5.632 \| 3.076 \| \| 95000 \| 5.849 \| 3.287 \| \| 105000 \| 6.265 \| 3.727 \| ### Online Accuracy of Online-Learning (no memory rehearsal) on CGLM \| Time Steps \| delay=10 \| delay=50 \| delay=100 \| \| --- \| --- \| --- \| --- \| \| 100 \| 0.000 \| 0.000 \| 0.000 \| \| 800 \| 0.463 \| 0.389 \| 0.263 \| \| 1500 \| 0.476 \| 0.319 \| 0.379 \| \| 2200 \| 0.531 \| 0.242 \| 0.257 \| \| 2900 \| 0.465 \| 0.196 \| 0.218 \| \| 3600 \| 0.459 \| 0.172 \| 0.179 \| \| 4300 \| 0.390 \| 0.188 \| 0.187 \| \| 5100 \| 0.419 \| 0.178 \| 0.158 \| \| 5900 \| 0.456 \| 0.253 \| 0.169 \| \| 6600 \| 0.504 \| 0.313 \| 0.175 \| The results clearly indicate the necessity of memory rehearsal: models on CLOC saturate at <6.5% for delay=10 and <4% for delay=50. In the case of CGLM dataset the performance collapses in all three delay scenarios <1%. --- Rebuttal Comment 1.2: Title: Responses Comment: I appreciate the authors taking the time to respond and help improve my understanding of the manuscript. From my perspective, both online accuracy during incremental learning and overall forgetting performance are crucially important metrics to evaluate continual learning methods, and their relative significance may depend on factors like the specific dataset and intended practical application (e.g. whether domain incremental, class incremental, or both). --- Rebuttal 2: Title: Response Comment: We thank the reviewer for their feedback. We fully agree with the claim that the significance of these metrics is _highly_ dependent on the exact application. In real world applications often times the final metric is high level, such as "customer satisfaction" or "viewer retention rate" that is difficult if not intractable, therefore these quantitative metrics are just proxies for such higher level objectives. What the research community deems important and relevant can, and indeed does change over time. Our message is simple: "regardless of the choice of the metrics, label delay is a very common problem that needs to be addressed". We truly appreciate the reviewer pointing out the dependency of the relative significance on the specifics of the actual problem. Regarding the notion of incremental learning, we would like to reiterate on our answer to the reviewer @HwKE, this is a positive signal that we will emphasise in the main manuscript: We argue that our experimental setup does not strictly fit the definition of the domain incremental learning, because it does have a changing distribution of the underlying labels as well, while in other benchmarks the labels distribution is static, such as Permuted MNIST, Rotated MNIST [An Empirical Investigation of Catastrophic Forgetting in Gradient-Based Neural Networks by Goodfellow et al.] and Clear [The CLEAR Benchmark: Continual LEArning on Real-World Imagery by Lin et al.] An outstandingly clear comparison is provided in Figure 1 of [CLAD: A realistic Continual Learning benchmark for Autonomous Driving by Verwimp et al.] that further refines the definitions of modern continual learning benchmarks. Finally, to ensure that the requirements of the reviewer are met, we would like to point out that we indeed report both metrics in our manuscript: - Online accuracy is reported in Figure 2, 3, 4 in the main manuscript and in Figure 5, 9, 11, 12, 13 and 18 in the supplementary material - Overall forgetting performance, mentioned by the reviewer is reported as the final score in Figure 19. The curve just gives a richer representation of the performance across all previous tasks. Please find the final scores below: ### CLOC Dataset \| Delay (d) \| ★ Naïve w/o delay \| ◆ Naïve \| ✚ S4L \| ✖ Pseudo-Label \| ▲ IWMS \| \|-----------------\|-------------------\|---------\|-------\|----------------\|--------\| \| d=10 (top) \| 6.9 % \| 6.0 % \| 6.4 % \| 6.7 % \| 6.6 % \| \| d=50 (middle) \| 6.9 % \| 6.2 % \| 6.9 % \| 6.8 % \| 6.9 % \| \| d=100 (bottom) \| 6.9 % \| 6.4 % \| 6.8 % \| 6.7 % \| 6.6 % \| ### CGLM Dataset \| Delay (d) \| ★ Naïve w/o delay \| ◆ Naïve \| ✚ S4L \| ✖ Pseudo-Label \| ▲ IWMS \| \|-----------------\|-------------------\|---------\|-------\|----------------\|--------\| \| d=10 (top) \| 32.2 % \| 34.2 % \| 28.8 % \| 33.1 % \| 57.2 % \| \| d=50 (middle) \| 32.2 % \| 25.0 % \| 33.1 % \| 31.8 % \| 56.2 % \| \| d=100 (bottom) \| 32.2 % \| 36.7 % \| 35.1 % \| 34.6 % \| 56.2 % \| ### FMoW Dataset \| Delay (d) \| ★ Naïve w/o delay \| ◆ Naïve \| ✚ S4L \| ✖ Pseudo-Label \| ▲ IWMS \| \|-----------------\|-------------------\|---------\|-------\|----------------\|--------\| \| d=10 (top) \| 53.6 % \| 56.4 % \| 56.8 % \| 57.4 % \| 55.4 % \| \| d=50 (middle) \| 53.6 % \| 56.7 % \| 56.6 % \| 57.2 % \| 55.6 % \| \| d=100 (bottom) \| 53.6 % \| 54.4 % \| 57.4 % \| 56.6 % \| 54.3 % \| ### Yearbook Dataset \| Delay (d) \| ★ Naïve w/o delay \| ◆ Naïve \| ✚ S4L \| ✖ Pseudo-Label \| ▲ IWMS \| \|-----------------\|-------------------\|---------\|-------\|----------------\|--------\| \| d=10 (top) \| 97.2 % \| 96.8 % \| 97.1 % \| 97.3 % \| 97.0 % \| \| d=50 (middle) \| 97.2 % \| 97.3 % \| 94.0 % \| 96.4 % \| 97.4 % \| \| d=100 (bottom) \| 97.2 % \| 94.9 % \| 91.0 % \| 94.8 % \| 95.5 % \| --- Rebuttal Comment 2.1: Comment: I appreciate the authors' thoughtful responses and detailed explanations. It addressed most of my concerns. I'm raising my score to 5. --- Reply to Comment 2.1.1: Comment: We thank the reviewer for raising the score and appreciate their recognition of the value of our work.	Summary: This paper introduces the problem of label delay in Online Continual Learning (OCL), where a model is trained continually on an unlabelled stream where labels are revealed with a fixed delay. The problem becomes learning from semi-supervised data with the objective of quickly adapting to the unlabeled distribution of most recent data. After analyzing the impact of label delay, the authors introduce their method: Importance Weighted Memory Sampling. The main idea is to favor replaying past samples which share similar features representation and labels with current unlabeled samples. Eventually, the authors compare various flagship strategies from different connected domains (Self-Supervsied Semi-Supervised Learning, Test-Time Adaptation and pseudo-labeling) and show superior performances of their approach compared through experiments. Strengths: - The presentation is very clear and the paper is well-written. - The motivations of introducing such a realistic problem are clearly defined and I believe the problem of label delay to be interesting to the community. - Thorough experiments have been conducted, with a specific care given to computation budget, which is often overlooked in OCL - The idea behind the presented approach is interesting, simple and effective - The figures are informative and clear Weaknesses: Major Weaknesses 1. My main concern regarding this paper is the evaluation metric. I understand that Online Accuracy is the usual metric in Online Learning. However, in Continual Learning, the usual metric of interest is the Average Accuracy. The problem with Online Accuracy is that only the performances on the last task are considered, meaning that you could have terrible performances on previous tasks. In fact, given that the proposed approach favors samples which are the most similar to the current distribution, I would expect the model to improve on the current task at the cost of worse performances on previous tasks (since the corresponding samples are less likely to be replayed). I have seen the Figure 19 of the appendix regarding backward transfer, but I am unsure that I understand these values correctly. Backward transfer can often be negative, but I might be unfamiliar with those specific datasets. I would like the authors to either: - Clarify why is Online Accuracy the only metric considered. Is it not important to maintain performances on previous tasks too? Is the lastly observed distribution the only one that is important? - Show the average accuracy across all tasks in the main draft to ensure that the gain in performances in current task does not come at the cost of lower performances on previous tasks. 2. I wonder why the authors did not compare to existing Continual Learning methods in supervised and semi-supervised cases. Such methods have been cited in appendix (CaSSLE and SCALE). Other methods such as [1,2,3] could also be considered. 3. Why should you have a limited memory size? It seems to me that all data are stored anyway for labeling. 4. The authors did not discuss the potential limitations of their approach. 5. The code is not accessible yet. Minor Weaknesses 6. To my understanding, a similar computational budget does not imply a similar training time. For example, one pass training with Cross-entropy is much faster than one pass with contrastive losses, as they require computing the Gram matrix. it could be enlightening to include training time of various compared methods. 7. This setup is a domain incremental learning setup, I believe this should be stated in the paper. 8. I don't think the memory buffer $M$ is defined in the text. If the authors can address my main concerns I would happily raise my score. Typos l191: One -> on ? l193 Is the memory really $2^{19}$? [1] Michel, Nicolas, et al. "Contrastive learning for online semi-supervised general continual learning." 2022 IEEE International Conference on Image Processing (ICIP). IEEE, 2022. [2] Wang, Liyuan, et al. "Ordisco: Effective and efficient usage of incremental unlabeled data for semi-supervised continual learning." Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2021. [3] He, Jiangpeng, and Fengqing Zhu. "Unsupervised continual learning via pseudo labels." International Workshop on Continual Semi-Supervised Learning. Cham: Springer International Publishing, 2021. Technical Quality: 3 Clarity: 4 Questions for Authors: See weaknesses. Confidence: 4 Soundness: 3 Presentation: 4 Contribution: 3 Limitations: I believe the authors have not addressed the limitations of their work. Please see weaknesses. Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We thank the reviewer for their recognition of our presentation, the problem we study and the solution we proposed, and appreciate the reviewer for the valuable suggestions. Here are our responses to the reviewer's concerns: ## Q1: Evaluation metric Using the Online Accuracy metric, in online continual learning, appears to be the common practice in most works, see [Online Continual Learning with Natural Distribution Shifts by Cai et al.] [Real-time evaluation in online continual learning: A new hope by Ghuneim et al] [Computationally budgeted continual learning: What does matter? by Prabhu et al] [Rapid Adaptation in Online Continual Learning by Hammoud et al.] . While evaluation on past tasks has been reported, they were often reported in supplementary as the main focus was the most recent task with OCL. The reason is among the motivation prior works provide, [Cai et al.], is that the top priority for a time-ordered stream is the ability to predict current task and future task (note that evaluation is done before training at each time step). Performance on past tasks is not relevant in case those tasks (which happens to be from the past) might never be presented again (think of images of old laptops or images of devices (for classification) that are no longer popular these days nor will they be in the future.). Thus the majority of works considering OCL, they investigate it on streams that are time ordered (unlike classical offline CL), where performance on next batch accuracy (online accuracy) is the top priority while still reporting the past task accuracies for completion is supplementary. To reiterate on our results, the reviewer can find the performance on past tasks metric described and evaluated on all settings of the main experiment in Section A.13, Figure 19. The backward transfer in Figure 19 is computed by the average accuracy of the last model on previous data following the evaluation protocol in [Online Continual Learning with Natural Distribution Shifts by Cai et al.] ## Q2: Comparison to existing Continual Learning methods As mentioned in line xx, our work is built on top of previous works of CL with fixed budget [Real-time evaluation in online continual learning: A new hope by Ghuneim et al] [Computationally budgeted continual learning: What does matter? by Prabhu et al] and delayed evaluation [Rapid Adaptation in Online Continual Learning by Hammoud et al.]. As consistently shown in these works, other existing semi-supervised CL methods like CaSSLE and most existing supervised CL methods like regularization based methods are not as effective as ER [On Tiny Episodic Memories in Continual Learning by Chaudhry et al]. [1] is a contrastive learning based method and [3] is a pseudo-labeling based method. We have tried our best to tune these two types of methods and include more advanced variants of them in section 6, although the modified version might be a little bit different from the two suggested ones. In conclusion, despite best efforts, since the computational complexity is constrained these methods failr terribly. This is already well documented in several other prior art that conducts these experiments under fixed computation [Rapid Adaptation in Online Continual Learning by Hammoud et al.] [GDumb: A Simple Approach that Questions Our Progress in Continual Learning by Prabhu et al.] where more sophisticated methods generally struggle. Furthermore, prior to this discussion, we tried using the publicly available implementation of [2] and failed to reproduce their results. ## Q3: Limited memory size We agree with the reviewer that we should not limit the buffer size since all data are stored anyway, and this is what we set for the experiments for small datasets, such as FMoW, Wild-Time and CGLM in our experiments section. The current buffer size set is more due to the experiment time and I/O time. We have tried our best to set the buffer size to be $2^{19}$ (line 193), which will be enough to store all the data in a small dataset like Yearbook, FMoW, and CGLM, but not enough for CLOC(39M, Supplementary A1). We further have buffer size analysis in supplementary A14, showing that our methods is robust to different buffer sizes and the performance is not relied on large buffer size. ## Q4: Potential limitations and code accessibility We agree with the reviewer that we should discuss the potential limitations of our approach. We will add this in the final version of the paper. We will also make our code accessible after the paper is accepted. ## Q5: Training time We agree with the reviewer that the training time can be influenced by various factors. Our current computation budget is based on the number of forward-backward passes, which is generally a good proxy for time and it has been widely used in CL literature[8,9]. While the training time metric could be sensitive to multiple other factors, e.g. code optimization, hardware, data I/O speed, and implementation. Here we add the training time for our method, ER, constrastive learning-based method, pseudo-labeling based method, and TTA methods with the same number of forward-backward pass in Table 1. Most of the experiments are using a single A100 GPU with 12 CPU. The training time is measured in hours. This table shows that the training time is could be entirely different due to various other factors. \| \| CLOC \| CGLM \| FMoW \| Yearbook \| \|---\|:---:\|:---:\|:---:\|:---:\| \| Naive \| 52 \| 3.6 \| 2.3 \| 0.2 \| \| ReSSL \| 67.3 \| 3 \| 4 \| 0.25 \| \| CoTTA \| 39 \| 5 \| 2.5 \| 0.2 \| \| Pseudo Labeling \| 111.3$^1$ \| 4.6 \| 2.5 \| 0.15 \| \| IWMs \| 61 \| 3.6 \| 3.5 \| 0.2 \| $^1$ the CPU allocation was 6 --- Rebuttal Comment 1.1: Title: Additional comments Comment: > The code is not accessible yet. Thank you for requesting the code, we believe that publishing code will help improve transparency reproducibility. To that end, following the NeurIPS 2024 Authors' guideline, we have submitted an anonymised link to the following two codebases: - an online interactive demo written in JavaScript, using the webcam as the input data stream in which we visualise how our experimental framework defines label delay and how IWMS selects samples from the buffer - the original experimental framework that was used for the entirety of the project. We were logging every experiment in Weights and Biases for reproducibility, however we could not transfer the project to an anonymous user to share it. Nevertheless, upon acceptance, we will make the project, with all our findings publicly accessible. > I don't think the memory buffer $M$ is defined in the text. We thank the reviewer for pointing this detail out. Our implementation of the memory is identical to [Online Continual Learning with Natural Distribution Shifts by Cai et al.]. We will update this in the paper alongside with the mentioned typos. > This setup is a domain incremental learning setup, I believe this should be stated in the paper. We argue that our experimental setup does not strictly fit the definition of the domain incremental learning, because it does have a changing distribution of the underlying labels as well, while in other benchmarks the labels distribution is static, such as Permuted MNIST, Rotated MNIST [An Empirical Investigation of Catastrophic Forgetting in Gradient-Based Neural Networks by Goodfellow et al.] and Clear [The CLEAR Benchmark: Continual LEArning on Real-World Imagery by Lin et al.] An outstandingly clear comparison is provided in Figure 1 of [CLAD: A realistic Continual Learning benchmark for Autonomous Driving by Verwimp et al.] that further refines the definitions of modern continual learning benchmarks. --- Rebuttal 2: Comment: I thank the authors for taking the time to address my concerns. I have read their rebuttal carefully. Evaluation Metric I get the authors point, although I respectfully disagree. The Online Accuracy is not the common practice in most work in OCL. See [1,2,3,4,5,6,7,8] for examples. Now, I understand that the metric is indeed very dependent on the specific application and that in your case of study you focus on Online Accuracy. While this makes sense, I believe this should be clarified in the manuscript, since I do not think that Online Accuracy is obviously the most important metric. Regarding Figure 19, I do not think this is the standard definition of Backward Transfer, or at least it differs from the definition of [1]. In any case, if my understanding is correct, IWMS indeed improves Online Accuracy at the cost of some marginal performance drop on previous tasks for specific cases (FMoW dataset). This does not in any case diminish this work's quality, although I believe that this potential drawback of the approach could be more clearly stated in the manuscript, in a limitation section for example. Comparison to existing OCL methods While your work is based on previous studies, their findings regard the fully supervised scenario. I still believe that including some existing work on Semi-Supervised OCL in your comparison would improve the manuscript. About the setup Thank you for clarifying. I understand the setup more clearly now. However, I am not sure to understand, if new classes appear in the stream, but are unlabeled yet, how do you predict them? To my understanding, you would not predict them before they are labeled. Training time Thank you for sharing the training time, I believe including it in the appendix would improve the paper. Code Thank you for sharing the code with a live demo. [1] Mai, Zheda, et al. "Online continual learning in image classification: An empirical survey." Neurocomputing 469 (2022): 28-51. [2] Guo, Yiduo, Bing Liu, and Dongyan Zhao. "Online continual learning through mutual information maximization." International conference on machine learning. PMLR, 2022. [3] Koh, Hyunseo, et al. "Online boundary-free continual learning by scheduled data prior." The Eleventh International Conference on Learning Representations. 2023. [4] Wang, Maorong, et al. "Improving Plasticity in Online Continual Learning via Collaborative Learning." Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2024. [5] Wei, Yujie, et al. "Online prototype learning for online continual learning." Proceedings of the IEEE/CVF International Conference on Computer Vision. 2023. [6] Gu, Yanan, et al. "Not just selection, but exploration: Online class-incremental continual learning via dual view consistency." Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2022. [7] Buzzega, Pietro, et al. "Dark experience for general continual learning: a strong, simple baseline." Advances in neural information processing systems 33 (2020): 15920-15930. [8] Prabhu, Ameya, Philip HS Torr, and Puneet K. Dokania. "Gdumb: A simple approach that questions our progress in continual learning." Computer Vision–ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part II 16. Springer International Publishing, 2020. --- Rebuttal 3: Comment: We thank the reviewer for engaging in this thoughtful conversation and bringing our attention towards richer representations of performance in OCL settings. We would like to respond to the raised points in order: ## First point > The Online Accuracy is not the common practice in most work in OCL. See [1,2,3,4,5,6,7,8] for examples. ### Comment on [1] In the referenced paper [1], under the "Evaluation Metrics" Average Accuracy is defined as follows: $$\mathrm{Average Accuracy}(A_i)=\frac{1}{i}\sum_{j=1}^ia_{i,j}$$ which is closely related to Online Accuracy, but we appreciate the fact that in such scenario performance on past iterations are re-evaluated in every step. In our setting evaluating and reporting this metric would be infeasible, as the number of steps on the CLOC, CGLM and FMoW dataset is 3-to-4 orders of magnitude larger than the examples provided in the survey (the maximum number of steps in [1] is 20, for comparison in our experiments the maximum number of "tasks" is 296 119). > Regarding Figure 19, I do not think this is the standard definition of Backward Transfer, or at least it differs from the definition of [1]. We truly appreciate the reviewer pointing out the difference in the definition. We recognise that the backward and forward transfer can, and should be reported (as defined in [1]) at given points during the training. To this end we have re-run the Naïve and IWMS experiments with $d=50$ on the two datasets where IWMS was performing the _best_ and the _worst_ to provide a full comparison against the baseline. We simplified the table representation by splitting the validation data into 100 equal sized ranges along the time axis, such that the ranges would correspond to the training data range: --- Rebuttal Comment 3.1: Comment: #### Accuracy matrix of Naïve on Yearbook @ $d=50$ \| Accuracy \| $te_{0 \rightarrow 12 }$ \| $te_{12 \rightarrow 25 }$ \| $te_{25 \rightarrow 37 }$ \| $te_{37 \rightarrow 50 }$ \| $te_{50 \rightarrow 62 }$ \| $te_{62 \rightarrow 75 }$ \| $te_{75 \rightarrow 87 }$ \| $te_{87 \rightarrow 100 }$ \| \| -------- \| ------------------------ \| ------------------------- \| ------------------------- \| ------------------------- \| ------------------------- \| ------------------------- \| ------------------------- \| -------------------------- \| \| $tr_{ 0 \rightarrow 12 }$ \| 0.99 \| 0.98 \| 0.94 \| 0.76 \| 0.56 \| 0.70 \| 0.89 \| 0.88 \| \| $tr_{ 12 \rightarrow 25 }$ \| 0.98 \| 0.99 \| 0.97 \| 0.79 \| 0.58 \| 0.72 \| 0.88 \| 0.86 \| \| $tr_{ 25 \rightarrow 37 }$ \| 0.99 \| 0.99 \| 0.96 \| 0.75 \| 0.55 \| 0.59 \| 0.77 \| 0.86 \| \| $tr_{ 37 \rightarrow 50 }$ \| 0.99 \| 1.00 \| 0.99 \| 0.83 \| 0.65 \| 0.74 \| 0.87 \| 0.90 \| \| $tr_{ 50 \rightarrow 62 }$ \| 0.94 \| 0.96 \| 0.96 \| 0.88 \| 0.88 \| 0.94 \| 0.96 \| 0.94 \| \| $tr_{ 62 \rightarrow 75 }$ \| 0.97 \| 0.99 \| 0.99 \| 0.96 \| 0.93 \| 0.93 \| 0.97 \| 0.97 \| \| $tr_{ 75 \rightarrow 87 }$ \| 0.99 \| 0.99 \| 0.99 \| 0.96 \| 0.93 \| 0.93 \| 0.96 \| 0.96 \| \| $tr_{ 87 \rightarrow 100 }$ \| 0.99 \| 0.99 \| 1.00 \| 0.95 \| 0.93 \| 0.96 \| 0.98 \| 0.97 \| #### Accuracy matrix of IWMS on Yearbook @ $d=50$ \| Accuracy \| $te_{0 \rightarrow 12 }$ \| $te_{12 \rightarrow 25 }$ \| $te_{25 \rightarrow 37 }$ \| $te_{37 \rightarrow 50 }$ \| $te_{50 \rightarrow 62 }$ \| $te_{62 \rightarrow 75 }$ \| $te_{75 \rightarrow 87 }$ \| $te_{87 \rightarrow 100 }$ \| \| -------- \| ------------------------ \| ------------------------- \| ------------------------- \| ------------------------- \| ------------------------- \| ------------------------- \| ------------------------- \| -------------------------- \| \| $tr_{ 0 \rightarrow 12 }$ \| 0.98 \| 0.99 \| 0.92 \| 0.74 \| 0.53 \| 0.71 \| 0.86 \| 0.86 \| \| $tr_{ 12 \rightarrow 25 }$ \| 0.99 \| 0.99 \| 0.97 \| 0.78 \| 0.57 \| 0.69 \| 0.86 \| 0.86 \| \| $tr_{ 25 \rightarrow 37 }$ \| 0.99 \| 0.99 \| 0.96 \| 0.73 \| 0.54 \| 0.61 \| 0.78 \| 0.86 \| \| $tr_{ 37 \rightarrow 50 }$ \| 0.99 \| 1.00 \| 0.99 \| 0.82 \| 0.64 \| 0.74 \| 0.87 \| 0.90 \| \| $tr_{ 50 \rightarrow 62 }$ \| 0.97 \| 0.98 \| 0.98 \| 0.91 \| 0.89 \| 0.94 \| 0.96 \| 0.94 \| \| $tr_{ 62 \rightarrow 75 }$ \| 0.99 \| 0.99 \| 0.99 \| 0.96 \| 0.94 \| 0.95 \| 0.98 \| 0.97 \| \| $tr_{ 75 \rightarrow 87 }$ \| 0.99 \| 1.00 \| 0.99 \| 0.96 \| 0.94 \| 0.95 \| 0.97 \| 0.96 \| \| $tr_{ 87 \rightarrow 100 }$ \| 0.99 \| 1.00 \| 1.00 \| 0.95 \| 0.92 \| 0.95 \| 0.98 \| 0.97 \| #### Accuracy matrix of Naïve on CGLM @ $d=50$ \| Accuracy \| $te_{0 \rightarrow 16 }$ \| $te_{16 \rightarrow 33 }$ \| $te_{33 \rightarrow 50 }$ \| $te_{50 \rightarrow 66 }$ \| $te_{66 \rightarrow 83 }$ \| $te_{83 \rightarrow 100 }$ \| \| -------- \| ------------------------ \| ------------------------- \| ------------------------- \| ------------------------- \| ------------------------- \| -------------------------- \| \| $tr_{ 0 \rightarrow 16 }$ \| 0.10 \| 0.27 \| 0.09 \| 0.07 \| 0.06 \| 0.05 \| \| $tr_{ 16 \rightarrow 33 }$ \| 0.14 \| 0.29 \| 0.24 \| 0.11 \| 0.11 \| 0.08 \| \| $tr_{ 33 \rightarrow 50 }$ \| 0.21 \| 0.38 \| 0.37 \| 0.26 \| 0.17 \| 0.13 \| \| $tr_{ 50 \rightarrow 66 }$ \| 0.23 \| 0.39 \| 0.39 \| 0.35 \| 0.25 \| 0.15 \| \| $tr_{ 66 \rightarrow 83 }$ \| 0.25 \| 0.40 \| 0.41 \| 0.38 \| 0.34 \| 0.22 \| \| $tr_{ 83 \rightarrow 100 }$ \| 0.15 \| 0.26 \| 0.26 \| 0.24 \| 0.24 \| 0.19 \| #### Accuracy matrix of IWMS on CGLM @ $d=50$ \| Accuracy \| $te_{0 \rightarrow 16 }$ \| $te_{16 \rightarrow 33 }$ \| $te_{33 \rightarrow 50 }$ \| $te_{50 \rightarrow 66 }$ \| $te_{66 \rightarrow 83 }$ \| $te_{83 \rightarrow 100 }$ \| \| -------- \| ------------------------ \| ------------------------- \| ------------------------- \| ------------------------- \| ------------------------- \| -------------------------- \| \| $tr_{ 0 \rightarrow 16 }$ \| 0.15 \| 0.39 \| 0.13 \| 0.11 \| 0.09 \| 0.08 \| \| $tr_{ 16 \rightarrow 33 }$ \| 0.24 \| 0.51 \| 0.45 \| 0.20 \| 0.17 \| 0.15 \| \| $tr_{ 33 \rightarrow 50 }$ \| 0.31 \| 0.55 \| 0.61 \| 0.41 \| 0.25 \| 0.20 \| \| $tr_{ 50 \rightarrow 66 }$ \| 0.35 \| 0.57 \| 0.63 \| 0.59 \| 0.38 \| 0.24 \| \| $tr_{ 66 \rightarrow 83 }$ \| 0.38 \| 0.59 \| 0.64 \| 0.62 \| 0.60 \| 0.32 \| \| $tr_{ 83 \rightarrow 100 }$ \| 0.40 \| 0.60 \| 0.65 \| 0.64 \| 0.64 \| 0.53 \| --- Rebuttal 4: Comment: > In any case, if my understanding is correct, IWMS indeed improves Online Accuracy at the cost of some marginal performance drop on previous tasks for specific cases (FMoW dataset). It is correct, thank you for highlighting this. We will address the limitations on backward transfer on FMoW. ### Comment on [2, 5, 6, 7] We would like to point out that Guo et al. details the evaluation metric in 6.1 as follows: > We first learn from the data stream of all tasks for each dataset, and then test the final model using the test data of all tasks. We report the average accuracy of all tasks from 15 random runs for each dataset We would like to argue that this metric is taking an excessively strong measure to remove noise from the metric. In our experiments, we experienced that rerunning the same training with different seed results in negligible (less than 0.01%) differences in the results. Running the experiments 15 times to evaluate the metric of [2] is infeasible for us. This similarly hold for [5] as well, since they report the Average Accuracy and Average Forgetting across 15 runs. Furthermore, although the main manuscript of [6] does not provide the detail about re-running the Average Accuracy, the corresponding code is set by default to 15 re-runs with different seeds. (follow this URL for reference: https://github.com/YananGu/DVC/blob/6f12984d10a4a1c4609f221b939f93d94fc8258e/general_main.py#L29 ) In [7], the number of random initialization is dropped to 10, otherwise they report the Average Accuracy as well. ### Comment on [3] We would like to point out that [3] introduces their own metric: Knowledge Loss/Gain Ratio [3] claiming that the metrics used by [2, 5, 6] are relying on the notion of task boundaries therefore they define a new objective that is "appropriate for periodic data distribution". In our paper we cannot make such assumptions about periodicity. ### Comment on [4] The accuracy metric proposed by [4], Learning Accuracy (LA) using Model Plasticity is formally defined for the $j$-th task as: $$ l_j = a_j^j $$ where $a_j^i$ is the accuracy evaluated on the test set of task $j$ after training the network from task 1 to task $i$. We would like to argue that this metric is similar to the _Online Accuracy_ metric apart from the notion that here the test samples are drawn from a different distribution, whereas the Online Accuracy is evaluated on the $j$-th batch of data before it is used for training. If we assume that both the test and the training batch is drawn from the same distribution at time-step $j$, the two metrics are arguably the same. (Please note that the training is only done on the batch after the evaluation in the case of Online Accuracy.) ### Comment on [8] The paper _"GDumb: A simple approach that questions our progress in continual learning"_ highlights the weakness of the then-standard metrics in Continual Learning such as Average Accuracy and Accuracy at end. While the paper aimed to steer the CL research community towards higher standards, these metrics stayed popular. In fact, [Online Continual Learning with Natural Distribution Shifts by Zai et al.] uses GDumb's argument to propose the Online Accuracy metric that recently increasing number of work adopted. ### Resolution > Now, I understand that the metric is indeed very dependent on the specific application and that in your case of study you focus on Online Accuracy. While this makes sense, I believe this should be clarified in the manuscript, since I do not think that Online Accuracy is obviously the most important metric. We recognize and fully agree with the reviewer's criticism that our choice of metric needs more justification, as we present it as the "go-to" metric in our narrative. We will definitely address these points in the updated manuscripts more carefully, if the reviewer agrees. ## Second point > While your work is based on previous studies, their findings regard the fully supervised scenario. I still believe that including some existing work on Semi-Supervised OCL in your comparison would improve the manuscript. Due to time limitations we could not fully reimplement [Contrastive learning for online semi-supervised general continual learning. by Michel et al.] to report scores in the rebuttal period, but we are happy to include them in the final manuscript. For reference, while we think the comparison would indeed interesting, we remain skeptical about contrastive methods outperforming IWMS, or even the Naïve baseline under our experiments' computational constraints, as we discuss the results of 6 different contrastive methods in detail in the Supplementary Material A.8. --- Rebuttal Comment 4.1: Comment: ## Third point > However, I am not sure to understand, if new classes appear in the stream, but are unlabeled yet, how do you predict them? To my understanding, you would not predict them before they are labeled. In our setup the model has to make a prediction at every time-step, even if an unseen category is presented for the first time (which is a realistic scenario). In such cases it is indeed theoretically impossible to correctly predict the class (apart from the accidental random-chance correct guesses), and the model is not able to perform well on the new class until the labels arrive. In our setup we simply penalize the model, regardless whether a class was seen or not. This can be also verified in the live demo (press the "`" key to open the Online and the Current Accuracy curve panel). ## Final note We would like to thank the reviewer for their insightful comments and opening up the discussion. We hope that we have addressed all concerns in detail. --- Rebuttal 5: Title: Thank you for the detailed responses Comment: I would like to thank the authors for taking the time to address all my points and for their detailed response. I agree that Average Accuracy is not the only metric considered in the paper I referenced, but it is still the most commonly used one for comparison. I guess we could argue on which is the most important one depending on the problem at hand, and again I understand that Online Accuracy might make more sense in your context. Overall conclusion The authors have addressed my concerns and I will increase my score to 6.	Summary: This manuscript delineates a novel method, termed as IWMS, which is devised to address the problem of label delay in online continual learning, where new data may not be labeled due to slow and costly annotation processes. The IWMS exhibits an innovative usage of fine-grained Gaussian Mixture prototypes, along with mutual information optimizing that endow the proposed method with competitive performance for unsupervised class incremental discovery. Strengths: S1. The paper's significance is underscored by its empirical evaluation. The evaluation is comprehensive, with comparisons to different baselines and ablation study providing a compelling demonstration of the strength of this work. S2. The proposed IWMS method is simple and easy to implement. The proposed memory sampling has the potential to inspire future research in this domain. Weaknesses: W1. Although this paper presents a simple method, further technical insights regarding the implementation and each module within the IWMS method would be beneficial. Further efforts regarding technical innovation and methodological novelty would also be beneficial. W2. The manuscript could delve deeper into the memory buffer (e.g., how to construct and update it, and the additional complexity by it), a factor which could be pivotal for fully understanding the proposed model. W3. It would be better to analyze and discuss the difference between online learning, continual learning, and online continual learning, and further clarify the topic of this work. Technical Quality: 2 Clarity: 3 Questions for Authors: See weaknesses above. Confidence: 4 Soundness: 2 Presentation: 3 Contribution: 2 Limitations: Addressed. Flag For Ethics Review: ['No ethics review needed.'] Rating: 5 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We thank the reviewer's recognition for our extensive experiments and the simplicity of our proposed method and the value suggestions. The following are our responses to the reviewer's concerns: Q1: Need further explanation regarding the implementation and each module within the IWMS method and technical innovation. Our implementation of IWMs is closely related to the experimental setup in Section 5. It is built upon the best performing method, ER, and we further improve it by sampling the buffer samples that are more similar to the unlabeled data. All the implementation details and "technical innovation" are detailed in full in Section 1, 3 and 4, in Figure 1 and in Algorithm 1 and 2. Line 149 - provides "technical insights regarding the implementation and each module within the IWMS method". Line 154 and 163 - "delves deeper into the memory buffer". Section A.12, Line 756 - Analyses the difference between online learning continual learning and online continual learning. As we point it out in the main paper in Section 4, L137, the novelty lies in how to learn the unlabeled distribution. Our experiments in Section 6 demonstrate that existing literature, which focus on operating unlabeled features to learn the most recent distribution, are often neither effective nor efficient. Our approach avoids unnecessary computation on the most recent unlabeled distribution, even if it closely matches the evaluation distribution. Instead, we sample previously labeled data similar to the current unlabeled data. This allows us to construct a pseudo-distribution close to the most recent distribution and conduct purely supervised learning on the pseudo-distribution. Q2: Need further explanation regarding the memory buffer. This is all detailed in Section 4, Line 154, but we will reiterate it here for completion. At time step $t$, we keep a memory buffer to store both the raw data and the features of the labeled data from time step $t'$($t'$< $t$) computed in $t'$. We sample the data according to the similarity between the features from the current unlabeled data $x^t$ and the features in the buffer. We then take the buffer data whose features are most similar to x^t to construct a new labeled batch. This batch is our buffer batch. Q3: Need to analyze and discuss the difference between online learning, continual learning, and online continual learning, and further clarify the topic of this work. We already had a Section 2 and A.12 in the paper discussing this in full depth. In fact, the title of the section is "Online Learning vs Online Continual Learning" and "Considering catastrophic forgetting". We provide further explanation below: Online Learning and Online Continual Learning both involve learning from data arriving sequentially, but Online Learning typically deals with single-task streams, often assumed to be from an i.i.d. distribution. In contrast, Online Continual Learning is more concerned with non-stationary streams that undergo frequent changes in distribution. Continual Learning, on the other hand, is a broader concept that encompasses any type of learning process where the data is revealed sequentially and the distribution of the data may change over time. Recent discussion in continual learning has focused on the scenario where the computation is fixed for fair comparison. Our work mainly focuses on the online continual learning scenario, where the distribution of the data changes over time and the data is revealed batch-wise. As motivated in line 26, we further consider the label delay problem, where the labels of the data are not available immediately, and most recent data distribution is unlabeled. To this end, we propose solutions from various literatures and propose our simple yet effective method, IWMs, to address the label delay problem in online continual learning. --- Rebuttal Comment 1.1: Comment: We would like to ask if we managed to address the concerns of the reviewer?	Summary: The paper addresses the problem of label delay in online continual learning. The proposed framework explicitly models this delay, revealing unlabeled data from the current time step and labels delayed by a specific number of steps. Extensive experiments demonstrate that increasing computational resources alone is insufficient to overcome performance declines caused by significant label delays. The authors introduce Importance Weighted Memory Sampling, a robust method that prioritizes memory samples similar to the newest unlabeled data, improve the performance under label delays scenario. Strengths: - The paper is well-written and easy to follow. - The proposed method is comprehensive and reasonable. In Section 4, the IWMS addresses the critical issue of outdated feature updates. - The experiments conducted in the paper are thorough and convincing, covering a variety of datasets and including ablation studies, sensitivity analyses, and more. Weaknesses: 1. The discussion of related work on the label delay problem is not comprehensive. The authors should consider additional relevant studies in the field of online learning with label delay, such as: - Heliou et al. "Gradient-free Online Learning in Continuous Games with Delayed Rewards." ICML 2020. - Wan et al. "Online Strongly Convex Optimization with Unknown Delays." Machine Learning 2022. Additionally, recent works on "asynchronous labels" (which is a general version of label delay, in which both the feature and label can delay) should be better considered: - Zheng et al. "Asynchronous Stochastic Gradient Descent with Delay Compensation." ICML 2017. - Qian et al. "Learning with Asynchronous Labels." TKDD 2024. 2. The paper considers only the simple case of a fixed delay time step $d$, which may not be practical for real-world scenarios. It would be more beneficial if the authors considered a more general case where the delay time step is variable. 3. Although previous research on online learning with delayed feedback focuses on the online learning scenario rather than continual learning, it would be more comprehensive if the authors compared their method with these online learning approaches. Technical Quality: 3 Clarity: 3 Questions for Authors: See Weakness above. Confidence: 4 Soundness: 3 Presentation: 3 Contribution: 2 Limitations: See Weakness above. Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We appreciate the reviewers' positive comments on our method, experiments, and paper presentation. We address the above concerns here. Q1: The discussion of related work on the label delay problem is not comprehensive. We thank the reviewers for pointing out the missing references. We will include the suggested references in the final version of the paper, covering both the label delay and asynchronous labels scenarios. Q2: The paper considers only the simple case of a fixed delay time step d. We acknowledge that this is an excellent suggestion and plan to add further ablations on varying delay time steps during the rebuttal period. However, we would like to point out to the reviewer the numerous experiments we have conducted where we vary $d$ across all experiments (as in Figure 2,3,4), which already demonstrate the superiority of IWMS over other methods. We do not generally expect the performance ranking to differ significantly with variable delay times. Q3: Comparison to the online learning approaches We have a thorough discussion on this point in the supplementary A12. Furthermore, We appreciate the reviewer’s suggestion to compare our work with existing research on online learning methods that handle delayed labels. We would like to point the reviewer to the supplementary material A12 (page 23, L756), which explains the primary reasons for the incompatibility of traditional online learning approaches with modern OCL settings. Briefly, the main issue is the general lack of information retention mechanisms in traditional methods, which are crucial for addressing the complexities of real world continual learning tasks, such as training a feature extractor that both learns new concepts faster (forward transfer) without losing the capability to perform well on already seen problems (backward transfer). To highlight that without rehearsing on memory samples the methods suffer significant performance degradation, we implemented the OL algorithm that is analogous to the mentioned papers in the special case in which all the labels (or feedback) arrives in order with a fixed constant $d$ delay. We ran new experiments (with identical experimental environment described in the main experimental section, Section 6) on the two largest datasets, CLOC (39M) and CGLM (580K), with computational budget $\mathcal{C}=2,8$ respectively, for $d=10,50$ and $d=10, 50, 100$ respectively. The results show extreme underperformance: [Online Accuracy of Online-Learning (no memory rehearsal) on CLOC] \| Time Steps \| delay=10 \| delay=50 \| \| --- \| --- \| --- \| \| 5000 \| 0.195 \| 0.163 \| \| 15000 \| 2.142 \| 1.354 \| \| 25000 \| 2.960 \| 1.793 \| \| 40000 \| 3.467 \| 2.157 \| \| 50000 \| 4.202 \| 2.451 \| \| 60000 \| 4.838 \| 2.699 \| \| 75000 \| 5.238 \| 2.898 \| \| 85000 \| 5.632 \| 3.076 \| \| 95000 \| 5.849 \| 3.287 \| \| 105000 \| 6.265 \| 3.727 \| [Online Accuracy of Online-Learning (no memory rehearsal) on CGLM] \| Time Steps \| delay=10 \| delay=50 \| delay=100 \| \| --- \| --- \| --- \| --- \| \| 100 \| 0.000 \| 0.000 \| 0.000 \| \| 800 \| 0.463 \| 0.389 \| 0.263 \| \| 1500 \| 0.476 \| 0.319 \| 0.379 \| \| 2200 \| 0.531 \| 0.242 \| 0.257 \| \| 2900 \| 0.465 \| 0.196 \| 0.218 \| \| 3600 \| 0.459 \| 0.172 \| 0.179 \| \| 4300 \| 0.390 \| 0.188 \| 0.187 \| \| 5100 \| 0.419 \| 0.178 \| 0.158 \| \| 5900 \| 0.456 \| 0.253 \| 0.169 \| \| 6600 \| 0.504 \| 0.313 \| 0.175 \| The results clearly indicate the necessity of memory rehearsal: models on CLOC saturate at <6.5% for delay=10 and <4% for delay=50. In the case of CGLM dataset the performance collapses in all three delay scenarios <1%. --- Rebuttal Comment 1.1: Comment: The authors' responses have addressed my concerns, so I have decided to raise my score. I would be happy to see a more comprehensive discussion of related works in the revised version! --- Reply to Comment 1.1.1: Comment: We thank the reviewer for recognizing the value of our proposal. We will provide an in depth comparison to the papers discussed in the rebuttal period in the final version.	null	NeurIPS_2024_submissions_huggingface	2,024	null	null	null	null	null	null	null	null
Are More LLM Calls All You Need? Towards the Scaling Properties of Compound AI Systems	Accept (poster)	Summary: LLM inference systems often generate multiple answers (function calls) to a query and then aggregate the answers with rules like vote, filter-vote. This paper investigates how the number of function calls influence the influence of performance of the compound system. More concretely, this paper found that there often exist a U-shape overall performance curve with an increasing number of function calls where easy queries benefit from more function calls and hard queries can be harmed by the function calls. Some theoretical results are given for toy settings and an optimal number of function calls can be derived accordingly. Empirical results are given for simple settings. Strengths: - Clarity: The paper is mostly clear and easy to understand. - Originality: The problem is less studied in the literature. Weaknesses: 1. The results are not surprising. The problem is quite simple and straightforward if we drop the LLM background. Basically, we are drawing samples uniformly at random from a distribution (function calls) and then apply majority vote to them, then with more samples, the uniform + majority vote leads to select the one sample with highest probability (if majority exists). If the majority is the correct answer, then the sample is "easy", otherwise it is "hard". Mixing easy and hard majority naturally leads to U-shape in the performance curve. 2. The model prediction may not outperform simple hyperparameter-search. The "optimal" function calls prediction is based on strong assumptions which may not hold true and estimate accurately in practice. Instead, simple hyperparameter search is usually much more robust performance. Technical Quality: 3 Clarity: 2 Questions for Authors: As stated in the weakness. Confidence: 3 Soundness: 3 Presentation: 2 Contribution: 2 Limitations: Not applicable. Flag For Ethics Review: ['No ethics review needed.'] Rating: 5 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We thank the reviewer for the feedback. We answer the questions as follows. *The results are not surprising: We appreciate your comment but we would like to respectfully argue that our results are interesting. First, many practitioners and researchers believe that they can consistently improve performance with more LLM calls (e.g. https://arxiv.org/abs/2402.05120). In this context, our finding that more LLM calls can hurt performance in specific settings is unexpected and a useful contribution to the literature. Furthermore, we provide an explanation of this counterintuitive phenomenon based on query difficulty and provide both mathematical and empirical justifications. The prediction based on our theory matches the empirical performance accurately. These results have not been not studied in existing work and opened the door to understanding and optimizing the design of compound LLM systems. model prediction is based on strong assumption: Thank you for this question. We would like to clarify that the algebraic formula (K) shown in Theorem 4 is used for a special case. On real-world datasets, we use the scaling law function empirically fitted on a small dataset for prediction purposes, which does not require strong assumptions. The scaling law function’s predictions are empirically accurate (see Figure 4(c)). We will highlight this in the revised paper. *Comparison with hyper-parameter search algorithms: Thank you very much for pointing this out! We conducted additional experiments to compare our proposed scaling laws with two popular hyper-parameter search algorithms, bayesian optimization and tree structured parzen estimator, with the same number of performance evaluation budgets. Specifically, we evaluate these hyper-parameter search algorithms on GPQA, AVERITEC, and MMLU PHYSICS, with 5 calls to the performance oracle (the same input to our scaling law method). As shown in the following table, our scaling law reaches the most accurate prediction and overall accuracy. We will add this discussion in the revised version. Notably, both the Parzan and Bayesian optimizers would have recommended the users to use a large number of LLM calls (>100), which would have incurred much more cost while performing worse than our estimator. \| Method \| GPQA \| \| AVERITEC \| \| MMLU PHYSICS \| \| \|:---------------------------------:\|:-----------:\|:-------:\|:-----------:\|:-------:\|:------------:\|:-------:\| \| \| LLM Calls \| Acc \| LLM Calls \| Acc \| LLM Calls \| Acc \| \| Our scaling law* \| 19 \| 0.317 \| 5 \| 0.367 \| 13 \| 0.542 \| \| Tree-structured parzen estimator \| 444 \| 0.307 \| 114 \| 0.361 \| 418 \| 0.530 \| \| Bayesian optimization \| 430 \| 0.307 \| 377 \| 0.360 \| 47 \| 0.536 \| \| Ground truth \| 13 \| 0.320 \| 4 \| 0.368 \| 10 \| 0.543 \| --- Rebuttal Comment 1.1: Title: Response Comment: I have read the authors responses and believe that the second and third point addressed my concerns. As for the first one, I am still not very convinced that theory contribution is significant as it seems to me to rediscover the math (the non-peer reviewed works also are not convincing too). Yet I am willing to increase the score. --- Reply to Comment 1.1.1: Title: Thank you for your response and increasing the score! Comment: Thank you for your response and increasing the score!	Summary: This paper evaluates the LM calls and the task performance based on two compound system designs: Vote and Filter-Vote, for performing multiple-choice selection tasks. The authors conduct theoretical analysis for the system designs and the scaling behavior by proposing formal notion of query difficulty and model behavior. They suggest that adding LM calls does not always lead to monotonic task performance. Basically, additional LM calls increase performance on easy queries and decrease performance on difficult queries; increasing the number of LM calls can lead to non-monotone behavior depending on task difficulty. The authors also conduct empirical analysis on different datasets including one synthetic dataset to validate the claims, and notice that the analytical scaling model can accurately predict the performance of Vote and Filter-Vote. Strengths: 1. The paper offers a novel insights to the relation of task difficulty, LM calls, model behavior. 2. The overall writing of the paper is very easy to follow. Throughout the writing, the paper has made the main scope clear. 3. Both the theoretical analysis and empirical experimentation are well conducted to support the main claim. Weaknesses: Although I found that the authors have made solid empirical experimentation to support the claim, the empirical results are not well presented. The authors experimented with several different datasets, but only showed a few case studies on a certain dataset in the main body of the paper. I would at least present the overall results of all datasets in the experiment section if the page limit permits. Technical Quality: 3 Clarity: 4 Questions for Authors: 1. Since all your datasets are multiple choice tasks, how did you extract the results of the LLM outputs for answer mapping? Perhaps the token logits? If that is the case, did you check the robustness of the results of the token logits comparing to the real textual outputs? 2. Did you (Or do you think whether it is worthwhile to) conduct additional robustness tests (e.g. by shuffling the option orders, or adding input perturbations) to validate the LLM outputs? Recent evaluation papers that could be relevant, they more or less discuss that the difficulty of the tasks could affect model robustness: https://aclanthology.org/2024.naacl-long.295/ https://aclanthology.org/2024.findings-naacl.130/ Confidence: 3 Soundness: 3 Presentation: 4 Contribution: 3 Limitations: The authors already talk about the Limitations in Appendix B. Flag For Ethics Review: ['No ethics review needed.'] Rating: 8 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We thank the reviewer for the helpful feedback and support of our paper! Please see below for our response. *present the overall results of all datasets in the experiment section if the page limit permits: Thanks for your suggestion! We will move more empirical results to the experiment section in the revised paper. answer extraction: We prompted the LLM to generate its final output in a specific format (i.e., “the answer is (X)”) and then used regular expression to extract the output. Did you (Or do you think whether it is worthwhile to) conduct additional robustness tests (e.g. by shuffling the option orders, or adding input perturbations) to validate the LLM outputs?: Thank you very much for pointing this out! Robustness studies are orthogonal but complementary to the focus of this paper. For example, it would be interesting to study how input perturbation affects query difficulty distribution. We will add a discussion in the revision. Recent evaluation papers*: Thanks for the reference. We will add a discussion to them in the revised paper. --- Rebuttal 2: Title: Thanks for response Comment: Dear authors, thanks for the response. I think this is overall very interesting work and my initial score should properly reflect the quality of this work. And I would like to see it included to the proceedings. One final remark: thanks to the author for pointing the use of regex to extract the response and I would suggest the author include the details of the extraction process in the appendix, as well as the rate of the missuccess (if any). Especially the template "the answer is (X)" reminds me of the title of a very recent paper from ACL this year that might be more or less relevant. Perhaps also worth reading: https://aclanthology.org/2024.findings-acl.441/ --- Rebuttal Comment 2.1: Title: Thank you for your response and support of our paper! Comment: Thank you for your response and support of our paper! We will make sure to include the answer extraction process in the appendix and add a discussion to the suggested ACL paper.	Summary: In this paper, the authors studied the scaling properties of compound inference systems. Theoretically and empirically, the authors answered the properties of multiple LM calls. Strengths: 1. Answered several important questions in multiple LM calls, which may benefit the compound inference systems. 2. Heuristic for Optimal Number of LM Calls. Weaknesses: 1. The authors only studied two simple natural inference system designs, i.e., vote and filter vote, which might reduce the significant of findings. Technical Quality: 3 Clarity: 2 Questions for Authors: 1. Is the optimal number of LM calls model-specific or not? Confidence: 4 Soundness: 3 Presentation: 2 Contribution: 3 Limitations: 1. According to definition 1, a query can only be categorized into easy and difficult according to the difficulty indicator. The is a simplified case of the real-world scenarios. Besides, the difficulty indicator is not model-agnostic, which may also reduce the significant of findings. 2. The optimal LM calls depends on query difficult while the difficult level is simply either difficult or easy. A more rational way is to measure the least number of calls for a query to get correct answer. Flag For Ethics Review: ['No ethics review needed.'] Rating: 5 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We thank the reviewer for the helpful feedback and support of our paper! Please see below for our response. *Simple System: We indeed focus on simple systems. There are two reasons. First, they both represent real-world systems. For example, the Cot@32 approach by Google Gemini is indeed a simple vote method. Second, the observed phenomenon is pretty surprising, and we want to understand it empirically and analytically in the simplest place it appears. We believe it is always important to find a minimal working example. Of course, it will be valuable to investigate in challening settings in future work. Is the optimal number of LM calls model-specific or not?: Thanks for this great question. We actually predict the optimal number of calls in two ways (K for a special case and the scaling law function G in general). They are both model-specific. This is an important point and we will emphasize it in the revision. *the least number of calls for a query to get correct answer*: Thank you so much for pointing this out. We think this is a great intuitive way to explain how to characterize query difficulty. It is very closely connected, in fact, to the characterization we currently use. It is a kind of soft version of our difficulty level. Indeed, let’s take Vote as an example. $$ d_V(x) = \max_a \Pr[G(x, \theta) = a] - \Pr[G(x, \theta) = y] $$ Here, the inverse of $Pr[G(x,\theta)=y]$ is the expected number of calls to obtain at least one correct answer, which is one formalization of the reviewer’s soft difficulty. We focus on the binary difficulty because it offers a natural explanation of the non-monotonic behavior. We appreciate the suggestion for providing this nice intuition and will put a sentence about it into the revision. --- Rebuttal Comment 1.1: Comment: Thanks for the clarification. While my concerns/questions are not fully addressed, I will keep my score. --- Reply to Comment 1.1.1: Title: Thank you for your response and support of our paper! Comment: Thank you for your response and support of our paper!	Summary: Recent state-of-the-art results in language tasks have been achieved using compound systems that make multiple calls to Language Models (LMs) and aggregate their responses. However, there is limited understanding of how the number of LM calls affects the performance of these compound systems. This paper studies the scaling properties of compound inference systems, specifically analyzing how the number of LM calls impacts the performance of two simple compound system designs, Vote and Filter-Vote, which use majority voting to aggregate LM responses, with or without applying filters. The analysis, both theoretical and empirical, reveals that the performance of Vote and Filter-Vote systems can initially improve but then decline as the number of LM calls increases. This non-monotonic behavior is attributed to the varying difficulty of queries within a task: more LM calls enhance performance on "easy" queries but reduce it on "hard" queries. Strengths: The findings of the authors are of definite interest, and the phenomenon that they identify is thought provoking, although mathematically intuitive. I think the paper has potential for good to high impact. The problem itself is conveyed very clearly, and the authors made the right choice in including the 'main contributions' box and Figure 1 on the first page of the paper itself. The theory and substance of the paper is strong, and worth the status of the conference. The study also provides a method to determine the optimal number of LM calls that maximizes system performance, based on a small number of samples, and develops an analytical scaling model for both systems. Experiments confirm that the scaling model can accurately predict the performance of Vote and Filter-Vote systems, identifying the optimal number of LM calls. Weaknesses: One weakness is in the insistence on using multiple choice benchmarks, while not being more expansive about using more benchmarks (for example, there are many common sense benchmarks that are also multiple choice). Another weakness is that I still don't quite buy the way in which difficulty was determined experimentally. However, the question itself is a provocative one, and I'm not sure that any answer is really perfect. The paper does a good job at starting the conversation and showing one way in which all of this can be formalized. Technical Quality: 3 Clarity: 4 Questions for Authors: Are there methods beyond Vote and Vote-Filter that can be applied here? And can this approach generalize if difficulties are probabilistic to begin with? Confidence: 3 Soundness: 3 Presentation: 4 Contribution: 3 Limitations: The paper focuses on analyzing and experimenting with the scaling behaviors of two specific instances of compound AI systems, though there are many other types. The experiments are conducted on relatively objective language tasks for ease of evaluation. It remains an open question how performance scales on more subjective tasks, such as generating poems and writing essays. Another open problem is how to predict the difficulty of queries without actually querying the language models. Flag For Ethics Review: ['No ethics review needed.'] Rating: 7 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We thank the reviewer for the helpful feedback and support of our paper! We have addressed the questions as follows. *Are there methods beyond Vote and Vote-Filter that can be applied here? Can this approach generalize if difficulties are probabilistic to begin with?: Thanks for bringing up this point. We indeed believe that our methods can apply beyond just Vote and Vote-Filter. We chose to focus on those two systems because … [paste] … We have in mind to do some followup work on more complex systems. One example is Ranking-Vote: Each LLM call provides a ranking of possible answers, and then we perform ranking aggregation (e.g., Borda count) to produce a final answer. Here it’s indeed important to consider probabilistic query difficulty. The effect will be significant because while in Vote and Vote-Filter we have a law of large numbers that governs the resulting prediction (polarizing the difficulty to a binary easy or had), in more complex models, we expect that the full distribution of query difficulty over a dataset will impact model performance. Can this approach generalize if difficulties are probabilistic to begin with?*: Yes. For example, one might take a Bayesian view: difficulties can be initially quantified as a prior distribution, and invoking more LLM calls offers a more accurate estimation of the posterior.	Rebuttal 1: Rebuttal: We thank all reviewers for their feedback! Please find our answers and clarifications in the individual responses.	NeurIPS_2024_submissions_huggingface	2,024	null	null	null	null	null	null	null	null
Practical Bayesian Algorithm Execution via Posterior Sampling	Accept (poster)	Summary: This paper proposes a new method within the Bayesian algorithm execution framework, which enables finding a target set of points in a highly efficient manner. Strengths: The problem tackled in this paper is highly important. The proposed method is sound and shows strong improvements vs. relevant benchmarks, in particular with respect to speed and sometimes also in terms of accuracy. Weaknesses: I don’t see any noteworthy weaknesses, but I am also not an expert in BO and related tasks (only about Bayesian inference more generally). Technical Quality: 4 Clarity: 4 Questions for Authors: The word “posterior sampling” seems to mean something very specific in this paper. However, in the Bayesian literature, this term is used for a much wider set of problems, which may be confusing to some readers of this paper. I think the mention the alternative term “Thompson sampling”. Could this be a good alternative name to use in relevant places? Confidence: 3 Soundness: 4 Presentation: 4 Contribution: 4 Limitations: I think the authors could discuss more limitations of this method. Are there any major limitations beyond the fact that the method only works for base algorithms that have a set of points as target? Flag For Ethics Review: ['No ethics review needed.'] Rating: 7 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Dear Reviewer BTqx, We sincerely appreciate your feedback and positive evaluation of our work. We have addressed your comments below and are ready to discuss any new questions or concerns that may arise during the discussion period. Additionally, given your positive assessment, we would deeply appreciate your support in championing our paper during the discussion period. Thank you in advance for your help! Q1. "The word 'posterior sampling' seems to mean something very specific in this paper. However, in the Bayesian literature, this term is used for a much wider set of problems… the alternative term 'Thompson sampling.' Could this be a good alternative name to use in relevant places?" A1. Thank you for pointing this out. We initially chose "posterior sampling" instead of "Thompson sampling" because we thought the latter might be too closely associated with optimization settings, whereas our work addresses a broader class of problems. However, in hindsight, we agree that "posterior sampling" may be too broad. We will revise our terminology accordingly in the final version. Q2. "I think the authors could discuss more limitations of this method. Are there any major limitations beyond the fact that the method only works for base algorithms that have a set of points as target?" A2. Our approach inherits several limitations from Bayesian optimization. Specifically, the good performance of PS-BAX depends on having a probabilistic model with reasonable predictive capabilities, which can be challenging in some applications. Since this limitation also applies to INFO-BAX and most probabilistic numerical methods, we did not initially highlight it. However, we will add a discussion on this in the revised manuscript. Additionally, as mentioned in Section 3, there is room for more sophisticated theoretical analyses of PS-BAX and INFO-BAX. While we see this as a future research direction enabled by our work rather than a limitation, we acknowledge the importance of developing novel theoretical tools to further analyze these methods. --- Rebuttal Comment 1.1: Comment: Thanks. I will keep my (positive) score. --- Reply to Comment 1.1.1: Comment: Dear Reviewer BTqx, Thank you once again for your thoughtful feedback and support of our work. Sincerely, The Authors	Summary: This paper introduces a scalable posterior sampling algorithm (aka PS-BAX) in the framework of Bayesian algorithm execution (BAX). Its fundamental basis is built on a key observation that the property of interest for many tasks is a target set of points, which for many scenarios such as standard bayesian optimization and level set estimation is usually the subset of the domain of the function. The authors propose PS-BAX with two key steps at each iteration: first the algorithm computes the target set using a sampled function from a posterior of Gaussian process; then it chooses the point with highest posterior variance (or entropy). In comparison to prior work (mainly INFO-BAX) based on expected information gain (EIG), PS-BAX shows superior computational efficiency. The authors have also established the asymptotic consistency of PS-BAX assuming the target set is stable. Strengths: The paper is well-written and the algorithm is validated through extensive and broad classes of experiments with comparisons to other methods. Weaknesses: This paper is an incremental work on INFO-BAX, which provides an alternative method to get around the expensive EIG computations by limiting the optimization to a target set per iteration. The posterior sampling is intuitive, however, it lacks rigorous development on its connection to INFO-BAX. Technical Quality: 3 Clarity: 3 Questions for Authors: 1. The authors claim that PS-BAX is scalable to high dimensional problems. The experiments contain examples only up to 10 dimensions. In the gene embedding experiment, the authors applied PCA to reduce the data to 5 dimensions. I suspect the Gaussian process approximations might be problematic in high dimensions. Can you justify the feasibility of applying PS-BAX to high dimensional data in numerical experiments? 2. INFO-BAX optimizes the mutual information over the entire domain X, whereas PS-BAX optimizes over a sampled target set X_n which is a subset of X. At each iteration (especially early iterations), one can imagine INFO-BAX should select a point that’s more informative on learning the task as it’s over the entire space. However, for some experiments (e.g. Figure 4), PS-BAX performs much better than INFO-BAX for early iterations. Can you provide more intuition on such phenomena? 3. The authors also claim that the PS-BAX method is easily parallelizable as one of the attractive contributions. However, as described in the main algorithm, PS-BAX is based on sequential computation of the posterior for the Gaussian process given samples selected from earlier iterations. Can you provide more details on how it can be parallelizable? Confidence: 2 Soundness: 3 Presentation: 3 Contribution: 2 Limitations: My major concern is that within the BAX framework minimizing the number of evaluations is of primary importance. However, PS-BAX seems to have a lower efficiency picking up points as compared to the INFO-BAX method given a limited number of iterations. The study of non-asymptotic results seems to be important to provide practical guidance as compared to asymptotic results. Flag For Ethics Review: ['No ethics review needed.'] Rating: 5 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Dear Reviewer Ta95, We thank you for your feedback and questions. We are glad that you found our paper "well-written" and our empirical evaluation "extensive." We hope to address your concerns in the following clarifications. Q1. "This paper is an incremental work on INFO-BAX… it lacks rigorous development on its connection to INFO-BAX" A1. We reiterate that PS-BAX offers superior computational efficiency and empirical performance compared to the state-of-the-art INFO-BAX. Furthermore, unlike INFO-BAX, PS-BAX enjoys a provable convergence guarantee. These facts alone demonstrate that our work is a substantial advancement. Additionally, though both policies use posterior samples of the target set, the design principles of INFO-BAX and PS-BAX are fundamentally different. INFO-BAX uses them to approximate the expected information gain (EIG) over the target set and then chooses the point maximizing this quantity. In contrast, PS-BAX draws a single posterior sample to identify a region likely to be the target set and then chooses the point with the highest posterior uncertainty within this set. While PS-BAX’s second step is also related to the notion of entropy, the EIG calculation required by INFO-BAX is generally much more computationally demanding, as discussed in Appendix B. We hope this further clarifies the algorithmic difference and highlights the novelty of our work. Q2. "The authors claim that PS-BAX is scalable to high dimensional problems… I suspect the Gaussian process (GP) approximations might be problematic…" A2. As noted, GP models struggle in high-dimensional settings. However, this is a limitation of GPs - PS-BAX can easily be paired with other probabilistic models. To demonstrate this and alleviate concerns, we conducted an additional experiment involving a top-K protein optimization task with an 80-dimensional input space using the publicly available GB1 dataset. Instead of a traditional GP, in this experiment, we use a deep kernel GP (Wilson et al. 2015), which scales better to higher-dimensional settings. As shown in Figure 1 of the PDF attached to our rebuttal’s overall response, PS-BAX and INFO-BAX perform similarly (both outperforming Random significantly). However, like in other problems, PS-BAX is much faster to compute. Indeed, in our paper, we imply that PS-BAX scales better to high-dimensional settings than INFO-BAX. Here, we referred specifically to the computational cost of PS-BAX and INFO-BAX in problems of moderate-to-high dimension. Maximizing the EIG, as required by INFO-BAX, becomes prohibitively expensive in such settings. For instance, in the Ackley 10-D problem, each iteration of INFO-BAX takes over 6 minutes. In contrast, PS-BAX, with its lightweight computation, remains feasible in moderate dimensions (only 15 seconds for Ackley 10-D). Q3. "...for some experiments (e.g. Figure 4), PS-BAX performs much better than INFO-BAX for early iterations." A3. The superior performance of PS-BAX over INFO-BAX is due to two key reasons. First, PS-BAX inherits posterior sampling’s strong exploration capabilities due to the stochastic nature of its sampling choices, allowing it to escape incorrect posterior beliefs of the target set’s identity. This is particularly true in challenging problems, where uncertainty remains significant throughout the experimentation loop. In contrast, INFO-BAX myopically tries to minimize the target set’s posterior uncertainty as much as possible at each iteration, causing it to get stuck in a wrong belief of the true target’s identity. This behavior is observed in Figure 3 (corresponding to the performance plot shown in Figure 4), where INFO-BAX fails to find a portion of the target set due to its lack of exploration. The second reason is computational. As discussed above, maximizing the EIG over high-dimensional spaces is challenging, meaning that despite significant computational efforts, INFO-BAX is potentially choosing to evaluate a local maximum of the EIG instead of the global maximum, thus deteriorating its performance. Q4. "The authors claim that PS-BAX is easily parallelizable…" A4. We apologize for any confusion. By "parallelizable," we mean that our algorithm can be generalized to the "parallel" or "batch" setting, where multiple points are selected at each iteration (Kandasamy et al. 2018). This generalization follows the approach of Kandasamy et al. (2018). Specifically, to select $q$ points at each iteration, we draw $q$ independent samples from the target set and then select the subset of $q$ points with the highest entropy. To demonstrate our algorithm's empirical performance in the batch setting, we include results for two test problems analyzing PS-BAX under three different batch sizes in Figures 2 and 3 of the PDF attached to our rebuttal’s overall response. Q5. "My major concern is that… PS-BAX seems to have a lower efficiency.. compared to INFO-BAX ... The study of non-asymptotic results seems to be important…" A5. We kindly ask the reviewer to consider that our empirical evaluation clearly shows the opposite: PS-BAX is more efficient than INFO-BAX, especially in challenging problems. In contrast, there is no evidence, theoretical or empirical, to substantiate that INFO-BAX is more efficient than PS-BAX. We agree that non-asymptotic results could provide guidance. However, this does not diminish the contributions of our work, which provides an algorithm with faster computation, significantly better empirical performance, and a convergence guarantee. References Kandasamy, K., Krishnamurthy, A., Schneider, J., & Póczos, B. (2018). Parallelised Bayesian optimisation via Thompson sampling. In International Conference on Artificial Intelligence and Statistics. Wilson, A. G., Hu, Z., Salakhutdinov, R., & Xing, E. P. (2016). Deep kernel learning. In International Conference on Artificial Intelligence and Statistics. --- Rebuttal Comment 1.1: Comment: Thanks for the response! The batch-based experiments make sense to me. I have raised the score to 5. --- Reply to Comment 1.1.1: Comment: Dear Reviewer Ta95, We are glad that our response addressed your concerns, and we sincerely thank you for raising your score. Thank you again for your valuable feedback and support of our work. Sincerely, The Authors	Summary: This paper utilizes the idea of Bayesian sampling in the Bayesian algorithm execution (BAX) framework. As claimed by the authors, this is the first extension of posterior sampling beyond the optimization setting. The idea is very simple and natural and the performance seems reasonable. While I am not familiar with the BAX literature, it seems that the theoretical justification is incorrect. Strengths: Overall, the paper is well structured and well written. Weaknesses: In terms of computation cost, if a Gaussian process is used, then the computation cost of posterior variance $\sigma_n(x)$ involves the inversion of a high dimensional matrix. Can you please comment on it? In terms of the theorem 1: The notation is very confusing. On the LHS ($P_n(X=\mathcal{O_ A}(f))$), $f$ is random with probability measure $P_n$. On the RHS ($1(X=\mathcal{O_ A}(f))$), $f$ becomes fixed. I suppose the author means that on the LHS, $f$ refers to the random sample, and on the RHS, $f$ refers to the true function. The proof of theorem 1 doesn't involve the prior specification or the true $f$ function, which seems incorrect. For instance, (1) if the prior distribution is a Dirac measure, then the posterior is also a Dirac measure, and the algorithm never works as the sampling repeatedly returns the same $f_n$. But it won't affect theorem 1? (2) if $f$ contains a flat region of global optimal and $\mathcal A$ aims to find optimums, (i.e., $\mathcal{O_ A}(f)$ is an uncountable set) and the Gaussian process is used, then with probability 1, $\mathcal{O_ A}(\tilde f_n)$ contains only one element for any $n$. The claimed convergence clearly doesn't hold. Theorem 2 claims that consistency fails in general, and Theorem 1 claims that consistency holds when we have the stable assumption. This implies that some restriction is necessary, and the stable assumption is a sufficient assumption, not a necessary assumption. Technical Quality: 1 Clarity: 3 Questions for Authors: Please see weakness. Confidence: 4 Soundness: 1 Presentation: 3 Contribution: 2 Limitations: The checklist mentions that it discusses the limitation in section 4.3, but I cannot find it. Flag For Ethics Review: ['No ethics review needed.'] Rating: 5 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Dear Reviewer FDTf, We sincerely thank you for your comments and questions. We are glad that you found our paper "well structured and well written." You expressed concerns related to the notation and clarity of our theoretical results. We hope our response below addresses these concerns. Since no other major issues were raised, we kindly ask you to consider raising your rating. Q1. "In terms of Theorem 1: The notation is very confusing. On the LHS…" A1. Our statement of Theorem 1 is indeed mathematically correct and we hope to clarify any misunderstandings here. Under a Bayesian perspective, $f$ is a random function with a given prior probability distribution. On both sides of the equation $\lim\_{n\rightarrow\infty}\mathbf{P}\_n(X = \mathcal{O}\_{\mathcal{A}}(f)) = \mathbf{1}\set{X = \mathcal{O}\_{\mathcal{A}}(f)}$, $f$ refers to the same random function. Furthermore, we note that $\mathbf{P}\_n$ is a random probability measure due to its dependence on the dataset $\set{(x_i, y_i)}\_{i=1}^n$, which in turn depends on $f$. Thus, the claim that “$\lim\_{n\rightarrow\infty}\mathbf{P}\_n(X = \mathcal{O}\_{\mathcal{A}}(f)) =\mathbf{1}\set{X = \mathcal{O}\_{\mathcal{A}}(f)}$ almost surely” should be interpreted as follows: the sequence of random variables $\set{\mathbf{P}\_n(X = \mathcal{O}\_{\mathcal{A}}(f))}\_{n=1}^\infty$ converges almost surely to the random variable $\mathbf{1}\set{X = \mathcal{O}\_{\mathcal{A}}(f)}$ under the probability measure induced by the prior on $f$. We will clarify this in the revised version of our manuscript. In measure-theoretic terms, Theorem 1 simply states that the random variable $\mathbf{1}\set{X = \mathcal{O}\_{\mathcal{A}}(f)}$ is $\mathcal{F}\_\infty$ -measurable, where $\mathcal{F}\_\infty$ is the sigma-algebra generated by the sequence of observations $\set{(x_n, y_n)}\_{n=1}^\infty$. Intuitively, this means that given the data $\set{(x_n, y_n)}\_{n=1}^\infty$, we can perfectly determine if the statement “ $\mathcal{O}\_{\mathcal{A}}(f)$ is equal to $X$” is true for any $X$. In particular, the following is a direct corollary of Theorem 1. Corollary. Let $\widehat{X}\_n = \arg\max\_{X\subset\mathcal{X}} \mathbf{P}\_n(X = \mathcal{O}\_{\mathcal{A}}(f))$ be the time-$n$ maximum a posteriori estimator of $\mathcal{O}_{\mathcal{A}}(f)$. Then, $\widehat{X}\_n$ converges to $\mathcal{O}\_{\mathcal{A}}(f)$ almost surely. Q2. "The proof of Theorem 1 doesn't involve the prior specification or the true $f$ function…" A2. We apologize for the confusion caused by our statement of Theorem 1. As discussed in A1, we follow a Bayesian perspective where $f$ is drawn from the prior distribution used by our algorithm. Our result assumes that the prior is well-specified. Such results are typical in the literature (see, e.g., Russo and Van Roy 2014, Bect et al. 2016, and Astudillo and Frazier 2021) and can be seen as the Bayesian counterpart of frequentist results, which typically assume that $f$ lies in the reproducing kernel Hilbert space of a user-specified kernel. Consistency in the case where the prior over $f$ is a Dirac measure does hold. Regarding the example where $f$ "contains a flat region of global optimum and A aims to find optima… and the Gaussian process is used," we agree that consistency does not hold because the prior is not well-specified in this case. We will revise our statement of Theorem 1 to clarify that $f$ is assumed to be drawn from the prior distribution to avoid this confusion in the future. Q3. "...the stable assumption is a sufficient assumption, not a necessary assumption." A3. We thank the reviewer for pointing this out. We will revise our usage of the word "necessary" in Line 168 to avoid this confusion. As noted, Theorem 1 shows that stability is a sufficient condition for asymptotic consistency, and Theorem 2 shows that asymptotic consistency cannot be guaranteed without stability. Q4. "...the computation cost of posterior variance $\sigma_n(x)$ involves the inversion of a high-dimensional matrix" A4. The computation of the posterior variance scales as $n^3$, where $n$ is the number of training points. While this is seen as a computational bottleneck in other applications involving Gaussian processes, it is not the case in Bayesian optimization and Bayesian algorithm execution tasks, where the maximum number of evaluated points rarely surpasses 1000 due to the assumed high cost of such evaluations. In contrast, the computation of the expected information gain required by INFO-BAX scales as $(n+m)^3$, where $m$ is the size of the target set, also due to the need to compute the inverse of a matrix. Even if $n$ is small, $m$ can be very large in real-world applications such as shortest-path problems and level-set estimation tasks, making INFO-BAX extremely computationally demanding. Q5. "The checklist mentions that it discusses the limitation in section 4.3…" A5. We apologize for the confusion. We meant Section 3 instead of Section 4.3. Additionally, in Section 2 (Lines 77-78), we point out that the range of problems that can be tackled with our PS-BAX strategy is narrower than those that can be tackled with INFO-BAX. References Astudillo, R., & Frazier, P. (2021). Bayesian optimization of function networks. Advances in Neural Information Processing Systems, 34, 14463-14475. Bect, J., Bachoc, F., & Ginsbourger, D. (2019). A supermartingale approach to Gaussian process based sequential design of experiments. Bernoulli, 25(4A), 2883-2919. Russo, D., & Van Roy, B. (2014). Learning to optimize via posterior sampling. Mathematics of Operations Research, 39(4), 1221-1243. --- Rebuttal Comment 1.1: Comment: Dear authors, Sorry for my late reply. I have read all the rebuttals, and have a follow-up regarding theorem 1. On the LHS (i.e., the term $\mathbf P_n(X=O_A(f))$), does $\mathbf P_n$ refer to the posterior measure of $f$? That is, LHS is a RANDOM probability quantity, which depends on random data $(x_i,y_i)$ On the RHS, does $f$ still follow the posterior distribution, where the posterior distribution is random as it depends on data $(x_i,y_i)$? If so, then, theorem 1 essentially prove that the posterior distribution of $O_A(f)$ converges to a Dirac measure. In other words, theorem 1 is a posterior concerntration result (posterior distribution concentrates toward certain limit), but NOT a posterior consistency result (posterior distribution concentrates toward truth). --- Rebuttal 2: Comment: Dear Reviewer FDTf, We sincerely thank you for taking the time to read our response and for your follow-up question. Our result is indeed a Bayesian posterior consistency result, and we hope our response below clarifies this. In short, you are right that $\mathbf{P}\_n$ denotes the posterior measure of $f$. You are also right that the LHS denotes a random quantity depending on the random data $\set{(x\_i, y\_i)}\_{i=1}^n$. However, $f$ on the RHS is not a function drawn from the posterior but rather a function drawn from the prior, as we explain below. Our result should be interpreted as follows. Suppose that a random function $f$ is drawn from a prior distribution $p$. This function will remain fixed throughout the entire data collection process, and the data will emanate from this fixed function in the sense that $y\_i = f(x\_i) + \epsilon\_i$. Further, suppose that $f$ is unknown to our algorithm, but the prior $p$ is known to our algorithm. Although $f$ is unknown, at each iteration $n$, our algorithm can form a posterior distribution over $f$ using the (correctly-specified) prior $p$ and the sequence of observations $\set{(x\_i, y\_i)}\_{i=1}^n$. Let $p\_n$ denote this posterior and, for any fixed $X\subset\mathcal{X}$, let $q\_n(X)$ denote the probability that $\mathcal{O}\_\mathcal{A}(f) = X$ computed from $p\_n$. Theorem 1 states that with probability one, for a function $f$ drawn from $p$, $q\_n(X)$ will converge to $1$ if $X=\mathcal{O}\_\mathcal{A}(f)$ and will converge to 0 otherwise. Consequently, our result is indeed a Bayesian posterior consistency result in the sense that the estimated posterior of $\mathcal{O}\_\mathcal{A}(f)$ converges to the Dirac measure of the ground truth. As a final note, we acknowledge that consistency results are typically stated in terms of convergence of estimators rather than convergence of posterior distributions. We chose to present the latter as it is a stronger result, but we now realize that such a result is more prone to cause confusion. As discussed in our original response, a direct corollary of Theorem 1 is that the maximum a posteriori estimator of $\mathcal{O}\_\mathcal{A}(f)$ is asymptotically consistent. We will add this corollary to our revised manuscript, along with a discussion of our result inspired by your question. We hope this explanation addresses your concern. If you have any further questions, please let us know, and we will do our best to respond within the remaining time of the discussion period. Thank you again for your valuable contribution to improving our work. Sincerely, The Authors --- Rebuttal Comment 2.1: Comment: Thanks for the explanation. I finally understand the notation (and the meaning that the prior needs to specify the $f$ well). A better presentation of theorem 1 is needed. Personally, if you say that $f$ follows a prior distribution $p$, and present the LHS as a conditional probability conditioned on data, I could understand it quickly. I would raise my score to 5, as I think the implicit assumption of compact $\mathcal X$ is not favorable. --- Reply to Comment 2.1.1: Comment: Dear Reviewer FDTf, We are glad to hear that our response has addressed your concern. We will revise the statement of Theorem 1 based on your feedback to enhance its clarity. Thank you again for your valuable feedback and support of our work. Sincerely, The Authors	Summary: This work proposes a posterior sampling algorithm for Bayesian algorithm execution, where the goal is to infer the output of an algorithm $O$ applied to an unknown function $f$. The algorithm is simple to implement and computationally more efficient than previous works based on mutual information maximization. The authors proved the algorithm is consistent with prior probability 1, and demonstrated empirically that it has comparable and sometimes better sample efficiency than previous work while being faster computationally. Strengths: - The problem studied is interesting and practically relevant. - The proposed method is simple, computationally efficient and empirically effective, thereby providing a robust baseline for the problem. - The paper is mostly well-written. Weaknesses: - I am uncertain about the correctness of the consistency proof (see question section below). - A less important point is about a design choice in the algorithm: when the sampled target set has multiple point the algorithm selects the point $x$ with the highest entropy $H[f(x)]$. This does not appear to be a universally optimal choice, for example if we want to estimate a level set $\\{x: f(x)\ge A\\}$ and there exists some $x_0\in \mathcal{X}$ for which the posterior $P(f(x_0)\mid D_n)$ has a very large mean and also a large variance: in such cases $H(f(x_0))$ could be the largest among the sampled level set but there may be little uncertainty that $x_0$ lives in the true level set. In addition, there may be scenarios where we need to focus on inferring the boundary of the level set instead of reducing the uncertainty about function values in the interior. Technical Quality: 2 Clarity: 3 Questions for Authors: My main question is about the proof of Theroem 1, in particular the claim on L762 that $Z\cap O_A(f)=\emptyset$ "by construction". I can see this may be true if we assume $P_\infty(O_A(f))$ always assigns positive probability to some set $X\subset \mathcal{X}$, but what if this is not the case, for example, if $P_\infty(O_A(f))$ is a distribution over bounded intervals $\subset \mathbb{R}=\mathcal{X}$ and its marginal distributions for the endpoints are atom-less? Clearly similar issues can happen if we have (for example) a GP prior with a continuous kernel and consider $P_n$ instead of $P_\infty$. There should be explanations on why the issue will not happen with $P_\infty$, or why the reasoning of L762 is still valid in the presence of such issues. Confidence: 3 Soundness: 2 Presentation: 3 Contribution: 3 Limitations: Yes. Flag For Ethics Review: ['No ethics review needed.'] Rating: 5 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Dear Reviewer qVHf, We sincerely appreciate your feedback and questions. We are glad that you found the problem of study "interesting and practically relevant," our method providing "a robust baseline for the problem," and our paper "well-written." Your major concern relates to the correctness of a specific claim in our proof of Theorem 1. Below, we explain that the claim holds true under an additional assumption (see A1) that we regrettably neglected to include in the statement of Theorem 1. Noticing your high scores for soundness, presentation, and contribution, we hope this clarification will lead you to consider raising your rating. We also address a minor concern related to the optimality of our algorithm in certain settings. Q1. "My main question is about the proof of Theorem 1, in particular the claim…" A1. Our proof of Theorem 1 assumes that $\mathcal{X}$ is a finite set. As noted, without this assumption, $\mathbf{P}(X=\mathcal{O}_{\mathcal{A}}(f))$ cannot be guaranteed to be positive for any subset $X \subset \mathcal{X}$. We conjecture that a suitable form of asymptotic consistency for PS-BAX holds even if $\mathcal{X}$ has infinite cardinality. However, we expect such analysis to be significantly more challenging given that techniques commonly used in Bayesian optimization to extend convergence guarantees to continuous input spaces, such as carefully chosen discretizations and Lipschitz assumptions (e.g., Bull et al. 2011 and Srinivas et al. 2012), do not naturally translate to settings where performance cannot be summarized by objective function values. We believe this is the reason why most theoretical results in level set estimation assume that the input space is finite (see, e.g., Gotovos et al. 2013 and Mason et al. 2022). Despite this limitation, our theoretical guarantee still covers a broad range of real-world applications; in particular, the shortest path, top-K, and drug discovery problems explored in our experiments, which are inherently discrete. We apologize for the confusion and hope this response alleviates any concerns regarding the validity of our theoretical result. Q2. "A less important point is about a design choice in the algorithm…" A2. PS-BAX comprises two steps: (1) drawing a sample from the posterior distribution on the target set and (2) selecting a point within the sampled target set. For the second step, we chose to select the point with the highest posterior variance within the sampled target set, which is a simple strategy that can ensure asymptotic consistency, inspired by active learning. Despite its simplicity, this strategy performs well across a broad range of tasks. However, we agree that other strategies tailored to specific applications could potentially improve performance further. We see this as a valuable direction for future research enabled by this work, and the demonstration that posterior sampling can be successfully applied to a broader range of tasks beyond optimization as our primary contribution. References Gotovos, A., Casati, N., Hitz, G., & Krause, A. (2013). Active learning for level set estimation. In International Joint Conference on Artificial Intelligence. Mason, B., Jain, L., Mukherjee, S., Camilleri, R., Jamieson, K., & Nowak, R. (2022). Nearly Optimal Algorithms for Level Set Estimation. In International Conference on Artificial Intelligence and Statistics. --- Rebuttal 2: Title: quick questions Comment: Thank you for your response. I agree my question will be addressed if we assume $\mathcal{X}$ is finite. Could you point out where the finiteness assumption is made? I also have a quick question regarding your response to Reviewer fDTf: you said > Regarding the example where "contains a flat region of global optimum and A aims to find optima… and the Gaussian process is used," we agree that consistency does not hold because the prior is not well-specified in this case. Could you explain why the prior is not well-specified in this case? Your subsequent response suggested that the issue will not appear if we restrict to prior-almost every $f$. But we can have a GP prior that assigns positive (indeed, total) mass to constant functions, in which case the example doesn't seem to be immediately ruled out by the restriction. --- Rebuttal 3: Comment: Dear Reviewer qVHf, Thank you for taking the time to read our response. We address your new questions in detail below. > I agree my question will be addressed if we assume $\mathcal{X}$ is finite. Could you point out where the finiteness assumption is made? As mentioned in our original response (fourth sentence of the first paragraph), we regrettably neglected to explicitly include this assumption. We apologize for this oversight. This assumption will be clearly stated in the revised version of our manuscript. > I also have a quick question regarding your response to Reviewer fDTf… Your subsequent response suggested that the issue will not appear if we restrict to prior-almost every $f$. But we can have a GP prior that assigns positive (indeed, total) mass to constant functions, in which case the example doesn't seem to be immediately ruled out by the restriction. We agree that Gaussian priors can place positive mass on constant functions. However, note that Reviewer fDTf mentions an example such that the true function “$f$ contains a flat region of global optima” and then adds that “[if a] Gaussian process is used, then with probability 1, $\mathcal{O}\_\mathcal{A}(\tilde{f}\_n)$ contains only one element…” The combination of these two statements necessarily means that Reviewer fDTf is considering an example where the prior is not well-specified. Indeed, if we use a Gaussian process prior such that with probability one $\arg\max\_{x\in\mathcal{X}}f$ is a singleton, this necessarily implies that, with probability one, $f$ will not contain a flat region of global optima. To address your concern more directly, below we show that if the prior is such that $f$ is constant with probability one and $\mathcal{O}\_\mathcal{A}(f) = \arg\max\_{x\in\mathcal{X}}f$, then asymptotic consistency holds. Although the proof is tautological, we hope it clarifies any misunderstanding. Additionally, we note that this result holds for any choice of sampling decisions $\set{x\_n}\_{n=1}^\infty$. Thus, asymptotic consistency in this specific situation not only holds for PS-BAX, but actually holds for any algorithm. Proposition. Suppose the prior is such that $f$ is constant with probability one and let $\mathcal{O}\_\mathcal{A}(f) = \arg\max\_{x\in\mathcal{X}}f$, then $\lim\_{n\rightarrow\infty}\mathbf{P}(X=\mathcal{O}\_\mathcal{A}(f)) = \mathbf{1}\set{X = \mathcal{O}\_\mathcal{A}(f)}$ almost surely under the prior for any $X \subset \mathcal{X}$. Proof. Since $f$ is constant with probability one, we have that $\mathcal{O}\_\mathcal{A}(f) = \mathcal{X}$ with probability one. Consequently, with probability one, $\mathbf{1}\set{X = \mathcal{O}\_\mathcal{A}(f)} = 1$ if $X=\mathcal{X}$ and $\mathbf{1}\set{X = \mathcal{O}\_\mathcal{A}(f)} = 0$ if $X\neq \mathcal{X}$. Moreover, observe that since the prior only puts mass on constant functions, the same is necessarily true for the posterior. Therefore, $\mathbf{P}\_n(\mathcal{O}\_\mathcal{A}(f) = \mathcal{X})=1$ for all $n$. This also implies that $\mathbf{P}\_n(\mathcal{O}\_\mathcal{A}(f) = X)=0$ for any $X\subset\mathcal{X}$ with $X\neq \mathcal{X}$. From the above, it follows that, for any $X \subset \mathcal{X}$, $\lim\_{n\rightarrow\infty}\mathbf{P}(X=\mathcal{O}\_\mathcal{A}(f)) = \mathbf{1}\set{X = \mathcal{O}_\mathcal{A}(f)}$ almost surely, as desired. $\square$ We hope this discussion addresses your concerns. Please let us know if any further clarification is needed. Sincerely, The Authors --- Rebuttal Comment 3.1: Comment: Dear Reviewer qVHf, As the end of the discussion period approaches, we would greatly appreciate it if you could confirm whether our response has adequately addressed your concerns. We also encourage you to review our recent discussion with Reviewer FDTf, as it addresses similar points and may provide further clarity on the issues you have raised. If any questions remain, please let us know, and we will do our best to respond within the remaining time. If your concerns have been resolved, we kindly ask you to consider raising your rating, as this would reflect the improvements made based on your valuable feedback. Thank you again for your time and efforts in reviewing our manuscript. Sincerely, The Authors --- Rebuttal 4: Comment: Thank you for your response. My concerns regarding correctness seem addressed and I will update the score accordingly. The finiteness assumption is somewhat unfortunate, especially since you only have a consistency result (as opposed to rates of contraction). You mentioned there are technical challenges with discretization. Is there any other scenario where the input space has a very large cardinality (e.g. in graph-related applications) and the consistency proof could still be relevant? --- Rebuttal 5: Comment: Dear Reviewer qVHf, We are glad that our response has addressed your primary concern, and we sincerely thank you for raising your score. Regarding your question about our consistency result and its finiteness assumption, we would like to emphasize the following points: - Our result applies to a broad range of critical real-world applications. Indeed, many real-world problems involve large, inherently discrete input spaces. For instance, in the drug discovery application discussed in our work (Section 4.5), the input space consists of a discrete set of 5,000 gene mutations. We also recently introduced a protein engineering application formulated as a top-$k$ optimization problem with an input space of similar size (please see A2 in our response to Reviewer Ta95). Moreover, similar problems in this area can involve input spaces exceeding 100,000 elements. Real-world shortest-path problems also often feature large transportation networks with thousands of nodes. - Our result is non-trivial even when the input space is small. In practical scenarios, observations are often corrupted by noise, which means that uncertainty regarding the true identity of the target set might persist even if the entire input space is evaluated. Our consistency result, which holds under noisy observations, ensures that PS-BAX can effectively mitigate such uncertainty in the long run. - The asymptotic consistency of adaptive algorithms like PS-BAX is not something that can be taken for granted. In Bayesian optimization contexts, popular algorithms have been shown to lack asymptotic consistency, even in discrete input spaces (see, e.g., Astudilo et al. 2023). The absence of asymptotic consistency often comes with erratic behavior, which may hinder performance in practical scenarios. Thus, the combination of our broad empirical evaluation, demonstrating the strong performance of PS-BAX, with our asymptotic consistency result, provides compelling evidence of PS-BAX’s potential to effectively address real-world challenges. We hope this discussion clarifies the significance of our asymptotic consistency result. Thank you again for your valuable feedback and support of our work. Sincerely, The Authors References Astudillo, R., Lin, Z. J., Bakshy, E., & Frazier, P. (2023). qEUBO: A decision-theoretic acquisition function for preferential Bayesian optimization. In International Conference on Artificial Intelligence and Statistics (pp. 1093-1114). PMLR.	Rebuttal 1: Rebuttal: Dear reviewers, We sincerely thank you for your thoughtful comments and questions. We are pleased that you found our paper well-written (qVHf, FDTf, Ta95, BTqx), addressing an interesting and practically relevant problem (qVHf, BTqx), and proposing a sound algorithm (qVHf, FDTf, BTqx) with clear performance improvements demonstrated through extensive empirical evaluation (qVHf, FDTf, Ta95, BTqx). Our paper's current ratings are as follows: * Reviewer BTqx provided a high appraisal with a rating of 7. * Reviewers qVHf and Ta95 both assigned a rating of 4. Despite the moderate ratings, their comments are fairly positive, and their scores for soundness, presentation, and contribution are high. We believe we have addressed their concerns effectively and hope they will consider raising their ratings. * Reviewer FDTf assigned a rating of 3. The only major concern raised was the correctness of our asymptotic consistency result. As we explained in our response, our theoretical result is correct, and the confusion arises from not explicitly stating that $f$ is assumed to be drawn from the prior distribution, a simple improvement we will include in the revised version of our manuscript. Since no other concerns were raised, we kindly ask this reviewer to consider raising their rating. We have responded to each reviewer’s questions individually, aiming to provide sufficient detail. However, we would like to use this space to globally comment on a few points: Contributions of our work: We introduce PS-BAX, a novel algorithm applicable to a wide range of real-world problems, offering superior computational efficiency and empirical performance compared to the state-of-the-art (INFO-BAX). Additionally, PS-BAX comes with a convergence guarantee (unlike INFO-BAX) and provides novel insights into posterior sampling methods. We believe these contributions are at least on par with the average NeurIPS paper, so we kindly ask reviewers to consider raising their ratings. Novelty and algorithm design choices: PS-BAX is the first application of posterior (Thompson) sampling to the Bayesian algorithm execution setting. Indeed, PS-BAX reduces to posterior sampling when the algorithm is simply selecting the argmax. While there are many ways to extend posterior sampling, we chose a simple yet effective approach: selecting the point with the highest uncertainty among the candidate points returned by the posterior sampling step. Exploring other design choices would be a great direction for future research enabled by our work. Theoretical results: We appreciate Reviewers qVHf and FDTf for requesting more clarification about our theoretical results. As detailed in our individual responses, the issues are minor and can be easily addressed. Thank you again for your insightful feedback. We look forward to a fruitful and engaging discussion. Sincerely, The authors Pdf: /pdf/633fae7f76a362ed2a8e4ccf3f1a0797bbfadd02.pdf	NeurIPS_2024_submissions_huggingface	2,024	null	null	null	null	null	null	null	null
Deep Graph Mating	Accept (poster)	Summary: In this paper, the authors present a method for learning-free and label-free model reuse for graph neural networks. They dub the task Deep Graph Mating (Grama). Without relying on costly fine-tuning or re-training, Grama aims to generate a child model by reusing and fusing knowledge from pre-trained parent models, particularly by managing pre-trained parameters. Compared with conventional model reuse tasks (knowledge distillation, and knowledge amalgamation) in the GNN domains, the authors claim in Tab 1 that Grama simultaneously enables multi-model reuse, annotation free, and training/fine-tuning free, leading to more resource-efficient model reuse. To achieve Grama, the authors begin by deriving two vanilla Grama methods, dubbed VPI and VAPI, motivated by the permutation invariance proposition in GNNs. However, both VPI and VAPI don’t work well, and the authors observe unique challenges for Grama, namely the increased sensitivity to weight misalignment and the accompanying topology-dependent complexities, substantiated by a series of propositions, lemmas, and conjectures. Motivated by the identified challenges, the authors introduce the proposed DuMCC framework, which includes a PMC and CMC. PMC is designed to identify topology-aware permutation matrices but suffered from increased susceptibility to over-smoothing. To solve this problem, the authors further designs CMC, which is to refine message statistics of the child network through a learning-free message normalization layer. Extensive experiments (7 benchmarks, 4 tasks, 5 architectures) in various graph applications have been performed to validate the effectiveness of the proposed framework, covering popular GNNs including GCN, GraphSAGE, GAT, GIN, and DGCNN. The authors also provide extensive supplementary information with detailed proofs in the appendix. Strengths: The positive traits of this paper include: - Well-organized paper with a very clear logic flow. The authors first give a well-defined problem of Grama with reasonable motivations. Then they present two vanilla methods as initial solutions to solve the Grama problem. Based on these methods, the authors identify the unique challenges from Grama with elaborated analysis and clear thoughts. Motivated by the identified challenges, a DuMCC framework is proposed. The way the authors present the entire paper make it very easy for following its development. - Interesting idea. To the best of my knowledge, Grama is the first to simultaneously enable multi-model reuse, annotation-free, and training/fine-tuning free capabilities. The authors clearly demonstrate their idea by comparing Grama with previous model reuse methods in Tab 1, which is very useful for understanding the main benefits of Grama. It is also interesting to explore the correspondence in the weight space for different GNNs, which also appears to be the first exploration in this field of GNN. - Solid intuitions and well-established algorithm. At first sight, the proposed approach appears a bit ad-hoc. However, the authors provide sufficient motivations for each component in DuMCC, making their method well-justified. This includes extensive discussions on vanilla methods and the over-smoothing issues from PMC, with a motivating example shown early in Fig 1, making the proposed method technically sound. - Experiments are conducted thoroughly. It is clear that the authors try to cover a wide range of tasks and various GNN architectures in their experiments. The tasks include node property prediction, graph property prediction, 3D object recognition, and 3D semantic segmentation, and the architectures include GCN, GraphSAGE, GAT, GIN, and DGCNN. The effectiveness of the proposed method and nearly all claims in the paper are thoroughly verified. - Good reproducibility. Extensive supplementary details are provided in the appendix, with source code and models provided in the supplementary material. Algorithm 1 is clear and useful for capturing the main procedure of the entire framework. Weaknesses: - In Sec 5.3, the authors introduce a learning-free message normalization (LFNorm) layer for statistics calibration in CMC, which is technically sound. However, the authors miss the discussions on the Grama case where pre-trained models already contained normalization layers, which is common in the area of graph neural networks. In that case, it is unclear whether the batch statistics are also recomputed in CMC. If they are, it should be specified when the recomputation occurs: simultaneously with LFNorm or after? - Because the authors have claimed multimodel reuse as a contribution of this work, I would advise the authors to include more experiments for the Grama cases with three or more pre-trained models. Otherwise, clear statements should be added to limit the scope of the study to the reuse of two pre-trained models. - The descriptions and explanations in the experimental results are not closely matched with the authors’ assumptions and propositions from the previous sections. It is suggested that the authors more closely relate the texts in the experimental section to the methods section, particularly by specifying which results validate which proposition or claim. This would make the paper more readable. - In the appendix, there is a minor problem where Eq 30 extends beyond the margin of the paper. Technical Quality: 3 Clarity: 3 Questions for Authors: 1. Can you clarify if the statistics in the original normalization layer are recomputed during Grama? If so, when would they be recomputed? 2. Have you experimented with using three or more pre-trained models in Grama? 3. Consider improving the experimental section to more closely relate the texts to the validated claims. Although the authors missed some key information, I feel these could be easily addressed. On balance, the paper is both novel and useful. The concept of training-free and label-free model reuse for GNN had not been done before. Confidence: 5 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: The paper has thoroughly discussed limitation in Sec 7 and social impact in appendix. Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: ### Response to Reviewer tRzs We appreciate the reviewer for the positive support and constructive comments. `W1. & Q1.` Statistics recomputation details >"The authors miss the discussions on the Grama case where pre-trained models already contained normalization layers, which is common in the area of graph neural networks. In that case, it is unclear whether the batch statistics are also recomputed in CMC. If they are, it should be specified when the recomputation occurs: simultaneously with LFNorm or after?" >"Can you clarify if the statistics in the original normalization layer are recomputed during Grama? If so, when would they be recomputed?" `Response:` We apologise for not having sufficiently highlighted the details regarding the recomputation of statistics, which might have led to them being overlooked by the reviewer. In fact, as detailed in line 313-314, we have indeed discussed the scenario where models are equipped with normalisation layers. We have explicitly explained that the running mean and variance of the original normalisation layers are recalculated for the child GNN. In our revised version, we will enhance the clarity of this detail further by also providing additional explanations in Sect. 5.3. But admittedly, we apologise for not having specified when the recomputation of statistics occurs for the original normalisation layers in our paper. We appreciate the reviewer's insightful comments on this issue. As specified in the source code provided in the supplementary material, we clarify that the recomputation of statistics is performed concurrently with that of the LFNorm layer for the child GNN. In the revision, we will provide a detailed clarification and add the discussion on the normalisation-related procedures described above in Sect. 5.3 and Sect. 6. --- `W2. & Q2.` Multi-model GRAMA results >"Because the authors have claimed multimodel reuse as a contribution of this work, I would advise the authors to include more experiments for the Grama cases with three or more pre-trained models. Otherwise, clear statements should be added to limit the scope of the study to the reuse of two pre-trained models." >"Have you experimented with using three or more pre-trained models in Grama?" `Response:` We appreciate the reviewer's constructive suggestion. As suggested, we have conducted additional experiments on disjoint partitions of the ogbn-arxiv dataset using the architecture described in Tab. 10 of the main paper, specifically for the proposed GRAMA in scenarios involving the simultaneous reuse of three pre-trained models. The results, presented below, demonstrate that our proposed DuMCC approach is also effective in multi-model GRAMA contexts. These results will be incorporated into the revised version of the paper. \| Models \| Performance: Dataset 1 \| Performance: Dataset 2 \| Performance: Dataset 3 \| \|:----- \| :----- \| :----- \| :----- \| \| Parent 1 \| 0.6645 \| 0.6476 \| 0.6040 \| \| Parent 2 \| 0.4796 \| 0.7044 \| 0.4651 \| \| Parent 3 \| 0.5805 \| 0.6078 \| 0.7728 \| \| Child \| 0.5574 \| 0.6360 \| 0.5478 \| --- `W3. & Q3.` Linking experiments to the claims to validate >"The descriptions and explanations in the experimental results are not closely matched with the authors' assumptions and propositions from the previous sections. It is suggested that the authors more closely relate the texts in the experimental section to the methods section, particularly by specifying which results validate which proposition or claim. This would make the paper more readable." >"Consider improving the experimental section to more closely relate the texts to the validated claims." `Response:` We would like to thank the reviewer for the constructive suggestion. In our revised version, we will strive to expand the text in Sect. 6. This will include more detailed descriptions that closely and explicitly connect the results presented in Figs. 2-3 and Tabs. 2-5 with the validated conjectures, propositions, and claims discussed in Sects. 4-5. --- `W4.` Minor problem >"In the appendix, there is a minor problem where Eq 30 extends beyond the margin of the paper." `Response:` Thank you for pointing out this formatting issue. We will address this issue in the revision by breaking Eq. 30 into three separate lines to ensure it fits within the margins.	Summary: This paper proposes Deep Graph Mating, a novel task for model reuse in non-Euclidean domains, specifically focusing on GNNs. The goal is to create a child GNN that combines knowledge from pre-trained parent GNNs without requiring re-training, fine-tuning, or ground-truth labels. The process firstly identifies the permutation invariance properties of GNNs, and accordingly explores two naïve methods: Vanilla Parameter Interpolation (VPI) and Vanilla Alignment Prior to Interpolation (VAPI), both of which were shown to be insufficient due to challenges such as parameter misalignment and topological dependencies. The contribution of this paper is to address this issue by proposing a Dual-Message Coordination and Calibration (DuMCC) methodology that optimizes the permutation matrices for parameter interpolation (PMC scheme) by incorporating topological information, and solves over-smoothing by calibrating the message statistics for child GNNs (CMC scheme). This results in performance on par with traditional re-training methods but without the associated training cost, which is validated on seven datasets across node and graph classification as well as semantic parsing tasks. Code and proofs are given in the appendix and the supplementary material. Strengths: I see several merits of this paper: Addressing the challenge of model reuse in GNNs is a significant topic, especially given the increasing scale of graph data and models in this era. The proposed GRAMA bypasses the need for re-training or fine-tuning, which is commonly required in existing model reuse approaches. GRAMA also removes the need for ground-truth labels, enhancing its generalizability in scenarios where such labels are unavailable. These properties make GRAMA suitable for graph analysis scenarios where resources are limited. The authors present a couple of effective schemes PMC and CMC to deal with the challenges in GRAMA, specifically focusing on parent models and child model. These proposed methods seem overall straightforward yet supported by strong motivations and solid theoretical analysis, such as the amplified sensitivity to parameter misalignment and the increased susceptibility to over-smoothing in child models. The paper also conducts a thorough analysis of these methods across seven graph benchmarks and demonstrates good performance. I see competitive values in Tables 4,5 for the large-scale point cloud benchmarks, without needing to re-training. Overall, the paper is well-written and the main ideas, arguments, and algorithm are very easy to follow, with ample details provided in the appendix. Weaknesses: I see the following weaknesses in this paper: While I appreciate the authors’ attempt to present the complex trajectory of the paper (from defining the problem and elaborating the task motivation, to developing two naïve methods and analyzing challenges, and finally proposing two schemes) in a narrative style for easier understanding, there is an imbalance in the organization of content. The current organization of the paper only allocates less than two pages to the experiment section, leaving many experimental details to the appendix. I suggest the authors move more experimental details from the appendix into the main body of the paper, reducing the current extensive focus on the motivations of the method, and instead, expanding more information about the datasets used, enhancing the details and settings of the comparative methods, and elaborating on the experimental procedures. Also, the connection between the main paper and the appendix is very limited. The authors vaguely direct readers to Section D for further details, without specifying specific tables or paragraphs, which makes it hard to locate relevant information. Also, the explanation regarding the heterogeneous GRAMA in Line 123 actually makes me confused. The statement is too vague. What exactly are the key differences between the homogeneous and heterogeneous cases of GRAMA? Does it specifically relate to the same or different architectures of the pre-trained models, or does it also involve models specializing in different tasks? For example, if there are two pre-trained models with the same architecture but trained for different tasks, would this scenario be classified as homogeneous or heterogeneous GRAMA? Furthermore, while the authors claim that “our initial investigation in this paper is confined to scenarios where pre-trained GNNs possess identical architectures yet are trained on separate datasets”, the authors should at least provide some insights or potential solutions for solving heterogeneous GRAMA, or perhaps consider removing this statement of heterogeneous GRAMA since this is not the contribution of this paper. Following this discussion of heterogeneous GRAMA, I think it is at least worth a try to apply GRAMA to two different GNN variants, such as GCN and GraphSage, since the key learnable parameters in these GNNs are all MLPs. The existing literature has demonstrated the importance of MLPs to GNN performance [1]. Exploring this avenue could be very interesting; if successful, the contributions of this paper would be significantly enhanced. Furthermore, the results may shed light on the connections between different GNN variants in the weight space. In Tables 2 to 5, the authors overlook an important comparative result: the performance of retraining a multi-dataset model that can jointly combine the expertise of both parent models. While this approach involves retraining, it can at least serve as an upper bound for GRAMA, showing the potential space for further improvement. It is also not very clear to me why the authors chose a 20%/80% splitting ratio in the experiments, as there are too few explanations about the splitting protocol. For example, is the split random? More details should be provided, as this concerns the pre-trained models and can reveal how knowledge from the two pre-trained models is distributed and when the proposed method is effective. Would other splitting, such as 10/90, work for GRAMA? It is necessary to conduct more experiments with more splitting ratios to make the authors’ claims more convincing. The current discussion of limitations in this paper appears somewhat constrained. The authors should consider expanding this section, either in the main body of the paper or in the appendix. Another limitation of this work, compared to previous approaches such as KA, is the combination of models working on different levels of tasks, such as node classification and graph classification. Weight-space combination for models addressing various tasks has recently been explored in the Euclidean domain such as ZipIt [2], making it intriguing to see whether similar results can also be achieved in the non-Euclidean domain. Minor issues in the paper: Line 26: "reduing" $\rightarrow$ "reducing"; Line 106: "presents" $\rightarrow$ "present"; Line 193: "are" $\rightarrow$ "is". [1] Han, Xiaotian, et al. MLPInit: Embarrassingly Simple GNN Training Acceleration with MLP Initialization. ICLR 2023. [2] Stoica, George, et al. ZipIt! Merging Models from Different Tasks without Training. ICLR 2024. Technical Quality: 3 Clarity: 4 Questions for Authors: See weaknesses. Overall, I believe this paper addresses a significant problem and proposes a novel approach for model reuse in GNNs. The motivation is strong, supported by solid theoretical analysis. Despite the merits, I see several weaknesses. Therefore, I rate the paper as borderline at this stage and look forward to the authors' responses and further discussions regarding my concerns in the weaknesses section. Confidence: 5 Soundness: 3 Presentation: 4 Contribution: 3 Limitations: Yes, limitations and broader impacts have been discussed. An additional limitation to consider would be the scenario where two pre-trained models address tasks at different levels, such as node and graph-level tasks. Flag For Ethics Review: ['No ethics review needed.'] Rating: 7 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: ### Response to Reviewer Uv5e (Part 1/2) We truly appreciate the reviewer's insightful comments, and would like to address them as follows. Due to character limitations, we have to split our response into two parts. The second part will be provided as a comment following our initial response. `W1.` Imbalanced organisation >"The current organization of the paper only allocates less than two pages to the experiment section, leaving many experimental details to the appendix." `Response:` We appreciate the reviewer for the suggestion. Given the novelty of the studied GRAMA task and the sophisticated rationale behind the proposed DuMCC, our organisational approach was intended to provide detailed explanations, thus facilitating easier comprehension for the readers. These include: - `Sect. 3`: Motivation, definition, and applications of the GRAMA task; - `Sect. 4.1`: Vanilla solutions for the GRAMA task described in `Sect. 3`; - `Sect. 4.2`: Challenges associated with the vanilla solutions from `Sect. 4.1`, leading to the motivations for our proposed DuMCC in `Sect. 5`; - `Sect. 5.1`: An overview of our proposed DuMCC, motivated by the challenges outlined in `Sect. 4.2`; - `Sect. 5.2`: The first component of DuMCC, PMC, including an analysis of its shortcomings, which subsequently motivates the introduction of CMC; - `Sect. 5.3`: The second component of DuMCC, CMC, motivated by the findings in `Sect. 5.2`. The sophisticated logic detailed above may require some pages to clarify potential confusion arising from our novel task and method. Admittedly, our preference for clarity, on the other hand, made us have to relocate some details to the appendix to adhere to page limits, exactly as the reviewer suggested. In the revision, we will strive to achieve a better balance by incorporating as many details as possible within the scope of the additional content page allowed. Our detailed plan is as follows: - Relocating portions of Sect. B and Sect. C to Sect. 6; - Reorganising the architectural details in Sect. D into narrative texts and transferring the associated experimental details from Sect. D to Sect. 6. --- `W2.` Limited connection between the main paper and the appendix >"The authors vaguely direct readers to Section D for further details, without specifying specific tables or paragraphs, which makes it hard to locate relevant information." `Response:` We apologise for any inconvenience caused by the insufficient linkage between the main paper and the appendix, which may have extended the time required for reviewing our paper. Following the transfer of experimental details from the appendix to the main paper in `W1`, we will endeavour to enhance the connectivity of the remaining implementation details within the appendix to the main paper in our revision. --- `W3.` Vague heterogeneous GRAMA descriptions >"The explanation regarding the heterogeneous GRAMA in Line 123 actually makes me confused. The statement is too vague. What exactly are the key differences between the homogeneous and heterogeneous cases of GRAMA?" `Response:` We apologise for not having made our heterogeneous GRAMA clearer, which may have confused the reviewer. We would like to clarify that the heterogeneous case of GRAMA serves as a complementary extension to the proposed homogeneous GRAMA framework in our paper, where pre-trained parent models share similar architectures and purposes. We will explicitly highlight and detail this distinction in Sect. 3 of our revised version. --- `W4.` Potential solutions for heterogeneous GRAMA >"The authors should at least provide some insights or potential solutions for solving heterogeneous GRAMA, or perhaps consider removing this statement of heterogeneous GRAMA since this is not the contribution of this paper." `Response:` We appreciate the advice from the reviewer. A potential solution could involve Partial GRAMA, which entails first identifying shared features between two parent models that have different architectures or handle varied tasks. Subsequently, this possible method would yield a multi-head child GNN that selectively integrates elements of the pre-trained parent GNNs. Although this method would increase the model size compared to full GRAMA, which combines the entire models, it would be expected to offer a more favourable balance and trade-off between model size and performance. We will incorporate this discussion in the revised paper. --- `W5.` Applying GRAMA to variants of GNNs >"I think it is at least worth a try to apply GRAMA to two different GNN variants, such as GCN and GraphSage, since the key learnable parameters in these GNNs are all MLPs." `Response:` The reviewer's point is very well taken. Our immediate-next goal is, exactly as the reviewer suggested, to adapt our proposed method for cross-architecture model reuse scenarios, including the combination of GCN and GraphSAGE as outlined in lines 348-349 of our paper. Here, to echo the reviewer's concern, we conducted a pilot study by performing additional experiments on GRAMA using both a GCN and a GraphSAGE model on non-overlapping partitions of the ogbn-arxiv dataset. These two parent models share a similar overall architecture but differ in their respective GCN and GraphSAGE layers. For the GraphSAGE model, we employed the official SAGEConv implementation from the Deep Graph Library (DGL), setting the 'aggregator_type' to 'gcn' for this preliminary study. This adjustment is due to the design of our homogeneous GRAMA, which requires alignment of parameters with the same shape and number. The results, as shown below, highlight the promising potential and feasibility of cross-architecture GRAMA applications. \| Parent Models \| Architectures \| Performance: Dataset 1 \| Performance: Dataset 2 \| \|:----- \| :----- \| :----- \| :----- \| \| Parent 1 \| GCN \| 0.6908 \| 0.5752 \| \| Parent 2 \| GraphSAGE \| 0.6181 \| 0.7508 \| \| Child \| GCN \| 0.6082 \| 0.5746 \| --- Rebuttal Comment 1.1: Title: Quesiton about homogeneous GRAMA Comment: I want to thank the authors for their detailed responses and the extensive efforts put into conducting new experiments. Most of my questions have been answered satisfactorily. However, I still find the clarification on heterogeneous GRAMA somewhat vague. As I mentioned in my review, in the case of two pre-trained GNN with exactly the same architecture but trained for different domain tasks, should this scenario be classified as homogeneous or heterogeneous GRAMA? I would appreciate further elaboration on this point. --- Reply to Comment 1.1.1: Title: Response to Reviewer Uv5e Comment: >"I want to thank the authors for their detailed responses and the extensive efforts put into conducting new experiments. Most of my questions have been answered satisfactorily. However, I still find the clarification on heterogeneous GRAMA somewhat vague. As I mentioned in my review, in the case of two pre-trained GNN with exactly the same architecture but trained for different domain tasks, should this scenario be classified as homogeneous or heterogeneous GRAMA? I would appreciate further elaboration on this point." `Response:` We appreciate the reviewer's follow-up discussion and would like to apologise for not explicitly defining and explaining heterogeneous GRAMA in our paper and the rebuttal. We would like to clarify that heterogeneous GRAMA refers to scenarios where pre-trained parent models have either diverse architectures or are designed for different domain tasks. Indeed, the case mentioned by the reviewer falls under heterogeneous GRAMA. In our revision, we will address the reviewer's comments by explicitly defining heterogeneous GRAMA. Specifically, we will replace the sentence in lines 123-124 of the main paper, "The exploration into varied architectures and tasks (i.e., heterogeneous GRAMA) remains a topic for subsequent future studies", with the following: "We reserve the exploration of more challenging heterogeneous GRAMA scenarios, where pre-trained parent models either have diverse architectures or are designed for different domain tasks, as a topic for future studies." --- Rebuttal 2: Title: Response to Reviewer Uv5e (Part 2/2) Comment: `W6.` Comparative upper bound results of re-training >"In Tables 2 to 5, the authors overlook an important comparative result: the performance of retraining a multi-dataset model that can jointly combine the expertise of both parent models." `Response:` We appreciate the reviewer's constructive suggestion. Following the reviewer's advice, we conducted additional experiments by re-training a model using combined parent datasets with ground-truth labels. The results are presented below. As the reviewer pointed out, these results can indeed establish an upper bound for GRAMA's performance. We will incorporate these results into Sect. 6 and provide the corresponding discussion in the revision. \| Tables \| Datasets \| Parent 1 \| Parent 2 \| Re-training \| \|:----- \| :----- \| :----- \| :----- \| :----- \| \| Tab. 2 \| ogbn-arxiv \| 0.7193 / 0.5516 \| 0.6564 / 0.7464 \| 0.6903 / 0.7268 \| \| Tab. 2 \| ogbn-products \| 0.7982 / 0.7308 \| 0.7626 / 0.7904 \| 0.7981 / 0.7787 \| \| Tab. 3 \| ogbg-proteins \| 0.7478 \| 0.7222 \| 0.7514 \| \| Tab. 3 \| ogbn-molbace, ogbg-molbbbp \| 0.7247 / 0.4681 \| 0.4067 / 0.6366 \| 0.6954 / 0.6087 \| \| Tab. 4 \| ModelNet40 \| 0.9159 / 0.8151 \| 0.8862 / 0.9275 \| 0.9390 / 0.9243 \| \| Tab. 5 \| S3DIS \| 0.8181 \| 0.8174 \| 0.8428 \| --- `W7.` Experiments with additional splitting ratios >"It is also not very clear to me why the authors chose a 20%/80% splitting ratio in the experiments, as there are too few explanations about the splitting protocol. For example, is the split random? Would other splitting, such as 10/90, work for GRAMA?" `Response:` We would like to thank the reviewer for the constructive comments. In fact, we have indeed addressed the choice of our dataset partition strategy in lines 299-302 of the main paper, following the well-established methodology of "Git Re-basin" by Ainsworth et al. [1], which involves random partitioning of the dataset. Following the reviewer's suggestion, we conducted further experiments with an additional 90%/10% random split on the relatively large-scale ModelNet40 dataset, used for 3D object recognition tasks. The results, presented below, further substantiate the effectiveness of our proposed method. Datasets \| Parent 1 \| Parent 2 \| Ours (w/o CMC) \| Ours (w/ CMC) \|:----- \|:----- \| :----- \| :----- \| :----- \| \| Dataset 1 \| 0.9246 \| 0.8393 \| 0.8451 \| 0.8782 \| \| Dataset 2 \| 0.7621 \| 0.9294 \| 0.8374 \| 0.8422 \| --- `W8.` Constrained limitation discussions >"The current discussion of limitations in this paper appears somewhat constrained. The authors should consider expanding this section, either in the main body of the paper or in the appendix. An additional limitation to consider would be the scenario where two pre-trained models address tasks at different levels, such as node and graph-level tasks." `Response:` The reviewer's point is well taken. Indeed, our current DuMCC framework does not accommodate scenarios where the parent models tackle tasks at different levels, a challenge that falls within the scope of heterogeneous GRAMA as outlined in `W3`. To address this, we will enhance our discussion on limitations by introducing a new Sect. G titled "Limitation Discussion" in the appendix. This section will offer a detailed analysis of this limitation and provide the corresponding potential solutions for such cases, as exemplified in `W4`. --- `W9.` Typos >"Line 26: 'reduing' - 'reducing'; Line 106: 'presents' - 'present'; Line 193: 'are' - 'is'." `Response:` We appreciate the reviewer's thorough examination of our paper. We will address these typographical errors in our revisions by correcting "reduing" to "reducing" in line 26, removing the extraneous "s" from "presents" in line 106, and adjusting "are" to "is" in line 193 to ensure grammatical correctness.	Summary: The paper tackles pre-trained model fusion for graph-centric tasks. The pre-trained models share the same architecture, but differ in the graph datasets they are trained on. The pipeline involves two core approaches. The first one matches parameters in pre-trained parent models by aligning the aggregated messages of the pre-trained parent models. The second one modifies the message statistics for the child model to correspond with the overall statistics of the pre-trained models. The authors have validated two approaches on diverse datasets and models and show nice visualization results. Strengths: S1. As far as I know, this is the first work to study model fusion for graph tasks. The idea is novel and has potential applicability to other networks like Transformers. S2. Throughout the paper, the author examines alternative approach and the challenges associated with each to more effectively justify their choices. S3. The authors show quantitative and qualitative results on diverse datasets for convincing validation throughout their experiments. I found the visualization results in the experiments to be nice and show the effectiveness of the approach. Weaknesses: W1. Discussions on existing model fusion approaches for CNNs are not adequate. The presentation will benefit from a more comprehensive literature review. W2. Lack of sensitivity analysis on the interpolation coefficient alpha. W3. It seems that each task uses only one network architecture and dataset partition. More experiments are needed with more network architectures. W4. It will be great to showcase more applications of the proposed approach besides training-free model reuse. Technical Quality: 3 Clarity: 2 Questions for Authors: Address weaknesses as mentioned above. Confidence: 4 Soundness: 3 Presentation: 2 Contribution: 3 Limitations: Yes. There are no obvious negative society impacts I can notice. Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: ### Response to Reviewer 856w We would like to thank the reviewer for the very helpful feedback. We will address each of the reviewer's comments in detail as follows. `W1.` Inadequate literature review >"Discussions on existing model fusion approaches for CNNs are not adequate. The presentation will benefit from a more comprehensive literature review." `Response:` We appreciate the reviewer for the advice. We will elaborate on the Model Merging section in Sect. 2 of the main paper's related work, and add a new section in Sect. E (extended related work) of the appendix to offer a more comprehensive discussion on existing model fusion techniques for CNNs and transformers. Specifically, we plan to broaden the discussions on model fusion for CNNs and transformers in the revision, covering, but not limited to, the following: - Singh and Jaggi [r1] frame the task of one-shot neural network merging based purely on model weights as model fusion. They propose an Optimal Transport Fusion (OTFusion) method that uses the Wasserstein distance to align weight matrices prior to performing parameter fusion, eliminating the need for re-training; - The subsequent work [r2] extends the application of OTFusion to Transformer-based architectures, aiming to enhance efficiency and performance through fusion; - Liu et al. [r3] approach the challenging task of model fusion as a graph matching problem, incorporating second-order parameter similarities to improve the effectiveness of model fusion; - Addressing the fusion of pre-trained models trained on disparate tasks, Stoica et al. [r4] develop ZipIt!, a novel method that utilises a "zip" operation for layer-wise fusion based on feature redundancy, creating a versatile multi-task model without additional training; - More recently, Xu et al. [r5] present the Merging under Dual-Space Constraints (MuDSC) framework. This approach optimises permutation matrices by mitigating inconsistencies in unit matching across both weight and activation spaces, targeting effective model fusion in multi-task scenarios. [r1] Singh and Jaggi. Model fusion via optimal transport. In NeurIPS, 2020. [r2] Imfeld, et al. Transformer fusion with optimal transport. In ICLR, 2024. [r3] Liu, et al. Deep neural network fusion via graph matching with applications to model ensemble and federated learning. In ICML, 2022. [r4] Stoica, et al. Zipit! merging models from different tasks without training. In ICLR, 2024. [r5] Xu, et al. Training-free pretrained model merging. In CVPR, 2024. --- `W2.` Sensitivity analysis on $\alpha$ >"Lack of sensitivity analysis on the interpolation coefficient alpha." `Response:` We appreciate the reviewer's comments. In fact, the results of sensitivity analysis and the corresponding detailed discussions have already been reported in Sect. C of the appendix. We have also referenced these results of sensitivity analysis in line 292-293 of the main paper. We will further highlight these results in the revision. --- `W3.` Experiments with more network architectures >"It seems that each task uses only one network architecture and dataset partition. More experiments are needed with more network architectures." `Response:` We appreciate the constructive feedback from the reviewer. Following the suggestions provided, we conducted additional experiments with various network architectures and dataset splits on the large-scale ModelNet40 dataset. The results, along with detailed descriptions of the architectures used, are shown below. We will incorporate these results into the revised version of the paper. \| Architectures \| Layers \| Feature Map Channels \| MLP \| \|:----- \|:----- \| :----- \| :----- \| \| Architecture-Main \| 8 \| [64, 64, 128, 256, 1024] \| [512, 256, 40] \| \| Architecture-Rebuttal \| 7 \| [32, 32, 64, 128, 512] \| [256, 40] \| \| Architectures \| Partition \| Datasets \| Parent 1 \| Parent 2 \| Ours (w/o CMC) \| Ours (w/ CMC) \|:----- \|:----- \|:----- \| :----- \| :----- \| :----- \| :----- \| \| Architecture-Main \| 10%/90% \| Dataset 1 \| 0.9246 \| 0.8393 \| 0.8451 \| 0.8782 \| \| Architecture-Main \| 10%/90% \| Dataset 2 \| 0.7621 \| 0.9294 \| 0.8374 \| 0.8422 \| \| Architectures \| Partition \| Datasets \| Parent 1 \| Parent 2 \| Ours (w/o CMC) \| Ours (w/ CMC) \|:----- \|:----- \|:----- \| :----- \| :----- \| :----- \| :----- \| \| Architecture-Rebuttal \| 20%/80% \| Dataset 1 \| 0.9184 \| 0.8920 \| 0.8236 \| 0.8846 \| \| Architecture-Rebuttal \| 20%/80% \| Dataset 2 \| 0.8175 \| 0.9299 \| 0.8279 \| 0.8550 \| --- `W4.` Additional applications >"It will be great to showcase more applications of the proposed approach besides training-free model reuse." `Response:` We would like to thank the reviewer for the constructive comments. Here, we demonstrate an additional application of the proposed DuMCC beyond its primary role in training-free model reuse, specifically as a pre-processing step that facilitates more effective graph-based knowledge amalgamation (KA). In the table below, "KA" refers to re-training a student model from random initialisations by amalgamating knowledge from two teachers pre-trained on distinct partitions of the ModelNet40 dataset. "KA + GRAMA" denotes the process of performing KA by fine-tuning from the child model obtained via the proposed DuMCC method. The results indicate that "KA + GRAMA" leads to improved knowledge amalgamation performance. Additionally, since GRAMA provides a superior initialisation compared to random initialisation, "KA + GRAMA" empirically achieves faster convergence speed. \| Methods \| Performance: Dataset 1 \| Performance: Dataset 2 \| Performance: Average \| \|:----- \| :----- \| :----- \| :----- \| \| KA \| 0.9246 \| 0.9270 \| 0.9258 \| \| KA + GRAMA \| 0.9221 \| 0.9318 \| 0.9269 \| --- Rebuttal Comment 1.1: Title: Response to the rebuttal Comment: Thanks for the well-structured rebuttal. I'm happy with the authors' reply and particularly appreciate the well-prepared and ready-to-incorporate text in the rebuttal, which strongly convinces me that the authors will make the corresponding necessary changes to improve the manuscript. Generally speaking, there is no doubt that the paper makes a clear contribution to both the fields of model fusion and gnn, with novel insights like the amplified parameters sensitivities due to topology, which adds new knowledge to the field. After careful consideration for my final justification, I have decided to raise my score to 6: weak acceptance. There is one minor issue remaining. When reviewing again the related work on model fusion, I noticed several works in the reference list have the wrong pubilcation year. For instance, the well-known git rebasin work ([1] in the paper) was actually published at last year's ICLR, NOT 2022; Zipit [57] was at this year's ICLR, not in 2023. Double check the reference list for such errors. --- Reply to Comment 1.1.1: Title: Response to Reviewer 856w Comment: >"There is one minor issue remaining. When reviewing again the related work on model fusion, I noticed several works in the reference list have the wrong pubilcation year. For instance, the well-known git rebasin work ([1] in the paper) was actually published at last year's ICLR, NOT 2022; Zipit [57] was at this year's ICLR, not in 2023. Double check the reference list for such errors." `Response:` We sincerely appreciate the reviewer for the positive support and the constructive comment regarding the issues in our reference entries. We have thoroughly reviewed all the entries in the reference section and will correct the entries accordingly in our revision as follows: Original references: " [1] Samuel Ainsworth, Jonathan Hayase, and Siddhartha Srinivasa. Git re-basin: Merging models modulo permutation symmetries. In ICLR, 2022. [26] Moritz Imfeld, Jacopo Graldi, Marco Giordano, Thomas Hofmann, Sotiris Anagnostidis, and Sidak Pal Singh. Transformer fusion with optimal transport. In ICLR, 2023. [57] George Stoica, Daniel Bolya, Jakob Brandt Bjorner, Pratik Ramesh, Taylor Hearn, and Judy Hoffman. Zipit! merging models from different tasks without training. In ICLR, 2023. [61] Hongyi Wang, Mikhail Yurochkin, Yuekai Sun, Dimitris Papailiopoulos, and Yasaman Khazaeni. Federated learning with matched averaging. In ICLR, 2019. " Revised references: " [1] Samuel Ainsworth, Jonathan Hayase, and Siddhartha Srinivasa. Git re-basin: Merging models modulo permutation symmetries. In ICLR, 2023. [26] Moritz Imfeld, Jacopo Graldi, Marco Giordano, Thomas Hofmann, Sotiris Anagnostidis, and Sidak Pal Singh. Transformer fusion with optimal transport. In ICLR, 2024. [57] George Stoica, Daniel Bolya, Jakob Brandt Bjorner, Pratik Ramesh, Taylor Hearn, and Judy Hoffman. Zipit! merging models from different tasks without training. In ICLR, 2024. [61] Hongyi Wang, Mikhail Yurochkin, Yuekai Sun, Dimitris Papailiopoulos, and Yasaman Khazaeni. Federated learning with matched averaging. In ICLR, 2020. "	Summary: The paper is interested in re-using Graph Neural Networks trained from one task to another task (e.g., transfer learning). This motivation is nice. We've seen such trends in computer vision (e.g., network trained on ImageNet used for CIFAR), or more recently, everyone is using LLMs for a variety of tasks they were not trained on (by prompt engineering, LoRA, etc). The paper proposes to do the same for graphs. Their specific methodology starts by combining multiple GNNs. To do so, they want to permute the internal channels of the graph neural network so that the networks are best-aligned, before combing them. Strengths: * Using Pre-trained networks can save compute resources * If the initial trained dataset is so large and/or private, it might not be practical to retrain the network (even with much compute resources). Using pre-trained networks would work well here. * Paper proposes to use multiple GNN (pre-trained) networks for a new task. The method is simple: just average the latent vectors of the GNNs. However, for this averaging to make sense, authors propose permuting the channels of the GNN latent vectors. This permutation is optimized using an objective (and algorithm). Weaknesses: While the paper is well-motivated and well-written, it nicely builds-up the reader's excitment to anticipate for the actual method. Once reader arrives to the well-anticipated Equation (3), reader discovers that the equation is not properly-written ("buggy"). I've written many papers with probably subtle bugs in them (e.g., perhaps some footnote or some text in some side-section has a bug). However, a bug in the main equation qualifies for an immediate rejection. ## Math incorrectness After Eq 1: "where $\mathbf{P}^∗$ represents the permutation matrices" -- what does that mean? The expression $\mathbf{P}^∗$ is not used in the Equation. I think it can imply one of two things. Option 1: $\mathbf{P}^∗ = [P^{(\ell)} ]_{\ell \in [L]}$; or Option 2: $\mathbf{P}^∗$ refers to the space of permutation matrices (e.g., a reshuffle of the identity matrix or its continuous relaxation which is probably any basis?) Equation 3 has several issues: 1. Where does $i$ come from? do you mean to add $\sum_i$ between the $\arg\min$ and the Frobenius norm? 2. The aggregation function "Agg" is not introduced. While the exact definition is unnecessary, however, the domain and range should be clearly specified. Does it take a many vectors and output one vector? If yes, then $P \cdot \textrm{Agg}$ should be a vector and therefore you should be minimizing the L2 norm (not Frobenius norm). If it acts on the whole graph (i.e., $\textrm{Agg}$ outputs a matrix with shape `NumNodes x FeatureDim`), then you are multiplying the permutation matrix from the wrong side. 3. In general, where do the $X$s come from? I was expecting data-independent scheme. In my reading, at this point, I think paper is "learning to align" (because you are doing gradient decent on the permutation matrix by looping over data, no?) -- with all honesty, I stopped reading the paper after arriving at Eq3 and skimming over the algorithm. Please address the above and I am happy to take another look during the rebuttal. That being said, I do trust that your implementation is correct, but the paper may not indeed reflect the implementation. Technical Quality: 3 Clarity: 3 Questions for Authors: - Does the exact form of Equation 2 depends on the architecture of the GNN model? I would expect to see different formulas depending on the model (e.g., GraphSAGE vs GCN of Kipf vs GAT vs MixHop etc) - Please respond to the main math weakness, which might cause you to be explicit about the notation Confidence: 5 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: I did not spot a Limitations section. Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: ### Response to Reviewer rPnJ (Part 1/2) We appreciate the reviewer's constructive comments and thoughtful suggestions, and we sincerely apologise for any confusion or ambiguity caused by our use of notations. We are committed to addressing each of the reviewer's concerns as outlined below. Due to character constraints, we have to divide our responses into two parts. The second part will be presented as a comment appended to our initial rebuttal. Additionally, we warmly welcome any further questions or comments the reviewer may wish to share. `W1.` Clarity on $\mathbf{P}$* >"After Eq 1: 'where $\mathbf{P}^$ represents the permutation matrices' -- what does that mean? The expression $\mathbf{P}^$ is not used in the Equation. I think it can imply one of two things. Option 1: $\mathbf{P}^* = \left[\mathbf{P}^{(\ell)}\right]_{\ell \in [L]}$ or Option 2: $\mathbf{P}^$ refers to the space of permutation matrices (e.g., a reshuffle of the identity matrix or its continuous relaxation which is probably any basis?) `Response:` We apologise for not having clearly defined the symbol $\mathbf{P}^$, which may have led to confusion and inadvertently extended the review process. We appreciate the reviewer's patience and constructive comments. We would like to clarify that $\mathbf{P}^$ denotes the set of all permutation matrices corresponding to each layer $\ell$ of the graph neural network (GNN), which is exactly Option 1 suggested by the reviewer. We would like to further clarify that while $\mathbf{P}^$ does not appear in Eq. 1, it is utilised in line 158, Eq. 3, Conjecture 4.1, and lines 208, 210, and 222. We intended to use this notation to simplify the discussions related to the collection of permutation matrices across different layers. Admittedly, we sincerely apologise for the oversight in not explicitly linking Eq. 1 with $\mathbf{P}^$ and for not adequately defining and elaborating on this symbol, which resulted in ambiguity. To address this issue, we will revise Eq. 1 in line with the reviewer's constructive suggestion and enhance the description of $\mathbf{P}^$ in our revision by incorporating the following: "$$ W^{(\ell)} = \alpha W_{a}^{(\ell)} + (1 - \alpha) P^{(\ell)} W_{b}^{(\ell)} (P^{(\ell-1)})^{T}, \quad P^{(\ell)} \in \mathbf{P}^, \tag{1} $$ where $\mathbf{P}^ = \left[{P}^{(\ell)}\right]_{\ell \in [L]}$ represents the set of all permutation matrices $P^{(\ell)}$ for each layer $\ell$ of the graph neural network (GNN). Here, $[L]$ refers to the set of indices corresponding to all layers in the GNN." --- `W2.` Issues on Eq. 3 `W2.1. & W2.2.` Clarity on $i$ and "Agg" >"Where does $i$ come from? do you mean to add $\sum_i$ between the $\arg \min$ and the Frobenius norm?" >"The aggregation function 'Agg' is not introduced. While the exact definition is unnecessary, however, the domain and range should be clearly specified. Does it take many vectors and output one vector? If yes, then $P \cdot \text{Agg}$ should be a vector and therefore you should be minimizing the L2 norm (not Frobenius norm). If it acts on the whole graph (i.e., Agg outputs a matrix with shape NumNodes $\times$ FeatureDim), then you are multiplying the permutation matrix from the wrong side." `Response:` We sincerely appreciate the reviewer's constructive feedback and apologise once again for the lack of clarity in our notations and the rigor of our equations. Since the issues concerning $i$ and $\text{Agg}$ are interrelated, we are combining our responses to W2.1 and W2.2 to address these two issues collectively. Originally, our intention was to use the symbol $i$ to universally represent the set of all nodes, rather than a single node. Consequently, $\text{Agg}$ was meant to act on the entire graph, outputting a matrix. Here, we would like to clarify that this output matrix was shaped as FeatureDim $\times$ NumNodes. This configuration was deliberately chosen to align with the conventions used in model merging within the Euclidean domain (e.g., Eq. 1 in [r1] and Eq. 13 in [r2], where the corresponding matrices are denoted as Dim $\times$ Num). This alignment was intended to facilitate a unified formulation and simplify conceptually intuitive comparisons across different domains. As a result, in our original Eq. 3, we multiplied the permutation matrix from the left side to match this dimension. This choice also explains our use of the Frobenius norm, as the operation is performed on the matrix. Admittedly, exactly as the reviewer kindly suggested, the definition and use of the symbol $i$ in this context lacks mathematical rigor, correctness, and explicitness, which subsequently led to confusion concerning $\text{Agg}$ and the Frobenius norm. We sincerely appreciate the reviewer's thorough advice on this issue. In our revision, we will address this issue by exactly following the reviewer's suggestion: defining $i$ as a single node identifier and introducing $\sum_i$ after $\arg \min$ to explicitly iterate over all nodes. This adjustment will result in the output of $\text{Agg}$ being a vector. Accordingly, we will replace the original Frobenius norm with the L2 norm to ensure consistency with these changes. We sincerely thank the reviewer once again for the insightful comments. We will revise Eq. 3 as described above, along with including the corresponding detailed descriptions for rigorous and explicit formulation and symbol definitions. Additionally, we will also update Eq. 3 in Alg. 1 accordingly. [r1] Ainsworth, et al. Git re-basin: Merging models modulo permutation symmetries. In ICLR, 2023. [r2] Li, et al. Deep model fusion: A survey. 2023. --- Rebuttal 2: Title: Response to Reviewer rPnJ (Part 2/2) Comment: `W2.3(a)` Clarity on $\mathbf{X}$s >"In general, where do the $\mathbf{X}$s come from? I was expecting data-independent scheme." `Response:` The reviewer's point is very well taken. In our paper, we initially introduced a fully data-independent approach for GRAMA, referred to as the vanilla VAPI method, described in lines 158-161: "One possible data-independent solution is to minimise the L2 distance between the weight vectors of the pre-trained models by solving a sum of bilinear assignments problem, similar to weight matching techniques described in [1]." However, as discussed in Sect. 4.2, our subsequent analysis reveals that GNNs are particularly sensitive to mismatches in parameter alignment, which are "contingent upon the topological characteristics inherent to each graph" (Conjecture 4.1). This sensitivity renders the data-independent solution less effective. Motivated by this, we propose integrating the topological characteristics inherent to each graph. To achieve this, we have to resort to passing the graph data to the pre-trained model to capture these graph-specific topological characteristics, utilising $\mathbf{X}$. Nevertheless, we clarify that our PMC and CMC methodologies require only a single forward pass of the unlabelled graph data to extract messages for alignment and calibration, respectively—eliminating the need for iterative training or ground-truth labels. Our immediate-next goal is, exactly as the reviewer suggested, to explore the possibility of an entirely data-independent GRAMA scheme, for example, by initially generating fake graphs as described in [r3]. We will include these discussions in the revised version. [r3] Deng and Zhang. Graph-free knowledge distillation for graph neural networks. In IJCAI, 2021 `W2.3(b)` Details of aligning >" In my reading, at this point, I think paper is 'learning to align' (because you are doing gradient descent on the permutation matrix by looping over data, no?)" `Response:` We apologise for any confusion caused by our previous lack of clarity and the omission of essential details. We would like to clarify that our method does not employ gradient descent for alignment. In the field of model merging within the Euclidean domain, the minimisation problem in the form of Eq. 3 is typically transformed into a maximisation problem to maximise an inner product (as derived from expanding Eq. 3), thereby fitting it within the framework of a standard linear assignment problem [r1, r2, r3, r4]. In our revision, we will include these crucial details and introduce a preliminary section on Euclidean model merging to ensure our paper is self-contained. [r1] Ainsworth, et al. Git re-basin: Merging models modulo permutation symmetries. In ICLR, 2023. [r2] Li, et al. Deep model fusion: A survey. 2023. [r3] Liu, et al. Deep neural network fusion via graph matching with applications to model ensemble and federated learning. In ICML, 2022. [r4] Stoica, et al. Zipit! merging models from different tasks without training. In ICLR, 2024. --- `Q1.` Model-specific variants of Eq. 2 >"Does the exact form of Equation 2 depends on the architecture of the GNN model? I would expect to see different formulas depending on the model (e.g., GraphSAGE vs GCN of Kipf vs GAT vs MixHop etc)" `Response:` We appreciate the constructive comments provided by the reviewer. To maintain clarity and simplicity, Eq. 2 in our paper is based on the most basic form of GNNs. Due to the short rebuttal period, we have only been able to develop specific formulas for GraphSAGE and Kipf's GCN as suggested by the reviewer. Our model formulation presented here follows the unified mathematical framework detailed in [r5], albeit with symbols adapted to those used in our paper. In the revised version, we are committed to developing more detailed, model-specific formulas. (We apologise for having to split each of our equations below into two lines for display purposes, as OpenReview does not support single-line presentation of long equations.) Eq. 2 tailored for GCN of Kipf is as follows: $$ \Delta F_i \approx \sigma'\left(W \sum_{j \in \mathcal{N}(i)} \frac{X_j}{\sqrt{\deg}_i \sqrt{\deg}_j}\right) $$ $$ \cdot \left(\epsilon \sum_{j \in \mathcal{N}(i)} \frac{X_j}{\sqrt{\deg}_i \sqrt{\deg}_j}\right) $$ Eq. 2 tailored for GraphSAGE is as follows: $$ \Delta F_i \approx \sigma' \left(W \ \text{Concat}\left(X_i, \text{Mean}_{j \in \mathcal{N}(i)} X_j\right)\right) $$ $$ \cdot \left(\epsilon \ \text{Concat}\left(X_i, \text{Mean}_{j \in \mathcal{N}(i)} X_j\right) \right) $$ [r5] Dwivedi, et al. Benchmarking graph neural networks. In JMLR, 2023. --- `L1.` Limitations >"I did not spot a Limitations section." We apologise for not making the limitations section more explicit in our paper. In fact, we have included a section on limitations in Sect. 7, titled "Conclusions and Limitations". We will highlight this section further in the introduction. --- Rebuttal Comment 2.1: Comment: Since the reviewers have responded to all my comments, I am increasing my score. Thank you for your good work!	null	NeurIPS_2024_submissions_huggingface	2,024	null	null	null	null	null	null	null	null
Practical Shuffle Coding	Accept (poster)	Summary: The design of compression mechanisms for compression of unordered objects is considered. The scenario models various applications of interest such as compression of unlabeled graphs and multisets. A novel recursive solution, called recursive shuffle, is introduced. An advantage of this compression mechanism, compared to previous plain shuffle method, is that it can be applied in one-shot compression scenarios. However, it is noted that since the mechanism requires computation of orbits of prefixes, it cannot be completed in polynomial time. To mitigate this, an alternative `sub-optimal' mechanism is introduced which does not jointly compress all of the order information, and handles part of the information separately. The mechanism is called incomplete recursive shuffle. Various simulations on real-world datasets are provided to compare the performance of the proposed mechanism with the state of the art methods. Strengths: - The compression problem considered in this work is of interest and is applicable to a wide range of graph compression problems. - The method, using a recursive process that interleaves encoding and decoding steps to allow for one-shot compression, and which avoids jointly compressing all order information to reduce run-time is novel and interesting. - The paper is well-written, and the proof arguments and explanations are complete and clear. - The manuscript includes comprehensive numerical simulations and comparisons with the state of the art methods both in terms of compression rate and speed of compression. Weaknesses: - As mentioned in the manuscript, the recursive shuffle mechanism requires computation of the orbits of the prefixes, which cannot be solved in polynomial time. The alternative of separately compressing part of the order information is suboptimal in terms of compression rates. - In several of the scenarios for graph compression, the experimental results do not seem to show significant gains compared to SZIP and Plane shuffle methods. For instance, in Table 5, for one-shot compression, SZIP outperforms the proposed methods in two of the datasets and performs comparably in the other four. Furthermore, while the compression speed is significantly faster than plain shuffle, it is still below that of SZIP in several cases as shown in Table 6. Technical Quality: 4 Clarity: 4 Questions for Authors: - It appears from Tables 5 and 6 that the performance of SZIP is comparable and even better than the proposed methods on average. Please comment on why and in what respect if any do the proposed methods outperform SZIP? - In Table 4, sometimes WL_1 and sometimes WL_2 hashing yields better results. I wonder if there is an intuitive explanation and a way to determine what type of hash should be used beforehand. Also, the proposed methods sometimes outperform SZIP and sometime do not. Please comment if there are specific conditions (such as graph statistics) that would cause one mechanism to outperform in different scenarios. Confidence: 3 Soundness: 4 Presentation: 4 Contribution: 3 Limitations: - The comparison with SZIP performance should be made more explicit and comprehensive in the discussions in Section 6. Flag For Ethics Review: ['No ethics review needed.'] Rating: 7 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We thank reviewer tkqf for taking the time to review our paper. It will be helpful to respond to part of your second question first: > In Table 4, sometimes WL$_1$ and sometimes WL$_2$ hashing yields better results. I wonder if there is an intuitive explanation and a way to determine what type of hash should be used beforehand. Also, the proposed methods sometimes outperform SZIP and sometime do not. We thank the reviewer for pointing this out. Using more Weisfeiler Leman hashing iterations with incomplete recursive shuffle coding will always result in equal or better rates in expectation over all possible initial messages, assuming that hash collisions are rare enough, a condition practically achieved with standard 64-bit hashing methods. Our paper was misleading in this sense, and will be amended, in two ways: - As detailed in our response to reviewer Htst and briefly mentioned in Appendix C, the AP model uses a variational approximation. In the paper we failed to clarify that this can lead to the rate depending on the initial message, making it appear stochastic. We will update the paper accordingly. - In Table 4 we only reported each compression rate result for the AP model based on a single run. This was misleading in the sense that WL$_2$ appeared worse than WL$_1$ for some graphs. As stated in the overall response, we have now repeated each of these experiments three times with different initial message seeds, and the new results, shown in Table S1 of the supplemental page, are compatible with the prediction that WL$_2$ has a rate at least as good as WL$_1$. We now also report empirical standard deviations to characterize the stochasticity of the AP model. We will now respond to the remaining questions together: > Also, the proposed methods sometimes outperform SZIP and sometime do not. Please comment if there are specific conditions (such as graph statistics) that would cause one mechanism to outperform in different scenarios. > It appears from Tables 5 and 6 that the performance of SZIP is comparable and even better than the proposed methods on average. Please comment on why and in what respect if any do the proposed methods outperform SZIP? As described in the overall response, our new results in Table S1 show that incomplete shuffle coding now comfortably outperforms SZIP both with AP/WL$_1$ and AP/WL$_2$ in terms of one-shot compression rate. Our proposed methods are entropy coding methods, with optimal rates for a model that can be easily inspected, improved, and swapped out. Shuffle coding methods can therefore be easily adapted to specific domain models, and automatically benefit from advances in generative graph modeling. The model update mentioned above leading to the drastically improved rates from Table S1 demonstrates this capability. This is not possible for SZIP, since it is not based on an explicit probabilistic model. If its implicit model fails, there is no known method to improve it. Similarly, it is difficult to reason about the strengths of SZIP. One thing that we do know is that for large enough graphs, SZIP cannot perform significantly worse than shuffle coding with an Erdős–Rényi graph model, due to the rate upper bound discussed in Section 5, $\log\frac{1}{P_\mathrm{ER}(g)} - n\log n + O(n)$. It is hard to know how large graphs have to be for this bound to have practical significance, since the constant factor for $O(n)$ is unknown. In contrast, we can easily inspect the explicit model used by shuffle coding. We know, for example, that preferential attachment models, like the Pólya urn models (PU / AP) used in our experiments, are suitable whenever there is a ‘rich get richer’ dynamic in the generating process, meaning the rate at which new neighbors are attached is approximately proportional to the number of neighbors already present, leading to skewed neighbor count distributions (see Severo et al. 2023 for more details). This dynamic is present in many natural contexts such as social networks or web graphs, providing a possible explanation for good performance of Pólya urn models on such graphs. We will extend the discussion on SZIP in Section 6 accordingly to be more comprehensive. ### References Severo, Daniel, et al. (2023): Random Edge Coding: One-Shot Bits-Back Coding of Large Labeled Graphs." arXiv preprint arXiv:2305.09705. --- Rebuttal Comment 1.1: Comment: I appreciate the authors' comprehensive response. My concerns regarding the comparison with SZIP have been fully addressed. Corrections have been made to the experimental setup which now yield consistent outcomes for WL1 and WL2 hashing. I have updated my score to reflect the improvements.	Summary: This paper proposes recursive shuffle coding, a general method for optimal compression of unordered objects using bits-back coding. And the paper further presents present incomplete shuffle coding, allowing near-optimal compression of large unordered objects with intractable automorphism groups. When combined, these methods achieve state-of-the-art one-shot compression rates on various large network graphs at competitive speeds. Strengths: 1. Incomplete recursive methods improve the speed of processing large-scale unordered multiple sets. 2. The paper is a solid contribution with solid theoretical foundations. Introducing group theory to consider this problem is very interesting. Weaknesses: 1. The two methods seem to address different issues and do not clarify their connection. 2. The symbols in the paper are too numerous, making it difficult to intuitively understand your intention. Technical Quality: 3 Clarity: 2 Questions for Authors: 1. Could you specifically explain more details of the Autoregression PU model in Appendix C? 2. Could you clarify the connection between recursive shuffle coding and incomplete shuffle coding? Confidence: 4 Soundness: 3 Presentation: 2 Contribution: 2 Limitations: 1. Less evaluation on unordered graphs with vertex and edge attributes. 2 Answer my questions Q1 and Q2. Flag For Ethics Review: ['No ethics review needed.'] Rating: 4 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We thank reviewer Htst for their review, comments, and questions. ## Weaknesses > The two methods seem to address different issues and do not clarify their connection. Your observation is correct, the two methods address different issues. They appear together in this paper since it is convenient to use recursive shuffle coding to implement incomplete shuffle coding, and their advantages can be combined this way. Specifically, we implement incomplete shuffle coding by modifying the orbits function to return orbits of an incompletely ordered object, approximating the orbits (and automorphism group) of the underlying ordered object. It would be feasible for incomplete shuffle coding to be based on plain shuffle coding instead, requiring a function that returns an approximate canonization and automorphism group for a given object. While this would be an interesting direction of research, readily available graph isomorphism libraries do not provide such functionality. Therefore, we leave implementing this idea for future work and focus on the approach based on recursive shuffle coding. > Less evaluation on unordered graphs with vertex and edge attributes. As stated in the overall response, we ran additional experiments on a vast array of graph datasets featuring vertex and edge attributes, confirming that the favorable properties of our method extend to attributed graphs. ## Questions > The symbols in the paper are too numerous, making it difficult to intuitively understand your intention. While we agree that this is a technical paper, we spent great effort on visualizations to help motivate and clarify the technical concepts required for formalizing our method. In particular, we strongly encourage the reviewer to revisit the example in Tables 1 and 2, as well as Figure 1. > Could you specifically explain more details of the Autoregression PU model in Appendix C? Yes. We will extend Appendix C and also give a detailed explanation here. The joint model we want to approximate is the Pólya urn model from Kunze et al. (2024). Its iterative generative process starts from an empty graph, into which a fixed number of edges are inserted iteratively. The two vertices forming the next inserted edge are sampled with probabilities proportional to the current number of vertex neighbors (+1). This process favors vertices that already have many neighbors in a ‘rich get richer’ dynamic, often called ‘preferential attachment’, with a skewed distribution over vertex neighbor counts. Herby, self-loops and redraws are disallowed. This breaks edge-exchangeability, leading to a 'stochastic' codec, meaning that the code length depends on the initial message. Shuffle coding is compatible with such models. In this more general setting, the ordered log-likelihood term in the optimal rate of Eq. 3 is replaced with a variational `evidence lower bound' (ELBO). The discount term is unaffected. The derivations in the main text are based on the special case of exchangeable models, where log-likelihoods are exact, for simplicity. They can be generalized with little effort and new insight. We will now describe the approximate Pólya urn (AP) model that is autoregressive over graph slices. A graph slice $f_i$ can be represented as the set of edges connecting the vertex $i$ with any following vertex, i. e. a subset of $\lbrace (i, i+1), (i, i+2), … (i, n-1) \rbrace$. Any autoregressive graph model will provide a distribution over slices $f_i$, given the graph prefix $f_{[i]}$ (that comprises all previous slices). We can break down the task into two parts: Given the prefix $f_{[i]}$, predict the number of edges $k_i \in [n - i]$ within the next slice $f_i$, and then from that, predict the edge positions within the slice. For the Pólya urn model, the distribution for the first part, $P(k_i \| f_{[i]})$, is not tractable and we instead approximate it by a truncated log-uniform distribution $Q(k_i \| f_{[i]})$ where $\lfloor \log k_i \| f_{[i]} \rfloor$ is (approximately) uniformly distributed within the possible range. The distribution of the second part, $P(f_i \| f_{[i]}, k_i)$, is tractable with the following generative process: Given the graph of the prefix $f_{[i]}$ and the number of edges $k_i$ in the next slice $f_i$, iteratively insert $k_i$ edges, with probabilities proportional to the neighbor count + 1 of the adjacent vertices $\{ i+1, i+2, … n-1\}$ without allowing repetition (i. e. zeroing out the probabilities where edges were already inserted). We then have the slice distribution $Q(f_i \| f_{[i]}) = \sum_{k_i=0}^{n-i} P(f_i \| f_{[i]}, k_i) Q(k_i \| f_{[i]})$, and finally the complete AP model $Q(f) = \prod_{i=0}^{n-1} Q(f_i \| f_{[i]})$. The approximation of $k_i$ increases the stochasticity of the compression rates, much like disallowing self-loops/redraws. As described in the overall response, we conducted further experiments to characterize this stochasticity. We report empirical means and standard deviations of compression rates for the experiments from Table 4 with the AP model, on the supplemental page in Table S1. It shows that the resulting stochasticity is significant, highlighting that our choice of $Q(f_i \| f_{[i]})$ is a crude approximation. Improving it, for example, by using information from the prefix $f_{[i]}$, is left for future work. > Could you clarify the connection between recursive shuffle coding and incomplete shuffle coding? This is answered in the first section of this response.	Summary: This paper proposed an entropy coding method of large unordered data structures. The newly proposed method allows one-shot compression and achieves competitive speed. The experimental results demonstrate the advantages. Strengths: This paper proposed a new entropy coding method for large unordered data sets with near-optimal compression. Specially, the new method allows one-shot compression. In addition, the new method has competitive speed, which is verified by the experiments. Weaknesses: As a contribution, the proposed method in this paper can work for one-shot compression while the SOTA works can not. Hence, the applications of one-shot compression should be presented and stressed in details. Technical Quality: 3 Clarity: 4 Questions for Authors: What are the applications of one-shot compression? Confidence: 2 Soundness: 3 Presentation: 4 Contribution: 3 Limitations: No. The comparison with other works for common cases (other than one-shot compression) should be discussed. Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We thank reviewer NPhs for taking the time to review and comment on our paper. ### Questions > What are the applications of one-shot compression? Example applications include storing/transmitting large social, web, network, or compute graphs, JSON files (nested multisets), machine learning datasets, and relational database tables with many rows (multisets). Large objects of these kinds are often sent/stored separately, making one-shot compression an interesting problem. For this reason, the initial bits problem of plain shuffle coding poses a fairly substantial limitation in practice. ### Limitations > The comparison with other works for common cases (other than one-shot compression) should be discussed. As described and discussed in the overall response, we performed additional experiments on all TU datasets comparing our method to plain shuffle coding outside of the one-shot regime, for graphs with vertex and edge attributes. Specifically, we compress the many graphs per dataset together in sequence. The experiments clearly show that outside of the one-shot regime, incomplete recursive shuffle coding can offer dramatically faster speeds for relatively small increases in compression rate. We hope that this addresses your concerns. --- Rebuttal Comment 1.1: Title: Thank the authors for the response to my comments. Comment: I think the rebuttal addressed my major concerns and I updated the score accordingly.	Summary: Coding of unordered structures is considered. This paper addresses the two main limitations of the entropy coding method proposed by Kunze et al. (2024): high cost of automorphism group calculation and poor compression of single unordered objects. To solve the first problem, it is suggested to approximate the object symmetries. The second problem is solved via a recursive coding, which requires an autoregressive probability model. Experiments are performed to demonstrate the merits of the proposed improvements in comparison to the original (``plain'') entropy coding method. Strengths: - The paper is well written, with all the required preliminary information included. The definitions, theorems, and proofs are mathematically rigorous. However, a great effort has also been made to facilitate the understanding of the article via various examples and clarifications, which is also commendable. - The proposed solutions to the issues raised are elegant and might admit a generalization to arbitrary groups of automorphisms. Weaknesses: - Recursive variant of the method requires an autoregressive probability model to encode the slices. Such probabilities can be untractable or hard to compute in case of complex models. This might limit the compression capabilities of the method in the case of complex structures. For example, the considered ER and PU random graph models might yield significantly suboptimal results for complex networks, but more advanced models might also be inapplicable due to intractable conditional probabilities. - The tradeoff between the approximation of the automorphisms group and encoding/decoding speed and compression rate is under-researched. Only two methods for approximate graph automorphisms calculation are considered. - It seems that the method can be abstracted from the ordered/unordered objects and permutations, and can be described in terms of equivalence relations solely. This approach should unify the "plain" and "approximate" methods and slightly simplify the article. I believe that the current state of the work already requires only the slightest changes to finalize this transition. - In the speed comparison table, the encoding speed values for SZIP are not measured on the same hardware as the rest of the experiments, but adopted from the original article. Technical Quality: 3 Clarity: 3 Questions for Authors: - Have you considered generalizing your method to any objects with symmetries inducing the groups of automorphisms? - Is it possible to extend your method to non-discrete symmetries? For example, one can consider translation and rotation invariance of images and 3D objects. Such symmetries can "disappear" after the floating point discretization due to the machine precision, but it might be possible to define "approximate" automorphisms. - Have you considered other random graph models, e.g., admitting the clustered structure? - Is it possible to conduct more experiments with variable "accuracy" of automorphisms calculation (from very poor to almost exact)? It seems interesting to vary the rate increase given in eq. 7 and measure the change in encoding/decoding speed. - Repetition in Table 4: ``All measurements are in bits per edge. All measurements are in bits per edge.'' - Is it possible to add a table, similar to Table 2, but for graph prefixes and other related concepts? Confidence: 3 Soundness: 3 Presentation: 3 Contribution: 2 Limitations: The authors briefly mention some of the limitations in the conclusion. However, I believe that the limitations related to the requirement of an autoregressive probability model should also be mentioned and discussed thoroughly. Flag For Ethics Review: ['No ethics review needed.'] Rating: 7 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We thank reviewer AUZQ for their thoughtful, thorough review. ## Weaknesses > It seems that the method can be abstracted from the ordered/unordered objects and permutations, and can be described in terms of equivalence relations solely. This approach should unify the "plain" and "approximate" methods and slightly simplify the article. I believe that the current state of the work already requires only the slightest changes to finalize this transition. The described generalization is possible and would allow compressing elements of arbitrary quotient sets. This class of methods can be summarized as bits-back coding (Townsend et al., 2019) restricted to models with deterministic conditional distributions $P(f_\sim \| g)$, which includes all shuffle coding variants. However, the existence of efficient algorithms for (approximate) orbit functions makes permutable classes something of a sweet spot in the trade-off between generality and practicality, hence the choice of framing for this paper. > In the speed comparison table, the encoding speed values for SZIP are not measured on the same hardware as the rest of the experiments, but adopted from the original article. We agree with the reviewer that having results on the same hardware would be valuable. Unfortunately, we have not found a public implementation for SZIP method, making such comparison difficult. ## Questions > Have you considered generalizing your method to any objects with symmetries inducing the groups of automorphisms? Yes. This generalization is straightforward, with the resulting rate being $\log \frac{1}{P(\bar{f})} = \log \frac{1}{P(f)} - \log \frac{\|G\|}{\|\text{Aut}(f)\|}$, where $G$ is the considered group, i. e. what’s $S_n$ in the paper. Hereby, the ratio inside the discount term computes the size of each coset of $\text{Aut}(f)$. All compelling applications that we found have objects that can be arbitrarily permuted. To keep this already quite technical paper more accessible, we chose to present the paper based on the most important case of a full symmetric group $G=S_n$. > Is it possible to extend your method to non-discrete symmetries? Continuous objects have infinite information content, i. e. no probability mass function can be defined over them. Therefore, they require discretization before any lossless compression method can be applied. Our method can be applied to such discretized objects, as usual. The discrete case is therefore sufficient for our practical purposes. In some cases, however, we can still generalize our rate for unordered objects to a ‘rate density’. For example, given a probability density function $p(f)$ over continuous objects f, and under the condition that the size of the cosets $c(f)$ of $\text{Aut}(f)$ is finite, we obtain a probability density function $p(\bar{f})$ over ‘unordered’ objects, via $p(\bar{f}) = p(f) \cdot c(f)$, and the corresponding ‘rate density’ $\log \frac{1}{p(\bar{f})} = \log \frac{1}{p(f)} - \log c(f)$. > Have you considered other random graph models, e.g., admitting the clustered structure? Yes. Generative modeling of graphs is an active area of research, see Zhu et al. (2022) and Maneth et al. (2015) for surveys. It is somewhat orthogonal to our work since any advance in graph modeling will allow better rates with shuffle coding. The models used in our paper have few parameters, with which we outperform competing methods. Finding and applying models that exploit more structure is a promising research direction to improve rates quickly. As explained in more detail in response to reviewer Htst, our simple Pólya models already exploit a ‘rich get richer’ generative process present in many natural contexts (such as social networks), where vertices with many neighbors have a proportionally higher probability to accumulate even more edges. > Is it possible to conduct more experiments with variable "accuracy" of automorphisms calculation (from very poor to almost exact)? It seems interesting to vary the rate increase given in eq. 7 and measure the change in encoding/decoding speed. Yes. As stated in the overall response, we ran an additional experiment with more variants of incomplete shuffle coding, reporting compression rates and speeds. For this, we chose three datasets with relatively small graphs because more WL iterations quickly become too slow on large graphs as local graph convolution features typically degenerate into global ones exponentially. The experiment shows that increasing WL iterations quickly approaches the optimal rate. In all 3 experiments, the exact optimal rate is achieved within 2 to 5 iterations. This corresponds to the fact that the Weisfeiler-Leman graph isomorphism test is effective on practical graphs. > Repetition in Table 4: "All measurements are in bits per edge. All measurements are in bits per edge." Thanks, this will be fixed! > Is it possible to add a table, similar to Table 2, but for graph prefixes and other related concepts? This is a fantastic idea, and we will include it in the final version of the paper. ## Limitations > I believe that the limitations related to the requirement of an autoregressive probability model should also be mentioned and discussed thoroughly. Agreed, will do! We want to emphasize here that we propose joint shuffle coding to mitigate exactly this restriction. Additionally, recent work on autoregressive models is promising (see Kong et al., 2023), and our results show that such models can lead to competitive rates even with few parameters. ## References Zhu, Yanqiao et al. (2022): A survey on deep graph generation: Methods and applications. Learning on Graphs Conference. PMLR. Maneth, Sebastian et al. (2015): A survey on methods and systems for graph compression. arXiv preprint. Townsend, James et al. (2019): Practical lossless compression with latent variables using bits back coding. ICLR. Kong, Lingkai, et al. (2023): Autoregressive diffusion model for graph generation. ICML. PMLR. --- Rebuttal Comment 1.1: Comment: Thank you for answering my questions and addressing the raised concerns. I am mostly satisfied with the answers and overall effort of the authors to improve their work. I would like to additionally commend the new experimental results the authors provide. However, I think that the results presented in the Figure S1 should be plotted in the log or log-log scale to better distinguish the converging lines. Additionally, in my opinion, the limitation regarding the complexity of obtaining conditional probabilities (required for encoding the slices in the recursive algorithm) for complex network models is still somewhat underdiscussed. In conclusion, I think that the paper is well-written, and the experimental evaluation is strong. Recognizing the results presented during the rebuttal, I would like to increase my score to "7: Accept". I still believe that the paper lacks some generalization, and some limitations can be discussed more explicitly, but these issues can be considered minor.	Rebuttal 1: Rebuttal: We are delighted that all reviewers agreed on the good soundness of the paper, with multiple pointing out its mathematical rigor, and most reviewers appreciate its presentation. Based on reviewers’ requests, we are excited to report striking new experimental results on our supplemental page, described here, and discussed further in the individual responses: - Table S1 updates Table 4 from the paper. We found that for incomplete recursive shuffle coding with the autoregressive Pólya urn (AP) graph model, to compress edge counts $k_i$ in slices, using a simple approximation based on a discretized Pareto distribution (specifically, $Q(k_i) \propto \frac{1}{2} k_i^{-2}$ for $k_i>1$ with $Q(k_i=0)=Q(k_i=1)=\frac{1}{3}$) leads to much-improved rates over the previously used log-uniform distribution. As a result, our method now comfortably outperforms SZIP in terms of one-shot compression rate for both the AP/WL$_1$ and AP/WL$_2$ configurations. This also improves all results in Table 5, which we will update in the final version. - We now repeat these AP experiments with varying seeds, and report means and empirical standard deviations. As explained further in the responses below, this highlights the stochastic nature specific to the AP model and supports the prediction that more WL iterations in practice lead to better or equal expected rates. Previous results for AP in Table 4 were misleading in these two respects. - In Table S2, we compare incomplete recursive with plain shuffle coding on all the 136 TU graph datasets (Morris et al., 2020) in six categories, ranging from molecules to social networks, with the majority of these datasets featuring vertex and edge attributes (see [chrsmrrs.github.io/datasets/docs/datasets](https://chrsmrrs.github.io/datasets/docs/datasets/) for details). Our method leads to a dramatic speedup over plain shuffle coding on some of these datasets, with a minimal increase in compression rate across all datasets, showing that these favorable properties extend to graphs with attributes. Maximum-likelihood parameters for a categorical distribution over vertex and edge attributes are inferred for and coded along with each dataset, as done and described in Kunze et al. 2024. - Figure S1 shows how the compression rate and speed varies with the automorphism group's approximation level, on three such graph datasets. This experiment further confirms that in practice, a low number of WL iterations is sufficient to get very close to the optimal rate. Crucially, the new results show that our method now achieves state-of-the-art rates in one-shot graph compression, as well as competitive rates and speeds on graphs with attributes, and explore the method’s rate-speed trade-off. We believe these results significantly strengthen the paper, particularly in the 'contribution' category. We hope that the reviewers agree, find our responses satisfactory, and that their concerns are being fully addressed. ### References Morris, Christopher et al. (2020): TUdataset: A collection of benchmark datasets for learning with graphs. Graph Learning and Beyond workshop, ICML. [chrsmrrs.github.io/datasets](https://chrsmrrs.github.io/datasets/) Kunze, Julius et al. (2024): Entropy Coding of Unordered Data Structures. ICLR. Pdf: /pdf/d888bba61fd6c37bf2dca99b0193b81784edc022.pdf	NeurIPS_2024_submissions_huggingface	2,024	null	null	null	null	null	null	null	null
Scanning Trojaned Models Using Out-of-Distribution Samples	Accept (poster)	Summary: The paper addresses the problem of detecting trojaned models. The paper proposes a trojaned model scanning method using out-of-distribution (OOD) samples. Specifically, it is observed that trojaned classifiers can erroneously identify adversarially attacked OOD samples as in-distribution (ID) samples. Therefore, the increased likelihood of perturbed OOD samples being classified as ID serves as a signature for trojan detection. The proposed trojan detection method can be applied in two scenarios: avaliability of clean samples and non-avaliability of clean ID samples. Extensive experiments demonstrate that proposed can achieve state-of-the-art performance compared with other scanning trojan detection methods. Strengths: The paper proposes a trojan detection method using out-of-distribution samples. This can be applied when no training in-distribution samples is available, which is effective in the real-word applications. The paper conducts sufficient experiments to demonstrate the effectiveness of proposed method. The trojan detection experiments are conducted compared with 7 state-of-the-art scanning methods on 5 datasets across various architectures including CNN and Transformer. The authors present both empirical results and theoretical analysis of proposed method. This is convincing. Weaknesses: The motivation (especially in Introduction Section) of proposed method is not presented clearly. For example, in line 73-74, the paper introduces the concept of near-OOD samples. However, the difference of far-OOD and near-OOD samples is not clearly explained. Also, the paper writes that "see the visual demonstration in Appendix Section 5". It is obvious there is no Section 5 in Appendix. The paper claims that current methods can not identify models trojaned using adversarial training (e.g. in line 6-7 and line 112-115). However, there is no detailed experiments to show the difference between detecting models trojaned using adversarial training and normal training. Some ablations can be done. For example, the number of used out-of-distribution samples, and the diversity of used OOD samples. Technical Quality: 3 Clarity: 2 Questions for Authors: Could you provide some details of used transformations? Could you compute FID between transformed samples and original samples? The paper only compares with scanning methods. How about other kinds of state-of-the-art trojan detection methods? Confidence: 4 Soundness: 3 Presentation: 2 Contribution: 3 Limitations: The limitations have been addressed. Flag For Ethics Review: ['No ethics review needed.'] Rating: 4 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thank you for the valuable comments. Responses to specific points are provided below: > W1: * We apologize for the referencing error. We intended to refer to Figure5 in SectionE. * We mentioned that near-OODs are those that share semanti/stylistic features with the IDs making them harder to distinguish (e.g., CIFAR10 vs. CIFAR100). On the other hand, far-OODs do not share any semantic or stylistic similarity with IDs (e.g., CIFAR10 vs. MNIST), making them easier to identify as OODs. For a more formal definition this Concepts, we kindly refer the reviewer to [1], as this is not the primary focus of our study; instead, we leverage these concepts for detecting trojaned classifiers. * The main motivation behind TRODO is to define and find a general and robust signature that distinguishes trojaned and clean classifiers, independent of architecture, label mapping strategy in trojaning, and the training strategy of the classifier (clean/adversarial). --- > W2: As stated in the caption of Table 1, all ACC* columns are related to adversarially trained models. Additionally, lines 287-290 mention, "Specifically, TRODO achieves superior performance with an 11.4% improvement in scenarios where trojan classifiers have been trained in a standard (non-adversarial) setting and a 24.8% improvement in scenarios where trojan classifiers have been adversarially trained," indicating a greater improvement on adversarially trained models. Moreover, Table 2 in our paper includes columns for rounds 3, 4, and 11 of TrojAI, which contain adversarially trained models. Table 1: Performance of TRODO compared with other methods, in terms of Accuracy on standard trained evaluation sets (ACC%) and adversarially trained ones (ACC%). \|LabelMapping\|Method\|MNIST\|CIFAR10\|GTSRB\|CIFAR100\|Pubfig\|Avg.\| \|-\|-\|-\|-\|-\|-\|-\|-\| \|\|\|ACC/ACC\\|ACC/ACC\\|ACC/ACC\\|ACC/ACC\\|ACC/ACC\\|*ACC/ACC\\| \|All-to-One\|NC\|54.3/49.8\|53.2/48.4\|62.8/56.3\|52.1/42.1\|52.5/40.2\|55.0/49.4\| \|\|ABS\|67.5/69.0\|64.1/65.6\|71.2/65.5\|56.4/54.2\|56.3/58.3\|63.1/62.5\| \|\|PT-RED\|51.0/48.8\|50.4/46.1\|58.4/57.5\|50.9/45.3\|49.1/47.9\|52.0/49.1\| \|\|TABOR\|60.5/45.0\|56.3/44.7\|69.0/53.8\|56.7/45.5\|58.6/44.2\|60.2/46.6\| \|\|K-ARM\|68.4/55.1\|66.7/54.8\|70.1/62.8\|59.8/50.9\|60.2/47.6\|65.0/54.2\| \|\|MNTD\|57.4/51.3\|56.9/52.3\|65.2/55.9\|54.4/48.8\|56.7/50.0\|58.1/54.7\| \|\|MM-BD\|85.2/65.4\|77.3/57.8\|79.6/65.2\|88.5/74.0\|65.7/48.3\|79.3/62.1\| \|\|UMD\|81.1/61.2\|77.5/54.7\|81.4/68.2\|69.0/56.3\|67.9/49.7\|75.4/58.0\| \|\|TRODO-Zero\|80.9/79.3\|82.7/78.5\|84.8/83.3\|75.5/73.7\|73.2/70.6\|79.4/77.0\| \|\|TRODO\|91.2/89.6\|91.0/88.4\|96.6/93.2\|86.7/82.5\|88.1/83.0\|90.7/87.3\| \|All-to-All\|NC\|26.7/21.6\|24.9/19.6\|31.6/23.2\|15.4/11.8\|16.8/12.3\|23.1/17.7\| \|\|ABS\|32.5/34.1\|30.7/28.8\|23.6/20.5\|34.3/34.8\|31.0/28.2\|30.4/29.3\| \|\|PT-RED\|41.0/33.5\|39.6/33.1\|45.4/43.9\|20.3/15.2\|12.6/9.8\|31.8/27.1\| \|\|TABOR\|51.7/39.7\|50.2/37.8\|48.3/39.5\|39.4/30.2\|38.6/30.8\|45.6/35.6\| \|\|K-ARM\|56.8/49.7\|54.6/47.6\|57.5/48.9\|51.3/45.0\|50.6/47.3\|54.2/47.7\| \|\|MNTD\|27.2/25.2\|23.0/18.6\|16.9/12.8\|29.8/31.0\|22.3/17.9\|23.8/21.1\| \|\|MM-BD\|54.3/40.4\|49.4/35.1\|57.9/44.0\|40.7/32.3\|41.2/34.1\|48.7/37.2\| \|\|UMD\|82.5/61.9\|74.6/60.1\|84.2/64.5\|70.6/49.9\|68.7/52.3\|76.1/57.7\| \|\|TRODO-Zero\|82.1/80.8\|80.4/77.3\|83.8/88.6\|74.8/72.3\|75.0/75.4\|79.2/78.8\| \|\|TRODO\|90.0/87.4\|89.3/87.5\|92.6/89.1\|82.4/85.0\|83.2/80.9\|87.5/86.1\| --- > W3: We aimed to provide similar ablations in Table 3 of the paper. However, to address your concerns, we conducted additional experiments. In the first experiment, we performed an ablation study on the number of samples in our validation set. By default, TRODO Zero uses the Tiny ImageNet validation dataset, which contains 200 classes, each with 500 samples. We explored the effect of varying the number of samples per class and investigated the ratio [0.0, 1.0]. \|LabelMapping\|Method\|OOD-Sample-Rate\|MNIST\|CIFAR10\|GTSRB\|CIFAR100\|Pubfig\|Avg.*\| \|-\|-\|-\|-\|-\|-\|-\|-\|-\| \|\|\|\|ACC/ACC\\|ACC/ACC\\|ACC/ACC\\|ACC/ACC\\|ACC/ACC\\|*ACC/ACC\\| \|All-to-One\|TRODO-Zero\|0.1%\|78.6/77.2\|80.5/76.3\|82.6/81.1\|73.2/71.4\|71.2/68.5\|77.3/75.0\| \|\|\|0.2%\|79.3/77.9\|81.2/77.1\|83.3/81.8\|73.9/72.1\|71.9/69.2\|78.1/75.7\| \|\|\|0.3%\|80.0/78.6\|81.9/77.8\|84.0/82.5\|74.6/72.8\|72.6/69.9\|78.8/76.4\| \|\|\|0.5%\|80.6/79.2\|82.5/78.4\|84.6/83.1\|75.2/73.4\|73.1/70.5\|79.3/77.0\| \|\|\|1% (default)\|80.9/79.3\|82.7/78.5\|84.8/83.3\|75.5/73.7\|73.2/70.6\|79.5/77.1\| Here, we utilized the DC metric [2] to measure the diversity based on your suggestion. A higher value indicates greater density and coverage of the validation set with respect to the evaluation set: \|Validation\|MNIST\|CIFAR10\|GTSRB\|CIFAR100\|PubFig\|DC Average\| \|-\|-\|-\|-\|-\|-\|-\| \|FMNIST\|0.45\|0.41\|0.52\|0.54\|0.47\|0.48\| \|SVHN\|0.50\|0.53\|0.60\|0.59\|0.57\|0.56\| \|STL-10\|0.55\|0.64\|0.63\|0.65\|0.62\|0.62\| \|TinyImageNet(default)\|0.60\|0.75\|0.70\|0.78\|0.80\|0.73\| --- > Q1: We apologize for the missing details about hard augmentations/transformations and kindly ask you to review our common response, where we have provided detailed answers to this question. --- > Q2: Here for each validation set, we crafted the OOD samples with hard transformations and then computed their distance using the FID metric: FMNIST\|SVHN\|STL-10\|TinyImageNet\|PubFig\| \|-\|-\|-\|-\|-\| \|43\|68\|59\|84\|75\| --- > Q3:** We apologize for any ambiguity/inconvenience caused by using different terms. However, we would like to clarify that our method is a post-training defense approach aimed at detecting whether an input model/classifier has been trojaned (backdoored). After reviewing the literature, we understand that researchers refer to this as both 'trojan detection methods' and 'scanning methods,' as the task involves scanning an input model. [1] Winkens Contrastive 2021 [2] Naeem Reliable 2020	Summary: The authors propose a trojan scanning technique that leverages the sensitivity of the network's confidence when near-OOD samples undergo an adversarial attack. The authors argue that the greater variation in confidence can be used to discriminate whether a network has been backdoored, and present extensive experiments to support the effectiveness of their method. Strengths: - The authors compare with a large number of baselines and a wide variety of datasets and trojan attacks, with standard and adversarial training on 3 architectures. - The concept is intuitive and the method seems to be empirically effective in a wide variety of cases. Weaknesses: - W0.1 (On writing) The concepts are very simple, but the presentation is hard to follow. Lines 40-89 refer to figures and concepts that will be displayed in following sections. I would suggest to anticipate significantly the figures. Furthermore, I would suggest providing immediately clear definitions of concepts, that are instead introduced in a handwavy way (e.g., blind spots, benign overfitting) before being properly discussed in related works. Probably completely restructuring Sections 1 and 2 to transfer most of the contents of 1 to 2 could help, having a good 'definitions' or 'preliminaries' section in Section 2. - W0.2 (On writing) The formatting of the theorems in the main paper (especially with lack of assumptions) is very unusual and may e confusing. - W1: The scanning procedure may be vulnerable to an adaptive trojan attack. Could the authors show what happens if attackers account for the author's defense in their attack strategy? How much does the effectiveness of their technique go down? The attackers may completely elude the proposed scanning method if they account for it in the design of their attack. - W2: How sensitive are the hyperparameters to the choice of the validation set? How computationally expensive is it to tune them? - W3: Results are adequately reported exclusively for Resnet18, and are not reported exhaustively for other architectures (or their presentation is not really immediately clear). Technical Quality: 2 Clarity: 1 Questions for Authors: - Q1: There must be a typo at line 161, I could not find information about hard transformations and therefore cannot verify claims made about them (Appendix Section 5). The typo "Attention Section" occurs repeatedly throughout the paper. . Q2: When using TinyImageNet, why did the authors not simply filter out the classes overlapping with CIFAR-10? This can be easily done and show the effectiveness of the method without applying G(.). Confidence: 4 Soundness: 2 Presentation: 1 Contribution: 3 Limitations: - L1: The method seems to assume to know the training distribution of the model so that it is possible to find near-OOD samples. However this may not always be possible or easy (near-ness could depend on many factors other than class similarity, e.g. may depend on resolution of the input, color range, other forms of naturally occurring covariate-shifts, types of shortcuts taken by the models etc.). The method significantly improves when this info is accessible. - L2: The previous limitation spills over to the need of having a good validation set. In some cases, the performance can go below the baselines if the validation set is not representative of the training set. Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thank you for your useful comments. Please find our responses below: >W0.1&W0.2: We appreciate your suggestions and will implement them to improve clarity and logical flow in the final manuscript based on your recommendations. ---- >W1: We conducted additional experiments to establish an adaptive Trojan attack. We have detailed these in our _common response_, which we kindly ask you to review. --- > W2: As stated in limitation section of our paper, for each new architecture, we have to firstly tune $\epsilon$ and then $\tau$. For each selection of architecture and validation dataset, first we have to train a surrogate classifier $g$ on the selected validation set and then find $\epsilon$ using DeepFool [1]. As the final stage for computing $\tau$ is not time consuming, we only report the computational cost of the first two stages of hyperparameters tuning in the tables below (the values are in terms of hours): \|\|ResNet18\|PreActResNet18\|ViT-B/16\| \|-\|-\|-\|-\| \|Training surrogate classifier $g$\|3\|5\|10\| \|DeepFool for $\epsilon$\|0.3\|0.7\|2\| * The experiments have been conducted on a RTX 3090. Regarding the sensitivity of TRODO to the validation set and hyperparameters, we have provided the values of hyperparameters for various selections of architecture and validation dataset: ResNet18 \|\|ϵ\|τ\| \|-\|-\|-\| \|FMNIST\|0.0491\|1.1625\| \|SVHN\|0.0476\|1.1338\| \|STL-10\|0.0488\|1.1571\| \|TinyImageNet\|0.0483\|1.1523\| PreActResNet18 \|\|ϵ\|τ\| \|-\|-\|-\| \|FMNIST\|0.0538\|1.0407\| \|SVHN\|0.0524\|1.0025\| \|STL-10\|0.0530\|1.0462\| \|TinyImageNet\|0.0527\|1.0179\| ViT-B/16 \|\|ϵ\|τ\| \|-\|-\|-\| \|FMNIST\|0.0621\|0.9341\| \|SVHN\|0.0598\|0.9106\| \|STL-10\|0.0611\|0.9246\| \|TinyImageNet\|0.0609\|0.9150\| Additionally, the performance results of our method using these hyperparameters can be found in Table 3 in the Ablation Study (Section 6) of the paper. This section demonstrates the robustness and effectiveness of our method across different datasets in the validation set. --- > W3: We apologize for the oversight and any confusion it may have caused. It seems there was an issue that resulted in the exclusion of results for PreActResNet18 and ViT-B/16 models from Appendix Section M. We will ensure that these results are included in the final updated version of the paper. Below are the detection results in terms of accuracy on standard trained evaluation sets (ACC %) and adversarially trained ones (ACC* %). PreAct ResNet-18 Architecture: \|LabelMapping\|Method\|MNIST\|CIFAR10\|GTSRB\|CIFAR100\|Pubfig\|*Avg.\| \|-\|-\|-\|-\|-\|-\|-\|-\| \|\|\|ACC/ACC\\|ACC/ACC\\|ACC/ACC\\|ACC/ACC\\|ACC/ACC\\|*ACC/ACC\\| \|All-to-One\|\|\|\|\|\|\|\|\| \|\|K-ARM\|69.9/54.3\|68.2/52.0\|72.0/62.8\|58.3/48.6\|61.3/48.6\|65.9/53.2\| \|\|MM-BD\|81.5/67.3\|75.6/57.1\|75.8/61.1\|86.1/72.7\|62.4/47.6\|76.3/61.2\| \|\|UMD\|79.5/58.8\|76.4/51.5\|79.2/64.7\|67.5/54.6\|64.5/47.0\|73.4/55.3\| \|\|TRODO-Zero\|85.0/78.5\|81.2/79.1\|85.2/83.9\|78.2/78.9\|73.6/72.4\|80.6/78.6\| \|\|TRODO\|92.6/88.7\|90.5/90.2\|93.4/90.1\|85.6/83.2\|80.2/78.8\|88.5/86.2\| \|All-to-All\|\|\|\|\|\|\|\|\| \|\|K-ARM\|56.8/49.7\|54.6/47.6\|57.5/48.9\|51.3/45.0\|50.6/47.3\|54.2/47.7\| \|\|MM-BD\|52.7/42.3\|49.3/35.1\|57.0/44.2\|41.3/32.0\|40.0/34.0\|48.1/37.5\| \|\|UMD\|80.7/61.5\|75.5/55.9\|83.5/64.9\|67.7/50.0\|66.7/47.5\|74.8/56.0\| \|\|TRODO-Zero\|83.9/76.8\|77.2/78.4\|87.5/83.7\|78.4/76.8\|76.0/70.3\|80.6/77.2\| \|\|TRODO\|90.8/87.9\|88.3/89.8\|92.0/87.9\|82.6/82.1\|81.0/76.6\|86.9/84.9\| ViT-B-16 Architecture: \|LabelMapping\|Method\|MNIST\|CIFAR10\|GTSRB\|CIFAR100\|Pubfig\|Avg.*\| \|-\|-\|-\|-\|-\|-\|-\|-\| \|\|\|ACC/ACC\\|ACC/ACC\\|ACC/ACC\\|ACC/ACC\\|ACC/ACC\\|*ACC/ACC\\| \|All-to-One\|\|\|\|\|\|\|\|\| \|\|K-ARM\|69.8/54.3\|68.2/52.0\|72.0/62.8\|58.3/48.6\|61.3/48.6\|65.9/53.2\| \|\|MM-BD\|72.9/58.4\|67.6/49.8\|67.5/52.5\|78.0/62.7\|54.4/39.5\|68.1/52.6\| \|\|UMD\|75.2/55.0\|69.0/45.4\|71.5/59.6\|62.5/48.6\|58.3/40.7\|67.3/49.8\| \|\|TRODO-Zero\|78.5/71.8\|72.9/71.5\|76.5/76.1\|71.6/70.2\|68.9/65.1\|73.7/71.0\| \|\|TRODO\|87.9/85.6\|82.5/84.4\|85.2/84.3\|80.3/79.6\|78.2/76.1\|82.8/82.0\| \|All-to-All\|\|\|\|\|\|\|\|\| \|\|K-ARM\|54.9/43.3\|50.5/43.1\|51.5/47.6\|46.4/39.9\|49.4/41.9\|50.5/43.2\| \|\|MM-BD\|51.9/40.7\|44.3/33.1\|57.8/41.6\|41.0/29.6\|40.4/28.0\|47.1/34.6\| \|\|UMD\|73.9/56.4\|69.4/48.6\|77.7/58.6\|58.8/43.4\|58.0/40.1\|67.5/49.4\| \|\|TRODO-Zero\|80.2/76.0\|70.4/73.3\|81.6/77.0\|74.2/69.8\|71.7/65.8\|75.6/72.4\| \|\|TRODO\|87.6/82.3\|82.6/84.3\|83.5/83.0\|79.8/77.3\|76.0/73.1\|81.9/80.0\| --- > Q1: We apologize for the missing details about hard augmentations/transformations and kindly ask you to review our common response, where we have provided detailed answers to this question. ----- >Q2: We appreciate your insightful question. Our primary objective is to develop a general method with a consistent pipeline regardless of the training data distribution. We aimed to avoid making dataset-specific modifications, such as filtering out overlapping classes for CIFAR-10 but not for others, such as GTSRB. By not removing common classes, we ensure that our approach remains uniform and applicable to various datasets without additional preprocessing steps. ----- >L1&L2:** Although our method can operate without access to any training data, we confirm that having access to such data from the input classifier can improve our detection performance. However, it should be noted that all existing detection methods, except for MM-BD, strongly rely on training data for operation. On the other hand, our method works in scenarios where training data is unavailable by just leveraging a small validation set. As shown in Table 3, by using only the FMNIST or SVHN dataset, we achieve superior performance, outperforming other methods by up to 10%. [1] Moosavi-Dezfooli et al. DeepFool 2015 --- Rebuttal Comment 1.1: Title: Response to rebuttal Comment: I have read the authors' rebuttals and the other reviewers' opinions. I think several of the concernes raised by other reviewers are fair, but the authors were convincing in their rebuttals. I appreciate the clarification, additional information and experiments performed by the reviewers, especially the introduction of an analysis of their method under adaptive attacks. While this may not be a final guarantee about the robustness of the proposed technique and the limitations pointed out are quite strong, I think this is reasonable and what an attacker would try out first, and the relative robustness of the technique to it indicates a significant research effort would be required in order to further degrade the performance of this technique. Therefore, I am happy to increase my score. --- Rebuttal 2: Title: Appreciation for Your Positive Feedback Comment: Thank you for your thoughtful feedback and for taking the time to review our rebuttal! We greatly appreciate your careful evaluation and positive response. Sincerely, The Authors	Summary: This paper proposes a general strategy for distinguishing between trojaned and clean models. The generality of the approach lies in its applicability to various types of trojan attacks, different label-mapping strategies, and its ability to work with and without clean training data. The authors claim that distorted areas of the learned decision boundary in trojaned models, referred to as blind spots, serve as a consistent signature for distinguishing between trojaned and clean classifiers, regardless of the trojan attack methodology. A key characteristic of these blind spots is that samples within these regions are expected to be out-of-distribution (OOD) relative to the clean training data, yet trojaned classifiers mistakenly identify them as in-distribution (ID) samples. To leverage this characteristic, the authors propose using adversarial attacks to perturb a given OOD sample towards an ID direction, followed by computing the difference in ID score (i.e., maximum softmax probability) before and after the adversarial attack. They found that trojaned models exhibit significantly larger differences in these scores. The effectiveness of the proposed detection method was empirically validated against eight trojan attacks and under different levels of access to clean training data. Strengths: 1. The paper is well-written and easy to read. 2. It includes clear explanations, supported by figures, to illustrate the intuition and proposed method. 3. The authors conduct extensive experiments across various types of trojan attacks and different levels of access to clean training data. 4. They propose a simple yet effective method to detect trojaned models. 5. Theoretical analysis is provided, demonstrating that a trojaned neural network is more sensitive to adversarial perturbations. Weaknesses: As mentioned in the limitation section, the selections of $\epsilon$ and $\tau$ rely on a validation set and a surrogate model. As a result, the quality of these hyperparameters depend on the choices of the validation set and the surrogate model. In particular, when detecting a trojaned model with a new model architecture or trained on a new domain, one might need to tune $\epsilon$ and $\tau$. Technical Quality: 4 Clarity: 4 Questions for Authors: 1. Could you clarify what it means for a trojaned classifier to be adversarially trained on poisoned training data? (For example, for poisoning samples generated by a label-consistent trojan attack, what does it mean by adversarially training the trojaned model on these data?) Additionally, are there any empirical results demonstrating the effectiveness of the proposed method in this specific scenario? 2. When a portion of clean training data is not accessible, the authors propose to use Tiny-ImageNet. In addition, it comes to my mind that given the accessibility to the model, one could actually reverse engineer the training data. In that way, you could obtain a portion of fake clean training data, followed by creating near-OOD samples.Consequently, the performance without clean training data could approach that of having a portion of real clean training data, thereby diminishing the performance gap between TRODO-Zero and TRODO. In particular, the reverse engineering of training data has been adopted in many backdoor works, such as [1] (see Figure 3). 3. Headings of paragraphs could be set in a uniform format. Eg., In Line 149, there is a ‘.’ after the paragraph heading; by comparison, in Line 168, there is not. [1] Liu, Y., Ma, S., Aafer, Y., Lee, W. C., Zhai, J., Wang, W., & Zhang, X. (2017). Trojaning attack on neural networks. Confidence: 4 Soundness: 4 Presentation: 4 Contribution: 4 Limitations: The performance of the proposed method drops by a large margin (over 10%) when no clean training data is accessible, compared to that with a portion of clean training data. Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thank you very much for your positive feedback on our paper. We greatly appreciate your insights and suggestions. > W1: We acknowledge that extracting hyperparameters may be considered a limitation of our method, as previously discussed in our limitations section. However, it is important to note that these hyperparameters must be selected based on the architecture where the validation is fixed. Moreover, our method demonstrates robustness with respect to the validation set, as evidenced in Table 3 of the Ablation Study. Specifically, using small datasets like FMNIST or SVHN for validation, our method outperforms others by up to 10%. --- > Q1: We sincerely apologize for any confusion caused. To clarify, the training dataset is initially compromised using a backdoor attack. The classifier is then adversarially trained using the PGD (Projected Gradient Descent) attack method. As mentioned in the caption of Table 1, all columns labeled ACC* pertain to adversarially trained models. Additionally, lines 287-290 states: "Specifically, TRODO achieves superior performance with an 11.4% improvement in scenarios where trojan classifiers have been trained in a standard (non-adversarial) setting and a 24.8% improvement in scenarios where trojan classifiers have been adversarially trained," indicating a significant improvement in adversarially trained models. Furthermore, Table 2 includes columns for rounds 3, 4, and 11 of TrojAI, which involve adversarially trained models. Table 1: Scanning performance of TRODO compared with other methods, in terms of Accuracy on standard trained evaluation sets (ACC %) and adversarially trained ones (ACC* %). \|Label Mapping\|Method\|MNIST\|CIFAR10\|GTSRB\|CIFAR100\|Pubfig\|*Avg.\| \|-\|-\|-\|-\|-\|-\|-\|-\| \|\|\|ACC/ACC\|ACC/ACC\|ACC/ACC\|ACC/ACC\|ACC/ACC\|ACC/ACC\| \|All-to-One\|\|\|\|\|\|\|\|\| \|\|NC\|54.3/49.8\|53.2/48.4\|62.8/56.3\|52.1/42.1\|52.5/40.2\|55.0/49.4\| \|\|ABS\|67.5/69.0\|64.1/65.6\|71.2/65.5\|56.4/54.2\|56.3/58.3\|63.1/62.5\| \|\|PT-RED\|51.0/48.8\|50.4/46.1\|58.4/57.5\|50.9/45.3\|49.1/47.9\|52.0/49.1\| \|\|TABOR\|60.5/45.0\|56.3/44.7\|69.0/53.8\|56.7/45.5\|58.6/44.2\|60.2/46.6\| \|\|K-ARM\|68.4/55.1\|66.7/54.8\|70.1/62.8\|59.8/50.9\|60.2/47.6\|65.0/54.2\| \|\|MNTD\|57.4/51.3\|56.9/52.3\|65.2/55.9\|54.4/48.8\|56.7/50.0\|58.1/54.7\| \|\|MM-BD\|85.2/65.4\|77.3/57.8\|79.6/65.2\|88.5/74.0\|65.7/48.3\|79.3/62.1\| \|\|UMD\|81.1/61.2\|77.5/54.7\|81.4/68.2\|69.0/56.3\|67.9/49.7\|75.4/58.0\| \|\|TRODO-Zero\|80.9/79.3\|82.7/78.5\|84.8/83.3\|75.5/73.7\|73.2/70.6\|79.4/77.0\| \|\|TRODO\|91.2/89.6\|91.0/88.4\|96.6/93.2\|86.7/82.5\|88.1/83.0\|90.7/87.3\| \|All-to-All\|\|\|\|\|\|\|\|\| \|\|NC\|26.7/21.6\|24.9/19.6\|31.6/23.2\|15.4/11.8\|16.8/12.3\|23.1/17.7\| \|\|ABS\|32.5/34.1\|30.7/28.8\|23.6/20.5\|34.3/34.8\|31.0/28.2\|30.4/29.3\| \|\|PT-RED\|41.0/33.5\|39.6/33.1\|45.4/43.9\|20.3/15.2\|12.6/9.8\|31.8/27.1\| \|\|TABOR\|51.7/39.7\|50.2/37.8\|48.3/39.5\|39.4/30.2\|38.6/30.8\|45.6/35.6\| \|\|K-ARM\|56.8/49.7\|54.6/47.6\|57.5/48.9\|51.3/45.0\|50.6/47.3\|54.2/47.7\| \|\|MNTD\|27.2/25.2\|23.0/18.6\|16.9/12.8\|29.8/31.0\|22.3/17.9\|23.8/21.1\| \|\|MM-BD\|54.3/40.4\|49.4/35.1\|57.9/44.0\|40.7/32.3\|41.2/34.1\|48.7/37.2\| \|\|UMD\|82.5/61.9\|74.6/60.1\|84.2/64.5\|70.6/49.9\|68.7/52.3\|76.1/57.7\| \|\|TRODO-Zero\|82.1/80.8\|80.4/77.3\|83.8/88.6\|74.8/72.3\|75.0/75.4\|79.2/78.8\| \|\|TRODO\|90.0/87.4\|89.3/87.5\|92.6/89.1\|82.4/85.0\|83.2/80.9\|87.5/86.1\| Table 2: Comparison of TRODO and other methods on all released rounds of TrojAI benchmark on image classification task. For each method, we reported scanning Accuracy and the average time scanning time for a given classifier. \|Method\|Round0 Accuracy/Time(s)\|Round1 Accuracy/Time(s)\|Round2 Accuracy/Time(s)\|Round3 Accuracy/Time(s)\|Round4 Accuracy/Time(s)\|Round11 Accuracy/Time(s)\| \|-\|-\|-\|-\|-\|-\|-\| \|NC\|75.1/574.1\|72.2/592.6\|-/>23000\|-/>23000\|-/>20000\|N/A/N/A\| \|ABS\|70.3/481.9\|66.8/492.5\|62.0/1378.4\|70.8/1271.4\|76.3/443.2\|N/A/N/A\| \|PT-RED\|85.0/941.6\|84.3/962.7\|58.2/>23000\|65.7/>25000\|66.1/>28000\|N/A/N/A\| \|TABOR\|82.8/974.2\|80.3/992.5\|56.2/>29000\|60.8/>27000\|58.3/>32000\|N/A/N/A\| \|K-ARM\|91.3/262.1\|90.0/283.7\|_76.0_/1742.8\|79.0/1634.1\|_82.0_/1581.4\|N/A/N/A\| \|MM-BD\|68.8/_226.4_\|73.2/_231.3_\|55.8/_174.3_\|52.6/_182.6_\|54.1/_178.1_\|_51.3_/_1214.2_\| \|UMD\|80.4/>34000\|79.2/>34000\|75.2/>18000\|61.3/>19000\|56.9/>90000\|N/A/N/A\| \|TRODO\|_86.2_/152.4\|_85.7_/194.3\|78.1/107.2\|_77.2_/122.4\|82.8/117.8\|61.3/984.3\| ----- > Q2:* We thank the reviewer for the suggestion. We should note that access to a portion of the training data would enable us to create near-OOD samples with higher quality compared to similar methods. Specifically, the mentioned method's strategy for exploiting data from a classifier leads to samples with artifacts and shortcuts, reducing its effectiveness as it becomes more apparent to the model that they do not belong to ID (considering them far OOD). However, to further explore your suggestion, we evaluated both the proposed method and the FastDFKD (Fang Data-free 2022) method for crafting data in situations where there is no access to training data: \|LabelMapping\|Method\|MNIST\|CIFAR10\|GTSRB\|CIFAR100\|Pubfig\|*Avg.\| \|-\|-\|-\|-\|-\|-\|-\|-\| \|\|\|\|ACC/ACC\\|ACC/ACC\\|ACC/ACC\\|ACC/ACC\\|ACC/ACC\\| \|All-to-One\|TrojaNN\|70.2/68.0\|72.3/70.1\|73.1/71.3\|65.5/63.1\|63.9/60.8\|69.0/66.7\| \|\|FastDFKD\|77.4/75.0\|79.2/75.5\|80.4/78.0\|72.9/70.1\|71.2/68.0\|76.2/73.3\| \|\|TRODO\|91.2/89.6\|91.0/88.4\|96.6/93.2\|86.7/82.5\|88.1/83.0\|90.7/87.3\| --- > Q3: Thank you for pointing out the formatting inconsistency. We apologize for this oversight and will ensure that all paragraph headings are uniformly formatted in the final manuscript. --- > L1:** We acknowledge that the limitation you mentioned is valid. However, the scenario of having no clean training data is a special case. Most research studies, including ours, typically utilize at least a small amount of clean data to validate their methods. --- Rebuttal Comment 1.1: Comment: W1: Thanks for the clarification. Q1: I still have some confusions about adversarially trained models. "To clarify, the training dataset is initially compromised using a backdoor attack. The classifier is then adversarially trained using the PGD (Projected Gradient Descent) attack method."—Could you clarify: 1) the data used for this adversarial training, 2) along with the objective function, and 3) the motivation for this adversarial training? Thanks. Q2: 1) So according to your claim, can we say that the samples generated by my mentioned's method are less effective than some randomly selected samples from a OOD dataset, e.g. Tiny ImageNet? 2) Is FastDFKD (Fang Data-free 2022) method using models to recover some training data? 2) I think my concern for using models to recover some training data when there is no clean training data arises from the large accuracy gap between TRODO and TRODO-Zero. (My intuition is that the samples generated in this way are more effective than some randomly selected samples from a OOD dataset) Thus, the comparison should be with TRODO-Zero, instead of TRODO. --- Reply to Comment 1.1.1: Comment: Thanks for your feedback. Here are the answers to your questions: Q1: 1) Each column in the table 1, above, shows the original training data. The data is first altered by the backdoor attack by adding the backdoor to the input and changing the corresponding output labels based on all-to-one (upper half) or all-to-all (lower half) . Then, adversarial training is applied on the entire dataset including the poisoned and clean samples. Here, each training data, which might be poisoned, undergoes an adversarial attack during the training. As in the regular adversarial training, we only perturb the input and keep the ground truth label unchanged. 2) The objective function in the adversarial training is the regular cross-entropy (line 593). 3) The main motivation for evaluating scanning methods on such models is that many previously proposed signatures for trojan scanning experience a detection accuracy drop in the case of an adversarially trained model (lines 37-39). For instance, in the table 1 above, UMD's detection acc. drops by almost 20% points on MNIST and CIFAR-10 on the adversarially trained models compared to the standard models (compare ACC with ACC). The possible reason behind this effect is that adversarial training makes shortcut learning difficult, potentially leading to harder trigger reverse-engineering and identification. While previous scanning methods struggle with this issue, our proposed method does not rely on such practices and shows insignificant drop under adversarially trained models. Q2: 2, 3) Please note that FastDFKD (Up to 100× Faster Data-free Knowledge Distillation by Fang et al.) is an efficient method of generating synthetic samples given access to a trained model (see Fig. 3 of that paper). Here, we use this model to create surrogates of the training samples and use them in our method TRODO-Zero to enhance it. So indeed the row mentioning "FastDFKD" is* the TRODO-Zero with the synthesized samples using FastDFKD. As can be seen, there is still a large gap between the two TRODO and TRODO-Zero (it even got slightly worse). The first row, TrojaNN, (Trojan Signatures in DNN Weights, by Fields et al.), which also assumes no access to the training data is included as a baseline. 1) Like we mentioned, we hypothesize that the generated samples, through methods such as FastDFKD, could contain shortcuts, potentially imperceptible, making them unideal OOD instances, compared to the clean tiny-ImageNet samples that lack such artifacts. Please note that such artifacts would help the classifier to classify the intended synthetic OOD sample as ID more easily, reducing the effectiveness of the method, because we take the gap between the ID score of the OOD sample before and after the attack as our detection score (line 179). Therefore, while your idea could generally be beneficial in other methods, it would be less so in our specific method that relies on measuring ID-ness of synthetic OOD samples before and after the attack.	Summary: The paper proposes TRODO, a method for identifying trojaned classifiers, which relies on the intuition that in presence of a backdoor it should be easier than for clean classifiers to make the model classify an out-of-distribution input as in-distribution by adding an adversarial perturbation. In practice, a PGD attack is run on OOD images to increase the confidence of the classifier predictions, and the average increase is used to detect backdoored models. In the experimental evaluation, the proposed method outperforms the baselines, being effective against several backdoor attacks even when combined with adversarial training. Strengths: - The proposed method is very general, as it doesn't make assumptions on the type of backdoor attacks or target architectures. Moreover, the paper proposes variants with and without access to training data. Finally, the proposed approach appears less expensive than the baselines (Table 2). - The experimental results support the proposed method, which is effective against several backdoor attacks even when combined with adversarial training. Weaknesses: - The proposed method seems to have brittle elements: - the PGD-10 attack used for computing the signature might be considered weak (only 10 iterations), then not able to fully optimize the target loss. Then, a stronger attack could further increase the confidence of the clean classifiers on OOD points, making them more similar to trojaned one (since the confidence is upper bounded, one could get the same score for all classifiers in the worst-case). Thus, the effectiveness of TRODO seems to rely to some extent on the attack not fully optimizing the target loss (which might be even exploited to bypass the detection mechanism). - the confidence of any classifier might be adjusted post-training by e.g. temperature rescaling without changing its decisions. In this way, it seems possible to bypass the detection by making a model under-confident. - adversarial training variants, e.g. on OOD data [60] or to have uniform predictions far from ID data [A], have been explored in prior works, and might be used to counter the proposed scanning scheme. - It's not clear how the effectiveness of TRODO correlates with the strength of the backdoor attacks. For example, one can imagine that using a lower poisoning rate might make the attack less detectable (but less effective). [A] https://arxiv.org/abs/1910.06259 Technical Quality: 3 Clarity: 3 Questions for Authors: - Why using a left-truncated normal distribution for estimating $\tau$ when the score S can take values only on a specific range (confidence is upper bounded by 1)? - While the proposed method provides good experimental results, I think it's important to address the concerns about its robustness (see above). Confidence: 3 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: Limitations are partially addressed. Flag For Ethics Review: ['No ethics review needed.'] Rating: 3 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thank you for your thoughtful review. We have provided the following response: >W1.1: We believe the epsilon value of the attack plays a key role compare to steps in our setup. If epsilon is large, as you mentioned, our signature for both clean and Trojan classifiers would be the same. This is the main reason we carefully estimate it using the validation set and Boundary Confidence Level to avoid such scenarios (see line 199). The number of steps in the attack has less effect as the attack radius has already been determined (Madry et al. Toward 2017). To further address your concern, we have provided an additional experiment where every component of our pipeline remains fixed except for the number of steps of the attack: \|Label Mapping\|Method\|n-step\|MNIST\|CIFAR10\|GTSRB\|CIFAR100\|Pubfig\|*Avg.\| \|-\|-\|-\|-\|-\|-\|-\|-\|-\| \|\|\|\|ACC/ACC\|ACC/ACC\|ACC/ACC\|ACC/ACC\|ACC/ACC\|*ACC/ACC\| \|All-to-One\|\|\|\|\|\|\|\|\| \|\|TRODO\|PGD-10 (default)\|91.2/89.6\|91.0/88.4\|96.6/93.2\|86.7/82.5\|88.1/83.0\|90.7/87.3\| \|\|\|PGD-100\|91.0/89.0\|90.2/87.6\|95.9/92.7\|86.0/81.9\|87.6/82.1\|89.9/86.6\| \|\|\|PGD-1000\|90.5/89.1\|90.5/87.6\|96.3/92.7\|86.1/81.7\|87.9/82.4\|90.5/86.4\| --- > W1.2: It is possible that temperature rescaling affects the regular softmax function, but the classifier remains more confident in in-distribution samples compared to OOD samples (Tajwar et al., No True 2021). This is consistent with the rationale and principle behind our method. As a result TRODO, would also detect trojaned classifiers with temperature scaling. To further clarify this concern, we kindly ask you to refer to our _common response_, where we considered worst-case scenarios involving adaptive attackers. Even when the attacker is aware of our defense mechanism, TOROD demonstrated robust and consistent performance. --- > W1.3: The mentioned papers, CCAT (Stutz et al., 2020) [A] and RATIO (Augustin et al., 2020) [60], aimed to enhance robust performance on in-distribution samples. Improving robustness on adversarially out-of-distribution (OOD) samples was not their main purpose. Although they considered the issue of overconfidence on OOD samples, their primary goal was robust in-distribution classification. They do not weaken principle behind TRODO, which is based on the vulnerability of existing robust methods, including CCAT, RATIO, and other SOTA methods, to perturbed OOD samples, especially when they are close to the in-distribution, referred to as near OOD in our paper (see line 88 and Table 5). This observation has also been demonstrated by Fort (Adversarial, 2022). Moreover, recently, RODEO (ICML, 2024) has shown that in the CIFAR10 vs. CIFAR100 OOD detection challenge, there is no better performance than random detection (i.e. less than 50% AUROC), which further supports our claim (see their Table 4-a of our paper). Additionally, we kindly ask you to check our _common response_, where we show that even in the worst-case scenario (adaptive attack), TRODO demonstrates promising performance. _Finally, to further address your concern, we trained input trojan models using the method proposed by RATIO instead of common adversarial training and reported TRODO's performance on them here. Other components of TRODO, including the training pipeline, remain fixed._ \|Label Mapping\|Method\|MNIST\|CIFAR10\|GTSRB\|CIFAR100\|Pubfig\|Avg.*\| \|-\|-\|-\|-\|-\|-\|-\|-\| \|\|\|ACC/ACC/ACC*\|ACC/ACC/ACC*\|ACC/ACC/ACC*\|ACC/ACC/ACC*\|ACC/ACC/ACC\|ACC/ACC/ACC\\*\| \|All-to-One\|\|\|\|\|\|\|\|\| \|\|TRODO-Zero\|80.9/79.3/78.4\|82.7/78.5/79.1\|84.8/83.3/82.8\|75.5/73.7/74.0\|73.2/70.6/70.2\|79.4/77.1/76.9\| \|\|TRODO\|91.2/89.6/89.4\|91.0/88.4/88.7\|96.6/93.2/92.4\|86.7/82.5/81.5\|88.1/83.0/83.4\|90.7/87.3/87.1\| --- >W2: Intuitively, increasing the poisoning rate enlarges the blind spots in trojaned classifiers, as these are boundary regions where the poisoned data causes the model to overfit. Consequently, this will increase the probability that TRODO detects the trojaned classifiers. However, our signature is based on the presence of blind spots in trojaned classifiers and shows consistent performance across different poisoning rates. In the paper, we considered a poisoning rate of 10% as it is common in the literature. In response to your comment, we have provided TRODO's performance for different rates below: _(Other components of TRODO remained fixed.)_ \|Label Mapping\|Method\|Poisoning-Rate\|MNIST\|CIFAR10\|GTSRB\|CIFAR100\|Pubfig\|Avg.*\| \|-\|-\|-\|-\|-\|-\|-\|-\|-\| \|\|\|\|ACC/ACC\|ACC/ACC\|ACC/ACC\|ACC/ACC\|ACC/ACC\|*ACC/ACC\\| \|All-to-One\|\|\|\|\|\|\|\|\| \|\|TRODO-Zero\|1%\|80.1/78.2\|81.3/77.0\|83.5/81.8\|74.6/72.7\|72.1/69.8\|78.3/75.9\| \|\|\|3%\|82.0/79.3\|82.5/78.1\|85.4/83.2\|76.7/74.4\|74.0/71.1\|80.1/77.2\| \|\|\|5%\|83.6/80.4\|84.0/79.5\|86.3/84.6\|77.5/75.3\|75.1/72.6\|81.3/78.5\| \|\|\|10% (default)\|80.9/79.3\|82.7/78.5\|84.8/83.3\|75.5/73.7\|73.2/70.6\|79.4/77.0\| \|\|TRODO\|1%\|89.5/87.4\|88.7/85.6\|94.9/91.0\|84.5/81.2\|86.4/80.3\|88.8/85.1\| \|\|\|3%\|91.0/89.2\|90.5/87.8\|96.5/93.1\|86.3/82.7\|87.8/82.4\|90.4/87.0\| \|\|\|5%\|92.8/90.6\|91.7/88.9\|97.0/94.1\|87.5/83.6\|89.1/84.5\|91.6/88.3\| \|\|\|10% (default)\|91.2/89.6\|91.0/88.4\|96.6/93.2\|86.7/82.5\|88.1/83.0\|90.7/87.3\| ---- > Q1: Thank you for your insightful question. You are correct that the score $S_i$ can take values only within a specific range, as confidence is upper bounded by $1$. We apologize for the oversight in the paper where we did not mention that we apply a transformation to the score $S_i$. In our method, we use $-log(1 - S_i)$ instead of $S_i$ directly. This transformation maps the original score to a new range $[0, ∞)$, making it suitable for fitting with a left-truncated normal distribution. We will include this clarification in the final manuscript to ensure there is no confusion. --- Rebuttal Comment 1.1: Comment: Thanks for the detailed response and additional experiments. > W1.2:** "the classifier remains more confident in in-distribution samples compared to OOD samples" If I understand it correctly, only OOD samples are used to compute the signature. Then, I think the absolute difference $\textrm{ID-Score}_f (x_i^{OOD}) − \textrm{ID-Score}_f (x_i^{OOD})$ could be (most likely) made arbitrarily small with temperature rescaling, so that it's in the range of the signature of a clean classifier. Similarly, since the confidence is upper bound by 1, also an over-confident classifier which has near 1 confidence on OOD data would lead to very small differences between adversarially perturbed and clean OOD points (since the ID-Score cannot increase much). Am I missing something? > W1.3:* "Improving robustness on adversarially out-of-distribution (OOD) samples was not their main purpose" RATIO loss includes a term which is a robust loss on OOD data, then it directly optimizes adversarial robustness on OOD points. What was the OOD data used for training the RATIO models? In general, these methods show that it is possible to control the (adversarial) confidence on OOD data (but might require some adaptation to the specific task of bypassing the detection mechanism). > common response I think the additional experiments with the new losses show that adaptive attacks have the potential to bypass the detection mechanism of TRODO: in fact, $L_{adaptive1}$ improves 12% compared to $L_{default}$ in the All-to-All setup, which seems significant. --- Rebuttal 2: Title: Response to Reviewer PnTa Comment: > Am I missing something? * We believe there has been some misunderstanding. When we mentioned that _'the classifier remains more confident in in-distribution samples compared to OOD samples,'_ we were referring to Fact-A. We would like to clarify that the principle behind our signature remains valid even when temperature rescaling is applied. The logical flow of our principle is as follows: Fact-A ⇒ Fact-B ⇒ Fact-C ⇒ Fact-D ⇒ Fact-E ⇒ Fact-F ⇒ Fact-G. * Furthermore, as indicated in Fact-A, the softmax output tends to resemble a uniform distribution for OOD samples. As a result, applying temperature rescaling will affect all logits equally and will not alter this uniform shape. Consequently the ID score for clean OOD samples would still be very low (e.g., approximately $\frac{1}{N}$, where $N$ is the number of ID classes). However, by perturbing these samples, the ID score can increase (up to one). Therefore, our signature would not be arbitrarily small as mentioned; it remains within the range of ([0,$1-\frac{1}{N}$]). Fact-A: A classifier remains more confident in IDs compared to OODs. Specifically, the output softmax of a classifier for OODs tends to be more uniform, while for IDs, it is more concentrated. (confidence: maximum softmax probablity) [1]. Fact-B: Previous research has primarily focused on perturbing IDs, shifting them, and using this as a signature to distinguish between clean and trojaned classifiers [2,3,4,5,6,7,8,9,10]. Fact-C: In scenarios where classifiers have been adversarially trained, perturbing ID samples for signature extraction becomes less effective. This is because adversarial training enhances the classifier's robustness to such perturbations, reducing their impact [9,10]. Fact-D: Instead of perturbing IDs, perturbing OODs toward the in-distribution is a viable approach, as classifiers are vulnerable to this shift. This approach remains effective even for adversarially trained classifiers, as they are still susceptible to perturbed OODs. This finding is also supported by parallel research in OOD detection [11,12,13,14]. Fact-E: Trojaned classifiers learn decision boundaries that include 'blind spots,' which are intuitively regions along the in-distribution boundary that have been altered to overfit on poisoned training data (data with triggers), thereby changing the geometry of the boundary. Fact-F: Perturbing OODs toward the classifier's decision boundary increases their ID scores, and this effect is particularly significant in trojaned classifiers. This is because the perturbations can exploit the blind spots within the decision boundary, mimicking the triggers used during the trojan attack. Fact-G: We use the difference in ID scores between a benign OOD sample (without perturbation) and an adversarially perturbed OOD sample as a signature to differentiate between clean and trojaned classifiers. --- >What was the OOD...?* Following their reported setting, we used Tiny ImageNet as the out-of-distribution dataset for training. --- >In general, these methods.. . Regardless of the adaptation strategy, existing classifiers are vulnerable to adversarial attacks on OODs, particularly when these samples are close to the in-distribution boundary. This vulnerability, as demonstrated recently by (Lorenz Deciphering, 2024), (Fort Adversarial Vulnerability, 2022), and others [11,12,13,14], limits them to act as counters to our method. To further illustrate this, we conducted additional experiments using RATIO as an OOD detector, where the MSP is employed as the ID score. In the first scenario, CIFAR-10 serves as the ID dataset and CIFAR-100 as the OOD, and vice versa in the second. While the method achieves good performance in clean scenarios , in adversarial scenarios—where perturbations are added to shift OODs to in-distribution and vice versa—its performance drops to below random detection, highlighting its vulnerability to perturbed OODs. \| Method\Benchmark \| _CIFAR-10 vs.CIFAR-100_ \|_CIFAR-100 vs. CIFAR-10_\| \|-\|-\|-\| \| \| AUC/AUC* \|AUC/AUC* \| \| RATIO \| 81.3/14.8 \|68.5/9.0 \| --- >I think the additional ... In the adaptive attack scenario, we evaluated various label mappings and adaptive strategies. The worst performance decrease we observed was 12% (from 86.0% to 74.1%). Despite this, our method still outperforms previous detection methods that were not subjected to adaptive attacks. For instance, UMD achieves 57.7%, while our approach achieves 74.1%. [1] Hendrycks, A baseline, 2016 [2] Xiang, UMD: Unsupervised , 2023 [3] Wang, Mm-bd: , 2024 [4] Wang, Neural, 2019 [5] Liu, Abs:Scanning, 2019 [6] Shen, Backdoor scanning for, 2021 [7] Guo, Tabor, 2019 [8] Hu, Trigger hunting, 2022 [9] Edraki, Odyssey, 2021 [10]Zhang, Cassandra, 2021 [11] Chen Robust OOD 2020 [12] Azizmalayeri, Your Detector 2022 [13] Chen, Atom ,2021 [14] Mirzaei, RODEO, 2024 --- Rebuttal Comment 2.1: Comment: > Furthermore, as indicated in Fact-A, the softmax output tends to resemble a uniform distribution for OOD samples. As a result, applying temperature rescaling will affect all logits equally and will not alter this uniform shape. Unless the softmax output is exactly uniform (i.e. all logits are identical), which seems very unlikely, temperature will change the distribution. > However, by perturbing these samples, the ID score can increase (up to one). Therefore, our signature would not be arbitrarily small as mentioned; it remains within the range of ($[0, 1 - 1/N]$). I meant arbitrarily small within its range of course (as the ID-Score after the attack cannot be lower than on the clean input). The point of using temperature >> 1, i.e. making the model under-confident, is that even when the attack is applied it will not increase much the confidence, in particularly since the hyperparameters of the attack such as $\epsilon$ and number of steps is fixed and calibrated on standard models. Also, one can consider the other extreme case with temperature = 0 (in practice, values close to 0 are sufficient), i.e. the softmax becomes the argmax function. If the argmax of the logits is unique, then the softmax output is a one-hot vector, and even for OOD points the difference of ID-Scores will be 0. This seems to bypass the detection mechanism. Is this correct? --- Rebuttal 3: Comment: We thank the reviewer for their thoughtful review and valuable feedback. * A key requirement for deploying deep neural networks in real-world applications is calibrated or approximately calibrated behavior on input samples. A miscalibrated classifier, particularly in cases where $T→0$, $T>>1$, can lead to poor decision-making. For example, in medical diagnosis, overestimating disease probability can result in unnecessary treatments, while underestimating it may lead to missed diagnoses. Miscalibrated models are especially unreliable in high-stakes scenarios such as finance, healthcare, or autonomous systems, where accurate probability estimates are crucial. Furthermore, the backdoor attack/defense literature generally assumes that Trojaned classifiers behave normally on samples without triggers, similarly to standard classifiers. Since standard classifiers typically exhibit approximately calibrated outputs, we implicitly assume that the Trojaned classifier would also retain this characteristic. * We acknowledge the challenges posed by the extreme temperature scenarios, specifically when $T→0$ or $T>>1$, as our ID-score would converge to zero in both cases. However, we believe this issue can be addressed with minimal modifications. In the backdoor attack setup (as detailed in our threat model), it is commonly assumed that the attacker has access to the input model. To mitigate the issue, we propose a minor extension to TRODO: applying its own softmax function to the logits of the input classifier instead of relying on the input classifier's final softmax output. This approach could help counter the adversarial temperature settings highlighted by the reviewer. * Moreover, since TRODO and TRODO-Zero both utilize an OOD set, another minor extension could involve evaluating the classifier's softmax output on these samples. If the output significantly deviates from a uniform distribution (e.g. measured by Kullback–Leibler divergence), the model could be rejected as extremely miscalibrated due to the attacker's manipulation. To further address your concern, we evaluated TRODO's performance on an all-to-one label mapping task using the CIFAR-10 and GTSRB datasets across different temperature values, while keeping other components unchanged. The results show that our designed detector effectively handles a reasonable range of temperature values, maintaining consistent performance throughout. \| CIFAR10 \| $T=0.5$ \| $T=0.7$ \| $T=1$(default) \| $T=1.2 $ \|$T=1.5$ \| $T=2$ \| \|-\|-\|-\|-\|-\|-\|-\| \| ACC/ACC* \| 90.5/87.1 \| 91.2/87.3 \| 91.0/88.4 \| 89.4/86.2 \| 88.4/85.7 \| 87.1/85.3 \| \| GTSRB \| $T=0.5$ \| $T=0.7$ \| $T=1$(default) \| $T=1.2 $ \|$T=1.5$ \| $T=2$ \| \|-\|-\|-\|-\|-\|-\|-\| \| ACC/ACC* \| 95.2/92.7 \| 94.8/93.0 \|96.6/93.2 \| 95.7/92.1 \|94.3/92.8 \|92.7/91.5 \| Title: Response to Reviewer PnTa --- Rebuttal Comment 3.1: Comment: Thanks for the additional reply. > A miscalibrated classifier [...] can lead to poor decision-making... In classification tasks as those reported in the paper, calibration is not a factor, since confidence is not used for classification (only argmax). > Since standard classifiers typically exhibit approximately calibrated outputs... I think this is not in general precise, that is it's not clear whether classifiers are calibrated before applying post-processing calibration techniques (see e.g. https://arxiv.org/abs/1706.04599). > However, we believe this issue can be addressed with minimal modifications... This might be true, but should be experimentally tested and discussed in the paper.   I think this simple approach (modifying the confidence of the model via temperature rescaling), which doesn't require designing new backdoor attacks, points to weaknesses of the proposed detection method, which can potentially be exploited by more sophisticated adaptive attacks. I think this can't be simply dismissed assuming models are well calibrated, and should be discussed in the paper. Overall, the rebuttal has confirmed that the proposed method, in the current form, is susceptible to temperature rescaling and (at least partially) other adaptive attacks ($L_{adaptive1}$). Thus, I think the paper would require significant improvements, at least discussing the current limitations and potential countermeasures. Then, since the original main concerns remain, I will vote for rejecting the paper. --- Reply to Comment 3.1.1: Title: Response to Reviewer PnTa Comment: We thank the reviewer for their feedback. We believe the discussion around temperature scaling has diverged from our primary focus. Our intention was to highlight the extreme cases of a miscalibrated classifier, such as those producing one-hot or uniform outputs, which, as the reviewer suggested, are uncommon in real-world scenarios due to their lack of explainability. This is why we did not address them in our paper. While these scenarios may be relevant in theoretical contexts, they are not directly applicable to practical situations. Nonetheless, our adaptive attacks demonstrate that even in the worst-case scenarios, TRODO performs consistently. To fully address the reviewer's concerns, as they agreed, we will consider applying the softmax function ourselves rather than relying on the classifier's softmax, and we will discuss this in our paper. >This might be true, but should be experimentally tested and discussed in the paper. Our experiments were conducted under the assumption that temperature scaling was not applied ($T=1$). Therefore, applying softmax ourselves instead of using the classifier’s built-in softmax would not change the results, and the experiments would yield identical outcomes. >Overall, the rebuttal has confirmed that the proposed method is susceptible to.... adaptive attacks We believe that a fair comparison should also consider the performance of other methods under adaptive attacks for a evaluation. However, even under strong adaptive attacks, our method continues to demonstrate superior performance compared to existing scanning methods. We remain open to further discussion to address any additional concerns the reviewer may have. Sincerely, The Authors	Rebuttal 1: Rebuttal: One of the common concerns raised by the reviewers was TRODO's performance against adaptive attacks. To address this concern, we conducted additional experiments and present the results here. Also there were concerns on the details of hard transformations, which we have also addressed in the following. We assure the reviewers that we will discuss these in our final manuscript and include the provided experiments. >TRODO under Adaptive Attack: We define two types of adaptive attacks targeting TRODO, which we will describe below: TRODO operates on the implicit assumption that a classifier is more confident on in-distribution (ID) samples than on out-of-distribution (OOD) samples, as shown in the OOD detection literature. The maximum softmax probability (MSP) assigned by the classifier to an input sample is considered as its confidence level, referred to as the ID-Score in our paper. An adaptive attacker knowing our defense mechanism and aiming to undermine the core principle of TRODO, we believe training a classifier with a new loss function that equalizes the ID-Score for both ID and OOD samples would be an adversarial approach to diminish our defense. As a result, we propose the following loss function for training a $k$-class classifier $f$: $L_{adaptive1}$=$E_{(x, y) ∼ D_{\text{in}}}$ [ $- \log f_{y}(x) $] $- λ_1 E_{(x, y) ∼ D_{\text{out}}}$ [ $H(U; f(x)) $] $+ λ_2 E_{(x, y) ∼ D_{\text{in}}}$ [ $H(U; f(x)) $] Here, $x, y$ are data samples and their labels, $f_{y}(x)$ denotes $y$-th output of the classifier, $U$ is the uniform distribution over $k$ classes, and $H$ is the cross-entropy. The first term of $L_{adaptive1}$ is the classification term (cross-entropy), while the other terms force the classifier to decrease MSP (ID-Score) for ID samples while increasing it for OOD samples. We set $λ_1=λ_2 = 0.5$, inspired by [1], to balance the importance of the first term. By this loss function, we hope the ID-Score for both OOD and ID samples will change while the classifier's decisions for them remain fixed. Moreover, we define another loss function for adaptive attacking. Our study relies on the difference in ID-Score between a benign OOD sample and a perturbed OOD sample as the scanning signature, where a higher signature indicates that the classifier is trojaned. In this adaptive attack, we explicitly target this by defining a new loss function that aims to decrease the signature for trojan classifiers, making it harder for our defense to distinguish them from clean ones: $L_{adaptive2}$=$E_{(x, y) ∼ D_{\text{in}}}$ [ $- \log f_{y}(x) $] $ - λ_3 E_{(x, y) ∼ D_{\text{out}}}$ [ $H(f(x); f(x^)) $] Here, $x^$ denotes the adversarially perturbed sample. We recreated our benchmark evaluation set, including both clean and trojaned classifiers, and trained trojan models using $L_{adaptive1}$ and $L_{adaptive2} $ and examined TRODO's performance in distinguishing them. We should note that in $L_{adaptive1}$ although the second and third terms act as regularizers to disrupt our core principle, the trojaned classifier still assigns relatively high confidence to ID samples as it is forced to classify IDs properly. Moreover, the designed loss functions $L_{adaptive1}$ and $L_{adaptive2}$ make the classifiers overfit to utilized OODs, as there is no pattern behind that loss function for to be learned by the model. Furthermore, we use a random process to create OODs in TRODO (selecting random hard transformations to obtain $G(.)$ ). This offers some remedy against such adaptive attacks. Here are the results for these experiments: \|LabelMapping\|Loss\|MNIST\|CIFAR10\|GTSRB\|CIFAR100\|Pubfig\|*Avg.\| \|-\|-\|-\|-\|-\|-\|-\|-\| \|\|\|ACC/ACC\\|ACC/ACC\\|ACC/ACC\\|ACC/ACC\\|ACC/ACC\\|ACC/ACC\\| \|All-to-One\|\|\|\|\|\|\| \|\|$L_{default}$\|91.2/89.6\|91.0/88.4\|96.6/93.2\|86.7/82.5\|88.1/83.0\|90.7/87.3\| \|\|$L_{adaptive1}$\|87.1/84.8\|87.1/84.5\|91.7/89.2\|79.8/78.5\|81.0/79.8\|85.3/83.4\| \|\|$L_{adaptive2}$\|87.3/86.3\|88.1/86.6\|93.0/90.8\|83.3/81.0\|83.7/81.1\|87.1/85.2\| \|All-to-All\|\|\|\|\|\|\| \|\|$L_{default}$\|90.0/87.4\|89.3/87.5\|92.6/89.1\|82.4/85.0\|83.2/80.9\|87.5/86.0\| \|\|$L_{adaptive1}$\|76.9/74.8\|78.2/76.8\|82.1/80.4\|73.0/71.3\|69.2/67.0\|75.9/74.1\| \|\|$L_{adaptive2}$\|84.4/83.3\|85.5/83.8\|85.6/84.1\|78.5/77.3\|79.7/78.4\|82.7/81.4\| ---- >Details of Hard Transformations:* We define a set $\mathcal{T} = $ { $ \{T_i\}$ } , with each $T_i$ representing a specific type of hard augmentation. For each ID sample $x$, a random subset of $k$ members from $\mathcal{T}$ is selected and permuted, $\{T_{j_1}, T_{j_2}, \ldots, T_{j_k}\}$, and the transformations are sequentially applied, resulting in $T_{j_k}(\ldots T_{j_1}(x))$. Each transformed training sample $x$ becomes a crafted OOD $x'$, with the transformation process denoted by $G(\cdot)$, i.e., $x' = G(x)$, where $G(\cdot) = T_{j_k}(\ldots T_{j_1}(x))$. We avoid using $k=1$, which would apply only a single hard transformation, because in some cases, the hard transformation does not significantly alter the semantics. For instance, applying rotation on a "Car" image yields an OOD as it is rare in natural images. However, some semantics are rotation invariant, such as "Flower" images. Therefore, we use $k>1$, and by rule of thumb, $k=3$ has been used. This ensures that the output of $G(\cdot)$ is sufficiently shifted from the ID. For creating the set of hard transformations, we use techniques that have been demonstrated to be harmful in the literature, including Jigsaw, Random Erasing, CutPaste, Rotation, Extreme Blurring, Intense Random Cropping, Noise Injection, Extreme Cropping, Mixup, Cutout, CutMix, and Elastic Transform [2,3,4,5]. These are the methods investigated in the literature for crafting OODs. Fig5 provides some examples of crafted OODs. [1] Hendrycks et al. Deep Anomaly 2019 [2] Miyai Rethinking 2023 [3] Kalantidis Hard 2020 [4] Li Cutpaste 2021 [5] Sinha Negative 2021	NeurIPS_2024_submissions_huggingface	2,024	Summary: This paper points out a limitation of existing backdoor model scanning methods: They fail to detect backdoored models trained with adversarial training. It propose a new backdoor model scanning method by utilizing adversarial shifts in Out-of-distribution samples. Experiments on MNIST, CIFAR-10, GTSRB, CIFAR-100, and Pubfig show the effectiveness of the proposed method. Strengths: * The investigated problem is interesting. * The motivation of this paper is good. * This paper provides theoretical analysis to support the proposed observation and method. * The proposed scanning method is general to different types of backdoor attacks. Weaknesses: * The paper states that experiments were conducted using ResNet18, PreActResNet18, and ViT-B/16 models. However, Table 1 only presents results for ResNet18. While the paper claims that results for other models are included in Appendix Section M, but this section also only contains ResNet18 results. It would be beneficial to explicitly present the detection accuracy for each model across various attack scenarios. * The discussion about the adaptive attacks to the proposed method is missing, where the attacker knows about the proposed defense strategy and actively tries to evade or overcome it. For example, the adaptive attacker might able to add a loss during the backdoored model construction phase to reduce the change of the ID-Score on the backdoored models. * While theoretical analysis of the proposed method is highly recommended, the paper would benefit from more intuitive explanations of this analysis. There appears to be a lack of detailed clarification regarding the connection between the theoretical analysis and the proposed method. For instance, the relationship between Theorem 2 and the fundamental principles of the proposed method is not clearly stated. Providing more intuitive explanations of these theoretical analysis would strengthen the paper. Technical Quality: 3 Clarity: 2 Questions for Authors: See Weaknesses. Confidence: 4 Soundness: 3 Presentation: 2 Contribution: 3 Limitations: The limitations have been discussed in this paper. Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We appreciate your insightful review. Here is our detailed response: >W1: We sincerely apologize for this oversight. There were some issues with the command that removed these tables from our submitted paper. Here, we have provided those tables and assure you that they will be included in the final manuscript. We should note that in the following experiments, all components of TRODO are fixed except for the architectures of the classifiers. _Below are the detection results in terms of accuracy on standard trained evaluation sets (ACC %) and adversarially trained ones (ACC* %). Due to character limits, we have included only the top competitors from the methods considered in Table 1 of our paper._ PreAct ResNet-18 Architecture: \|LabelMapping\|Method\|MNIST\|CIFAR10\|GTSRB\|CIFAR100\|Pubfig\|*Avg.\| \|-\|-\|-\|-\|-\|-\|-\|-\| \|\|\|ACC/ACC\\|ACC/ACC\\|ACC/ACC\\|ACC/ACC\\|ACC/ACC\\|*ACC/ACC\\| \|All-to-One\|\|\|\|\|\|\|\|\| \|\|K-ARM\|69.9/54.3\|68.2/52.0\|72.0/62.8\|58.3/48.6\|61.3/48.6\|65.9/53.2\| \|\|MM-BD\|81.5/67.3\|75.6/57.1\|75.8/61.1\|86.1/72.7\|62.4/47.6\|76.3/61.2\| \|\|UMD\|79.5/58.8\|76.4/51.5\|79.2/64.7\|67.5/54.6\|64.5/47.0\|73.4/55.3\| \|\|TRODO-Zero\|85.0/78.5\|81.2/79.1\|85.2/83.9\|78.2/78.9\|73.6/72.4\|80.6/78.6\| \|\|TRODO\|92.6/88.7\|90.5/90.2\|93.4/90.1\|85.6/83.2\|80.2/78.8\|88.5/86.2\| \|All-to-All\|\|\|\|\|\|\|\|\| \|\|K-ARM\|56.8/49.7\|54.6/47.6\|57.5/48.9\|51.3/45.0\|50.6/47.3\|54.2/47.7\| \|\|MM-BD\|52.7/42.3\|49.3/35.1\|57.0/44.2\|41.3/32.0\|40.0/34.0\|48.1/37.5\| \|\|UMD\|80.7/61.5\|75.5/55.9\|83.5/64.9\|67.7/50.0\|66.7/47.5\|74.8/56.0\| \|\|TRODO-Zero\|83.9/76.8\|77.2/78.4\|87.5/83.7\|78.4/76.8\|76.0/70.3\|80.6/77.2\| \|\|TRODO\|90.8/87.9\|88.3/89.8\|92.0/87.9\|82.6/82.1\|81.0/76.6\|86.9/84.9\| ViT-B-16 Architecture: \|LabelMapping\|Method\|MNIST\|CIFAR10\|GTSRB\|CIFAR100\|Pubfig\|Avg.*\| \|-\|-\|-\|-\|-\|-\|-\|-\| \|\|\|ACC/ACC\\|ACC/ACC\\|ACC/ACC\\|ACC/ACC\\|ACC/ACC\\|*ACC/ACC\\| \|All-to-One\|\|\|\|\|\|\|\|\| \|\|K-ARM\|69.8/54.3\|68.2/52.0\|72.0/62.8\|58.3/48.6\|61.3/48.6\|65.9/53.2\| \|\|MM-BD\|72.9/58.4\|67.6/49.8\|67.5/52.5\|78.0/62.7\|54.4/39.5\|68.1/52.6\| \|\|UMD\|75.2/55.0\|69.0/45.4\|71.5/59.6\|62.5/48.6\|58.3/40.7\|67.3/49.8\| \|\|TRODO-Zero\|78.5/71.8\|72.9/71.5\|76.5/76.1\|71.6/70.2\|68.9/65.1\|73.7/71.0\| \|\|TRODO\|87.9/85.6\|82.5/84.4\|85.2/84.3\|80.3/79.6\|78.2/76.1\|82.8/82.0\| \|All-to-All\|\|\|\|\|\|\|\|\| \|\|K-ARM\|54.9/43.3\|50.5/43.1\|51.5/47.6\|46.4/39.9\|49.4/41.9\|50.5/43.2\| \|\|MM-BD\|51.9/40.7\|44.3/33.1\|57.8/41.6\|41.0/29.6\|40.4/28.0\|47.1/34.6\| \|\|UMD\|73.9/56.4\|69.4/48.6\|77.7/58.6\|58.8/43.4\|58.0/40.1\|67.5/49.4\| \|\|TRODO-Zero\|80.2/76.0\|70.4/73.3\|81.6/77.0\|74.2/69.8\|71.7/65.8\|75.6/72.4\| \|\|TRODO\|87.6/82.3\|82.6/84.3\|83.5/83.0\|79.8/77.3\|76.0/73.1\|81.9/80.0\| ----- > W2: Thank you for suggesting this evaluation, and we apologize for missing it in our submitted version. To address the mentioned concern, we have provided different scenarios of adaptive attacks in our _common response_, which we kindly ask you to review. --- > W3:** We acknowledge that the paper lacks an intuitive explanation connecting the theoretical analysis to the proposed method. The theory section aims to provide high-level intuition for the method's core principle: _trojaned models are more susceptible to adversarial perturbations, especially in near-OOD regions_. As deducible from our experiments, there is a clear performance gap between TRODO and TRODO-Zero, and the main reason for this gap is the utilization of near-OOD samples instead of arbitrary (mostly far-OOD) ones. We have also illustrated this phenomenon on Figure 2 of the paper. Using near-OOD samples enhances the change of ID-Score, resulting in more recognizable signature for trojan scanning. To further validate this intuition, we have provided Theorem 1, in which we prove that the adversarial risk on near-OOD data is more evident. In Theorem 2, we analyzed a simplified scenario using a least-square loss and two-layer networks. Despite the simplicity, these cases are indicative of the general scenario because other complex losses used in practice yield similar optimization outcomes. Additionally, two-layer networks are known to be universal approximators capable of learning any function, analogous to more complex architectures like ResNet with MSE loss. This theoretical foundation helps to explain the observed phenomena in more intricate setups, thereby reinforcing the validity of our proposed method. Regarding the connection of this Theorem with our work, it is noteworthy that the core principle of TRODO is to use the difference of ID-Score of OOD samples before and after attack as the scanning signature. This change of ID-Score is equivalent to the adversarial risk which we have defined in Section 4. According to this theorem, this risk is linearly bounded by the trigger norm ($\|\|t\|\|$), which is non-zero for backdoored models and 0 for clean ones, making them distinguishable by our signature. In future revisions, we will elaborate on this connection to provide readers with a clearer understanding of how our theoretical analysis supports and motivates the proposed method. --- Rebuttal Comment 1.1: Comment: Thanks for your detailed rebuttal. As most of my concerns are addressed, I will increase my rating to 6. --- Reply to Comment 1.1.1: Title: Appreciation for Your Feedback and Review Comment: Thank you for your valuable review and positive feedback! We are pleased to hear that your concerns have been addressed. Warm regards, The Authors	Summary: The paper introduces a novel trojan scanning method named TRODO (TROjan scanning by Detection of adversarial shifts in Out-of-distribution samples). TRODO leverages the concept of "blind spots," where trojaned classifiers mistakenly identify out-of-distribution (OOD) samples as in-distribution (ID). The method scans for these blind spots by adversarially shifting OOD samples towards in-distribution, using the increased likelihood of these perturbed samples being classified as ID as a signature for trojan detection. TRODO is both trojan and label mapping agnostic, effective even against adversarially trained trojaned classifiers, and applicable even when training data is absent. Strengths: The writing is clear and the pictures are crisp and clear. Weaknesses: Undefined Threat Model: The threat model is not clearly defined, and the adversary's capabilities are not explicitly listed. Lack of Comparison with SOTA Baselines: There is an absence of comparison with SOTA baselines, such as FreeEagle. The results listed in Table 2 (baseline-NC) differ significantly from those in FreeEagle's Table 4. Additionally, many baselines from FreeEagle's Table 4, such as STRIP[2] and ANP[3], are not discussed. [1]Fu, C., Zhang, X., Ji, S., Wang, T., Lin, P., Feng, Y., & Yin, J. (2023). {FreeEagle}: Detecting Complex Neural Trojans in {Data-Free} Cases. In 32nd USENIX Security Symposium (USENIX Security 23) (pp. 6399-6416). [2]Gao, Y., Xu, C., Wang, D., Chen, S., Ranasinghe, D. C., & Nepal, S. (2019, December). Strip: A defence against trojan attacks on deep neural networks. In Proceedings of the 35th annual computer security applications conference (pp. 113-125). [3]Wu, D., & Wang, Y. (2021). Adversarial neuron pruning purifies backdoored deep models. Advances in Neural Information Processing Systems, 34, 16913-16925. Technical Quality: 3 Clarity: 3 Questions for Authors: 1. Can you clarify your threat model? 2. Why do you have so many missing baselines? Confidence: 4 Soundness: 3 Presentation: 3 Contribution: 2 Limitations: Design is too simple Flag For Ethics Review: ['No ethics review needed.'] Rating: 5 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thank you for your valuable comments. Our responses to each point are provided below: >Q1&Undefined Threat Model Sorry for the confusion. We have briefly stated our threat model in lines 212-226 of the paper. We further present our threat model more clear in depth here. We assure the reviewer that we will improve the clarity of this section in our final manuscript. ### Attacker Capabilities - Data Poisoning and Training Influence: Attackers can poison training data [1, 3] and influence the training process [2, 4] to embed backdoors into models. - Trigger Visibility and Coverage: - Stealthy to Overt Modifications: Attackers can deploy triggers that range from undetectable to more noticeable modifications. - Local and Global Coverage: Triggers can affect specific parts of a sample [1, 4] or the entire sample [5, 6]. - Sample-Specific Attacks: Attacks can be tailored to specific samples [7], complicating detection. - Label-Consistent Mechanisms: Attackers can use poisoned inputs labeled according to their visible content, which leads to misclassification during inference [8, 5]. - Attack Types: Attacks can be either All-to-One or All-to-All. In the former, a single target class is selected and whenever the input contains the trigger, the classifier outputs the selected target class. In the latter however, there is more control over the attack. For each source class (the class to which the clean input actually belongs), an arbitrary target class can be chosen so that the presence of the trigger causes the model to classify the input as that target class. - Model Training: Models can be trained adversarially or non-adversarially. ### Attacker Goals - Embed Backdoors: Ensure the model contains backdoors that lead to misclassification during inference. - Maintain Stealthiness: Create triggers that may be undetectable or hard to detect, even under scrutiny. - Evade Detection: Implement attacks that complicate detection efforts, especially those that are sample-specific or label-consistent. ### Defender Capabilities - Model-Only Detection: The defender receives the model and may (TRODO) or may not (TRODO-Zero) have access to a small set of clean samples from the same distribution as the training data. - No Prior Knowledge Required: The detection mechanism operates without any prior knowledge of the specific type of attack or the nature of the trigger involved. ### Defender Goals - Detect Backdoors: Identify if the model has been compromised with a backdoor. - Adapt to Various Scenarios: Effectively scan and detect backdoors both in scenarios with and without clean training samples. --- >Q2&Lack of Comparison with SOTA Baselines We have compared TRODO with eight strong Trojan detector methods in our paper. To further address your concerns regarding the comparison with other baselines, we have conducted additional experiments, including comparisons with FreeEagle, STRIP, ANP, Ex-Ray, and DF-TND. The results of these comparisons are presented in the following table: \|LabelMapping\|Method\|MNIST\|CIFAR10\|GTSRB\|CIFAR100\|Pubfig\|*Avg.\| \|-\|-\|-\|-\|-\|-\|-\|-\| \|\|\|ACC/ACC\\|ACC/ACC\\|ACC/ACC\\|ACC/ACC\\|ACC/ACC\\|*ACC/ACC\\| \|All-to-One\|\|\|\|\|\|\|\|\| \|\|DF-TND\|56.6/54.8\|49.1/49.7\|60.2/57.5\|49.6/45.5\|47.8/42.2\|52.7/49.9\| \|\|STRIP\|63.3/71.0\|68.4/64.4\|71.2/61.2\|59.4/53.5\|57.2/55.6\|63.9/61.1\| \|\|ANP\|78.5/59.7\|66.5/51.9\|73.4/65.8\|76.2/69.3\|61.8/47.7\|71.3/58.9\| \|\|Ex-Ray+ABS\|66.6/46.5\|63.3/51.4\|76.9/56.5\|60.5/46.4\|60.5/50.2\|65.6/50.2\| \|\|FreeEagle\|80.2/72.9\|82.0/73.2\|81.0/82.3\|73.2/66.9\|65.0/66.0\|76.3/72.3\| \|\|TRODO-Zero\|80.9/79.3\|82.7/78.5\|84.8/83.3\|75.5/73.7\|73.2/70.6\|79.4/77.0\| \|All-to-All\|\|\|\|\|\|\|\|\| \|\|DF-TND\|23.8/20.7\|28.9/26.7\|30.9/26.6\|13.8/10.1\|15.2/12.5\|22.5/19.3\| \|\|STRIP\|33.1/28.2\|26.5/24.3\|22.8/21.1\|33.8/29.9\|29.4/27.6\|29.1/26.2\| \|\|ANP\|52.4/47.9\|44.5/40.8\|52.7/48.4\|42.5/38.2\|36.1/32.5\|45.6/41.6\| \|\|Ex-Ray+ABS\|47.4/45.4\|48.3/44.7\|52.6/50.9\|39.3/35.7\|38.2/33.3\|45.2/42.0\| \|\|FreeEagle\|79.8/75.2\|54.9/50.2\|55.2/52.9\|56.5/52.7\|48.0/46.1\|58.9/55.4\| \|\|TRODO-Zero\|82.1/80.8\|80.4/77.3\|83.8/88.6\|74.8/72.3\|75.0/75.4\|79.2/78.8\| Although FreeEagle aims to be effective against various types of trojan attacks and performs better than other baselines in this table, it is particularly vulnerable in All-to-All settings, where samples from each class can be mapped to different target classes in the presence of a trigger. FreeEagle primarily addresses the scenario where a single source class is mapped to a single target class and attempts to identify such pairs, if they exist. Thus, it can only perform well in All-to-One scenarios. --- >The results listed in Table 2 (baseline-NC) differ significantly ...** Regarding the results of NC on the TrojAI benchmark (Table 2 in our paper), it is important to note that our results are specific to the TrojAI benchmark, which is not covered in FreeEagle’s Table 4. We have included baseline results from K-arm's [9] Table 1, which is a baseline primarily focusing on this benchmark. We will ensure these details and the additional comparative analysis are explicitly included in the revised version to enhance clarity and comprehensibility. [1] Gu et al. Badnets 2017 [2] Nguyen et al. Wanet 2021 [3] Chen et al. Targeted 2017 [4] Nguyen et al. Input-aware 2020 [5] Barni et al. New backdoor 2019 [6] Wang et al. Bppattack 2022 [7] Li et al. Invisible 2021 [8] Turner et al. Label-consistent 2019 [9] Shen et al. Backdoor scanning 2021 --- Rebuttal Comment 1.1: Comment: Thank you for the clarification and additional experiments. The author should consider adding this content to the article. I have decided to raise my score. --- Rebuttal 2: Title: Thank You for Your Positive Feedback Comment: Thank you for your positive feedback and for considering a higher score for our work! We will ensure that these experiments are incorporated to further enhance the manuscript. Sincerely, The Authors	null	null	null	null
Generalization Error Bounds for Two-stage Recommender Systems with Tree Structure	Accept (oral)	Summary: This paper presents the first generalization analysis of the learning algorithm driving the two-stage recommender systems. Specifically, it considers a representative two-stage recommender system with a tree structure, which consists of an efficient tree-based retriever and a more precise yet time-consuming ranker. An error decomposition framework is proposed, based on which the Rademacher complexity is applied to derive generalization upper bounds for various tree-based retrievers using beam search, as well as for different ranker models under a shifted training distribution. The upper bounds indicate that increasing the branches in tree-based retrievers and harmonizing distributions across stages can enhance the generalization performance of two-stage recommender systems. Furthermore, this theoretical finding is validated by experiments on real-world datasets. Strengths: This paper studies a timely topic. The two-stage recommender systems are widely adopted in industry and achieve remarkable empirical performance in balancing the computational efficiency and recommendation accuracy yet lack theoretical understanding on why it works so well. It fills in this gap and provides a nice attempt for this topic. The derived generalization upper bounds are meaningful, revealing valuable insights. The derived generalization upper bounds indicate that optimizing the design choice of two-stage recommender systems. For example, increasing the branches in tree-based retrievers and harmonizing distributions across stages can enhance the generalization performance of two-stage recommender systems. Theoretical findings are aligned with empirical experiments on real-world datasets. This further demonstrates that the generalization upper bounds are useful and powerful. The generalization upper bounds are technically nontrivial and the proof idea looks general and systematic. They nicely characterize the impact of various design choices of two-stage recommender systems on the generalization performance. They have a high potential to inspire the generalization analysis of other variants of two-stage recommender systems. Weaknesses: This paper lacks a discussion on the potential limitations of the derived generalization upper bounds. I believe that this would make a good complement to the contribution. Technical Quality: 3 Clarity: 4 Questions for Authors: Can the authors comment on the potential limitations of the derived generalization upper bounds? Confidence: 5 Soundness: 3 Presentation: 4 Contribution: 3 Limitations: The authors use the Rademacher complexity as a tool for proving the generalization upper bounds. The potential limitations of this work lie in the limitations of the Rademacher complexity tool itself. I do not think it has a negative impact on the contribution of this work. Flag For Ethics Review: ['No ethics review needed.'] Rating: 7 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We greatly appreciate your high recognition of our work, and are eager to share our thoughts with you. Response to Questions 1: > Questions 1: Can the authors comment on the potential limitations of the derived generalization upper bounds? Thank you for your question. While the derived generalization bounds offer valuable insights, they come with some limitations. The tradeoff between model complexity and computational efficiency is a key consideration, as more complex models can lead to higher computational costs despite improved generalization. In addition, aligning the training distribution with the inference distribution reduces distributional bias, but also reduces the number of training samples, which may weaken the generalization guarantee. Thus, in practical applications, there is a need for careful management of these tradeoffs. --- Rebuttal Comment 1.1: Title: Thank you for the authors' responses Comment: I appreciate the authors' response. My concerns have been addressed satisfactorily, and I maintain my positive evaluation of this paper.	Summary: This paper studies two-stage recommender systems. The authors focus on analyzing the generalization error of the retriever and ranker components within a two-stage recommendation model, specifically examining the Rademacher complexity. The findings are supported by both theoretical analysis and empirical studies Strengths: 1. The paper addresses a common structure used in recommender systems, i.e., the two-stage model. It is also used in other machine learning tasks, demonstrating broader impact. 2. The analysis of generalization errors in two-stage models fills a gap in the existing literature, which has predominantly focused on efficiency-related issues such as convergence rates. This paper enhances the understanding of this model. 3. The authors investigate the impact of different scoring models on the tree retriever model. They demonstrate that different model structures result in varying levels of generalization errors. This experimental validation inspired by the theoretical analysis further supports the findings. Weaknesses: 1. Although this paper is generally well-written, I suggest the authors create a separate "Experiments" section and include a list of notations to enhance clarity. 2. Please provide more detailed explanations of "Harmonized distribution" and "Harmonized model". Technical Quality: 3 Clarity: 2 Questions for Authors: None Confidence: 3 Soundness: 3 Presentation: 2 Contribution: 2 Limitations: Although limitations are mentioned in theoretical result discussion, the authors are encouraged to create a separate "Limitations" section. Flag For Ethics Review: ['No ethics review needed.'] Rating: 7 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thank you for your recognition of our work and providing the valuable suggestions and constructive comments. Response to Weakness 1: > Weakness 1: Although this paper is generally well-written, I suggest the authors create a separate "Experiments" section and include a list of notations to enhance clarity. Thanks for your suggestion. Due to space limitations, we initially decided to organize the paper as it is. In future revisions, we will consider creating a separate "Experiments" section to better organize the content. Additionally, we will include a list of notations to enhance readability and understanding of the paper. Response to Weakness 2: > Weakness 2: Please provide more detailed explanations of "Harmonized distribution" and "Harmonized model". Thanks for your comment. In our work, the "Harmonized distribution" arises within the context of the ranker model. During model inference, the ranker model deals with a distribution of data that has been filtered by the retriever. We name the distribution of successfully retrieved data the "Harmonized distribution," denoted as $\mathcal{D^\prime}$ in the paper. This distribution is harmonized with the retriever model, and its importance lies in the fact that the accuracy of the ranker in this stage is directly influenced by this data distribution. The term "Harmonized model" in our work refers to a two-stage model where the ranker model is trained on the Harmonized distribution. Unlike typical two-stage models that are trained independently, the Harmonized model eliminates the additional bias caused by different target data distributions, leading to better overall performance.	Summary: This paper theoretically analyzes the generalization bounds of two-stage recommender systems using Rademacher complexity. It examines the generalization bounds of the tree-structured retriever and the subsequent ranker, respectively. The paper concludes that the more branches a tree-structured retriever has, the tighter the corresponding generalization bound. Additionally, the smaller the discrepancy between the training distribution and the inference distribution of the ranker, the tighter the generalization bound. The authors also conducted experiments to validate their theoretical findings. Overall, this is a solid work, providing convincing theoretical evidence and offering insightful suggestions. Strengths: 1. The paper has a clear motivation, aiming to analyze the generalization bounds of two-stage tree-structured recommender systems. The writing is clear and understandable. 2. Two-stage recommender systems are indeed one of the mainstream structures in the current field of recommender systems, especially in industry. However, there has been a lack of theoretical guarantees and guidance for such systems. This paper effectively highlights the issues related to the number of branches in tree structures and the training data for rankers, which is innovative and high-quality. 3. The paper theoretically studies the relationship between the generalization of tree-structured Retrievers and the number of branches, as well as the relationship between the generalization of rankers and the difference between their training data distribution and the Retriever's predicted distribution. These conclusions are reasonable and align with empirical knowledge. 4. The experimental results nicely validate the theoretical findings, making the paper cohesive and solid. Weaknesses: 1. The theoretical results regarding the generalization bounds of tree-structured Retrievers are highly similar to the results for hierarchical multi-class classification in reference [1], with the main difference being the consideration of the Beam Search algorithm. Additionally, the analysis of the Rademacher complexity for linear models and MLP models has been previously established. These should be referenced in the main text of the paper. 2. The paper mentions: "In our experiments, we found that a recall rate of more than 10% is typically required to see an improvement effect" in line 271. Is there a corresponding experimental analysis for this conclusion? For example, how much does the performance of the ranker model improve with different recall rates (i.e., different data volumes)? Moreover, can this conclusion be theoretically justified as well? 3. There are some typos that need to be checked, especially in the use of bold and regular fonts. For instance, the subscript 'c' in the first formula on page 4 is not bolded, and the subscript 'v’' in formula 10 in Appendix A is not bolded. Technical Quality: 3 Clarity: 3 Questions for Authors: See weaknessnes. Confidence: 4 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: NA. Flag For Ethics Review: ['No ethics review needed.'] Rating: 7 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thank you for your recognition of our work and providing the constructive comments. We will try our best to address your concerns with planned revisions based on your valuable feedback. Response to Weakness 1: Thank you for your detailed feedback. Our analysis of tree-structured models indeed draws inspiration from previous work, particularly in the context of hierarchical multi-class classification[1]. Our contribution, in addition to the main difference of extending the analysis to the Beam Search algorithm, also provides a more refined estimate specifically tailored to the tree model. Specifically, we use the mapping $c(f, \boldsymbol{x}, y)=\left(\boldsymbol{v}, \boldsymbol{v}^{\prime}\right)$, which establishes a correspondence between a sample point $(x,y)$ and both the target node and the Top-K scoring nodes along the search path within the tree. This mapping allows us to partition the $m$ sample points into several disjoint sets $ \lbrace (x_i,y_i): c\left(f, \boldsymbol{x}_i, y_i\right)=\left(\boldsymbol{v}, \boldsymbol{v}^{\prime}\right) \rbrace $, each corresponding to different node pairs in the tree. Unlike prior work[1], the proof of Theorem 1 in [1], where a trivial upper bound of $m$ is used for the size of these sets, we provide individual estimates for each set’s size. By combining these estimates, we derive a tighter overall result, leading to a key conclusion that increasing the number of branches helps to reduce the generalization error—an insight not directly captured in [1]. Regarding Rademacher complexity, our model introduces an additional tree structure, distinguishing it from traditional linear models and MLPs. Our analytical approach involves separating the traditional scoring model from the tree structure component, corresponding to the term $\mathcal{T}$ in the upper bound of the Rademacher complexity in our analysis. This reduction maps the problem to the scoring function space of traditional models, making existing analysis techniques applicable. We appreciate your comments regarding our references, and we will ensure that these works are properly cited in future revisions. [1] Rohit Babbar et al. Learning taxonomy adaptation in large-scale classification. JMLR, 17(1):3350–3386, 2016. Response to Weakness 2: Thank you for your observation. We will include experimental results in the revised manuscript to demonstrate the ranker model's performance at different recall rates. In our experiments, we varied the number of items retrieved by the model to adjust the recall rate. With $ K=40 $ fixed during inference, from Tables 1 and 2, it can be observed that the ranker model's performance significantly declined when the recall rate was as low as 7.5\% due to insufficient training data. However, when the recall rate exceeded 10.7\%, the model consistently showed improvement. The improvement in the ranker model's performance depends on sufficient training data and alignment with the target distribution. It can be observed that once the recall rate reaches a sufficient threshold, further increases in the recall rate actually cause the performance of the trained ranker model in the two-stage classification process to gradually decline, approaching the performance of the ranker model trained on the original distribution, i.e., the complete dataset. This decline occurs because the training data distribution starts to deviate from the target distribution. The most optimal setting may be to keep the number of retrieved items consistent between training and inference, provided the recall rate is relatively high in this scenario. Experimental results show that limited data initially hinders performance, but as data and alignment improve, performance increases. However, excessive data leads to misalignment and a subsequent performance decline, consistent with theoretical predictions. For a quantitative result in theory, if we use a sampling method to estimate the error between distributions, defined as $err_{\mathcal{D}} := \mathbb{E} _{ (\boldsymbol{x}, y) \sim \mathcal{D}}\|1-\frac{P'(\boldsymbol{x}, y)}{P(\boldsymbol{x}, y)} \| $, we have $$ err_{\mathcal{D}} \approx \frac{1}{m}\sum_{i=1}^m \mathbb{I}[y_i \notin \mathcal{B}(x_i)] + \left(\frac{1}{m^\prime} - \frac{1}{m}\right) \sum_{i=1}^m \mathbb{I}[y_i \in \mathcal{B}(x_i)] , $$ and if we aim to achieve an $ \epsilon $ error between the generalization error and the empirical error with probability $1-\delta $, the estimated number of training samples required can be expressed as: $$ m \geq \left(\frac{4 c_{\Phi} N(K+1) B_{\text{model}} + B_{\Phi} \sqrt{2 \log (2 / \delta)}}{\epsilon - {err}_{\mathcal{D}}}\right)^2. $$ By further comparing this estimate with the total number of samples, we can estimate the required recall rate. It is worth noting such an estimate is typically conservative, and we still recommend using the results from the experiments. Table 1: Model Performance on the Mind Dataset \| Recall \| Accuracy \| Improvement (Above 0.6500) \| \| ----------- \| -------- \| ----------------------------------- \| \| 16.1% \| 0.6717 \| Yes \| \| 25.0% \| 0.6844 \| Yes \| \| 36.2% \| 0.6685 \| Yes \| \| 49.1% \| 0.6550 \| Yes \| Table 2: Model Performance on the Movie Dataset \| Recall \| Accuracy \| Improvement (Above 0.3516) \| \| ----------- \| -------- \| ----------------------------------- \| \| 7.5% \| 0.2581 \| No \| \| 10.7% \| 0.3548 \| Yes \| \| 17.4% \| 0.3562 \| Yes \| \| 24.6% \| 0.3547 \| Yes \| Response to Weakness 3: Thank you for identifying the typos. We will correct these issues and ensure consistency in formatting throughout the manuscript.	Summary: This paper analyzes the generalization error of two-stage recommender systems with a tree structure, which consist of an efficient tree-based retriever and a more precise but time-consuming ranker. The authors use Rademacher complexity to establish generalization error upper bounds for various tree-based retrievers using beam search, as well as for different ranker models under a shifted training distribution. The key findings are that increasing the number of branches in tree-based retrievers and harmonizing distributions across stages can enhance the overall generalization performance of two-stage recommender systems, as validated through both theoretical insights and practical experiments. Strengths: The paper provides a comprehensive theoretical analysis of the generalization error bounds for two-stage recommender systems with a tree structure. This is an important contribution, as previous theoretical research in this area has been limited. The paper analyzes the generalization upper bounds for various tree-based retriever models using beam search, including linear models, multilayer perceptrons, and target attention models. This provides valuable insights into the learnability and generalization capabilities of these widely used retriever models. The paper analyzes the generalization upper bounds for ranker models under shifted training distributions. This is an important consideration, as the data distribution during inference can often differ from the training distribution in real-world recommender systems. The theoretical and empirical findings on harmonizing distributions across stages are valuable insights. The theoretical insights and guidelines derived in this paper can inform the design and development of more robust and generalizable two-stage recommender systems, with significant implications for a wide range of industries and applications. The analytical techniques and the established error decomposition framework can serve as a foundation for future research in this domain. Weaknesses: The analysis of tree-based retriever models is comprehensive, but the paper does not explore the generalization properties of other types of retriever architectures, such as deep learning-based models beyond the target attention model. Expanding the analysis to a broader range of retriever models could provide a more holistic understanding of two-stage recommender systems. The paper primarily focuses on the generalization performance, but does not delve into the computational complexity and inference latency of the proposed two-stage recommender systems. Providing a more comprehensive analysis of the efficiency and scalability aspects would further strengthen the practical relevance of the work. Technical Quality: 3 Clarity: 3 Questions for Authors: The analysis of tree-based retriever models is comprehensive, but does the paper explore the generalization properties of other types of retriever architectures, such as deep learning-based models beyond the target attention model? Would expanding the analysis to a broader range of retriever models provide a more holistic understanding of two-stage recommender systems? While the paper primarily focuses on the generalization performance, does it delve into the computational complexity and inference latency of the proposed two-stage recommender systems? Would providing a more comprehensive analysis of the efficiency and scalability aspects further strengthen the practical relevance of the work? Confidence: 4 Soundness: 3 Presentation: 3 Contribution: 4 Limitations: NA Flag For Ethics Review: ['No ethics review needed.'] Rating: 8 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We appreciate your recognition of our work and the opportunity to address the concerns raised in your review. We value your insightful feedback and would like to share our thoughts in response. Response to Questions 1: > Questions 1: The analysis of tree-based retriever models is comprehensive, but the paper does not explore the generalization properties of other types of retriever architectures, such as deep learning-based models beyond the target attention model. Expanding the analysis to a broader range of retriever models could provide a more holistic understanding of two-stage recommender systems? Thanks for your question. We fully agree that exploring retriever architectures beyond tree-based models could provide a more comprehensive understanding of two-stage recommender systems. This is a promising and broad area of research. In this work, we chose to begin with an analysis of the generalization bounds for tree-based retriever models, which are commonly used in current two-stage models, with the hope of contributing to research in this broader area. We will continue to work in this direction to advance and refine research in this area in the future. Response to Questions 2: > Questions 2: While the paper primarily focuses on the generalization performance, does it delve into the computational complexity and inference latency of the proposed two-stage recommender systems? Would providing a more comprehensive analysis of the efficiency and scalability aspects further strengthen the practical relevance of the work? Thanks again for your question. We also share the same view that efficiency and scalability are key considerations in practical recommender systems. In terms of efficiency, specifically computational complexity and inference latency, which are critical concerns in practical recommender systems, our work points to the impact of the number of branches on the performance of tree-based models. While the generalization-related conclusions may not directly correlate with efficiency, we highlight that increasing the number of branches in the tree can improve model performance, but also increases the computational complexity and inference latency of the retriever. In an extreme case, when the number of branches equals the number of items, the tree structure becomes ineffective because it requires traversing all items during inference. At this point, the retriever model essentially degenerates into a ranker model, which is more precise yet more time-consuming. The number of branches can thus be viewed as a tradeoff between performance and efficiency. In terms of scalability, the two improvement strategies derived from our theoretical analysis, increasing the number of tree branches and adjusting the training distribution of the ranker, are feasible in practice across different models. Although the conclusions may vary depending on the specific network architecture, these strategies can still provide valuable guidance for model design.	null	NeurIPS_2024_submissions_huggingface	2,024	null	null	null	null	null	null	null	null
GLinSAT: The General Linear Satisfiability Neural Network Layer By Accelerated Gradient Descent	Accept (poster)	Summary: This paper presents a new approach to enforce constraints on the output of neural networks. Termed GLinSAT, this paper extends the application scope of LinSAT from positive linear constraints to general linear constraints. The authors also adopt ideas from OptNet/Cvxpylayers to derive implicit gradients in order to save GPU memory. Experiment study on all 3 experiments involved in LinSAT paper and one more case study with negative constraint terms show the proposed method works as expected to enforce constraints, while at the same time, offering better efficiency in GPU memory and (in many cases) running time. Strengths: * Enforcing constraints to the output of neural networks is an important research direction given the fact that deep neural networks are being extended from "traditional" classification and regression tasks to more complicated real-world decision-making with complicated constraints. * The proposed approach seems technically sound. It was well-grounded with existing work (including LinSAT and OptNet) and the derivation of Lagrange duals and adoption of ADPAGD sounds novel and interesting. * The experiment study shows that the proposed GLinSAT is able to encode the constraint as expected, reaching similar (and marginally better) accuracy than LinSAT. At the same time, it is more GPU memory efficient, which is important for scaling up the applications. Weaknesses: * The major drawback of GLinSAT over LinSAT is that its inference time is longer if LinSAT is set with its default 100 iterations. It will be helpful for the readers if there is a "discussion" section about what are the practical tips and suggestions for choosing the method for encoding constraints. Technical Quality: 4 Clarity: 3 Questions for Authors: * In the discussion of the limitation of LinSAT, the authors mentioned the bin packing problem as an example. Can you elaborate on why the bin-packing constraint has negative components? * Can we derive a similar form by maximizing $\mathbf{c}^\top\mathbf{x}$? I think it is more intuitive that larger values are preserved to be larger after an activation layer (like GLinSAT). * How to compute $\partial x/\partial y$ in Eq (11)? * Please make sure the code is released and easily accessible to the research community if this paper is accepted. Confidence: 4 Soundness: 4 Presentation: 3 Contribution: 4 Limitations: N/A Flag For Ethics Review: ['No ethics review needed.'] Rating: 8 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thanks for the thorough review and valuable comments. We are encouraged with your acknowledgment of our motivation, methodology, empirical results, and contribution. Below we respond to your specific comments. W1: Inference time of GLinSAT is sometimes longer than LinSAT with its default 100 iterations and required for discussion on practical tips As shown in Section A.8 in the Appendix, we have illustrated the reason for the speed difference. When the maximum iteration number of LinSAT is set to 100, LinSAT often reports warnings like "non-zero constraint violation within max iterations". After we increase the max iteration number to 500, the number of warnings has decreased, but there are still some warnings that indicate LinSAT has not converged. At this time, we already find that GLinSAT-Dense has already become faster than LinSAT-Dense from Table A.1. If we make a fair comparison, namely setting the maximum number of iterations of LinSAT to $\\infty$ just like GLinSAT, then LinSAT sometimes will not converge and will be always iterating. In addition, thanks for your suggestion on adding a discussion section about practical tips. In our new manuscript, we will supplement the following tips for using GLinSAT. First, the regularization coefficient $1/\\theta$ controls the smoothness of the outputs. The smaller $1/\\theta$ is, the more the outputs tend to be at the extreme point of the feasible region. A recommended way is to start with $0.1$. Using different values for $1/\\theta$ in the inference stage is also worth a try. Users can also use grid search to tune the parameter for their specific tasks. Second, for large-scale problems, it will be better to use the sparse version and implicit differential for saving GPU memory. Q1: Is there any negative component in bin-packing problem? In the bin-packing problem, there are indeed negative components in the constraint. Let's take the following classic bin-packing problem as an example [RQ1.1]: $$ \\begin{array}{c} \\min\\limits_{y_j,x_{ij}} \\sum\\limits_{j = 1}^n{y_j}\\\\ s.t.\\sum\\limits_{i\\in I}{s(i)x_{ij}}\\le B{y_j}\\\\ \\sum\\limits_{j=1}^n{x_{ij}}=1\\\\ y_j\\in\\{0,1\\},x_{ij}\\in\\{0,1\\} \\end{array} $$ where $I$ is the set of items, $s(i)\\in\\mathbb{Z}^+$ is item size, $B\\in\\mathbb{Z}^+$ is bin capacity, $y_j=1$ if bin $j$ is used and $x_{ij}=1$ if item $i$ is put into bin $j$. To see the negative coefficients clearly, we can reformulate the second linear constraint into the standard form $Ax\\leq b$ as follows: $$ \\sum\\limits_{i\\in I}{s(i)x_{ij}}-B{y_j}\\le0 $$ Due to the negative coefficient in front of $y_j$, LinSAT cannot be directly used for such a constraint. Similarly, even for a simple constraint $x\\leq y$, there will be negative coefficients in front of $y$ in the canonical form. Moreover, for the more complicated 2D [RQ1.2] or 3D [RQ1.3] bin-packing problems, there will be more negative coefficients in the constraints. [RQ1.1] Bin-packing problem, Chapeter 8 in Knapsack Problems: Algorithms and Computer Implementations, John Wiley & Sons, Inc., 1990. [RQ1.2] The two-dimensional bin packing problem with variable bin sizes and costs. Discrete Optimization, 2005. [RQ1.3] Machine Learning for the Multi-Dimensional Bin Packing Problem: Literature Review and Empirical Evaluation. arXiv 2023. Q2: Can we derive a similar form by maximizing $\\boldsymbol{c}^⊤\\boldsymbol{x}$? Yes, of course! Making larger values to be larger after an activation layer sounds more reasonable. Actually, the reason why we initially use $\\min \\boldsymbol{c}^⊤\\boldsymbol{x}$ is affected by optimal transport where the objective is to minimize the total transportation cost. But in the field of neural networks, maximizing $\\boldsymbol{c}^⊤\\boldsymbol{x}$ is indeed a more suitable choice. To maximize $\\boldsymbol{c}^⊤\\boldsymbol{x}$, all we need to do is to flip the signs with respect to $c$ in the derivative and function expressions. In our new version of this paper, we will use the maximization syntax and revise the corresponding formulas to provide a more intuitive understanding. Thanks for your suggestion! Q3: How to compute $\\partial\\boldsymbol{x}/\\partial\\boldsymbol{y}$ in Eq (11)? In the actual computation process, the matrix $\\frac{\\partial\\boldsymbol{x}}{\\partial\\boldsymbol{y}}$ is not explicitly formulated for saving GPU memory. Therefore, in the Appendix of the original manuscript, we only provide the formula of $\\frac{\\partial l}{\\partial\\boldsymbol{x}}\\frac{\\partial \\boldsymbol{x}}{\\partial\\boldsymbol{y}}$ in Eq (A.17e) instead of $\\frac{{\\partial \\boldsymbol{x}}}{{\\partial \\boldsymbol{y}}}$. In our revised manuscript, we will also supplement how $\\frac{\\partial\\boldsymbol{x}}{\\partial\\boldsymbol{y}}$ is calculated to provide a better understanding on the derivative calculation. When $\\boldsymbol{x(y)}=\\boldsymbol{u}\\circ\\boldsymbol{\\sigma(-\\theta\\boldsymbol{u}\\circ(\\boldsymbol{c-A}^T\\boldsymbol{y}))}$, we can calculate the derivative with respect to $\\boldsymbol{y}$ as follows: $$ \\frac{\\partial x_q}{\\partial y_p}=\\theta x_q(u_q-x_q)A_{pq} $$ By writing the above equation into matrix form, we have: $$ \\frac{\\partial\\boldsymbol{x}}{\\partial\\boldsymbol{y}}={\\rm{\\boldsymbol{diag}}}(\\theta\\boldsymbol{x}\\circ (\\boldsymbol{u-x}))\\boldsymbol{A}^T $$ where $\\circ$ represents the element-wise multiplication, ${\\rm{\\boldsymbol{diag}}}(·)$ maps a vector to its corresponding diagonal matrix. Q4: Please make sure the code is released and easily accessible to the research community if this paper is accepted. Thanks for your recognition for our contribution. We are currently reorganizing our codes so that the research community can easily use them. Once this paper is accepted, our source code will be released in Github. ------ We hope this response could help address your concerns, and wish to receive your further feedback soon. --- Rebuttal Comment 1.1: Comment: Thanks for the rebuttal and I have no further questions. --- Reply to Comment 1.1.1: Title: Thanks for your time and effort in reviewing our paper! Comment: Glad to know your questions have been addressed! We sincerely appreciate your valuable engagement and the time and effort you dedicated to reviewing our work!	Summary: The paper introduces "GLinSAT," a novel neural network layer designed to enforce general linear and bounded constraints on neural network outputs using a differentiable approach. The method leverages entropy-regularized linear programming, transforming it into an unconstrained convex optimization problem that avoids matrix factorization and is efficiently solvable on GPUs using accelerated gradient descent. The authors present experimental results that demonstrate the effectiveness of GLinSAT compared to existing solutions across several constrained decision-making applications. Strengths: 1. The paper addresses the challenge of imposing general linear constraints on neural network outputs in a differentiable manner. The approach of using entropy-regularized linear programming and transforming it into an unconstrained optimization problem is novel and technically sound. 2. The reformulation from a projection problem to an unconstrained optimization problem not only ensures differentiability but also boosts computational efficiency, as demonstrated in the forward and backward processes through dual formulation-based optimization and implicit differentiation. 3. The experiments cover a diverse set of applications, including the traveling salesman problem, graph matching, portfolio allocation, and power system unit commitment. The proposed GLinSAT outperforms several baselines over those problems in terms of constraint satisfaction and computational efficiency. Weaknesses: 1. Besides differentiable layer-related works, the paper could benefit from a more extensive review of other works about NN feasibility over linear constraints, such as [1], [2], and [3]. 2. The motivation for using a dot product in the objective function (Eq. 3) as opposed to a standard L2/L1 norm minimization for projection problems is unclear. Clarification on the rationale behind this choice, its advantages, or specific scenarios where it is particularly beneficial would be valuable. Additionally, is dot-product-based projection the only one that admits the entropy-regularized formulation and the unconstrained dual problem? 3. The comparison between the entropy-regularized and the unregularized projection problems could be more explicitly detailed, especially how the optimality gap between the two problems is affected by $\theta$. [1] "Tordesillas, J., How, J. P., & Hutter, M. (2023). Rayen: Imposition of hard convex constraints on neural networks. arXiv preprint arXiv:2307.08336.", [2] "Zhao, T., Pan, X., Chen, M., & Low, S. (2023). Ensuring DNN solution feasibility for optimization problems with linear constraints. In The Eleventh International Conference on Learning Representations." [3] "Tabas, D., & Zhang, B. (2022, June). Computationally efficient safe reinforcement learning for power systems. In 2022 American Control Conference (ACC) (pp. 3303-3310). IEEE." Technical Quality: 3 Clarity: 3 Questions for Authors: 1. For L2 projection layer, the projection distance will be no larger than the approximation error; thus, the universal approximation ability of NN+L2-projection will be maintained. Meanwhile, with the dot-product-based projection layer, will the approximation ability of NN+Dot-Product-projection be affected? 2. A deeper discussion on the motivation of this entropy regularization design for linear programming and its related works would be appreciated. For example, is it the first time to formulate such a dot-product projection problem with entropy regularization in literature? 3. Providing more detailed information about the problem sizes used in experiments (e.g., dimensions of m and n) would help in assessing the scalability and practical applicability of GLinSAT. 4. Since GLinSAT is designed for continuous problems, the experiments contain 3 discrete problems. It would be beneficial for the authors to explain in detail how GLinSAT is applied to these discrete problems and how feasibility over discrete decision variables is recovered. 5. Why not compare GLinSAT with CvxpyLayers/OptNet with the L2 projection? 6. Why can the feasibility of GLinSAT not be guaranteed in Table 6? Confidence: 4 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: The authors have discussed limitations in Appendix A3. Flag For Ethics Review: ['No ethics review needed.'] Rating: 7 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thanks for the thorough review and valuable comments. We are encouraged with your acknowledgment of our methodology, empirical results, and contribution. Below we respond to your specific comments. W1: Some related works need to be reviewed Thanks for sharing these works other than differentiable-optimizer-based layers for linear constraint satisfaction. The first work, Rayen, calculates null space for equality constraints in the offline stage and then use the basis for convex constraint satisfaction in the online stage. However, its efficiency may be affected when constraints are not fixed. As the basis for the null space of a matrix is usually dense, calculating null space for large matrices may face efficiency and memory challenges. The second work shrinks the feasible region of inequality constraints and decides DNN size for guaranteeing feasibility through solving a complicated bilevel programming. However, as it can only handle inequality constraints, equality constraints must be removed through Gaussian elimination, which may lead to denser matrices. The third work uses gauge function to map the neural network outputs from a $\\infty$-norm unit ball into a given polyhedral. Despite its success in the field of control, this method may encounter difficulties in handling equality constraints since the polyhedral need to contain the origin as an interior point. All these works are very insightful and provide excellent ideas for constraint satisfaction. We will include these works in our new version of literature review. W2 & Q2: Motivation for entropic regularization and dot-product based projection instead of L2/L1 normalization Thanks for the comment. We provide the details in the general response. W3: Comparison between entropy-regularized and unregularized problems Here, based on existing results in Table 6, we conducted more detailed tests and comparisons on unit commitment problems under different $1/\\theta$ values in the validation stage. The results after rounding operations are shown as follows: \|$\\;1/\\theta$\|Feasible Ratio\|Average Gap for Feasible Cases\| \|:----------:\|:------------:\|:----------------------------:\| \|0.01\|86.23%\|0.1119%\| \|0.005\|95.41%\|0.1381%\| \|0.001\|98.17%\|0.1109%\| \|0.0005\|100%\|0.1114%\| \|0.0001\|100%\|0.1114%\| \|0\|100%\|0.1114%\| From the table, we can find that no matter what the values of $1/\\theta$ are, the average optimality gaps are all around 0.1%. We can also find that in this case, as $1/\\theta$ decreases, the feasible ratio will increase. When $1/\\theta\\le0.0005$, our method can reach a 100% feasible ratio and the average gaps all tend to 0.1114%. Q1: Universal approximation ability of NN+L2-projection and NN+Dot-Production-projection As pointed in the answer to W2 & Q2 in the general response, L2-projection and softmax function can be also regarded as dot-product projection methods. We believe the key to the universal approximation ability is not the projection type, but the nonlinearity of the input-output mapping. Since the optimal solution should satisfy the KKT condition with respect to the input parameters, according to implicit function theorem, the optimal solution can be viewed as a nonlinear function of inputs. In this way, universal approximation theorem still holds. Q3: Detailed Problem Size The table below shows the size of constraint matrices when we stack constraints into block diagonal form to exploit parallelism. We will also include this information in our new version of paper. \|\|$\\hspace{1.5em}$TSP$\\hspace{1.5em}$\|Partial Graph Matching\|Portfolio Allocation\|Unit Commitment\| \|:--:\|:--------:\|:--------------------:\|:------------------:\|:-------------:\| \|$m$\|~40,000\|~2,500\|~250\|~1,000,000\| \|$n$\|~400,000\|~13,000\|~60,000\|~2,000,000\| Q4 & Q6: Integer constraint satisfaction & Feasibility issues in Table 6 Thanks for the comment. As shown in Section A.8, A.9, A.11 in Appendix, when it comes to integer constraints, we need to do some post-processing since GLinSAT is originally designed for linear constraint satisfaction. We restate the details in the general response. Moreover, from our response to W3, we can find that in this case, when $1/\\theta\\le0.0005$, our method can reach a 100% feasible ratio. Q5: Comparison with L2-projection As shown in the answer to W2 & Q2 in the general response, L2-projection is just one special case of dot-product-based projection where the quadratic regularization coefficient need to be $0.5$. In Section 3, we have already used OptNet where the quadratic regularization coefficient is set to $0.1$. Here, we also provide the result when using L2-projection. Since CvxpyLayers and OptNet is time-consuming, we only provide the results of partial graph matching and portfolio allocation. \|$\\hspace{2em}$Problem Type$\\hspace{2em}$\|Mean F1 using L2-projection\|Mean F1 when Quadratic-$1/\\theta=0.1$\| \|:--------------------:\|:-------------------------:\|:-----------------------------------:\| \|Partial graph matching\|0.614\|0.619\| \|$\\hspace{2em}$Problem Type$\\hspace{2em}$\|S. Ratio using L2-projection\|S. Ratio when Quadratic-$1/\\theta=0.1$\| \|:------------------:\|:--------------------------:\|:------------------------------------:\| \|Portfolio allocation\|2.504\|2.553\| From the tables above, we can find that using a larger regularization coefficient in these two cases didn't produce better results than before. ------ We hope this response could help address your concerns, and wish to receive your further feedback soon. --- Rebuttal Comment 1.1: Title: Thanks for your rebuttal Comment: The rebuttal provided has addressed my concerns and questions. I will adjust my evaluation scores accordingly. --- Reply to Comment 1.1.1: Title: Thanks for increasing your rating! Comment: Thank you very much for increasing your rating! Glad to know your concerns and questions have been addressed. Thank you for your time and effort in reviewing our paper!	Summary: This paper proposes a new architecture named GLinSAT that projects the output of a neural network to the feasible region of bounded and general linear constraints. The approach is based on a gradient-based approach for solving an entropy-regularized linear program and can be implemented with backpropagation/differentiation. Numerical results are reported for several tasks. Strengths: 1. Outputing solutions satisfying linear constraints is an important topic and the idea in this paper is reasonable and smooth. 2. The method works for general linear constraints and can be implemented efficiently (free of matrix factorization and can be differentiated with backpropagation). 3. The numerical results look nice compared with previous works. Weaknesses: The authors do not discuss how integer constraints are satisfied when some entries of $x$ are constrained to be integers. Note that several tasks in the numerical part are optimization over integer variables. Technical Quality: 3 Clarity: 3 Questions for Authors: How do you choose $\theta$ in practice? Confidence: 2 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: The authors clearly discuss limitations in Appendix A.3. Flag For Ethics Review: ['No ethics review needed.'] Rating: 5 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thanks for the valuable comments, constructive advice, and recognition of our idea and empirical results. Below we respond to your specific comments. W1: How to make it possible for integer constraint satisfaction? Thanks for the comment. As shown in Section A.8, A.9, A.11 in Appendix, when it comes to integer constraints, we need to do some post-processing since GLinSAT is originally designed for linear constraint satisfaction. We restate the details about integer constraint satisfaction in the general response. Q1: How to choose $θ$ in practice? Actually, the regularization coefficient $1/\\theta$ controls the smoothness of the outputs. The smaller $1/\\theta$ is, the more the outputs tend to be at the extreme point of the feasible region. A recommended way is to start with $0.1$, but users can also use methods like grid search to tune the parameter for their specific tasks. From the empirical results shown in Table 3-6, using a small value of $1/\\theta$ usually yields better results in these tasks. In addition, during the inference phase, using different values for $1/\\theta$ than what was used in the training phase is also worth a try. Users may also choose the best parameter value through grid search. ------ We hope this response could help address your concerns, and wish to receive your further feedback soon.	Summary: The main contribution of this paper is a method for enforcing arbitrary linear constraints on the outputs of a neural network. This is achieved in a differentiable way so that the entire pipeline can be trained end to end to solve constrained optimization problems. This is achieved by viewing the problem of projecting the outputs of the neural network to the feasible region as an entropy-regularized LP. This in turn can be reformulated as a convex unconstrained optimization problem with Lipschitz gradients. The problem is solved via adaptive primal-dual gradient descent and the authors provide a more efficient way to perform the backward pass than by just calling autodiff. The method is significantly more efficient than previous work and it works well on several problems. Strengths: - I think this is an interesting approach to satisfying linear constraints on the outputs of a neural network and it builds nicely on ideas from the OT literature. - The proposed method is considerably more efficient than previous works and it improves nicely over previous work on positive linear satisfiability via OT. - Empirically, the proposed method seems to improve over previous work. Weaknesses: - In terms of contribution, I would argue that this is somewhat incremental, given that previous works can handle some of the same problems. The method here is not limited by positivity like Linsatnet,which is nice, but I am not convinced it's a sufficiently strong contribution. However, there is another stated benefit of the method which is its computational efficiency. This leads me to my next point. - My main concern regarding the experimental evaluation is the scale of those experiments. For example, the TSP results are on quite small instances (20 nodes) which is quite far from anything that could be considered practical. How far can the method scale and how well does it do compared to a solver in those cases? The introduction of the paper argues that solvers can be quite expensive in order to motivate a neural approach but how does Gurobi do on those same problems. How much time does it take to solve them? Is there any setting in which this approach is competitive with Gurobi? I am asking this because for some combinatorial optimization problems (including classical TSP) there are neural network-based approaches that can have competitive results vs Gurobi given a certain time budget. How does the method here do against Gurobi for different budgets? Overall, I like the direction of the paper and the approach it takes but I am not convinced by the empirical results. I start with a tentative score which I will update after the rebuttal. Technical Quality: 3 Clarity: 3 Questions for Authors: see above Confidence: 3 Soundness: 3 Presentation: 3 Contribution: 2 Limitations: n/a Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thanks for the thorough review and valuable comments. We are encouraged with your acknowledgement to our idea, experiments and contribution. Below we respond to your specific comments. W1: Is the contribution compared with previous works incremental? Although existing satisfiability layers, such as LinSAT, CvxpyLayers and OptNet, can handle a few problems, they have significant drawbacks in either applicable scenarios or computational efficiency. Our proposed GLinSAT is the first satisfiability layer that can fully use GPU's parallel computing capability through matrix-factorization-free operations while supporting general linear constraints at the same time. We formulate the linear satisfaction problem as entropy-regularized linear programming and prove that it can be transformed into unconstrained convex optimization with Lipschitz continuous gradient. Thus it can be solved by gradient descent based algorithms. Compared with LinSAT, our proposed GLinSAT supports general linear constraints and thus enhances the expressiveness of satisfiability layers. Even for a simple constraint $x\\le y$, namely $x-y\\le0$, LinSAT cannot be used due to the negative coefficient in front of $y$, which shows the limited expressiveness of LinSAT. Moreover, we want to emphasize that making neural network outputs satisfy general linear constraints is of great significance for practical application, since negative coefficients occur frequently in real-life constrained decision-making problems, such as startup shutdown constraints in unit commitment [RW1.1], packing constraints in bin packing [RW1.2], flow conservation constraints in network flow [RW1.3], etc. Although other satisfiability layers, such as CvxpyLayers and OptNet, can be used for general linear constraint satisfaction, they may encounter efficiency issues. Take the unit commitment case as an example. In Table A.5, GLinSAT is ~10 times faster than OptNet and ~100 times faster than CvxpyLayers, which achieves significant acceleration. Considering we need ~1 day to train with GLinSAT, then we will need ~10 days with OptNet and ~100 days with CvxpyLayers, which is totally unacceptable. Our proposed GLinSAT have dramatically shortened the training time for neural networks with satisfiability layers, making what was previously impractical now viable and efficient. In addition, compared with LinSAT, our proposed GLinSAT can achieve a lower time complexity. In Section A.6, we further show the time complexity of GLinSAT is $O(\\sqrt{\\theta/\\epsilon})$, which is lower than LinSAT's $O(\\theta/\\epsilon^2)$. Moreover, our proposed GLinSAT is guaranteed to converge while LinSAT sometimes is not. From empirical results in Section A.8, we found LinSAT often reports warnings like "non-zero constraint violation within max iterations" when maximum iteration number is 100. If we make a fair comparison, namely setting the maximum iteration number in LinSAT to $\\infty$ just like GLinSAT, then LinSAT sometimes cannot converge. Therefore, our proposed GLinSAT is more reliable. [RW1.1] On mixed-integer programming formulations for the unit commitment problem. INFORMS Journal on Computing, 2020. [RW1.2] Machine Learning for the Multi-Dimensional Bin Packing Problem: Literature Review and Empirical Evaluation. arXiv 2023. [RW1.3] Machine learning based approaches to solve the maximum flow network interdiction problem. Computers & Industrial Engineering, 2022. W2: Scalability issues We conduct experiments on 500-node TSP to show the ability of GLinSAT on large-scale problems. Despite the supplementary TSP experiments, in the original paper, the unit commitment experiments can also demonstrate such an ability. As the number of nodes increases, the number of edges and decision variables can grow quadratically. In the 500-node TSP, there are about 250,000 decision variables, leading to large-scale problems. We train neural networks using GLinSAT-Sparse with $1/\\theta=0.1$. In the validation stage, beamsearch post-processing is used for obtaining feasible tours. We also use Gurobi with 32 threads to solve the MTZ formulation of TSP. The results of Gurobi under different time budgets and our approach are shown as follows: \|$\\hspace{11em}$TSP-StartEnd$\\hspace{8em}$TSP-Priority$\\hspace{4.4em}$\| \|:----------------------------------------------------------:\| \|$\\hspace{2em}$Method$\\hspace{2em}$\|Mean Time\|Mean Tour Length\|Mean Time\|Mean Tour Length\| \|:--------------------:\|:------------:\|:--------------:\|:------------:\|:--------------:\| \|GLinSAT\|3.225s\|35.323\|3.275s\|36.005\| \|Gurobi-1200s\|1200s\|52.820\|1200s\|82.242\| \|Gurobi-2400s\|2400s\|35.266\|2400s\|45.700\| From the table above, we can find that, solving 500-node TSP via Gurobi is extremely time-consuming while our method can generate feasible solution in few seconds. Since solving via Gurobi requires a large amount of time and computer threads, we can only compare the performance on 100 randomly generated cases. In TSP-StartEnd, the performance of our method is similar to Gurobi using 32 threads within 2400 seconds. As for the more complicated TSP-Priority, the solution from Gurobi within 2400s cannot outperform GLinSAT's solution in average. Moreover, the empirical results from the unit commitment experiment can also demonstrate the scalability of the proposed method. In this experiment, we use a real-life bulk power system which consists of about 360 units while the total number of considered time steps $T$ is set to 96. Consequently, when we stack constraints into block diagonal form to exploit parallelism, there are about 1,000,000 rows and 2,000,000 columns in the coefficient matrix. Experimental results in Table A.5 of the Appendix also demonstrate the applicability and efficiency of our proposed method in such a large case. ------ We hope this response could help address your concerns, and wish to receive your further feedback soon. --- Rebuttal Comment 1.1: Comment: OK, that's much better. Thank you! I will update my score accordingly. --- Reply to Comment 1.1.1: Title: Thanks for increasing your rating! Comment: Thank you very much for increasing your rating, and for acknowledging our response and additional experimental results! We sincerely appreciate your time and effort in reviewing our work!	Rebuttal 1: Rebuttal: Dear Chairs and Reviewers, We greatly appreciate the reviewers' time, valuable comments, and constructive suggestions. We are delighted that all the reviews have expressed a positive inclination towards accepting our submission. Overall, the reviewers acknowledge our methodology as "interesting" (6WNJ, U49f), "nice" (6WNJ, MtC5), "novel" and "sound" (TMC5, U49f), with empirical results showing better efficiency than existing methods (6WNJ, MtC5, TMC5, U49f). In the author response period, we make every effort to address reviewers’ concerns. Below are answers to some common questions. (1) Motivation for entropic regularization and dot-product based projection in Eq. (3) instead of L2/L1 normalization First, if L1-norm is used as the objective, then the optimization problem will become a linear programming (LP). As pointed by [R1.1], the optimal solution to an LP may not be differentiable (or even continuous) with respect to its parameters. As a result, the non-differentiability of L1-norm may make neural networks unable to train. As for L2-norm, we can show that it is also a dot-product based projection, but with an additional quadratic regularization term since $\\min\\frac{1}{2}\\left\\\|\\boldsymbol{y-x}\\right\\\|_2^2$ $\\Leftrightarrow$ $\\min-{\\boldsymbol{y}^T}\\boldsymbol{x}+\\frac{1}{2}\\sum{x_i^2}$, where $\\boldsymbol{y}$ is the output of the previous layer, $\\boldsymbol{x}$ is the decision variable and the dot product can be regarded as a measure of vector similarity. From above we can see that the main difference between Eq. (3) and L2-norm lies in the regularization term. In Section 2.1 of our paper, we have already mentioned the motivation for using dot-product based projection and entropic regularization, which is inspired by optimal transport. The formulation of optimal transport is given as follows [R1.2]: $$ \\begin{array}{c} \\min\\limits_{X_{ij}\\ge0}\\sum\\limits_{i,j}{(c_{ij}X_{ij}+\\eta X_{ij}\\log X_{ij})}\\\\ {\\rm{s.t.}}\\boldsymbol{X1=r},{\\boldsymbol{X}^T}\\boldsymbol{1=c} \\end{array} $$ A similar example is the softmax function since it is the solution to the following optimization: $$ \\begin{array}{c} \\min\\limits_{\\boldsymbol{x\\ge0}}-\\boldsymbol{y}^T\\boldsymbol{x}+\\boldsymbol{1}^T(\\boldsymbol{x}\\circ\\boldsymbol{\\log x})\\\\ {\\rm{s.t.}}\\;\\boldsymbol{1}^T\\boldsymbol{x}=1 \\end{array} $$ In Remark 1 of Section 2.1, we also show that the main benefit of using entropic regularization terms instead of quadratic regularization terms mainly comes from $\\boldsymbol{x\\log x+(1-x)\\log(1-x)}$ being an essentially smooth function [R1.3]. Therefore, the infimum in Eq. (4) can be attained only on a stationary point and we can derive an unconstrained dual optimization problem. Moreover, as pointed by reviewer U49f, making larger values to be larger after an activation layer sounds more reasonable. Therefore, in our new version of paper, we will use the $\\min-\\boldsymbol{c}^T\\boldsymbol{x}$ syntax to provide a more intuitive understanding. As for whether the proposed approach is the only way to formulate unconstrained dual optimization problems, so far we are not sure. Perhaps this could become a direction for future research. [R1.1] Melding the data-decisions pipeline: Decision-focused learning for combinatorial optimization, AAAI 2019. [R1.2] Sinkhorn distances: Lightspeed computation of optimal transport. NeurIPS 2013. [R1.3] Convex analysis, Chapter 26. Princeton university press, 1997. (2) How to make it possible to satisfy integer constraints? As shown in Section A.8, A.9, A.11 in Appendix, when it comes to integer constraint, we need to do some post-processing since GLinSAT is originally designed for linear constraint satisfaction. Here, we restate these methods more clearly. For the TSP problem, we have used two post-processing techniques, one is rounding and the other one is beam search. In the validation stage, direct rounding on the outputs of GLinSAT can satisfy all constraints in about 94% of cases, while beam search can satisfy all constraints in 100% cases. The results can be found in Table 3 and Table A.2. For the partial graph matching problem, we use Hungarian algorithm and greedy strategy for post-processing. We regard the cost of matching a pair of nodes as the outputs of GLinSAT, then use Hungarian algorithm to obtain a maximum matching. Finally, we use greedy strategy to preserve pairs with $p$-highest matching scores for constraint satisfaction. The mean F1 scores can be found in Table 4. For the unit commitment problem, we want to predict the optimal values of unit status while satisfying the logical constraint and the minimum up-time and down-time constraints. After we obtain the outputs of GLinSAT, we round the outputs to 0 or 1. Although GLinSAT can guarantee the feasibility of linear constraints prior to rounding operations, the inevitable rounding off errors may lead to infeasibility in few cases after rounding operations. However, as shown in Table 6, the feasible ratio will increase as $1/\\theta$ decreases and reach 100% when $1/\\theta=0$. The reason is that when $1/\\theta$ gets smaller, the output of GLinSAT gets closer to the extreme point of the feasible region. As pointed by [R2.1], when we consider the logical constraint and the minimum up-time and down-time constraints, these constraints formulate a convex hull so that the extreme points of the corresponding feasible region are binary. In this case, the output of GLinSAT will tend to be binary when $1/\\theta\\to0$. [R2.1] Minimum up/down polytopes of the unit commitment problem with start-up costs. IBM Research Report RC23628, 2005. ------ In our individual responses, we provide detailed answers to all the specific questions raised by the reviewers. We hope these responses could help address the reviewers' concerns, and further discussions are welcomed towards a comprehensive evaluation of our work.	NeurIPS_2024_submissions_huggingface	2,024	null	null	null	null	null	null	null	null
AR-Pro: Counterfactual Explanations for Anomaly Repair with Formal Properties	Accept (poster)	Summary: The paper addresses the anomaly explanation task by repairing the normal appearances of input inputs. Specifically, this paper designs four properties to guide the repair process that works in both image and time series domains. To demonstrate the effectiveness of the proposed method, this paper conducts experiments on the VisA and SWaT datasets. Strengths: The motivation is clear, thus, repairing the normal appearances for better anomaly explanation. The designed four properties to guide the repair process are reasonable. The proposed metrics are reasonable and the proposed method achieves significant improvements versus the utilized baseline. Weaknesses: This paper exploits fixed anomaly detection methods, i.e., Fastflow and GPT2, to guide the anomaly repair process. The two selected methods are out-of-date, and the influence of the utilized anomaly detection methods should also be investigated. More datasets should be included but not only VisA and SWaT. For example, for industrial image anomaly detection, MVTec AD should also be included. For me, Property 1 is a combination of Property 3 and 4, and maybe only Property 2, 3, and 4 are enough. The figure quality should be improved, especially for Figure 3. For the fourth metric M_{1-w}, should it be the absolute value? Or it will encourage lower anomaly scores for normal regions in the fixed results. Sec 3.3 is quite important but is not clear. The author may condense Sec. 2 a lot and then extend Sec 3.3, at least elaborating on the details of the generation process and adding proper references. ------------post rebuttal response-------- the authors have addressed all my concerns, so I raise my rate from 5 to 6. Technical Quality: 3 Clarity: 3 Questions for Authors: See the Weaknesses. Confidence: 4 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: See the Weaknesses. Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We thank the reviewer for their useful suggestions on how to improve our experiments and presentation. We will incorporate these changes into our manuscript, and we believe that they will help us greatly improve the quality of our work. We have included some results of our work-in-progress experiments in the supplemental material, namely the addition of new models (EfficientAD [1] for vision, Llama-2-7b [2] for time-series) and datasets (MVTec [3] for vision, WADI [4] and HAI [5] for time-series). Below, we respond to the reviewer's comments and questions. * Additional Models and Datasets. We are in the process of incorporating more recent models (EfficientAD, Llama-2-7b) and datasets (MVTec, WADI, HAI) to strengthen our experiments. In particular, WADI [4] is a time-series dataset that focuses on water distribution systems and contains anomalies related to sensor malfunctions and system faults. HAI [5] is another time-series dataset used for industrial control systems and includes anomalies such as cyber-attacks and operational failures. Like SWaT, both WADI and HAI have ground-truth feature-level annotations of anomalies. Due to limits in computing resources, we only include a sample of the ongoing experiments in the supplemental PDF. Importantly, we observe similar trends as our other experiments. * Relation Between Property 1 and Properties 3 and 4. The reviewer is correct in noting the relation between Property 1 and Properties 3 and 4. After some deliberation, we felt that it was simpler for the exposition to keep them separate, particularly in their loss function encodings of Section 3.2. We tried some other formulations but could not reach a satisfactory balance of a smooth exposition and notational compactness. We are open to suggestions on how to rework the presentation, and we would be glad to hear if the reviewer might have some ideas. * Figure 3 Quality. We have provided an updated Figure 3 in our supplementary PDF. * Metric of $M_{1-w}$. We intended for this to be without the absolute value. Intuitively, the repair can have a lower anomaly score in the normal region since fixing the anomalous region may reduce the overall score of nearby features. * Clarity of Section 3.3. We thank the reviewer for identifying this weakness. We will revise this section for greater clarity and detail. [1] Batzner, Kilian, Lars Heckler, and Rebecca König. "Efficientad: Accurate visual anomaly detection at millisecond-level latencies." Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. 2024. [2] Touvron, Hugo, et al. "Llama 2: Open foundation and fine-tuned chat models." arXiv preprint arXiv:2307.09288 (2023). [3] Bergmann, Paul, et al. "MVTec AD--A comprehensive real-world dataset for unsupervised anomaly detection." Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. 2019. [4] Ahmed, Chuadhry Mujeeb, Venkata Reddy Palleti, and Aditya P. Mathur. "WADI: a water distribution testbed for research in the design of secure cyber physical systems." Proceedings of the 3rd international workshop on cyber-physical systems for smart water networks. 2017. [5] Shin, Hyeok-Ki, et al. "{HAI} 1.0:{HIL-based} Augmented {ICS} Security Dataset." 13Th USENIX workshop on cyber security experimentation and test (CSET 20). 2020. --- Rebuttal Comment 1.1: Title: Response Comment: Dear Authors, Thank you so much for your efforts. Explaining anomalies by repairing them is an interesting direction, and your responses have addressed my concerns. I would suggest trying to repair semantic anomalies, like those anomalies in MVTec LOCO in the future since the structural anomalies are typically easy for users to understand the reason to be anomalous, but in comparison, users can suffer from explaining semantic anomalies without proper prior information. --- Reply to Comment 1.1.1: Title: Response to Reviewer k8Af Comment: We thank the reviewer for their encouragement and dataset suggestions. We will try to incorporate MVTec LOCO into our paper. Meanwhile, if the reviewer has any additional comments, questions, or requests, please do let us know.	Summary: The paper proposes an anomaly repair technique. Based on proposed four properties, it trains a generative model that can fix anomaly data to a benign one. The proposed properties include similar to the data, etc., which are globally applicable to any dataset. The evaluation is performed on VisA dataset and the GPT2 on SWat Dataset. The results show that the proposed method can effectively repair the anomaly data. Strengths: The paper can properly formulate the proposed properties and mix them into the training of the generative model. The proposed method can effectively repair the anomaly data. The paper is well-written and easy to follow. Weaknesses: * I do not get the motivation of fixing anomaly data. If it is detected and considered as anomaly, our typical action is to determine if it is a false positive and then improve the detection model and following downstream models. What is the rationale of fixing the anomaly data? * Properties 3 and 4 seem to be generalized version of 1 and 2. Typically, this should be solved in the training of models, either a vision model or time-series. Why not directly embedding these properties into the training of final model instead of training a separate model? * The method assumes the availability of anomaly map, which seems to be impractical in real-world scenarios. How can the method be applied to real-world scenarios where anomaly maps are not available? In the real world, we tend to only have individual anomaly data points, not a map. Similarly, it would be great if you can extend the discussion to the region selector. Technical Quality: 2 Clarity: 2 Questions for Authors: What is the motivation of fixing the anomaly data? Confidence: 4 Soundness: 2 Presentation: 2 Contribution: 1 Limitations: The paper does not explicitly discuss the limitations and potential negative. As for me, my concern is this can be used as a tool to evade anomaly detection models. The authors should discuss this in the paper. Flag For Ethics Review: ['No ethics review needed.'] Rating: 3 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We thank the reviewer for their feedback on problem motivation and potential risks. We will revise our manuscript to address these concerns. Below are our responses to the comments and questions. * Motivation for Fixing Anomalous Inputs. Anomaly repair is useful when the input data is noisy and needs to be cleaned [1]. This is studied in the context of image data [2, 10], time-series signals [3,4], graph data [5], and also heterogeneous data [6]. A common application is to improve the quality of the training data, while another is to recover from distribution shifts and allow downstream tasks to operate on data that is more in-distribution, i.e. rainy conditions in autonomous driving [7], geolocation data [8]. There are different ways to perform the repair, for example, [9] repairs anomalies through semantic-preserving transformations on images, while our work uses diffusion models for images and time-series, guided by formal specifications. In addition, we see a good opportunity to use repairs as an explainability method to improve the interpretability of black-box machine-learning models. Our proposed framework for counterfactual explanations is a step towards helping users better diagnose whether anomaly "hits" are indeed false positives by revealing what a non-anomalous input should have been like. Counterfactual explanations are especially relevant if the user is inexperienced or the data is complex [11]. * Properties 3 and 4 vs. Properties 1 and 2. Although Properties 3 and 4 are similar to Properties 1 and 2, their respective loss function encodings (Section 3.2) are different. Namely, the loss for Property 1 may be negative, while those for Property 2, 3, and 4 may not be. Moreover, Property 2 concerns similarity, while Property 4 concerns the anomaly score. We will clarify the exposition around this part. * Embedding Properties into Training. We apologize for not fully understanding the reviewer's question, and we would appreciate some clarification. In the meantime, we hope the following can serve as a partial answer. We have previously tried to encode these properties into a repair model's training objective. In particular, we attempted to perform anomaly repair on image data with VAEs. Despite our best efforts, the repair models often produced blurry or incorrect outputs. It was only when we switched to iterative diffusion-style methods that we could attain high-quality repairs. We will expand our discussions on this. * Availability of the Anomaly Map. The availability of the anomaly map is indeed a concern, especially if the detector were proprietary or closed-source. For open-sourced models implemented with libraries like PyTorch, we found that the anomaly map was usually available. For instance, the anomaly map was available for many detectors in Anomalib [12]. Nevertheless, the reviewer raises a valid point that the availability of the anomaly map is dependent on software implementation and should not be assumed. We will update our manuscript to address this limitation. * Region Selector. Our anomalous region selector is the same as the commonly used thresholding methods. The user specifies a threshold vector $\tau \in \mathbb{R}^{n}$ to specify the anomaly threshold of each feature. * Evasion of Anomaly Detectors. Since our work uses the detector's anomaly score in an optimization objective, attackers could potentially misuse our techniques. We will update our manuscript to discuss these risks. However, if attacks on anomaly detectors become common, it would likely spur interest in developing robust detectors, similar to the advancements in adversarial image classification and LLM jailbreaking defenses that followed extensive study of defensing against attacks. [1] Yang, J., Zhou, K., Li, Y., & Liu, Z. (2024). Generalized out-of-distribution detection: A survey. International Journal of Computer Vision. [2] Eduardo, S. F. L. M. (2023). Data cleaning with variational autoencoders. [3] Wang, X., & Wang, C. (2019). Time series data cleaning: A survey. Ieee Access, 8, 1866-1881. [4] Zhang, A., Song, S., Wang, J., & Yu, P. S. (2017). Time series data cleaning: From anomaly detection to anomaly repairing. Proceedings of the VLDB Endowment, 10(10), 1046-1057. [5] Akoglu, L., Tong, H., & Koutra, D. (2015). Graph based anomaly detection and description: a survey. Data mining and knowledge discovery, 29, 626-688. [6] Eduardo, S., Nazábal, A., Williams, C. K., & Sutton, C. (2020, June). Robust variational autoencoders for outlier detection and repair of mixed-type data. In International Conference on Artificial Intelligence and Statistics. PMLR. [7] Filos, A., Tigkas, P., McAllister, R., Rhinehart, N., Levine, S., & Gal, Y. (2020, November). Can autonomous vehicles identify, recover from, and adapt to distribution shifts?. In International Conference on Machine Learning. PMLR. [8] Corizzo, R., Ceci, M., & Japkowicz, N. (2019). Anomaly detection and repair for accurate predictions in geo-distributed big data. Big Data Research. [9] Lin, V., Jang, K. J., Dutta, S., Caprio, M., Sokolsky, O., & Lee, I. (2024, June). DC4L: Distribution shift recovery via data-driven control for deep learning models. In 6th Annual Learning for Dynamics & Control Conference. PMLR. [10] Pirnay, Jonathan, and Keng Chai. "Inpainting transformer for anomaly detection." In International Conference on Image Analysis and Processing. Cham: Springer International Publishing, 2022. [11] Verma, Sahil, John Dickerson, and Keegan Hines. "Counterfactual explanations for machine learning: A review." arXiv preprint arXiv:2010.10596 2 (2020) [12] Akcay, Samet, Dick Ameln, Ashwin Vaidya, Barath Lakshmanan, Nilesh Ahuja, and Utku Genc. "Anomalib: A deep learning library for anomaly detection." In 2022 IEEE International Conference on Image Processing (ICIP). IEEE, 2022. --- Rebuttal Comment 1.1: Title: Follow up Comment: Could you explain why encoding the properties as training objectives fail? Thanks. --- Reply to Comment 1.1.1: Title: Explanation of why encoding as training objectives fail Comment: Thank you for your response. We found that augmenting the training loss often failed to produce effective counterfactuals, either due to limitations in the model architecture or because of the model's natural training loss. After many attempts, we discovered that guided diffusion was a straightforward algorithm capable of achieving good performance. Our eventual choice of algorithm was primarily influenced by the available generative models for images (VAEs, GANs, Diffusion), and we elaborate on the challenges associated with each below. We initially tried VAEs for image repair but found the generated counterfactuals were often blurry, which is a known limitation of VAE [1,2]. Although later work describes methods for high-resolution VAE-based image generation [3,4], these methods often fell short in producing the same level of sharpness and detail compared to GANs and diffusion models—especially within the anomalous regions. So why are VAEs commonly used for reconstruction-based anomaly detection despite the blurry outputs? This is likely because although their ELBO loss tends to favor images that are the "average" of a distribution (i.e., blurry), this often suffices for detecting anomalous regions. Given the limitations of VAEs, we turned to GANs and diffusion models. However, GAN training was plagued by well-known issues of instability and non-convergence [5,6], thereby leading us to focus on diffusion models. With diffusion models, we found that incorporating the four properties into the training loss did not enhance counterfactual quality. This was likely because the diffusion model's noise prediction loss often led to static-like reconstructions (very noisy images) that were unsuitable for evaluation against our properties, especially at larger time steps. It was with guided diffusion methods [7] that we achieved high-quality counterfactuals, and these methods now form the foundation of our present methodology. We thank the reviewer for these questions, and we will update our manuscript to include a discussion of the lessons learned with various model architectures. Additional References [1] Tomczak, Jakub, and Max Welling. "VAE with a VampPrior." In International conference on artificial intelligence and statistics, pp. 1214-1223. PMLR, 2018. [2] Dai, Bin, and David Wipf. "Diagnosing and enhancing VAE models." arXiv preprint arXiv:1903.05789 (2019). [3] Razavi, Ali, Aaron Van den Oord, and Oriol Vinyals. "Generating diverse high-fidelity images with vq-vae-2." Advances in neural information processing systems 32 (2019). [4] Liu, Zhi-Song, Wan-Chi Siu, and Yui-Lam Chan. "Photo-realistic image super-resolution via variational autoencoders." IEEE Transactions on Circuits and Systems for video Technology 31, no. 4 (2020): 1351-1365. [5] Saxena, Divya, and Jiannong Cao. "Generative adversarial networks (GANs) challenges, solutions, and future directions." ACM Computing Surveys (CSUR) 54, no. 3 (2021): 1-42. [6] Lu, Yuzhen, Dong Chen, Ebenezer Olaniyi, and Yanbo Huang. "Generative adversarial networks (GANs) for image augmentation in agriculture: A systematic review." Computers and Electronics in Agriculture 200 (2022): 107208. [7] Dhariwal, Prafulla, and Alexander Nichol. "Diffusion models beat gans on image synthesis." Advances in neural information processing systems 34 (2021): 8780-8794.	Summary: Paper proposes a method for anomaly repair that goes one step beyond an anomaly detection and/or an anomaly localization method. While anomaly detection focuses on identifying which objects (images, time series, etc.) are anomalous, and anomaly localization focuses on identifying regions within the object (image or time series) which is anomalous, anomaly repair focuses on producing the normal object that the anomalous object is derived from. Authors identify properties for the repair, and develop a generative model that can take an anomalous object and produces the corresponding normal (and repaired) object. Paper describes experimental results to demonstrate the effectiveness of the proposed approach on different data sets. Evaluation is done both quantitatively and qualitatively. Strengths: Paper is well written except for some minor notational inconsistencies (see my questions). The idea is interesting and novel and targets an important and practical issue of anomaly repair. Experimental evaluation is robust and provides evidence regarding the effectiveness of the proposed method. Weaknesses: A primary weakness of this paper is that it does not state the assumptions regarding the scope of the methods upfront. The analysis holds for methods that follow the reconstruction-based anomaly detection recipe, i.e., each input is reconstructed, and the anomaly score is calculated using the difference between the input and reconstruction. While that is true for many methods, there is still a vast majority of methods for which this is not applicable. In fact, even in the reconstruction based methods, the analysis holds for those in which the scoring function can be decomposed over the individual features. Again, this is not true for all reconstruction-based methods. It would be good if the authors can make this clear in the beginning to avoid confusion. Technical Quality: 4 Clarity: 4 Questions for Authors: - In Section 2, the definition of anomaly map has a term $\hat{x}$, which has not been defined. From the figure 2 it appears that $\hat{x}$ is the reconstructed version of the original data point. Does that mean that this analysis framework is applicable only to the class of anomaly detection methods that involve reconstruction. While those methods are certainly capable, there is a large class of anomaly detection methods that do not necessarily have a reconstruction step involved. In fact, does the whole approach rely on availability of a base anomaly detector? - Can this method be applied of multivariate instances (no spatial or temporal relationships). It is unclear how a single threshold ($\tau$) in line 106 be applied to a case where each feature could have a different scale. - What is $\alpha_i(x)$ in Definition 2.1? Is it the absolute difference between the actual value and the reconstructed value for the $i^{th}$ feature? - In Section 3.1, what does a "normal input" mean? Does it refer to an entire observation (an image) that does not have any anomalous features or does it refer to the non-anomalous parts of an image? - In Figure 5(a), why do the authors infer that the proposed method is not adding a spurious signal, when the plot shows that it is different from the normal signal Confidence: 4 Soundness: 4 Presentation: 4 Contribution: 4 Limitations: I think the paper does not quite scope the method - identify what kind of problems will it work on and what it won't. Flag For Ethics Review: ['No ethics review needed.'] Rating: 7 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We thank the reviewer for their positive review. Their feedback will help us greatly improve the clarity throughout the manuscript. We address the reviewer's comments and questions below. * Scope of the assumptions. Although we use reconstruction-based anomaly detection as a motivating example, we are not restricted to such methods. Rather, our framework focuses on anomaly detectors with a linearly decomposable score (Definition 2.1). Linear decomposability includes many reconstruction-based methods (e.g., VAEs) as well as maximum likelihood-based ones (e.g., FastFlow). Despite this, it does not cover all anomaly detectors, such as clustering-based methods. We will improve our manuscript to better discuss the scope and limitations of our work to avoid confusion. * Definition of $\hat{x}$ in Section 2. $\hat{x}$ denotes the reconstructed input obtained from reconstruction-based methods. We will update our manuscript to clarify this. * Availability of base anomaly detector. We assume that the base anomaly detector is available. This is because we take the anomaly score into consideration when performing repairs. We will update our manuscript to clarify this. * Extension to multivariate instances. Our framework can generalize to the multivariate case. For an input $x \in \mathbb{R}^n$, we use the $n$-dimensional thresholding vector $\tau \in \mathbb{R}^n$ to allow for fine-grained control of the anomaly region selector $\omega(x) \in \\{0,1\\}^n$ at each feature. * $\alpha_i (x)$ in Definition 2.1. $\alpha_i (x)$ is the $i$th coordinate value of the anomaly map. In the given example, it is the absolute value between the reconstruction and original image on the $i$th feature. We will update our exposition to clarify this. * Definition of “normal input”. We will update Section 3.1 of our manuscript to clarify that “normal input” means that the entire input (e.g., whole image) is considered non-anomalous. * Spurious/normal signal in Figure 5(a). This figure compares our method (green) with the baseline (red) and shows that we generate qualitatively better (less “spurious”) signals. We will improve the wording in the experiments section to avoid confusion for future readers. --- Rebuttal Comment 1.1: Comment: Many thanks for your clarifications. I stand by my positive rating.	null	null	Rebuttal 1: Rebuttal: We thank the reviewers for their time and feedback. Their comments and suggestions will allow us to greatly improve our manuscript in exposition narrative, technical details, and experiment results. We are in the process of running additional experiments involving newer models like EfficientAD [1] and Llama-2 [2], as well as datasets like MVTec [3], WADI [4], and HAI [5]. We have attached some preliminary results in our supplemental PDF. [1] Batzner, Kilian, Lars Heckler, and Rebecca König. "Efficientad: Accurate visual anomaly detection at millisecond-level latencies." Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. 2024. [2] Touvron, Hugo, et al. "Llama 2: Open foundation and fine-tuned chat models." arXiv preprint arXiv:2307.09288 (2023). [3] Bergmann, Paul, et al. "MVTec AD--A comprehensive real-world dataset for unsupervised anomaly detection." Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. 2019. [4] Ahmed, Chuadhry Mujeeb, Venkata Reddy Palleti, and Aditya P. Mathur. "WADI: a water distribution testbed for research in the design of secure cyber physical systems." Proceedings of the 3rd international workshop on cyber-physical systems for smart water networks. 2017. [5] Shin, Hyeok-Ki, et al. "{HAI} 1.0:{HIL-based} Augmented {ICS} Security Dataset." 13Th USENIX workshop on cyber security experimentation and test (CSET 20). 2020. Pdf: /pdf/e2edaa8a33262f8f2d4c66a341e718ddbe41b3f9.pdf	NeurIPS_2024_submissions_huggingface	2,024	null	null	null	null	null	null	null	null
Should We Really Edit Language Models? On the Evaluation of Edited Language Models	Accept (poster)	Summary: The paper explores the general abilities of post-edited language models. Concretely, the paper performs a comprehensive empirical evaluation no various model editing methods and language models. The paper summarizes key findings based on the number of edits, the scale of the language model, the type of tuning, and the safety. The paper shows various numerical metrics to support these findings. The supplementary provides detailed results for experiments and the benchmark construction. Strengths: 1. The paper is well-written, precise, and easy to follow. All findings are summarized and listed. 2. The experiments are comprehensive and detailed. The related dataset, implementations, and evaluation metrics are provided in detail and well-organized. 3. The evaluation includes several latest language models and editing methods which are useful for later research. Weaknesses: 1. Though the experiments are comprehensive and organized, the findings are more about empirical observations rather than systematic analysis. The paper does not distill or form a theory or systematic justification based on these findings. Thus, it is hard to judge academic contributions, especially in terms of the criterion of Neurips. 2. Though the experiments are comprehensive and organized, the findings (line155~159) are more about empirical observations that do not form systematic logic. 3. The paper does not propose any new algorithms to improve the sequential model editing. Technical Quality: 4 Clarity: 4 Questions for Authors: 1. Is there any insight to improve the sequential model editing based on these findings? 2. In the experiments of sequential editing, the relation of these edits is not presented. For example, if there are 20 edits, does the correlation of edits in different order influence the conclusion and results? Confidence: 2 Soundness: 4 Presentation: 4 Contribution: 3 Limitations: The author does not provide any limitations for the evaluation method. Flag For Ethics Review: ['No ethics review needed.'] Rating: 5 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thank you for your time devoted to reviewing this paper and your constructive suggestions. Here are our detailed replies to your questions. > Q1: Though the experiments are comprehensive and organized, the findings are more about empirical observations rather than systematic analysis. The paper does not distill or form a theory or systematic justification based on these findings. Thus, it is hard to judge academic contributions, especially in terms of the criterion of Neurips. Reply: Thank you for your valuable questions. I want to address your concerns in two parts: (1) Purpose and Motivation: The primary objective of this paper is not to introduce a novel editing method. Instead, we aim to provide a detailed evaluation of existing knowledge editing methods' impact on LLM performance under the sequential editing setting. By investigating the potential factors influencing the performance of edited LLMs, we hope to offer insights and guidance for future research in this area. (2) A more systematic and general view of our findings:"Elasticity and Plasticity Trade-off" We have re-organized the findings of our paper and made the following analysis and observation , which we term the Elasticity and Plasticity Trade-off (EPT): First, we define two crucial concepts in sequential knowledge editing: plasticity, referring to the model's ability to retain newly acquired knowledge through editing, and elasticity, denoting the model's ability to retain its original knowledge after editing. We observe that during large-scale sequential editing, simultaneously maintaining high plasticity (retention of updated knowledge) and high elasticity (retention of inherent knowledge) is challenging. These two objectives are inherently in conflict. We refer to this phenomenon as the Elasticity and Plasticity Trade-off, drawing parallels to the Stability-Plasticity Dilemma observed in catastrophic forgetting during neural network fine-tuning. Existing work primarily focuses on the detrimental effect of large-scale sequential editing on elasticity, neglecting the interplay between elasticity and plasticity. Our work highlights this crucial trade-off, emphasizing the inherent conflict between these two objectives. Experimental results with different editing methods on Llama2-7B (see Table 1) demonstrate that when applying editing, maintaining a balance between elasticity and plasticity requires limiting the number of edits within a reasonable range. Exceeding it can lead to model collapse, characterized by the destruction of the model's intrinsic knowledge structure. Results in Table 1 also show that some methods can maintain the EPT over thousands of edits. However, this capability is not unlimited. With a sufficiently large number of edits, model collapse still occurs. We introduce the concept of editing endurance to quantify the maximum number of edits a model can undergo while preserving the EPT. \| # Edits \| MMLU \| BBH \| TrivialQA \| Efficacy \| Generalization \| Locality \| \| :------ \| :---- \| :---- \| :-------- \| :------- \| :------------- \| :------- \| \| 0 \| 0.459 \| 0.400 \| 0.525 \| 17.4 \| 19.2 \| 86.1 \| \| 10 \| 0.459 \| 0.401 \| 0.523 \| 100 \| 95.6 \| 81.3 \| \| 100 \| 0.459 \| 0.396 \| 0.521 \| 100 \| 95.5 \| 81.6 \| \| 500 \| 0.456 \| 0.392 \| 0.499 \| 99.4 \| 95.4 \| 80.7 \| \| 1000 \| 0.457 \| 0.392 \| 0.490 \| 99.1 \| 95.4 \| 80.2 \| Table 1: Edit result for PMET To provide an intuitive understanding, we examine the weight changes with different edits in edited layer (see Table 2). This analysis aims to demonstrate that maintaining EPT necessitates keeping weight changes within a reasonable range like PMET (neither too big nor too small). \| Method \| 0 \| 10 \| 100 \| 1000 \| \| :----- \| :------ \| :------ \| :------ \| :------- \| \| ROME \| 117.053 \| 118.265 \| 269.497 \| overflow \| \| MEMIT \| 116.579 \| 116.911 \| 121.681 \| 9291.127 \| \| PMET \| 116.579 \| 116.558 \| 116.667 \| 117.749 \| \| MEND \| 126.209 \| 126.198 \| 126.198 \| 126.199 \| Table 2: L2 Norm of edited layer for different edits on Llama2-7B model. Please refer to general response part for more detailed explanation or experiments of EPT. In summary, we believe that the research questions we address are of significant importance to editing. This work will enhance the understanding of knowledge editing for the research community. > Q2: Though the experiments are comprehensive and organized, ... form systematic logic. Reply: We would like to clarify that our intuition is to (lines 155-159) pose research questions aimed at exploring common factors that might influence the performance of edited models. To enhance the systematic nature of our work, we unify our findings under the framework of the Elasticity and Plasticity Trade-off (as elaborated in the response to Q1). Consequently, the five questions posed in lines 155-159 can be considered an analysis and understanding of the editing endurance of the model. > Q3: The paper does not propose any new algorithms to improve the sequential model editing. Reply: We would like to clarify that the primary objective of this paper is to provide a comprehensive evaluation of the impact of existing knowledge editing methods on LLM performance within the context of sequential editing. We aim to investigate potential factors influencing performance, offering valuable insights and guidance for future research in this area. Our focus is not on proposing specific new methods. Given the primary area of submission for this paper is evaluation, and considering the requirements of this area, we believe that our focus on evaluating and analyzing existing methods, rather than improving specific algorithms, is appropriate. We kindly request that you take this into careful consideration. --- Rebuttal 2: Title: Kindly Reminder Comment: Dear Reviewer FFCf: Thank you very much for your dedicated review. In the rebuttal period, we provide detailed responses to all your comments and questions point-by-point. A brief summary of our responses includes: + Q1: A more systematic and general view of our findings: "Elasticity and Plasticity Trade-off". + Q2: A boarder view of research questions. + Q3: Why are there no new algorithms in our paper. If there are any remaining issues or further clarifications needed, please let us know. We are more than happy to provide additional information or explanations to help resolve any concerns. Thank you for your time and valuable feedback. Best regards --- Rebuttal Comment 2.1: Title: Response to the rebuttal Comment: Thanks for the clarification and additional results! The results have addressed my questions. I choose to raise my score from 4 to 5.	Summary: The work evaluates the impact of various editing methods on LLMs. Specifically, it investigates how different editing techniques affect the general abilities of models, considering factors such as the number of edits, model scale, safety, and different aspects of model capabilities. Interesting findings and conclusions are suggested by the results. Strengths: + The work effectively outlines the problem statement related to editing language models and provides a clear framework for evaluating the impact of different editing methods on model performance. Moreover, the study includes a thorough literature review on LLM knowledge editing. + The empirical studies conducted in the research are solid and rigorous. The study offers reliable findings that can inform future research and development in the field. + Safety is a great perspective to study the current challenges of LLM editing. Weaknesses: - The outcome of the empirical study is somehow limited and can be further extended. The results are pretty intuitive and do not surprise people much from the current understanding. The authors can provide further insights on the following aspects. 1) Among different editing methods, what are really being traded off when number of editing increases (e.g., does PMET trade editing performance to preserve more general capabilities? And why?). It’d be great to see some hypothesis/quantification to identify and push the boundaries of performance tradeoff. The outcome will be generalizable to future development of methods. 2) Also, given the results, the audience may be further interested in why the muting effect could happen and why instruction-tuned models are more robust. Performing deeper analysis and find associations on the effects would add more value to the results. - The studied problem can be better motivated. If edit needs to be done at scale, e.g., more than 1k edit, is LLM editing still desired or a refreshed fine-tuning could already do the job better? Longer context and retrieval based generation for time sensitive knowledge is also usually consider as common solutions, which may weaken the need for sequential editing at scale. - For methods such as PMET and MEND that are robust within 10^3 editing, could the authors further extend the scale of # edits to verify if the a similar drastic performance drop, as well as the muting effect happens? - The term “intrinsic knowledge structure” appears multiple times to explain the potential reason. The claims can be more solid if a definition and detail discussion can be provided regarding the knowledge structure. Technical Quality: 3 Clarity: 3 Questions for Authors: Discussed in weaknesses. Confidence: 4 Soundness: 3 Presentation: 3 Contribution: 2 Limitations: n/a Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thank you for your time devoted to reviewing this paper and your constructive suggestions. Here are our detailed replies to your questions. > Q1: The outcome of the empirical study is somehow .... would add more value to the results. Reply: Thanks for your suggestion. we have re-organized the findings of our paper and made the following analysis and observation , which we term the Elasticity and Plasticity Trade-off (EPT): First, we define two crucial concepts in sequential knowledge editing: plasticity, referring to the model's ability to retain newly acquired knowledge through editing, and elasticity, denoting the model's ability to retain its original knowledge after editing. We observe that during large-scale sequential editing, simultaneously maintaining high plasticity and high elasticity is challenging. These two objectives are inherently in conflict. We refer to this phenomenon as the Elasticity and Plasticity Trade-off, drawing parallels to the Stability-Plasticity Dilemma observed in catastrophic forgetting during neural network fine-tuning. When editing, maintaining a balance between elasticity and plasticity requires limiting the number of edits within a reasonable range. Exceeding this limit can lead to model collapse. We introduce the concept of editing endurance to quantify this limit. Please refer to general response part for detailed experiments and analysis. > Q2: The studied problem can be better motivated. If edit needs to be done at scale, e.g., more than 1k edit, is LLM editing still desired or a refreshed fine-tuning could already do the job better? Longer context and retrieval-based generation for time-sensitive knowledge is also usually considered as common solutions, which may weaken the need for sequential editing at scale. Reply: We would like to clarify that the primary objective of this paper is to provide a comprehensive evaluation of the impact of existing editing methods on LLM performance. We aim to investigate potential factors influencing performance, offering valuable insights and guidance for future research in this area. Indeed, each approach like FT, and RAG possesses unique advantages, necessitating careful selection based on the specific application scenario. Our work strives to equip researchers and engineers with a thorough understanding of knowledge editing's strengths, limitations, and performance implications, enabling them to choose the most appropriate approach for modifying model behavior. Here's a brief summary of the pros and cons of the related techniques you mentioned: Fine-tuning: * Pros: Suitable for integrating a large amount of new knowledge into the model. * Cons: Resource-intensive and prone to catastrophic forgetting. Retrieval Augmented Generation (RAG): * Pros: Enables rapid knowledge updates, provides domain-specific knowledge. * Cons: Can lead to hallucinations when retrieved knowledge conflicts with the model's internal knowledge. Retrieved text may not always be relevant to the desired topic. Knowledge Editing : * Pros: Offers fine-grained and efficient modification of model knowledge. * Cons: Existing methods only support a limited number of knowledge edits. Edited knowledge is difficult to utilize for reasoning tasks. Long-Context : * Pros: Allows models to retrieve and cross-verify information within a long context, potentially correcting errors or inconsistencies and enabling more complex reasoning. * Cons: Requires additional training, making it extremely resource-intensive. Can also lead to hallucinations when prompt knowledge conflicts with the model's inherent knowledge. > Q3: For methods such as PMET and MEND that are robust within 10^3 editing, could the authors further extend the scale of # edits to verify if a similar drastic performance drop, as well as the muting effect happens? Reply: Thank you for your question. To further explore the limitations of the PMET and MEND methods, we extended the number of experiments to 3K on Llama2-7B model. The results are as follows: \|Method\| # Edits \| MMLU \| GSM8K \| BBH \| CSQA \| \|:-:\|:------:\|:--:\|:--:\|:---:\|:----:\| \|w/o Edit\|0 \| 0.4587 \| 0.1440 \| 0.4000 \| 0.5921 \| \| PMET \| 1000 \| 0.4572 \| 0.1391 \| 0.3921 \| 0.5823 \| \|PMET \| 2000 \| 0 \| 0 \|0 \| 0\| \| MEND \| 1000 \| 0.4571 \| 0.1600 \| 0.3978 \| 0.5864 \| \|MEND \| 2000 \| 0.4581 \| 0.1501 \|0.4014 \| 0.5905\| \|MEND \| 3000 \| 0.4574 \| 0.1539 \|0.3903 \| 0.5667\| Table 1: scaling edits to 3k As we can see from above table, PMET shows the muting effect. Although MEND still maintains relatively good performance, this does not indicate that it has good performance. Instead, it has high elasticity and low plasticity. > Q4: The term “intrinsic knowledge structure” ... structure. Reply: We would like to clarify that "Intrinsic knowledge structure" is an intuitive understanding of how knowledge is stored and organized within LLMs. Extensive existing work [1,2,3,4] suggests that parameter in LLMs implicitly encode knowledge. Manipulating these weights consequently affects the stored knowledge. A line of works[5,6,7] precisely alter the behavior of model by change parameters. Such alteration can cause weight distribution shift. When weight distribution changes induced by such manipulations remain within a reasonable range, the model retains its original knowledge. However, if these weight changes become too drastic, the model's distribution collapses, making it challenging to extract new knowledge. --- ## Reference [1] Locating and Editing Factual Associations in GPT [2] Transformer Feed-Forward Layers Are Key-Value Memories [3] Knowledge Neurons in Pretrained Transformers [4] Kformer: Knowledge Injection in Transformer Feed-Forward Layers [5] Mass-Editing Memory in a Transformer [6] PMET: Precise Model Editing in a Transformer [7] Editing Common Sense in Transformers --- Rebuttal Comment 1.1: Title: Thank you for your rebuttal Comment: Dear Authors, Thank you for your thorough rebuttal. I will update my review to raise my score. However, I have some lingering questions that you might consider addressing in your paper to provide more clarity and understanding: Is model editing necessary in the age of APIs? If so, who is going to perform model editing at all? In what situations should model editing be even considered aside from FT or RAGs? --- Reply to Comment 1.1.1: Title: Thanks for response Comment: Dear Reviewer mS2B: We are glad that our responses address some of your concerns. Regarding the question you raised about the use cases for model editing: We believe that in the era of large models and APIs, model editing has become increasingly necessary, especially when resources are limited or when there are few errors that don’t justify full fine-tuning but still need correction. Consider these typical scenarios: 1. If the entity deploying the model is different from the one training it, the deploying entity might lack sufficient computational resources to correct the model's errors through fine-tuning. 2. If the same organization is responsible for both training and deploying the model, but the model only has minor errors that don’t immediately warrant fine-tuning. 3. In situations that require quick responses, such as news updates or emergency events, model editing offers a rapid method of correction without the need for a complete fine-tuning process. 4. In highly sensitive or tightly regulated fields (e.g., finance, healthcare), model editing provides a precise and controlled way to correct specific errors without affecting other parts of the model. Regarding the use cases for RAG (Retrieval-Augmented Generation) and FT (Fine-Tuning): We believe that fine-tuning is suitable for correcting a large number of errors or injecting domain knowledge that the model has not previously learned. This approach can alter the model's overall behavior or style. RAG, on the other hand, is ideal for injecting domain knowledge into a model without fine-tuning, allowing for dynamic knowledge injection, especially for frequently changing information, without impacting the model's overall capabilities. We also welcome new suggestions/comments from you! Best regards, Author of paper 4258	Summary: [Post rebuttal update] The authors have addressed my main concern about providing an overarching framework for their evaluation, using the Elasticity - Plasticity tradeoff. I will raise my score from 5 to 6. This paper evaluates the influence of several model editing methods on the models' general capabilities. Their key dimensions of analysis include performance deterioration after several sequential single edits, the robustness of instruction-tuned models, the effect of LLM size and how editing mechanisms affect model safety. The paper stresses that current editing approaches are suitable only for minor updates and highlights the need for more practical and reliable editing techniques. They reveal that while a few edits have minimal impact, extensive edits can lead to significant degradation, including a phenomenon called muting effect, where models tend to produce empty outputs. Strengths: 1. The study provides a thorough assessment of various model editing methods across different language models, offering valuable insights into their effects on general capabilities. 2. By highlighting the pitfalls of current editing methods, such as knowledge distortion and catastrophic forgetting, the paper addresses critical issues that need to be resolved for model editing approaches to work in practical deployments of LLMs. The research emphasizes the impact of editing on the overall performance of language models, including world knowledge, reading comprehension, reasoning, and safety, rather than just on specific tasks. This broader perspective is crucial for understanding the real-world implications of model editing. 3. The study’s exploration of how editing affects the safety of models is particularly valuable, as it addresses a key concern for deploying language models in sensitive applications. 4. The paper claims to be the first to comprehensively evaluate and analyze the general abilities of edited language models. If this claim is true, then it makes it a timely contribution. Weaknesses: 1. The main weakness of the paper is in the exhaustive nature of the evaluation. While the paper evaluates multiple models, the findings may not generalize across all types of language models or specific applications, potentially limiting the broader applicability of the results. It is unfair to ask the authors to work with tons of LLMs but an application stratified view is missing. 2. The paper focuses more on empirical evaluation rather than providing deep technical insights into why certain methods cause performance degradation or safety issues. A more detailed technical analysis could enhance understanding and drive innovation. Note that by technical analysis, I do not mean theoretical analysis, but rather a deeper insight as to what issues are causing these failure modes. Are we not sanitizing the edits well? Are we overfitting? Is there an early stopping criterion that may work here? While pointing out flaws in existing methodologies is useful, since there are many such instances in the LLM world, well grounded research has to provide insights as to why the methodologies are flawed. 3. As an example of (2) above: The observation of the "muting effect," where models produce empty outputs after extensive edits, is concerning. However, the paper does not propose concrete insights as to why this issue might occur, leaving an important problem unaddressed. 4. There are also several different dimensions to consider for an analysis such as this: Scale of the model, prompting conditions or methodologies, whether models benefit from RAGs or not. Since the authors stress practical deployments, they should ensure their results are consistent with practical deployments as well. Technical Quality: 3 Clarity: 4 Questions for Authors: In addition to points above, I wonder what the author's thoughts are on what their best insights are as to why the model performance degrades? How would a practitioner deploying a complex LLM API go about figuring this out? consider cases where they have full models access or just API access. Confidence: 4 Soundness: 3 Presentation: 4 Contribution: 3 Limitations: In addition to addressing the weaknesses above, I encourage the authors to think of a more comprehensive evaluation framework to showcase their work. Evaluations of LLMs are themselves brittle -- hence any evaluation needs to have sufficient power to make a solid claim. In the current situation, the paper is indeed well written, but not statistically strong nor comprehensive. Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thank you for your time devoted to reviewing this paper and your constructive suggestions. Here are our detailed replies to your questions. > Q1: The main weakness ... stratified view is missing. Reply: Thank you for your valuable feedback. We address your questions and concerns regarding the scope of our evaluation below: (1) Can findings generalize to other types of language models? Currently, the vast majority of methods in editing like ROME, MELLO, IKE, focus on decoder-only models (GPT, Llama). The exploration of knowledge editing for other encoder-only (BERT) and encoder-decoder (T5) models remains largely under-explored at present, making it infeasible to evaluate all kinds of model. (2) Can findings generalize to specific application scenarios? We want to clarify that editing is not tied to specific application scenarios. Similar to pre-training, SFT, and RLHF, it is a stage in the model whole life pipeline. Therefore, we did not evaluate on specific application scenarios. In fact, knowledge editing can be applied to LLMs across various applications, making our evaluation general and broadly applicable. To justify our conclusion in specific application scenarios, we chose QA task as example , conducting evaluation on TriviaQA dataset with Llama2-7B. The results are shown below, note the un-edited model gets 0.5252 on TriviaQA: \| Method \| 10 \| 50 \| 100 \| 1000 \| \|--------\|----\|----\|-----\|------\| \| ROME \| 0.4887 \| 0.0035 \| 0 \| 0 \| \| MEND \| 0.5288 \| 0.5248 \| 0.527 \| 0.5282 \| \| PMET \| 0.5237 \| 0.5199 \| 0.5209 \| 0.4904 \| \| MEMIT \| 0.5247 \| 0.5189 \| 0.2121 \| 0 \| \| KN \| 0 \| 0 \| 0 \| 0 \| The trends in the table are consistent with the conclusions in our paper. > Q2: The paper focuses more on ... the methodologies are flawed. Reply: Thanks comment. We answer your question in two folds: (1) Distinction between editing and FT: Knowledge editing differs from FT. Most editing methods in this paper are gradient-free, without something like early stop in FT. These methods typically bypass backpropagation and gradient descent, instead directly calculating and substituting specific parameters to achieve the desired behavior change. (2) Deeper understanding and systematic analysis: Please refer to Q3 for detailed explanation. > Q3: As an example of Q2 above .... Reply: Existing work suggests that parameter weights in LLMs represent specific knowledge. Manipulating these weights consequently affects the stored knowledge. Through extensive experimentation, we have made the following analysis and observation regarding editing, which we term the Elasticity and Plasticity Trade-off (EPT): First, we define two crucial concepts : plasticity, referring to the model's ability to retain newly acquired knowledge through editing, and elasticity, denoting the model's ability to retain its original knowledge after editing. We observe that during large-scale sequential editing, simultaneously maintaining plasticity and high elasticity is challenging. These two objectives may inherently in conflict. We refer to this as the Elasticity and Plasticity Trade-off, drawing parallels to the Stability-Plasticity Dilemma in catastrophic forgetting during fine-tuning. When editing, maintaining a balance between elasticity and plasticity requires limiting the number of edits within a reasonable range. Exceeding this limit can lead to model collapse. We introduce item of editing endurance to quantify this limit. Please refer to general response part for more detailed explanation or experiments. > Q4: There are also several different ... as well. Reply: Thanks for comment. Here are some clarifications: Part A: Other dimensions for analysis. (1) Scale of the model: This factor is actually the RQ3 of our paper, with detailed analysis and experiments presented in Section 4.3. (2) Prompting conditions: We adopted widely used prompt settings and hyperparameters from existing literature and technical reports (e.g., CoT usage, number of few-shot prompting). These details are fully listed in Appendix D.1. (3) Whether models benefit from RAGs or not: This paper does not involve the use of RAGs. Part B: Fairness and reliability of our evaluation To ensure a fair and comprehensive evaluation, we utilize widely adopted commonly used benchmarks like MMLU and BBH. We adopt standard prompt settings and hyperparameters from existing literature and technical reports (e.g., CoT usage, few-shot prompting) and utilize inference acceleration frameworks (vLLM) to align with practical application. These details are thoroughly documented in Appendix D.1, ensuring fairness and reproducibility in evaluation. > Q5: In addition ... just API access. Reply: We will answer your question from two aspects: (1) Why performance degradation occurs: We believe that the performance degradation during sequential editing stems from the disruption of the LLM's "intrinsic knowledge structure." Specifically, the model's parameter distribution undergoes changes. With a limited number of edits, the parameter distribution shift is minimal, allowing for knowledge updates while preserving performance. However, extensive editing leads to significant changes in the parameter distribution, disrupting the knowledge encoded within the parameters and resulting in performance decline. (2) Detecting performance degradation with different access levels: API Access: Users with API access cannot directly edit the model's knowledge. However, they can assess potential performance degradation by querying the model and evaluating its responses. Full Weight Access: Users with access to model weights can perform all operations described in our paper. > Q6: In addition ... nor comprehensive. Reply: Thank you for your advice. We will add our new finding and analysis in the revision. --- Rebuttal 2: Title: Kindly Reminder Comment: Dear Reviewer 2vUP: Thank you very much for your valuable comments. In the rebuttal period, we have provided detailed responses to all your comments and questions point-by-point. A brief summary of our responses includes: + Q1: Can our findings generalize to other language models or other application scenarios. + Q2: Distinction between knowledge editing and FT + Q3: Deeper understanding and systematic analysis: Elasticity and Plasticity Trade-off + Q4: Fairness and reliability of our evaluation & Other dimensions for analysis. + Q5: Why performance degradation occurs & Detecting performance degradation with different access levels. Please let us know if any remaining issues or further clarifications are needed. We are more than happy to provide additional information or explanations to help resolve any concerns. Thank you for your time and valuable feedback. Best regards	null	null	Rebuttal 1: Rebuttal: ## General Response to All of Reviewers We appreciate all the reviewers for their thoughtful comments and suggestions on our paper. We are very glad to see that the reviewers find our focused problem is important and useful (R1, R2, R3) ,the insights are valueable and reliable (R1, R2) and the experiments are comprehensive and organized (R1, R2, R3). We are pleased that the reviewers find our writing is very clear and easy to understand (R2, R3). We have tried our best to address the reviewers' comments and concerns in individual responses to each reviewer. The reviews allowed us to improve our draft. In the following part, we would like to provide more detailed version of important questions: > Q1: Motivation and goal of our work Reply: The primary objective of this paper is to provide a comprehensive evaluation of the impact of existing knowledge editing methods on LLM performance within the context of sequential editing. We aim to investigate potential factors influencing performance, offering valuable insights and guidance for future research in this area. Our focus is not on proposing specific new methods. > Q2: Deeper understanding and analyis of our findings Reply: We have re-organized the findings of our paper and made the following analysis and observation , which we term the Elasticity and Plasticity Trade-off (EPT): First, we define two crucial concepts in sequential knowledge editing: plasticity, referring to the model's ability to retain newly acquired knowledge through editing, and elasticity, denoting the model's ability to retain its original knowledge after editing. Prior to editing, the model's knowledge solely comprises its inherent, intrinsic knowledge. Post-editing, the model's knowledge encompasses two components: updated knowledge and inherent knowledge. We observe that during large-scale sequential editing, simultaneously maintaining high plasticity (retention of updated knowledge) and high elasticity (retention of inherent knowledge) is challenging. These two objectives are inherently in conflict. We refer to this phenomenon as the Elasticity and Plasticity Trade-off, drawing parallels to the Stability-Plasticity Dilemma observed in catastrophic forgetting during neural network fine-tuning. Existing work primarily focuses on the detrimental effect of large-scale sequential editing on elasticity, neglecting the interplay between elasticity and plasticity. Our work highlights this crucial trade-off, emphasizing the inherent conflict between these two objectives. Experimental results with different editing methods on Llama2-7B (see Table 1,2,3) demonstrate that when applying editing, maintaining a balance between elasticity and plasticity requires limiting the number of edits within a reasonable range. Exceeding it can lead to model collapse, characterized by the destruction of the model's intrinsic knowledge structure. Results in Table 1-3 also show that some methods can maintain the EPT over thousands of edits. However, this capability is not unlimited. With a sufficiently large number of edits, model collapse still occurs. We introduce the concept of editing endurance to quantify the maximum number of edits a model can undergo while preserving the EPT. \| # Edits \| MMLU \| BBH \| TrivialQA \| Efficacy \| Generalization \| Locality \| \| :------ \| :----- \| :----- \| :--------- \| :------- \| :------------- \| :------- \| \| 0 \| 0.459 \| 0.400 \| 0.525 \| 17.4 \| 19.2 \| 86.1 \| \| 10 \| 0.457 \| 0.396 \| 0.524 \| 100 \| 94.4 \| 92.2 \| \| 100 \| 0.443 \| 0.377 \| 0.212 \| 100 \| 94.8 \| 91.8 \| \| 500 \| 0 \| 0 \| 0 \| 98.2 \| 92.5 \| 80.7 \| \| 1000 \| 0 \| 0 \| 0 \| 98.4 \| 92.3 \| 80.4 \| Table 1: Edit result for MEMIT \| # Edits \| MMLU \| BBH \| TrivialQA \| Efficacy \| Generalization \| Locality \| \| :------ \| :---- \| :---- \| :-------- \| :------- \| :------------- \| :------- \| \| 0 \| 0.459 \| 0.400 \| 0.525 \| 17.4 \| 19.2 \| 86.1 \| \| 10 \| 0.459 \| 0.401 \| 0.523 \| 100 \| 95.6 \| 81.3 \| \| 100 \| 0.459 \| 0.396 \| 0.521 \| 100 \| 95.5 \| 81.6 \| \| 500 \| 0.456 \| 0.392 \| 0.499 \| 99.4 \| 95.4 \| 80.7 \| \| 1000 \| 0.457 \| 0.392 \| 0.490 \| 99.1 \| 95.4 \| 80.2 \| Table 2: Edit result for PMET \| # Edits \| MMLU \| BBH \| TrivialQA \| Efficacy \| Generalization \| Locality \| \| :------ \| :---- \| :---- \| :-------- \| :------- \| :------------- \| :------- \| \| 0 \| 0.459 \| 0.400 \| 0.525 \| 17.4 \| 19.2 \| 86.1 \| \| 10 \| 0.457 \| 0.394 \| 0.528 \| 99.6 \| 51.2 \| 54.2 \| \| 100 \| 0.458 \| 0.394 \| 0.527 \| 92.4 \| 47.3 \| 54.7 \| \| 500 \| 0.457 \| 0.399 \| 0.528 \| 86.4 \| 42.4 \| 57.6 \| \| 1000 \| 0.457 \| 0.397 \| 0.528 \| 68.3 \| 46.5 \| 53.8 \| Table 3: Edit result for MEND To provide an intuitive understanding of the EPT, we examine the weight changes with different edits in edited layer (see Table 4). This analysis aims to demonstrate that maintaining both elasticity and plasticity necessitates keeping weight changes within a reasonable range like PMET (neither too big nor too small). MEND lose the plasticity, while MEMIT lose the elasticity. \| Method \| 0 \| 10 \| 100 \| 1000 \| \| :----- \| :------ \| :------ \| :------ \| :------- \| \| ROME \| 117.053 \| 118.265 \| 269.497 \| overflow \| \| MEMIT \| 116.579 \| 116.911 \| 121.681 \| 9291.127 \| \| PMET \| 116.579 \| 116.558 \| 116.667 \| 117.749 \| \| MEND \| 126.209 \| 126.198 \| 126.198 \| 126.199 \| Table 4: L2 Norm of edited layer for different edits on Llama2-7B model. ---- We appreciate your comments and time! We have tried our best to address your concerns. Would you mind checking and confirming if there are any unclear parts?	NeurIPS_2024_submissions_huggingface	2,024	null	null	null	null	null	null	null	null
Do LLMs Build World Representations? Probing Through the Lens of State Abstraction	Accept (poster)	Summary: This work investigates what kind of abstractions LLMs use to encode the world, distinguishing goal-oriented abstractions (discarding world dynamics that are not necessary for achieving the goal) from world-general abstractions (including dynamics irrelevant for the goal). The authors note that prior work looking at LLM world models don't make this distinction, leading to conflicting results. For a text-based planning task, the authors probe LLMs doing the task in-context as well as fine-tuned, finding that from the latter goal-oriented abstractions can be recovered with a relatively higher accuracy than from the former. Strengths: - It's a neat idea and good contribution to formalise the types of abstractions LLMs can use to represent the world through state abstraction theory. I believe using this framework will make future work around these questions more interesting and grounded. - The paper is well-written and easy to understand - The authors design a synthetic task that is both easy to understand, has modular and distinct state abstractions, and is somewhat complex for LLMs to perform (requiring planning to perform optimally). Weaknesses: LLMs can only do this task after task-specific fine-tuning, in which case it is unsurprising that the representations are goal-oriented. This doesn't mean that LLMs doing tasks out of the box (or in-context) well won't use world-general representations The main result is that LLMs fine-tuned for a task start forming more useful goal-oriented state abstraction representations. This is unsurprising, as when fine-tuning the LLM for a task you're essentially training it to discard irrelevant information and encode relevant information for the goal. I would expect world-general abstractions to only be present in a general model that can do a task well out of the box without task-specific fine-tuning. However, the result is presented in the paper like LLMs do not use world-general representations (e.g. from the abstrcct: "Our experiments reveal that LLM-built representations tend to favor goal-oriented abstractions during decoding, prioritizing task completion over recovery of the world’s state and dynamics."). For this kind of claim to be sufficiently substantiated I would like to see the framework applied to a task that LLMs can do out of the box, or this task using LLMs that can do it in-context (e.g. larger models or better models). Alternatively, I'd like to see the text indicate that the claims are about task-specific (i..e. fine-tuned) LLMs. Missing baselines It's difficult to interpret the results without a simple baseline that learns a probe on top of a non-LLM trained to do this task. Now, we don't know whether the accuracy of around 20% for world-general predicates is because the LLM has picked up general knowledge from pretraining or because you can learn to represent these things from this task (either in-context or through fine-tuning.) Technical Quality: 3 Clarity: 3 Questions for Authors: Questions - For section 6.1, it would be useful to know what percentage of random moves would be legal. It seems like the ICL performance of llama2 and mistral are very low / close to random? - I would expect more world-general abstractions to be represented in earlier layers, and goal-oriented abstractions in the later layers. Did you look at the earlier layers for world-general predicate accuracy as well (first half)? Suggestions - I would caveat the abstract (and results mentioned elsewhere) sentence: "Our experiments reveal that LLM-built representations tend to favor goal-oriented abstractions during decoding, prioritizing task completion over recovery of the world’s state and dynamics.". You only find this for fine-tuned LLMs, and it might be entirely different for LLMs that can do a task out-of-the box. - Should the predicate be subgoal(l=1, j) and subgoal(l>1, j) in line 270 - 271? It says in line 252 that l is the relative distance and j the j-th container to be visited. Confidence: 4 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: Discussed and addressed in the appendix. Flag For Ethics Review: ['No ethics review needed.'] Rating: 7 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thank you for your time and constructive feedback. We address all of your concerns below. ### Weakness 1: Results are unsurprising > (unsurprising finding) LLMs fine-tuned for a task start forming more useful goal-oriented state abstraction representations… is unsurprising … * It's indeed surprising when viewed in the context of existing work$\color{red}{^1}$. We provide detailed elaboration in the [Global Rebuttal](https://openreview.net/forum?id=lzfzjYuWgY&noteId=i9JnaUoNpz), with other surprising findings and broader significance of our framework. We sincerely invite you to review it. > (it's unsurprising because) you're training it to discard irrelevant information and encode relevant information for the goal… * World dynamics are relevant to the goal since the optimal plan is derived based on predicting future states (illustrated in Fig.3). * On the other hand, LLMs-sft could have been abstracted away $Q^$ abstraction since the $\pi^$ abstraction alone is sufficient for predicting optimal actions. However, we successfully recovered $Q^$ abstraction from LLM-sft representations. It reveals that LLMs-sft preserve the long-term impact of each possible action, despite being fine-tuned on next-token prediction only. To the best of our knowledge, this interesting finding is novel, and we believe it can inspire future research to conduct a more in-depth investigation. > (experiments on a task that LLMs do well in context)* I would expect world-general abstractions to only be present in a general model that can do a task well out of the box without task-specific fine-tuning… * Given the space of possible tasks is extraordinarily huge, it's intractable to do an exhaustive search to identify a task that simultaneously satisfies (1) complex enough to differentiate the spaces of various abstractions, and (2) existing open-source LLMs can perfectly solve without fine-tuning. Even though we find such a task, it is hard to rule out the possibility of data contamination, especially considering that it still remains unknown if LLMs can solve a novel task that involves modeling the underlying world described by the text. * Due to these challenges, we instead start with a simple enough task. As explained in L179-184, the RePlace task closely resembles the [gripper problem](https://shorturl.at/j0wp5), one of the most basic planning tasks, where the transition is deterministic, the state can be derived without uncertainty, the goal must be achievable and the optimal plan can be searched in O(n). Therefore, we cannot make the task simpler while still having distinctive state space for different abstractions. > (experiments with pre-trained LLMs) …using LLMs that can do it in-context (e.g. larger models or better models) To fully address your concern, we conduct the same set of probing experiments with Phi3-17b, a brand-new SOTA LLM that achieves 19.94% success rate on GRIPPER with ICL only$\color{red}{^2}$. In response to your question that whether better LLMs are more likely to use general abstractions, we employ two smaller and weaker pre-trained LMs for comparative study. The new results indicate that (1) goal-oriented abstractions are probed from pre-trained Phi3-17b with significantly higher recovery rate, and (2) more advanced pre-training leads to higher priority of encoding goal-oriented abstractions over a more general one. Please refer to the Global Rebuttal for detailed results and in-depth analysis. > (clarification needed) I'd like to see the text indicate that the claims are about task-specific (i..e. fine-tuned) LLMs. Thank you for your suggestion! Due to space constraints, we summarize the findings for LLMs-sft/icl in one sentence in the abstract/Intro. In the experiments section (Sec 6.2), however, we have reported and analyzed the findings for LLMs-sft/icl separately (L320-321, L331). We will do the same thing for the abstract/Intro in the revised version to make it clear, which will be flexible given more space. ### Weakness 2: learns a probe on top of a non-LLM Thank you for your suggestion! To address your concern, we train two 6-layer decoder-only Transformers on our datasets, which achieves around 30% success rates. The average recovery rate of each abstraction is: \|Datasets\| Raw \| World \| $Q^$ \| $\pi^$\| \| --- \| --- \|--- \|--- \|--- \| \|GRIPPER\| 27.7 \| 29.9 \| 52.5 \| 28.3\| \|COOK\| 23.9 \| 23.8 \| 32.5 \|28.7\| Interestingly, the raw state and world-irrelevant abstractions can be probed with a much similar recovery rates as from LLMs-sft (\~30% for both). The main difference lies in goal-oriented abstractions, i.e. $Q^$-irrelevant abstraction (\~73% from LLMs-sft) and $\pi^$-irrelevant abstraction (\~83%). It suggests that pre-training mainly improves the encoding of goal-oriented abstractions. We will include this result in Fig. 5 in our paper. Thanks again for your valuable suggestion! ### Questions * Legal rate of random move: The legal rate of a random move is 21.65%. All LLMs-icl have significantly higher legal rates than this baseline. * Probing earlier layers: We treat the layer index as a hyperparameter (detailed in L620-L622) and use the last 6th layer based on the overall performance on the validation set. We have also experimented with using early layers. For instance, the average recovery rate (RR) of each abstraction from the 4th layer of Llama2-13b-sft is: \| Layer \| Raw \| World \| $Q^$ \| $\pi^$\| \| --- \| --- \|--- \|--- \|--- \| \| 4th \| 13.7 \| 13.0 \| 14.7 \| 7.02 \| The RR of both the general- and goal-oriented abstractions are much lower. ### Suggestions * Thank you for the suggestion! We will summarize the findings for LLMs-icl/sft separately in the abstract/Intro to avoid confusion, which would be an easy fix. * Thank you for your question! `subgoal(l, j)` denotes that the container to be visited at $j$-th step is at relative distance $l$. We will clarify it in Line 270. --- Rebuttal 2: Title: Footnotes and Reference list for Rebuttal Comment: ### Footnotes $\color{red}{^1}$ We mainly focus on reporting the findings for LLMs-sft since it is a common interest [1, 2, 3, 4] in this field that if Transformers/LLMs trained on next-token predictions develop internal world models . $\color{red}{^2}$ While this result isn't very satisfactory, it is acceptable given that it's the best ICL performance we've seen with open LLMs. We didn't conduct probing experiments on COOK, because Phi3-17 still falls short of this dataset without fine-tuning (\~3% success rate). It's expected as COOK has a more complicated structure of containers, and larger space of actions. ### Reference [1] Li, Kenneth, et al. "Emergent World Representations: Exploring a Sequence Model Trained on a Synthetic Task." ICLR, 2023. [2] Li, Belinda Z., Maxwell Nye, and Jacob Andreas. "Implicit Representations of Meaning in Neural Language Models." ACL, 2021. [3] Hazineh, Dean S., Zechen Zhang, and Jeffery Chiu. "Linear Latent World Models in Simple Transformers: A Case Study on Othello-GPT." arXiv preprint arXiv:2310.07582 (2023). [4] Kim, Najoung, and Sebastian Schuster. "Entity Tracking in Language Models." ACL, 2023. --- Rebuttal 3: Title: Thank you for your review; we're ready to answer any further questions Comment: Dear reviewer, Thank you again for your review! We really appreciate your recognition of our methodologies, new resources, and presentation. Hopefully, our additional experiments in response to your comments and in-depth elaboration on the surprisal of our findings and the broader significance of our framework can address most of your concerns. If not, please don't hesitate to ask us any further questions. We are always ready to respond. We're very looking forward to an engaged and productive discussion! --- Rebuttal 4: Comment: Thanks for the detailed rebuttal, both personal and the general one. I will respond below. ## Using FT-ed models instead of base models The fact that world dynamics are still relevant for the goal and the fact that you recovered the $Q^$ abstraction from LLM-sft representations is more convincing to me and alleviates some of my concerns. However, in the general rebuttal you claim your findings are more relevant in light of related work claiming that your finding: "undermines an increasingly widespread belief that an implicit world model can emerge from next-token prediction (NTP) training [1]. This belief is encouraged by recent studies demonstrating that world states can be recovered from Transformer representations [2,3]."* However, [3] explicitly controls for the effect on representations of pretraining and fine-tuning, from their paper section 4.2: "To determine whether the advantage is conferred by LM pretraining or fine-tuning, we ablate either open-domain pretraining, in a -pretrain,+fine-tune ablation, or in-domain finetuning, in a +pretrain,-fine-tune ablation. [...] While both fine-tuning and pretraining contribute to the final probe accuracy, pretraining appears to play a much larger role: semantic state can be recovered well from models with no in-domain fine-tuning." So even though for your task it might be the case that world-general representations can be useful, the results seemingly can't say much about whether or not pre-trained LLMs or next-word predictors learn world models. In your rebuttal, you say: ""Given the space of possible tasks is extraordinarily huge, it's intractable to do an exhaustive search to identify a task that simultaneously satisfies (1) complex enough to differentiate the spaces of various abstractions, and (2) existing open-source LLMs can perfectly solve without fine-tuning. Even though we find such a task, it is hard to rule out the possibility of data contamination, especially considering that it still remains unknown if LLMs can solve a novel task that involves modeling the underlying world described by the text. Of course, I am not suggesting you do an exhaustive search over all possible tasks, as I don't even know what that would mean in practice, but it seems highly unlikely to me that there exists no task open source LLMs can perform in-context that affords both types of abstractions. Additionally, it's not necessary for the model to perform the task perfectly without fine-tuning. Finally, data contamination can be equally an issue in your fine-tuning setup, so I don't fully understand that argument. Finally, respectfully, I disagree that it's a good reason to not include the fact that your results are only on fine-tuned LLMs in the abstract due to space constraints. The line that is misleading is the following: ""Our experiments reveal that LLM-built representations tend to favor goal-oriented abstractions during decoding, prioritizing task completion over recovery of the world’s state and dynamics."." Which can be changed to "Our experiments reveal that LLM-built representations when fine-tuned tend to favor goal-oriented abstractions during decoding, prioritizing task completion over recovery of the world’s state and dynamics.". ## Learning a probe Thanks for these additional insights. I believe it's important that these, together with the results on a pretrained model you have in the general rebuttal, are a part of the paper, as without it it remains hard to say much about which part contributes to what (pretraining vs finetuning). ## Summary To summarise, some of my concerns are alleviated, and I am still of the opinion that making the distinction between world-general and goal-oriented abstractions is a really cool contribution and makes for an interesting paper. The additional baselines and results on pretrained LLMs are interesting. This makes that I'll raise to score to a 5. The reason I did not raise it further just yet is as follows: I remain of the opinion that the results are somewhat unsurprising and do not necessarily go against prior findings. For example, you also find world-general abstractions in LLMs, just more goal-oriented ones. All in all I believe the paper requires quite a substantial rewrite to include the right baselines (non-LLMs), the experiments using a pretrained LLM from the general rebuttal, and to change the language about what exactly LLMs (when finetuned or not) do learn and how that relates to prior findings. --- Rebuttal Comment 4.1: Comment: Dear reviewer, We've just saw your most updated response. We truly appreciate your active participation in the discussion phase, especially during what might also be a busy week for you. We're grateful to hear that we've addressed some of your concerns and that you find our findings interesting. This feedback boosts our confidence and excitement to share our work with a broader community. Thank you once again for your invaluable feedback; it has greatly enhanced the quality of our work! We will be sure to incorporate the promised changes in the revised version. --- Rebuttal 5: Title: Further Clarifications (1/3) Comment: Thank you for your detailed feedback. We truly appreciate your engagement in further discussions and your acknowledgment that the facts we clarified in our rebuttal (the recovery of $Q^$-abstractions and the relevance of world dynamics for the task) make our findings more surprising. It is also gratifying to know that our additional experiments echoing our original findings, provide deeper insights into world abstractions in both pre-trained and fine-tuned LLMs. We address every single one of your remaining concerns below. --- ### Clarification about the finding based on SFT and its relation to a common belief shared by the community* In the Global Rebuttal, we stated that "(Our $\color{blue}{\text{finding}}$ that SFT mainly enhances goal-oriented abstractions) undermines an increasingly widespread belief that an implicit world model can emerge from next-token prediction (NTP) training* [1]"* and we cited two representative works [2,3] that encourage this belief. However, you raised concerns, arguing that our finding is insufficient because [3] "explicitly controls for the effect on representations of pretraining and fine-tuning" and they found "While both fine-tuning and pretraining contribute to the final probe accuracy, pretraining appears to play a much larger role". Indeed, it doesn't negate the fact that our $\color{blue}{\text{finding}}$ presents a sufficient challenge to the belief discussed above, as we will elaborate below. (1) The next-token prediction (NTP) training mentioned in the belief is not limited to pre-training alone; the findings that encourage this belief go beyond just pre-training, including conventional supervised training for particular tasks (i.e., SFT or training from scratch). As you noted, even [3] has claimed that fine-tuning enhances world modeling. Moreover, the work [2] in this area that attracted wide attention, along with more recent follow-up work, e.g. [4], has probed from Transformer models trained on synthetic tasks with NTP. Therefore, our $\color{blue}{\text{finding}}$ directly addresses the belief since SFT on NTP is the de facto standard approach to adapt LLMs to a particular task. (2) Furthermore, the results you quoted from [3], "semantic state can be recovered well from models with no in-domain fine-tuning", don't undermine our finding that pre-trained LLMs struggle to maintain a general world abstraction (in RePlace), as a reanalysis of it [5] has already revealed that the high accuracy in [3] was based on trivial cases, and the entity status indeed cannot be recovered from pre-trained LMs with reasonable performance (please refer to section 2 in [5])$\color{red}{^1}$$\color{red}{^2}$. (3) Although we don't make such a claim in the paper, it's reasonable that one may expect some insights into the world modeling capabilities of pre-trained LLMs from the findings on LLMs-sft, as several recent studies suggest that it's very rare that fine-tuning alters the pre-trained capabilities [6,7,8]; rather, it tends to accentuate them. This may indicate that SFT could potentially enhance LLMs's prioritization of specific types of word abstractions that emerge during pre-training$\color{red}{^3}$. The findings from our new experiment with pre-trained LLMs, combined with the original ones, seem to align with this general hypothesis. With all that being said, we agree with you that the $\color{blue}{\text{finding}}$ based on LLMs-sft should not be considered universally applicable to all LLM variants (as we consciously warned in L495-498), including pre-trained LLMs. In response to your suggestion, we have conducted additional experiments, as detailed in the Global Rebuttal, to provide a more comprehensive assessment of the belief in question. Even if the findings from LLMs-sft do not fully address the world abstractions in pre-trained LLMs, hopefully, we can agree that SFT is an important type of NTP training, and our new experiments probing pre-trained LLMs address this concern. --- Rebuttal 6: Title: Further Clarifications (2/3) Comment: ### Clarification about the conclusions from our experiments >So even though for your task it might be the case that world-general representations can be useful, the results seemingly can't say much about whether or not pre-trained LLMs or next-word predictors learn world models. With great respect, we have to point out that this is NOT true. 1. In our original findings, world-irrelevant abstractions are mostly missing in both LLMs-sft and LLMs-icl, which are fine-tuned and pre-trained on NTP, respectively. While you might consider the LLMs-icl results uninformative due to their poor performance on the task, the LLMs-sft results clearly demonstrate that fine-tuning on NTP doesn't result in an internal world model. 2. Further experiments, prompted by your suggestion, on pre-trained LLMs and non-LLM transformers suggest that pre-training on NTP doesn't inherently produce internal world models—instead, more advanced pre-training leads to a higher priority for encoding goal-oriented abstractions. --- > I remain of the opinion that the results … do not necessarily go against prior findings. For example, you also find world-general abstractions in LLMs, just more goal-oriented ones. This is NOT true, either. 1. Figure 6 in our paper and Plot B in the Global Rebuttal clearly show that the predicate cannot be reliably recovered unless it pertains to the goal-oriented abstractions. 2. Conversely, the predicates uniquely tied to world-irrelevant abstraction, namely `store_u`, `store_g`, and `held_g`, are probed with a recovery rate of less than 12.5% across all LLM variants (even lower than the one of `boxName`, which is unrelated to both task completion and world dynamics). Both types of NTP training—pre-training and fine-tuning—make little difference in this regard. It suggests that LLMs tend to discard predicates that serve solely for preserving world dynamics, the most critical ingredient of world models [9]. To be more concrete, LLM representations do not indicate how the world will change when an agent manipulates a certain object. 3. These results directly go against the claim from previous work that Transformers/LLMs trained on NTP develop internal world models. 4. Moreover, we want to emphasize a meta-result that diverges from previous work: probing without considering world abstractions leads to biased evaluation and unnecessary conflicts, as reiterated throughout the paper and rebuttals. --- ### Why we don't work on a simpler task (such that LLMs can do well in context) > Of course, I am not suggesting you do an exhaustive search over all possible tasks, as I don't even know what that would mean in practice, but it seems highly unlikely to me that there exists no task open source LLMs can perform in-context that affords both types of abstractions. We agree it's reasonable to expect a task that open-source LLMs can perform well in context, consisting of distinguishable world abstractions at various levels. However, finding such a task can be exhausting and tricky. Our task is already simple enough, as elaborated in our rebuttal and paper, so we didn't tweak the task setting until the LLM could achieve a reasonable performance. We were not implying that you suggested searching through all possible tasks. Apologize if there is any confusion. –-- > Finally, data contamination can be equally an issue in your fine-tuning setup, so I don't fully understand that argument. Sorry for the confusion. By data contamination, we're referring to the possibility that the LLM might have been pre-trained on very similar or almost identical tasks. However, we're not suggesting that the LLM has encountered the exact same test samples during pre-training—this is a general concern across all setups. This possibility makes it unclear whether experimental results reflect SFT or pre-training effects. It's not an issue for fine-tuning setup, as we're fully aware that the LLMs are fine-tuned specifically for this task. If this still doesn’t seem like an issue to you, we can temporarily set it aside, as it doesn't affect the validity of our findings or how we've addressed your concern. Simply put, these three points in our rebuttal were to explain why we initially didn't conduct experiments on pre-trained LLMs. However, these were relatively minor, as we later identified a better model during rebuttal, Phi3 (released after the submission deadline), which achieves acceptable performance without fine-tuning. Therefore, the additional experiments in the Global Rebuttal primarily address your concerns regarding pre-trained LLMs. --- Rebuttal Comment 6.1: Title: Further Clarifications (3/3) Comment: ### Making the scope of findings clear in abstract/introduction > I disagree that it's a good reason to not include the fact that your results are only on fine-tuned LLMs in the abstract due to space constraints. To clarify the findings in our experiment section (Finding 1): (1) goal-oriented abstractions are much more effectively recovered from LLMs-sft representations than world-irrelevant abstractions, and (2) world-irrelevant abstractions are largely absent in both LLMs-icl and LLMs-sft representations. Thanks to your thoughtful reminder, we realize that we should have summarized the findings for LLM-sft/icl separately in the abstract and introduction. As stated in our rebuttal, we will ensure this is clearly clarified in these two sections. We are also open to addressing any remaining confusion. --- ### Incorporating the new results into the paper > I believe the paper requires quite a substantial rewrite to include the right baselines (non-LLMs) and the experiments using a pretrained LLM from the general rebuttal In our humble opinion, these would be easy adjustments. First, the new results support our original findings, and second, incorporating them involves simply: (1) adding the new Plots A and B from the Global Rebuttal, presented in the same format as Figures 5 and 6 in our paper, along with the corresponding descriptions (points 1-3 in the Global Rebuttal); and (2) including the results with non-LLM baselines in Plot A and Figure 5, represented by one additional bar each. We sincerely hope the clarifications above encourage you to reassess the value of our work. We're eager to address any further concerns you may have. Thank you again for your time and dedication in the review process. --- ### Footnotes $\color{red}{^1}$ We are not intentionally downplaying the validity of [3]. We know it's one of the earliest works in this space, raising an important question and inspiring many follow-up works, including ours. $\color{red}{^2}$ We agree that it's important to distinguish the effects of pretraining from fine-tuning when drawing conclusions. Therefore, we compare probing results from LLMs-sft and LLMs-icl (L342-350, Finding 2) to investigate how SFT impacts the encoding of world abstractions. The new results (again, thanks to your suggestion) using pre-rained LLMs and non-LLM models help clarify the effects of pre-training, as discussed in the Global Rebuttal. $\color{red}{^3}$ To clarify, we're just explaining that this assumption is valid and inspired by existing work, and hence findings based on LLMs-sft could be informative and insightful. We're not suggesting that experimenting with LLMs-sft is sufficient for drawing conclusions for pre-trained LLMs. ### Reference [1] Kenneth Li, "Do Large Language Models learn world models or just surface statistics?", The Gradient, 2023. [2] Li, Kenneth, et al. "Emergent World Representations: Exploring a Sequence Model Trained on a Synthetic Task." ICLR, 2023. [3] Li, Belinda Z., Maxwell Nye, and Jacob Andreas. "Implicit Representations of Meaning in Neural Language Models." ACL, 2021. [4] Hazineh, Dean S., Zechen Zhang, and Jeffery Chiu. "Linear Latent World Models in Simple Transformers: A Case Study on Othello-GPT." arXiv preprint arXiv:2310.07582 (2023). [5] Kim, Najoung, and Sebastian Schuster. "Entity Tracking in Language Models." ACL, 2023. [6] Prakash, Nikhil, et al. "Fine-Tuning Enhances Existing Mechanisms: A Case Study on Entity Tracking." ICLR, 2024. [7] Panigrahi, Abhishek, et al. "Task-specific skill localization in fine-tuned language models." ICML, 2023. [8] Zhou, Chunting, et al. "Lima: Less is more for alignment." NeurIPS, 2024. [9] Ha, David, and Jürgen Schmidhuber. "Recurrent world models facilitate policy evolution." NeurIPS, 2018.	Summary: This paper proposes a new framework for studying world state abstractions from LLM representations. The framework, based on state abstraction theory, focuses not on assessing whether a model has a single world representation but assessing different levels of possible abstractions. These are each roughly functions of states: world-irrelevant abstractions, q-irrelevant abstractions, and pi-irrelevant abstractions. A probe is used to assess whether an LLM's representation is encoding each abstraction. Specifically, the authors design a planning task called REPLACE (where each abstraction is known to the researcher but not to the model), and train a probe on a transformer's representation to predict the abstraction when prompted with a textual description corresponding to the abstraction. In experiments, the authors prompt LLMs to perform the REPLACE task and assess which abstractions each LLM is encoding. They have four findings: 1. LLM representations reflect goal-oriented abstractions during decoding 2. Supervised-fine tuning increases the level of this goal-oriented abstraction 3. We don't see the other abstractions present during decoding 4. LLMs do not appear to build complete world representations. Strengths: I think the main strengths of the paper are as follows: 1. World abstractions: The idea of probing for different types of state abstractions is interesting. This is a new perspective for studying emergent world representations in LLMs, which have typically focused on all-or-nothing notions of world representations. 2. Clarity: The paper is mostly well-written and clearly structured. The authors do a good job of explaining their framework and describing the experimental methodology. 3. The planning task proposed in the paper (REPLACE) is an interesting task, and may be useful as a testbed for future research on analyzing world representations in language models. Weaknesses: While the framework proposed in the paper is interesting, the main weakness of the paper is that the experimental results are limited: REPLACE consists of two related planning task based on a limited set of containers and objects. Since the paper is focused on assessing world representations generally, there needs to be more evaluation settings and datasets. For example, the paper frequently refers to Othello as a testbed for assessing LLM world representations, but doesn't include any experiments involving Othello. The findings in the last section of the paper are based on insufficient evidence since they're only using the two related planning tasks that make up REPLACE. Do the same trends hold for game datasets? At the very least, more settings for REPLACE should be considered to provide more ablations of where each abstraction is recovered. Similarly, the paper could benefit from more extensive model comparisons to make sure the takeaways are robust. While the authors do cover multiple LLMs, they're all from the Llama family or Mistral. Using models from different families (or at least the larger sizes of the Llama/Mistral models) would help make the findings more compelling. Less importantly, there are a couple of writing and presentation suggestions that would improve the paper. The text in Figure 5 is difficult to read because it is small. Moreover, although Q* has a standard definition for anyone familiar with RL, it is not defined in the paper even though RL problems are defined. Technical Quality: 3 Clarity: 3 Questions for Authors: See questions above relating to robustness. Confidence: 4 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: Yes Flag For Ethics Review: ['No ethics review needed.'] Rating: 4 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thank you for your time and detailed review. We address all of your concerns below. ### Major Weakness: Running more experiments with more datasets/LLMs > REPLACE … planning task based on a limited set of containers and objects. * We acknowledge that RePlace has predefined domains of containers and objects due to the synthetic nature of the datasets. However, this synthetic setup is crucial for faithful probing experiments. (1) It allows us to manipulate and track the underlying world state for probing experiments. (2) We can deliberately balance the dataset based on various attributes, such as the order of containers and the initial location of agents and objects, to eliminate potential biases. However, it's hard to do the same thing for real-world datasets. * Due to the above reasons, almost all existing work on probing world representations uses synthetic datasets. [1,2] use game script datasets synthesized with the rule of the Othello game. [3,4] also employ synthesized text datasets based on limited sets of containers and objects, inspiring our datasets' design (mentioned in L175-176). However, we additionally have an agent exploring the environment and manipulating objects with some physical constraints to accomplish a goal, which makes the task more challenging and complex enough to differentiate the spaces of various abstractions. Furthermore, we have introduced different dataset variants to create a more diverse and realistic task setup (detailed in L227-234). * In addition, our new task spans a broad domain (elaborated in L179-182). It is closely related to the Gripper problem$\color{red}{^1}$ and the EntityTracking task [4,5], which is widely adopted in planning and NLP interpretability research. This task also involves tracking entities, aggregating multi-hop information, and recognizing textual entailment, essential skills for many real-world NLP tasks. > Since the paper is focused on assessing world representations generally, there needs to be more evaluation settings and datasets. * Our framework, the central contribution of this work, is generic and can be seamlessly applied to other LLMs and tasks wherever the abstract state spaces are distinguishable. However, we are not aiming to reach a conclusion that applies universally to all LLMs and tasks (discussed in L494-498). This is also impossible, given that there is even a lack of available testbeds. * We didn't use other datasets employed by existing work as they don't meet the necessary criteria for faithful and feasible probing (explained in L179-184, L488-493): (1) processing text data; (2) the abstractions across various levels are distinguishable; (3) the world state can be easily manipulated and tracked. Our new task/datasets fill this gap. > … doesn't include any experiments involving Othello * In the paper (L166-170), we explain why the Othello dataset is not a faithful testbed: it's not complex enough to differentiate abstract states at various levels. * Nonetheless, the Othello dataset was originally used to probe a small-scale Transformer trained on the game script data. But our work focuses on LLMs, and game script is very different from natural text that LLMs are trained on. > … more settings for REPLACE should be considered to provide more ablations … * We have done such ablation. Due to the limit of space, we highlight the conclusion in the main body of the paper (L337-L340) and provide more details in Appendix K. To sum up, we have created a new setting where an originally goal-unrelated predicate (e.g. color info) belongs to goal-oriented abstractions. As such, the recovery rate of color information has significantly increased from zero. This further consolidates our finding that SFT mainly enhances goal-oriented abstractions. * During the rebuttal period, we have done another ablation to answer Reviewer 4mUW's question, which shows that LLMs fine-tuned with sub-optimal demonstrations still prioritize goal-oriented abstractions, echoing our main finding. For the sake of not repeating, we sincerely invite you to refer to the rebuttal to Reviewer 4mUW for more details and results of this ablation study$\color{red}{^2}$. > … could benefit from more extensive model comparisons … all from the Llama family or Mistral… * We have already adopted four LLMs for experiments, each of which is experimented with both ICL and SFT, while existing work like [1,2] uses only one and [3] uses two. Even though some of the experimented LLMs are from the same family, they are at different scales. Llama2 and Llama3 have different tokenizers trained on different data and post-training methods (DPO, etc.). * To address your concern fully, we adopt another SOTA LLM, phi3-17b, and conduct new probing experiments. The results show that this SOTA LLM, whether or not it has been fine-tuned, tends to maintain a more goal-oriented abstraction of the underlying world. It is totally consistent with our original findings. Please refer to the [Global Rebuttal](https://openreview.net/forum?id=lzfzjYuWgY&noteId=i9JnaUoNpz) for more details and analysis. * Consequently, we end up experimenting with 12 LLM variants (5 SFT + 4 ICL + 3 pre-trained), which, to the best of our knowledge, is the most extensive comparison$\color{red}{^3}$ so far in the area of probing world representations. ### *Minor Weakness: text in Figure 5 too small & formal definition of $Q^\$** Thank you for your suggestions! We will increase the font size in Fig.5 and quote the formal definition of $Q^$ in the revised version. They would be very easy fixes. --- Rebuttal Comment 1.1: Comment: Thank you for the rebuttal. While I appreciate the additional experiments and clarifications, I still have my original concerns about whether these conclusions can be made from just REPLACE. Therefore I will keep my original score. --- Rebuttal 2: Title: Footnotes and Reference List for rebuttal Comment: ### Footnotes* $\color{red}{^1}$ Gripper problem is a classic AI/planning task featured in the first ICAPS's competition [6]. $\color{red}{^2}$ In case you can't find it quickly, we've copied the details here. We synthesize another variant of GRIPPER, where ground-truth action sequences are sub-optimal, including some random legal moves that do not lead to the goal. We compare the average recovery rates for each abstraction from Llama3 models, fine-tuned on both the original and new datasets: \| Models \| Raw \| World \| $Q^$ \| $\pi^$\| \| --- \| --- \|--- \|--- \|--- \| \| Llama3-sft (original, reported in paper) \| 28.5 \| 30.6 \| 73.0 \| 87.8 \| \| Llama3-sft (new) \| 25.5 \|27.4 \| 63.3\| 72.1 \| The results show that the recovery rate of goal-oriented abstractions decreases, which is expected as the models imitate a sub-optimal policy. However, the margin between goal-oriented and general abstractions is still substantial. More importantly, the recovery rate of world-irrelevant abstraction doesn't increase at all, implying that training with non-optimal ground truths doesn't lead to a more general world abstraction. These new results echo our main findings, and we'll add them to the revised version. $\color{red}{^3}$ While existing work focuses on one or two predicates, our probing experiments involve ten different predicates, as our framework requires a comprehensive assessment of various types of abstractions. Also, the main results reported in the paper are averaged across multiple runs. In total, we completed \~500 runs of training and testing probes across all LLMs and predicates. ### Reference [1] Li, Kenneth, et al. "Emergent World Representations: Exploring a Sequence Model Trained on a Synthetic Task." ICLR, 2023. [2] Hazineh, Dean S., Zechen Zhang, and Jeffery Chiu. "Linear Latent World Models in Simple Transformers: A Case Study on Othello-GPT." arXiv preprint arXiv:2310.07582 (2023). [3] Li, Belinda Z., Maxwell Nye, and Jacob Andreas. "Implicit Representations of Meaning in Neural Language Models." ACL, 2021. [4] Kim, Najoung, and Sebastian Schuster. "Entity Tracking in Language Models." ACL, 2023. [5] Prakash, Nikhil, et al. "Fine-Tuning Enhances Existing Mechanisms: A Case Study on Entity Tracking." ICLR, 2024. [6] McDermott, D.M., 2000. The 1998 AI planning systems competition. AI magazine, 21(2), pp.35-35. --- Rebuttal 3: Title: Thank you for your review; we're ready to answer any further questions Comment: Dear reviewer, Thank you again for your review! We really appreciate your recognition of our methodologies, new resources, and presentation. We hope our rebuttal clearly explains why our new task representative and the existing testbeds are not ideal for faithful probing. Along with the two ablation studies and additional experiments on new families of LLMs, we hope that your suggestion for more extensive experiments has been adequately addressed. If there are any specific setups you believe might challenge our claims/findings that we haven't covered, please let us know. We'd be glad to accommodate them if feasible. We're very looking forward to an engaged and productive discussion! --- Rebuttal 4: Comment: Thank you for your feedback. We regret that your concern about experiments on other tasks remains, although we believe we have thoroughly addressed the need for more extensive model comparisons in our rebuttal. If possible, we would like to respectfully reiterate several points for your kind consideration. (1) Existing testbeds are NOT available for feasible and faithful probing; (2) Due to (1), we develop a new one, creating TWO datasets in one effort; (3) We have conducted two ablations with different task variants, fully addressing your request for "more settings for ablations"; (4) We aren't aware of any work analyzing LLMs/neural models that proposes two or more tasks simultaneously.	Summary: This paper investigates whether different levels of world abstractions can be decoded from LLM representations. The study is performed using a synthetic dataset of simple planning problems involving moving objects between containers. The study takes inspiration from RL to define different levels of abstraction, from specific goal-oriented states to general world-related states. The LLMs (Llama and Mistral models) are initially adapted to the task through in-context learning and fine-tuning. The problem is encoded in natural language and the hidden states are extracted and used to train probing classifiers to determine whether each level of abstraction can be recovered. The results show that the recovery of goal-related states is much more successful than general world abstractions. Strengths: 1. The presentation is rigorous and comprehensive. 2. The synthetic dataset is very good - simple but to the point. Upon release, it could be helpful for various other research papers. 3. The experiment performed using the dataset is clear and well-designed. 4. The question of whether LLMs build state abstractions is very interesting and currently an open question. Progress in this area would lead to more interpretable models. Weaknesses: Major comments: 1. Given that the models are first fine-tuned to the task, I wonder whether the results are not a self-fulfilling prophecy of the setup. If full problems and solutions were presented during adaptation, then it seems expected that the model would produce representations that are relevant to this task - the paper result would be, in a sense, a measure of the LLM having adapted to the fine-tuning, as opposed to reflecting the general inner workings of an LLM faced with a realistic or novel problem. A more convincing experiment would involve no task adaptation. 2. It is not clear how the hidden states were selected for probing. Only the "last hidden states" were selected (section 3.2). This again seems to invite the paper's results - it seems like economically preserving only relevant representations in the last layers is a measure of the LLM adaptation to the task. How about early layers of the LLM - can general abstractions be recovered there? Additionally, can any abstractions be recovered at all before adaptation? Minor comments: 3. Some of the figures lack clarity. Figures 5-6 have unreadably small fonts, tables have barely readable fonts. "Best viewed in color" is a useful warning, but better colour selection would be ideal. 4. Typos: L326 "the most coarsest". Additionally, expressions like L318 "LLMs prefer maintaining a goal-oriented world abstraction over a more general one" are excessive LLM personification. Technical Quality: 3 Clarity: 3 Questions for Authors: See above. Confidence: 3 Soundness: 3 Presentation: 3 Contribution: 2 Limitations: Adequately discussed, but only in the Appendix, not in the main text. Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thank you for your thoughtful and encouraging feedback. We address all of your concerns below. ### Major Weakness 1: Findings are expected > (Findings unsurprising) If full problems and solutions were presented during adaptation, then it seems expected that the model would produce representations that are relevant to this task * World dynamics are also relevant to the task as the optimal plan is derived based on the prediction of future states (illustrated in Figure 3), but they are almost absent from LLM representations. * On the other hand, LLMs-sft could have omitted $Q^$ abstraction since the $\pi^$ abstraction alone is sufficient for predicting optimal actions. However, we successfully recovered $Q^$ abstraction from LLM-sft representations. It reveals that LLMs-sft preserve the long-term impact of each possible action, despite being fine-tuned on next-token prediction only. To the best of our knowledge, this exciting finding is novel and can inspire future research to conduct more in-depth analysis. In the [Global Rebuttal](https://openreview.net/forum?id=lzfzjYuWgY&noteId=i9JnaUoNpz), we further elaborate on the broader significance of our framework and explain why our findings are surprising. We sincerely invite you to review it. > (Experiments with pre-trained LLMs) A more convincing experiment would involve no task adaptation. To completely address your concern, we conduct new probing experiments on top of Phi3-17b, a brand-new SOTA LLM that achieves 19.94% success rate on GRIPPER without any fine-tuning. From its representations, goal-oriented abstractions can be probed with a much higher recovery rate than the world-irrelevant abstraction. We further include two smaller and weaker pre-trained LMs for comparison. The results suggest that more advanced pre-training leads to a higher tendency to prioritize goal-oriented abstractions over a more general one. Please refer to the [Global Rebuttal](https://openreview.net/forum?id=lzfzjYuWgY&noteId=i9JnaUoNpz) for detailed results and in-depth analysis. ### Major Weakness 2: Clarity issue > not clear how the hidden states were selected … How about early layers of the LLM… * To clarify, the last hidden states refers to the hidden states at the last step before decoding, following the common practice of existing work. We treat the layer index as a hyperparameter (detailed in L620-L622) and use the last 6th layer for all experiments. * We have also experimented with using early layers. For instance, the average recovery rate (RR) of different abstractions from the 4th layer of Llama2-13b-sft is as follows: \| Layer \| Raw \| World \| $Q^$ \| $\pi^$\| \| --- \| --- \|--- \|--- \|--- \| \| 4th \| 13.7 \| 13.0 \| 14.7 \| 7.02 \| \| Last 6th \| 28.2 \|30.2 \| 71.1\| 85.6 \| The RR for both the general- and goal-oriented abstractions significantly decreases. Therefore, the world abstraction cannot be recovered using earlier layers. ### Minor Weaknesses * small fonts in the table and better color selection: Thank you for your feedback! We will enlarge the font in the figure and use deepened color in the revised version. * Typos and expression: Thank you for your careful reading! We will fix the typos and grammar in the revised version. They would be very easy fixes. --- Rebuttal 2: Title: Thank you for your review; we're ready to answer any further questions Comment: Dear reviewer, Thank you once again for your review! We are truly grateful for your appreciation. We hope our rebuttal clearly explains why our findings are surprising and important, and that our additional experiments with pre-trained LLMs, along with our clarification about the selection of hidden states, address all of your concerns. Please don't hesitate to ask any further questions. We are always ready to respond, and we're very looking forward to an engaged and productive discussion! --- Rebuttal 3: Comment: Dear Reviewer, We appreciate the effort you've put into reviewing our paper and offering really helpful feedback. We're wondering if you've had a chance to go through our responses. As the discussion period is nearing its close, we want to ensure that any remaining questions or concerns are fully addressed. If you feel that the main concerns (suprisal of findings & the selection of hidden states) have been satisfactorily resolved, could you please consider increasing the score to reflect that? We would be truly grateful.	Summary: This work makes the observation that prior work in probing LLM for planning tasks comes to different conclusions on whether there are internal state abstractions in LLM hidden layers. This work hypothesizes that the disagreement comes from that these works are probing LLM with different tasks, which may necessitate LLM to learn abstraction at different granularities best suited for next-token prediction. Hence probing without controlling the abstraction level might give rise to different recovery rates of abstractions from hidden layers for different tasks. To support this claim, the paper borrows from RL literature the concept of world-/Q-/$\pi^$-irrelevant abstractions to designate different abstraction levels and aims to show whether they can recover these abstractions from hidden layers of Llama and Mistral models (both via in-context prompting and LoRA fine-tuning) Strengths: 1. The paper is really well-motivated and written. The problem is a varying interesting one and the observation is timely for the field. 2. It's quite novel to break down abstractions through the lens of q value function and policy to capture aspects that are relevant for planning. 3. The design of the abstraction together with the RePlace experiment is rigorous and well-thought-out. 4. Figures are super helpful to ground analysis in concrete examples. Weaknesses: The biggest weakness I find is that the experiment may not be sufficient to support their central hypothesis. While it is necessary to show high recovery rate of Q-/$\pi^$-irrelevant abstractions for RePlace tasks, it is not sufficient to substantiate the claim that LLM will learn different abstractions given different tasks. RePlace tasks (both Gripper and Cook) is just one category of tasks whose solution requires forming abstractions such as Q value. Necessary additional experiments should show figure 6 on other tasks where 1. world-irrelevant features are important and 2. either Q abstractions or $\pi$ abstractions are important but not both to solve the task. If figure 6 shows different recovery rates for these other categories of tasks, there is sufficient evidence to support your claim. Technical Quality: 4 Clarity: 4 Questions for Authors: A lot of the authors' findings (e.g. finding 2 "Supervised fine-tuning mainly enhances a goal-oriented world abstraction") are not necessarily true because they only fine-tune LLM on RePlace tasks that are inherently goal-oriented. Finetuning with other categories of tasks that are not goal-oriented may not lead to goal-oriented world abstractions. For example, if you modify your RePlace domains to generate a dataset of non-optimal plans (potential back-and-forth legal moves but no shortest paths to goals), fine-tuning with this dataset might not lead to purely goal-oriented abstractions. Confidence: 4 Soundness: 4 Presentation: 4 Contribution: 3 Limitations: The authors make the remark that "one cannot determine whether the success of probing stems from the LLM’s preference for learning a general world model or from the necessity to recover the world state while learning the optimal policy" in the critique of prior work using the Othello game. However, the proposed probing method might not prevent supervised fine-tuning on a particular task from washing out LLM's innate preference either. In other words, the probing results only reflect abstractions suited to solve the tasks they fine-tune on. Perhaps one can only use ICL results to probe LLM's innate preference for abstraction learning, and to that end, one might need to come up with tasks that LLM are better at solving zero-shot. Flag For Ethics Review: ['No ethics review needed.'] Rating: 7 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thank you for your time and review. We appreciate the opportunity to clarify two critical aspects that may have been misinterpreted: our claims/hypothesis and the concept of goal-oriented abstractions. ### Weakness > This work hypothesizes that the disagreement comes from that these works are probing LLM with different tasks, which may necessitate LLM to learn abstraction at different granularities best suited for next-token prediction (from summary) … the claim that LLM will learn different abstractions given different tasks (from weakness) … Indeed, these are NOT our hypotheses and claims, and NONE of our contributions (listed in L74-83) & findings (Sec 6.2) is about LLMs learning different abstractions for different tasks. We clarify our main claims & proofs below. * The disagreement among existing works is likely due to (1) they probe only the raw world state on different tasks (L30-31), and (2) the raw world state might equal to or overlap with varying levels of abstraction depending on the task (L34-L38). It's, therefore, plausible that LLMs consistently preserve certain types of abstractions, but previous studies fail to capture this pattern. * More critically, probing WITHOUT the notion of world abstractions leads to biased evaluations, which may underestimate LLMs' world modeling when there is no overlap between the raw state and any abstraction or overestimate when the raw state aligns with goal-oriented abstractions (illustrated in L49-54, L164-170). Our framework directly addresses this limitation by probing different levels of abstractions$\color{red}{^1}$. * Within our framework, we primarily investigate which types of world abstractions are encoded in LLM representations (mentioned throughout the paper, L71-73, 77-79, 145, 243-245). Sec 6.2 lists all findings. Moreover, the results prove that probing only the raw state causes biased evaluations and unnecessary disagreement (emphasized in L374-379, also elaborated in the Global Rebuttal). Due to the lack of faithful testbeds (explained in L162-170)$\color{red}{^2}$, we CANNOT further test if LLMs preserve different types of abstractions for other tasks$\color{red}{^3}$ . > RePlace tasks … is just one category of tasks whose solution requires forming abstractions such as Q value. Necessary additional experiments should show figure 6 on other tasks where 1. world-irrelevant features are important and 2. either Q abstractions or $\pi$ abstractions are important but not both to solve the task. This question doesn't hold up because (1) The world-irrelevant abstraction is indeed important in RePlace as it's necessary for predicting future states, from which the optimal plan is derived (illustrated in Fig.3). (2) A goal-oriented abstraction unimportant for a task doesn't exist$\color{red}{^4}$. State abstractions are task-specific and derived from the task's characteristics (i.e., transition dynamics and reward function). A state abstraction is a function that maps raw states to a smaller space, preserves the interested statistics (such as $Q^$ and $\pi^$), and omits other irrelevant information. Therefore, if a predicate is not important (no impact on the $Q^$ or $\pi^$) for a task, it should be excluded from the goal-oriented abstractions. Please refer to Sec 3.1 for formal definition and explanation of state abstractions as well as pointers to relevant literature. ### Questions > Finetuning with other categories of tasks that are not goal-oriented may not lead to goal-oriented world abstractions. To clarify, we interpret "tasks that are not goal-oriented" as fine-tuning with sub-optimal ground truths, given that any RL task has a goal that maximizes the accumulated reward. We address this concern below. > For example, if you modify your RePlace domains to generate a dataset of non-optimal plans (potential back-and-forth legal moves but no shortest paths to goals), fine-tuning with this dataset might not lead to purely goal-oriented abstractions. This is an interesting question! To answer it, we synthesize another variant of GRIPPER, where ground-truth action sequences are sub-optimal (as you suggested, with some random legal moves that do not lead to the goal). We compare the average recovery rates (RR) for each abstraction from Llama3 models, fine-tuned on both the original and new datasets: \| Dataset \| Raw \| World \| $Q^$ \| $\pi^$\| \| --- \| --- \|--- \|--- \|--- \| \| original (reported in paper) \| 28.5 \| 30.6 \| 73.0 \| 87.8 \| \| new \| 25.5 \|27.4 \| 63.3\| 72.1 \| On the new dataset, the RR of goal-oriented abstractions decreases, which is expected as the model imitates sub-optimal actions. However, the margin between goal-oriented and general abstractions is still substantial. More importantly, the RR of world-irrelevant abstraction doesn't increase at all, which proves that training with non-optimal ground truths doesn't lead to a more general world abstraction. These new results echo our main findings, and we'll add them to the revised version. ### Limitations > However, the proposed probing method might not prevent supervised fine-tuning on a particular task from washing out LLM's innate preference either… It's completely flexible for our framework to verify this assumption by directly probing pre-trained LLMs. To demonstrate that, we conduct the same probing experiments with three different pre-trained LMs. The results suggest that more advanced pre-training increases the tendency to prioritize goal-oriented abstractions over more general ones, and that SFT further enhances this trend. Please refer to the [Global Rebuttal](https://openreview.net/forum?id=lzfzjYuWgY&noteId=i9JnaUoNpz) for detailed results and in-depth analysis. Moreover, we thoroughly discuss all aspects of our work's limitations in Appendix A, which includes the scope of our experimental findings. --- Rebuttal 2: Title: Footnotes for rebuttal Comment: $\color{red}{^1}$ To clarify, we don't need to use different tasks to probe different levels of abstraction. Instead, we derive the abstraction functions, such as world-irrelevant abstraction and $Q^$-irrelevant abstraction, for the same task and then probe the abstract state at various levels from LLM representations. All we need is to carefully select a task where the spaces of abstract states at different levels differ (highlighted in L162-166). This generic framework is introduced in Section 3, and we demonstrate how to apply it to our new task, RePlace, in Section 5. $\color{red}{^2}$ In particular, there are three necessary criteria for faithful and feasible probing (explained in L179-184, L488-493): (1) processing text data; (2) the abstractions across various levels are distinguishable; (3) the world state can be easily manipulated and tracked. $\color{red}{^3}$ However, our findings can explain and resolve conflicts among existing works. Specifically, the contradiction may stem from various degrees of alignment between the raw states and goal-oriented abstractions across tasks. For instance, [1] successfully probes the Othello board state, which is already the goal-oriented abstraction for predicting legal moves (the task Transformers are trained on), whereas [2] struggles to probe the status of entities, as they have little overlap with goal-oriented abstractions of the task that LLMs are fine-tuned on. $\color{red}{^4}$ As explained earlier, however, it's entirely possible for a predicate to pertain to the goal-oriented abstractions of Task A yet remain unrelated to Task B (i.e., Task A and B share the same state space but have very different reward functions). Our ablation study confirms that under this scenario, this predicate can be probed from LLMs adapted to Task A but not Task B. Concretely, we have created a new variant of GRIPPER where an originally goal-unrelated predicate (e.g., information of objects' color) pertains to $Q^$ abstraction. As such, the recovery rate of color information has significantly increased from zero. This further consolidates our conclusion that SFT mainly enhances goal-oriented abstractions. Due to the limited space, we summarize this finding in the main body of the paper (L337-L340) and provide details in Appendix K. --- Rebuttal 3: Title: Thank you for your review; we're ready to answer any further questions Comment: Dear reviewer, Thank you once again for your review! We really appreciate your recognition of all aspects of our work, including our addressed problem, framework, presentation, and the design of our task and experiments. We believe the major concern might be due to some misunderstandings, and hopefully, our rebuttal has already cleared it up. Meanwhile, we hope that our additional experiments in response to your concerns and suggestions will make you more convinced about the soundness of our findings. Please don't hesitate to let us know if you have any concerns that are unaddressed or if we need to clarify something further. We're always on standby to respond, and we're very looking forward to an engaged and productive discussion! --- Rebuttal Comment 3.1: Title: Response to Rebuttal Comment: Thanks for the really well-done rebuttal! Your response clarified a major misunderstanding on my end and thus cleared many of my resulting concerns about your experiments and results. I especially appreciate the added experiment on SFT with sub-optional data following my question. Therefore, I am happy to raise my scores. --- Reply to Comment 3.1.1: Comment: Dear Reviewer, We're truly heartened and thrilled to know that our responses are able to address your concerns. This is incredibly important to us! Your interesting question has inspired us to uncover more insights. We will definitely incorporate them in the revised version.	Rebuttal 1: Rebuttal: We thank all reviewers for their time and feedback. We are encouraged that all aspects of our work are widely recognized. The reviewers found our addressed problem to be new and important (R1, R2, R3), our framework novel and neat (R2, R4, R5), our task/datasets thoughtfully designed and thus useful for future research (R1, R2, R3, R4, R5), experiments well-designed (R3), findings sound (R1), and our paper well-written (R2, R3, R4, R5). Most of the concerns regard two interrelated aspects, namely the findings' surprisal (R1, R3, R5) and the lack of probing experiments on pre-trained LLMs without ICL/SFT (all reviewers$\color{red}{^1}$). We address both of them below. ### 1. Surprisal of findings One primary concern is that the finding, SFT mainly enhances goal-oriented abstraction, is not entirely surprising$\color{red}{^2}$. Below, we elaborate on this and other surprising findings, as well as the broader significance of our framework. * This finding is surprising when viewed in the context of existing work. While not entirely unexpected in hindsight, it undermines an increasingly widespread belief that an implicit world model can emerge from next-token prediction (NTP) training [1]. This belief is encouraged by recent studies demonstrating that world states can be recovered from Transformer representations [2,3]. When we use a testbed that can differentiate abstraction at various levels, however, we find that LLMs tend to discard the transition dynamics, essential ingredient of world models, when possible. This finding also explains and hence resolves the conflicting conclusions in existing works. For instance, [2,4] successfully probes the Othello board state, which is already the goal-oriented abstraction for predicting legal moves (the task Transformers are trained on), whereas [5] struggles to probe the entity status, which falls outside the goal-oriented abstractions for which LLMs are optimized. * It’s more surprising how probing different sets of predicates WITHOUT the notion of world abstractions can lead to directly opposing conclusions. Imagine two researchers independently assessing an LLM's world representation on RePlace, using current methodologies. As demonstrated in our experiments (Sec 6.2, illustrated in L374-379), Researcher A selecting `agentLoc`, `nearby`, and `held_u`, might infer that SFT enforces LLMs to develop an implicit world model, while Researcher B, focusing on `store` and `boxName`, comes to a totally different conclusion, arguing that LLMs-sft have little awareness of the underlying world. This disparity reflects the current state of the field$\color{red}{^3}$. Thus, our framework provides a more grounded basis for future research, avoiding blind spot conflict and enabling others to dissect and critically evaluate claims in a principled way. * It's noteworthy that *LLM-sft representations encode the $Q^\$ abstraction, which could have been abstracted away** as $\pi^\$ abstraction is sufficient for deriving optimal actions. It reveals that LLMs-sft preserve the long-term impact of actions even when fine-tuned with NTP only. This finding is novel in this area, and it leaves many interesting questions open for future work: How does this abstraction emerge from SFT? Could different training methods enhance or diminish it? Does it affect only the immediate prediction, or do LLMs continuously revisit previous abstractions during decoding? ### 2. Probing experiments on pre-trained LMs* Similarly, another concern is that the pre-trained LLMs, especially those that can perform the task well without SFT, may tend to maintain general abstractions, but SFT fundamentally alters this pattern. To best address it, we employ Phi3-17b [6], a brand-new open LLM that hits a 19.94% success rate on GRIPPER without SFT. To further understand the impact of pre-training on the encoding of world abstractions, we employ two weaker and smaller LLMs, Phi3-3.8b and Pythia-70m, which reach near-random success rates$\color{red}{^4}$, for comparative analysis. We additionally fine-tune another Phi3-17b. Plot A in the attached pdf reports the average recovery rate ($\color{red}{\text{RR}}$) of predicates within each type of world abstractions from different LLMs, and Plot B reports the RR of each predicate across all LLMs. We highlight three key findings below. 1) Pre-trained Phi3-17b prioritizes maintaining goal-oriented abstractions over a more general one. Plot A clearly shows that. And SFT substantially strengthens this tendency. 2) As LLMs increase in scale and capability (Pythia<Phi3-3.8b<Phi3-17b), they are more likely to maintain goal-oriented abstractions over a more general one, thereby widening the gap between their RR. This is apparent from Plot B, where the RR variance among different LLMs is mainly found in predicates that pertain to $\pi^$- and $Q^$-irrelevant abstractions. In particular, the predicates that uniquely pertain to goal-oriented abstractions, such as `nearby` and `nextObjDirect`, are probed with remarkably higher recovery rates from Phi3-17b than from Pythia. In contrast, the RR of `store` and `held_g`, essential for world-irrelevant abstraction, are almost identical across all LLMs variants. It suggests that more advanced pre-training doesn't necessarily enhance the encoding of world dynamics. 3) Advanced pre-training is NOT sufficient for efficient world modeling. Interestingly, probing from Phi3-17b has much higher RR of `boxName`, which is neither relevant to task completion nor to the transition dynamics. This verifies our Finding 4 (L366-371) that LLMs are limited in building world representations. We thank the reviewers for their suggestion of adding this experiment. It turns out to be a good chance to gain new insights about SOTA LLM's world representation which supplement and consolidate our original findings. We will include these results in the revised version, which would be an easy fix. Pdf: /pdf/4a69ed23bd70d4c3c3e58062808975b49f1ae6ee.pdf	NeurIPS_2024_submissions_huggingface	2,024	Summary: In this work, the authors attempt to examine whether large language models (LLMs) possess representations that can work as the world model. To this end, they target different state abstraction levels, world-irrelevant abstraction, $Q^$-irrelevant abstraction, and $pi^$-irrelevant abstraction, which are based on state abstraction in reinforcement learning. Their objective is to figure out how much of each state abstraction can be recovered from the last hidden representations of LLMs, and the authors propose a text planning task named RePlace. Given the objective, RePlace is designed to have distinct state abstractions for the three abstraction levels as well as the raw state. The authors empirically confirm that both pre-trained LLMs and LLMs fine-tuned on this task fail to recover the world-irrelevant abstraction, whereas the $Q^$-irrelevant and $pi^$-irrelevant abstractions exhibit higher recovery rates. Post-Rebuttal/Discussion: I find the author response fair and am raising my score to 7. Strengths: - Given the extensive use of transformer models these days, establishing sound findings in this regard (whether LLMs can provide representations useful for recovering world dynamics information) can be an important problem. - The proposed task, RePlace, is carefully designed with different levels of state abstractions in mind. It allows the authors to conduct the empirical analysis that shows the differences in the recovery rates for different state abstractions. Weaknesses: - The finding that the fine-tuned language models do not preserve the features needed for recovering the world dynamics may not be entirely surprising, as fine-tuning enforces the bias needed for the specific problem to the models and much of irrelevant aspects often end up being ignored. - In addition to my first point, I believe more interesting findings can be derived with pre-trained LLMs, as there is a possibility that the pre-training provides enough signals to the LLMs to catch the world dynamics information. While the authors do analyze pre-trained LLMs with in-context learning in the proposed benchmark, according to Table 1, their performance is quite low in the first place, which makes it difficult to examine this hypothesis. Technical Quality: 3 Clarity: 3 Questions for Authors: Please take a look at the Weaknesses section above. Confidence: 4 Soundness: 3 Presentation: 3 Contribution: 2 Limitations: The authors provide a fair list of limitations in Appendix A. Flag For Ethics Review: ['No ethics review needed.'] Rating: 7 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thank you for your detailed and positive feedback. We address all of your concerns below. ### Weakness 1: finding not entirely surprising >(Unsurprising findings) The finding that the fine-tuned language models do not preserve the features needed for recovering the world dynamics may not be entirely surprising… * It's indeed surprising when viewed in the context of existing work. We provide detailed elaboration in the [Global Rebuttal](https://openreview.net/forum?id=lzfzjYuWgY&noteId=i9JnaUoNpz), with other surprising findings and broader significance of our framework. We sincerely invite you to review it. > (It's unsurprising because) fine-tuning enforces the bias … and much of irrelevant aspects often end up being ignored * In fact, the world dynamics is useful for the task, albeit in a less goal-oriented way. As illustrated in Figure 3, the optimal plan is based on predicting future states. Generally speaking, planning is searching for the shortest action sequences induced by which the predicted future state matches the goal. * Nevertheless, *Although $\pi^\$ abstraction is sufficient for predicting the optimal actions, we still have probed $Q^\$ abstraction with a high recovery rate* (Figure 5&6, discussed in L324-330). It reveals that LLMs-sft preserves the impact of each possible action, despite being fine-tuned only on next-token prediction. To the best of our knowledge, this interesting finding is novel and can inspire future research to conduct more in-depth analysis. ### Weakness 2: Probing with pre-trained LLMs > (Experiments with pre-trained LLMs) more interesting findings can be derived with pre-trained LLMs … (but) their performance is quite low in the first place, which makes it difficult to examine this hypothesis. * You are correct that we didn't conduct probing experiments using pre-trained LLMs, given their near-random performance without fine-tuning. Also, it's a common practice [1, 2, 3, 4] that probing from Transformers/LLMs that are trained or fine-tuned on next-token prediction to see if implicit world models emerge, as this is a common interest in the field. * To completely address your concern, we conduct new probing experiments on top of Phi3-17b, a brand-new SOTA LLM that achieves 19.94% success rate on GRIPPER without any fine-tuning. From its representations, goal-oriented abstractions can be probed with a significantly higher recovery rate than the world-irrelevant abstraction. We further include two smaller and weaker pre-trained LMs for comparison. The results suggest that more powerful pre-training leads to a higher tendency to prioritize goal-oriented abstractions over a more general one. Please refer to the [Global Rebuttal](https://openreview.net/forum?id=lzfzjYuWgY&noteId=i9JnaUoNpz) for detailed results and in-depth analysis. [1] Li, Kenneth, et al. "Emergent World Representations: Exploring a Sequence Model Trained on a Synthetic Task." ICLR, 2023. [2] Li, Belinda Z., Maxwell Nye, and Jacob Andreas. "Implicit Representations of Meaning in Neural Language Models." ACL, 2021. [3] Hazineh, Dean S., Zechen Zhang, and Jeffery Chiu. "Linear Latent World Models in Simple Transformers: A Case Study on Othello-GPT." arXiv preprint arXiv:2310.07582 (2023). [4] Kim, Najoung, and Sebastian Schuster. "Entity Tracking in Language Models." ACL, 2023. --- Rebuttal 2: Title: Thank you for your review; we're ready to answer any further questions Comment: Dear reviewer, Thank you once again for your review! We are truly grateful for your appreciation. We hope our rebuttal clearly explains why our findings are surprising and important and that our additional experiments with pre-trained LLMs yield some other surprising and interesting findings and address all of your concerns. Please don't hesitate to ask any further questions. We are always ready to respond, and we're very looking forward to an engaged and productive discussion! --- Rebuttal 3: Comment: Dear Reviewer, We appreciate the effort you've put into reviewing our paper and offering really helpful feedback. We're wondering if you've had a chance to go through our responses. As the discussion period is nearing its close, we want to ensure that any remaining questions or concerns are fully addressed. If you feel that the main concerns (surprisal of findings and probing with pre-trained LLMs) have been satisfactorily resolved, could you please consider increasing the score to reflect that? We would be truly grateful.	null	null	null	null	null	null
Unified Domain Generalization and Adaptation for Multi-View 3D Object Detection	Accept (poster)	Summary: The paper proposes an Unified Domain Generalization and Adaptation(UDGA) scheme for multi-view 3D object detection. The main components of the proposed method are (1) depth inconsistency-based constraints from multi-view and (2) an efficient domain adaptation scheme (LEDA). The multi-view depth inconsistency constraints improve the domain gap caused by geometric gaps when it comes to perspective view changes. LEDA offers adaptation with fewer amounts of labels while preserving the knowledge from the source domain. Strengths: The paper clearly addresses which problem it focuses on, and each proposed module has solid goals to address existing limitations. The experimental results from multiple datasets also seem convincing. Weaknesses: I have a few concerns about the demonstration of the proposed method as well as addressing existing works. (1) The method part needs to be improved. For instance, specifying every input and output tensor dimension (i.e., depth estimation network, etc.) would significantly help readers understand what each module is doing more quickly. (2) Using the depth inconsistency seems similar to DETR3D's[1] inconsistency constraint on RGB features. In DETR3D, it is even mentioned that using the RGB feature is better than having explicit depth estimation. Based on this, the depth estimation network can introduce additional parameters for optimization while not being very helpful. I cannot see experiments or comparisons with DETR3D. (3) More related works need to be addressed. Currently, the paper focuses on multi-view-based domain generalization for 3D detection. However, there is also another active line of research with LiDAR-based domain adaptation for 3D detection, pioneered by ST3D[2]. Clearly addressing the difference between those researches will make the reader recognize the line of research that the paper focuses on. (4) The effectiveness of LEDA is not well demonstrated, as claimed in the paper, compared to existing approaches. [1] Detr3d: 3d object detection from multi-view images via 3d-to-2d queries, Wang, Yue, et al. CORL, 2022 [2] St3d: Self-training for unsupervised domain adaptation on 3d object detection, Yang, Jihan. et al., CVPR 2021. Technical Quality: 3 Clarity: 2 Questions for Authors: I have a few questions that correspond to the weakness that I mentioned. (1) Is there a justifiable reason why using Depth inconsistency constraint is superior to using deep RGB feature-based inconsistency ( as in DETR3D)? (2) As one of the main proposed contributions claimed, a detailed analysis of LEDA seems missing. How much does LEDA improve the proposed system compared to existing approaches in terms of efficiency and accuracy? Addressing those two questions properly will improve the quality of the paper and, subsequently, the reviewer's rating. Confidence: 4 Soundness: 3 Presentation: 2 Contribution: 2 Limitations: The limitations are addressed Flag For Ethics Review: ['No ethics review needed.'] Rating: 4 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: ### (1) Discussion with RGB feature-based methods Thank you for your constructive suggestion. As reported by DETR3D [4], CVT [5] and BEVFormer [6], RGB feature-based methods are considerably robust against calibration noise. However, these methodologies experience significant performance degradation in dynamic view shift environments, such as cross-domain scenarios where calibration changes dramatically, as shown in table below (Lyft $ \rightarrow $ nuScenes). \| Method \| Backbone \| Neck \| Source(NDS/mAP) \| Target(NDS/mAP) \| # Params(M) \| \|-----------\|-----------\|----------------\|:-----------------:\|:-----------------:\|:------------:\| \| BEVDepth \| ResNet50 \|LSS \| 0.684 / 0.602 \| 0.213 / 0.102 \| 51.7 \| \| CVT \| ResNet50 \|Transformer \| 0.658 / 0.563 \| 0.231 / 0.066 \| 51.2 \| \| DETR3D \| ResNet50 \|Transformer \| 0.650 / 0.552 \| 0.179 / 0.087 \| 32.3 \| \| BEVFormer \| ResNet50 \|ST Transformer \| 0.624 / 0.506 \| 0.138 / 0.008 \| 33.5 \| \| Ours \| ResNet50 \|LSS \| 0.702 / 0.630 \| 0.421 / 0.281 \| 51.7 \| Additionally, we explore how domain shift hinders stable 3D recognition in main text Sec. 3.2, Appendix C and Table 7. Although, fair comparisons are challenging due to architectural differences, we validate that our proposed modules quantitatively mitigate dramatic view changes in the cross-domain (up to 20.9% NDS gain). ### (2) In-depth explanation of LEDA We hope that additional explanation of LEDA in global response will help you understand better. To improve the development of 3D object detection, we design Label-Efficient Domain Adaptation (LEDA) framework that successfully adapts to novel targets by only a few data and extra parameters, introducing a practical solution for real-world Autonomous Driving. To this end, we conduct structural experiments (Rebuttal PDF Table 2 and 3) to explore the optimal architecture that is both effective and efficient. Also, we validate that our depth constraint method encourages transfer learning from pre-trained source potentials to novel targets as shown in Rebuttal PDF Table 1. Especially, by comparing our methodology with existing state-of-the-art models and evaluating NDS by efficiency across various setups (Rebuttal PDF Figure 1), we quantitatively showcase that our proposed methods successfully bridge the gap between the source and novel target using a limited amount of training budget (i.e., training parameters and data). It is noteworthy that our proposed methodology outperforms existing works by 1% on novel targets with only less than 20% extra parameters. Finally, we will obviously present and analyze in the revised version. ### (3) Status of LiDAR-based cross-domain object detection LiDAR object detection research includes addressing performance degradation caused by domain shifts. Wang et al.[1] proposed statistical normalization to mitigate the differences in object size distribution across various datasets. ST3D[2] leveraged domain knowledge through random object scale augmentation, and their self-training pipeline refined the pseudo-labels. STAL3D[3] extended ST3D by incorporating adversarial learning. These methods were all proposed under the assumption of target-aware conditions, which presents the limitation that they cannot be applied without prior access to the target domain. The table (Waymo $\rightarrow$ KITTI) below extracts a subset of experimental results from STAL3D. LiDAR-based methods were less affected by domain shifts compared to those of cameras, though their performance did not reach the oracle level. \| Method \| BEV AP \| 3D AP \| \|-----------------\|--------\|-------\| \| _Oracle_ \| 83.29 \| 73.45 \| \| _Direct Transfer_ \| 67.64 \| 27.48 \| \| SN[1] \| 78.96 \| 59.20 \| \| ST3D[2] \| 82.19 \| 61.83 \| \| STAL3D[3] \| 82.26 \| 69.78 \| Unlike LiDAR modalities, these approaches fail to address the domain shift problem due to the poor quality of pseudo-labels resulting from the inaccurate geometric information in image-based 3D detection. To address the domain generalization for camera-based systems, we propose a depth constraint module and introduce the LEDA module to achieve oracle-level performance with label efficiency in the target domain. Plus, we will clearly address the weaknesses (1) More details of our methods and (3) More related works in the revised version. Thank you. ------ [1] Wang, Chen, et al. Train in Germany, Test in The USA: Making 3D Object Detectors Generalize, Computer Vision and Pattern Recognition, 2020. [2] Yang, Shi, et al. ST3D: Self-training for Unsupervised Domain Adaptation on 3D Object Detection, Computer Vision and Pattern Recognition, 2021. [3] Zhang, Zhou, STAL3D: Unsupervised Domain Adaptation for 3D Object Detection via Collaborating Self-Training and Adversarial Learning, IEEE TRANSACTIONS ON INTELLIGENT VEHICLES, 2024. [4] Wang, Yue, et al. "Detr3d: 3d object detection from multi-view images via 3d-to-2d queries." Conference on Robot Learning. PMLR, 2022. [5] Zhou, Brady, and Philipp Krähenbühl. "Cross-view transformers for real-time map-view semantic segmentation." Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. 2022. [6] Li, Zhiqi, et al. "Bevformer: Learning bird’s-eye-view representation from multi-camera images via spatiotemporal transformers." European conference on computer vision. Cham: Springer Nature Switzerland, 2022. --- Rebuttal 2: Comment: I appreciate the authors' effort to answer my questions. I am convinced by the authors' responses about "RGB-features" and "In-depth explanation of LEDA". However, the authors' response to "Status of LiDAR-based cross-domain object detection" made me sceptical about what the authors meant by "(1) These methods were all proposed under the assumption of target-aware conditions, which presents the limitation that they cannot be applied without prior access to the target domain." and (2) "LiDAR-based methods were less affected by domain shifts compared to those of cameras, though their performance did not reach the oracle level." in the rebuttal. Regarding (1), did the authors mean that LiDAR-based self-training methods require target statistics, such as average object size, and self-training requires access to the target domain? From the reviewer's understanding, the proposed method LEDA even requires labels directly from the target domain(although it's a comparably small amount, i.e. 1%- 5%), whereas the LiDAR-based self-training methods do not require such direct labels. In the context that LEDA requires target labels and LiDAR-based methods require self-training, it seems to me that they both require prior access to the target domain. Regarding (2), if the authors meant that LiDAR-based methods haven't met the Orcal performance in Waymo to KITTI adaptation task, I must say that this is false. For instance, DTS [1] and so on already outperformed Oracle with the justifiable reasons in Waymo to KITTI adaptation task. I understand that the multi-view-based and LiDAR-based Domain adaptation/generalization methods have different characteristics and cannot be directly comparable. However, as a similar branch of work, the LiDAR-based domain adaptive 3D detection also needs to be addressed with proper description so that readers can understand the lines of similar works. Specifically, it does not sound very convincing to advertise that multi-view-based domain generalization is more challenging than LiDAR-based domain generalization because of the above-mentioned reasons for (1) and (2). I personally think that just mentioning LiDAR-based domain adaptive/generalization 3D detection works as relevant works without comparing which one is more challenging would make it clearer for readers to know what the relevant lines of research are. [1] Density-Insensitive Unsupervised Domain Adaption on 3D Object Detection, Hu et al. CVPR 2023 --- Rebuttal Comment 2.1: Comment: Dear Reviewer 2JqY, We sincerely appreciate your valuable efforts and professional feedback, which have indeed improved the quality of the final version of our manuscript. Above, we've prepared answers to your remaining concerns regarding the clear description of the relationship between lidar- and camera-based domain adaptation literature. If it appears to be reasonable, could you please update your rating of our work, as you indicated in your initial review? As the authors-reviewers discussion period is nearing its end, we would like to politely request your final verdict on this matter. Best regards, The Authors --- Rebuttal 3: Title: Revision of the Status of LiDAR-based Cross-Domain Object Detection Comment: We genuinely appreciate the reviewer's thoughtful and detailed feedback. Our proposed generalization technique, Multi-view Overlap Depth Constraint, does not require prior access to the target, unlike Statistical Normalization (SN) [1]. However, due to its limited performance, we designed LEDA to efficiently learn novel target knowledge, and we agree with your observation that LEDA indeed falls under the category of methods requiring direct labels from target domain. Furthermore, we acknowledge that our lack of insight into UDA for LiDAR 3D Object Detection led to an insufficient comparison in the answer (3) 'Status of LiDAR-based Cross-Domain Object Detection'. To address these issues, we will revise rebuttal phrases (1) and (2) as follows: Regarding (1): “These methods were all proposed under the assumption of target-aware conditions, which presents the limitation that they cannot be applied without prior access to the target domain.” $\rightarrow$ “These methods adopt 'prior access to target' approaches (e.g., SN [1], self-training [2, 3]) to effectively align the domain gaps between the source and target.” Regarding (2): “LiDAR-based methods were less affected by domain shifts compared to those of cameras, though their performance did not reach the oracle level.” $\rightarrow$ “~LiDAR-based methods were less affected by domain shifts compared to those of cameras, though their performance did not reach the oracle level.~” More importantly, we observe that the challenges of LiDAR-based methods are also aligned with our problems in a certain aspect (i.e., label-efficient learning). LiDAR-based approaches utilize self-training strategies with pseudo-labeling; however, our UDGA framework introduces a method for fine-tuning LEDA using a small subset of novel target labels, and applying LEDA to LiDAR-based 3D object detection represents a promising direction for future work. We will clearly mention this discussion in the final manuscript and agree that this explanation will indeed enhance the readability of our paper. Thanks for your thorough and professional comments on this comparison. ---- [1] Wang, Yan, et al. "Train in germany, test in the usa: Making 3d object detectors generalize." Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2020. [2] Yang, Jihan, et al. "St3d: Self-training for unsupervised domain adaptation on 3d object detection." Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. 2021. [3] Hu, Qianjiang, Daizong Liu, and Wei Hu. "Density-insensitive unsupervised domain adaption on 3d object detection." Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2023.	Summary: This paper is about Unified Domain Generalization and Adaptation for outdoor 3D object detection. To address the geometric misalignment between the source and target domains, Multi-view Overlap Depth Constraint that leveraging the strong association between multi-views and Label-Efficient Domain Adaptation are proposed. Experiments are done with several cross-dataset settings. Strengths: 1. The domain generalization and adaption is important for 3D object detection. 2. The writing is easy to follow. Weaknesses: 1. In the past works, the comparative methods, such as DG-BEV, and PD-BEV, reported the performance on both source and target domains. Yet it seems that the reported metrics in Table 1 do not align with the reported results in the two papers. What is the reason for such changes? 2. It is noteworthy that the NDS is specific for nuScenes and different for other datasets. It is important to discriminate such differences. 3. The authors claim that this work is toward 3D universal detection, yet there are other works toward 3D universal detection, such as [1][2]. It is better to illustrate the differences over these methods, both in discussion and experimental comparison. Besides, there are other works devoted to multi-dataset training for 3D detection for generalization, also missed in the paper. Can the proposed method be applied in multi-source training and multi-target testing? 4. The Domain Adapter design seems simple, yet the motivation and potential of such design is missed. It is important since it is one of the two components of this work. [1] Towards universal LiDAR-based 3D object detection by multi-domain knowledge transfer. ICCV 2023. [2] Cross-Dataset Sensor Alignment: Making Visual 3D Object Detector Generalizable. PMLR 2023. Technical Quality: 2 Clarity: 3 Questions for Authors: The experimental comparison and the novelty, importance of the work. Confidence: 4 Soundness: 2 Presentation: 3 Contribution: 2 Limitations: The limitation part has been addressed. Flag For Ethics Review: ['No ethics review needed.'] Rating: 5 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: ### (1) Explanation of Table 1 Thank you for your insightful review. We understand that there was some confusion regarding the misaligned results in Table 1 due to insufficient explanation. Since previous methods did not adopt consistent experimental methods, a fair comparison is only possible on the target domain (i.e., they did not evaluate the same model on both the source and target domains). To address this issue, we adopt their highest scores for comparison, as detailed in Appendix Table 5. We will make the necessary revisions to clarify this information and enhance the overall clarity of our manuscript. ### (2) Discussion with other metrics Note that, to ensure fair comparison, we adopt unified evaluation metric NDS as reported in DG-BEV and PD-BEV. Waymo and Lyft calculate 3D AP by directly measuring the intersection over union (IoU) between predicted and ground truth bounding boxes. In contrast, NDS measures differences between predicted and ground truth objects leveraging various indicators such as translation error, size error, orientation error, and BEV AP. This approach enables capturing plausible failure cases (e.g., translation fails or orientation fails) and allows for practical analysis of domain shift issues, as shown in main paper Sec. 3.2 L#126-130, Sec. 4.2 L#207-209, Sec. 4.3 L#244-245 and L#250-251, and Appendix C L#505-513. ### (3) Discussion with the 3D universal detection Thank you for your meaningful discussion. [1] and [2] explore multi-domain learning and generalization to improve 3D object detection. Here, we present a brief overview as follows. \| Method \| Modality \| View \| Target-aware \| Approach \| Multi-source \| Multi-target\| \|------\|----------\|:------:\|:------------:\|------------\|:------------:\|:--------:\| \| [1] \| LiDAR \| multi \| X \| out-domain \| O \| O \| \| [2] \| Camera \| single \| O \| out-domain \| O \| O \| \| Ours \| Camera \| multi \| X \| in-domain \| O (a few extra targets) \| O \| The Universal 3D Object Detection task is quite similar to our framework, and our proposed methods showcase considerable potential in this field (multi-source training and multi-target testing), as shown in main paper table 2. However, while universal training methods generalize external domain knowledge, we leverage internal knowledge, which do not rely on large-scale and diverse datasets. Furthermore, our framework can generalize without prior access to the target, making it more feasible to develop and deploy 3D detection in real-world scenarios. ### (4) More details of Label-Efficient Domain Adaptation We hope that this rebuttal will be helpful to you. We provide more details of LEDA in the global response, improving the motivation and potential of our proposed method. Although existing methods aim to generalize the domain shift between the source and the novel target, they often fail to provide a practical solution mainly due to unsatisfied generalization performance and show room for improvement. Additionally, costly resources exacerbate these issues and hinder the expansion of multi-view 3D object detection. To tackle these issues, we design a de-facto framework leveraging a efficient and effective learning strategy LEDA for multi-view 3D object detection. First, we show that our depth constraint method smoothly deals with sensor misalignment between source and target domains and effectively boost the adaptation capability of LEDA (as shown in Rebuttal PDF Table 1). Also, our LEDA successfully transfers their own potentials to novel targets without forgetting previously learned knowledge as shown in main text Table 5. As a result, we note that our proposed framework impressively releases costly resource issues and unstable performance as shown in Rebuttal PDF Figure 1 and presents practical solutions for real-world autonomous driving. We sincerely appreciate your constructive review and will address the points discussed above in the revised manuscript. Thank you. -------- [1] Wu, Guile, et al. "Towards universal LiDAR-based 3D object detection by multi-domain knowledge transfer." Proceedings of the IEEE/CVF International Conference on Computer Vision. 2023. [2] Zheng, Liangtao, et al. "Cross-Dataset Sensor Alignment: Making Visual 3D Object Detector Generalizable." Conference on Robot Learning. PMLR, 2023. --- Rebuttal 2: Comment: Dear Reviewer gahV, We greatly appreciate your valuable efforts and professional feedback, which have indeed improved the quality of the final version of our manuscript. We’ve provided answers to your remaining concerns above, and it would be great to hear your feedback on our rebuttal so that we can further improve the final version. Although the authors-reviewers discussion period is nearing its end, we are fully prepared to address any further questions you may have. Best regards, The Authors	Summary: The paper presents an adaptation of 3D object detectors to varying target environments using two major strategies. The proposed multi-view overlap depth constraint leverages associations across views to learn view-invariant features. Additionally, a LORA-like structure is designed for parameter-efficient adaptation, accommodating scenarios with limited target data. Experiments on three benchmark datasets demonstrate the effectiveness of the proposed approach, with minimal modification of parameters. Strengths: + The paper addresses a significant and practical issue in 3D object detection, aiding in the development of robust models for dynamically changing testing environments. + The flowchart clearly illustrates the core components, making it easy for readers to grasp the main idea. + Extensive experiments have been conducted, and the proposed strategy significantly outperforms the pre-trained source model and even surpasses full fine-tuning strategies. Weaknesses: - The two proposed strategies appear somewhat disconnected. While the multi-view idea is interesting, the technical details are confusing. Equation (5) includes three different objectives, but their importance or sensitivity is not discussed. The latter adaptation strategy resembles existing works and lacks specific design for the 3D detection task. This is also not correlated with the view-transform. The isolated adaptation modules seem more like a stacking of existing techniques rather than a cohesive addition. - More discussion and comparisons with previous multi-view augmentation strategies would help clarify the merits and innovations of the proposed approach. The current claim that existing strategies are poorly generalizable is somewhat unconvincing. Technical Quality: 2 Clarity: 2 Questions for Authors: Please refer to the weakness section. Confidence: 3 Soundness: 2 Presentation: 2 Contribution: 2 Limitations: The authors have not provided a discussion on the limitations of their work. Flag For Ethics Review: ['No ethics review needed.'] Rating: 5 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: ### (1) More details of LEDA. Disconnected strategies We hope that further explanations of LEDA in global response will enhance your understanding. We empirically observe that direct fine-tuning approaches (with a small fraction of data) often fail to align between source and target, mainly due to dynamic perspective shifts. To mitigate this issue, our proposed depth constraint method softly addresses perspective gaps and enhances adaptation capacity to novel targets. In Rebuttal PDF Table 1, LEDA (w/o $\mathcal{L_{ov}}$ and $\mathcal{L_{p}}$) struggles to bridge with novel target knowledge and show only 20.4% / 24.4% NDS / mAP gain. However, LEDA (w/ $ \mathcal{L_{ov}} $ and $ \mathcal{L_{p}} $) significantly transfer pre-trained knowledge to novel targets (34.2% / 42.5% NDS / mAP gain). As a result, we note that $ \mathcal{L_{ov}} $ and $ \mathcal{L_{p}} $ enable efficient learning by encouraging stable convergence even with tiny datasets. Ablation of objectives Additionally, we study the importance of our Optimization Objectives ($\mathcal{L_{ov}} $ and $ \mathcal{L_{p}} $). Although, our depth constraint method effectively addresses domain gaps, it suffer from narrow overlap regions between adjacent views. To tackle this drawback, $\mathcal{L_{p}}$ effectively boosts stable recognition during both pre-training and fine-tuning phase. Specifically, $\mathcal{L_{p}}$ yields max 1.9% / 1.9% NDS / mAP gain during pre-training and max 2.1% / 1.6% NDS / mAP gain during fine-tuning. As a result, we demonstrate that our proposed two methods synergistically enhance UDGA performance. In-depth explanation of LEDA. In this paper, we advocate that depth consistency play a pivotal role to bridge domain gaps. Especially, we show that UDGA stably releases geometric differences from perspective view changes and encourages optimal BEV representation learning as shown in Rebuttal PDF Table 3. Precisely, UDGA achieves remarkable performance improvements of up to 10.0% in NDS and 9.5% in mAP in view transformation and BEV encoder layers, demonstrating its effectiveness. ### (2) Comparison with Augmentation strategies. We also measured generalizability of previous multi-view augmentation strategies as shown in Rebuttal PDF Tab.4. We adopt conventional augmentation strategies to multi-view 3D object detection as follows: - GT sampling effectively address unbalanced labels by sampling ground truth objects in the 3D object detection. - 2D aug. directly augment multi-view inputs (i.e., image resize, crop and paste, contrast and brightness distortion). - 3D aug. globally rotate, re-scale and translate multi-view inputs and ground truths. - Extrinsic aug. also adopts global yaw rotation towards the random direction - CBGS re-balances to solve unbalanced ground truths. These methods significantly enhance geometric understanding from input noises. However, in dynamic view changes (i.e., cross-domain), they still suffer from geometric inconsistency and show poor generalization capability. Moreover, various 2D approaches do not guarantee geometric alignments between 2D images and 3D ground truths and relevant studies have not been explored well, as reported in DG-BEV and [1]. To tackle these issues, we present a novel Multi-view Overlap Depth Constraint that effectively mitigate dynamic view changes in the cross domain. We also applied top 3 methods 2D, 3D, and extrinsic augmentation to our method and achieved state-of-the-art performance on both source and target. Unfortunately, we were unable to find any additional comparable papers for generalized multi-view 3D object detection. If you could provide us with relevant papers, we would be happy to analyze them together. We sincerely appreciate your thoughtful review and will address above discussion in the revised version. Thank you. ---- [1] Zhao, Yunhan, Shu Kong, and Charless Fowlkes. "Camera pose matters: Improving depth prediction by mitigating pose distribution bias." Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2021. [2] Zheng, Liangtao, et al. "Cross-Dataset Sensor Alignment: Making Visual 3D Object Detector Generalizable." Conference on Robot Learning. PMLR, 2023. --- Rebuttal Comment 1.1: Title: Thanks for the reponse. Comment: I would like to appreciate the detailed response from the authors and most of my questions have been addressed. As I agree with other reviewers, the proposed multi-view approach does not seem specifically designed for DG problem. However, based on the impressive improvement compared with the commonly seen augmentation strategy, I think the proposed module has the merit to facilitate future work. Thus, I would like to increase my score to borderline accept. --- Reply to Comment 1.1.1: Title: Thank you for the positive feedback Comment: We are delighted that our responses to your questions have been well received and led to a positive evaluation of our work (4 → 5). As you acknowledged, our practically motivated framework achieves state-of-the-art performance. We also hope that our work provides new perspectives on the unified view of domain generalization and adaptation for future research.	Summary: This paper focuses on multi-view 3D object detection. The authors proposed a unified domain generalization and adaptation-based detection method. To enhance the detection model for unseen datasets and address the geometric misalignment problem, the authors proposed a multi-view overlap depth constraint module and a label-efficient domain adaptation technique. The comprehensive experiments on lar-scale datasets, including nuScenes, Lyfy and Waymo have demonstrate the effectiveness of the proposed method. Strengths: - The storyline is clear, the authors provided detailed motivation analysis in the introduction and present their innovation clearly. - The performance is strong. The method shows SOTA performance in variance benchmarks with fast speed. - Most of the figures are clear. This paper is also well-written. Weaknesses: - There is no source code for review. Technical Quality: 3 Clarity: 4 Questions for Authors: This paper seems well-shaped to me. However, I'm not an expert in this field, so I'm open to discussion with other reviewers if they hold opponent opinions to me. Confidence: 3 Soundness: 3 Presentation: 4 Contribution: 3 Limitations: Please refer to the above. Flag For Ethics Review: ['No ethics review needed.'] Rating: 7 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: ### Code release We plan to release the source code upon acceptance. Additionally, in this rebuttal, we provide further explanations, analyses, and results (please refer to global response and Rebuttal PDF). We hope that these materials will be helpful to you. If you have any questions or need further discussion, please feel free to ask. Thank you.	Rebuttal 1: Rebuttal: We sincerely appreciate your effort to review our work. We have carefully read and considered all of the comments and suggestions provided by the reviewers. To assist with your understanding, we provide detailed analyses and additional experiments of Label-Efficient Domain Adaptation(LEDA) in the global response. Furthermore, we have addressed each reviewer's comments accordingly, and have provided detailed responses and references for each comment. We hope that this rebuttal will be helpful in clarifying any misunderstandings or concerns and will contribute to the overall evaluation of our work. Thank you again for your time and consideration. We look forward to hearing your feedback. ---- ### The motivation of LEDA There exist practical challenges in developing and deploying multi-view 3D object detectors for safety-critical self-driving vehicles. Each vehicle and each sensor requires its own model that can operate in dynamic weather, location, and time conditions. Furthermore, collecting large-scale labels in diverse environments is extremely expensive and inefficient. Among those, we are particularly motivated to address the following issues: 1. Stable performance 2. Efficiency of training 3. Preventing catastrophic forgetting 4. Minimizing labeling cost To satisfy these practical requirements, we carefully design an efficient and effective learning strategy, Label-Efficient Domain Adaptation (LEDA). In Rebuttal PDF Figure 1, we evaluate the efficiency of LEDA compared to existing methods, particularly in terms of domain adaptation (DA) performance. LEDA achieves the highest accuracy with low parameters and data cost, demonstrating its practicality and effectiveness in real-world applications. ### Technical details of LEDA Label-Efficient Domain Adaptation is a novel strategy to seamlessly bridge domains gaps leveraging a small amount of target data. To this end, we add extra parameters $\mathcal{A}$ consisting of bottleneck structures (i.e., projection down $\phi_{down}$ and up $\phi_{up}$ layers). \begin{equation} \mathcal{A}(x) = \phi_{up}(\sigma(\phi_{down}(BN(x)))), \end{equation} where $\sigma$ and $BN$ indicates activation function and batch normalization. We parallelly build $\mathcal{A}$ alongside pre-trained operation blocks $\mathcal{B}$ (e.g., convolution, and linear block) in main paper Figure 3 (ii) and below Equation; \begin{equation} y = \mathcal{B}(x) + \mathcal{A}(x), \end{equation} Firstly, we feed $x$ into $\phi_{down}$ to compress its shape to $[H/r, W/r]$, where $r$ is the rescale ratio, and then utilize $\phi_{up}$ to restore it to $[H, W]$. Secondly, we fuse each outputs from $\mathcal{B}$, and Adapter by exploiting skip-connections that directly link between the downsampling and upsampling paths. By doing so, these extensible modules allow to capture high-resolution spatial details while reducing network and computational complexity. Plus, it notes worthy that they are initialized by a near-identity function to preserve previously updated weights. Finally, our LEDA lead to stable recognition in both source and target domains, incrementally adapting without forgetting pre-trained knowledge. To analyze the impact of different architectural choices on the performance of our model, we conduct two ablation studies on the structure of adapters. - In Rebuttal PDF Table 2, we compare variations of the projection up and down layers. - In Rebuttal PDF Table 3, we compare different locations where the adapter can be attached. This allows us to understand how the structure of the adapter affects the model's performance. ### Optimization Objective We optimize our proposed framework UDGA using the total loss function $\mathcal{L}_{total}$ during both phases (i.e., generalization and adaptation). $$\mathcal{L_{total}} = \mathcal{\lambda_{det}L_{det}} +\mathcal{\lambda_{ov}L_{ov}} +\mathcal{\lambda_{p}L_{p}}$$ where we grid-search $\lambda_{det}$, $\lambda_{ov}$ and $\lambda_{p}$ to harmonize $\mathcal{L_{det}}$, $\mathcal{L_{ov}}$ and $\mathcal{L_{p}}$. Specifically, $\mathcal{L}_{total}$ supervises $\mathcal{B}$ during pre-training and $\mathcal{A}$ during fine-tuning, respectively. In Rebuttal PDF Table 1, we highlight the importance and the sensitivity of each objective and demonstrate the validity and the connection of both methods. Additionally, we present more comparison with existing multi-view augmentation in Rebuttal PDF Table 4. Among 5 dominant methods, we applied the top 3 methods, 2D, 3D, and extrinsic augmentation, to our method. ### Limitations There are two limitations of our proposed UDGA for multi-view 3D object detection task. These limitations include: 1. Our proposed method for calculating depth transformation between multi-view images relies on the presence of overlapping regions between images. The width of the overlap region impacts the accuracy of depth estimation, as demonstrated in Appendix C, Fig.6. 2. It is structurally challenging to apply depth constraint techniques to query-based networks. This is because cross-attention networks, including BEVFormer and DETR3D, often project from 2D to 3D without passing through a depth net, making it difficult to obtain depth features and calculate correlations. Pdf: /pdf/51a2303376be2aa0eba923b44ec5a5da4eb99d4d.pdf	NeurIPS_2024_submissions_huggingface	2,024	null	null	null	null	null	null	null	null
Predictive Attractor Models	Accept (poster)	Summary: The paper deals with training generative sequence models under a few constraints to make it biologically plausible. Specifically, states consist of a sparse binary representation (i.e. neurons that are on/off and sequence training and generation are performed using operations that are described as "local and synaptic". The proposed model seems to be a modification / generalization of Hopfield networks (a form of recurrent network) such that weights are asymmetric and there is an orthogonal "hidden" dimension to the representation (hidden in that is does not affect the output directly) that captures prior history and, in effect, turns the model in to an HMM. State transitions involve a linear transition matrix, lateral inhibition (within the hidden state dimension using a max operation) and output-dependent filtering. Output is generated through sampling from a Gaussian mixture model of attractors. The method is backed with some theoretical guaranty of optimality of the observation generation. The model is compared against others of it's kind in online and offline (teacher-forced) settings and assessed with respect to memory capacity (number of sequences and their lengths), robustness to noise, correlation between states, training speed and parameterized hidden state dimension. The proposed model outperforms in every way. Strengths: The proposed model seems like a natural extension of Hopfield networks to the sequence and hidden-state domains. The state, training and inference all seem to adhere to the constraints that make the method "biologically plausible". The various (synthetic and "real") are used well to show the model streangths. Performance is very clearly better than the models against which it is compared. The presentation is very clear. The accompanying video-based presentation is superb. The supplementary material is very extensive and provides additional background, proofs and implementation details at the same level of attention to clarity and detail. Weaknesses: * It is not clear to me what constitutes a biologically plausible mathematical operation. Are the various models equally biologically plausible? Is there a scale of plausibility? With that in mind, I wonder how the models that appear in this study compare against state of the art sequence memorization / generative models that are unconstrained (e.g. transformers?), whether those models are better and are they provably non-plausible. It would have been interesting to read a discussion of whether biological plausibility can help us produce better models or whether (/ how) it is helpful in explaining observations from studies of real biological networks. * Clearly, both memory capacity and training efficiency are highly dependent on the level of sparcity and the number of parameters (be they weights or neurons). The comparison across models is somewhat limited in its treatment of model complexity (albeit the proposed PAM models are compared across several hidden-dimension sizes). I'd expect that some models would match PAM performance in some respects as they are scaled up. It may be useful for the analysis to be extended to cover this aspect of the models better. Technical Quality: 3 Clarity: 4 Questions for Authors: The state transition and the output mechanisms in PAM are different. One uses attractor dynamics and the other uses a deterministic transition and a max operation. Wouldn't it be more biologically plausible to have operations use similar mechanics? It is unclear whether the models presented are useful from a machine learning point of view. Taking the word generation as an example, the level of complexity of that task is orders of magnitude below what language generation models are capable of. Assuming this assertion is correct, what does that imply about the state of what we consider biologically plausible compared with what real brains are capable of and what unconstrained ML methods can achieve. Is there an unaddressed gap? Confidence: 4 Soundness: 3 Presentation: 4 Contribution: 3 Limitations: It is unclear how the methods compare agains general ML methods. It is unclear whether the assumed biolgical plausibility helps understand biological neural mechanisms. Flag For Ethics Review: ['No ethics review needed.'] Rating: 7 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thank you for the review and the great remarks on PAM’s strengths and the clarity of the presentation, supplementary material, and video. We appreciate the positive rating! Weaknesses * Biological plausibility is typically referred to as imposing constraints such as local learning rules [1] (e.g., Hebbian, anti-Hebbian, etc.) and sparsity of representations. For example, backpropagation approaches and methods that copy weights/gradients are known to be highly implausible. Therefore, approaches like Predictive Coding and Associative Memory Models are considered biologically plausible despite some contradictions to biology [1,2] (e.g., synapses cannot magically turn from excitatory to inhibitory; inhibition occurs through interneurons). Methods that map their approaches to the anatomy of the brain can be considered higher on the scale of plausibility. We do not intend to accurately map every detail in PAM to the cortical anatomy; however, some aspects of PAM (e.g., lateral inhibition in minicolumns, afferent thalamocortical input to layer IV, contextual information on basal dendrites, associative connections) can be mapped to the cortical column and subcortical structures. We will provide a discussion in the supplementary material on mapping to anatomy and hierarchical processing with PAM. Unconstrained models (e.g., transformers, LSTMs, CNNs) that use backpropagation to train their weights can learn unique contexts to solve prediction tasks. However, they completely fail in tasks such as continual learning since they overwrite previous knowledge. We provide additional results (rebuttal attachment; figure 1D) comparing PAM to a 2-layer transformer on continually learning protein sequences. Backward transfer results show low performance of transformers in remembering older sequences after training on new sequences. Table 1 (in the rebuttal attachment) supplements these results by showing detailed sequence-by-sequence performance values. Transformers excel at learning the current task (diagonal values) but suffer in remembering previous tasks (lower triangle values). Approaches that use sparsity and local learning rules (i.e., PAM) can easily learn new knowledge without overwriting previous knowledge. * Yes, some models perform well in sequence capacity. For example, AHN with a polynomial degree of 2 and sparsity of 50% seems to outperform PAM with $N_k=4$ (Figure 1A; main paper); however, it quickly loses capacity as soon as correlation is introduced (Figure 1B; main paper). Additionally, PAM with $N_k=1$ seems to consistently perform worse than other methods (Figure 2), as expected. Moreover, PAM requires sparse representations to represent multiple possibilities, and unions of SDRs will not be useful if the representations are not unique (low overlap in bits). This failure mode is specific to PAM; however, we argue that sparsity is the biological solution to continual learning and multiple possibilities representation and that models should be using sparsity to enhance their performance. Questions * While both mechanisms may seem quite different, they both rely on Hebbian learning and synaptic excitation/inhibition to learn associations between patterns. State transition learns to associate patterns at time $t$ to patterns at time $t+1$, while the emission model builds attractors by associating the firing neurons of each pattern and inhibiting predicted neurons of other patterns. It is important to note that the cortical column has a distinct structure of minicolumns where axons are bundled together [3]. This may force competitive learning rules (through lateral inhibition) across the pyramidal neurons of each minicolumn. This structure is not assumed for associative memory models. In other words, the morphology of the minicolumn may play a role in defining the learning mechanics. * It is important to note that the word generation task presented in PAM is very different from LLMs. PAM performs regression-based prediction; we do not assume knowledge of a dictionary of tokens/characters. PAM regresses the full future representation. However, current LLMs are deterministic in nature. Therefore, they assume knowledge of a set of possible tokens and predict a probability distribution over the token IDs, from which they sample a single ID that retrieves the token from the stored dictionary. Current LLMs are trained in a contrastive manner as classification models. While this approach can work for LLMs (to a certain extent), they are not suitable for tasks that do not assume knowledge of the entire set of possible predictions (e.g., video frame predictions). As a side note, this approach to LLMs does not allow building compositional models with higher-order predictions (e.g., sentence predicting sentence) because we simply cannot tokenize all possible sentences. However, in PAM, we can predict frames, regress multiple possibilities, and build true compositional models. We hope to have addressed all of your concerns and look forward to any additional discussions References [1] Salvatori, Tommaso, et al. "Brain-inspired computational intelligence via predictive coding." arXiv preprint arXiv:2308.07870 (2023). [2] Chrysanthidis, Nikolaos, Florian Fiebig, and Anders Lansner. "Introducing double bouquet cells into a modular cortical associative memory model." Journal of computational neuroscience 47.2 (2019): 223-230. [3] Peters, A., and C. Sethares. "The organization of double bouquet cells in monkey striate cortex." Journal of neurocytology 26.12 (1997): 779-797.	Summary: The paper presents a novel sequence memory architecture focused on multistability, ie: generating multiple possible futures from every present context. They propose to train this architecture in a biologically plausible way via predictive coding under the Free Energy Principle, in the online learning setting where data items are seen once instead of repeated in mini-batches. Section 2 reviews background, and Section 3 describes the Predictive Attractor Model formally. In Section 4 the paper evaluates PAMs for their storage capacity, robustness against catastrophic forgetting, multistability, and noise robustness, comparing to symmetric and asymmetric Hopfield Networks and temporal predictive coding (tPC). Strengths: The paper attempts and achieves a high degree of biological plausibility by focusing on specific learning mechanisms, observed in the brain, for each computation that must be carried out. To my knowledge it is the first sequence memory architecture to explicitly consider a mixture likelihood in order to consider multiple possible continuations given every context. I have a hard time evaluating the quality of the experimental results, since I have not kept up with the state of the art in sequential memory models this past year or two. Weaknesses: Clarity is the main aspect in which the paper lacks. The paper uses a lot of concepts without either initially laying them out together or introducing them explicitly the first time they are needed and then referencing back to them when they are used. A table of notations somewhere would be quite helpful. For instance, presenting terms as \log 1/p(x) sows a bit of confusion over whether a likelihood ratio/Radon-Nikodym derivative is the intended concept being conveyed, or just \log 1/p(x) = -\log p(x) giving an entropy or cross-entropy. Technical Quality: 3 Clarity: 2 Questions for Authors: Can the authors provide a unified presentation of the different biologically-informed learning rules/computations they propose, and how those unite back with the variational inference perspective introduced at the start of the paper? I would suggest that the authors provide some hyperlink to Table 1 (in the Supplementary material, giving notations) in the main text, or even move Table 1 into the main text. Regarding the neuroscience, why does the paper refer to a cortical minicolumn (rather than a laminar column or cortical microcircuit?) rather than a hippocampal circuit? Sequential memory architectures, particularly those akin to Hopfield networks, are more typically considered as models of hippocampal circuitry. Confidence: 3 Soundness: 3 Presentation: 2 Contribution: 4 Limitations: The authors freely admit that their model relies on having a pretrained model, such as an autoencoder, to convert sensory-level inputs into sparse distributed representations. Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thank you for the positive remarks. Yes, to our knowledge, we are the first to model multiple possibilities as a GMM in a state space model and prove that maximizing the likelihood of a query observation under the mixture model is equivalent to a Hopfield recall function (Theorem 1). Weaknesses * We will provide a link to the notation table in the supplementary for clarity. Due to space constraints in the main paper, we moved the derivation of the variational free energy (VFE) equation to the supplementary material (Derivation 2; appendix B1). The log term of a conditional probability in VFE refers to the log-likelihood, whereas the expected log of a marginal probability refers to the entropy of the distribution. Questions * In PAM, we only use Hebbian plasticity as the biologically plausible learning rule. The variational free energy equation minimizes the KL divergence between the intractable true posterior and the approximate posterior in the Bayesian inference framework by simply minimizing the negative log-likelihood (NLL) terms (i.e., latent and observation terms in equation 3); see derivation 2 in appendix B1. Since we assume Gaussian form for the latent term and a mixture of Gaussians for the observation term, minimizing the prediction errors will maximize the likelihood and, therefore, push the approximate posterior towards the true posterior. We use Equation 7 to minimize the prediction error for the latent state term, which uses local Hebbian rules to correct its predictions by modifying the synaptic weights. For the observation state, we correct the model's predictions by forming attractors around the observation (in a contrastive manner), effectively learning the parameters of the GMM. During inference, applying the Hopfield recall function iteratively maximizes the likelihood of the query observation under the learned mixture model (Theorem 1). Therefore, using Hebbian rule in equations 7 and 8, minimize the variational free energy, which minimizes the KL divergence between $q(z_t)$ and $p(z_{t} \| z_{t-1}, x_{t})$ (i.e., the goal of variational inference). * Thank you for the suggestion. We will link to the table in the supplementary or move it to the main paper. * While associative memory models are typically considered models of hippocampal circuitry, spatiotemporal sequence modeling and integration can be considered a function of the neocortex [1,2]. The architecture of PAM follows the same anatomy mapping of HTM [3], in which sequence modeling is performed in layer IV of the cortical column. Many connections between layers in a single cortical column (IV ↔︎ VI) and across columns (III ↔︎ III) are considered bidirectional associative connections; therefore, it is biologically plausible for associative connections to exist in the neocortex. We will provide more discussions on mapping of PAM to the neocortex anatomy in the camera-ready submission. We hope to have addressed all of your concerns and look forward to any additional discussions References [1] Oberländer, Jette, Younes Bouhadjar, and Abigail Morrison. "Learning and replaying spatiotemporal sequences: A replication study." Frontiers in integrative neuroscience 16 (2022): 974177. [2] Synaptic inhibition in the neocortex: Orchestration and computation through canonical circuits and variations on the theme --- Rebuttal Comment 1.1: Comment: [3] Hawkins, Jeff, and Subutai Ahmad. "Why neurons have thousands of synapses, a theory of sequence memory in neocortex." Frontiers in neural circuits 10 (2016): 174222.	Summary: This paper examines a biologically plausible model for sequential memory which overcomes various issues (capacity, forgetting, size of context window) in previous models. The proposed model is essentially an extension of Hierarchical Temporal Memory which allows the model to generate predictions. Strengths: * Memorizing sequences is a fundamental ingredient of cognition and remains a challenge for neuroscience, while no theoretical model has been completely satisfactory. * Hierachical Temporal Memory remains a compelling model of sequence memorization and attempts to expand its capabilities are worthwhile. * The model presented here is simple, well-explained, and composed of familiar ingredients. * The experiments are clear and seem appropriate. Weaknesses: * This paper's contribution seems to be a bit weak. What it builds upon Hierarchical Temporal Memory [1] is the ability for the model to generate predictions, which can be used to sample autoregressively or robustly handle noisy input. While the sampling capability in particular could be interesting, key properties of such a function -- perhaps most obviously, which distribution does it produce? Can it learn the input statistics? -- are not explored. * In relation to the above, it inherits a (slight) weakness from HTM: reliance on the ability of dendritic computation to put neurons into a "predictive" state, which is necessary (but not sufficient) for firing. We should be leery of overstating the biological plausibility of what remains a hypothetical mechanism. * The model is biologically plausible in that it relies on local updates, but various other aspects of it (especially the configuration of the minicolumns) seems to need to be set up precisely in advance, and dense connectivity is assumed. * The theoretical results amount to showing that sampling (an observation given context) can be written as a Hopfield recall problem. A deeper theoretical analysis of this model -- touching on capacity, number of rounds for convergence, and characterizing the distribution which is sampled from -- would significantly strengthen this paper's contribution. [1] Hawkins, Jeff, and Subutai Ahmad. "Why neurons have thousands of synapses, a theory of sequence memory in neocortex." Frontiers in neural circuits 10 (2016): 174222. Technical Quality: 3 Clarity: 2 Questions for Authors: * What does $\mathcal N(x; \mu_c, \Sigma_c)$ symbolize in Eq. (4)? * It would be nice to examine a crucial step of the generative algorithm -- lines 7-8 in algorithm 2, which combine a Hopfield step with influence from the context -- in more detail, theoretically and/or in experiment. If something like this has been explored in the literature before, it would also be good to point to it. * Please define Jaccard index/IoU in the main text. Confidence: 4 Soundness: 3 Presentation: 2 Contribution: 2 Limitations: Since the paper claims biological plausibility, I would appreciate a more thorough and insightful discussion of what features are plausible and which are less so, with references to the relevant neurophysiology literature. Flag For Ethics Review: ['No ethics review needed.'] Rating: 4 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thank you for your positive remarks on the simplicity of the model and the importance of the challenge being solved by PAM. Weaknesses * PAM extends both HTM and Hopfield Network (HN). It allows HTM to autoregressively generate sequences (enabling HTM to do more than just anomaly detection) and also provides a framework for HN to perform temporal predictions without the need for asymmetric weights (unstable and lack capacity as shown in the experiments; Figure 1B in the main paper). The learning of attractors is performed alongside the sequence modeling in a novel contrastive manner (different from how HN weights are learned; summing up patterns). Therefore, the attractors “push away” only the patterns that appear as a possibility in the same context during prediction (not all the other patterns), improving the capacity of HN. In our mathematical framework, we have used a Gaussian Mixture Model (GMM) to represent the learned possibilities and proved that maximizing the likelihood of an observation under the mixture model is equivalent to performing recall in a HN framework (Theorem 1). We also showcase the model’s ability to generate full words despite multiple valid possibilities. Being able to generate a recall of close to 80% (Figure 2D in the main paper) of the full dataset means that the model learned the input statistics because it can cover most of the data when randomly sampling from the data distribution. * Biological plausibility in PAM mainly refers to excluding the obvious biologically implausible learning rules, such as backpropagation or copying of weights. This allows models like PC and HN to be referred to as biologically plausible despite clear contradictions with anatomy [1, 2]. Moreover, PAM also constrains its learning and representations to sparse and binary cell assemblies, which improves the plausibility of the model. Therefore, a biologically plausible model here refers to the ability to continually learn using only local learning rules (e.g., Hebbian, anti-Hebbian, STDP) and sparse binary representations. However, we do not intend to accurately map every detail in PAM to the anatomy of the neocortex. * HTM relies on lateral inhibition such that when a depolarized cell fires, it inhibits the firing of other cells in the same minicolumn. The configuration of the minicolumns can be seen in every cortical column as vertical bundles of myelinated axons; each bundle has a single inhibitory double bouquet cell (i.e., interneuron) [3]. Therefore, HTM assumes interneurons perform lateral inhibition during prediction but does not learn the synapses of the interneurons. It can be very hard to accurately map the details of any learning algorithm to the anatomy since “no computer simulation can fully replicate the intricate workings of the brain in every respect” [1]. We provide additional results (rebuttal attachment figure 1C) on varying the connection density and show that PAM can still perform relatively well after removing some of the connections. We will extend this discussion and add it to the supplementary material. * Thank you for these great suggestions for additional theoretical analyses. We show in Theorem 1 that minimizing the prediction error by forming attractors around the observations (Hopfield Network) can be characterized as a mixture of Gaussians distribution where each mode represents a center of an attractor (single observation). Questions * This is the Gaussian distribution for the component $c$ (of the GMM) at mean $\mu_c$ and covariance $\Sigma_c$. When substituting the point $\boldsymbol{x}$ in this function, it returns a high number if it is close to the mean of this component (attractor center). Therefore, normalizing by the sum of distances to means of all components gives a similarity measure. This similarity measure becomes a normalized score for each component, denoting how close the observation is to each of them. * This step ensures that the Hopfield iterations stay bounded by the model’s beliefs (predicted union of SDRs). It effectively minimizes the number of attractors during generation. If removed, the noisy observation will be pushed towards the closest observation regardless of what is predicted by the model. Thank you for the suggestion of future experiment. To our knowledge, we are the first to implement such a rule for an associative memory model. * We will define it in the main text. Thank you. We hope to have addressed all of your concerns and look forward to any additional discussions References [1] Salvatori, Tommaso, et al. "Brain-inspired computational intelligence via predictive coding." arXiv preprint arXiv:2308.07870 (2023). [2] Chrysanthidis, Nikolaos, Florian Fiebig, and Anders Lansner. "Introducing double bouquet cells into a modular cortical associative memory model." Journal of computational neuroscience 47.2 (2019): 223-230. [3] Peters, A., and C. Sethares. "The organization of double bouquet cells in monkey striate cortex." Journal of neurocytology 26.12 (1997): 779-797. --- Rebuttal Comment 1.1: Comment: Thank you for the thorough response. One point of clarification: "Being able to generate a recall of close to 80% (Figure 2D in the main paper) of the full dataset" only indicates that these distributions have similar _support_. There are other questions one might ask, for instance if the model samples a next token starting from some context, how close is the (conditional) distribution over tokens to the frequency with which those tokens appeared in the training data after that context? For some purposes it might be beneficial for the model's distribution to differ from the training distribution, but in any case it is important to quantify what the learned distribution actually is, not just its support. --- Reply to Comment 1.1.1: Title: Reply on PAM's Learned Distribution Comment: Thank you for your reply and useful suggestions. Yes, we agree. It would be beneficial to further study additional characteristics of the learned conditional distribution. In theory, the learning rule (eqn. 8) is designed such that the attractors (GMM components) are more defined with training data. Therefore, the frequency of seeing a specific pattern controls the energy of that attractor center; more frequent patterns will have lower energy attractors and wider attractor basins. That being said, other factors can affect this rule. For example, clamping the synaptic weights prevents some frequent patterns from dominating the distribution. --- Rebuttal 2: Comment: Thank you for your positive remarks. We will incorporate your suggestions in the paper. We will also include an extended discussion on the learning rules (learned distribution), biological plausibility and the mapping of PAM to cortical anatomy. Your suggestions have greatly improved our paper, please let us know if you have other questions or concerns.	Summary: The paper presents Predictive Attractor Models (PAM), designed to improve sequence memory by resolving issues like catastrophic forgetting. PAM processes inputs in real-time, remembering each just once, using lateral inhibition in cortical minicolumns to preserve old memories. By generating future predictions from previously trained possibilities, PAM leverages synaptic adjustments and competitive learning, inspired by Hierarchical Temporal Memory systems. This combination enhances memory robustness and versatility across different data types, marking a significant advance in creating biologically plausible, computationally efficient memory systems for cognitive science and AI research. Strengths: 1. By processing each input only once and using mechanisms such as lateral inhibition, PAM effectively retains old memories without them being overwritten by new ones. 2. PAM's ability to handle inputs in real-time allows for dynamic updating of memory without the need for batch processing or retraining. 3. It can generate multiple future predictions based on a set of learned possibilities, providing flexibility and adaptiveness in response generation. 4. PAM has been tested and shown to be effective across various types of data, including text, visual inputs, and protein sequences, demonstrating its applicability in diverse fields. Weaknesses: 1. Need for Detailed Comparison with HTM: There is a necessity for a more thorough comparison between PAM and Hierarchical Temporal Memory (HTM). By detailing the differences and improvements PAM offers over HTM, the unique contributions of PAM can be better understood and appreciated. This analysis would help delineate PAM's novel features and demonstrate its advancements over HTM in handling sequence memory. 2. Risk of Overfitting: PAM's robust capacity to utilize past information and generate predictions from learned attractors might lead to overfitting. This risk is particularly high in environments with highly variable or noisy data. Additionally, if the sequence itself contains repetitive elements meant to convey information, PAM may struggle to capture this, potentially losing the ability to extract meaningful patterns from repetitions in the data. 3. Parameter Sensitivity: The performance of PAM heavily relies on the precise configuration of its parameters, including the strength of synaptic connections and the degree of lateral inhibition. Incorrectly calibrated parameters can significantly impair the model's functionality, making it less effective and reducing its overall reliability. Technical Quality: 3 Clarity: 2 Questions for Authors: 1. The implementation of PAM involves complex mechanisms, such as lateral inhibition in cortical minicolumns and synaptic adjustments, which can be more intricate compared to traditional memory models. Moreover, the dimensionality of matrices involved in PAM (e.g., matrix A and B's dimension being $(N_c \times N _k)^2+N_c^2$ contributes further to this complexity. Such high complexity demands more computational resources and sophisticated programming techniques, which may limit PAM’s accessibility and scalability in practical applications. 2. In the introduction, it is emphasized that different contexts in a sequence should lead to different perceptions of the same stimulus. However, there is a concern that PAM might merely memorize the entire sequence without effectively considering the variability in sequence content. This approach could potentially overlook the importance of recognizing repetitive structures within sequences, which are crucial for efficient sequence memory. The ability to identify and utilize such repetitions can greatly enhance memory efficiency and predictive accuracy, suggesting a potential area for improvement in PAM's design. Confidence: 2 Soundness: 3 Presentation: 2 Contribution: 2 Limitations: The article compares related work in the introduction and mentions some limitations. Flag For Ethics Review: ['No ethics review needed.'] Rating: 4 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thank you for highlighting the strengths of PAM in retaining old memories, continual learning, and generating multiple possibilities. Weaknesses 1) PAM extends HTM by providing a mechanism for filtering noisy inputs and generating sequences in an autoregressive manner. * The current implementation of HTM predicts multiple possibilities as a union of SDRs. This union entangles the possibilities and therefore can only be used for anomaly detection (comparing new observation to the union prediction). PAM adds generative ability to HTM, enabling it to perform autoregressive sequence generation (i.e., LLMs, video completion, etc.) * HTM does not use its predictions to clean the input observations from noise. The noisy observations are used as groundtruth. PAM applies its learned beliefs and predictions to filter the noisy observations. Note that observations are filtered based on the attractors from within the predicted unions, not the entire attractor space (see the intersection in line 8 of algorithm 2). * PAM provides a mathematical formulation that unites both HTM and HN in a single Bayesian framework. * PAM also extends Hopfield Networks by enabling temporal sequence processing without the need for asymmetric weights. Therefore, it inherits all the convergence and stability guarantees of vanilla HN [1,2]. 2) We believe that the tendency of PAM to push noisy observations towards familiar observations (attractors) matches behavioral experiments of “filling in” information [3]. In an online learning system, we can determine how close the prediction is to the observation; if the discrepancy is high, we learn new temporal dependencies. Otherwise, we can remove the noise from the observation, which allows us to make predictions based on our version of the observation (i.e., attractor center). Being able to make predictions after seeing noisy observations can be considered generalization and contributes to adversarial robustness. We agree that PAM’s ability to capture long context may not allow it to extract sub-patterns in repetitive data. This problem can be handled in a hierarchical structure of PAM blocks, where low-level repetitions are handled by a lower-level PAM block that sends a summary of the repeated pattern to the higher-level PAM block for higher-order predictions. In this work, we focused on implementing a modular PAM block, and we will continue to work on a hierarchical architecture to capture the compositional structure in data. 3) It is important to note that we do not tune hyperparameters for specific tasks. Instead, we use the same set of parameters for all modalities (i.e., vision, text, protein sequences) and all tasks (i.e., sequence capacity, noise robustness, catastrophic forgetting, and autoregressive sequence generation). Therefore, we believe that the model is not sensitive to parameters. In addition to extensively evaluating the performance of the model at different SDR sizes (i.e., $N_c$) and number of neurons per minicolumn (i.e., $N_k$) in the main paper, we provide additional results on varying the connection density (rebuttal attachment Figure 1C) and show that PAM performs relatively well after randomly removing some of the connections. Questions 1) We agree that the implementation involves additional mechanisms than a simple associative memory model; however, the structure of cortical columns is complex. Minicolumns do exist in every cortical column and are usually referred to as a vertical bundle of myelinated axons, where each minicolumn has a single inhibitory double bouquet cell [4], which can directly inhibit the pyramidal neurons in the minicolumn. This results in an intricate mechanism which allows for accurate temporal predictions that cannot be achieved with a simple memory model (HN). In figure 1D of the main paper, we show that PAM is significantly more efficient than temporal predictive coding (tPC) and is performed on a CPU. 2) Please refer to the answer to point 2 in the weaknesses section. We hope to have addressed all of your concerns and look forward to any additional discussions References [1] Hopfield, John J. "Neural networks and physical systems with emergent collective computational abilities." Proceedings of the national academy of sciences 79.8 (1982): 2554-2558. [2] Ramsauer, Hubert, et al. "Hopfield networks is all you need." arXiv preprint arXiv:2008.02217 (2020). [3] Papenmeier, Frank, Alisa Brockhoff, and Markus Huff. "Filling the gap despite full attention: The role of fast backward inferences for event completion." Cognitive Research: Principles and Implications 4 (2019): 1-17. [4] Peters, A., and C. Sethares. "The organization of double bouquet cells in monkey striate cortex." Journal of neurocytology 26.12 (1997): 779-797. --- Rebuttal Comment 1.1: Comment: I have reviewed the comments from other reviewers and appreciate the authors' thoughtful responses to them. I will be maintaining my original score. --- Reply to Comment 1.1.1: Comment: Thank you for your reply. Please let us know if you have any other questions or concerns.	Rebuttal 1: Rebuttal: We thank the reviewers for their comments and useful suggestions. We think that these comments have helped us better refine the paper. We provide additional supplementary results to help address some of the concerns on scaling, connection sparsity and comparing to transformer architecture. We refer to these experiments in the individual rebuttals. Pdf: /pdf/80e2677a04b42848a5ca304e09016ff6a6395eec.pdf	NeurIPS_2024_submissions_huggingface	2,024	Summary: This paper introduces Predictive Attractor Models (PAM), a novel biologically-inspired approach for sequential memory and prediction. PAM combines a predictive model based on sparse distributed representations (SDRs) with an attractor model for generating future predictions. Key contributions include an online learning algorithm that encodes temporal context without catastrophic forgetting, the ability to represent and sample from multiple possible future predictions, noise robustness through learned attractors, and formulation as a variational inference problem within a state space model framework. The authors evaluate PAM on various sequence learning tasks, demonstrating superior performance compared to temporal Predictive Coding (tPC) and Asymmetric Hopfield Networks (AHN) in terms of sequence capacity, avoiding catastrophic forgetting, generating multiple possibilities, and noise robustness. Strengths: - Novelty in combining predictive coding with attractor dynamics for sequential memory. - Strong theoretical grounding within a variational inference framework. - Comprehensive evaluation on various tasks and datasets. - Biological plausibility through local learning rules and sparse binary representations. - Addressing key challenges in sequential memory (catastrophic forgetting, multiple possibility generation, noise robustness). Weaknesses: - Limited comparison to modern sequence learning approaches (e.g., transformers). - Scalability concerns for very long sequences or high-dimensional inputs. - Lack of hierarchical processing discussion or implementation. - Insufficient analysis of hyperparameter sensitivity. - Limited discussion of failure modes and limitations. Technical Quality: 3 Clarity: 3 Questions for Authors: 1. How does PAM's performance change with increasing sequence length? 2. Can PAM be extended to handle variable-length or incomplete sequences? 3. How does PAM perform with more structured types of noise? 4. Can you provide more insight into the attractor model's learning of multiple possibilities? 5. Have you explored more sophisticated SDR encoding schemes for complex inputs? 6. How might PAM incorporate top-down feedback or attention mechanisms? 7. What specific neural circuit predictions arise from your model? Confidence: 3 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: - The requirement for sparse binary representations limits direct application to dense inputs. - The use of a separate autoencoder for vision experiments adds complexity. - The paper lacks thorough discussion of computational complexity and memory requirements for larger-scale problems. Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thank you for your thorough review and positive remarks on novelty, theoretical grounding, and comprehensive evaluation of key challenges in sequential memory. We have provided additional figures to address your concerns. Weaknesses * Similar to other papers on biologically-plausible sequential memory [1], we have mainly focused on comparing to biologically plausible approaches that do not employ backpropagation for learning (i.e., only local learning rules). Transformers and deep learning based approaches (e.g., LSTMs, MAMBA, etc.) use backpropagation through time. While these modern sequence learning approaches use backpropagation to successfully learn meaningful context representations, they still struggle against key challenges such as catastrophic forgetting, generating multiple possibilities, and adversarial robustness. In the rebuttal attachment (Figure 1D), we compare PAM against a 2-layer transformer on the task of continually learning multiple protein sequences. We show that transformers completely forget older sequences after training on new sequences, whereas PAM successfully learns the sequences without catastrophic forgetting. We also show detailed results of the transformer in Table 1. Transformers excel at learning the current sequence (diagonal values) at the expense of forgetting previously learned sequences (lower triangle values). * PAM scales well with the size of the SDR $N_c$ and the number of neurons per minicolumn $N_k$. In the rebuttal attachment (Figure 1A), we show that PAM scales up exponentially with the size of the SDRs. In our experiments, we only use $N_c$ to be 100 neurons for fair comparison with previous approaches; however, a cortical column typically contains close to 1500 pyramidal cells in layer IV only [2]. Please note that Figure 1A in the main paper shows the capacity on a log scale and that while some other approaches seem to be competitive to PAM (i.e., AHN d=2), they lose all their capacity when provided with correlation in the sequence as shown in Figure 1B in the main paper. * In the future, we plan on implementing a hierarchy of PAM blocks. However, a hierarchy in PAM is different from simply stacking layers of transformer blocks (or CNN filters). We believe in an explicit part-whole hierarchy where the higher-level performs higher-order predictions in a true compositional manner. Higher-level predictions in vision can be predicting objects after seeing multiple objects (i.e., context), whereas in NLP, it would be predicting sentences or phrases instead of tokens. In PAM, each level would need to send a summary representation of the current sequence (i.e., latent state) [3,4] to the higher level for recursive processing. We will add a detailed discussion in the supplementary for hierarchical processing and higher-order predictions in PAM. * The main hyperparameters in PAM are $N_c$ and $N_k$, which we test extensively throughout the paper. We provide additional results on testing the effect of connection density (Rebuttal attachment figure 1C). While neurons in PAM are currently densely connected (similar to Hopfield Network), we show that PAM can still perform relatively well in a sparser setup. It is important to note that we do not tune any hyperparameters for specific tasks. We use the same model and parameters for testing sequence capacity, noise robustness, catastrophic forgetting, and sequence generation with different modalities (i.e., text, vision, protein sequences, synthetic). * PAM’s performance degrades when not using enough neurons to represent the context or when SDR representations become dense such that a union of two SDRs doesn’t provide unique information about either SDR. In figures 2A, 2B, 2C, and 2D (main paper), we show the effect of low values of $N_k$. At one neuron per minicolumn, the model fails to capture context and performs worse than other models. Also, Figure 12 in the supplementary material shows how the performance of the generated words degrades with lower $N_k$ (4). Questions 1) We provide additional results (Rebuttal attachment Figure 1B) on the IoU performance of a small PAM model ($N_c=50$, $N_k=4$) with increasing sequence length. In offline generation, incorrect predictions accumulate quickly, leading to a sharp drop in performance. However, online generation attempts to correct inaccurate predictions. Therefore, performance smoothly degrades with sequence length. Note that both experiments (i.e., online and offline) use identical seeds/sequences for a fair comparison. 2) PAM can handle variable-length and incomplete sequences. The task of the protein sequence is one example of dealing with variable lengths. Each protein sequence has a different length; PAM can learn each of them and generate them in an offline, auto-regressive manner. 3) We only introduce noise into the SDRs to simulate neuronal noise caused by firing rate (i.e., misfiring, random spikes, or bursts). Different types of neurons and interneurons can generate uncorrelated spikes [2]. The collective behavior of neurons in a sequence model should be capable of filtering out and inhibiting noisy action potentials. We show that the introduced attractors can effectively remove noise. We have not tested other types of noise in this work, but in future work, we plan on testing the model’s performance on sequences after the order of observations has been completely or partially randomized. Replies to the rest of the questions are available in a follow-up comment --- Rebuttal 2: Title: Follow-up to rebuttal Comment: Follow-up replies to questions 4) The predictive component of PAM allows it to learn these multiple valid possibilities as unions of representations. This can only be achieved because of the sparsity of representations (i.e., SDRs). If a dense network (e.g., transformer) attempts to generate multiple possibilities in a regression predictive task, they will both be averaged in Euclidean space, or one will overwrite the other (i.e., catastrophic forgetting). While the union of SDRs in PAM represents multiple possibilities, they are entangled. The attractor model disentangles this union by forming an attractor for each possibility. Descending the energy landscape towards an attractor inhibits the other possibilities and only excites one of the possibilities. Note that we do not learn these attractors by simply summing up all the existing patterns (as in Hopfield network), but this behavior is learned in a continual manner alongside the predictor during sequence memory learning, which is a novel approach to learning the attractors. The supplementary video provides a visual explanation of attractors learning. 5) Yes, we have explored other approaches to encode complex inputs into SDRs. One option is to use a scalar encoder and spatial pooler [5]. This pipeline can be used to turn an input into a dense binary representation; then the spatial pooler prunes the representations using top-k and boosting tricks. While this approach can effectively turn complex inputs (e.g., images) into SDRs, it struggles to convert the SDR back to the input. After exploring different approaches, we found that the sparse autoencoder performs the best (reconstruction images shown in supplementary Figure 7). Note that the sparse autoencoder is a novel approach for converting between dense representations and SDRs; we are the first to introduce the tricks involved in training such a sparse and binary representation. 6) Top-down feedback between hierarchical regions (e.g., V1 and V2) in the neocortex is received on the apical dendrites of pyramidal neurons. The feedback pathway originates at layer VI(a) of a higher region (e.g., V2) and forms en-passant synapses in layer VI(a) of the lower region in the hierarchy (e.g., V1). These connections reach Layer I and provide feedback on the apical dendrites of other layers [6]. In PAM, we envision similar feedback connections from a higher-level PAM block. The attractor in the higher-level block can reach a decision on the object being seen (e.g., human) based on partial observation (V1: layer III → V2: layer IV), and therefore, the feedback inhibitory connections on the apical dendrites can rule out some lower-level possibilities (e.g., paws). A hierarchy of PAM blocks remains future work, but we will add a discussion on compositionality and top-down feedback in the supplementary. 7) PAM inherits assumptions from HTM [8] that the sequence modeling process takes place in Layer IV of the cortical column. Layer IV receives thalamocortical driving input (i.e., afferent sensory input) from the thalamus on the proximal dendrites of the pyramidal neurons of Layer IV. The context information is received on the basal dendrites and primes a specific context by depolarizing neurons. The depolarized neurons fire first and inhibit others in the same minicolumn through lateral inhibition. Please note that we do not explicitly model interneurons but assume that the inhibition is caused by Double Bouquet Cells (DBC) [2]. There is exactly one DBC for each vertical bundle of myelinated axons forming a minicolumn [9]. Interneurons are also assumed to perform the inhibitory tasks of the attractor model [10]. References in a follow-up comment We hope to have addressed all your concerns and look forward to any additional discussions --- Rebuttal 3: Title: Follow-up References Comment: References [1] Tang, Mufeng, Helen Barron, and Rafal Bogacz. "Sequential memory with temporal predictive coding." Advances in neural information processing systems 36 (2024). [2] Markram, Henry, et al. "Interneurons of the neocortical inhibitory system." Nature reviews neuroscience 5.10 (2004): 793-807. [3] Mounir, Ramy, Sujal Vijayaraghavan, and Sudeep Sarkar. "STREAMER: Streaming representation learning and event segmentation in a hierarchical manner." Advances in Neural Information Processing Systems 36 (2024). [4] LeCun, Yann. "A path towards autonomous machine intelligence version 0.9. 2, 2022-06-27." Open Review 62.1 (2022): 1-62. [5] Cui, Yuwei, Subutai Ahmad, and Jeff Hawkins. "The HTM spatial pooler—A neocortical algorithm for online sparse distributed coding." Frontiers in computational neuroscience 11 (2017): 272195. [6] Stettler, Dan D., et al. "Lateral connectivity and contextual interactions in macaque primary visual cortex." Neuron 36.4 (2002): 739-750. [7] Schuman, Benjamin, et al. "Neocortical layer 1: an elegant solution to top-down and bottom-up integration." Annual review of neuroscience 44.1 (2021): 221-252. [8] Hawkins, Jeff, and Subutai Ahmad. "Why neurons have thousands of synapses, a theory of sequence memory in neocortex." Frontiers in neural circuits 10 (2016): 174222. [9] Peters, A., and C. Sethares. "The organization of double bouquet cells in monkey striate cortex." Journal of neurocytology 26.12 (1997): 779-797. [10] Chrysanthidis, Nikolaos, Florian Fiebig, and Anders Lansner. "Introducing double bouquet cells into a modular cortical associative memory model." Journal of computational neuroscience 47.2 (2019): 223-230. --- Rebuttal Comment 3.1: Comment: I have reviewed the responses from the authors and discussion with other reviewers. I appreciate the author's in-depth responses, including the detailed biological motivation and citation. Many of the weaknesses and questions I had were addressed. I am not fully convinced on whether the Layer IV circuit described actually implements a PAM-like mechanism, but understand tying biologically-plausible models to direct circuits is difficult and was curious to hear potential implementations. I think this is model takes an interesting direction, and look forward to the completed version of the paper. I will be maintaining my positive review. --- Reply to Comment 3.1.1: Comment: We appreciate your positive review and remarks. Your comments, especially on additional experimental results and discussions, have significantly refined our paper. We will be adding these results to the paper along with a discussion on biological plausibility and cortical circuitry with respect to PAM's architecture.	null	null	null	null	null	null
Search for Efficient Large Language Models	Accept (poster)	Summary: This paper introduces a neural architecture search method for Large Language Models (LLMs) comprising three steps: inheriting the most salient weights from the original model to form the initial sub-network, using an evolutionary algorithm to search for the optimal sub-network, and reconstructing the original network’s output with calibration samples. Strengths: Overall, I like the application of evolutionary algorithm to pruning problem. 1. The method overcomes the limitations of uniform sparsity across all layers—a common but sub-optimal constraint in previous structured pruning methods. 2. It avoids reliance on back-propagation, enhancing its applicability to larger models within a constrained memory budget. Weaknesses: Here are some technical concerns: 1. The candidate evaluation is conducted based on perplexity on WikiText2. This is one metric on one task, then the searching algorithm might be biased towards the task used for candidate evaluation, resulting in a model favoring one task (e.g., language modeling) over others. How to ensure the performance of the searched model on other tasks? 2. Based on my experience, using perplexity as a metric for selecting pruning candidates can result in good performance on WikiText but sub-optimal performance on the MMLU dataset. It would be great if authors could validate their pruned (or searched) model on MMLU. 3. Table 5. The performances on QA tasks should also be presented. If the space is not enough, authors can put them in the Appendix and link it in the main text. 4. Figure 7. Tokens/s is not a very rigorous measurement for speed. The generated tokens per second can depend on many factors, e.g., floating point precision, KV Cache, flash attention, batch size, etc. I suggest authors use other metrics such as MACs or at least provide more specifications. Technical Quality: 3 Clarity: 2 Questions for Authors: In addition, here are some points about writing: 5. The term "inheriting ratio" needs to be rigorously defined in an early section. In addition, authors would better add a sentence describing the difference between "inheriting ratio" and "sparsity" / "pruning ratio". In my understanding, "inheriting ratio" = 1 - "sparsity". 6. Figure 2 Left: it is better to add an annotation showing which axis is the layer index. I assume it is the horizontal axis. 7. Section 3.2, Lines 139-149: The writing of this part needs to be improved. It gives me a lot of confusion when I read this part. (i) Please define what $M$ and $P$ are in Line 140. (ii) Are the masks smooth or binary? If binary then $\{0, 1\}$ should be used instead of $\mathbb{R}$ in Line 140. (iii) Line 141, why do different layers of Attn / MLP share the same mask? If so, would the sparsity of each layer be the same? However, according to Figure 1, the sparsity should be different across layers. Also, in Line 146, authors "set the same inheriting ratios for the masks in all building blocks". So, these claims and figures are very chaotic. Additionally, what does "align the internal computations" mean? (iv) Line 144, Are Eq. (2) and (3) jointly optimized (e.g., take the sum) or separately? Are the masks of different layers jointly optimized or separately? 8. Section 3.3.1, Line 165: $\gamma$ should be rigorously defined. Confidence: 4 Soundness: 3 Presentation: 2 Contribution: 3 Limitations: 1. Section 5: The limitations provided by the authors are too trivial. It is obvious that larger models require more time for pruning/searching. Please add more in-depth discussions on limitations. Authors can refer to the points listed in the "Weaknesses" section for more substantial issues that need addressing. Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thanks for the suggestions from the reviewer. ### Weakness 1. Potentially biased evaluation To mitigate the potential biased evaluation problem, we do not use the data from specific downstream tasks (QA datasets such as MMLU or ARC). Instead, for our candidate evaluation and reformation, we use the general dataset WikiText2, which represents generic verified and featured articles on Wikipedia and makes sure that our experiments remain actually zero-shot since no task-specific downstream data is seen during our search. As the model never sees any downstream data (so that it does not know how to be biased), it incurs less biased evaluation for downstream tasks. This setting is consistent with previous works such as FLAP and SliceGPT. To ensure a fair comparison, we use the same calibration dataset (i.e. WikiText2) for all methods including the baselines. We believe our evaluation is fair and comprehensive. Our method can achieve a better performance on language modeling and QA datasets. Furthermore, FLAP investigates the performance with different calibration datasets (see Appendix C.3 and Table 7 of FLAP). It observes that C4 or WikiText2 as the calibration data, results in a fluctuation of about ±1\% in average accuracy for zero-shot tasks, which is not significant. ### Weakness 2. MMLU performance To make a fair comparison, we follow the setting in existing works (such as FLAP and SliceGPT) to use the data from WikiText2 for the candidate evaluation. Besides, it is super efficient to compute the perplexity of a few samples compared with testing the accuracy on the whole test set. Following the suggestion, we demonstrate the performance on MMLU dataset for our method and the baseline FLAP in Table 9 in the global rebuttal. Our method can outperform the baseline in terms of accuracy on MMLU. As we mentioned in our response to Weakness 1, we use the general dataset WikiText2 for candidate evaluation, which represents generic text data from Wikipedia and makes sure that our experiments remain actually zero-shot since no task-specific downstream data is seen during our search. Thus, our model is less biased for different downstream tasks. ### Weakness 3. QA performance in Table 5 Following the suggestion, we further provide the results on the common sense reasoning datasets for the 50\% inheriting ratio in Table 8 provided in the global rebuttal. As observed, our method can lead to lower perplexity and higher downstream evaluation accuracy performance. ### Weakness 4. Speed evaluation Thanks for the suggestion. For the generation speed test, we feed sentences consisting of 64 tokens into the model and adopt float16 with KV cache. The flash attention is not activated. Our results in Figure 7 demonstrate that our smaller models lead to practical inference acceleration. We agree that the inference speed may be different under different evaluation settings. Following the suggestion, we further show the computation cost in GMACs in Table 10 below. Table 10 \| Inheriting Ratio \| 100% \| 90% \| 80% \| 70% \| 60% \| 50% \| \|-------------------------\|-------\|-------\|-------\|-------\|-------\|-------\| \| Computation Cost (GMACs)\| 424.02\| 377.89\| 333.49\| 293.29\| 249.28\| 211.88\| ### Question 1. Writing issues Yes, the reviewer’s understanding is correct, and "inheriting ratio" = 1 - "sparsity". We provide an explanation for the inheriting ratio in Line 116-117. ### Question 2. Annotation Thanks, we will add the annotation. ### Question 3. Clarification M denotes the number of heads in attention, P denotes the number of channels in MLP (intermediate size). Masks are binary, it should be {0,1} rather than R. We will make it more clear in the revision ### 3.a. Mask sharing The mask is shared for the query, key, value, and output projection within a single self-attention module, as well as for the up, gate, and down within a single MLP module, rather than being shared across different blocks. We discuss the mask sharing in detail in the above global rebuttal. ### 3.b. Align internal computations As discussed in Line 141-142, for multiple layers inside each module (self-attention or MLP), they have to use the same mask to ensure the aligned internal computations. For example, if the query and key have different masks and thus different/unaligned dimensions, they can not be multiplied elementwise and the attention operation can not be correctly performed. In Appendix A, we demonstrate the mechanism for the self-attention and MLP modules, highlighting that the layers inside the same module need to share the same mask to ensure successful computations. But layers in different modules or blocks do not need to share the same mask. We are sorry for any confusion and will make this more clear in the revision. ### 3.c. Joint optimization of masks They are jointly optimized to determine the initial masks as discussed in Line 147-149. Later during our search, all masks are also jointly optimized to satisfy the parameter count constraint. ### Question 4. Gamma $\gamma_{attn}^i$ denotes the inheriting ratio for the $i^{th}$ head in attention. Each attention module can have multiple heads. Thus the inheriting ratio for the attention module is a collection from multiple heads as the inheriting ratio for each head can be different. $\gamma_{mlp}$ denotes the inheriting ratio of the MLP module. It does not have the concepts of multiple heads like the attention module and we just have one single ratio for the whole MLP. ### Limitation: Search cost As we discussed in the global rebuttal, the search cost can be reduced to 20 epochs within 2 hours and is similar to those baselines. --- Rebuttal Comment 1.1: Title: Side comment Comment: Dear authors, Thank you for your response. Before we launch further discussions, I would like to ask an important question: Did you submit any global response, or any one-page pdf containng tables / figures? It is not visible to me at the moment. If yes, please contact the AC or PC to fix the systematic error. To be clear, this comment will not affect my ratings. I just want to make sure that I don't miss any further experiment results or responses from you. --- Rebuttal 2: Title: Global rebuttal invisible to reviewers Comment: Dear reviewer, We sincerely appreciate you for spending time reviewing and providing feedback on our paper. We have submitted one global rebuttal which includes important explanations and additional results following the instructions of one global rebuttal for the paper. However, we find that the global rebuttal is only visible to program chairs and authors. The reviewers are not able to read the global rebuttals. We just raised the problem to the committee and asked if they can fix this issue. Thanks, Authors	Summary: This paper introduces an architecture search method based on mask mutation and candidate evaluation to find a subnet with better performance in the LLM. An evolution-based algorithm is applied to globally search the subset with a special initialization from the evaluation of parameter importance. After the search from LLM, a reformation algorithm is proposed to rectify the weights in the inherited LLM. Experimental results show that the compressed LLM can achieve better results compared to the baselines. Strengths: 1. A novel method for searching a subnet in the LLM. The search of subnet in LLM is not well-explored, and it would benefit the community. 2. The overall approach is reasonable and sufficient. The entire method’s pipeline is very clear, and the design for each component is reasonable and complete. 3. The experimental results are comprehensive and show its benefits compared to the baselines (SliceGPT and LLM-Pruner). A large and steady improvement is observed. 4. The paper is well-organized and easy to follow. Weaknesses: 1. My main concern is about the extra cost of searching the subnetworks. Since it still takes around 5 hours to search for the architecture, it introduces extra cost for pruning the model. What if we take these 5 hours to post-train the compressed model? Would the results be better or worse than the subnetwork search from the LLMs? 2. Since the training-free structured pruning is not a very useful setting, post-training the LLM with an acceptable amount of resources is more realistic. Thus, compared to just pruning and searching the mode, a more realistic setting is also to train the pruned model. And I’m not sure whether the searched sub-LLM would be better than the pruned one. Some previous work to prune CNNs suggest that with sufficient training, magnitude pruning would be a better choice than any manually designed metrics. Thus, it would be better to examine the performance of the compressed LLM with post-training. Technical Quality: 3 Clarity: 3 Questions for Authors: 1. For Mask Sharing: why does the same module in different modules share the same mask (Line 141)? In line 147 - 149, from my understanding, the mask is not the same for blocks in different layers. Using different masks for different layers seems to still have the aligned internal computations. I’m not sure if I understand correctly in this part. 2. What about the results if you post-train the compressed models? What are the results if compared with LLM-Pruner? Confidence: 4 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: N/A Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thanks for the suggestions from the reviewer. ### Weakness 1. Search cost As we discussed in the global rebuttal, we can reduce the search epoch number from 50 to 20 for LLaMA-7B, which costs 2 hours and can still achieve better performance than other methods. For LLaMA-65B model, we only adopt 20 epochs for search, which costs 5 hours. Besides, compare to other methods, our search cost is similar to those baselines. ### Weakness 2. Post training performance To improve the performance, we adopt an efficient reformation method in Section 3.4 without the need of expensive training or finetuning. As shown in our Figure 6, our reformation can significantly improve the perplexity within a few minutes. We highlight that without training or finetuning, our method already outperforms baselines which are based on recovery training, such as LLM-Pruner and SliceGPT, as shown in our experimental results and Table 8 provided in global rebuttal. To further evaluate our performance with continual training, we train our searched compact models with LoRA following the same setting of recovery training in LLM-Pruner, which can be finished within another 3 hours for LLaMA-7B. The results are demonstrated in Table 8 provided in global rebuttal. With recovery training, our method can further improve the perplexity and accuracy performance. For more sparse models (50\% inheriting ratio) suffering from more significant accuracy loss, recovery training can lead to larger improvements. Recovery training can further improve the accuracy, requiring more data and computations. Different from recovery training, our reformation can efficiently and effectively lead to better performance than the baselines, with little data and computations. We agree that it is an attractive topic to investigate the pruned model performance with sufficient training. But for LLMs, it is still an open question how to define sufficient training, in terms of model architecture, data quality, data amount, GPUs, and so on. We leave this as our future research. ### Question 1. Mask sharing The weights in the same module of each block share the same mask, but different blocks do not share masks. Each block has its own masks, which are not shared with other blocks. The reviewer’s understanding is correct. We discuss mask sharing in detail in the global rebuttal. Line 141 shows that there are two masks for each block, meaning there will be $2k$ masks for $k$ blocks, as the masks of different blocks are not shared. ### Question 2. Fine-tuned compressed model compared to LLM-Pruner We show the results compared to the fine-tuned LLM-Pruner in Table 8 provided in the global rebuttal. Results show that our method can perform better than LLM-Pruner even without finetuning. --- Rebuttal Comment 1.1: Comment: Thanks for the detailed response. It solves my concern and I prefer to keep my score unchanged.	Summary: The paper introduces a new model architecture search method for LLMs that does not require additional training. The proposed architectures have structural sparsity and reach better performance than SOTA pruning baselines. Strengths: - The proposed method does not require additional training. - The output model has structural sparsity, which is useful for hardware. - The paper is clear and well-written. Weaknesses: - Why isn't there a comparison to SparseGPT? Technical Quality: 3 Clarity: 3 Questions for Authors: Why isn't there a comparison to SparseGPT? Confidence: 2 Soundness: 3 Presentation: 3 Contribution: 2 Limitations: The authors discuss limitations. Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thanks for the suggestions from the reviewer. ### Weakness & Question: Comparison with SparseGPT The pruning or search granularity of our method and SparseGPT are different, thus they should not be directly compared. Our method searches smaller compact models, which is more like the structural pruning method to remove rows and columns of weight matrices. Different from our structure search, SparseGPT is an irregular pruning method to remove weights without strong constraints (such as, the removed weights should be in the same rows or columns). Thus in the experiments, we compare with other structure pruning methods such as SliceGPT and FLAP, without comparing to SpareGPT. Furthermore, the baseline SliceGPT already provides certain comparisons with SparseGPT, as shown in Table 1 of SliceGPT. As mentioned above, it is hard to make a fair comparison between structure and irregular pruning. In Table 1 of SliceGPT, the sparsity of SparseGPT is 50\%, while the sparsity of SliceGPT is 25\% or 30\%. The results in their Table 1 can serve as a reference, and our method can outperform SliceGPT non-marginally. --- Rebuttal Comment 1.1: Title: response to rebuttal Comment: Thank you for the response. I maintain my score for acceptance.	Summary: This paper proposes a technique for searching for efficient LLM subnets for fast inference, while still attaining strong performance. The proposed technique involve a training-free pruning stage based on a genetic algorithm, followed by a ``weight rectification'' stage that improves the resulting subnet. Empirically, the proposed method outperforms existing LLM pruning baselines. Strengths: - Empirically, the proposed pruning technique appears to outperform a handful of existing pruning baselines, as shown in Figure 1 and Table 2. - The proposed technique appears to be general enough to work for a wide variety of existing LLMs, which appears to be a s hortcoming of existing LLM pruning techniques. - The proposed method is fairly simple in that it combines LLM pruning with a genetic algorithm for search. Weaknesses: - From the abstract, it is unclear what the actual proposed method does beyond standard neural network pruning techniques. - The term ``reformation algorithm'' is used throughout the paper, and it is unclear what this means until much later in t he text. The intro and abstract should clarify this. - The main evaluation is done in terms of sparsity levels, which makes sense for this particular type of search space, how ever, the authors should also empirically validate the efficiency of the search process compared to existing pruning techn iques. E.g., how do these pruning techniques compare, subject to the same budget? Moreover, it would be interesting to see how the proposed method performs as a function of the search budget. - The authors should compare to other techniques beyond pruning, such as quantization which may achieve similar inference speed and memory improvements with better overall performance. In general, it seems that LLM pruning techniques take a sub stantial performance hit even at fairly modest sparsity levels (I'm mainly referring to Table 7). Technical Quality: 2 Clarity: 1 Questions for Authors: - The authors mention in the limitations that the search cost scales with the model size. Can the authors characterize thi s a bit more? Do other pruning methods suffer from the same problem, and can the authors comment on what this scaling curv e actually looks like? - Why was the search budget set to 5 hours? Have the authors explored this choice empirically, and does this apply to all of the models and data settings that were evaluated? Confidence: 2 Soundness: 2 Presentation: 1 Contribution: 2 Limitations: The authors adequately address limitations in the conclusion. Flag For Ethics Review: ['No ethics review needed.'] Rating: 4 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thanks for the suggestions from the reviewer. ### Weakness 1. Difference from network pruning Our proposed search method significantly differs from standard neural network pruning techniques. Specifically, our method includes a search initialization process to identify the effective initial architecture and an evolution architecture search with special mask mutation and efficient candidate evaluation, which are not typical in standard weight pruning techniques. We are the first to incorporate the architecture search technique for LLMs to find efficient architectures. There are no other search baselines and thus we mainly compare with pruning works. Although our method significantly differs from traditional pruning, both our search and pruning can lead to efficient lightweight LLMs. ### Weakness 2. Reformation Thanks for the suggestion, we will explain more about the reformation in the abstract and introduction. ### Weakness 3. Search cost We highlight that our search cost is just similar to the baselines. Details can be found in the above global rebuttal. ### 3.a. Performance as a function of search cost. Figure 5 clearly demonstrates how the proposed method performs as a function of the search budget (number of epochs and thus search hours). As we discussed in the global rebuttal, we can reduce the search cost without significantly sacrificing the performance. Note that, with 20 epochs, for LLaMA-7B, we can still achieve 6.23 and 7.21 perplexity on WikiText2 under 90\% and 80\% inheriting ratios, respectively, which are better than all baselines shown in Table 2. ### 3.b. Lower cost for memory and data in the search Besides the training time on GPU to finish the search, our method achieves a better performance with less memory and data. (a) Compared to LLM-Pruner, which identifies weight importance using backward passes, our method relies solely on inference. This significantly reduces GPU memory consumption, enhancing efficiency. (b) Compared to the SliceGPT, it requires 1024 data samples for calibration. Our method requires significantly fewer calibration samples (128) for weight optimization after the search, leading to better performance on various datasets. ### Weakness 4. Quantization Quantization methods are not direct competitors but rather complementary to our search approach. Existing quantization methods can be applied to the subnets identified by our method, further enhancing acceleration through integer matrix multiplications. In practice, quantization methods require specialized computation kernels to handle integer matrix multiplications, which can introduce additional computational overhead and complexity in implementation, particularly during inference. This can pose challenges when deploying quantized models on diverse hardware platforms which may not support the quantized model. Our approach, on the other hand, maintains the original computation framework of LLMs, without the need for additional operations or specialized implementations. Regarding the performance hit, we would like to clarify that the results in Table 7 were generated with a sequence length of 128, rather than 2048, for sequence length ablation study. Given that the dense LLaMA-7B model achieves the perplexity of 12.62 at 128 sequence length, our method, with a 90\% ratio, results in the perplexity of 13.40. This indicates a relatively small performance gap. With a larger sequence length such as 2048, the performance of all methods can be better and we can still achieve the best performance as shown in Table 2 and Table 3. ### Question 1. Search cost and model size We discuss the search cost in detail at the global rebuttal. For LLaMA-7B, we still can achieve better performance than other methods when searching only with 20 epochs within 2 hours. For LLaMA-65B, we only adopt 20 epochs to get the subnets, which takes 5 hours. ### Question 2. Search cost with 5 hours As we mentioned in global rebuttal, for LLaMA-7B on one A100 GPU, it takes 5 hours for 50 epochs during the search process. The search already converges at 20 epochs as demonstrated in Figure 5, which only takes 2 hours and already achieves better performance than other methods. Besides, for large models such as LLaMA-65B, we only adopt 20 epochs for faster search, which takes just 5 hours.	Rebuttal 1: Rebuttal: We thank the reviewers for acknowledging that our work overcomes prior research limitations and benefits the community (Reviewer Xurt, BSTA), our method is novel, general, and high-performing (Reviewer n31G, Qp2U, Xurt, BSTA), our experiments are comprehensive (Reviewer Xurt), and our paper is well-written (Reviewer Qp2U, Xurt). ### 1. Search Cost We can reduce our search cost with similar performance. For the search cost, it usually takes only 20 epochs for the convergence, and more epochs beyond 20 lead to marginal improvements (see Figure 5). For LLaMA-7B, we run 50 epochs (5 hours) for better results. When searching for 20 epochs, the perplexity for LLaMA-7B with 10\% and 20\% sparsity on WikiText2 are 6.23 and 7.21, which are still better than all baselines in Table 2. Besides, for large models such as LLaMA-65B, we only adopt 20 epochs for faster search, which takes 5 hours. Besides, we highlight that our search cost is just similar to the baselines. For LLM-Pruner, it is very difficult to scale to larger models such as LLaMA-30B and LLaMA-65B, as it requires backward propagation to compute the gradients for identifying the weight importance, incurring much larger memory and computation cost. It only has the results for 7B models (such as LLaMA-7B, and Vicuna-7B) and LLaMA-13B. The 7B models require 3 hours to finish the compression. For SliceGPT, it requires 5 hours to finish the pruning and fine-tuning of LLaMA-2 70B, which is similar to ours with 5 hours of search for LLaMA-65B. SliceGPT relies on PCA (Principal Component Analysis) to identify and remove less significant weights, which computes the eigenvectors and eigenvalues of the covariance matrix of the signal matrix. For large models, this step is computationally demanding and time-consuming because it requires processing a substantial amount of data to compute these matrices accurately. Besides, this work notes that using double precision for eigenvector calculations in PCA is necessary to avoid numerical errors that can degrade performance. This choice further increases the computational load and time required for pruning. It is data-hungry and requires to use much more calibration samples (1024) while our method only uses 128 samples. ### 2. Mask Sharing As discussed in Line 124-128, the model has multiple blocks (i.e., layers of LLaMA model). Each block has two modules, the self-attention module and MLP module. Each module contains multiple weights, such as the query, key, value, output projection in the self-attention module and up, gate, down in MLP module (Line 124-128). Mask sharing means that different weights in the same module share the same mask. However, modules of different blocks do not necessarily share the same mask. Each block has two masks corresponding to its self-attention module and MLP module, respectively. Different blocks do not share masks, and they have their own masks, which are not necessarily the same. Just as Reviewer BTSA mentioned, Figure 2 middle shows that different blocks can have different masks/sparsity. Initially, different masks have the same inheriting ratios. But the masks can still be different although their sparsity are the same. After we begin our search, the inheriting ratios of different masks can be different, although they start from the same value. ### 3. Additional Results We further provide more results with lower inheriting ratios and fine-tuning in Table 8 as follows, Table 8 \| Method \| Inheriting Ratio \| Wiki PPL↓ \| PTB PPL↓ \| BoolQ \| PIQA \| HellaSwag \| WinoGrande \| ARC-e \| ARC-c \| OBQA \| Average Acc.↑ \| \|------------------\|------------------\|-----------\|----------\|-------\|------\|-----------\|------------\|-------\|-------\|------\|----------------\| \| LLM-Pruner \| 80% \| 11.97 \| 55.68 \| 59.39 \| 75.57\| 65.34 \| 61.33 \| 59.18 \| 37.12 \| 39.80\| 56.82 \| \| LLM-Pruner (LoRA)\| 80% \| 8.41 \| 41.78 \| 67.78 \| 76.36\| 68.14 \| 64.32 \| 63.41 \| 36.87 \| 40.10\| 59.57 \| \| Ours \| 80% \| 6.89 \| 36.06\| 70.98 \| 74.92\| 67.29 \| 64.64 \| 64.23 \| 36.52 \| 39.40\| 59.71 \| \| Ours (LoRA) \| 80% \| 6.57 \| 34.32 \| 71.19 \| 75.01\| 68.43 \| 64.12 \| 63.88 \| 35.44 \| 40.03\| 59.73 \| \| LLM-Pruner \| 50% \| 126.02 \| 460.71 \| 51.97 \| 60.23\| 35.89 \| 49.07 \| 32.76 \| 25.67 \| 34.91\| 41.50 \| \| LLM-Pruner (LoRA)\| 50% \| 51.56 \| 198.76 \| 60.98 \| 69.12\| 47.83 \| 55.78 \| 46.92 \| 28.56 \| 35.64\| 49.26 \| \| Ours \| 50% \| 15.48 \| 117.06\| 61.27 \| 69.13\| 51.18 \| 57.78 \| 52.16 \| 29.34 \| 35.45\| 50.79 \| \| Ours (LoRA) \| 50% \| 11.34 \| 83.48 \| 63.49 \| 73.03\| 56.93 \| 59.21 \| 48.87 \| 36.12 \| 36.23\| 53.41 \| We also provide the results of our method on MMLU dataset compared to FLAP in Table 9 as follows, Table 9 \| Inheriting Ratios \| LLaMA-7B FLAP \| LLaMA-7B Ours \| LLaMA-13B FLAP \| LLaMA-13B Ours \| LLaMA-30B FLAP \| LLaMA-30B Ours \| \|-------------------\|---------------\|---------------\|----------------\|----------------\|----------------\|----------------\| \| 100% \| 34.9 \| 34.9 \| 46.9 \| 46.9 \| 58.2 \| 58.2 \| \| 90% \| 31.3 \| 31.6 \| 40.3 \| 40.3 \| 52.8 \| 52.9 \| \| 80% \| 29.1 \| 29.2 \| 35.7 \| 36.2 \| 47.3 \| 47.6 \| \| 70% \| 26.8 \| 27.3 \| 33.4 \| 33.7 \| 41.5 \| 42.4 \| \| 60% \| 25.9 \| 26.3 \| 28.4 \| 29.5 \| 36.1 \| 37.4 \| \| 50% \| 24.2 \| 25.1 \| 27.9 \| 29.1 \| 34.5 \| 35.1 \|	NeurIPS_2024_submissions_huggingface	2,024	null	null	null	null	null	null	null	null
Controlling Multiple Errors Simultaneously with a PAC-Bayes Bound	Accept (poster)	Summary: The authors suggest an approach for controlling errors of various types (simultaneously) within a PAC-Bayes framework. They derive a high probability bound on the KL divergence between the empirical and distribution risk vectors. This generalizes earlier work in the binary case by Maurer, 2004 and Begin et al 2016. The authors propose a method for using this bound as a (differentiable) training objective via considering a linear combination of the elements of the risk vector and inverting the resulting binary KL divergence. Strengths: - The main result seems to be a creative and potentially useful extension of the PAC-Bayesian framework. - Results are generally introduced with some explanation, and given more context after the result, which improves readability. Weaknesses: - The framework considered would be better motivated by providing a specific example (ideally with an experiment) indicating the usefulness of the method (and illustrating why existing approaches with a union bound are insufficient). - The authors included the NeurIPS checklist from last year (and not from this year). - The proof of theorem 2 is very long and either an outline/proof sketch should be provided at the beginning, or the proof should be broken into lemmas/propositions to improve readability. Technical Quality: 4 Clarity: 3 Questions for Authors: - Is it really not possible to upper bound arbitrary linear combinations of different types of risk with a union bound? I would think that bounding finitely many would place linear constraints on other linear combinations, allowing at least some bounds to be derived. It isn’t immediately obvious to me how sharp the resulting bounds would be, or if indeed this approach leads to non-vacuous bounds, but I’d appreciate the authors commenting on this and adding some discussion on the topic if the answer to my question is that it is possible. - How would $\ell$ be chosen in practice when defining the objective function? It seems the whole purpose of the framework is to provide useful bounds simultaneously for many types of predictive errors. But it also seems the choice of $\ell$ will have a large impact on how well a method trained with proposed approach will perform on tasks that weight certain types of errors heavily. Confidence: 3 Soundness: 4 Presentation: 3 Contribution: 3 Limitations: - The authors acknowledge limited experiments as a limitation of the current work. I agree that this is a limitation. In particular, I wonder whether similar results can be obtained using held-out data, and whether or not the bounds derived this way are competitive with this approach. In the one experiment provided, half of MNIST was used to define the prior, while have was used for optimization of the bound. It seems possible that a similar breakdown of data between a training and validation set (trying directly via cross-entropy loss) might lead to similar model performance and tighter bounds on the risk (as in the cases considered in Foong et al). It isn’t immediately obvious whether it is easy to directly invert the multinomial CDF to derive tight bounds on various datasets as in the binary case using holdout data. But I’m curious to hear the authors’ thoughts on this approach. - The authors point out that the bound is limited to discrete cases. This seems reasonable, and there are many discrete problems where the bound might be useful. I think finding a concrete problem (dataset, with a plausible decision depending on balancing error types) where this bound might be useful would be more convincing to address this limitation than extending the work to continuous problems. Minor points: - Line 147: $S_{m,n}^{>0}$ isn’t the interior of $S_{m,n}$ in a topological sense, assuming $S_{m,n}$ is considered with the discrete topology, which seems the natural choice. Otherwise, I don’t know what is meant by interior. - Line 198: "Different different" Flag For Ethics Review: ['No ethics review needed.'] Rating: 7 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thank you so much for your thorough review and many insightful comments! We are especially glad you find our results creative and the paper well-written. You mention the paper would be better motivated by an explicit example showing the utility of our method above existing ones. We did not initially believe a comparison to existing methods could be made, since - existing methods bound simply the error rate (lumping all errors together), or (via a union bound), bound each error probability, whereas; - our method bounds the whole distribution of errors, allowing us to derive bounds on every linear combination of the error probabilities. However, your (and the other reviewers') insightful comments have clarified that with bounds on the individual error probabilities one can in fact derive bounds on linear combinations of them. Thank you for this astute observation, which enables direct comparison of our result to existing ones! Thus we conducted two quick experiments to demonstrate the utility of our method above existing results. First, as a quick illustration that our method can beat a union bound, we construct a synthetic example. We take the number of error types to be $M = 25$ (corresponding to one error type for each entry of a $5 \times 5$ confusion matrix), $\delta = 0.05$, and suppose that the empirical risk vector is $\mathbf{R}_S(Q) = (1/25, \dots, 1/25) \in \mathbb{R}^{25}$ and $KL(Q\|P) = 0$. We then sample $10^8$ points $\mathbf{p}$ uniformly from the simplex. By counting what proportion of the sampled $\mathbf{p}$ could be values of $\mathbf{R}_D(Q)$ compatible with the union bound and our bound respectively, we obtain the following estimates (with 95% confidence intervals) for the volumes of the two confidence regions: \| $M$ \| $m$ \| Vol. Union CR \| Vol. Our CR \| \|-------\|------\|--------------------------------\|----------------------------------\| \| \| 100 \| 0.3191 (0.3190, 0.3192) \| 0.1758 (0.1757, 0.1758) \| \| $5^2$ \| 300 \| 1.313e-3 (1.306e-3, 1.320e-3) \| 3.710e-4 (3.672e-4, 3.748e-4) \| \| \| 1000 \| 5.665e-8 (1.090e-8, 1.024e-7) \| 3.7336e-8 (2.422e-9, 7.225e-8) \| One can see that our confidence region has a smaller volume in all three cases tested (although in the third case the confidence intervals overlap). Second, we conducted a more realistic experiment using the HAM10000 dataset, a dataset of images of cancerous and non-cancerous skin marks. We obtain an empirical risk vector $\mathbf{R}_S(Q) = (0.7718, 0.0548, 0.1735)$ for error types $E_0 = $ correct, $E_1 = $ false negative, and $E_2 = $ false positive, by employing our Theorem 2 to do simultaneous minimisation of the Type I and Type II errors with a loss vector $\ell = (0, 1, 2)$, weighting Type II errors as twice as bad as Type I errors (more time would allow professional estimates of the loss vector to be obtained). Again, the results are positive, showing our confidence region to be smaller. \| $M$ \| $m$ \| Vol. Union CR \| Vol. Our CR \| \|-------\|------\|--------------------------------\|----------------------------------\| \| 3 \| 500 \| 0.03067 (0.03056, 0.03077) \| 0.02985 (0.02975, 0.02996)\| Another excellent observation you make is that the choice of the loss vector may have a large impact on the training procedure and thus the types of errors the final classifier is liable to make. Unfortunately the choice of a scalar metric to optimise cannot be avoided, the question is just which one to choose. We chose the scalar metric to be the "total risk" $\ell \cdot \mathbf{R}_D(Q)$ for a given $\ell$ because that seems the most natural; while one might be uncertain about the future cost of different error types (for example new medical knowledge may change the relative severity of a type II error in skin cancer diagnosis), one presumably has a rough idea. Our bound then in essence gives you the most assurance with respect to this loss vector and ones close to it. Nevertheless, one can always default to putting a uniform weighting across all error types (except correct classification, where you should pick 0). Do let us know if this makes sense! You raise the point of calculating a bound using held-out data and mention the very interesting work of Foong et al. It is indeed the case that PAC-Bayes bounds in the literature are usually looser than test set bounds. Given more time we will make this comparison; it is an important and frequently neglected one in the PAC-Bayes literature. Nevertheless, PAC-Bayes bounds, including ours, are worth pursuing for two reasons: - First, while the optimal test set bound is known (the Binomial tail bound), PAC-Bayes bounds are still an open research direction and the community has seen a dramatic improvement in the tightness of PAC-Bayes bounds. There is a hope that PAC-Bayes bounds will eventually beat test-set bounds, and we hope that our bound is a stepping stone along this path; by showing how to generalise existing results, it may provide a recipe to generalise any future tighter PAC-Bayes bounds. - Second, at least half the value of PAC-Bayes bounds lies in shedding light on generalisation rather than getting tight empirical bounds. For example the framework can help answer questions on sample complexity, as in "Pac-bayes, mac-bayes and conditional mutual information: Fast rate bounds that handle general VC classes" by (Grunwald, Steinke, Zakynthinou, 2021). Our result makes progress on the sample complexity of bounds on the the kl(), relating it to the number of error types. Finally, thank you for your two minor points! We are grateful that you read the paper so carefully and took the time to relay even these minor points to us, which we have now corrected. We hope we have managed to address all of your questions and that you may even consider raising your score. --- Rebuttal Comment 1.1: Comment: The authors have addressed the primary concerns raised in my review. I have increased my score by a point, as I think the paper should be accepted. --- Reply to Comment 1.1.1: Comment: We are glad we have addressed the important points you raised, and to read that you believe the paper should be accepted. Thank you again for taking the time to write such a thorough review, and for your suggestions and insights which we believe have allowed us to improve the paper!	Summary: The notion of a set of "error-types", introduced by this work, is a user-defined partition of the product space of predictions and responses, generalizing well-known summaries of the erring behavior of predictors. This work presents a PAC-Bayes bound on the divergence between and empirical and true distributions over a set of "error types" for a choice of posterior prediction scheme. The phrasing of the bound in terms of KL divergence naturally generates bounds on arbitrary linear combinations of the true error type probabilities, which hold simultaneously. The authors present a training objective founded on the minimization of the worst-case value of a pre-defined weighting of error types consistent with the empirical error-type profile of a given choice of posterior over predictors. Strengths: The paper strikes a good balance between practicality and theory. The problem setting is well-motivated, and the generalization of the confusion matrix through error types allows for the use of the bounds in a variety of non-standard settings. The PAC-Bayes bound is novel. While PAC-Bayes has been previously used to control the spectral norm between between empirical and true confusion matrices, the bounding distributions over error types in terms of KL divergence straightforwardly generates bounds on the expected cost (over the true distribution over error types) which hold simultaneously, something of notable practical use. The paper is generally well-written. Weaknesses: One of the main strengths of this work the simultaneous control of arbitrary linear combinations of the error type probabilities. However, the ability of the previous literature to generate some semblance of these guarantees is not particularly well-exposited. Having not previously read any of this line of work, I think an explicit example of how one can construct a bound on a single linear combination might do the paper some good (and might motivate this work past the simultaneous control of all linear combinations, given that the task seems somewhat tricky given a spectral norm bound). Technical Quality: 3 Clarity: 3 Questions for Authors: Is there a quick example of to control a single linear combination of confusion matrix probabilities using a previous bound? In line 331, what is an "empirically tighter" bound? Confidence: 3 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: Yes Flag For Ethics Review: ['No ethics review needed.'] Rating: 7 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thank you for your kind review, especially for noting that our approach is much more flexible than that of Morvant et al. (2012), as one is not limited to the confusion matrix and can instead consider arbitrary user-specified error types. As for your question on previous results, we have now run an experiment to compare our bound to that of Morvant et al. (2012). Since their result considers the full confusion matrix, a fair comparison requires requires choosing $M = L^2$ error types, where $L$ is the number of labels. To compare, we estimated the volume of our confidence region, which is a kl-ball, against the volume of Morvant's confidence region, which is a spectral-norm-ball. Our Monte Carlo estimates show that Morvant's result is either not applicable or their confidence region is much larger than ours and essentially takes up the entire simplex, hence the volume estimate of 1.000 in the below table. The reason their bound is sometimes inapplicable is because it requires every class to contain at least $8L$ instances (in the $L=5, M=25, m=100$ case this would require each class to contain at least $5 \times 8 = 40$ instances which is impossible with $m=100$ samples). Here the table of the results: \| $M$ \| $m$ \| Vol. Morvant CR \| Vol. Our CR \| \|-------\|------\|--------------------------------\|----------------------------------\| \| \| 100 \| NA ($m$ too small) \| 0.1758 (0.1757, 0.1758) \| \| $5^2$ \| 300 \| 1.00 (1.000, 1.000) \| 3.710e-4 (3.672e-4, 3.748e-4) \| \| \| 1000 \| 1.00 (1.000, 1.000) \| 3.7336e-8 (2.422e-9, 7.225e-8) \| This comparison of the volumes of our confidence region and Morvant's clearly shows that bounds on linear combinations of error probabilities derived from our bound will be much tighter that Morvant's, which will be almost as large as they theoretically can be and therefore provide no utility. We were surprised to see that Morvant's highly interesting theoretical bound unfortunately performs so poorly empirically. It is certainly a comparison we will add to the final paper if it is accepted, and we thank you for the very excellent advice which we believe improves the exposition of our result. On your advice, we are in the process of adding an explicit example of how to calculate a bound on a linear combination using our Theorem 2 and using Morvant's bound. We agree with your observation that this would clarify the import of our quite technical theorem. Once again we kindly thank you for taking the time to carefully read our paper and give highly constructive and helpful feedback! We believe this has improved the exposition of our paper and helped elucidate our contributions. --- Rebuttal Comment 1.1: Comment: Thanks for further elucidation via experiment. I retain my previous score. --- Rebuttal 2: Comment: We are pleased to read that you found the experiment clarifying! Thank you again for taking the time to write a careful review and for your insightful comments, which we believe have improved the work.	Summary: This paper introduces a novel PAC-Bayes bounds that extend classical kl-based bounds to vector-valued losses that can control several error-types simultaneously. The bound is converted into a differentiable minimization objective and details for the practical implementation of the bound are provided in the Appendix. Strengths: 1- The paper is clearly written and theoretically solid. 2- The contributions have a potential impact in scenarios where simultaneous control over multiple error types is needed. Weaknesses: 1- The proof techniques and the main contribution (Proposition 4) are straightforward extensions of scalar kl-bounds to the multidimensional case. I fear this contribution alone might not be significant enough. 2- The experiment in Section 7 is too simple, further empirical evaluation of the bounds (e.g. simultaneous minimization of Type I and Type II errors in more realistic scenarios) would be very beneficial for the impact of the paper. 3- Wu, Y. S., & Seldin, Y. (2022). Split-kl and PAC-Bayes-split-kl inequalities for ternary random variables. Advances in Neural Information Processing Systems, 35, 11369-11381. is significant related work not discussed in the paper. In general, I don't think this is a bad paper, but the theoretical contributions are not very original and I feel that the paper needs more experimental backup to be a well-rounded contribution. Technical Quality: 4 Clarity: 3 Questions for Authors: See weaknesses. Confidence: 3 Soundness: 4 Presentation: 3 Contribution: 2 Limitations: The limitations are properly discussed in Section 6. Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thank you kindly for taking the time to carefully review our paper. Thank you especially for pointing out the unfortunate omission of Wu, Y. S., & Seldin, Y. (2022) from our related work section! The paper is indeed very related and we have now incorporated a discussion of it into our paper. As for the main contributions of our paper, we emphasise that these are Theorems 1 and 2, not Proposition 4. We hope that you take this into consideration with your final score as we feel that Proposition 4 represents a small portion (20%?) of our theoretical contribution. Indeed, as you point out, the proof technique for Proposition 4 is similar to existing results and comes to one page, whereas the techniques required for Theorems 1 and 2 are extensive, novel, and require an additional 12 pages of proof, mostly found in the appendices. The meat of our paper is: - Theorem 1 This specialises Proposition 4 to the $d(\cdot, \cdot) = kl(\cdot∥\cdot)$ case and then slightly loosens it to a tractable form so that it can be easily evaluated. The proof is in two parts: - Proving Corollary 1 (lines 286 - 299) by specialising Proposition 4 to the $d(\cdot, \cdot) = kl(\cdot∥\cdot)$ and applying some combinatoric arguments; - Proving Theorem 1 (lines 304 - 316) by loosening the bound to a tractable form. Because we want to keep the bound as tight as possible, we expend some effort in loosening it as little as possible. For this require the technical Lemma 3, which itself requires two helping Lemmas found in Appendix 8.3. These three Lemmas require a further two pages to prove. We split this up into chunks for readability, but appreciate that this might obfuscate what the main contributions are. Hopefully this clears it up with respect to Theorem. - Theorem 2 This constructs a differentiable training objective from the bound given in Theorem 1. Its proof requires five pages and is found in Appendix 8.5 that does not follow the lines of any existing proof we are aware of. We hope you can agree that altogether this represents a non-trivial contribution! Thank you for highlighting, as we should have realised, that the binarised MNIST dataset we used is not particularly realistic to our stated use-case. On your advice, we conducted an experiment using the HAM10000 dataset, a dataset of images of cancerous and non-cancerous skin marks. We obtain an empirical risk vector $\mathbf{R}_S(Q) = (0.7718, 0.0548, 0.1735)$ for error types $E_0 = $ correct, $E_1 = $ false negative, and $E_2 = $ false positive, by employing our Theorem 2 to do simultaneous minimisation of the Type I and Type II errors with a loss vector $\ell = (0, 1, 2)$, weighting Type II errors as twice as bad as Type I errors (more time would allow professional estimates of the loss vector to be obtained). We will be able to write up the experiment in full in the camera-ready version if accepted, but the results are positive; a Monte Carlo estimate shows that the resulting Confidence Region for $\mathbf{R}_D(Q)$ takes up only $3\%$ of the simplex, which compares favourably to the Confidence Region formed by taking a union bound. Here is a table of results, showing the volume of the confidence regions normalised by the volume of the simplex (in brackets are the confidence intervals from the MC estimates): \| $M$ \| $m$ \| Vol. Union CR \| Vol. Our CR \| \|-------\|------\|--------------------------------\|----------------------------------\| \| 3 \| 500 \| 0.03067 (0.03056, 0.03077) \| 0.02985 (0.02975, 0.02996)\| Do please let us know whether you find the new setting of the HAM10000 dataset to be more realistic, and any further details you would like to know! Thank you again for investing the time to review our paper and provide very helpful feedback which we believe has improved the presentation of the paper and the convincingness of our results. --- Rebuttal Comment 1.1: Comment: Thank you for your clarifying comments, with my doubts regarding the theoretical contributions solved, having read the rest of the rebuttals and with more realistic experiments, I am happy to increase my rating to 6. Including these extra experiments in the camera-ready version will make for a solid contribution. --- Reply to Comment 1.1.1: Comment: We are glad that you found the additional experiment useful! Thank you for raising your score in light of this, and for your suggestions which we believe have improved the work.	Summary: In the standard statistical learning theory control of generalization error typically means studying deviations of the empirical risk from the risk (for instance, with high probability over the sample). In the language of statistics such control doesn't make a difference between type I & II errors. This paper goes beyond as it looks at a multi-objective risk, i.e. 'deviation' of the vector or M empirical risks from the corresponding risks. In the context of the paper, the vector of empirical risks lives on a probability simples and the paper shows how to control a KL divergence between such a vector and a risk probability vector. Then, the paper goes on and studies a PAC-Bayes formulation of this problem (Theorem 1). In section 4 the paper proposes how to construct a differentiable objective for this problem. Strengths: This is an important problem, especially in the context of multiobjective errors which are commonplace in practice. To some extent these results also extend results of Maurer (2004) (PAC-Bayes bounds on little-kl) to simpleces (while original results were shown for Bernoulli case). Weaknesses: The price for having such a multiobjective setting is M log(M), whereas the standard union bound would lead to a log(M) cost --- note that there is a (significant) gap. However, it should be noted that this would give a bound on a different objective (not kl() between points on the simplex). I think that the paper should argue why such kl() is interesting or at least design analytical instances and/or some experiments which would demonstrate that there exist instances where kl() leads to a better/tigther results (perhaps a performance on a downstream task?). Technical Quality: 3 Clarity: 3 Questions for Authors: Not a question really, but I'd be interested to hear authors' thoughts on the weakness against a union bound. E.g. how would one approach this comparison problem / experiment design. Confidence: 4 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: Seems to be adequate. Flag For Ethics Review: ['No ethics review needed.'] Rating: 5 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thank you for your thoughtful review and for taking the time to give constructive feedback on our work! We are especially grateful for your insight that it is difficult to compare our bound against the standard union bound since we have bounded a different quantity - the vector rather than scalar kl. The reason we did not initially make a comparison was because it seemed to us that the standard union bound was not suitable for our use-case; we are considering a situation such as outlined in paragraph 3 of the introduction, where one cannot anticipate the future costs of certain error types, such as the type I & type II errors you mention. In such a case, a union bound would simultaneously bound the type I & II errors individually, whereas our bound simultaneously bounds every linear combination of these two errors. This gives greater assurance when the costs of different error types may change (for example with better medical knowledge), and may be especially useful when there are more than two ways of mis-classifying data. Your and the other reviewers' insightful comments however, have helped elucidate the fact that a union bound applying to each error probability individually can indeed be used to bound linear combinations of errors, and thereby also constrains the entire vector of error probabilities. For this reason both our bound and a union bound form confidence regions around the empirical risk vector $\mathbf{R}_S(Q)$, and these regions can be compared. One way to do this is to compare their volume, where a smaller volume indicates greater control over the true risk vector $\mathbf{R}_D(Q)$. As a quick experiment, we take the number of error types to be $M = 25$ (corresponding to one error type for each entry of a $5 \times 5$ confusion matrix), $\delta = 0.05$, and suppose that the empirical risk vector is $\mathbf{R}_S(Q) = (1/25, \dots, 1/25) \in \mathbb{R}^{25}$ and $KL(Q\|P) = 0$. We then sample $10^8$ points $\mathbf{p}$ uniformly from the simplex. By counting what proportion of the sampled $\mathbf{p}$ could be values of $\mathbf{R}_D(Q)$ compatible with the union bound and our bound respectively, we obtain the following estimates (with 95% confidence intervals) for the volumes of the two confidence regions (normalised by the volume of the simplex): \| $M$ \| $m$ \| Vol. Union CR \| Vol. Our CR \| \|-------\|------\|--------------------------------\|----------------------------------\| \| \| 100 \| 0.3191 (0.3190, 0.3192) \| 0.1758 (0.1757, 0.1758) \| \| $5^2$ \| 300 \| 1.313e-3 (1.306e-3, 1.320e-3) \| 3.710e-4 (3.672e-4, 3.748e-4) \| \| \| 1000 \| 5.665e-8 (1.090e-8, 1.024e-7) \| 3.7336e-8 (2.422e-9, 7.225e-8) \| One can see that our confidence region has a smaller volume in all three cases tested (although in the third case the confidence intervals overlap). We can therefore see that the cost to the bound of Mlog(M), rather than log(M) for the union bound, can indeed be outweighed by the fact that we are bounding a different quantity. We thus outperform the union bound method. Thank you again for taking the time to carefully review our work, and giving us feedback we believe has enabled us to strengthen the presentation of our result! We hope that our demonstration that our bound on the kl() can lead to tighter bounds than the union bound approach convinces you of the utility of our result, and that if so you will kindly consider raising your score. --- Rebuttal Comment 1.1: Comment: Thanks for your explanations, and empirical evaluations which are indeed interesting. I guess, for the final revision it would be interesting to see a more complete empirical suite (even if it's synthetic) to see where the proposed approach "breaks" and the union bound becomes better. Otherwise it is hard to see a full picture. Overall, I think this is interesting and proofs have some new ideas, so I'm in favor of accepting this paper. --- Rebuttal 2: Comment: We are glad you found the evaluations interesting and are pleased that you are in favour of accepting the paper! We will indeed expand the evaluations in the final version if it is accepted. Thank you for your detailed review and suggestions, which we believe have improved the paper!	Rebuttal 1: Rebuttal: We genuinely thank all four reviewers for their many insightful comments, constructive feedback, and astute observations. We have responded to all four reviewers individually. We hope they let us know in the discussion phase if there is anything we should clarify, or any further results they would like to see, and we will do our best to get back to them promptly. Thank you again investing the time to write these thorough reviews.	NeurIPS_2024_submissions_huggingface	2,024	null	null	null	null	null	null	null	null
Unifying Generation and Prediction on Graphs with Latent Graph Diffusion	Accept (poster)	Summary: This paper proposes a Latent Graph Diffusion model that unifies multiple tasks such as classification, regression, and generation in a generative task. Additionally, this model achieves feature generation in multiple categories, including node-level, edge-level, and graph-level. Strengths: 1. This paper attempts to address a crucial task in the field of graphs, which is to propose a model capable of handling various tasks such as classification, regression, and generation at different levels, including node-level, edge-level, and graph-level. 2. The writing in this paper is relatively clear. Weaknesses: 1. The evaluations conducted by the authors are insufficient. For example, in the generation task, they avoid comparing their proposed method with classical diffusion-based graph generation models such as DiGress [1], HGGT [2], and DruM [3], which perform better in terms of metrics like Validity on QM9. Additionally, they avoid comparisons on larger datasets like MOSES [4] or GuacaMol [5]. Furthermore, they do not compare with commonly used metrics for graph generation tasks, such as FCD and NSPDK. 2. Regarding the approach of treating classification and regression tasks as generation tasks, the authors should provide a simple baseline (which can be seen as an ablation study) to demonstrate that they have not unnecessarily complicated simple and intuitive tasks. For example, they could use condition features as input and train a baseline prediction model with the same architecture as the denoising model, using the corresponding masked global features as supervised signals. This would help assess the benefits brought by the generation model. 3. The authors aim to develop a model that can address graph learning tasks at all levels (node, edge, and graph) and types (generation, regression, and classification), which is hard to achieve. The authors are suggested to provide a clear table to include detailed information about the source and size of the training dataset, as well as the number of model architecture and parameters for reference. 4. The explanation of how the decoder of generation tasks reconstructs the structural information, i.e., A, of the graph is not clear. 5. Figure 1, mentioned in the Section 5, is displayed in the Appendix. [1] C. Vignac, I. Krawczuk, A. Siraudin, B. Wang, V. Cevher, and P. Frossard. DiGress: Discrete denoising diffusion for graph generation. In International Conference on Learning Representations, 2022. [2] Y. Jang, D. Kim, and S. Ahn. Graph generation with K2-trees. In The Twelfth International Conference on Learning Representations, 2023. [3] Jaehyeong Jo, Dongki Kim, and Sung Ju Hwang. Graph generation with diffusion mixture., ICML 2024. [4] D. Polykovskiy, A. Zhebrak, B. Sanchez-Lengeling, S. Golovanov, O. Tatanov, S. Belyaev, R. Kurbanov, A. Artamonov, V. Aladinskiy, M. Veselov, et al. Molecular sets (moses): a benchmarking platform for molecular generation models. Frontiers in pharmacology, 11:565644, 2020. [5] N. Brown, M. Fiscato, M. H. Segler, and A. C. Vaucher. Guacamol: benchmarking models for de novo molecular design. Journal of chemical information and modeling, 59(3):1096–1108, 2019. Technical Quality: 2 Clarity: 3 Questions for Authors: 1. Whether the Encoder and Decoder are permutation equivariant? 2. Will the authors release the source code and checkpoints? Confidence: 4 Soundness: 2 Presentation: 3 Contribution: 2 Limitations: Yes. Flag For Ethics Review: ['No ethics review needed.'] Rating: 5 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We thank the reviewer for the useful comments. We now address your concerns as follows. 1. Experiments (weakness 1). * Originally, LGD is designed to be a single framework that can do both generation and regression/classification, where we put our experimental focus on demonstrating that latent diffusion models can achieve state-of-the-art regression/classification results. Now, we have rerun our generation experiments on QM9 and (i) compare our results with state-of-the-art diffusion-based graph generation methods including DiGress [1], HGGT [2] and GruM [3]; (ii) use metrics including FCD and NSPDK to better evaluate the quality of generation. In our new experiments, we train a more powerful encoder/decoder by incorporating positional encodings and more complex reconstruction tasks to obtain better latent graph representations, and carefully choose the hyper-parameters of the diffusion model. The results are shown in Table 1 of the pdf in the global response. It is remarkable that our new results are highly competitive or superior than the baselines in nearly all metrics, which reveals the strong generation ability of LGD. Especially, LGD achieves a better FCD score (in favor of generating graphs similar to the training data) than the three strong recent baselines [1,2,3] while sacrificing the least novelty. * We are conducting new experiments on the larger dataset MOSES. Since the time is limited, the results are not yet ready. We will update the results as soon as possible, using comments and the link in the global response. We promise to include the new results in our camera-ready paper. It is also notable that most recent studies including HGGT [2] and GruM [3] also do not include these datasets at scale. 2. Clarifications on treating classification and regression tasks as generation tasks (weakness 2). Actually, in Appendix B of our paper, we already provided solid theoretical analysis and clarifications on why diffusion models can perform better than traditional deterministic models in classification and regression tasks. In detail, Theorem B.4. provides error bounds of diffusion models on regression tasks, equation (52) explicitly states the condition where diffusion models outperform regression models, and Corollary B.6. provides guarantees that there exists at least one latent diffusion model whose MAE is not larger than the autoencoder. Combining all these pieces, we can use arbitrary existing prediction models as the autoencoder, and the diffusion model trained on top this autoencoder is able to perform better than this prediction model. To empirically verify this, we utilize the ZINC dataset and train a regression model with our proposed graph attention architecture (the same as the denoising model) using the masked global features (regression target) as the supervising signal. This prediction model has $0.084\pm 0.004$ test MAE on ZINC, while the diffusion model trained upon the autoencoder with the same architecture achieves $0.065\pm0.003$ MAE. This results verifies that we are not complicating simple tasks---instead, diffusion models can provably outperform traditional prediction models. 3. Clarification on training (weakness 3). First of all, LGD is a unified framework (and not necessarily a single model) for all levels of graph learning tasks. As explained in both the limitation section and the experiment section of our paper, we currently train our models separately for each task. However, we are indeed able to train a single model for all tasks using LGD. For example, we can use OFA [4] as the autoencoder, which is a single model that embeds graphs from various domains into a unified latent space through a powerful LLM-GNN mixture architecture. We can then train a diffusion model in this latent space to tackle tasks of all levels and all types, utilizing the unified formulation in our paper. We believe the unified formulation of our paper is a significant contribution and a necessary stage towards graph foundation models. As for the model architectures, details are provided in Appendix D.1 and D.3.1. As a reference, the total number of parameters of LGD (including both the autoencoders and the diffusion model) is 1,825,481 for the regression task on ZINC, and 6,605,545 for the generation task on QM9. 4. Decoding graph structure (weakness 4). As clearly explained in our main text, the latent space contains both the structural information and edge features, since the encoder treats absence of an edge as a special feature. We have a latent representation $W\in \mathbb R^{n\times n\times d}$ for adjacency matrix $A$ (line 145), which is exactly the diffusion target. After obtaining the generated latent edge-augmented representations $\hat W\in \mathbb R^{n\times n\times d}$, we directly apply the pretrained decoder (which is actually a linear layer) to reconstruct both the graph structure ($A$) and edge features. In summary, LGD is a one-shot generation method that generates graph structures and features simultaneously. 5. Figure (weakness 5). Figure 1 is presented in the Appendix due to limited space. We will move the figure to the main context in the camera ready version. 6. Equivariance (question 1). The encoder and decoder are indeed permutation equivariant. We can either use existing models such as MPNN and graph Transformers, or use our specially designed graph attention mechanisms as encoder/decoder. All of the above models are permutation equivariant. Actually, we also use MPNN and graph transformers for the denoising network, making our diffusion model also permutation equivariant (in terms of each denoising step). 7. Source code (question 2). We provide our code through the anonymous link in the global response. We will definitely release all the source code and checkpoints in the final version. --- Rebuttal 2: Comment: References [1] C. Vignac, I. Krawczuk, A. Siraudin, B. Wang, V. Cevher, and P. Frossard. DiGress: Discrete denoising diffusion for graph generation. In International Conference on Learning Representations, 2022. [2] Y. Jang, D. Kim, and S. Ahn. Graph generation with K2-trees. In The Twelfth International Conference on Learning Representations, 2023. [3] Jaehyeong Jo, Dongki Kim, and Sung Ju Hwang. Graph generation with diffusion mixture., ICML 2024. [4] Liu, H., Feng, J., Kong, L., Liang, N., Tao, D., Chen, Y., and Zhang, M. One for all: Towards training one graph model for all classification tasks. In The Twelfth International Conference on Learning Representations, 2024 --- Rebuttal 3: Title: We are looking forward to your reply Comment: Dear reviewer, While we have addressed all your concerns and answered all your questions, we are looking forward to your reply. As shown in the global response, we rerun the experiments on QM9 and introduce FCD and NSPDK metrics, showing that LGD is competitive with all the strong baselines you mentioned in terms of all metrics. We also conduct experiments on large scale dataset MOSES and show that LGD outperforms DiGress. The answers to your other questions are included as well in our response. While other reviewers have replied positively, we are looking forward to your reply and we are happy to answer any further question you may have. Thank you. --- Rebuttal Comment 3.1: Title: thank you for your rebuttal Comment: Dear authors, thank you for your response. After reviewing the additional experimental results, I decided to increase my score to 5. --- Reply to Comment 3.1.1: Title: Thank you for your reply Comment: Dear reviewer, we sincerely appreciate your constructive comments and kind reply. We believe that this paper has great contribution to the graph learning and the machine learning community.	Summary: The paper proposes Latent Graph Diffusion to generate node, edge, and graph-level features to meet the need of different tasks under the unified framework. Extensive experiments show the competitive performance across various graph-based tasks. Strengths: 1. It is good to reformulate regression and classification problems with diffusion generation task. 2. The paper is technically sound, and the introduction of diffusion in latent space is well-motivated. 3. The unified framework has theoretical support. Weaknesses: 1. Readers have to jump to Appendix over and over again to have a general understanding of your proposed model. 2. Despite the unified formulation as generation, other parts of the paper lack novelty. For example, there is a lack of new breakthroughs in specific model design and diffusion models. The underlying diffusion and denoising models largely build upon existing methods. 3. No code or demo available. Technical Quality: 3 Clarity: 2 Questions for Authors: 1. How does the LGD framework scale with very large graphs? 2. In the task of node classification from Planetoid, is the LGD model capable of generating a new graph almost the same as Cora or PubMed? 3. The graph will be too dense when considering no link as a special type of edge? How to deal with this problem Confidence: 4 Soundness: 3 Presentation: 2 Contribution: 2 Limitations: 1. The lack of detailed analysis of time and space complexity makes the performance of the model in practical applications less transparent. 2. The sensitivity analysis of parameters in the conditional generation task is insufficient, especially since the influence of important hyperparameters is not studied deeply. Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We thank the reviewer for the comments and acknowledgement. 1. Writing (weakness 1). We will refine the writing of the paper and provide as much important information as possible in the main text. 2. Novelty (weakness 2). We believe that our unified formulation is novel and beneficial. Regarding model design, we design novel self-attention and cross-attention mechanisms for graphs with augmented edges, as shown in Appendix D.1. For the diffusion model, we explained in our main text that we use the latent diffusion to tackle the problems brought by the discrete structure, feature misalignment and task variety. It is novel to use a latent diffusion model to perform graph learning tasks of all levels and all types. Furthermore, while architecture and diffusion model design are not our main focus, our LGD framework is extremely flexible, being able to incorporate various model designs and diffusion processes (e.g. DDPM, DDIM, Rectified flow etc.). 3. Source code (weakness 3). We provide the code through the anonymous link in the global response. We will definitely release all the code and checkpoints in the camera-ready version. 4. Scalability of LGD (question 1 and 3). We have explained the scalabiltiy of LGD in multiple places in appendix, such as architecture design in Appendix D.3, theoretical complexity in Appendix E.2, and experimental verifications in Appendix D.2 (line 887-891). We do not have to always model the dense graph where no link is a special type of edge---actually when we do not need to generate structures, we can utilize the sparse graph and only consider the real edges. We can thus use efficient MPNN based models to scale LGD to extremely large graphs such as OGBN-Arxiv, see Table 3 in the paper. 5. Node classification (question 2). In classification and regression tasks, generating the entire graphs is not our main focus---LGD mainly aims to generate the class labels of nodes. However, since LGD actually refines the representation in the latent space, it is still able generate the whole graph. LGD outputs the refined representation, which contains the information of both input (graph structures and known features) and class labels to be predicted, thus it could generate new graphs similar to the training sets. --- Rebuttal Comment 1.1: Comment: Thanks to the author for addressing my questions. I have to say that generating the refined representation is one thing, and generating new graphs (graph structure and node features) is one another. Anyway, I will maintain my positive score. --- Reply to Comment 1.1.1: Title: Thank you for your reply Comment: We sincerely thank the reviewer for the positive reply. Regarding the question, it is true that generating refined representation and generating new graphs are different things. However, since the target of node classification is not generating new graphs, the models are trained in a conditional generation manner, which is different from unconditional sampling for generating new graphs. We believe the capability of LGD to generate new graphs has been verified in unconditional generation experiments presented in the original submission and our rebuttal. Please inform us if you have any questions.	Summary: This paper proposes a model framework, LGD, which addresses multiple types of graph tasks and can simultaneously handle both generative and predictive tasks. Specifically, the LGD framework employs a method similar to stable diffusion, using a pretrained graph encoder to obtain latent representations and then training a diffusion model in the latent space. To use the diffusion model for predictive tasks, the authors incorporate the label in regression and classification tasks as a partially masked feature, using the rest of the training set as a condition to predict the masked label. Finally, the authors validate their approach through experiments for prediction, classification, and regression tasks. Strengths: - The idea of using a conditional diffusion model for predictive tasks is novel. Unlike traditional representation learning approaches, LGD can directly generate labels using the diffusion model, providing a new perspective for the community. - The experiments cover a comprehensive range of tasks, datasets, and baselines. The model shows promising results on most datasets. - The authors provide theoretical support for using latent diffusion models in regression tasks and discuss the advantages of converting deterministic tasks into generation tasks. Weaknesses: - The model functionality described in the introduction is somewhat misleading. In lines 28-29, the authors claim to construct a graph model capable of solving all task types. However, the implementation does not use a single foundation model to accomplish all tasks. Since the proposed framework still cannot solve the feature misalignment problem, specific tasks require training specific encoders and decoders. This makes the introduction seem like an overclaim, as LGD is just a framework that can handle multiple tasks but includes many components that need task-specific training. From a training cost perspective, this is similar with using separate models for different tasks (and potentially more expensive due to the need for GT and encoder-decoder). It is noteworthy that the authors mention this drawback in the limitations section. - The experiments lack details on training and potential unfairness. (1) In Appendix C, the authors describe that the encoder $\mathcal{E}$ was pretrained with both unsupervised and supervised methods. However, the main text does not specify which datasets were used for the encoder's training and whether they are the same as those used for training the diffusion model or if additional datasets were used. For instance, in the Conditional generation task, if the encoder was supervised-trained on the QM9 dataset, the results in Table 2 would lack fairness. This is because other baselines do not have access to supervised training information when providing conditions, giving the encoder a clear advantage. The authors should clarify the training data details for the encoder. (2) In line 762, the authors state that the latent diffusion model is shared across different tasks, but the paper does not provide details on what data this shared diffusion model was trained on. Was it trained on the entire dataset across all tasks? This also lacks detailed description. - Some of the authors' statements are overclaimed or inaccurate: (1) In the analysis of the unconditional generation task (QM9), the authors claim that generation based on 3D information is easier, highlighting their method's effectiveness. This premise is incorrect. First, RNN-based models [1], which generate based on SMILES, can achieve a validity of 99% and high uniqueness without 3D information. 3D generation actually targets a different need—conformation generation—related to molecular dynamics and has broader research and practical value, hence the use of 3D information. It is not that generation based on 3D information is easier. Moreover, the three metrics the authors compare are not the best indicators of molecular generation quality. I suggest the authors use more effective metrics such as QED and TPSA. Previous works, especially those focusing on molecular generation tasks, do not limit evaluations to these two metrics due to their limited persuasiveness. (2) In line 767 of the appendix, the authors claim their work is the first to achieve in-context learning on graph pretrained models. In fact, there are several prior works on graph prompts, such as PRODIGY [2]. [1] Molecular de-novo design through deep reinforcement learning. Journal of Cheminformatics. [2] PRODIGY: Enabling In-context Learning Over Graphs. NeurIPS 2023. - The discussion in lines 111-112 about the capabilities of generative models lacks citations. Technical Quality: 3 Clarity: 3 Questions for Authors: - Did the authors use a shared pretrained diffusion model for different tasks? If so, what data was it trained on? - What do the authors think about the feature misalignment problem in graph pretraining? Confidence: 4 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: Please refer to the Weakness and Questions above. Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We sincerely thank the reviewers for the constructive comments and suggestions. We address your concerns as follows. 1. Clarification on the introduction section (weakness 1). The introduction section mainly states the motivation to build a graph foundation model, but we do not claim that we have already done so (even in lines 28-29). We will surely rewrite the paper to avoid misunderstanding. Actually, LGD is indeed a unified framework to handle multiple tasks. Although we do not train a single model currently (as mentioned in our paper), we can indeed build a single graph foundation model under the LGD framework. For example, we can use OFA [1] as the autoencoder, which is a single model that embeds graphs from various domains into a unified latent space through a powerful LLM-GNN mixture architecture. We can then train a diffusion model in this latent space to tackle tasks of all levels and all types, utilizing the unified formulation in our paper. We believe the unified formulation of our paper is a significant step and a necessary stage towards graph foundation models. 2. Clarification on experiments (weakness 2). Appendix C actually describes the potential to build a complete pretrain, finetune and in-context learning framework using LGD. The contents are mainly discussions on future work about LGD, where we have not fully implemented them. We will move Appendix C to the discussion and future work section in the revised version of our paper and make it clearer. * For your question 2.(1) regarding the conditional generation task on QM9, the encoder is pretrained through the reconstruction loss of nodes and edges as described in the main context, without targets as supervision signals. The diffusion model still generates the molecules, and we follow the prior practice to use another pretrained model to evaluate the properties of all generated molecules, so our method does not introduce any unfairness. In addition, we would like to emphasize that even the encoders are trained in a supervised manner, it still does not introduce unfairness. Denote the conditions (properties) as $y$ and the molecules as $x$, there is a ground truth distribution $p(x\|y)$. In training we sample $x\sim p(x\|y)$ and denote it as a probabilistic mapping $f(y)$. The traditional conditional generation methods model $\hat x:=\theta(y)$ with a supervision signal $f(y)$. In comparison, our latent diffusion model learns the distribution of latent representation $\mathcal E(\hat x, \hat y):=\theta(y)$ with a supervision signal $\mathcal E(x, y)=\mathcal E(f(y), y)$. We can observe that both schemes include a supervision signal containing $y$ for generative models, except that we use an encoder to better represent the desired property. This is actually an advantage: utilizing representation learning, latent diffusion can better learns conditional distribution without introducing unfairness. * For your questions 2.(2), we train a separate encoder/decoder and diffusion model for every dataset for now. However, encoders and diffusion models trained across multiple datasets may have better performance and generalization, which we leave as future work. 3. Clarification on some statements. We thank the reviewer for the advice. There are indeed some statements that can be refined - we will correct them in the revision. (1) It is true that 3D generation is not necessarily easier. Our recent experiments also find that 3D generation can sometimes be harder due to the difference in data distribution and internal network architectures. We will correct related discussions in the paper. We also acknowledge the practical value of 3D generation and leave it as future work. Regarding the evaluation metrics, we additionally introduce the persuasive FCD and NSPDK metrics in our new experiments for better evaluation, and LGD achieves highly competitive performance. We do not include QED and TPSA for the QM9 dataset, since drug-likeliness is not suitable for the small molecules in QM9. (2) We are aware of PRODIGY and did not mention it since it works in a slightly different manner compared with the in-context learning in NLP: PRODIGY works by measuring the "similarity" between test and context examples, which is different from the conditional generation as in NLP. However, it is true that PRODIGY is also one sort of ICL, and we will make sure to include this citation and discussion in the revised version. 4. We will add the discussion concerning generalization ability of generative models, including LLM (Llama-3 and GPT-4 etc.) and diffusion models [2-3]. 5. Answers to questions. (1) We train diffusion models separately as explained in 2. (2) Feature misalignment is indeed one of the obstacles to build a graph foundation model. LGD provides the potential simultaneously perform generation and regression/classification as long as graphs with different features are mapped to a unified latent space. Actually, OFA [1] is a single model that embeds graphs from various domains into a unified latent space through a powerful LLM-GNN mixture architecture. Combining LGD and OFA, we can then train a generative model in the latent space to overcome the difficulty of feature misalignment and task unification problems, which is left for future work. Overall, we believe that LGD is a significant contribution to the community and could inspire future work on graph foundation models. [1] Liu, H., Feng, J., Kong, L., Liang, N., Tao, D., Chen, Y., and Zhang, M. One for all: Towards training one graph model for all classification tasks. ICLR, 2024. [2] Zahra Kadkhodaie, Florentin Guth, Eero P Simoncelli, and Stephane Mallat. Generalization in diffusion models arises from geometry-adaptive harmonic representations. ICLR, 2024. [3] Puheng Li, Zhong Li, Huishuai Zhang, and Jiang Bian. On the generalization properties of diffusion models. NeurIPS, 2023. --- Rebuttal Comment 1.1: Comment: Thanks for the response. I've read through other reviewers' feedback and responses as well. Most of my concerns have been addressed. Therefore, I will increase my rating to `6`. The authors should implement the revision promised in their response. --- Reply to Comment 1.1.1: Title: Thank you for your reply Comment: We sincerely thank the reviewer for the kind reply. We will definitely include these results in the revision.	Summary: This paper presents Latent Graph Diffusion (LGD), a graph generation framework adept at handling a variety of tasks including generation, regression, and classification at different objects—node, edge, and graph levels. LGD approaches regression and classification tasks as conditional generation challenges and employs a latent diffusion model within the latent space to concurrently generate features for nodes, edges, and graphs. Strengths: 1. The paper introduces the Latent Graph Diffusion as a novel approach to address the challenges of using a unified framework to achieve classification, regression and generation tasks on graph. 2. This paper clearly shows its idea. Weaknesses: 1. While GDSS and DiGress are designed to handle various types of graph data beyond molecular diffusion, it is unclear if LGD has conducted similar graph generation experiments. The article lacks relevant general graph generation experiments and evaluations using metrics like MMD (Maximum Mean Discrepancy) to assess the effectiveness of graph generation. 2. The authors have not included some successful practices of diffusion models. Recent studies [1,2] have demonstrated better performance than the proposed method on the QM9 dataset. 3. The QM9 dataset used for graph generation is relatively small, and it would be beneficial for the authors to validate their proposed model on larger benchmark datasets, such as Moses [3], which also provide a wider range of evaluation metrics. 4. The description of the model's architecture details is unclear and could benefit from further clarification. 5. The submission does not include an open-source implementation of the proposed model. [1] DiGress: Discrete denoising diffusion for graph generation. ICLR2022. [2] Graph generation with diffusion mixture., ICML 2024. [3] Molecular sets (moses): a benchmarking platform for molecular generation models. 2020. Technical Quality: 2 Clarity: 3 Questions for Authors: 1. It would be valuable to compare the computational efficiency of the evaluated methods during training and sampling to provide a comprehensive analysis. Confidence: 3 Soundness: 2 Presentation: 3 Contribution: 2 Limitations: N/A. Flag For Ethics Review: ['No ethics review needed.'] Rating: 5 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We thank the reviewer for the useful comments and suggestions. As presented in the global response, we have conducted extensive new experiments to improve the paper. Here we address your concerns as follows. 1. General graph generation (weakness 1). As explained in our paper, LGD is capable of generating arbitrary graphs through the powerful continuous latent space, hence is able to handle various types of graph data including generic and synthetic graphs. We conduct generic graph generation experiments and report the results in the global response. We use SBM, Planar and Community-20 datasets for evaluation. LGD performs better than most traditional graph generation methods on these datasets (although it is not always the best). However, other methods often have hard set rules or structural information in the diffusion process, while LGD performs a pure diffusion process in the Euclidean latent space without input structural information, requiring the encoder to be powerful enough to encode structural information. While MPNN and graph transformers are known to have limited expressivity, they may fail to encode enough structural information like degree, clusters, orbits and eigenvalues. Moreover, the autoencoder is trained through reconstruction loss, which has limited supervision on high-order structures. Therefore, if expressive higher-order GNNs or proper structural encodings are used, or more pretraining tasks related to complex structures are adopt, we will be able to obtain a more powerful encoder and thus better generation qualities, which we leave as future work. Please note that LGD has superior performance in real-world tasks where node/edge features are often more important than high-order structures, making it easier to obtain more powerful and suitable encoders for diffusion. 2. Baselines and metrics (weakness 2). We rerun our experiments on QM9 and compare our results with state-of-the-art methods including DiGress [1], HGGT [2] and GruM [3]. We also include the FCD and NSPDK metrics to better evaluate the quality of generation. The results are shown in Table 1 of the pdf in the global response. In our new experiments, we train a more powerful encoder/decoder by incorporating positional encodings and more complex reconstruction tasks to obtain better latent graph representations, and carefully choose the hyper-parameters of the diffusion model. It is remarkable that our new results now are highly competitive or superior than the baselines in nearly all metrics. Especially, it achieves a better FCD score (in favor of generating graphs similar to the training data) than the three strong recent baselines [1,2,3] while sacrificing the least novelty. 3. Larger datasets (weakness 3). We are conducting new experiments on the larger benchmark dataset MOSES. The results are not yet ready due to the limited time. We will update the results as soon as possible, using comments and the link in the global response. We will include the new results in our revised paper. It is also notable that most recent studies including HGGT [2] and GruM [3] also do not include these datasets at scale. 4. Model architecture details (weakness 4). Actually, we have included our model architecture details in the appendix of our submission. Please refer to Appendix D.1 for our architecture design and implementation details, and Appendix D.3.1 for the detailed architectures utilized in each experiment. 5. Open-source implementation (weakness 5). We provide the code through the anonymous link in the global response. We will definitely include the open-source code in the final version of our paper. 6. Computation efficiency (question 1). We have actually included comprehensive analysis on computation efficiency in Appendix E.2 of our submission. In summary, the complexity of LGD depends on the specific architecture of the autoencoder and the denoising network, which is generally $O(n^2)$ for generation tasks and $O(n)$ for regression/classification tasks. The training stage of the diffusion model is fast, and the inference stage can also be accelerated through various methods (such as DDIM, which we already implemented). As a reference, training LGD requires about $0.2$s/epoch on Cora and about $16$s/epoch on ZINC. [1] C. Vignac, I. Krawczuk, A. Siraudin, B. Wang, V. Cevher, and P. Frossard. DiGress: Discrete denoising diffusion for graph generation. In International Conference on Learning Representations, 2022. [2] Y. Jang, D. Kim, and S. Ahn. Graph generation with K2-trees. In The Twelfth International Conference on Learning Representations, 2023. [3] Jaehyeong Jo, Dongki Kim, and Sung Ju Hwang. Graph generation with diffusion mixture., ICML 2024. --- Rebuttal Comment 1.1: Title: Thank you for your response. Comment: After reading the response, I decide to raise my score. The additional result should be included in the revised version. --- Reply to Comment 1.1.1: Title: Thank you for your considered review and kind reply Comment: We sincerely appreciate the reviewer for the kind reply. We will surely include these new results in the revised version.	Rebuttal 1: Rebuttal: We sincerely thank the reviewers for their insightful comments and suggestions. To address the concerns, we conduct new experiments as follows. * For QM9 generation task, we (i) add more persuasive evaluation metrics including FCD and NSPDK to better evaluate the generation quality; (ii) take recent state-of-the-art baselines into account; and (iii) rerun our experiments with better configurations. In detail, we pretrain a more powerful encoder which takes nove-level and edge-level positional encodings (PE) as inputs, which is trained through more complex reconstruction tasks (decoding bond type with two connected atoms and decoding atom type with connected bonds). We then train a larger diffusion model with carefully searched hyper-parameters. Our new model achieves highly competitive and even state-of-the-art performance in all metrics except novelty. The convincing results validate the strong generation capability of LGD. Detailed results are shown in the pdf. * We conduct generic graph generation experiments and report the results in the attached PDF. We use SBM, Planar and Community-20 datasets for evaluation. LGD performs better than most traditional graph generation methods on these datasets (although it is not always the best). However, other methods often have hard-set rules or additional structural information in the generation process, while LGD performs a pure diffusion process in the Euclidean latent space without explicitly input structures or hard rules, which requires the encoder to be very powerful to distinguish subtle structural differences. Since MPNN and graph transformers are known to have limited expressivity, they may fail to encode some structural information like clusters, orbits and eigenvalues, resulting in lower generation quality in pure-structure tasks. Please note that LGD has superior performance in real-world tasks where node/edge features are often as important as structures, making it easier to obtain more powerful and suitable encoders for diffusion. Overall, LGD is designed to handle various tasks on real-world graphs as shown by the other parts of our experiments. * We are conducting experiments on larger datasets including MOSES. Due to the limited time, the final results are not yet ready. We will update the results as soon as the experiments are finished, through both official comments and the anonymous link below: https://anonymous.4open.science/r/NeurIPS2024Rebuttal-0290/. * We also provide the code through the same anonymous [link](https://anonymous.4open.science/r/NeurIPS2024Rebuttal-0290/) above. We will release our source code in the final version. Pdf: /pdf/fede1ad45dfcda42ea432af0377381e34665f70f.pdf	NeurIPS_2024_submissions_huggingface	2,024	null	null	null	null	null	null	null	null
MotionCraft: Physics-Based Zero-Shot Video Generation	Accept (poster)	Summary: In this work the authors propose MotionCraft, a new zero-shot video generator to craft physics-based and realistic videos. MotionCraft is able to warp the noise latent space of an image diffusion model, such as Stable Diffusion, by applying an optical flow derived from a physics simulation. The authors show that warping the noise latent space results in coherent application of the desired motion while allowing the model to generate missing elements consistent with the scene evolution, which would otherwise result in artefacts or missing content if the flow was applied in the pixel space. The authors compare the method with the state-of-the-art Text2Video-Zero reporting qualitative and quantitative improvements, demonstrating the effectiveness of the approach to generate videos with finely-prescribed complex motion dynamics. Strengths: 1. This method does not need extra training, which is efficient. 2. This method introduces explicit phisyics control to the field of video generation, which is novel. 3. This method finds out that the optical flow is consistent between pixel space and latent space, which is interesting. Weaknesses: 1. The main concern is about the experiments. This paper only have 5 video results in total, which is not sufficient. I am worrying that the method is highly unstable and not robust, thus the author cannot present more video results. If this is the case, I think this manuscript is not suitable for publication. If this is not the case, I think the authors should provide more generated results and I will be glad to raise my score. 2. How long does it take to generate a single video? If the paper claims that the video can be generated within minutes, I think generating tens of videos in the supplemtary material would be a good idea. Technical Quality: 2 Clarity: 2 Questions for Authors: See the weakness. Confidence: 4 Soundness: 2 Presentation: 2 Contribution: 3 Limitations: See the weakness. Flag For Ethics Review: ['No ethics review needed.'] Rating: 5 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: > The main concern is about the experiments. This paper only have 5 video results in total, which is not sufficient. I am worrying that the method is highly unstable and not robust, thus the author cannot present more video results. If this is the case, I think this manuscript is not suitable for publication. If this is not the case, I think the authors should provide more generated results and I will be glad to raise my score. We understand the reviewer's concern regarding the number of video results presented in our paper. We would like to assure you that our method is stable and robust. To address this concern, we have generated additional video results, which are included in the Additional Results PDF. These supplementary videos showcase a diverse set of scenarios and dynamic simulations, further demonstrating the reliability and consistency of our approach. For example, we applied it to a different diffusion model (SDXL) without modifications, resulting in a video with higher resolution and smoothness (example 4 in Additional Results PDF). We also generated a video 200 frames long (example 6) showcasing the stability of the method. Furthermore, in example 2, we tested our method with a complex optical flow, estimated directly from a real video. However, we are limited to a 1-page additional PDF in the rebuttal, so we included all the videos that we could fit on this page. More video will be added in the supplementary materials upon acceptance. We believe that these additional results will provide a comprehensive validation of our method and alleviate any concerns regarding its stability and robustness. Code is also available to generate any additional videos. > How long does it take to generate a single video? If the paper claims that the video can be generated within minutes, I think generating tens of videos in the supplementary material would be a good idea. The claims in the paper regarding the time to generate a video refer only to the video generation itself (all and only the steps in Pseudocode 1 in the manuscript). However, as explained in the methodology, there is a certain degree of manual input in setting up the physical simulation and conditioning it on the starting frame. For example, setting the type of simulation, the boundary conditions, or the initial state of the fluid often takes time beyond the actual runtime of the generative model. --- Rebuttal 2: Comment: Thank you for the rebuttal. I have read the additional results. The rating has been updated. I hope the authors could show more generated results in the suppmentary materials upon the acception of this manuscript.	Summary: The paper presents MOTIONCRAFT, a novel zero-shot video generation method that leverages physical simulations to create realistic and physically plausible videos. Unlike traditional video diffusion models that require extensive training and large datasets, MOTIONCRAFT uses a pre-trained image diffusion model, such as Stable Diffusion, and warps its noise latent space with optical flow derived from physical simulations. This approach ensures coherent motion and the generation of missing elements consistent with scene evolution. Key Contributions: 1. Innovative Approach: Introduction of a zero-shot video generation method that uses optical flow from physical simulations to warp the noise latent space of a pre-trained image diffusion model. 2. Experimental Validation: Demonstrates the effectiveness of MOTIONCRAFT through both qualitative and quantitative comparisons with the state-of-the-art Text2Video-Zero method, showing significant improvements. 3. Theoretical Insights: Provides an analysis of the correlation between optical flow in the image space and the noise latent space, supporting the proposed method. 4. Versatility: Showcases the ability of MOTIONCRAFT to generate videos with complex dynamics, including fluid dynamics, rigid body physics, and multi-agent interaction models, without additional training. 5. Technical Details: Describes key techniques such as multi-frame cross-attention and spatial noise map weighting to ensure temporal and spatial consistency in the generated videos. Overall, MOTIONCRAFT represents a significant advancement in zero-shot video generation, combining the strengths of physical simulations and image diffusion models to produce high-quality, dynamic videos. Strengths: Refer to Summary. Weaknesses: - There are now many approaches to zero-shot video generation, such as [https://openreview.net/forum?id=zOjW6yVYkE](https://openreview.net/forum?id=zOjW6yVYkE). The authors only compared their method with T2V0 (a relatively earlier method), which may make the experimental results insufficient. It would be better to include a more comprehensive comparison. - The current zero-shot video generation methods generally cannot achieve a very coherent video effect and can only generate keyframes. Although the method proposed in the paper largely ensures content consistency between consecutive frames, the generated videos still fail to achieve a highly coherent effect as seen in the demonstration. - Optical flow-based strategies often face limitations in certain specific situations. In complex environments, optical flow might not be effective. This aspect should be discussed and analyzed in the paper's main text. Moreover, the performance in different scenarios should be thoroughly evaluated using a variety of experimental results presented in the paper, instead of merely showcasing the method's validity through three handpicked examples. Technical Quality: 3 Clarity: 3 Questions for Authors: - Is it possible to achieve better controllable video generation by handling optical flow in a manner similar to that mentioned in the Generative Image Dynamics ([https://generative-dynamics.github.io/](https://generative-dynamics.github.io/)) paper? - Can more coherent videos be generated through interpolation in optical flow? - Is it possible to generate long videos? Confidence: 4 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: Refer to Weakness Flag For Ethics Review: ['No ethics review needed.'] Rating: 5 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: > There are now many approaches to zero-shot video generation, such as https://openreview.net/forum?id=zOjW6yVYkE. The authors only compared their method with T2V0 (a relatively earlier method), which may make the experimental results insufficient. It would be better to include a more comprehensive comparison. We appreciate the reviewer's suggestion. Besides the work linked by the reviewer, we are unsure what other specific papers the reviewer has in mind for zero-shot video generation with diffusion models. At the time of submission, we ensured to include references to all relevant works available. Regarding the method mentioned by the reviewer, we note that this work primarily compares with T2V0 using a limited quantitative metric (text-image CLIP score) without considering frame alignment or temporal coherence. Additionally, since it does not have any code or videos released, we are unable to reproduce their results and compare them within our evaluation setting. Nonetheless, we acknowledge the importance of a broader comparative analysis. We will include the referenced work in our related work section in the camera-ready version of our paper. > The current zero-shot video generation methods generally cannot achieve a very coherent video effect and can only generate keyframes. Although the method proposed in the paper largely ensures content consistency between consecutive frames, the generated videos still fail to achieve a highly coherent effect as seen in the demonstration. It is indeed a challenge for zero-shot video generation methods to maintain high temporal coherence. In particular, we found that the 64x64 resolution of the latent space is one limitation. This resolution restricts the accuracy and smoothness of the motion since it prevents sub-pixel movements during warping, resulting in less coherent transitions between frames. However, we have explored the use of Stable Diffusion XL (SDXL), which offers a larger latent space (128x128), allowing for more precise and smooth motion. By increasing the spatial resolution of the latent space, SDXL mitigates the limitations observed with the standard 64x64 latent space. See example 4 in the Additional Results. We will deepen the discussion of these aspects in the revised version of our manuscript. > Optical flow-based strategies often face limitations in certain specific situations. In complex environments, optical flow might not be effective. This aspect should be discussed and analyzed in the paper's main text. Moreover, the performance in different scenarios should be thoroughly evaluated using a variety of experimental results presented in the paper, instead of merely showcasing the method's validity through three handpicked examples. We agree with the reviewer that optical flow in the pixel domain is not adequate for generating realistic video (for example, it is difficult to handle occlusions). However, the same optical flow in the latent space can produce realistic videos. Indeed, MotionCraft is able to exploit the diffusion model in order to fix the simplistic nature of the optical flow while maintaining its explainability and simulation ability. We will add some videos generated by warping the pixel space to prove this claim. Regarding the limited number of examples, we provide videos with new contents and dynamics in the Additional Results PDF (we are limited to 1 page as per NeurIPS FAQ). These additional examples have been evaluated and confirm the quantitative results presented in the paper. > Is it possible to achieve better controllable video generation by handling optical flow in a manner similar to that mentioned in the Generative Image Dynamics paper? The Generative Image Dynamics method involves training a motion prior by modeling it as a dense Fourier volume. This representation of optical flow is particularly suited for oscillatory motions, while MotionCraft is able to handle both non-oscillatory and oscillatory motion (as seen in the sea waves in example 6). However, it should be possible to achieve better controllable video generation in a manner similar to GID, at the cost of learning an optical flow generator, because this module should be input dependent. This is, however, outside the scope of this work. > Can more coherent videos be generated through interpolation in optical flow? We found that the coherence of generated videos is affected by both the temporal resolution of the provided optical flows as well as the spatial resolution of the latent space of the diffusion model. Concerning the temporal resolution, it is possible to run the physics simulator with a sufficiently small time step to match the desired frame rate. However, as mentioned in our response to Weakness 2, the primary limitation we encountered to represent fine-grained motion is the 64x64 spatial size of the latent space in the current Stable Diffusion model. Plase, see our response to Question 2 for further comments. > Is it possible to generate long videos? Yes, it is possible to generate long videos using our method. From a theoretical point of view, the method complexity is linear in the number of frames to be generated, as explained in the manuscript. From a practical point of view, we demonstrate this capability in example 6 in the Additional Results, generating a coherent video with 200 frames. The key to generating long videos lies in the cross-attention mechanism we use, which ensures global consistency by attending to the first frame in the sequence and local consistency by attending to the previous frame. However, this cross-attention approach may introduce challenges when there are significant scene changes, as the mechanism maintains consistency with the initial frame. To address this, the frame-to-attend can be shifted periodically, or a weighted attention mechanism that decays over time can be used. These strategies can help enhance the video generation capability of our method for very long videos. --- Rebuttal Comment 1.1: Title: Please discuss Comment: Dear reviewer, The discussion period is coming to a close soon. Please do your best to engage with the authors. Thank you, Your AC	Summary: The paper works on the zero-shot video generation task and proposes, MotionCraft. It uses physics simulations to generate optical flow that follows physical dynamics. Then, optical flow is applied to warp the noise in the latent space with the stable diffusion model. This approach ensures coherent motion application and consistent scene evolution, avoiding artefacts and missing content typical in pixel space flow applications. Compared to the state-of-the-art Text2Video-Zero, MotionCraft shows both qualitative and quantitative improvements in generating videos with complex motion dynamics. Strengths: 1. The paper is well-motivated and well-written. 2. The idea of using physics simulator to generate the optical flow which is then applied in latent space is very interesting. 3. The qualitative results are impressive. Weaknesses: 1. More quantitative comparison with baselines. Table 1 only reports the comparison with T2V0 on the generated videos. However, it is not clear which benchmark it is. Is it possible to compare with other baselines on more benchmarks, like MUG, MHAD? 2. The method seems limited by specific types of dynamics, like fluid dynamics. It is not clear how to generate more general dynamics in real world. This may limit the potential application of the proposed method. 3. The method is assumes that "Optical Flow is preserved in the Latent Space of Stable Diffusion" based on the observation of average correlations 0.727 between optical flows estimated in the RGB and noise latent spaces. Does this assumption hold true for generating realistic, pixel-wise precise motion in video with only 0.72 cosine simlarity? Technical Quality: 3 Clarity: 3 Questions for Authors: How do you specify the region for simulating physical dynamics? Will the type of dynamic physics simulator affect the quality of the generated video? Confidence: 4 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: The method is limited to specific types of dynamic simulators and may be hard to apply to generic real-world video generation. Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: > More quantitative comparison with baselines. Table 1 only reports the comparison with T2V0 on the generated videos. However, it is not clear which benchmark it is. Is it possible to compare with other baselines on more benchmarks, like MUG, MHAD? We appreciate the reviewer’s suggestion to include more quantitative comparisons with additional baselines. Our experimental setup follows that of the seminal paper of T2V0, which, at this time, is the only baseline for zero-shot video generation. Moreover, the nature of our approach introduces specific challenges in directly comparing with benchmarks such as MUG and MHAD. Our method, MotionCraft, utilizes optical flows derived from physical simulations to generate videos. As no benchmarks of this kind are available, we decided to compare T2V0 and MotionCraft on a new set of physics-based videos. This set of videos is composed of the 5 examples reported in Table 1. Additionally, among other newly generated videos in the Additional Results PDF, example 2 shows that MotionCraft can handle facial expression change videos similar to those present in the MUG dataset. > The method seems limited by specific types of dynamics, like fluid dynamics. It is not clear how to generate more general dynamics in real world. This may limit the potential application of the proposed method. We argue that focusing on specific types of dynamics, like fluid dynamics, is a novel approach to the problem of zero-shot video generation. Moreover, fluid dynamics can be applied even to videos with content different from liquids or gases, as shown in the example 5 of Additional Results PDF, where the crowd moves according to a fluid simulation. However, the proposed approach based on physical simulations may not apply to all kinds of content to be generated. But, as we argue in the paper, the approach could be generalized by thinking of replacing the physics simulator with an animation software or 3D engine to generate the optical flows. One could even consider learning a dedicated optical flow generator. These are all interesting directions that require substantial future work that can leverage the evidence of effectiveness for the case of physics simulations shown in this paper. We will deepen the discussion of these aspects in the revised version of our manuscript. > The method is assumes that "Optical Flow is preserved in the Latent Space of Stable Diffusion" based on the observation of average correlations 0.727 between optical flows estimated in the RGB and noise latent spaces. Does this assumption hold true for generating realistic, pixel-wise precise motion in video with only 0.72 cosine similarity? We agree with the reviewer that the value of the cosine similarity needs a comment. This cosine similarity is computed between the flow estimated in the pixel space (between the two RGB frames) and, independently, the flow estimated between the corresponding latents. We argue that these two flows cannot be perfectly aligned since the VAE module of Stable Diffusion downsamples and encodes the image to the latent space. In addition, we think that 0.72 is a surprisingly high value, as it is measured in the noisy regime of the diffusion process ($\tau=400$ over $1000$ steps). To address the concern about generating pixel-wise precise motion, we conducted an additional experiment (Additional Results, experiment 2) to further validate our approach. We extracted the optical flow from a video and used it as input to MotionCraft. The results demonstrated that our method could reconstruct the initial video with higher fidelity than the reconstruction in the pixel space. > How do you specify the region for simulating physical dynamics? We thank the reviewer for this comment that allows us to clarify a point. As explained in the methodology, there is a certain degree of manual input in setting up a physical simulation conditioned on the starting frame. For example, in the fluid simulation of the filling glass in the supplementary material, we extracted a semantic map from the latent of the first frame and labeled the glass walls as boundary regions. > Will the type of dynamic physics simulator affect the quality of the generated video? We compared different types of physics simulators in appendix C, where the same dynamic is simulated with both an Eulerian and a Lagrangian numerical solver, resulting in two different videos. As we can see in this example, both methods provide plausible frames. This showcases the ability of MotionCraft to operate on very different types of physical simulators seamlessly, leveraging the image prior of the diffusion model to correct potential misalignments and produce visually coherent and high-fidelity videos. --- Rebuttal Comment 1.1: Title: Please discuss Comment: Dear reviewer, The discussion period is coming to a close soon. Please do your best to engage with the authors. Thank you, Your AC	null	null	Rebuttal 1: Rebuttal: Dear Reviewers and Area Chair, We appreciate the valuable feedback and suggestions provided by the reviewers. We have carefully addressed the concerns raised and integrated additional experiments and clarifications to expand the original material in the suggested ways. Please find below a detailed point-by-point response to the reviewers' comments. We hope that these revisions and the additional results provided will demonstrate the robustness and effectiveness of our method. Since the most critical point highlighted by the reviewers is the limited number of generated videos, we would like to note that we are limited to a 1-page additional PDF in the rebuttal, so we included all the videos that we could fit on this page. More videos will be added in the supplementary materials upon acceptance. Thank you for your consideration. Sincerely, Authors of submission 16937 Pdf: /pdf/ef8706fcb3c8f2d6ad757eceec7bfdc401aac430.pdf	NeurIPS_2024_submissions_huggingface	2,024	null	null	null	null	null	null	null	null
EEGPT: Pretrained Transformer for Universal and Reliable Representation of EEG Signals	Accept (poster)	Summary: The paper presents EEGPT, a 10-million-parameter pretrained transformer model designed for universal EEG feature extraction. The model employs a dual self-supervised learning method for efficient feature extraction and demonstrates state-of-the-art performance on various downstream tasks with linear probing. Strengths: 1. The proposed EEGPT model introduces a dual self-supervised learning method combining spatio-temporal representation alignment and mask-based reconstruction, which enhances representation quality and model robustness. 2. EEGPT achieves state-of-the-art performance on a range of downstream tasks such as motor imagery classification, ERP detection, and sleep stage detection. 3. The hierarchical structure for decoupled processing of spatial and temporal information reduces computational complexity and enhances flexibility for BCI applications. 4. Comprehensive experiments demonstrate the superior performance of EEGPT across multiple EEG tasks. Weaknesses: 1. The function of the adaptive spatial filter is unclear. Why is this module not included during pre-training? An ablation study is needed to evaluate the effectiveness of this component. 2. In Table 2, the authors use different metrics than in Table 3. Cohen's Kappa should be used consistently across both tables. Additionally, the results of LaBraM on these two datasets should be reported to provide a more comprehensive comparison, even if EEGPT, which uses linear probing, might not surpass LaBraM. 3. The technical contribution of the paper appears incremental. The framework is largely based on the context autoencoder (CAE) with modifications such as the adaptive spatial filter, multiple summary tokens, and rotary temporal embeddings. 4. During pre-training, EEGPT uses EEG data of the same configuration. In contrast, LaBraM leverages various datasets with different configurations, which might make EEGPT less universal. 5. The ablation study is conducted on a single dataset, which is not persuasive enough to demonstrate the effectiveness of the proposed modules and their scalability. 6. EEGPT uses different model settings for different downstream datasets, which may limit its generalizability. Technical Quality: 3 Clarity: 3 Questions for Authors: 1. It is unclear whether the baselines in Table 3 are based on linear probing or fully fine-tuned models. Clarification is needed. Confidence: 4 Soundness: 3 Presentation: 3 Contribution: 2 Limitations: The authors have discussed the limitations. Flag For Ethics Review: ['No ethics review needed.'] Rating: 5 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We thank the reviewer for the helpful feedback. We have uploaded a revision with the changes marked as blue. Our detailed responses are as follows: --- #### W1: The function of the adaptive spatial filter is unclear. Why is this module not included during pretraining? An ablation study is needed to evaluate the effectiveness of this component. The adaptive spatial filter is designed for the downstream task, adapting the EEG signal for each task/data distribution to the input of the encoder. In the downstream task, the model learns to adjust the parameter weights of the adaptive spatial filter, which maps (scale and project) the downstream task data to the same distributions as the pretrain data. For example, it can adapt to the different EEG references (e.g., for the epilepsy classification task, the double banana bipolar montage reference [2] with the features are more significant, and adaptive spatial filters allow for such a reference.). The adaptive spatial filter is designed only for downstream tasks. It can be of limited use during pretraining, since we have uniformly pre-processed the pretraining dataset and assumed that these data are in the same distribution. On the other hand, during pretraining, since the strategy of masking in both the spatial and temporal dimensions is used, the channels are not consistent in each input, and consequently leads to distributional drift and reduces training efficiency [3]. In the ablation experiments, we added the results of the ablation experiments that exclude the adaptive spatial filter in the downstream task, as shown in Revision Appendix A.3 Table 8 & author rebuttal Table 2. #### W2: In Table 2, the authors use different metrics than in Table 3. Cohen's Kappa should be used consistently across both tables. Additionally, the results of LaBraM on these two datasets should be reported to provide a more comprehensive comparison, even if EEGPT, which uses linear probing, might not surpass LaBraM. We added the Cohen's Kappa metrics to Table 3 and added the LaBraM-Base results, see Revision Appendix A.6. [Table: Results of experiments on TUAB] \|Methods\|Model Size\|Balanced Accuracy\|AUROC\| \|-\|-\|-\|-\| \|BIOT\|3.2M\|0.7959±0.0057\|0.8815±0.0043\| \|LaBraM\|5.8M\|0.8140±0.0019\|0.9022±0.0009\| \|Ours-Tiny\|4.7M\|0.7959±0.0021\|0.8716±0.0041\| \|Ours\|25M\|0.7983±0.0030\|0.8718±0.0050\| [Table: Results of experiments on TUEV] \|Methods\|Model Size\|Balanced Accuracy\|Weighted F1\|Cohen's Kappa\| \|-\|-\|-\|-\|-\| \|BIOT\|3.2M\|0.5281±0.0225\|0.7492±0.0082\|0.5273±0.0249\| \|LaBraM\|5.8M\|0.6409±0.0065\|0.8312±0.0052\|0.6637±0.0093\| \|Ours-Tiny\|4.7M\|0.5670±0.0066\|0.7535±0.0097\|0.5085±0.0173\| \|Ours\|25M\|0.6232±0.0114\|0.8187±0.0063\|0.6351±0.0134\| LaBraM employs a larger pretraining dataset than ours (reference to [1] Appendix D), which also contains the TUEG dataset that is similar to the TUAB and TUEV distributions, and whose paper illustrates the scale effect of the amount of pretraining data, which may have facilitated the learning of a richer and more downstream task-adaptable representation for LaBraM. #### W4: During pretraining, EEGPT uses EEG data of the same configuration. In contrast, LaBraM leverages various datasets with different configurations, which might make EEGPT less universal. In this paper, we unified the pretraining dataset into 58 channels, each channel has the same amount of data, making each channel of equal importance. In addition to this, LaBraM uses more eeg reference configuration (the double banana bipolar montage reference [2]) for epilepsy data, which may lead to stronger generalization (for epilepsy tasks). We conducted initial experiments using a limited dataset and will gradually expand to larger datasets (e.g., epilepsy dataset) to enhance the generalizability of EEGPT. #### W5: The ablation study is conducted on a single dataset, which is not persuasive enough to demonstrate the effectiveness of the proposed modules and their scalability. In the ablation experiment, we added test results for the BCIC-2B and KaggleERN datasets, as shown in the table below, and see Revision Section 3.4. See author rebuttal Table 1 for more details. #### Q1: It is unclear whether the baselines in Table 3 are based on linear probing or fully fine-tuned models. Clarification is needed. The baseline in Table 3 uses the full fine-tuning method, which we have supplemented, see Revision Section 3.2. [1] Jiang, W. B., Zhao, L. M., & Lu, B. L. (2024). Large brain model for learning generic representations with tremendous EEG data in BCI. arXiv preprint arXiv:2405.18765. [2] Rosenzweig, I., Fogarasi, A., Johnsen, B., Alving, J., Fabricius, M. E., Scherg, M., ... & Beniczky, S. (2014). Beyond the double banana: improved recognition of temporal lobe seizures in long-term EEG. Journal of Clinical Neurophysiology, 31(1), 1-9. [3] Tian, K., Jiang, Y., Diao, Q., Lin, C., Wang, L., & Yuan, Z. (2023). Designing bert for convolutional networks: Sparse and hierarchical masked modeling. arXiv preprint arXiv:2301.03580. --- Thanks again for your comments. Hope our rebuttal has addressed all your concerns. --- Rebuttal Comment 1.1: Comment: I appreciate the comments by the authors. Some of my concerns are addressed by the clarification and additional experiments. However, regarding the technical contribution and universibility (W3 and W4), I maintain my standpoints.	Summary: The authors propose a novel pretraining strategy, EEGPT, that is essentially a multi-task self-supervision loss consisting of a masked autoencoder-style reconstruction objective, and an alignment loss that is reminiscent of knowledge distillation approaches such as data2vec. EEGPT is applied to representation learning of electroencephalography (EEG) data, and the authors demonstrate the pretrained model flexibility across a variety of downstream EEG decoding tasks (motor imagery, sleep staging, ERP detection). Strengths: a. Originality: The individual components of EEGPT are not original. The masking strategy and reconstruction loss are based on Masked Autoencoder which was originally developed for vision, and has since been applied to time-series data (including EEG) previously. Similarly, the alignment loss is highly similar to that of data2vec, except that data2vec uses a smoothed L1 loss. However, the combination of these two objectives into a multi-task self-supervision loss is novel (to my knowledge). According to the ablation results in Table 5, it does seem that the two objectives compete against each other, but in a way that is constructive for the downstream decoding tasks. b. Quality: The conclusions made by this paper are mostly sound. The authors demonstrate the improvements of EEGPT over other SOTA approaches across a comprehensive list of downstream EEG decoding tasks. Also of note, the authors made an effort to demonstrate the design choices of EEGPT in their ablation analysis. c. Clarity: The paper is overall clear in its writing. There are some missing details, which I have detailed below in the Questions section. d. Significance: The pretraining approach is likely to be reused in other works. The model generalizability across a diverse set of EEG decoding tasks is especially compelling, since it suggests that this model may be used as a generalized foundation model for EEG. Weaknesses: b. Quality: One weakness of this work is that the proposed model has significantly more parameters than the SOTA approaches it is comparing against. It remains unclear if the authors have developed a better pretraining approach, or if the improvements in the downstream tasks are due to higher expressivity in the base model architecture. Could the authors add additional comparisons where the number of parameters EEGPT matches the number of parameters in the other SOTA approaches? Another weakness is that the comparison to pretraining approaches such as BIOT is unfair. The authors state in Appendix E: “the BIOT model applies an FFT on 1-second length patches, which is overly extensive and results in significant information loss". However, this does not mean that the pretraining recipe in this paper is better than BIOT. It could be the case that BIOT performs similarly, it just needs to be retrained with different patch length and/or patch stride. I understand that retraining BIOT with new parameters may not feasible in the rebuttal timeframe, but I think this should be clarified in the text. Technical Quality: 2 Clarity: 3 Questions for Authors: 1. Line 59: The authors claim that the low SNR of EEG “...make it challenging to learn abstract features using masked autoencoders”. However, there have been multiple works that use MAE-style pretraining to successfully learn features from EEG, such as: [1] Chien, et al. MAEEG: Masked Autoencoder for EEG Representation Learning. arXiv:2211.02625. [2] Wu, et al. Neuro-BERT: Rethinking Masked Autoencoding for Self-supervised Neurological Pretraining. IEEE JBHI. 2. Eqn 9: Did the authors try using a weighted loss with weight $\lambda$? Or does each individual loss take on similar values? $\mathcal{L} = \mathcal{L}_A + \lambda \mathcal{L}_R$ 3. Section 2.4: Can the authors clarify which “Features” are passed to the linear classification head? Are these the summary tokens (passed to the Predictor during pretraining)? Or are they the non-summary tokens? 4. Section 3.1: Were any notch filters used to suppress electrical noise in datasets used for pretraining or downstream tasks? 5. Section 3.2: Why were 58 electrodes used? Is it because these 58 electrodes were common across the pretraining datasets? Also, what happens at inference time if a new electrode not belonging to the original set of 58 locations appears? 6. Section 3.2: Is there any normalization applied to the input signal? I saw in the appendix that each dataset has a different normalization (EA on entire session, z-score on each input, etc.). Does that mean that the normalization is dataset-dependent? 7. Line 197: Missing reference for OneCycle learning rate 8. Tables 2&3: The proposed EEGPT model has ~8X as many parameters as the second largest model (BIOT). It would be a fairer comparison if the authors also reported BAC for a similarly sized EEGPT model, such as tiny3 or little (from Table 6). 9. Table 5: Can more description be added to this table? For example, BAC stands for Balanced Accuracy I assume? And which dataset is BAC being reported for here? Or is it an average across datasets? 10. (Minor comment) The title “EEGPT” is slightly misleading, since it makes the reader think that the model is being pretrained using an autoregressive reconstruction objective such as in the popular GPT language models (e.g. predict the next token given the previous tokens). I would recommend changing to a different name to avoid confusion. 11. (Minor comment) Line 186: Spelling error “Appendix” 12. (Minor comment) Line 499: What does EA in “EA normalization” stand for? I tried to find it elsewhere in the paper, but could not find it.. Confidence: 4 Soundness: 2 Presentation: 3 Contribution: 3 Limitations: Yes. Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We thank the reviewer for your appreciation and constructive comments. We have uploaded a revision and used blue to mark the new changes. --- #### W1&Q8 See Revision Section 3.3 for test results of model variants with comparable number of parameters. [Table: Results of experiments on TUAB] \|Methods\|Model Size\|Balanced Accuracy\|AUROC\| \|-\|-\|-\|-\| \|BIOT\|3.2M\|0.7959±0.0057\|0.8815±0.0043\| \|Ours-Tiny\|4.7M\|0.7959±0.0021\|0.8716±0.0041\| \|Ours\|25M\|0.7983±0.0030\|0.8718±0.0050\| [Table: Results of experiments on TUEV] \|Methods\|Model Size\|Balanced Accuracy\|Weighted F1\|Cohen's Kappa\| \|-\|-\|-\|-\|-\| \|BIOT\|3.2M\|0.5281±0.0225\|0.7492±0.0082\|0.5273±0.0249\| \|Ours-Tiny\|4.7M\|0.5670±0.0066\|0.7535±0.0097\|0.5085±0.0173\| \|Ours\|25M\|0.6232±0.0114\|0.8187±0.0063\|0.6351±0.0134\| The number of parameters in the table for Ours-Tiny is 4.7M comparable to the BIOT model. Compared to BIOT, our Tiny model improves about 4% on the TUEV dataset and achieves comparable performance with it on the TUAB dataset. Compared to the BIOT method, we used a smaller pretraining dataset and did not use more datasets containing epilepsy samples, but still achieved good results on both datasets. #### W2 We agree with you that on some tasks it is possible that the pretraining scheme in this paper performs similarly to the BIOT. The statement in Revision Appendix F aims to explain the possible reasons for the poor performance of the BIOT on the BCIC-2A and BCIC-2B motor imagery tasks (see Section 3.3, Table 4). The "significant information loss" is mainly reflected in the fact that the BIOT model only retains the spectral energy information for 1s of each patch after the FFT, and discards the phase information. We have provided additional explanations for this, see Revision Appendix F. #### Q1 While other work has validated the ability of MAE pretraining methods to learn features in corresponding EEG tasks, our work would like to emphasize that the low signal-to-noise ratio of EEG prevents the model from learning "high-quality" representations through MAE pretraining methods. Our ablation experiments verify this, see Revision Section 3.4. See author rebuttal Table 1 for more details. In the absence of $L_A$ loss, the model's performance degradation on all datasets is significant (6% to 9%). #### Q2 In our experiments, despite the presence of two loss functions, no weights were explicitly introduced during the optimization process to balance the two losses. This approach also simplifies the training process of the model. #### Q3 "The summary tokens (passed to the Predictor during pretraining)" are passed to the linear classification head. We provide additional clarification, see Revision Section 2.4. #### Q4 In Revision Appendix C we describe in detail how we preprocessed each dataset. For all datasets we did not target the use of notch filters to suppress electrical noise. #### Q5 This is because these 58 electrodes are the electrodes that are present in all pretraining datasets and are the intersection of their electrode sets. These 58 electrodes cover as much as possible the electrodes in the international 10-10 EEG system standard [1]. During the pretraining phase, we removed channel data from the data sample that were not these 58 electrodes. If a new electrode that does not belong to the 58 pretrained electrodes appears in the downstream task, the signal from the new electrode can be mapped to the neighbouring/similar electrode input of the encoder model by using the adaptive spatial filter (see Revision Section 2.4), which is done by training the adaptive spatial filter on the downstream task dataset. #### Q6 In Revision Appendix B, we describe in detail the normalization and preprocessing approach for each dataset. The normalization and preprocessing approach is dependent on different EEG paradigms. #### Q7 We have added references to OneCycle Learning Rate [2], see Revision Section 3.2 & Appendix C.2. #### Q9 We added more detailed descriptions for the tables, and in Table 5 of Revision Section 3.4, we changed the headers to BCIC-2A-BAC, BCIC-2B-AUROC, and KaggleERN-AUROC. #### Q10 Thanks for commenting on the potential confusion with "EEGPT". We chose this to highlight the application of pretrained Transformers on EEG signals, despite not using autoregressive task pretraining. It showcases the ability of to learn universal pattern from massive data. However, we've clarified that our model isn't pretrained using autoregressive objectives in our paper's introduction (Revision Section 1) to prevent misunderstandings, while maintaining our original title intent. #### Q11 We have fixed spelling errors, see Revision PDF. #### Q12 We have added a citation and description of EA normalisation [3], see Revision Appendix C.2. EA normalization is a common normalization method applied to EEG data from motor imagery tasks. This approach align EEG trials from different subjects in the Euclidean space to make them more similar, and hence improve the learning performance for a new subject. [1] The five percent electrode system for high-resolution EEG and ERP measurements. [2] Super-convergence: Very fast training of neural networks using large learning rates. [3] Transfer learning for brain–computer interfaces: A Euclidean space data alignment approach. --- Rebuttal Comment 1.1: Comment: Thank you for thoroughly addressing my comments & questions. See below for further comments New comment: I see that the authors have added a comparison to LaBraM in A.6 showing that LaBraM outperforms EEGPT significantly on TUEV and TUAB datasets. However, Table 4 demonstrates that EEGPT is better than LaBraM at other tasks (BCIC, SleepEDF, etc.). Do the authors have any explanation for this? W1: Thanks for running this! Indeed, it seems like when the number of parameters is matched between EEGPT and BIOT the performance is better (TUEV) or similar (TUAB). W2: I appreciate that you added the comment about possible reasons for why BIOT might be worse in the Appendix. I still feel the comparison is slightly unfair, and re-running the BIOT pretraining with smaller token sizes should be subject of future work to determine whether EEGPT is truly a better pretraining strategy. Q12: Can you define the acronym for EA somewhere in the text? I assume it stands for Euclidean Alignment? --- Reply to Comment 1.1.1: Comment: #### New comment Our dual self-supervised method improves the quality of EEG representations, therefore it performs better on other tasks. On the TUAB and TUEV tasks, LaBraM may achieve better performance for two reasons: firstly, LaBraM uses 8-second pre-training data segments (while EEGPT uses 4 seconds), and secondly, LaBraM employs TUEG datasets (which are similar to the data distribution of TUAB and TUEV) as pretraining datasets, thus it is able to better capture the long-term features in the EEG data. We use a 4-second pretraining dataset because all the pre-training data we use are task-state data segments, which are almost 4 seconds in length to ensure data quality. #### W2 In this paper, our work mainly focuses on showing that, under the same conditions, our proposed dual self-supervised method demonstrates significant performance improvement compared to the MAE self-supervised method. Moreover, our model exhibits comparable or better performance on a wider range of tasks/datasets compared to other SOTA general EEG pretrained models (not pretraining methods), suggesting that our model may have stronger generalizability. The main improvement of BIOT lies in the Biosignal Tokenization method, which is an improvement in the tokenizer, such as the introduction of normalization and FFT to retain more task-related features; while LaBraM's main improvement is the proposal of the Neural Tokenizer inspired by VQ-VAE, and using the tokens encoded by the Neural Tokenizer as the prediction target, which is an improvement in the prediction target, making the prediction target contain more task-related features (such as spectral features); in contrast, our work focuses on an enhancement to the self-supervised method itself, using a dual self-supervised method to enhance the quality of the representations. EEGPT, BIOT, and LaBraM may each have their own design advantages in the pretraining method. More work may be needed in the future for a more fair comparison, and they can learn from and integrate with each other. #### Q12 Thank you for your comments. EA stands for 'Euclidean space EEG data alignment' approach, which we have declared in the new revision.	Summary: The paper presents a new pretrained model called EEGPT (EEG pretrained transformer), designed to improve the analysis of EEG (electroencephalography) signals. EEGPT is a model with over 10 million parameters (up to 100M for the largest model) that aims to solve common problems in EEG analysis, where the challenges are low signal quality and high variability between individuals. The model introduces a special method called "dual self-supervised learning" to extract better features from EEG data. This method includes two main techniques: the usual masked reconstruction loss, and additionally the alignment loss. The alignment loss is a special loss introduced inside the masked auto-encoder to “force” the learned features of the masked patches inside the encoder (and predictor) to be aligned with spatial and temporal patches (all masked and unmasked). This is achieved by adding a momentum encoder and computing the loss between the predictor (global integration of summary patches of each time j) and momentum encoder (spatial integration of all unmasked patches for each time j). The model was trained on several large EEG datasets and tested on various tasks like classifying motor imagery, detecting event-related potentials, and identifying sleep stages. EEGPT performed better than existing models in these tasks, showing it is effective and scalable. Strengths: 1. The model introduces the “dual self-supervised technique” and to my knowledge it is novel. 2. The model performs better in the proposed tasks presented in the paper. However, the model is also significantly larger than the competition. 3. The paper performs an ablation study showing the importance of their decision of using the alignment loss (L_A). Without L_A, accuracy drops from 58.46 to 52.87. 4. The paper presents a range of experiments besides the model training and evaluation, providing a better understanding of the model design and behaviour. This includes methods like scaling laws, relationships of the channel embeddings (which are learned), and visualizations on the correlations of the motor imagery tasks and the channels for the BCI competition (BCIC2A dataset). 5. The downstream tasks are performed only with linear probing (no need for full finetuning) and the model still outperforms other models. Weaknesses: 1. The experiments done for the model design and behaviour such as scaling laws, ablation studies, channel embedding relationships, etc, are done only on one downstream task: namely the BCIC2A dataset. How does the model perform on the other tasks, what is the reason for not presenting the results on the other datasets? 2. The paper presents interesting scaling law experiments, showing potential in the method to scale to larger model size. However the experiments are too limited to be able to state the scaling laws as given in the paper (e.g. the ACC = (33.6*N)^0.029). I would argue one has to test it either to more tasks, or have a unified metric that uses many datasets for the scaling law. Also, how does the law depend on the amount of training data used? 3. The design of the model was at first confusing and I still have trouble understanding the main motivation from some of the decisions. The ablation studies are a great help here, but I think a few more ablation studies or explanations on the design of the model would clarify the paper better (see Questions section). 4. I have a few doubts on the evaluation of the baselines. Although I am not certain what are the best achievable accuracies in many of the tasks, for TUAB we have many different models that might perform better: We have models up to 86% (ChronoNet) in Table 1 in here (https://www.sciencedirect.com/science/article/pii/S2213158223001730) and above 90% in here (https://arxiv.org/pdf/2401.10283). I know there are certain differences: accuracy vs balanced accuracy and the second paper has a different evaluation for the highest accuracy. However, the TUAB dataset is quite balanced to my knowledge and still in the second paper in Table 1, there are multiple models close to 90%. Why are these models not included in the comparison, and since we have this discrepancy in this task, do other models exist also for other tasks as well? 5. Why is the pretraining set very small? In terms of participants, the downstream set is much bigger than the pretraining set. In addition, the TUEG dataset is bigger than all the datasets used here (combined). Is there any reason why the authors chose these datasets in particular? It is not clear from the paper why this is the case. Usually, during pretraining the dataset is much bigger compared to the downstream task. In this paper this does not seem to be the case. In fact, many of these datasets (all?) used during pretraining do have labels (i.e. why self-supervised). 6. In some sections it was not clear which evaluation dataset has been used to produce the results. For example in Section 3.4 there is Table 5 but it is not clear which model size and which dataset has been used to produce the results. Because of the accuracy number 58.46 I could infer that it was BCIC2A, but it is not mentioned anywhere. For scaling laws it is mentioned in the text, but it is better to have it on the table description. 7. Many minor annoyances, which make me wonder how much care the authors took to write this paper: - Misspellings, for example: Accuarcy vs accuracy, Appandix vs Appendix, etc. - The references are messed up. Just look at the first two references [1] and [2], where every author is listed at least 2 times, sometimes even 3 times! - You need to put a space in front of a reference, like this [1]. You almost always have no space, but not even this is consistent, because sometimes you do. Technical Quality: 3 Clarity: 2 Questions for Authors: Even though there is the ablation study for the alignment loss, it is not clear from the paper what is the main motivation behind it, and why this particular design is used. To be more concrete, we have the summary tokens from the encoder which are only created from patches at specific time point j. By design we have now an encoder that integrates all the patches spatially (for each timepoint). The confusing part for me is the predictor: on the one hand it tries to capture global features between all tokens in the time domain, on the other hand the alignment loss forces the tokens to only capture the local features at specific timepoint. What is the main motivation behind this? Note that from the experiments it is shown that the method works quite well, but it was confusing for me in the beginning to understand how did you come up with this design choice. Stated differently: how does the model perform if we do the alignment loss before the predictor, or if we remove the predictor. How did you evaluate the TUAB dataset since your model can only work 4s windows. The TUAB dataset can have many minutes to hours of EEG data. Same goes for TUEV. The evaluation for TUAB seems to not include in the comparison the best performing models in the literature (see limitations). Are there also other models for the other tasks that perform better and are not included in the paper? Confidence: 4 Soundness: 3 Presentation: 2 Contribution: 2 Limitations: The model is only trained with 4s of EEG data and the pretraining set is very small. Although the initial experiments are promising, It is not clear how the model generalises in other datasets and downstream tasks. The model is not evaluated how it performs if we finetune the model (instead of linear probing). Flag For Ethics Review: ['No ethics review needed.'] Rating: 5 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We thank the reviewer for the constructive feedback. We have uploaded a revision and used blue to mark the new changes. Our detailed responses are as follows. #### W1 We added more test results for more datasets in the ablation experiment and scale law experiment, see Revision Section 3.4 & Appendix A. Channel embedding relationships are computed based on channel embeddings extracted from the pretrained model and are independent of the downstream task. #### W2 In the scale law experiment, we added test results for the TUAB dataset and the BCIC-2B dataset, see Revision Appendix A.4. The results on the TUAB dataset and the BCIC-2B dataset show that they present a scale law for the performance metrics with respect to the model size. We added pretraining experiments using 100%, 50%, 25%, and 12.5% of the training data and tested them on the downstream tasks of BCIC-2A and BCIC-2B, see Revision Appendix A.5. The results show that their performance metrics exhibit a scale law with respect to the amount of data. #### W3&Q1 The design motivation of the predictor: (1) on the one hand, it predicts the representations of the masked "50% time patches" (equivalent to the BERT-style task), which allows the encoder to extract features in each local patches, that contribute to the task of completing masked patches in the time dimension; (2) on the other hand, it avoids the collapse of representation due to the direct alignment of the outputs from the encoder and momentum encoder (i.e., the parameters of the two are difficult to change during pretraining), and in this way, it is similar to the predictor in BYOL [1]. Also due to the design of (1) we can avoid predictor learning to be a identity map. We conducted a pretraining experiment after removing the predictor, the variation of $L_R$ loss with the number of iteration steps during training is shown in Figure 6 of the Revision Appendix A.1. In Figure 6, the $L_R$ loss of the model without the predictor does not decrease, which indicates that directly aligning the outputs of the encoder and the momentum encoder does lead to the problem of representation collapse, resulting in the model can not learn meaningful representation. #### W4&Q3 (1) The difference in accuracy on TUAB is due to the different sample lengths used by the different methods. For a fair comparison, we use the same evaluation approach on TUAB as in the LaBraM and BIOT papers, which both use 10-second samples for classification. ChronoNet [2] uses a window of different sizes of at least 60s for classification and achieves an accuracy of 86%. In the second paper [3] it used 1-minute samples for classification. This shows that different sample lengths can have a huge impact on the experimental results. (2) Our aim is to develop better pretraining methods so that the models can achieve better performance (more generality) on a wide range of EEG tasks. Therefore on other tasks we only compare with similar pretrained models. On some tasks, there exist complex models that may achieve better results using handcrafted features or dedicated designs, but they are developed for specific tasks and may be more difficult to generalize across tasks compared to pretrained models. #### W5 TUEG belongs to the clinical dataset and the data are related to patients and brain diseases, the channels of TUEG dataset are mostly 23 channels 10-20 standard electrodes, not diverse enough. We trained our model on a limited number of pretraining datasets, which includes more brain-computer interface related tasks such as motor imagery, error-related potentials, sleep stage detection, and identity recognition. These datasets also contain more channels. The TUEG dataset was chosen as the downstream task dataset in order to better compare with other pretrain models such as BIOT. Despite the abundance of labeled data for epilepsy and sleep task, many works [4] [5] [6] have used self-supervised methods to achieve performance that exceeds that of supervised methods. \|Methods\|Model Size\|Balanced Accuracy\|AUROC\| \|-\|-\|-\|-\| \|Ours (no pretrained)\|25M\|0.7553±0.0014\|0.8260±0.0018\| \|Ours\|25M\|0.7983±0.0030\|0.8718±0.0050\| We added test results using random initialization parameters on TUAB (see Appendix A.7), and according to our experimental results, even with a small amount of pretraining data, the pretrained model works better on the TUAB dataset compared to the unpretrained model. #### W6 We added more detailed descriptions for the tables, see Revision PDF Section 3.4. In Table 5, we changed the headers to BCIC-2A-BAC, BCIC-2B-AUROC, and KaggleERN-AUROC. In Table 6, we update the header to BCIC-2A-BAC. #### W7 We have fixed all spelling errors and regularized the citation reference format, see Revision PDF. #### Q2 In Revision Appendix C.2.6 Table 12&13, we show the architecture of our model used in the TUAB and TUEV. [ Model architecture for TUAB & TUEV ] \|In Size\|Opt\|kernel\|stride\|groups\|padding\| \|-\|-\|-\|-\|-\|-\| \|23xT\|conv1d,norm,gelu\|1\|1\|1\|0\| \|20xT\|conv1d,norm,gelu,dropout\|K\|1\|20\|K/2\| \|20xT\|encoder\|64\|64\|\|\| \|(T/64)x4x512\|flatten,linear\|\|\|\|\| The 23-channel input is first to reduce the number of channels to 20 by the convolution (K=15 for TUAB and 55 for TUEV). Then, the eegpt-encoder uses the 20 channel embeddings and maps 64-length window segments of the input signals to 4 (number of summary tokens)x512-dimensional features. Finally, the flatten and linear layers are used to output the final classification score. #### L1 See author rebuttal ablation experiment. [1] Bootstrap your own latent-a new approach to self-supervised learning. [2] ChronoNet: A deep recurrent neural network for abnormal EEG identification. [3] Window Stacking Meta-Models for Clinical EEG Classification. [4] Channel-Aware Self-Supervised Learning for EEG-based BCI. [5] MAEEG: masked auto-encoder for EEG representation learning (2022). [6] BIOT: Cross-data biosignal learning in the wild.	null	null	Rebuttal 1: Rebuttal: We thank all the reviewers for your time and constructive feedback. During the rebuttal, we have prepared a revision and used blue to mark the new changes. Below are the results of the added experiments as the response. [Table 1: Ablation study for pretraining methods (Appendix A.2)] In the ablation experiment, we added test results for the BCIC-2B and KaggleERN datasets, as shown in the table below. The 'BAC' is the Balancing Accuracy, and more detailed descriptions of the metrics can be found in Revision Appendix D. \| Variants \| $\mathcal{L}_{A}$ \| $\mathcal{L}_{R}$ \| BCIC-2A-BAC \| BCIC-2B-AUROC \| KaggleERN-AUROC \| \| ------------------------ \| ----------------: \| ----------------: \| ----------------: \| ----------------: \| ----------------: \| \| A: w/o $\mathcal{L}_{A}$ \| 37.13 \| 0.57 \| 0.5287±0.0086 \| 0.7264±0.0381 \| 0.5752±0.0164 \| \| B: w/o $\mathrm{LN}$ \| 0.15 \| 0.002 \| 0.5567±0.0088 \| 0.7920±0.0012 \| 0.5891±0.0227 \| \| C: w/o skip \| 0.12 \| 0.56 \| 0.5796±0.0011 \| 0.7702±0.0122 \| 0.6356±0.0296 \| \| D: with all \| 0.24 \| 0.56 \| 0.5846±0.0070 \| 0.8059±0.0032 \| 0.6621±0.0096 \| The results of the ablation experiments with the addition of the dataset are still in line with the conclusions in the original paper:(1) the model performance decreases significantly on all datasets without the $L_A$ loss (6% to 9%); (2) the performance of the variant B model without layer normalization of the target patches decreases by 3%, 1%, and 7% on BCIC-2A, BCIC-2B and KaggleERN, respectively; (3) the performance of the variant C model with the removal of skip connection was reduced by 1%, 3%, and 3% on the three datasets, respectively. [Table 2: Ablation experiments of fine-tuning methods (Appendix A.3)] In the ablation experiments, we added experiments comparing the linear probing method with the full fine-tuning method, see Revision Appendix A.3, as shown in the table below. \| Variants \| ASF \| L-P \| BCIC-2A-BAC \| BCIC-2B-AUROC \| KaggleERN-AUROC \| \| -------- \| ---- \| ---- \| ----------------- \| ----------------- \| ----------------- \| \| A \| \| \| 0.5774±0.0072 \| 0.7871±0.0054 \| 0.6078±0.0101 \| \| B \| ☑️ \| \| 0.5183±0.0155 \| 0.7541±0.0083 \| 0.6110±0.0019 \| \| C \| \| ☑️ \| 0.5586±0.0089 \| 0.7974±0.0030 \| 0.6463±0.0081 \| \| D \| ☑️ \| ☑️ \| 0.5846±0.0070 \| 0.8059±0.0032 \| 0.6621±0.0096 \| In the above table, ASF denotes with adaptive spatial filter, instead of directly feeding the signal into the model; L-P denotes using linear probing, instead of using the full fine-tuning of the model. Model variants A and C are models with full fine-tuning and linear probing, respectively, after excluding the adaptive spatial filter. Model variants B and D are models with full fine-tuning and linear probing after using adaptive spatial filter, respectively. The results show that on the BCIC-2B and KaggleERN datasets, variants C and D tested with linear probing achieve better results than variants A and B using full fine-tuning; on the BCIC-2A dataset, variant A with full fine-tuning and no adaptive spatial filter is close to variant D with linear probing and adaptive spatial filter. Overall, the variants C and D (using linear probing) outperform A and B. Also, the results show that variants B and D with adaptive spatial filter tested on the BCIC-2A, BCIC-2B and KaggleERN datasets achieve better results compared to variants A and C without adaptive spatial filter. Pdf: /pdf/e1dc93f516729dc10958fd5648704c86f15217b1.pdf	NeurIPS_2024_submissions_huggingface	2,024	null	null	null	null	null	null	null	null
UniBias: Unveiling and Mitigating LLM Bias through Internal Attention and FFN Manipulation	Accept (poster)	Summary: This paper addresses internal biases in LLMs that cause prompt brittleness. Unlike previous external adjustment methods, it examines the roles of MLPs and attention heads in creating these biases. The proposed UniBias method identifies and masks these biased components, thereby enhancing the model's in-context learning ability. Experiments across several NLP tasks demonstrate that UniBias effectively improves ICL performance. Strengths: * This paper explores the internal contributions of FFN vectors and attention heads to LLM bias, a relatively unexplored area. * The proposed method is straightforward and does not introduce any additional inference time cost, unlike post-calibration methods. Weaknesses: The main weakness of this paper is that there seems to be some disconnect between the problem it aims to solve (the prompt sensitivity of large language models, or LLMs) and two of the objectives of the preliminary analysis (how internal components generate common token bias and selection bias towards labels). Specifically, the challenge the paper proposes to address is the sensitivity of LLMs to input prompts (example selection, format, and order), and this issue is also illustrated in Figure 1. However, two of the targets of the exploration section are the biases present when LLMs use labels from inputs. The authors do not explicitly state why suppressing the model's bias toward labels can simultaneously alleviate the model's bias toward prompts. For example, intuitively, suppressing the label's preference for option A does not necessarily suggest an improvement in the model's bias towards a certain format or example. Given this lack of clarity, and since the criteria for detecting biased components also focus on labels, it is unclear why the proposed method can reduce prompt brittleness. Technical Quality: 3 Clarity: 2 Questions for Authors: NA Confidence: 4 Soundness: 3 Presentation: 2 Contribution: 3 Limitations: Limitations: Yes; Broader impacts: NA Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Dear Reviewer CnNd, We deeply appreciate your thorough review and valuable feedback. Below is a summary of our answers (A) to the weaknesses (W) you raised. --- [W1]: There seems to be some disconnect between the problem it aims to solve and two of the objectives of the preliminary analysis. [A1]: Thank you for raising this important issue! We apologize for the lack of clarity in linking these aspects in our paper. The connection between prompt sensitivity of LLMs and their bias towards labels is well-established in the LLM bias mitigation literature and we mentioned this in Line 21 of our paper. However, thanks to your insightful feedback, we realize the need to more thoroughly strengthen and elucidate this connection to enhance understanding of our method's effectiveness. A detailed analysis is provided below: The Connection Established by Literature: Firstly, the literature demonstrates that the prompt sensitivity arises from LLMs' inherent bias towards predicting certain answers. Therefore, the prompt brittleness can be addressed by mitigating LLMs' bias towards labels. For example, in the classic study of "Calibrate Before Use" [1] it explicitly states that: > We demonstrate that this instability of prompt arises from the bias of language models towards predicting certain answers. They detailed this analysis in Section 4 of their paper. Building on this foundation, following studies address the bias of LLMs toward to address prompt brittleness. For example, a domain-label bias is identified and mitigated to alleviate prompt brittleness in [2]. Additional Analysis on Our Method: We further analyze why mitigating model's bias towards labels can alleviate prompt brittleness in our method. Due to the inherent bias of LLMs, different prompts can lead to varying biases towards labels. For example, due to recency bias, placing a negative sentiment analysis sample at the end of a prompt can make LLMs tend to predict 'negative', incorrectly classifying positive samples and thus degrading ICL performance. Various bias factors lead to different direction and extend of bias, resulting in different changes in ICL performance and leading to the prompt brittleness. In contrast, our UniBias method effectively mitigates various potential biases inherent in LLMs. By doing so, it minimizes the introduction of bias towards labels regardless of the difference in prompts, leading to more stable and accurate ICL performance across different prompt configurations. Experimental Validations: We further provide experimental results to demonstrate how mitigating bias towards labels can alleviate prompt brittleness in ICL. We analyze the average prediction probabilities of 'negative' and 'positive' across all test samples in various prompt formats on the SST2 dataset. The ICL performance corresponding to these formats is depicted in Figure 1 of our paper. As shown in the table below, the average prediction probabilities for each format are represented as $(P_{\text{ave}}(\text{neg}), P_{\text{ave}}(\text{pos}))$. Our results reveal that different prompt formats introduce varying biases, leading to different levels of label bias which, in turn, can destabilize ICL performance. Conversely, our UniBias method produces balanced probabilities across labels and significantly mitigates label bias, resulting in more stable and accurate ICL performance and reducing prompt brittleness. \|Method\|Format 1\|Format 2\| Format 3\|Format 4\|Format 5\| \|--\|--\|--\|--\|--\|--\| \| Vanilla ICL\| (0.34, 0.55) \| (0.35, 0.61) \| (0.31, 0.40) \| (0.33, 0.39)\|(0.33, 0.55)\| \| UniBias \| (0.49, 0.49) \| (0.47, 0.48) \| (0.35, 0.35) \|(0.40, 0.39) \|(0.46,0.48)\| We appreciate your constructive comments, which have significantly enhanced the clarity of our paper. We will make sure to incorporate the above discussions into the revised manuscript. --- Moreover, we would like to highlight other enhancements made during the rebuttal process: * Expanded LLM Evaluation: We conduct additional experiments to evaluate UniBias on more LLMs, including GPT-J (6B) and GPT2-XL (1.5B). Detailed results can be found in Table 1 of the attached one-page PDF. * Exploration of Common Biased Components Across Tasks: We identify and eliminate the common biased components within LLMs and evaluate its performance on multiple tasks. Results are detailed in Table 4 of the new PDF. This experiment demonstrates the potential of our work in stimulating diverse bias mitigation methods in a novel perspective of LLM inner structure manipulation. * Enhancement in Grid Search Process: We optimize the grid search process used in our method by providing alternatives of implementing one-time grid search and using unlabeled samples in place of labeled ones. The results of these optimizations are depicted in Table 2 and Figure 1 of the rebuttal PDF. * In-depth Component Analysis: We further visualize and analyze the distribution of identified biased attention heads, as shown in Figure 2 in the PDF. --- We are encouraged to find that these additioanl analysis and experiments, prompted by your valuable insights, have substantially strengthened our paper. We look forward to hearing your feedback! --- References: [1] Calibrate before use: Improving few-shot performance of language models. ICML 2021. [2] Mitigating Label Biases for In-context Learning. ACL 2023. --- Rebuttal Comment 1.1: Comment: Thank you for your comprehensive response. I believe these discussions and explanations indeed serve as excellent supplements to clarity. However, before I finalize my decision, I feel there is still one point that hasn't been covered by these discussions. Even after we've determined the label names, examples, and their order, there's still significant room for variation in the other parts of the prompt beyond these elements. For instance, the way queries and labels are connected, the method of separating examples, the overall format of the input, and so on. These variations, beyond what has already been discussed, can also lead to substantial fluctuations in model performance [1]. However, this situation cannot be explained by any of the three types of bias discussed in this work. [1] Sclar, Melanie, et al. "Quantifying Language Models' Sensitivity to Spurious Features in Prompt Design or: How I learned to start worrying about prompt formatting." The Twelfth International Conference on Learning Representations. --- Rebuttal 2: Title: Response to Reviewer CnNd Comment: Dear Reviewer CnNd, Thank you for your prompt response and for recognizing the discussions presented in our rebuttal! We greatly appreciate your insightful feedback regarding the broad spectrum of prompt variations that can lead to fluctuations in model performance. We are glad to further discuss this important issue. We completely agree with your insight about the significant room for variation across different parts of a prompt. Actually, this insight has been a key motivation for the development of our UniBias method. We believe that merely identifying and addressing a limited number of specific bias factors based on superficial observations is not optimal or entirely effective for mitigating LLM bias. Indeed, due to the significant room in variations in prompts/models/data corpus, it is nearly impossible to identify and address all bias factors from just analyzing external effects of prompt variations on model performance. Therefore, the key challenge and limitation of current bias mitigation methods lie in designing a method that can universally address all potential factors contributing to LLM bias and fluctuations ICL performance. This challenge motivates the rationale for our UniBias method. UniBias is designed to address all potential biases, whether arising from prompt variations, data corpus discrepancies, or model variations. This is feasible because, fundamentally, all these superficial variations lead to biased behaviours internally in either the FFN vectors or the attention heads—components where nearly all LLM parameters reside. By directly identifying and mitigating biases within these FFN vectors and attention heads, UniBias provides a foundational approach to address all forms of bias, representing a poineering work in achieve this objective and a significant advancement in LLM bias mitigation. The three types of bias factors discussed in the paper are intended to demonstrate the internal mechanisms of these biases observed superficially. Given the vast potential for variations in prompt-induced biases, as you noted, it would be impractical to discuss the internal mechanisms for every possible bias factor. Thank you for highlighting this crucial point, which not only validates a very important strength of our method but also aligns perfectly with our motivation for developing it. We intend to include a discussion on this issue in our revised manuscript. --- Rebuttal Comment 2.1: Comment: Thank you for your further clarification. I agree that addressing as many types of bias as possible is important, and consequently, debiasing the models internally might be the most fundamental approach. I also agree that it's impossible to analyze every type of bias. However, I believe there may be some misunderstanding about my concerns. My remaining concern is specifically about the prompt format bias that I mentioned in my previous comment, which neither your provided literature nor your additional response has addressed. I believe this bias is quite significant but not covered by the three biases discussed in your paper, as mainly evaluated in [1]. By saying this bias is significant, I mean that models are also highly sensitive to prompt formatting, not that there is too much room for prompt formatting to discuss in detail. Therefore, when the claim states "effectively mitigating the models' prompt sensitivity," I assume this type of bias is also addressed. I think the following questions might help you understand my concerns and the discussion I expect to see in the paper more clearly: 1. Does the prompt sensitivity issue that the paper claims to mitigate cover the prompt format bias? 2. If so, since this bias is not discussed, what is your speculation as to why the proposed method can mitigate this? Or are you unsure whether the method indeed mitigates this type of bias since you haven't analyzed it in the paper, and thus the method is not explicitly addressing this? 3. If not, I think it's necessary to discuss more about the scope of prompt sensitivity, explicitly stating what is not addressed by the proposed method. --- Reply to Comment 2.1.1: Title: Response to Reviewer CnNd Comment: Dear Reviewer CnNd, Thank you for your follow-up and for clarifying your concerns regarding prompt format bias. We are pleased to the common ground we share and appologize for any misunderstanding. Your insights are invaluable in refining the depth and clarity of our work! We appreciate the questions you provided, as they are very helpful in facilitating our discussion. We would like to respond by addressing these questions. --- Firstly, our method can mitigate prompt format bias, which is supported by experimental evidence in our paper. We evaluate the ICL performance of both the vanilla LLM and the LLM employing our UniBias method under various prompt formats. These experimental results are depicted in the middle figure of Figure 1 in the paper, with different prompt templates used in this experiment detailed in Table 5 of the Appendix. The prompt formats we evaluated include many of the formats discussed in [1], such as differences in separators (': ' vs ' '), spacing ('\n' vs ' '), casing (Positive/positive—with 'positive' used in prompts in Table 4 of the Appendix), label names (Positive/good/Yes), question formats, and the overall format of the prompts. From Figure 1 of the paper, we can observe that these variations in prompt formats indeed lead to significant fluctuations in ICL performance of the vanilla LLM, as suggest by you and [1]. In contrast, despite these challenging prompt format variations, our UniBias method consistently achieve much more accurate and stable ICL performance across all tested formats. --- Next, we address the important question of why UniBias effectively mitigate prompt format bias: - As concluded in [2] and analyzed our rebuttal, prompt instability arises from the LLMs' bias towards predicting certain answers. Different prompt formats can induce biases of varying directions and extents, leading to different effects of bias, and consequently, fluctuations in ICL performance. This is also evidenced by the table provided in the rebuttal, which shows average prediction probabilities for the five prompt formats evaluated in the prompt formatting experiment in Figure 1 of the paper. - On the other hand, recalling our analysis on mechanistic interpretability of LLMs (detailed in Section 2.1 of the paper): each attention layer FFN layer contribute to the final prediction by adding their output hidden states to the residual stream. These outputs can be viewed as the sum of their respective attention heads and FFN vectors. Integrating these two insights, we can learn that the bias of the overall LLM introduced by prompt format bias can be attributed to biased behaviors internally in either the FFN vectors or the attention heads—components where nearly all LLM parameters reside. Therefore, by directly identifying and mitigating biases within these FFN vectors and attention heads, UniBias provides a foundational approach to address prompt format bias and its resultant fluctuations in ICL performance. Because of the same rationale, UniBias is not only effective against the three types of bias discussed in our paper but is also capable of mitigating prompt format bias, example selection bias (as analyzed in Figure 1), and other potential bias factors. This reflects our objective, as discussed in this discussion thread, of designing a method capable of universally addressing all bias factors that contribute to LLM bias. Finally, we want to emphasize that the your insightful question is crutial in demonstrating why Unibias works. We intend to include a detailed subsection in of our revised manusprict to thoroghly discuss this issue. Thank you again for your thoughtful engagement, which is instrumental in refining our presentation and deepening our analysis. We truly appreciate these inspiring discussions. --- [1] "Quantifying Language Models' Sensitivity to Spurious Features in Prompt Design or: How I learned to start worrying about prompt formatting." ICLR 2024. [2] Calibrate before use: Improving few-shot performance of language models. ICML 2021.	Summary: This paper aims to address prompt brittleness in in-context learning methods introduced by seemingly inconsequential changes such as vanilla label bias, recency bias, and selection bias. They use the logit lens technique, where the linear function is applied from the final layer of a decoder-only transformer (AKA the unembedding matrix) on the intermediate layers to interpret the attention activations and FFN embeddings. They detect biased activations and FFN vectors by their relatedness, residual bias, and low variance. Applying their method results in unchanged or improved performance compared to the baseline. Strengths: - The paper introduces a novel application of mechanistic interpretability to mitigate biases in in-context learning and demonstrates that specific attention heads and FFNs in the decoder can be isolated as responsible for prompt brittleness. - The methodology and findings are insightful and bring consistent performance improvements. - The experiments are extensive, covering various datasets and settings, providing strong evidence for the effectiveness of the proposed methods to support the conclusions. - The paper is in general well-written, with clear explanations of the methodology and results. Weaknesses: - The paper mentions some tasks could share biased components (line 318-321) – based on their experiments, I think it would have been possible to identify these components in this work, and I would have liked to see this result. - The analysis of their results could be expanded further. In table 2 the performance of FFN-only, attention-only, and their full UniBias method is inconsistent – sometimes FFN-only performs the best, sometimes attention-only, with no pattern on dataset task. - It would be interesting to see how bias exists across different layers of the model. Analyzing the distribution and impact of biased components in each layer may provide deeper insights into how biases are propagated and amplified throughout the model. - Some of the mathematical formulations, particularly those related to the identification of biased components, could be more detailed to enhance reproducibility and understanding. The authors could provide more thorough mathematical derivations and examples illustrating how the bias criteria are identified. This would make the methodology clearer and easier to understand. Technical Quality: 4 Clarity: 3 Questions for Authors: - Do you have any insights on why UniBias improves performance on certain tasks but not others? - For vanilla label bias, how do you choose the alternative label names? - It is interesting that the unembedding matrix can also be applied to the attention hidden state. Did you notice any discrepancies in the logit distribution between the FFN and the attention head of the same layer? - The sections and overall flow of Sections 2 and 3 may be organized more clearly - Sections 3.3 and 3.4 seem very small (3 and 11 lines respectively) and could be combined with other sections. - It is unclear how the criterion they introduce (relatedness, bias, low variance) in lines 195-201 relate to the three bias conditions introduced in line 218 - In general, Sections 2 and 3 seem related to each other but the writing does not tie them together clearly (eg. lines 202-205 seem to be describing the importance of low variance to identifying the biases in section 2.2, but it is not stated very clearly) - There are further minor writing issues that could be revised, including various minor typos. Confidence: 4 Soundness: 4 Presentation: 3 Contribution: 3 Limitations: Yes. Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Dear Reviewer 8E5c, We deeply appreciate your thorough review and valuable feedback. Below is a summary of our answers (A) to the weaknesses (W) and questions (Q) you raised. --- [W1]: Show shared biased components across tasks. [A1]: We greatly value your suggestion! Inspired by your insights, we not only list the common biased components, but also explore eliminating common biased components for bias mitigation. Listing of Common Biased Components: The shared biased attention heads with their frequency of occurrence (>=3) are detailed in Table 3 of the rebuttal PDF. The shared biased FFN vectors are not discussed because they are associated with label words, which typically vary across tasks. Eliminating Common Biased Components: Inspired by your feedback, we explore the potential of eliminating common biased components and apply it to all tasks, rather than handling each task individually. We conduct additional experiments on multiple tasks to assess the effectiveness of directly elinimate these components. Experimental results in Table 4 of the new PDF indicate that although not as effective as our full Unibias method, it outperforms the vanilla ICL by a large margin. Notably, this is a free gain in performance as it involves merely the direct masking of the biased components identified in our work and is applicable to diverse tasks. --- [W2]: Result analysis could be expanded for Table 2. [A2]: We deeply appreciate your suggestion and provide a detailed analysis here. As an ablation experiment, we find that both FFN-only and attention-only methods outperform the standard ICL, demonstrating their effectiveness. However, as you suggested, there is no consistent pattern on when these two ablation methods perform better. The removal of biased FFN vectors predominantly addresses uncontextual biases associated with label words, while the removal of biased attention heads targets contextual biases, such as recency bias. Therefore, the relative effectiveness of these methods depends on the dominance of either type of bias in the given task and the specifics of the prompt examples, which can vary unpredictably. Despite these variations, it is notable that our comprehensive UniBias method consistently yields the best performance, underscoring the effectiveness and robustness of UniBias. --- [W3]: It would be interesting to see how bias exists across different layers of the model. [A3]: Thanks for your insightful suggestion! We visualize the identified biased attention heads in Figure 2 of the new PDF and provide analysis below. Biased FFN vectors are not visualized due to the high dimensionality of FFN layers. In Figure 2, the intensity of color corresponds to the frequency with which an attention head is identified as biased across 12 datasets, each repeated 5 times. The visualization reveals that biased attention heads mostly occur in the middle layers. This observation aligns with the established understanding of LLM architecture: early layers tend to encode shallow and discrete concepts such as lexical features, while higher layers are responsible for high-level semantics and complex reasoning. Therefore, discrete concepts at early layers are less likely to directly contribute to the bias. In contrast, middle layers, where simple reasoning begins to form are more likely to introduce biases (e.g., "A sentence is always positive rgardless of its content"). The higher layers are less likely to produce this kind of bias associated with simpler, shallow reasoning. --- [W4]: Some mathematical formulations can be more detailed. [A4]: We appreciate the valuable suggestion! The formulations for identifying biased components involve measuring the sum of label logits, the bias of label logits, and the coefficient variance of FFN vector coefficients or attention head label logits. They are corresponding to the relatedness, biased, and low variance criterions. Based on your feedback, we will include more comprehensive mathematical derivations and examples in the revised paper to clearly illustrate how these bias criteria are identified and applied. --- We apologize for the brevity of the following responses due to space limitations. --- [Q1]: Unibias performance [A5]: Thank you for this important question. To address it, it's essential to consider that a more appropriate comparison for UniBias would be between standard ICL and UniBias, as this comparison contrasts ICL based on vanilla LLMs with LLMs after the elimination of biased components. In this case, our UniBias consistently improve performance across all 12 datasets we evaluated. On the other hand, calibration methods are under different methodology and perform better on a limited number of datasets. This may be attributed to the specific characteristics of each dataset. --- [Q2]: Alternative label names for vanilla label bias [A6]: They are generated by GPT-4. --- [Q3]: Discrepancies in logit distribution between FFN and attention head [A7]: Thank you for the insightful question. Generally, they are consistent across the middle and higher layers. However, we observe discrepancies in some individual components. --- [Q4]: Organization of Sections 2 and 3. [A8]: Thank you for the detailed suggestions! We will revise Sections 2 and 3 carefully. This includes merging Sections 3.3 and 3.4, enhancing coherence between the criteria and conditions, and strengthening connections between 3.3 and 3.4. We will also ensure a thorough review for typos and writing issues as suggested in Q5. --- We wish to highlight that we also extend the evaluation of UniBias to include GPT-J and GPT2-XL, which are detailed in Table 1 of the new PDF. --- We are encouraged to find that these experiments and enhancements, prompted by your valuable insights, have substantially strengthened our paper. We look forward to hearing your feedback!	Summary: This paper seek to address prediction bias in ICL through intervening feedforward vectors and attention heads. Strengths: * This paper studies a critical problem of in ICL. * The proposed method is well-motivated. Weaknesses: The major problem with the proposed method is that it requires approximately 20 labeled instances per class. This leads to the following concerns: * The comparison between some baselines and the proposed method is unfair. The proposed method requires labeled instances for searching thresholds, while baselines like DC require no labeled instances. * The generalizability of the proposed method is also questionable. Can we search thresholds once and apply them to all datasets? If not, it is impractical for real-world application. If we have that many labeled instances and access to the model parameters for the target task and dataset, we can always use them to finetune the model to achieve better performance. * Moreover, the cost will be intractable when the number of classes increase. Technical Quality: 1 Clarity: 2 Questions for Authors: See weaknesses. Confidence: 4 Soundness: 1 Presentation: 2 Contribution: 2 Limitations: This paper does not have a limitation section. Flag For Ethics Review: ['No ethics review needed.'] Rating: 5 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Dear reviewer LH5i, We deeply appreciate your thorough review and valuable feedback. We appreciate your acknowledgment of the significance of the problem we investigated and motivation behind our approach. Below is a summary of our answers (A) to the weaknesses (W) you raised. --- [W1]: Grid search with 20 labeled samples per class can lead to concerns in fairness in comparison with baselines, generalizability of the method, and increased cost when the number of classes is large. [A1]: Thank you for raising this important concern! It prompted us to further enhance our method. Initially, we choose a limited number of labeled samples for grid search because it is typically manageable given the small size of labeled data. Additionally, it is only needed for the initial set up to identify biased LLM components. During inference stage, the computation cost of UniBias is completely identical to that of the vanilla LLMs. We believe this is a reasonable trade-off for the substantial improvements in ICL performance and the novel perspective Unibias offers on mitigating bias through LLM inner components. However, we understand the concern that using labeled samples can be challenging when labeled data are scarce or the number of classes is large. Therefore, we have optimized our method in light of your valuable feedback: Replacing Labeled Samples with Unlabeled Ones: To Address the potential challenge in accessing labeled samples and ensure a fair comparison with baselines, we further explore the alternative of using unlabeled samples during grid search. In our method, labeled samples are used to ensure each class is represented proportionally in the grid search, without direct use of the specific label information. Therefore, for balanced datasets, it is equally effective to employ a slight larger pool of unlabeled samples. Our experimental findings, illustrated in Figure 1 of the rebuttal PDF, indicate that approximately $40 \times \text{number of classes}$ unlabeled samples achieves performance comparable to that obtained with labeled samples. This number is significantly lower than the quantity required by the PC baseline, which requires at least $200 \times \text{number of classes}$ unlabeled samples. Optimization of Grid Search: Inspired by your feedback, we streamline the grid search process. We now provide an alternative to perform grid search on a single dataset to obtain a set of threshold values, which are then applied universally across other datasets. This alternative eliminates the need for repeated grid searches across each dataset, significantly enhancing efficiency and scalability of the method. By directly adopting the fixed threshold, it also addesses the scalability issues associated with an increasing number of classes. Our experimental results in Table 2 of the rebuttal PDF confirm that using a fixed set of thresholds maintains good performance, slightly below the original UniBias method but still superior to vanilla ICL and other baseline methods. This also illustrate the robustness of our method in terms of threshold selection. Exploration of Eliminating Common Biased Components: In response to your feedback, we have delved deeper into the potential of eliminating common biased components and apply it to all tasks, rather than handling each task individually. In Section 4.4 of our paper, we discuss the presence of shared biased LLM components across different tasks. Inspired by this finding, we conduct additional experiments to directly elinimate these common biased components and evaluate on different tasks. Experimental results in Table 4 of the rebuttal PDF indicate that although not as effective as our full Unibias method, it outperforms the vanilla ICL by a large margin. Notably, this is a free gain in performance as it involves merely applying a direct masking to the biased components identified in our work and is applicable to diverse tasks. Finally, following your suggestion, we evaluate the performance of fine-tuning LLMs using 20 labeled samples per class and find that their performance is significantly lower than that of UniBias. This results further validates the effectiveness of our method. \| \|SST2 \|MNLI\|Trec \|--\|--\|--\|--\| \| ICL \| 87.22 \| 53.83 \|72.92\| \| ICL after SFT \|88.64 \| 53.95 \|73.2\| \| UniBias \| 94.54 \|54.97\|80.80\| The enhancements suggested by your feedback have markedly improved the efficiency and scalability of our approach. Furthermore, the third experiment underscores the potential of our work in stimulating diverse bias mitigation methods in a novel perspective of LLM inner structure manipulation. --- During rebuttal, we also extend the evaluation of UniBias to include GPT-J and GPT2-XL, which are detailed in Table 1 of the new PDF. --- We would like to further highlight the contributions of our work. * A New Insight for LLM Bias Mitigation: Unlike existing works based on external adjustments of LLM outputs, we mitigate LLM bias through manipulation of LLM internal structure. This novel perspective could potentially stimulate future research on LLM bias mitigation from inner structure manipulation, offering a new direction for the field. * Unveil Internal Mechanisms of LLM Bias: We conduct a thorough exploration of the internal mechanisms underlying biases in LLMs, providing deep insights that go beyond superficial observations, revealing the inner causes of these biases. * Extensive Evaluation: UniBias is evaluate across 12 datasets and 4 LLMs, consistently demonstrating its effectiveness. * Addressing Prompt Brittleness: Our experimental results show that our method effectively mitigates prompt brittleness, a critical issue in ICL. --- We are encouraged to find that these experiments and enhancements, prompted by your valuable insights, have substantially strengthened our paper. We look forward to hearing your feedback!	Summary: The paper explores the influence of FFNs and attention heads in LLMs on biases, resulting in model predictions that exhibit favoritism towards specific labels. The authors propose UniBias, an inference-only technique designed to detect and mitigate biased components within LLMs by analyzing and manipulating FFN vectors and attention heads. Experiment results on 12 diverse NLP datasets show that UniBias significantly improves the in-context learning performance of LLMs and reduces their sensitivity to prompt design. Strengths: - The paper introduces UniBias, a novel method designed to mitigate bias in LLMs by manipulating internal model components, specifically FFN vectors and attention heads. This approach represents a unique contribution to bias reduction within LLMs. - UniBias demonstrates significant improvements over standard ICL and state-of-the-art calibration methods, indicating its potential to become a leading technique in bias mitigation for LLMs. - The authors conduct an in-depth exploration of the internal mechanisms underlying biases in LLMs, providing insights that delve beyond superficial observations to uncover the root causes of these biases. Weaknesses: - While the method shows promising results, it is crucial to assess how well these findings generalize across different types of LLMs and datasets (e.g., machine translation tasks) beyond those tested. This paper only conducts experiments on Llama and classification tasks. - The reliance on grid search with a limited number of labeled training samples for identifying biased components may limit scalability and efficiency. Technical Quality: 3 Clarity: 3 Questions for Authors: - Is this analytical method compatible with LLMs that have undergone SFT? If so, how would the fine-tuning process potentially influence the identification and mitigation of biases within the model's internal mechanisms? - Can the grid search process be automated or optimized to minimize the need for manual parameter tuning? Confidence: 3 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: None Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Dear Reviewer BhZE, We deeply appreciate your thorough review and valuable feedback. Below is a summary of our answers (A) to the weaknesses (W) and questions (Q) you raised in the review. --- [W1]: Assess how well the findings generalize across different types of LLMs and datasets. [A1]: We greatly value your insightful suggestion and we have conducted additional experiments to validate our method. Model Generalization: Following your suggestion, we have conducted additional experiments to evaluate more LLMs, including GPT-J (6B) and GPT2-XL (1.5B). The experimental results are detailed in Table 1 of the rebuttal PDF. Together with the Llama2-7B and Llama2-13B models evaluated in our paper, we rigorously evaluate our method across four popular LLMs of different sizes and demonstrated the effective of our method across these models. Task Coverage: Our work is currently focus on classification tasks, consistent with the prevailing LLM bias mitigation literature. Our work evaluates extensive array of tasks compared to the exsiting studies. To provide context, current works typically focus either on mitigating bias in general classification tasks [1-2] or exclusively address bias in multiple-choice questions (MCQs) in reasoning tasks [3-4]. In contrast, our study covers 12 datasets spanning 5 distinct tasks, including both general classification tasks and MCQs in reasoning tasks. This breadth not only aligns with but surpasses the scope of tasks examined in the literature. Moving forward, we are enthusiastic about adapting our method to a wider variety of tasks as you suggested. We appreciate your insights, which significantly strengthened our experimental validation. --- [W2]: Grid search with a limited number of labeled samples may limit scalability and efficiency [A2]: Thank you for raising this concern, which prompted us to further enhance our method. Although using a small set of labeled samples is typically manageable, we recognize the potential for improvement. In response, we conduct experiments to address your concern from three aspects. Optimization of Grid Search: To streamline the grid search process, we now provide an alternative to perform grid search on a single dataset to obtain a set of threshold values, which are then applied universally across other datasets. This alternative eliminates the need for repeated grid searches across each dataset, significantly enhancing efficiency and scalability of the method. Our experimental results in Table 2 of the rebuttal PDF confirm that using a fixed set of thresholds maintains good performance, slightly below the original UniBias method but still superior to vanilla ICL and other baseline methods. This experiment also illustrate the robustness of our method in terms of threshold selection. Replacing Labeled Samples with Unlabeled Ones: To Address the potential challenge in accessing labeled samples, we further explore the alternative of using unlabeled samples during grid search. In our method, labeled samples are used to ensure each class is represented proportionally in the grid search, without direct use of the specific label information. Therefore, for balanced datasets, it is equally effective to employ a larger pool of unlabeled samples. Our experimental findings, illustrated in Figure 1 of the rebuttal PDF, indicate that approximately $40 \times \text{number of classes}$ unlabeled samples can achieve performance comparable to that obtained with labeled samples. Exploration of Eliminating Common Biased Components: In response to your feedback, we have delved deeper into the potential of eliminating common biased components and apply it to all tasks, rather than handling each task individually. In Section 4.4 of our paper, we discuss the presence of shared biased LLM components across different tasks. Inspired by this finding, we conduct additional experiments to directly elinimate these common biased components and evaluate on different tasks. Experimental results in Table 3 of the rebuttal PDF indicate that although not as effective as our full Unibias method, it outperforms the vanilla ICL by a large margin. Notably, this is a free gain in performance as it involves merely applying a direct masking to the biased components identified in our work and is applicable to diverse tasks. The enhancements suggested by your feedback have markedly improved the efficiency and scalability of our approach. Furthermore, the third experiment indicates the potential of our work in stimulating future bias mitigation methods in a novel perspective of LLM inner structure manipulation. --- [Q1]: Compatibility with SFT [A3]: Thank you for the interesting question. We would like to respond in two aspects: the performance of our method on fully fine-tuned models and the impact of SFT on biased components. Compatibility with SFT: UniBias is fully compatible with LLMs that have undergone SFT. We applied UniBias to Llama2-7b fine-tuned on the SST2 and MNLI datasets, with following experimental results: \|\|SST2\|MNLI \|--\|--\|--\| \|ICL after SFT\|94.19\|64.78 \|Unibias\|94.72\|65.88 Impact of SFT on Biased LLM Components: Algorithmically, the processes for identifying and mitigating biases in both SFT and non-SFT models are identical. However, we observed a reduction in the coefficients of biased FFN vectors and the magnitudes of label logits of biased attention heads post-SFT in some cases, suggesting that the SFT process may suppress the biased components in LLMs. --- [Q2]: Can grid search process be automated? [A4]: Yes, the grid search process in our method is automated. We will release our code to facilitate reproduction and further exploration. --- We are encouraged to find that these experiments and enhancements, prompted by your valuable insights, have substantially strengthened our paper. We look forward to hearing your feedback! --- Rebuttal Comment 1.1: Title: References Comment: The references in the rebuttal are as follows: [1] Mitigating Label Biases for In-context Learning. ACL 2023. [2] Prototypical Calibration for Few-shot Learning of Language Models. ICLR 2023 [3] Large Language Models Are Not Robust Multiple-Choice Selectors. ICLR 2023 [4] Large Language Models Sensitivity to The Order of Options in Multiple-Choice Questions. Findings of NAACL 2024.	Rebuttal 1: Rebuttal: Dear Reviewers, We would like to express our gratitude for the time and effort you've dedicated to reviewing our paper. We are deeply grateful for your recognition of our work: * The proposed UniBias method is novel (Reviewer BhZE, 8E5c), well motivated (Reviewer LH5i), straightforward and does not introduce any additional inference time cost (Reviewer CnNd), and has potential to become a leading technique in LLM bias mitigation (Reviewer BhZE). * The exploration on internal mechanisms of underlying biases in LLMs is a relatively unexplored area (Reviewer CnNd) and provides insights that uncover the root causes of these biases (Reviewer BhZE). * The methodologies and findings are insightful (Reviewer 8E5c). * Experiments are extensive (Reviewer 8E5c), demonstrating significant improvements over standard ICL and SOTA calibration methods (Reviewer BhZE). * A critical problem in ICL is investigated in this paper (Reviewer LH5i) and the work represents a unique contribution to bias reduction within LLMs (Reviewer BhZE). * The paper is well-written (Reviewer 8E5c). --- In response to your valuable feedback, we have carefully responded to each comment and wish to outline the main changes as follows: * Expanded LLM Evaluation: We conduct additional experiments to evaluate UniBias on more LLMs, including GPT-J (6B) and GPT2-XL (1.5B). Detailed results can be found in Table 1 of the attached one-page PDF. * Exploration of Common Biased Components Across Tasks: We identify and eliminate the common biased components within LLMs and evaluate its performance on multiple tasks. Results are detailed in Table 4 of the attached PDF. This experiment demonstrates the potential of our work in stimulating diverse bias mitigation methods in a novel perspective of LLM inner structure manipulation. * Enhancement in Grid Search Process: We optimize the grid search process used in our method by providing alternatives of implementing one-time grid search and using unlabeled samples in place of labeled ones. The results of these optimizations are depicted in Table 2 and Figure 1 of the rebuttal PDF. * In-Depth Analysis: We further analyze why mitigating LLMs' bias towards labels can alleviate prompt brittleness in our method. We also visualize and analyze the distribution of identified biased attention heads, as shown in Figure 2 in the PDF. --- Moreover, We would like to further highlight the contributions of our work. * A New Insight for LLM Bias Mitigation: Unlike existing works based on external adjustments of LLM outputs, we mitigate LLM bias through manipulation of LLM internal structure. This novel perspective could potentially stimulate future research on LLM bias mitigation from inner structure manipulation, offering a new direction for the field. * Unveil Internal Mechanisms of LLM Bias: We conduct a thorough exploration of the internal mechanisms underlying biases in LLMs, providing deep insights that go beyond superficial observations, revealing the inner causes of these biases. * Extensive Evaluation: UniBias is evaluate across 12 datasets and 4 LLMs, consistently demonstrating its effectiveness. * Addressing Prompt Brittleness: ur experimental results show that our method effectively mitigates prompt brittleness, a critical issue in ICL. --- In light of your valuable feedback, we find our work has been significantly enhanced during the rebuttal. Again, we would like thank you for your dedication throughout the review process. Details of our rebuttal are outlined in our individual responses and the attached rebuttal PDF. Please do not hesitate to contact us for any further suggestions or discussions. With Gratitude, Authors of Paper 12087 Pdf: /pdf/0dfd939d8809bafc4f6ae173338981dbb2b55859.pdf	NeurIPS_2024_submissions_huggingface	2,024	null	null	null	null	null	null	null	null
Latent Intrinsics Emerge from Training to Relight	Accept (spotlight)	Summary: This paper proposes a fully data-driven relighting method applicable to images and real scenes. The approach requires only paired images of the same scene under different illumination as inputs. The trained model can also produce albedo-like maps even though it is not trained without such supervision. The experimental results show that it significantly outperforms unsupervised baselines while being competitive with baselines that use supervision. Strengths: 1. The idea and the overall achieved goal are quite novel. Training models that can do relighting without any supervision on ground-truth intrinsic properties (albedo, depth, normal, lighting, etc.) is quite challenging and useful. 2. The presentation is clear. The overall methodology is quite straightforward, with all the details (inputs, outputs, intermediate steps, losses) well-explained and self-contained. 3. The experiments are sufficient. Weaknesses: 1. The paper proposes several novel regularizations, e.g. Eq 6 and Eq 9, which are not ablated. It would be good to see how much these losses improve the model's accuracy. Technical Quality: 4 Clarity: 4 Questions for Authors: Please see the weaknesses. Confidence: 4 Soundness: 4 Presentation: 4 Contribution: 4 Limitations: The limitations and potential negative societal impact are discussed properly. Flag For Ethics Review: ['No ethics review needed.'] Rating: 7 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We appreciate Reviewer n7pv's recognition of the novelty and usefulness of our work, particularly in training models for relighting without supervision on ground-truth intrinsic properties. We are also grateful for the positive feedback on the clarity and straightforwardness of our presentation, and the comprehensiveness of the methodology and experiments. Regarding Ablation of regularization terms Thank you for suggesting additional ablation experiments. In response, we trained a model using our optimal configuration ($\alpha=0.5$) for the relighting experiment, excluding the regularization terms in Eq. 6 and Eq. 9 from the training objectives. The results, presented in the attached tables, clearly show that the proposed regularization terms significantly improve performance in both relighting and albedo estimation tasks. \|Relight Config \|Raw Output\|Color Correction\| \|---\|---\|----\| \| \| RMSE SSIM\| RMSE SSIM\| \|w/ regularization \| 0.297 0.473 \| 0.222 0.571\| \|w/o regularization \| 0.315 0.462 \| 0.232 0.550\| Table 1: Ablation experiments of regularization term for image relighting. \_ \|Albedo Config \|$\delta = 0.1$\|optimal $\delta$\| \|---\|---\|----\| \|w/ regularization \| 31.81 \| 19.53 \| \|w/o regularization \| 32.56 \| 19.67 \| Table 2: Ablation experiments of regularization term for unsupervised albedo estimation. --- Rebuttal Comment 1.1: Comment: Thanks for the update! The above reply resolves my concerns and I have no further questions. I will keep my score as accept (7). --- Reply to Comment 1.1.1: Comment: Thank you for acknowledging our response, and we're glad we were able to address your concerns. We will include the results of the ablation studies in our final version.	Summary: This paper proposes a fully data driven method to perform scene relighting that does not require any groundtruth lighting supervision. As such, it allows for scene relighting by training only on real paired images of the same scene under different illuminations rather than requiring synthetic data that contains the groundtruth lighting environments, which improves the generalizability and performance on real-world scenes. The method involves learning both intrinsic and extrinsic features from each scene, where the intrinsics represent the albedo and geometry and the extrinsics represent the lighting. The authors design a fairly simple but well thought out pipeline with several loss functions that encourage the intrinsics to be consistent for the same scene under different illumination, self reconstruction losses that reconstruct the original image using its own intrinsic and extrinsic features, and relighting loss functions that involve swapping extrinsic features between image pairs to perform scene relighting. The authors also design a constrained scaling mechanism that prevents too much information outside of the lighting information in the target image from being transferred to the source image. Experiments demonstrate that the relighting performance greatly outperforms other unsupervised methods that don't require groundtruth lighting and is comparable or slightly better depending on the metric than the existing state of the art supervised scene relighting methods. Ablations for the constrained scaling are also provided, which is helpful to understand the importance of this technical contribution. Strengths: Experiments in the paper are quite thorough and demonstrate state-of-the-art performance for scene relighting, most notably being comparable to methods that require full light supervision and greatly outperforming unsupervised methods. Ablation studies on the constrained scaling are also helpful for understanding this contribution. The model and pipeline itself is quite simple and straightforward yet effective and easy to follow. The loss functions align with intuition and are carefully designed. The ability to estimate albedo from a scene is an important feature that has many implications for downstream tasks, and the authors achieve state-of-the-art performance for albedo estimation. Weaknesses: One experiment that would be interesting to see is the level of albedo consistency. Since the authors design a loss function that encourages intrinsic features of the same scene under different illuminations to be similar, it would be good to see how well that design performs empirically, both quantitatively and qualitatively. Some citations are missing for the image-based relighting domain, namely: 1. Towards High Fidelity Face Relighting with Realistic Shadows (CVPR 2021) 2. Learning to Relight Portraits for Background Replacement (SIGGRAPH 2021) 3. Face Relighting with Geometrically Consistent Shadows (CVPR 2022) 4. Lumos: Learning to Relight Portrait Images via a Virtual Light Stage and Synthetic-to-Real Adaptation (SIGGRAPH Asia 2022) 5. DiFaReli: Diffusion Face Relighting (ICCV 2023) Technical Quality: 3 Clarity: 3 Questions for Authors: As I mentioned above, the experiments in this paper are quite thorough and I'm mostly interested to see whether the estimated albedo is consistent for the same scene under different illuminations. It would be good to compare this with other methods as well. Otherwise, please add the missing citations for the image-based relighting domain as there are many recent methods in that area. Confidence: 4 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: The limitations section is well thought out and touches on many important points. Relighting often does not have significant negative societal impacts since only the illumination is edited, but as with any other editing work there is always the potential to generate fake content. Perhaps this could be mentioned in the potential negative societal impact section. Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We appreciate Reviewer t1Cn's feedback and are grateful for their recognition of the simplicity and effectiveness of our proposed method, the thoroughness of our experiments, and our approach to unsupervised training. We are pleased to see the acknowledgment of our design choices, the clarity and straightforwardness of our paper's writing, and the practical implications of our method for downstream tasks. We now respond to the reviewer's questions. Regarding Consistency of Our Albedo Predictions We conducted an experiment to evaluate the consistency of albedo predictions under varying lighting conditions. Specifically, we estimated albedo for the same scene under different lighting settings, as provided in the MIT dataset. Qualitative visualizations are included in the attached PDF. For comparison, we also evaluated the state-of-the-art supervised intrinsic estimator, Intrinsic Image Diffusion (IID) [1]. Notably, IID requires 10 independent estimations followed by averaging to produce a single image estimation. Due to the computational expense of this process, we assessed albedo consistency under only five different lighting settings. Our method, on the other hand, produces consistent results with just a single forward pass. Our findings demonstrate that our approach achieves stable albedo estimation despite not utilizing any albedo-like maps as supervision. Also, while IID often exhibits significant color drift due to its reliance on synthetic training data, our method maintains color fidelity, as shown in our qualitative figures. To quantify the albedo stability under varying lighting conditions, we report the mean deviation and standard deviation on albedo map, normalized per scene, \|Results \|Mean Deviataion\| Standard Deviation\| \|---\|---\|---\| \|IID \| 0.054\| 0.063\| \|Ours \| 0.046\| 0.053\| Our method produces more stable albedo estimations under varying lighting conditions compared to the supervised IID approach. [1] Kocsis et al., Intrinsic Image Diffusion for Indoor Single-view Material Estimation. CVPR 2024. On Missed citations Thank you for bringing these papers to our attention. We will ensure that these citations are added to the related work section in our final draft. --- Rebuttal Comment 1.1: Comment: Thank you for your detailed rebuttal! My concerns are addressed and I am convinced by the authors' rebuttal with regard to other reviewers' concerns as well. I will maintain my Weak Accept rating. Please be sure to include missing details in the final version, especially the ablation tables. --- Reply to Comment 1.1.1: Comment: Thank you for acknowledging our response, and we're glad we could address your concerns. We will incorporate your suggested references, along with the additional visualizations and results, in our final version.	Summary: The paper proposes a 2D relighting pipeline based on latent space manipulation, which is purely data-driven without explicit intrinsic representations such as geometry and materials. Given a single image input, the proposed model recovers latent variables representing scene intrinsic properties and latent variables representing lighting, enabling applications like 2D relighting based on a single reference image and albedo estimation. The proposed methods are validated on a held-out dataset, demonstrating its generalization capability and precision in real-world scenarios. Strengths: 1. The paper is well-written and easy to understand. 2. The idea of relighting images with latent latent intrinsic and without explicit explicit intrinsic representations is interesting and novel. The paper shows impressive results on the test datasets and outperforms the baselines for relighting and albedo estimation. 3. The proposed method doesn't need explicit lighting supervision and is purely data-driven, which indicates the method may have the potential to be scaled up in the future. Weaknesses: 1. using latent intrinsic to relight means the users lose the ability to control the relighting in a fine-grained way. For example, the user can't precisely control the lighting intensity and directions. 2. The proposed method doesn't show enough generalization ability. In Sec. 4.2, the authors first mention that they train the model on the Multi-illumination training dataset and then test it on the Multi-illumination test dataset. Then, when the authors validate the method on StyLitGAN images, the authors mention that they train the model again on StyLitGAN images. (as said in L213). It seems to mean that the trained model can't be used to test out of the distribution images directly. To show enough generalization ability, the model should be able to test on unseen real images directly after training ends. Technical Quality: 3 Clarity: 2 Questions for Authors: What if the authors use the trained model to test some unseen images on the internet? Will the model still show good relighting ability? Can the checkpoint trained on the Multi-illumination dataset be used to infer StyLitGAN images? Some results would be appreciated. Confidence: 4 Soundness: 3 Presentation: 2 Contribution: 3 Limitations: The limitations have been well discussed by the authors in the paper. Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We appreciate Reviewer VUQQ's feedback and are grateful for their recognition of the novelty of our proposed method, our approach to unsupervised training, and the quality of our results and paper writing. We now address the specific concerns and questions raised by the reviewer. Concern: Lack of Fine-Grained Control in Relighting because of latent lighting representation While it is true that using latent lighting representation for relighting may limit the ability to achieve fine-grained control over specific lighting parameters such as intensity and direction, this limitation is not unique to our approach. Current explicit parametric methods, like spherical harmonics or spherical Gaussians, also face challenges in providing precise control over these aspects, often due to the complexity and computational cost involved. Our approach, however, offers a significant advantage by simplifying the relighting procedure. By learning a high-dimensional latent representation, we bypass the need for detailed manual tuning of lighting parameters, making the process more efficient and accessible. Additionally, we have demonstrated the ability to interpolate in the latent space to achieve various lighting effects, as shown in Figure 4. This capability suggests that while explicit fine-grained control is limited, our method provides a practical and scalable solution for many real-world applications where such precision may not be critical. Future work will explore ways to enhance control, potentially by mapping latent representations to explicit lighting parameters for users who require finer-detailed adjustments. Concern: Generalization Ability We respectfully disagree with the reviewer's assessment regarding the generalization capabilities of our model. Our paper demonstrates two key aspects of the model's generalization ability, which we detail below: 1) Generalization to Unpaired Images: While our model is trained on paired images captured under the same scene, we evaluate its performance on unpaired images. This is illustrated in the first two columns of Figure 3, where the input and reference images originate from different scenes. We opted for paired training data due to practical scalability considerations, as using unpaired data necessitates the calibration of exact extrinsic lighting conditions across different scenes. Our model significantly outperforms comparable approaches in accurately rendering relit patterns while preserving environmental details. 2) Generalization to Out-of-Distribution Images: We trained our model on the Multi-Illumination dataset and directly tested it on the IIW dataset for albedo estimation, without any additional training or prior exposure to the IIW dataset. Despite the considerable distribution shift between these datasets, our method consistently outperforms other approaches in albedo estimation, even without using any albedo-like supervision. To further substantiate our model's generalization capability, we provide additional results as suggested by the reviewer. We tested the relighting performance of models trained on different datasets, including demonstrating plausible relighting of StyLitGAN and IIW images using models trained on the MIT Multi-Illumination dataset. These experiments show the model's ability to generalize to unseen images using two sources of extrinsics: (a) estimating extrinsics from in-distribution images and (b) estimating extrinsics from out-of-distribution images. Visualizations of these results are included in the attached PDF. While these results demonstrate plausible relighting, we acknowledge that they are not perfect, primarily due to the distinct nature of the distributions. The MIT Multi-Illumination dataset does not feature visible luminaires or light sources, whereas StyLitGAN images may include visible luminaires, with the quality dependent on StyleGAN's generative capabilities. We believe that expanding our training dataset will enhance rendering quality, as our approach can easily scale due to its lack of reliance on supervision. We appreciate the reviewer's feedback and questions, and we will incorporate these clarifications into the final version of our manuscript for improved clarity. --- Rebuttal Comment 1.1: Comment: Thanks for the authors' response. My concerns have been addressed, and I will raise my score to 6. I also have some further suggestions for the authors, which are purely recommendations and do not affect the evaluation of this paper: 1. I suggest that the authors consider building a project website in the future. This would allow them to showcase more results and, more importantly, present video results of Latent Extrinsic Interpolation (as shown in Fig. 4), which could significantly enhance the paper's presentation. (BTW, can the current methods get consistent video results?) 2. Regarding the "Generalization to Out-of-Distribution" problem, as the authors have acknowledged, there is a noticeable performance drop when dealing with data that has different distributions. In addition to expanding the dataset, as the authors mentioned, combining the current model designs with large generative models, such as Stable Diffusion, might improve performance and generalization. This could be an interesting direction for future work. --- Reply to Comment 1.1.1: Comment: Thank you for acknowledging our response, and we’re pleased that we were able to address your concerns. We also appreciate your suggestion of creating a project webpage for an interactive demonstration. Your follow-up idea of leveraging the richer structural priors of the generative model to enhance performance sounds both interesting and promising—thank you for proposing it.	null	null	Rebuttal 1: Rebuttal: We thank the reviewers for their positive and constructive feedback on our paper. Reviewers consistently highlighted the "novelty" [VUQQ, n7pv] of our approach, particularly appreciating the method's ability to perform relighting using latent intrinsic properties without explicit representations, described as "interesting and novel" [VUQQ]. The paper was noted for being "well-written and easy to understand" [VUQQ], with a "clear" [n7pv] and "straightforward" [n7pv] methodology. Reviewers also commended the "thorough" [t1Cn] and "sufficient" [n7pv] experiments, which demonstrate "state-of-the-art performance" [t1Cn] in scene relighting and albedo estimation and "outperforms baselines" [VUQQ, t1Cn]. We are extremely delighted to receive such feedback. Thank you! Further positive points include the "simplicity and effectiveness" [t1Cn] of our model and pipeline, the careful design of loss functions that "align with intuition" [t1Cn], and the important feature of estimating albedo, which has significant "implications for downstream tasks [t1Cn]." The method's "data-driven" nature, which "doesn't need explicit lighting supervision," suggests potential for future scalability [VUQQ]. Overall, the reviewers appreciated the challenging and useful nature of training models for relighting without supervision on ground-truth intrinsic properties [n7pv]. With the clarifications provided (detailed responses to each reviewer below), we hope our paper stands as a meaningful contribution to the community. Here is a summary of our responses to the reviewers' questions and concerns, along with summaries of results and figures included in our rebuttal PDF: - For VUQQ: We provide visualizations of relighting unseen images to demonstrate the generalization and robustness of our approach. - For t1Cn: We offer results on albedo prediction and its variance under various lighting conditions to demonstrate the stability of the learned intrinsic representation. - For n7pv: We present an ablation study excluding regularization losses, demonstrating their necessity for improved relighting and albedo prediction. Pdf: /pdf/181e758f3cbcf885c0f7ed9b9d0895481625664d.pdf	NeurIPS_2024_submissions_huggingface	2,024	null	null	null	null	null	null	null	null
Exploiting the Replay Memory Before Exploring the Environment: Enhancing Reinforcement Learning Through Empirical MDP Iteration	Accept (poster)	Summary: When using the Bellman update with incomplete data, the estimation error is hard to eliminate. To solve this problem, the authors develop a novel framework EMIT, which can be used to enhance existing RL algorithms by iteratively solving a current empirical MDP for stable finite-time performance, and can progressively approach a solution to the original MDP. Strengths: 1. The experiment is performed on both MuJoCo and Atari. 2. Propositions 3.2-3.4 explain the estimation error in RL. Weaknesses: 1. The running time of the different RL methods should be compared. 2. The proposed method should be compared with recent RL methods, such as TD7 and CrossQ [1, 2]. 3. A head-to-head comparison between your memory buffer and the existing ones should be given [3]. 4. From Algorithm 1, it is hard to see the difference between the proposed method and the existing ones. If the authors can present their contributions clearly and open-source the code, I will increase my rating. [1]Fujimoto S, Chang W D, Smith E, et al. For sale: State-action representation learning for deep reinforcement learning[J]. Advances in Neural Information Processing Systems, 2024, 36. [2] Bhatt A, Palenicek D, Belousov B, et al. CrossQ: Batch Normalization in Deep Reinforcement Learning for Greater Sample Efficiency and Simplicity[C]//The Twelfth International Conference on Learning Representations. [3] Schaul T, Quan J, Antonoglou I, et al. Prioritized experience replay[J]. arXiv preprint arXiv:1511.05952, 2015. Technical Quality: 2 Clarity: 2 Questions for Authors: 1. Can EMIT improve the performances of the recently proposed RL methods? 2. For continuous control, e.g., MuJoCo, in most cases, the current state-action pair will not be in the memory buffer D. Thus, $R = -\infty$ for this state-action pair. How to deal with this issue? Confidence: 5 Soundness: 2 Presentation: 2 Contribution: 1 Limitations: N/A Flag For Ethics Review: ['No ethics review needed.'] Rating: 5 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: - Can EMIT improve the performances of the recently proposed RL methods such as TD7 and CrossQ? We instantiate EMIT with TD7, and results are shown in Fig.1(b) in the rebuttal pdf. We find that EMIT can enhance the performance of TD7 on Ant and get similar performance on HalfCheetah. It may because TD7 is already a very strong baseline on HalfCheetah. We will discuss TD7 and CrossQ to provide a more comprehensive comparison in the revised manuscript. - For continuous control, e.g., MuJoCo, in most cases, the current state-action pair will not be in the memory buffer D. Thus, $R=-\infty$ for this state-action pair. How to deal with this issue? For the in-sample Bellman update as in eq.(3), we sample (s,a,r,s',a') in the buffer for the update. This r is not $-\infty$. - The running time of the different RL methods should be compared. We compare the running time of EMIT with DQN and TD3 using Frame Per Second (FPS) in the table below. EMIT consumes nearly double the time of DQN and TD3. \|\| EMIT-DQN \| Rainbow \| DQN \| IQN \| C51 \| \|-\| :-: \| :-: \|:-: \|:-: \|:-: \| \|FPS\| 103.1 $\pm$ 4.2 \| 124.0 $\pm$ 5.8 \| 194.4 $\pm$ 5.7 \| 179.1 $\pm$ 2.9 \| 170.6 $\pm$ 5.6 \| \| \| EMIT-TD3 \| SAC \| TD3 \| XQL \| TRPO \| PPO \| \| :-: \| :-: \| :-: \|:-: \|:-: \|:-: \|:-: \| \| FPS \| 42.2 $\pm$ 0.2 \| 70.4 $\pm$ 0.3 \| 79.4 $\pm$ 0.3\| 68.2 $\pm$ 0.5 \| 108.9 $\pm$ 3.1 \| 166.6 $\pm$ 3.5 \| Despite this limitation, as demonstrated in Figure 3 of our manuscript and summarized in the table below, EMIT still outperforms DQN and TD3 when using the same amount of computation, underscoring its efficacy. \| \| EMIT-DQN (5e6) \| DQN (1e7) \| \| :-----\| ----: \| :----: \| \| Asteroids \| 1338.9 $\pm$ 27.8 \| 948.6 $\pm$ 61.9\| \| Atlantis \| 1599569.9 $\pm$ 231875.6\| 910471.2 $\pm$ 35727.5\| \| Breakout \| 266.2 $\pm$ 40.4\| 103.3 $\pm$ 8.0\| \| Gravitar \| 949.7 $\pm$ 14.9\| 523.5 $\pm$ 68.3\| \| \| EMIT-TD3 (1e6) \| TD3 (2e6) \| \| :-----\| ----: \| :----: \| \| Ant \| 6059.5 $\pm$ 427.4\| 5216.2 $\pm$ 720.1\| \| HalfCheetah \| 13100.6 $\pm$ 639.5\| 11481.2 $\pm$ 294.8\| \| Hopper \| 3587.0 $\pm$ 225.0\| 2968.1 $\pm$ 1145.1\| \| Humanoid \| 5526.9 $\pm$ 378.3\| 5876.5 $\pm$ 292.2\| - A head-to-head comparison between your memory buffer and the existing ones should be given [3]. Our first-in-first-out (FIFO) memory buffer is identical to the ones utilized in DQN and TD3. We have not incorporated any advanced memory buffer. - From Algorithm 1, it is hard to see the difference between the proposed method and the existing ones. Thank you for pointing this out. We will provide a more detailed description in the revised manuscript. - code In compliance with the rebuttal guidelines, which should not contain any links, we sent an anonymized link to the AC in a separate comment. We will open-source the code after the review process. --- Rebuttal Comment 1.1: Comment: I thank the authors for their detailed responses. I've increased my score. --- Reply to Comment 1.1.1: Comment: Thank you for taking the time to review our responses and for increasing your score. We greatly appreciate your feedback and support.	Summary: The paper introduces the Empirical MDP Iteration (EMIT) framework, which enhances online reinforcement learning by regularizing algorithms with a sequence of empirical MDPs derived from replay memory data. By focusing on in-sample bootstrapping, EMIT ensures stable and unique convergence of Q-functions, leading to monotonic policy improvement (shown for deterministic MDPs). EMIT can be integrated with existing RL algorithms, effectively acting as a regularizer. Experiments with DQN and TD3 on Atari and MuJoCo benchmarks demonstrate that EMIT significantly reduces estimation errors and improves performance. Strengths: The paper is generally easy to follow and clearly written. The proposed approaches appear to be simple and relatively original. Good analysis and numerical results. Weaknesses: Reproducibility: The code was not provided in the supplemental material. Representation: The text in Figure 2(a) is very small and unreadable. Technical Quality: 3 Clarity: 3 Questions for Authors: Can you plot $\Delta (Q, \hat Q) $ to see how far is Q from the empirical $\hat Q$? In the discussion on page 4, in particular, when you state that "Q neither converges to Q* nor $\hat Q^*$, does this depend on the initialization of Q? How did you initialize Q? Confidence: 4 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: Added computational complexity from learning two Q-functions Flag For Ethics Review: ['No ethics review needed.'] Rating: 7 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: - Can you plot to see how far is Q from the empirical Q? We plot the difference in Fig.4 in the rebuttal pdf. We find $\Delta(Q,\widehat Q)$ is similar to $\Delta(Q,\widehat Q^)$ in the later stage of training since $\widehat Q$ will converge to $\widehat Q^$. - when you state that "Q neither converges to $Q^$ nor $\hat Q^$, does this depend on the initialization of Q? How did you initialize Q? There is no dependency. In the extreme case where Q is initialized precisely to $Q^$, then Q will converge to $Q^$. Otherwise, Q has no guaranteed convergence to $Q^*$ whatever its initialization is. In our experiments, we initialize the Q as a randomly initialized neural network. - code In compliance with the rebuttal guidelines, which should not contain any links, we sent an anonymized link to the AC in a separate comment. We will open-source the code after the review process. - text in Figure 2(a) is very small and unreadable. Thank you for pointing this out. We will increase the font size in the revised manuscript. --- Rebuttal Comment 1.1: Comment: Thank you for your response.	Summary: This paper introduces a novel framework called Empirical MDP Iteration (EMIT) to improve the stability and performance of reinforcement learning algorithms. Traditional reinforcement learning algorithms optimize a Markov Decision Process (MDP) using the Bellman equation, which can lead to unstable optimization when function approximation is used. And the EMIT framework addresses this by constructing a sequence of empirical MDPs from the replay memory and using an in-sample Bellman update to learn an estimated Q-function, denoted as $\widehat{Q}$. As claimed in the paper, this method restricts updates to in-sample data, ensuring convergence to a unique optimal $\widehat{Q}$ function and inducing monotonic policy improvement. The paper demonstrates that EMIT can be integrated with existing reinforcement learning algorithms like DQN and TD3, acting as a regularizer to enhance their performance. The experimental results on Atari and MuJoCo benchmarks show that EMIT significantly reduces estimation errors and improves the performance of these algorithms. The authors provide theoretical analysis and extensive experiments to support their claims, highlighting the advantages of in-sample bootstrapping over traditional Bellman updates in reducing estimation errors and improving policy learning stability. Strengths: 1. Comprehensive Experimental Evaluation: The paper presents extensive experimental results on both discrete action space environments (Atari) and continuous control tasks (MuJoCo). The diverse set of benchmarks provides a solid empirical foundation to validate the effectiveness of EMIT across different types of reinforcement learning tasks. 2. Improved Stability and Performance: The empirical results show that EMIT significantly enhances the stability and performance of existing reinforcement learning algorithms like DQN and TD3. By integrating EMIT as a regularizer, the paper demonstrates a clear reduction in estimation errors and notable policy improvements across various benchmarks. Weaknesses: Limited Novelty in Core Ideas: While the EMIT framework introduces in-sample bootstrapping for empirical MDPs, similar ideas have been explored in offline reinforcement learning and distributional reinforcement learning. The paper would benefit from a deeper differentiation from existing works, such as Implicit Q-Learning and In-Sample Actor-Critic. The authors should elaborate on how EMIT offers significant advancements over these methods beyond just integrating with online reinforcement learning algorithms. I recommend the authors clearly delineate the unique contributions of EMIT compared to existing methods like Implicit Q-Learning and In-Sample Actor-Critic. Providing a comprehensive discussion on the theoretical and practical advancements offered by EMIT would strengthen the paper’s claim of novelty. Scalability Concerns: The proposed framework requires maintaining and updating two Q-functions (Q and $\widehat{Q}$) simultaneously, potentially doubling the computational and memory requirements. This scalability issue is not sufficiently addressed in the paper. The authors should discuss how EMIT can be efficiently scaled to more complex environments or larger state-action spaces without significantly increasing the computational burden. Hyperparameter Sensitivity: The paper lacks an in-depth analysis of the sensitivity of EMIT to its hyperparameters, particularly the regularization parameter $\alpha$ and the exploration bonus $\delta(s, a)$. Understanding the robustness of EMIT to different hyperparameter settings is crucial for its practical applicability. The authors should provide more insights or guidelines on how to select these hyperparameters in various scenarios. Technical Quality: 2 Clarity: 3 Questions for Authors: 1. How does EMIT fundamentally differentiate itself from existing methods like Implicit Q-Learning and In-Sample Actor-Critic? 2. What are the limitations of the current exploration mechanism used in EMIT, and how does it compare to other advanced exploration strategies? 3. Can EMIT be effectively applied to real-world reinforcement learning problems, and what are the potential challenges in such applications? I am willing to discuss and update my score until all the concerns are addressed. Confidence: 3 Soundness: 2 Presentation: 3 Contribution: 2 Limitations: The authors have discussed the limitations of their work, especially concerning the scalability and computational cost of maintaining and updating two Q-functions. However, the paper would benefit from a more detailed analysis of these limitations and their broader implications. Additionally, the potential societal impact of this research has not been thoroughly addressed. Flag For Ethics Review: ['No ethics review needed.'] Rating: 5 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: - Limited Novelty in Core Ideas. How does EMIT fundamentally differentiate itself from existing methods like Implicit Q-Learning and In-Sample Actor-Critic? The main contributions shown in our work is that in online RL, iteratively solving a sequence of empirical MDPs is better than just solving the original MDP, especially when the data is incomplete. We develop a novel framework EMIT to enhance existing online RL algorithms by iteratively solving current empirical MDPs. We showcase the effectiveness of our framework by using it to combine established methods for in-sample learning (i.e., Implicit Q-Learning, In-Sample Actor-Critic) and well-known online RL algorithms (i.e., DQN, TD3). - limitations of the current exploration mechanism used in EMIT, how does it compare to other advanced exploration strategies? The objective of exploration in EMIT is to grow the empirical MDP to better approximate the original MDP. Our current exploration mechanism employs the `principle of optimism in the face of uncertainty`, which has proven a useful heuristic in practice. However, its efficiency may wane when the environment is too hard to explore, or when estimated uncertainty is unreliable (e.g., in environments with a 'Noisy TV' problem [1]). In such instances, exploration methods like Go-Explore [2] or intrinsic motivation exploration [1,3,4] might be more suitable. We will discuss this in revised manuscript. [1] Burda, Yuri, Harrison Edwards, Amos Storkey, and Oleg Klimov. "Exploration by random network distillation." arXiv preprint arXiv:1810.12894 (2018). [2] Ecoffet, Adrien, Joost Huizinga, Joel Lehman, Kenneth O. Stanley, and Jeff Clune. "First return, then explore." Nature 590, no. 7847 (2021): 580-586. [3] Zhang, Tianjun, Huazhe Xu, Xiaolong Wang, Yi Wu, Kurt Keutzer, Joseph E. Gonzalez, and Yuandong Tian. "Bebold: Exploration beyond the boundary of explored regions." arXiv preprint arXiv:2012.08621 (2020). [4] Jarrett, Daniel, Corentin Tallec, Florent Altché, Thomas Mesnard, Rémi Munos, and Michal Valko. "Curiosity in Hindsight: Intrinsic Exploration in Stochastic Environments." arXiv preprint arXiv:2211.10515 (2022). - Can EMIT be effectively applied to real-world reinforcement learning problems, and what are the potential challenges in such applications? EMIT is designed to improve the performance of existing online RL algorithms. Thus, it should lead to improved performance on any `real-world` problems where existing online RL algorithms have shown to be effective. It is an interesting question on what kind of problems EMIT's advantage would disappear or unable to observe. We will try elaborate more on the limitations of EMIT in our revision. While EMIT could still encounter similar challenges as other RL algorithms, such as facing issues related to sample efficiency and safety, it is encouraging to note that EMIT not only boosts the sample efficiency of tested RL algorithms but also yields a more conservative estimate of the Q-function. This conservative estimation could potentially enhance the safety of the derived policy, positioning EMIT as a more preferable solution for real-world applications. - Scalability Concerns: The proposed framework requires maintaining and updating two Q-functions, potentially doubling the computational and memory requirements. We recognize the added computation cost in the "Limitations" section of our manuscript. However, the good news is that EMIT does not asymptotically make existing online RL algorithms more expensive over time. We thus believe EMIT enjoys similar `scalability` property as the online RL algorithms it intends to enhance. Future work to further improve EMIT's efficiency include designing an exploration strategy that is directly based on in-sample Bellman update, or improving runtime using two processes to update Q and $\hat Q$ in parallel. Despite this limitation, as demonstrated in Figure 3 of our manuscript and summarized in the table below, EMIT still outperforms DQN and TD3 when using the same amount of computation, underscoring its efficacy in the current format. \| \| EMIT-DQN (5e6) \| DQN (1e7) \| \| :-----\| ----: \| :----: \| \| Asteroids \| 1338.9 $\pm$ 27.8 \| 948.6 $\pm$ 61.9\| \| Atlantis \| 1599569.9 $\pm$ 231875.6\| 910471.2 $\pm$ 35727.5\| \| Breakout \| 266.2 $\pm$ 40.4\| 103.3 $\pm$ 8.0\| \| Gravitar \| 949.7 $\pm$ 14.9\| 523.5 $\pm$ 68.3\| \| \| EMIT-TD3 (1e6) \| TD3 (2e6) \| \| :-----\| ----: \| :----: \| \| Ant \| 6059.5 $\pm$ 427.4\| 5216.2 $\pm$ 720.1\| \| HalfCheetah \| 13100.6 $\pm$ 639.5\| 11481.2 $\pm$ 294.8\| \| Hopper \| 3587.0 $\pm$ 225.0\| 2968.1 $\pm$ 1145.1\| \| Humanoid \| 5526.9 $\pm$ 378.3\| 5876.5 $\pm$ 292.2\| - Hyperparameter Sensitivity We searched the best regularization parameter $\alpha$ in the range of [0.05, 0.1, 0.5] shown in Fig.3 in the rebuttal pdf. We find that a small value such as 0.05 or 0.1 suffices for the effective operation of EMIT. Regarding the exploration bonus, it does not introduce any hyperparameter. As demonstrated in the manuscript Fig.6(b), incorporating the exploration bonus notably enhances performance. --- Rebuttal Comment 1.1: Comment: Thanks for your rebuttal and I will maintain my score! --- Reply to Comment 1.1.1: Title: Follow-Up on Rebuttal and Feedback Comment: Dear Reviewer, As you mentioned 'I am willing to discuss and update my score until all the concerns are addressed', we kindly ask if our responses have addressed your concerns. We appreciate your feedback and are happy to provide further clarification if needed. Many thanks, The Authors	Summary: The authors transfer insights from IQL to online RL, demonstrating the performance of online RL can be enhanced by leveraging a Q-function that performs a max only over actions in the replay buffer when updating the Q-network. They use this network to: * encourage exploration by driving the agent towards state-action pairs where the value predictions between a traditional Q-network and the IQL style Q-network differ * they regularise the value predictions of a Q-network with a term equal to the difference between a traditional Q-network and an IQL style Q-network's value prediction Overall, I like the paper and think it should be published. Strengths: * The authors demonstrate improved empirical performance on an impressive number of environments. * They also perform ablation studies to analyse why their approach improves on previous methods (by considering how for example it reduces policy churn or how it improves the accuracy of Q-network value predictions). * They provide theoretical results surrounding the convergence of EMIT and how it relates to Q-learning Weaknesses: * prior to reading this paper, I did not know how IQL worked. For example, it was initially confusing to me how an update like equation 2 could be done in continuous state spaces where you are unlikely to encounter the same state twice. I think a little bit more explanation of how this approach works (for example around equation 3) would help a reader like me that is not very familiar with the related work. * it seems there are some insights that initially I thought we generated originally from your work but now I understand (e.g. Eq(2)) builds of prior work (IQL). Maybe that would be obvious to some but I think it would be good to make explicit your contributions compared to IQL etc * Pretty hard for me to parse figure 4. I think something more suitable would show some summary statistics aggregated over all environments (with the bar charts in the appendix). Also why no error bars in figure 4? * Would be interesting to see how it compares to basic regularization techniques (e.g. weight decay) and slightly more sophisticated methods for exploration (e.g. those that use intrinsic rewards). For example on venture it appears that the exploration term is contributing a lot to the improvement (possibly because it is a difficult environment in terms of exploration). I am not suggesting that EMIT must beat these methods, I just think it would add some context for the reader. Technical Quality: 3 Clarity: 3 Questions for Authors: Why integrate into TD3 instead of SAC when as far as I know SAC seems to outperform TD3? Confidence: 3 Soundness: 3 Presentation: 3 Contribution: 4 Limitations: The authors do discuss the limitations briefly in the conclusion, but I think having a larger limitation section where they are discussed more thoroughly would improve the paper. Flag For Ethics Review: ['No ethics review needed.'] Rating: 7 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: - Why integrate into TD3 instead of SAC when as far as I know SAC seems to outperform TD3? We integrate EMIT with DQN and TD3 mainly because they are directly built upon optimizing value-based Bellman equations, which aligns with our theoretical analysis and is exactly we aim to improve with EMIT. SAC, on the other hand, is originated from maximum entropy reinforcement learning with an entropy term. It is possible to implement EMIT with SAC. To show that, we've experimented EMIT with SAC, discovering that EMIT similarly enhances SAC's performance (Fig.1(a) in the rebuttal pdf). - a little bit more explanation of how IQL works We appreciate the suggestion and will incorporate a more detailed explanation of IQL in the revised manuscript. - it would be good to make explicit your contributions compared to IQL etc EMIT is a novel framework that iteratively solves a sequence of empirical MDPs, enhancing existing online RL algorithms. Other existing offline methods that solve empirical MDPs alternatively than IQL can also be adopted when instantiating EMIT. We will make this point more explicit in the revised manuscript. - summary statistics aggregated over all environments, and error bars in figure 4 We summarize the aggregated results over all environments in Fig.2 in the rebuttal pdf. We will add error bars in fig.4 in the revised manuscript. - how it compares to basic regularization techniques (e.g. weight decay) and slightly more sophisticated methods for exploration (e.g. those that use intrinsic rewards). We appreciate the suggestion. In our current study, we concentrate on the theoretical analysis and empirical validation of the EMIT framework, employing a straightforward instantiation to showcase its efficacy. Indeed, We aim to explore more sophisticated methods for exploration and regularization in our future research. - a larger limitation section We value your suggestion and will elaborate more on the limitations of this work.	Rebuttal 1: Rebuttal: We would like to thank all the reviewers for their comments, suggestions for improvement, and interest in the paper. Besides the detailed responses to each reviewer's comments, we submit a rebuttal pdf to supplement our responses. This pdf includes four figures to address the reviewers' concerns. - Figure 1 shows the performance of EMIT with SAC and TD7, demonstrating that EMIT can also enhance the performance of other state-of-the-art RL algorithms. - Figure 2 summarizes the aggregated results over all environments, providing a clearer comparison between EMIT and the baselines. - Figure 3 shows the parameter sensitivity analysis of the regularization parameter $\alpha$, showing that a small value such as 0.05 or 0.1 suffices for the effective operation of EMIT. - Figure 4 plots the difference between the Q-function and the empirical Q-function, showing that $\Delta(Q,\widehat Q)$ is similar to $\Delta(Q,\widehat Q^)$ in the later stage of training since $\widehat Q$ will converge to $\widehat Q^$. Pdf: /pdf/3d0daef5111534f556859d74fd33a6dabe3e6f57.pdf	NeurIPS_2024_submissions_huggingface	2,024	null	null	null	null	null	null	null	null
CODA: A Correlation-Oriented Disentanglement and Augmentation Modeling Scheme for Better Resisting Subpopulation Shifts	Accept (poster)	Summary: To accommodate the presence of spurious correlations and group imbalance, the paper introduces a method CODA to reliably recognize the same object across a range of spurious attributes. The paper disentangled invariant and causal features concerning the object identification from spurious features. The performance of the proposed scheme is tested on extensive experiments. Strengths: The presentation of the paper is clear and easy to follow. Problem is well-motivated and each component of the proposed method is clearly justified. The paper introduces an innovated scheme to disentangle the spurious features and causal features related to the predefined spurious attributes. The authors conducted a variety of experiments showing the effectiveness and robustness of the method. Weaknesses: 1, Compared to disentanglement of the latent variables, the binary classification of the latent space has practically no transfer learning ability and limited for downstream tasks. Please discuss this aspect of the limitation further. 2, The average accuracies in the experimental results sometimes are not as good as the baseline models. In comparison, the worst group accuracy is higher for the proposed method. Please explain this phenomena in details. Technical Quality: 3 Clarity: 3 Questions for Authors: Please see Weakness Confidence: 4 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: The authors adequately addressed the limitations. Flag For Ethics Review: ['No ethics review needed.'] Rating: 7 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We appreciate your insightful comments and positive assessment. Response to W1: Thank you for this interesting comment, which inspires us to think of CODA under transfer learning. To achieve disentanglement, the variance encoder is regularized by an auxiliary label prediction task, which is problem-specific. Therefore, compared with classical machine learning models, in transfer learning, CODA needs to consider the variance encoder-auxiliary label predictor together with the remaining structure for the downstream classifier. Hypothetically, given a source domain and multiple target domains, the CODA can be first trained based on source domain data with source domain spurious attributes identified to produce a CODA backbone model. To transfer such a backbone model in to target domains, the specific target domain spurious attributes need to be first identified to fine-tune the auxiliary label predictor. Next, the target domain data can be supplied to customize the CODA backbone towards the target domain task. This is a quick thought of helping the CODA gain transfer learning ability. Meanwhile, we highly appreciate this great comment because both extending CODA to gain transfer learning ability and innovating transfer learning to accommodate CODA are interesting and valuable problems which will be considered in our future research. Meanwhile, we will update our limitation discussions on this item in the revised paper. Response to W2: Thank you for this insightful comment. We appreciate the opportunity of clarifying the observed phenomenon regarding the average accuracies and the worst group accuracies in our experimental results. We acknowledge that, on CelebA, the CODA method in terms of accuracy might not be always higher that of the baseline models (see Table 3). However, on ColoredMNIST and its variants, our method achieves higher average accuracy compared to the baseline models (see Tables 3 and 5 in Appendix). More importantly, our method consistently achieves higher worst group accuracy across all datasets than baseline models (see Table 3 and Table 4). This trade-off between average accuracy and worst group accuracy is a well-documented phenomenon in the literature on subpopulation shifts, which commonly considered the worst-group accuracy as the golden standard while average accuracy as a reference in evaluation. CODA performances can be interpreted from following two aspects: 1. Modeling Focus and Learning Mechanism: The primary focus of our work is to develop a robust classification model that performs well across all subpopulations, particularly those that are underrepresented. Traditional models often optimize for overall average accuracy, which can lead to suboptimal performance on minority groups. In contrast, baseline methods such as RWG and GDRO in the paper apply techniques that upweight the sampling probabilities of minority groups (while inevitably compromising the sampling probabilities of majority groups). This drives the model to focus more on minority groups, resulting in higher worst-group accuracy, but potentially lower average accuracy. Our proposed learning mechanism may further emphasize minority groups to ensure no subpopulation is disproportionately disadvantaged. 2. Dataset Characteristics and Group Proportion Shifts: According to the dataset statistics in Table 1, the degree of group proportion shifts (DGPS) between training and testing sets is low for CelebA and high for ColoredMNIST and its variants. In ColoredMNIST and its variants, the majority/minority groups in the training set become the minority/majority groups in the test set. While in CelebA, the majority/minority groups in the training set remain the majority/minority groups in the testing set. DGPS and the insight from point 1 explain the lower average accuracy of our method on CelebA. On the other hand, high DGPS has less impact on worst group accuracy since group-wise accuracy is not sensitive to changes in sample size. In summary, the higher worst group accuracy of CODA, despite the occasional lower average accuracy, aligns with our goal of developing a robust and fair classification model. This trade-off is justified and beneficial for ensuring equitable performance across all subpopulations, especially in critical applications like healthcare, where fairness and robustness are paramount, and where the cost of misclassification is high for certain groups. Thank you once again for this valuable feedback. --- Rebuttal Comment 1.1: Comment: I thank the authors for their reply and incorporating the feedback given. I look forward to seeing an updated version of the manuscript. --- Reply to Comment 1.1.1: Title: Thank you! Comment: Thank you very much again for your encouragement and enlightening comments! We will revise the manuscript to incorporate the discussions.	Summary: This paper proposes a novel framework for addressing subpopulation shifts. The framework contains two key components. The first one is Correlation-Oriented Disentanglement, which separates the spurious information from class information. Then, with disentangled spurious and class embeddings, the framework can augment the original dataset with synthesized images and train a robust classifier for improved worst-group performance. Strengths: (1) The idea of translating the research problem from overcoming the effects of spurious attributes to utilizing them, so that the model is encouraged to perform equally well on them, is clever, inspiring, and promising. (2) The design of the Correlation-Oriented Disentanglement is a good contribution and well-suited for the purpose of the proposed framework. (3) I enjoyed the writing of this paper very much; it is clear and easy to follow. Weaknesses: The major weakness of this paper is its evaluation: (1) The qualitative results do not seem to support the disentanglement very well. In Figure 3, while the results for MNIST are pretty good, the gender consistency on CelebA is not so obvious, which leads to my question (1). (2) The paper could be strengthened by evaluating datasets that have more complex spurious information, like the popular Waterbird dataset, which is also used as an example in the introduction. This would better help readers evaluate how well the proposed method addresses complex real-world problems. Technical Quality: 3 Clarity: 3 Questions for Authors: (1) Did the authors investigate the relationship between the disentanglement performance and the worst-group classification accuracy? For example, does higher disentanglement performance always lead to more robust worst-group accuracy? If there is a trade-off concerning the quality of the synthesized images due to high KL loss, what would be the best strategy for selecting the hyper-parameter γ (the coefficient for the KL term) for different datasets, especially more complex ones like CelebA? (2) Did the authors study the diversity and fairness of the learned spurious embeddings? Could it also be biased? For example, could the VAE reconstruct images of the major group better? How could this affect the final classification performance? Confidence: 3 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: N/A Flag For Ethics Review: ['No ethics review needed.'] Rating: 7 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thank you for your thoughtful feedback and encouraging evaluation. Response to W1: Thank you for your observation regarding the concern about the gender consistency on CelebA presented in Figure 3. The reconstruction results suggest that causal and label-irrelevant features such as background, posture, hairstyle, and hair color are determined by the leftmost samples, while the variance embeddings primarily capture the facial characteristics, which are crucial for determining biological gender in this face dataset. This implies that the variance embeddings carry significant information related to the gender. One factor contributing to the less consistent gender representation could be the heavily regularized mutual information term I(x;z), controlled by the hyperparameter γ. The variance embeddings are expected to carry the most essential information about gender. However, if I(x;z) is overly minimized during training time due to a high γ value, it may lead to reduced expressive power of the variance embeddings, thereby affecting the consistency and quality of the reconstructions. Furthermore, CelebA is a significantly more complex dataset compared to ColoredMNIST. The variety of attributes such as age, hairstyle, facial expressions, and lighting conditions adds to the complexity, making disentanglement more challenging. Careful tuning of the hyperparameter γ is essential to prevent overly minimizing I(x;z) and ensure the variance embeddings retain sufficient expressive power for accurate reconstruction. Conducting experiments to find an optimal balance for γ that maintains the expressive power of the variance embeddings while achieving effective disentanglement could lead to better results. Response to W2: We highly appreciate your constructive suggestion of evaluating CODA via additional datasets with more complex spurious information, such as the popular Waterbirds dataset. Yet, given the limited rebuttal time, we decided to first conduct computational experiments to clarify concerns on CODA performances based on datasets considered and we faced insufficient time of conducting a thorough evaluation on additional datasets. However, we acknowledge the importance of your suggestion and we plan to include this evaluation in our future research. Response to Q1: Thank you for the input. Higher disentanglement performance, primarily assessed through visual inspection of reconstructed images, generally leads to improved worst-group classification accuracy. The key is to drive synthesized samples to retain as much information as possible from the original image with spurious attributes excluded. While disentanglement performance is sensitive to the hyperparameter y (the coefficient for the KL term), reconstruction quality is less so. The y-weighted KL loss limits the expressive power of the variance embeddings, but the other KL loss term, being less regularized, helps maintain reconstruction quality, which is crucial for robust classification (see response to Q2). To better control reconstruction quality, assigning extra weight to the reconstruction loss term can be considered. Selecting y can start with a higher value if spurious attributes carry minimal information (e.g., colors in the ColoredMNIST dataset) and a lower value if they include more complex patterns. Next y can be adjusted with decreasing it if the variance embeddings capture insufficient variation of spurious attributes and increasing it if they capture more than just the spurious attributes. Response to Q2: Thank you for highlighting the potential bias in the learned spurious embeddings. To address this concern, we conducted an evaluation of the pixel-wise reconstruction loss for each group on both validation and test sets of both datasets. Results are reported in the following table: \|\|Group 1\|Group 2\|Group 3\|Group 4\| \|-\|-\|-\|-\|-\| \|CelebA\|\|\|\|\| \|Validation Set\|0.0128\|0.0119\|0.0134\|0.0128\| \|Test Set\|0.0129\|0.0121\|0.0128\|0.0129\| \|ColoredMNIST\|\|\|\|\| \|Validation Set\|0.00057\|0.00054\|0.00057\|0.00054\| \|Test Set\|0.00056\|0.00058\|0.00056\|0.00056\| The pixel-wise reconstruction loss is relatively consistent across all groups on both the validation and test sets. While there are minor variations, they do not suggest a significant bias toward any particular group. However, poor reconstruction quality can indeed decrease the final classification performance. We applied a simple 3-CNN-layer architecture for encoders (less powerful than the Resnet18 encoders used in the paper) on the ColoredMNIST dataset, the average pixel-wise reconstruction losses were 0.06451 for the validation set and 0.05424 for the test set, both significantly higher than those of Resnet18 encoders. Notably, the reconstruction images indicate lower reconstruction quality but good disentanglement performance, comparable to that of Resnet18 Encoders. Classification results are presented in the below table: \|\|Avg. Acc. (std)\|Worst. Acc. (std)\| \|-\|-\|-\| \|3-CNN-layer Encoders\|\|\| \|CODA+ERM\|68.89% (0.59%)\|67.50% (1.08%)\| \|CODA+RWG\|70.89 (0.31%)\|69.83% (0.38%)\| \|CODA+GDRO\|71.36% (0.48%)\|70.51% (0.52%)\| \|Resnet18 Encoders\|\|\| \|CODA+ERM\|72.20% (0.57%)\|71.74 (0.24%)\| \|CODA+RWG\|73.20% (0.12%)\|72.11 (0.51%)\| \|CODA+GDRO \|73.02% (0.23%)\|71.98 (0.57%)\| We observed that the classification performance of CODA using ResNet18 encoders is higher than that of the 3-CNN-layer encoders. This implies that poor reconstruction quality does indeed hinder the final classification performance. Building on this insight, we argue that bias towards embeddings of certain groups (i.e., poor reconstruction quality for specific groups) will negatively impact the final classification performance. --- Rebuttal Comment 1.1: Comment: I appreciate the author's detailed response in addressing my concerns. My concerns have been addressed, and I have increased my score. The authors are encouraged to add the additional experiments and discussion to the appendix. --- Reply to Comment 1.1.1: Title: Thank you! Comment: Thank you very much for your valuable time in reading our rebuttal and supporting this work! We will revise the Appendix of this work to incorporate additional experiments and discussions.	Summary: This paper proposes CODA (Correlation-Oriented Disentanglement and Augmentation), a novel framework for addressing subpopulation shifts caused by spurious correlations and group imbalance (SC-GI). The key contributions are: 1. A correlation-oriented disentanglement (COD) method that learns to separate variant (spurious) and invariant (causal) features using a bi-branch encoder architecture with a decoy classifier. 2. A strategic sample augmentation technique that generates synthetic samples by mixing disentangled features, combined with a reweighted consistency loss. 3. Extensive experiments on ColoredMNIST and CelebA datasets demonstrating CODA's effectiveness in improving worst-group accuracy and reducing maximum group accuracy gap compared to existing robust classification methods. Strengths: 1. Novel approach that utilizes spurious attributes constructively rather than treating them as impediments 2. Theoretically motivated disentanglement objective that avoids complex density estimation 3. Comprehensive experiments across multiple datasets and varying degrees of subpopulation shifts 4. Compatibility with existing robust classification methods like ERM, RWG, and GDRO Weaknesses: 1. Reliance on pre-identified spurious attributes, which may require costly labeling 2. Increased computational cost during training due to data augmentation processes 3. Limited theoretical analysis of the proposed method Technical Quality: 3 Clarity: 3 Questions for Authors: 1. While the paper provides some intuition behind the proposed method, could you elaborate on the theoretical understanding of the disentanglement process? How does CODA ensure that the learned representations are truly disentangled? 2. The experiments focus on datasets with binary spurious attributes. How well does CODA scale to scenarios with multiple or continuous spurious attributes? Have you conducted any experiments in such settings? 3. The paper shows a sensitivity analysis for the reweighted consistency loss parameter. Are there other critical hyperparameters in CODA? How sensitive is the method to their tuning? 4. The paper mentions increased computational cost during training. Can you quantify this overhead compared to standard training procedures? Confidence: 3 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: NA Flag For Ethics Review: ['No ethics review needed.'] Rating: 7 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thank you for your insightful questions and positive evaluation. Response to W1: Thank you for this insightful comment. Labeling spurious attributes (SA) indeed require extra efforts. Yet, we would like to justify its necessity: 1. Robust methods without using SA in training would suffer severe performance drops if SA were not provided in the validation set for model selection [1]. 2. There was a significant performance gap between methods with and without considering SA [2]. [1] Liu, Evan Z., et al. "Just train twice: Improving group robustness without training group information." International Conference on Machine Learning. PMLR, 2021. [2] Yang, Yuzhe, et al. "Change is hard: A closer look at subpopulation shift." arXiv preprint arXiv:2302.12254 (2023). Response to W2: Thank you for your concern on extra computational cost due to Data Augmentation Process (DAP). Although DAP induces an increase of computational costs at training, the long-term benefits of a robust and fair model, such as reducing the frequency of model retraining and corrections, outweigh additional training costs caused by DAP and ultimately save resources. Furthermore, additional cost of CODA scales linearly and is controllable via hyperparameter L. A detailed analysis is offered in our response to your Q4. Response to W3: Thank you for pointing out this limitation. As a pioneering study of the CODA idea, this work aims at showing the empirical effectiveness of CODA. Although preliminary theoretical exploration has been offered in this work, we would like to conduct a deeper theoretical analysis in future studies with directions detailed in our response to Q1 to reinforce the credibility and applicability of CODA. Response to Q1: Thank you for your question. Eq. (9) justifies the disentanglement process (DP). The prediction loss and the mutual information I(x;z) weighted by large γ aim to learn z for predicting SA while minimizing the expressiveness of z on x. The reconstruction loss and the mutual term I(x;t) aim to emphasize more on the contribution of t in reconstructing x, pushing t to better describe x. The effectiveness of CODA in DP has been numerically justified in Table 1. Yet, ensuring true disentanglement enquires more theoretical establishment, which is our next research pursuit. Our sketchy plan includes 1) developing a disentanglement measure (DM) tailored for CODA; 2) studying disentanglement performance bounds based on developed DMs; and 3) pursuing disentanglement performance bounds under a more generalized problem setting. Response to Q2: Thank you for this interesting question. For the continuous case, defining groups g= (a, y) can be challenging. One possible approach is to discretize the continuous values by bins, transforming it into a classification task. We have evaluated the case of multiple SA values on MultipleColoredMNIST. The task is to predict digits 0-9, and we define ten RGB colors 0-9, resulting in 100 groups. In the training set, each sample is painted with the color a=y with prob 85%, and randomly assigned another for the remaining 15%. Labels are flipped with prob 25%. The training set is highly group imbalanced, with colors spuriously correlated with labels. The majority group constitutes 9.44% of the population, while the minority group only 0.13%. The worst group accuracy of the hypothetical optimal digit classifier f can be as low as 62.07% (only 62.07% of the samples whose digits match the labels due to randomness in label flipping and color assignment). Thus, for validation and test sets, we used a different setup, no label flipping and uniform color assignment. The worst group accuracy of f is 100% for both sets. Experimental results next indicate that CODA scales well to scenarios with multiple SA values, demonstrating effectiveness in solving SA. \|Method\|Avg. Acc. (std)\|Worst Acc. (std)\| \|-\|-\|-\| \|ERM\|16.60% (1.23%)\|0% (0%)\| \|RWG\|95.54% (0.24%)\|85.46% (0.44%)\| \|GDRO\|94.99% (0.33%)\|82.16% (4.60%)\| \|CODA+ERM\|97.06% (0.11%)\|91.85% (0.68%)\| \|CODA+RWG\|96.94% (0.08%)\|91.10% (1.16%)\| \|CODA+GDRO\|96.85% (0.05%)\|90.42% (1.39%)\| Response to Q3: Thanks for the insightful question. Another critical hyperparameter is L, which controls the number of synthesized samples for each data point. A larger L introduces more variability to synthesized samples, leading the model to "forget" the SA in decision-making. A sensitivity analysis of L on MultipleColoredMNIST is offered next: \|Methods\|Avg. Acc. (std)\|Worst Acc. (std)\| \|-\|-\|-\| \|ERM\|16.60% (1.23%)\|0% (0%)\| \|RWG\|95.54% (0.24%)\|85.46% (0.44%)\| \|GDRO\|94.99% (0.33%)\|82.16% (4.60%)\| \|CODA+ERM\|\|\| \|L=1\|96.90% (0.29%)\|88.78% (0.87%)\| \|L=2\|97.12% (0.07%)\|90.36% (0.68%)\| \|L=4\|97.06% (0.11%)\|91.85% (0.68%)\| \|CODA+RWG\|\|\| \|L=1\|96.92% (0.04%)\|90.87% (0.93%)\| \|L=2\|96.66% (0.28%)\|91.35% (0.70%)\| \|L=4\|96.94% (0.08%)\|91.10% (1.16%)\| \|CODA+GDRO\|\|\| \|L=1\|96.68% (0.03%)\|89.88% (0.50%)\| \|L=2\|96.70% (0.16%)\|90.10% (1.10%)\| \|L=4\|96.85% (0.05%)\|90.42% (1.39%)\| It is observable that increasing L generally improves the worst-group accuracy. However, a larger L also implies higher computational costs. Furthermore, excess L may not introduce more variability (depends on the complexity of the dataset), e.g., L greater than 10 in this case can result in redundant synthesized samples. Therefore, the trade-off between classification performance and computational efficiency should be considered carefully. Response to Q4: The increased computational cost in training the robust classifier is caused by augmented samples. Assume that C is the computational cost per sample. The standard training cost is C * N where N is the number of original samples. The training cost of CODA is then C * (L+1) * N. Here the cost of the RWC loss in Eq. (10) is ignored, as it is much smaller than that of classifying and computing the cross-entropy loss. Thus, the cost increases linearly with L, resulting in an overhead of L times standard training cost.	null	null	null	NeurIPS_2024_submissions_huggingface	2,024	null	null	null	null	null	null	null	null
ShowMaker: Creating High-Fidelity 2D Human Video via Fine-Grained Diffusion Modeling	Accept (poster)	Summary: This paper proposed a pipeline for 2D human motion retargeting, based on 2D key points and face encoding via fine-grained diffusion modeling. Strengths: The paper is well-written, and the demo presents satisfying hand and face modeling results. Weaknesses: 1. Missing recent baselines in the experimental comparison. E.g. Diffusion-based method: MagicPose from ICML 2024, GAN-based methods: TPS from CVPR2022 etc. 2. The novelty of the proposed method is limited. It seems to me that the only novel component of this paper compared to previous works is the face encoding and key point-based hand modeling. The ReferenceNet for identity control and Pose Encoder for pose control are all from previous works (ReferenceNet from Animate Anyone, MagicAnimate, MagicPose, Champ, and ControlNet from DisCo, MagicAnimate, Animate Anyone, MagicPose). Temporal Attention the same. 3. I would like to see the performance of the proposed method compared to baselines on the TikTok dataset since this is a public dataset, and all previous works reported metrics on it. I'm not implying the authors are hiding something but I think it's necessary to follow previous settings and report the performance on the new dataset (talkshow). 4. Lack of Long Video Generation Results. The presented videos in the supplementary materials are very short compared to MagicAnimate's videos on TikTok dataset. It's necessary to provide longer visualization. 5. Lack of motion retargeting visualization while the condition of human DW-Pose is significantly different than the Human Pose in the reference image. The visualization in both supp and main paper shows only the cases where condition and reference image share similar human poses. This also raises the concern about face encoding and key point-based hand modeling: will such network design preserves human identity and ensure temporal coherence when the poses have a huge gap between reference and condition? Technical Quality: 3 Clarity: 3 Questions for Authors: See weakness. Major questions as following: 1. Examples of Long Video Generation. Ideally, the generated video of the TikTok test set as MagicAnimate shows. (~at least 20 seconds or longer) 2. Motion retargeting visualization while condition and reference have quite different human poses. Please see the previous paper, MagicPose Figure 5,7,17, for reference. 3. Previous work, e.g., MagicPose, has shown that the model can synthesize the back of the human subject. Can this model do the same? Given that the front of the human is the reference image and another pose condition, how can the model learn to distinguish the front/back of the human from such conditions? 4. Can this model provide a satisfying result if the identity reference is a non-frontal image,e.g., the left or right side of the human? 5. Is there any visualization on out-of-domain motion retargeting? E.f. cartoon-style animation? 6. More detailed explanation of the novelty. Confidence: 5 Soundness: 3 Presentation: 3 Contribution: 2 Limitations: The authors include a limitation section. Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: # Response to Reviewer xzB4 Thanks for such a detailed review. Here are my responses. According to the provided reviews, we believe there are a few misunderstandings on the task setting that need to be clarified. __Task \& Setting Issues__ &emsp; (1) Unlike the approaches concentrating on holistic body movement synthesis (e.g., AnimateAnyone and MagicPose), our ShowMaker focuses on __half-body 2D talking avatar__ creation, which aims to upgrade the traditional 2D talking head avatar task to the next level with richer movements (e.g., hand gestures and upper body movements). In this scenario, the training data usually contains frontal images with very few side-view cases, therefore learning to distinguish the front/back of the human is not necessary (Question 3). &emsp; (2) Our approach follows the __person-specific setting__ as Make-Your-Anchor rather than the one-shot animation setting used in AnimateAnyone. We provide our train/test data information in Table 1 in the pdf file. Under the person-specific setting, we focus on self-driving and cross-driving performance on trained IDs instead of out-of-domain data (e.g., cartoon-style images (Question 5)). Considering this person-specific setting and our task scenario, we conduct experiments on the TALKSHOW dataset and our collected dataset instead of the TikTok dataset (Weakness 3). &emsp; (3) The dual-branch design with a ReferenceNet is an appropriate backbone for generative tasks, but we did not involve them in our contributions. Our technical contributions are summarized below (Weakness 2 and Question 6): &emsp; &emsp; (a) We propose an effective hand recovery module for fine-grained synthesis by using a positional encoding strategy and a carefully designed key point-based texture codebook. Our approach shows obvious superiority in the hand synthesis over other dual-branch baselines. &emsp; &emsp; (b) We propose a Face Recapture module to improve facial details and identity-preserving ability, especially under the cross-driving setting with pose and shape gaps. The results in our supplementary video and the pdf file also show the temporally consistent facial texture and identity. In addition to the task and setting issues, there are also some experimental issues. __Experimental Issues__ - Q1. Comparison with TPS from CVPR2022 and MagicPose from ICML 2024. (Response to Weakness 1) A1. Thanks for your suggestion. We compared the TPS method in our paper, please see Table 1 and Figure 3 in the main paper for more details. In terms of MagicPose, since it is a concurrent work that was accepted by ICML in May 2024 (very close to the submission date), we involve another work MagicAnimate (similar to MagicPose) in comparison. &emsp; Here, we finetune MagicPose on our training data and provide the comparison results in Figure 4 in the pdf file. Similar to AnimateAnyone and MagicAnimate, MagicPose suffers from inaccurate generation and texture degradation in both facial and hand regions. Please zoom in and see the details in the red boxes. The results demonstrate the necessity of our proposed Face Recapture module and Fine-Grained Hand Modeling. - Q2. About longer visualization. (Response to Question 1) A2. In our supplementary video, we have provided demo videos of 10-15 seconds, which is comparable to the length of videos in the TikTok dataset. During our exploration, we managed to generate videos longer than 30 seconds with high-fidelity and temporally consistent texture. Since we cannot provide external links, we will add longer-generated videos to our supplementary video. - Q3. Pose gaps between the reference and driving images & Non-frontal images. (Response to Question 2, Question 4) A3. We provide two examples in Figure 5 in the pdf file to illustrate how our ShowMaker performs when dealing with challenging cases with pose gaps or non-frontal images. &emsp; (1) For pose gaps, we first provide an example by using a reference image from a male and driving signals from a female. As shown in Figure 5 (left part), there is a huge gap in both pose and shape. Please also refer to A4 for Reviewer jdWr, we adopt our pose alignment strategy to first align the driving signals with that of the reference image for better synthesis and the results show the effectiveness of our approach under huge pose gaps. &emsp; (2) Non-frontal images. Since most of the videos in our dataset are in the front view, so we took some time to look for a side-view reference image in the test set, and the generated result is shown in Figure 5 (right part). The results demonstrate that our network design can preserve human identity and hand texture even when processing non-frontal reference images. --- Rebuttal Comment 1.1: Comment: Thanks for the author's detailed response to my questions. My concern on visualization and experiments has been well-addressed. Hence, I'm raising my score. For out-of-domain data, I still suggest the authors explore training/finetuning the trained model on real-human data on half-body 2D talking avatar of cartoon-style videos since it would be an outstanding contribution to demonstrate the generalization ability, and it would be interesting to see the results. For MagicPose ICML 2024, kindly remind that it's released around Sep 2023, which is close to Animate-Anyone/MagicAnimate, and it's a prior work rather than a concurrent one. But this is a tiny point since ShowMaker has already demonstrated better visualization quality for gestures. --- Reply to Comment 1.1.1: Comment: We greatly appreciate your response and the valuable insights. In our revised manuscript, we will update the discussion and comparison of MagicPose in the related work and experiments sections, respectively. In terms of out-of-domain data, we agree with your suggestion and we will make efforts to demonstrate its generalization ability on multiple styles of conversational videos.	Summary: This paper proposes a novel conversational human video generation framework named showmaker based on the fine-grained diffusion model. This paper proposes a novel Key Point-based Fine-grained Hand Modeling module and construct a key point-based codebook to handle the challenging hand generation. Meanwhile, this paper extracts facial texture features and global identity features from the aligned face and integrates them into the diffusion process for face enhancement. Sufficient experiments demonstrate the superiority of the proposed framework, and the generated human videos are of high quality. I have read the comments of other reviewers and the response of the authors, I tend to give a accept score. Strengths: 1.The paper is well written and clearly explains the motivations for the design as well as important technical details. 2.The human video generation framework introduced in this paper holds significant practical value, particularly in conversational scenarios like TV shows. 3.The proposed Face Recapture module contributes to generate accurate face regions. 4.The paper has conducted sufficient experiments that demonstrate the superiority of the proposed model, and the generated videos are highly satisfactory. Weaknesses: 1.Although the author compares the proposed method with talking head synthesis in the demo, it would be more beneficial to include this comparison in the main text or as an appendix, as it is crucial for understanding the contribution. 2.The presence of several writing errors in the paper impedes the correct comprehension of its content. 3.The details regarding some network structures and inference processes, such as the pose encoder, are insufficiently described and require further elaboration. Technical Quality: 3 Clarity: 3 Questions for Authors: 1. The input for the Face Recapture Module in Figure 1 incorporates a reference pose, whereas the input in Figure 2 lacks such a reference pose. 2. Shouldn't line 140 on page 3 be replicated F times along the temporal dimension? 3. What is the rationale behind selecting a VAE encoder as the texture extractor in the Face Recapture Module? 4. Could you elaborate on the shapes of Gh and Gf? Additionally, why was a discrete codebook chosen to represent hand information? What advantages does this design offer? Confidence: 5 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: Authors adequately discuss the limitations and potential negative societal impact Flag For Ethics Review: ['No ethics review needed.'] Rating: 7 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: # Response to Reviewer R54L Thanks very much for your valuable suggestions and for pointing out the typos in the manuscript. Here are my responses. - Q1. About the comparison with talking head synthesis. A1. Thanks for your suggestion, we will add the talking head comparison in the demo to the main manuscript to reflect that our task is an extension of the talking head. - Q2. About the structure of the pose encoder. A2. For a detailed description of the Pose Encoder, please refer to Figure 2 in the pdf. - Q3. About some typoes.(Response to Weakness 2 and Questions 1, 2) A3. Sorry for the confusion. The "Reference Pose" in Figure 1 (d) should be removed. It should be repeated F times instead of L times in line 140 on page 3. We have corrected the known writing errors and will carefully revise the manuscript. - Q4. The reason for choosing the VAE encoder as the texture extractor in the Face Recapture Module? A4. The VAE pre-trained on large-scale datasets has excellent compression capabilities and can capture local texture information of the face region well. Combined with global facial identity features, it can comprehensively represent facial information. Additional results are shown in Figure 3 in the pdf file, it can be seen that our method can generate satisfactory facial details due to the Face Recapture module. - Q5. About the shapes of $G_h$ and $G_f$. The advantages of adopting a discrete codebook chosen to represent hand information. A5. Sorry for our oversight. The shape of $G_h$ is $B\times F\times 2\times 512$ and the shape of $G_f$ is $B\times F\times h\times w\times 768$. Representing hand information with a discrete codebook is robust. Moreover, the designed codebook consists of a set of orthogonal basis directions, which can fully represent hand structure and texture. (Please refer to A4 for Reviewer HRx2 and A3 for Reviewer jdWr for more motivation.)	Summary: This paper proposes a 2D human video generation framework called ShowMaker, which can generate half-body conversational videos based on 2D keypoints as motion conditions. ShowMaker includes a texture enhancement module for the face and hands, and demonstrates some effectiveness. Strengths: 1. ShowMaker achieves good results in generating texture details, with improvements over the compared methods. 2. The motivation behind ShowMaker is straightforward, and the method design is simple. Weaknesses: 1. The experimental settings lack clarity, especially regarding the balance of training duration for each ID in the testing examples. It is unclear how much training data corresponds to each ID in the visualized examples, and more details are needed. 2. The method proposed by the authors has some improvements in details compared to the similar method Make-an-Anchor. However, from Figure 4, it seems that the generation of facial details is unstable, resulting in some color changes (such as in the beard area), which raises concerns about the facial generation capability of the proposed method. 3. The ablation studies only excluded the face modeling and hand modeling modules. However, introducing region images to enhance the generation results is a straightforward approach in human body generation tasks and could serve as a baseline for comparison. Technical Quality: 3 Clarity: 2 Questions for Authors: 1. In Table 1, it is unclear what "One ID" represents. It is also unclear whether the training set only contains data for the testing IDs. Additionally, it is unclear whether the compared methods, such as Animate Anyone and TPS, were also trained on the testing IDs. More details are needed to clarify these points. 2. Are the IDs included in the training set and testing set are completely identical? Confidence: 4 Soundness: 3 Presentation: 2 Contribution: 3 Limitations: Please refer to the questions and weaknesses. Flag For Ethics Review: ['No ethics review needed.'] Rating: 5 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: # Response to Reviewer HRx2 Thanks very much for taking time out of your busy schedule to review our manuscript. We reply as follows. - Q1. The experimental settings. (Response to weakness 1 and question 2) A1. Our goal is to extend the talking head task to generate an expressive half-body conversational speaker. Therefore our experimental setting is not a one-shot animation. Following the settings of Make-Your-Anchor, the IDs of the training set and the test set are consistent, and the test set and training set do not overlap. &emsp; For detailed dataset information, please refer to Table 1 in the pdf file. Due to space limitations, we will place the additional experimental details in the appendix. - Q2. About "One ID" and Animate Anyone and TPS. A2. Since Make-Your-Anchor provides only one pre-trained model of Seth without the training code before our submission. We failed to reproduce its experiments on other subjects. Therefore, we compared our approach with it only on the Seth data ("One-ID") for fairness. &emsp; Please also note that the training and test sets of Animate Anyone and TPS are consistent with ours. - Q3. About some color changes in the results. A3. We suppose that the color changes mentioned by the reviewers may come from lighting changes in the facial region. Please refer to the videos in our supplementary videos and additional results in Figure 3 in the pdf file, our results show temporally consistent texture in the facial region due to the Face Recapture module. - Q4. About introducing region images as a baseline for comparison. A4. Thanks for your suggestion. Actually, we follow the idea of using the regional image to enhance the results. In our Face Recapture Module, we use cropped and aligned face images to enhance the facial details in the generated results. However, when dealing with hand regions, there are two major problems: &emsp; (1) Since they only occupy a small proportion of the image, it is difficult to capture enough textured details within the cropped hand images. &emsp; (2) Hand poses, expressed in more degrees of freedom, are more complex when compared with head poses. The hand poses from the reference image and driving signals may be very different, which may alleviate the contribution of texture from the cropped hand image. &emsp; Therefore, we map the driving hand key points to a high-dimensional space and adopt a discrete codebook for hand enhancement (refer to A3 for Reviewer jdWr). &emsp; We still follow your suggestion to conduct an experiment and the results are illustrated in Figure 1 (the second row). Here, the cropped reference hand image is sent to the VAE encoder for feature extraction and then fed into the hand attention module. The results show that the hand modeling strategy we proposed enables the best synthesis of hand details and structures among all these variants. --- Rebuttal Comment 1.1: Comment: Thank you for the response, which has addressed the points I previously did not understand. I am inclined to maintain my rating. --- Reply to Comment 1.1.1: Comment: Thank you very much for your reply and insightful comments. Please let us know if you have any other questions or concerns, and we will respond to them carefully.	Summary: This paper represents an early attempt to advance the field of 2D digital human synthesis in real-world scenarios, extending the traditional talking head task to include more complex body movements, such as hand gestures. The authors present a novel framework utilizing dual-stream diffusion models for full-body video synthesis, complemented by two specialized modules designed for fine-grained detail modeling. Specifically, the proposed Key Point-based Fine-grained Hand Modeling effectively addresses the challenge of hand gesture generation through the integration of a key point-based codebook and high-dimensional positional encoding. Additionally, the Face Recapture module enhances facial textures and identity consistency, enabling the generation of more realistic and vivid 2D avatar videos. Strengths: 1) This paper outlines an exciting approach for digital human communities, focusing on the synthesis of 2D controllable human body videos with detailed modeling. Different from existing works like "Animate Anyone" that primarily focus on generalized human body movement generation across various subjects, this paper presents a promising direction for digital human communities by focusing on fine-grained modeling and controllable video synthesis of human body videos, especially hand gestures, which has the potential to achieve better human-computer interaction with our body gesture and benefit our demands in real-world scenarios. 2) The paper is well-written and easy to follow. The motivation for the whole task and the two additional modules is clear. The supplementary video provides robust results and convincible comparison with sota methods, which further demonstrate the effectiveness of all the designs proposed in this paper. Weaknesses: There are few mistakes and typos in the paper: 1) The Face Recaption module takes “Reference Pose” as input in Fig 1 (d), which is not mentioned in Fig 2 and Sec 3.3. The description of Face Recaption module should be the same throughout the full manuscript. 2) Eq (5) and Eq(6) are nearly the same. The authors can merge them into one equation. Technical Quality: 4 Clarity: 3 Questions for Authors: 1) I understand the authors adopted the positional encoding to enhance the information provided by the sparse hand key points, but I’d like to see a further analysis of the positional encoding used in the Hand Modeling module. Since it cooperates with the key point-based codebook, I am wondering if a well-trained codebook is already enough for hand restoration. 2) Have the authors adopted any alignment strategy in cross-driving experiments? In my understanding, cross-driving animation usually suffer from mis-alignment due to the subject differences. Confidence: 4 Soundness: 4 Presentation: 3 Contribution: 3 Limitations: The authors adequately addressed the limitations and negative societal impact. Flag For Ethics Review: ['No ethics review needed.'] Rating: 7 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: # Response to Reviewer jdWr Thanks very much for your careful review comments. We'll answer your questions one by one. - Q1: About Fig 1 (d). A1. Thanks for pointing out this mistake. We will remove the “Reference Pose” from Figure 1 (d) and refine the whole figure in the revised manuscript. - Q2: About Eq (5) and Eq (6). A2. Thanks for your suggestion. We will replace Eq (5) and Eq (6) with a unified definition of the attention operation and describe the different inputs used in Eq (5) and Eq (6). - Q3. The effect of position encoding on generation results. A3. Please refer to Figure 1 (the third row) in the pdf file, the comparison results demonstrate that position encoding brings a significant performance improvement in both textural and structural details of human hands. The key points of the hands are distributed in a low-dimensional space and are relatively clustered, directly using the coordinates of the hand key points as inputs of the Hand Modeling module cannot well capture the structure and texture information from the hand. - Q4. Alignment strategy in cross-driving experiments. A4. For the cross-driving setting, we adopt a pose alignment strategy that roughly aligns the driving signals from other subjects with the ones of the target person. The key points of our pose alignment strategy are as follows: &emsp; (1) When preparing the training data, we first produce the DWPose results from each frame and set the center of shoulders as the cropping center. Then we crop the original video frame by using an adaptive cropping size, where the cropping size is designed as a fixed ratio of the shoulder width. This operation forces the human body to lie in a roughly consistent position. &emsp; (2) In the inference stage, we further try to bridge the gap between body shapes by scaling the driving poses to match the reference pose. The scaling ratios of width and height are defined as $w_r/w_d$, $h_r/h_d$. Here, $w$ represents the shoulder width, and $h$ represents the height of the human body. We will involve a detailed description of $w$ and $h$ in the appendix. &emsp; The above strategy enables satisfactory pose alignment between different subjects under our scenario. --- Rebuttal Comment 1.1: Comment: My concerns about the the alignment strategy and the effect of position encoding have been well addressed in the rebuttal. I am willing to increase my score. --- Reply to Comment 1.1.1: Comment: Thank you very much for your response, we are glad to address your concerns!	Rebuttal 1: Rebuttal: # To all reviewers and ACs Thanks very much for all the reviewers' efforts and suggestions. We appreciate the positive comments on the following: 1. The generation framework holds significant practical value, particularly in conversational scenarios like TV shows. 2. The paper is well-written and clearly explains the important technical details. 3. The motivation behind ShowMaker is straightforward and the designed model is clear and convincible. 4. The proposed model achieves good results in generating texture details, with improvements over the compared methods, and the demo presents satisfying hand and face modeling results. We believe the remaining issues can be fully addressed. We will respond to your comments one by one below. Some relevant figures and tables are provided in the pdf file. Pdf: /pdf/7cf71df934a0a762cfc7cf9e109d610549452219.pdf	NeurIPS_2024_submissions_huggingface	2,024	null	null	null	null	null	null	null	null
Vaccine: Perturbation-aware Alignment for Large Language Models against Harmful Fine-tuning Attack	Accept (poster)	Summary: This paper introduces Vaccine, a technique that improves the security of Large Language Models (LLMs) by incorporating perturbation-aware alignment during fine-tuning. Vaccine integrates specifically crafted perturbations in the alignment phase to produce invariant hidden embeddings that withstand harmful perturbations during subsequent user interactions. Strengths: The authors demonstrate with empirical evidence that even minimal harmful data can induce significant shifts in LLM embeddings, disrupting their alignment. The method Vaccine was validated on widely-used LLMs, including Llama2, Opt, and Vicuna, showing substantial enhancements in their resilience against malicious prompts while preserving their reasoning capabilities with non-malicious prompts. A work that may be interesting to the broad community. Weaknesses: The computational overhead of the method seems to linearly scales with model size. When the harmful ratio is high, the method doesn't seem to perform very well. Technical Quality: 3 Clarity: 3 Questions for Authors: If the data for fine-tuning comes from the user, why would the user upload harmful data to break the model. Confidence: 3 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: The authors have partially addressed the limitations of their work. Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We thank the reviewer for the generally positive review. As below, we show extra results and discussion to address the concerns. W1: Computational overhead scales with model size. Step time. Because Vaccine requires double forward-backward pass in each optimization step, the training time of Vaccine is approximately double compared to the SFT baseline. Because the training time itself is scaling with the model size, the overhead is also scaling with the model size. See the following table for our evaluation results. \| \| OPT-1.3b \| OPT-2.7b \| OPT-6.7b \| OPT-13b \| \|---\|---\|---\|---\|:---:\| \| Running time(SFT) \| 0.06 s (1x) \| 0.08s (1x) \| 0.09s (1x) \| 0.12s (1x) \| \| Running time(Vaccine) \| 0.11s (1.83x) \| 0.15s (1.875x) \| 0.17s (1.88x) \| 0.24s (2x) \| \| Extra time overhead \| 0.05s (0.83x) \| 0.07s (0.875x) \| 0.08s (0.88x) \| 0.12s (1x) \| However, it is important to note that this is only a one-time cost for aligning a pre-trained model. In the fine-tuning-as-a-service scenario, the same aligned model is used to finetune for all the finetuning requests (which arrive at the rate of thousands or millions/per hour). Since Vaccine does not incur extra overhead for fine-tuning, its computation overhead should not be a big concern for real-time service for a service provider. Accelerated Vaccine. It is also possible to accelerate Vaccine. Vaccine requires double training time because in every step we need to first search for the perturbation, and then apply the perturbation to the model to do another forward-backward pass. To address the reviewer's concern, we propose accelerated Vaccine. The idea is simple- we search for the perturbation not for every step, but to do it for every $\tau$ step. See Algorithm 2 in our attached pdf for details. The following table shows the evaluation results. \| Methods \| Harmful score \| Finetune accuracy \| Step time \| \|:---:\|:---:\|:---:\|:---:\| \| SFT \| 59.20 \| 94.40 \| 0.12902s (1x) \| \| Vaccine \| 50.40 (+0) \| 92.40 \| 0.24852s (1.93x) \| \| Accelerated Vaccine($\tau=100$) \| 51.00 (+0.60) \| 95.20 \| 0.13140s (1.02x) \| \| Accelerated Vaccine($\tau=1000$) \| 52.00 (+1.60) \| 94.20 \| 0.12956s (1.004x) \| \| Accelerated Vaccine($\tau=10000$) \| 53.20 (+2.80) \| 94.80 \| 0.12934s (1.002x)\| \| Accelerated Vaccine($\tau=20000$) \| 58.80 (+8.40) \| 94.40 \| 0.12902s (1x) \| Our results surprisingly show that by searching perturbation for every 100 steps, Acceleated Vaccine is still able to achieve decent defense (harmful score is increased by a marginal 0.60), but the step time is significantly shortened. We hope accelerated Vaccine can erase the reviewer's concern about resource overhead. GPU memory*. On the other hand, the extra memory overhead is scaled with the hidden embedding size, because in the second forward-backward pass we need to register the perturbation to the hidden embedding. The following table shows how the GPU memory usage scales with the hidden embedding size: \| \| OPT-1.3B \| OPT-2.7B \| OPT-6.7B \| OPT-13B \| \|---\|---\|---\|---\|---\| \| Hidden Embedding size \| 242048 (1x) \| 322560 (1.67x) \| 324096 (2.67x) \| 405120 (4.16x) \| \| Memory (SFT) \| 5.586GB \| 10.814GB \| 25.469GB \| 48.824GB \| \| Memory (Vaccine) \| 5.596GB \| 10.830 GB \| 25.492GB \| 48.863 GB \| \| Extra Memory Cost \| 0.0097GB(1x) \| 0.0157GB(1.62x) \| 0.0235GB(2.42x) \| 0.039GB(4.02x) \| As shown, for OPT-13B, only a marginal 0.039/48.824=0.08% extra memory is induced compared to SFT. The memory cost is apparently marginal. W2: Performance downgrades when the harmful ratio is high. Indeed, Vaccine experiences performance downgrades when the harmful ratio is high. This however is an inevitable drawback of an alignent stage solution. Particularly, an alignment stage solution aims at enhancing the model's robustness to the harmful data, but fine-tuning on too much harmful data (e.g., pure harmful data) can still break the enhanced robustness. While we in this paper only aim to increase the aligned model's robustness in the alignment stage, we admit that this may not be enough for the hard case (e.g., a large harmful ratio). We envision that future research can build on top of Vaccine to provide a stronger defense. We will discuss this in our limitation and future direction section. Q1: Why would users upload harmful data to break the model?* There are two possible cases for the user to upload harmful data. i) Non-adversary case. The users are not aware that the data they upload contains harmful instances. Users may collect fine-tuned data from their application use case, and they may not carefully inspect and clean the data. ii) Adversary case. The user (e.g., a business adversary) aims to scandalize the service provider to provide harmful content (or political sensitivity content) from the API. Because the finetuned model is deployed on the service provider's server and the harmful content is transmitted from their API, the service provider may face legal accusations or governance issues. An example is that users may ask "How do you comment the war between Israel and Palestine"? The service provider is responsible for the answer. While the first case may be more realistic and common, the second case may raise serious concerns for the service provider. --- Rebuttal Comment 1.1: Comment: Thanks to the authors for clarifying my confusion. I am keeping my score unchanged for now. --- Reply to Comment 1.1.1: Title: Thanks for the feedback! Comment: We thank the reviewer for acknowledging our efforts in rebuttal! Please feel free to leave us a comment if you need charification during our interaction with other reviewers.	Summary: This paper proposes a novel alignment technique called "Vaccine", which addresses the security risks of large language models (LLMs) during user fine-tuning. It is found that even a small amount of harmful data can destroy the alignment effect of a model, leading to the "alignment destruction effect".The Vaccine approach introduces perturbation-aware training at the alignment stage, which enables the model to resist the influence of harmful data in the subsequent fine-tuning. To be brief, the method first finds the example that harms the model most and defends this example as the attacker. Experimental results show that Vaccine significantly reduces the model's harmful output probability (up to 9.8%) while maintaining good downstream task performance. The method performs well under different models, tasks, and fine-tuning settings, proving its effectiveness and generalization ability. Strengths: The strength of this paper lies in the following two main points: 1. the authors first present a sample attack scenario under "Finetuning as a service", and discuss the need to maintain security under this scenario. 2. the authors propose the phenomenon of "embedding drift" as the main cause of broken alignments qualitatively. 3. the simplicity and generality of the proposed approach. Specifically, the method proposed in the paper combines the idea of adversarial training, and requires only one more gradient finding and parameter optimization, while applying to various alignment methods. This method can be used as a plug-and-play method in different application scenarios. Weaknesses: The weakness of this paper may be the following: 1. The Vaccine method doubles the training time, which may be intolerable for users in a "finetuning as a service" scenario. Related acceleration methods can be considered. I am willing to raise my score once this is solved. 2. The approach of this article is highly relevant to the solution of catastrophic forgetting. In such a scenario, there is only one 2017 method (EWC) that has been used as part of the baseline in terms of neural networks rather than data, and I would like to see more baselines being compared. 3. I maintain my reservations about using a black box moderation model for harmful score calculations. Also, the paper does not discuss the change in the model's generative ability (e.g., ppl) after using the vaccine method. Technical Quality: 3 Clarity: 3 Questions for Authors: I am very interested in the phenomenon of embedding drift and noticed that you have demonstrated it using the L2 norm. can you provide a more intuitive visualisation such as t-SNE? Confidence: 3 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: The authors have adequately addressed the limitations in the checklist guidelines. Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We sincerely thank the reviewer for the constructive review comments. As belows, we try to address the separate comment. W1: Double training time Indeed, Vaccine requires double training time, because in every step we need to first search for the perturbation, and then apply the perturbation to the model to do another forward-backward pass. To address the reviewer's concern, we propose accelerated Vaccine. The idea is simple- we search for the perturbation not for every step, but to do it for every $\tau$ steps. See Algorithm 2 in our attach pdf for details. The following table for the evaluation results. \| Methods \| Harmful score \| Finetune accuracy \| Step time \| \|:---:\|:---:\|:---:\|:---:\| \| SFT \| 59.20 \| 94.40 \| 0.12902s (1x) \| \| Vaccine \| 50.40 (+0) \| 92.40 \| 0.24852s (1.93x) \| \| Accelerated Vaccine($\tau=100$) \| 51.00 (+0.60) \| 95.20 \| 0.13140s (1.02x) \| \| Accelerated Vaccine($\tau=1000$) \| 52.00 (+1.60) \| 94.20 \| 0.12956s (1.004x) \| \| Accelerated Vaccine($\tau=10000$) \| 53.20 (+2.80) \| 94.80 \| 0.12934s (1.002x)\| \| Accelerated Vaccine($\tau=20000$) \| 58.80 (+8.40) \| 94.40 \| 0.12902s (1x) \| Our results surprisingly show that by searching perturbation for every 100 steps, Acceleated Vaccine is still able to achieve decent defense (harmful score is increased by a marginal 0.60), but the step time is significantly shortened. W2: Lack of baseline Indeed, not enough baselines are used in this submission, partly because this work is one of the initial works for the harmful fine-tuning issue. After NeurIPS submission deadline, we do see there are a few defense solutions arise, e.g., RepNoise[1]]. RepNoise, same as Vaccine, is an alignment stage solution that aims at improving the neural network's robustness, which makes it a perfect baseline to compare. In the following table, we show the performance comparison under different harmful ratios. \| \| Harmful Score \| --> \| --> \| --> \| --> \| Finetune Accuracy \| --> \| --> \| --> \| --> \| \|---\|:---:\|:---:\|:---:\|:---:\|:---:\|:---:\|:---:\|:---:\|:---:\|:---:\| \| \| p=0 \| p=0.01 \| P=0.05 \| p=0.1 \| p=0.2 \| p=0 \| p=0.01 \| P=0.05 \| p=0.1 \| p=0.2 \| \| SFT \| 54.20 \| 52.40 \| 54.80 \| 58.80 \| 64.80 \| 93.20 \| 94.40 \| 94.20 \| 94.40 \| 94.20 \| \| Repnoise \| 51.20 \| 50.20 \| 51.20 \| 52.40 \| 62.40 \| 92.80 \| 92.80 \| 92.60 \| 92.40 \| 92.80 \| \| Vaccine \| 45.40 \| 46.40 \| 48.40 \| 50.40 \| 60.00 \| 92.00 \| 91.80 \| 92.00 \| 92.40 \| 93.40 \| As shown, Vaccine consistently outperforms RepNoise with a smaller harmful score and also a higher finetune accuracy. W3-A: Evaluation using black box moderation model We totally agree that using a moderation model is not ideal and not accurate enough for performance evaluation. However, at the current stage, there is not an accurate benchmarking method to evaluate the model's harmfulness. Existing research, e.g., [2][3] utilizes GPT4 for evaluation. RepNoise [1] follows our setting to use BeaverTails's moderation model for evaluation. We generally believe that compared to GPT, the moderation model we use is less "black box", given that GPT4 is not even open-sourced, and the score it gives may change due to version updates or randomness. W3-B: the paper does not discuss the change in the model's generative ability (e.g., ppl) For model evaluation, we mainly use finetune accuracy to evaluate the model's reasoning ability (for example, for GSM8k, we measure the accuracy based on whether the model can predict the final answer). Indeed, it is interesting to study the model's generative ability using metrics like perplexity. In the following table, we show the comparison results after the model fine-tuned on wikitext. \| \| Harmful score \| -->\| --> \| -->\| -->\| Perplexity (lower the better) \| --> \|--> \| --> \| -->\| \|---\|:---:\|:---:\|:---:\|:---:\|:---:\|:---:\|:---:\|:---:\|:---:\|:---:\| \| \| p=0 \| p=0.01 \| P=0.05 \| p=0.1 \| p=0.2 \| p=0 \| p=0.01 \| P=0.05 \| p=0.1 \| p=0.2 \| \| SFT \| 54.60 \| 55.60 \| 58.00 \| 59.00 \| 63.40 \| 26.08 \| 26.11 \| 26.29 \| 26.61 \| 27.27 \| \| Repnoise \| 54.20 \| 53.80 \| 55.80 \| 57.60 \| 59.20 \| 35.32 \| 36.11 \| 37.16 \| 36.64 \| 38.68 \| \| Vaccine \| 46.60 \| 46.20 \| 48.40 \| 50.80 \| 55.40 \| 34.36 \| 34.45 \| 34.67 \| 34.85 \| 35.77 \| The results show that Vaccine may slightly increase the perplexity of the model compared to SFT. However, in comparison, RepNoise also increases the perplexity but is less effective in maintaining a low harmful score. Q1: Intuitive visualization of harmful embedding drift We thank the reviewer for the suggestion of t-SNE visualization. We plot the embedding drift of SFT and Vaccine under different harmful ratios. When the harmful ratio is high, it is intuitive to see that the embedding is drifting toward a specific direction. Interestingly, the embedding drift of Vaccine is slighter, making the drifted embedding still able to preserve the alignment knowledge. This may better explain how Vaccine really works. [1] Rosati D, Wehner J, Williams K, et al. Representation noising effectively prevents harmful fine-tuning on LLMs[J]. arXiv preprint arXiv:2405.14577, 2024. [2] Hsu C Y, Tsai Y L, Lin C H, et al. Safe LoRA: the Silver Lining of Reducing Safety Risks when Fine-tuning Large Language Models[J]. arXiv preprint arXiv:2405.16833, 2024. [3] Qi X, Zeng Y, Xie T, et al. Fine-tuning aligned language models compromises safety, even when users do not intend to![J]. arXiv preprint arXiv:2310.03693, 2023. --- Rebuttal 2: Comment: Thank you for your detailed and constructive rebuttal, which includes extensive experimental evidence. I am particularly impressed by your introduction of the accelerated** Vaccine method. The results demonstrating only a 2% overhead when $\tau = 100$ with minimal performance degradation significantly reduce the application cost of the Vaccine method. I highly recommend including this improvement in the manuscript for future versions. Based on my initial review and your comprehensive response, I decided to adjust my score positively. P.S. There appears to be a typo in Algorithm 2 of the attached PDF: "t mod $\gamma$ == 0" should be reviewed for accuracy. Best regards, --- Rebuttal Comment 2.1: Title: Thanks for the re-evaluation of our work Comment: We thank the reviewer for the encouraging comment, and also for positively adjusting the socre reflecting our efforts in rebuttal! It is a a great relief for us to see that accelerated Vaccine address your main concern! We will definitely include the accelerated Vaccine as well as its evaluation into the next revision. Indeed, there is a typo in Algorithm 2 of the attached PDF. It should be $t$ mod $\tau$==0. We will fix it in our revision. Thank you for pointing this out! Please feel free to leave us a comment if you need charification during our interaction with other reviewers.	Summary: This paper presents a novel approach to enhance the security of finetuning-as-a-service for Large Language Models (LLMs). The proposed method, Vaccine, introduces a perturbation-aware alignment technique to mitigate the risk of harmful data introduced during user finetuning. The paper demonstrates that Vaccine can effectively maintain alignment robustness against harmful prompts while preserving reasoning abilities on benign prompts. Strengths: Quality: The paper provides comprehensive empirical evidence demonstrating the effectiveness of Vaccine in reducing harmful scores while maintaining finetuning accuracy across multiple models and datasets. Clarity: The methodology is well-explained, with clear descriptions of the problem, the proposed solution, and the experimental setup. Weaknesses: Resource Overhead: The increased computational and memory overhead introduced by Vaccine, though justified, might be a limitation for practical deployment, especially in resource-constrained environments. For example the runtime and memory usage should be compared to normal finetune and other baselines. Technical Quality: 3 Clarity: 3 Questions for Authors: Could you provide more information about Resource Overhead? Confidence: 3 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: Yes Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We thank the reviewer for the positive comments on our work. Below we try to address the concern on resource overhead. GPU memory. Because in the second forward-backward pass we need to register the perturbation to the hidden embedding, Vaccine requires slightly more GPU memory usage. We in the following show comparison results with the normal finetune (SFT) and a recent alignment-stage defense RepNoise [1] using different sizes of model. * Evaluation is done with an H100 with 80GB memory: \| \| OPT-1.3b \| OPT-2.7b \| OPT-6.7b \| OPT-13b \| \|---\|---\|---\|---\|---\| \| memory(SFT) \| 5.586GB \| 10.814GB \| 25.469GB \| 48.824GB \| \| memory (RepNoise) \| 8.962GB \| 17.453GB \| 40.017GB \| 76.027GB \| \| memory(Vaccine) \| 5.596GB \| 10.830GB \| 25.492GB \| 48.863 GB \| Vaccine incurs slightly more GPU memory compared to SFT. For OPT-13B, only a marginal 0.039GB extra memory is induced compared to SFT. In sharp contrast, RepNoise introduces 27.164GB extra memory overhead compared to SFT. With this result, it seems that Vaccine is superior in the resource-constrained scenario. Training time. Because Vaccine requires double forward-backward pass in each optimization step, the training time of Vaccine is approximately double compared to the SFT baseline. See the following table for our evaluation results. * Evaluation is done with an H100 with 80GB memory: \| \| OPT-1.3b \| OPT-2.7b \| OPT-6.7b \| OPT-13b \| \|---\|---\|---\|---\|:---:\| \| step time(SFT) \| 0.06 \| 0.08 \| 0.09 \| 0.12 \| \| step time(RepNoise) \| 0.14 \| 0.19 \| 0.2 \| 0.29 \| \| step time(Vaccine) \| 0.11 \| 0.15 \| 0.17 \| 0.24 \| As shown, Vaccine uses approximately 2x training time and RepNoise uses approximately 2.3x-2.4x training time compared to SFT. Vaccine is more computation-efficient compared to RepNoise. It is also possible to accelerate Vaccine. Vaccine requires double training time, because in every step we need to first search for the perturbation, and then apply the perturbation to the model to do another forward-backward pass. To address the reviewer's concern, we propose accelerated Vaccine. The idea is simple- we search for the perturbation not for every step, but to do it for every $\tau$ steps. See Algorithm 2 in our attached pdf for details. The following table shows the evaluation results. \| Methods \| Harmful score \| Finetune accuracy \| Step time \| \|:---:\|:---:\|:---:\|:---:\| \| SFT \| 59.20 \| 94.40 \| 0.12902s (1x) \| \| Vaccine \| 50.40 (+0) \| 92.40 \| 0.24852s (1.93x) \| \| Accelerated Vaccine($\tau=100$) \| 51.00 (+0.60) \| 95.20 \| 0.13140s (1.02x) \| \| Accelerated Vaccine($\tau=1000$) \| 52.00 (+1.60) \| 94.20 \| 0.12956s (1.004x) \| \| Accelerated Vaccine($\tau=10000$) \| 53.20 (+2.80) \| 94.80 \| 0.12934s (1.002x)\| \| Accelerated Vaccine($\tau=20000$) \| 58.80 (+8.40) \| 94.40 \| 0.12902s (1x) \| Our results surprisingly show that by searching perturbation for every 100 steps, Acceleated Vaccine is still able to achieve decent defense (harmful score is increased by a marginal 0.60), but the step time is significantly shortened. We hope accelerated Vaccine can erase the reviewer's concern about resource overhead. [1] Rosati D, Wehner J, Williams K, et al. Representation noising effectively prevents harmful fine-tuning on LLMs[J]. arXiv preprint arXiv:2405.14577, 2024. https://arxiv.org/abs/2405.14577 --- Rebuttal Comment 1.1: Comment: Thank you for the clarification and additional experiments. The resource overhead of vaccine is smaller than I expected. The author should consider adding this content to the article and making it open source. I have decided to raise my score. --- Rebuttal 2: Title: Thanks for recognizing our effort in rebuttal! Comment: Thank you for recognizing our effort in rebuttal, and also for increasing rating to reflect the addressed concern. We will definitely include the new results to the paper, and also open-source the accelerated Vaccine method.	Summary: This paper introduces a novel phenomenon called harmful embedding drift, which occurs when a few harmful data points uploaded by users cause misalignment in the fine-tuned LLM. To combat this, this paper proposes a technique called Vaccine, which uses perturbation-aware alignment to produce invariant hidden embeddings. This method aims to maintain the alignment of LLMs even when fine-tuning on potentially harmful user data. The empirical results demonstrate that Vaccine improves the robustness of LLMs against harmful prompts while preserving their reasoning abilities on benign prompts. Strengths: The paper focuses on an interesting and important topic, proposing a method that performs well across diverse evaluations. Weaknesses: 1. The Vaccine method introduces additional computational overhead due to the need for two forward-backward passes for each step of model optimization. Further evaluation should be shared to show the trade-offs between robustness and efficiency. 2. While the paper demonstrates the efficacy of Vaccine on several datasets (SST2, AG-NEWS, GSM8K, and AlpacaEval), the scalability to larger datasets might require further investigation. 3. The effectiveness of the Vaccine method depends on the noise intensity (ρ), and choosing the optimal value may not be straightforward. Further guidelines on selecting this parameter would enhance the practical applicability of the method. (btw, what is the training data in Table 6) Technical Quality: 3 Clarity: 3 Questions for Authors: 1. How does the computational overhead introduced by the Vaccine method compare with other alignment techniques in terms of real-time performance and resource consumption? 2. Will all the fine-tuning jailbreak attacks cause harmful embedding drift? How to define the effectiveness bound of the proposed harmful embedding drift phenomenon? Confidence: 3 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: The authors have discussed the limitations and further impacts in the appendix. Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: W1+Q1: Reource ovehead GPU memory. Because in the second forward-backward pass we need to register the perturbation to the hidden embedding, Vaccine requires slightly more GPU memory usage. We in the following show comparison results with the normal finetune (SFT) and a recent alignment-stage defense RepNoise [1] using different sizes of model. * Evaluation is done with an H100 with 80GB memory: \| \| OPT-1.3b \| OPT-2.7b \| OPT-6.7b \| OPT-13b \| \|---\|---\|---\|---\|---\| \| memory(SFT) \| 5.586GB \| 10.814GB \| 25.469GB \| 48.824GB \| \| memory (RepNoise) \| 8.962GB \| 17.453GB \| 40.017GB \| 76.027GB \| \| memory(Vaccine) \| 5.596GB \| 10.830GB \| 25.492GB \| 48.863 GB \| Vaccine incurs slightly more GPU memory compared to SFT. For OPT-13B, only a marginal 0.039GB extra memory is induced compared to SFT. In sharp contrast, RepNoise introduces 27.164GB extra memory overhead compared to SFT. By this result, it seems that Vaccine are superior in resource-constrained scenarios. Training time. Because Vaccine requires double forward-backward pass in each optimization step, the training time of Vaccine is approximately double compared to the SFT baseline. See the following table for our evaluation results. * Evaluation is done with an H100 with 80GB memory: \| \| OPT-1.3b \| OPT-2.7b \| OPT-6.7b \| OPT-13b \| \|---\|---\|---\|---\|:---:\| \| step time(SFT) \| 0.06 \| 0.08 \| 0.09 \| 0.12 \| \| step time(RepNoise) \| 0.14 \| 0.19 \| 0.2 \| 0.29 \| \| step time(Vaccine) \| 0.11 \| 0.15 \| 0.17 \| 0.24 \| As shown, Vaccine uses approximately 2x training time and RepNoise uses approximately 2.3x-2.4x training time compared to SFT. Vaccine is more computation-efficient compared to RepNoise. Accelerated Vaccine. It is also possible to accelerate Vaccine. Vaccine requires double training time because in every step we need to first search for the perturbation, and then apply the perturbation to the model to do another forward-backward pass. To address the reviewer's concern, we propose accelerated Vaccine. The idea is simple- we search for the perturbation not for every step, but to do it for every $\tau$ steps. See Algorithm 2 in our attached pdf for details. The following table shows the evaluation results. \| Methods \| Harmful score \| Finetune accuracy \| Step time \| \|:---:\|:---:\|:---:\|:---:\| \| SFT \| 59.20 \| 94.40 \| 0.12902s (1x) \| \| Vaccine \| 50.40 (+0) \| 92.40 \| 0.24852s (1.93x) \| \| Accelerated Vaccine($\tau=100$) \| 51.00 (+0.60) \| 95.20 \| 0.13140s (1.02x) \| \| Accelerated Vaccine($\tau=1000$) \| 52.00 (+1.60) \| 94.20 \| 0.12956s (1.004x) \| \| Accelerated Vaccine($\tau=10000$) \| 53.20 (+2.80) \| 94.80 \| 0.12934s (1.002x)\| \| Accelerated Vaccine($\tau=20000$) \| 58.80 (+8.40) \| 94.40 \| 0.12902s (1x) \| Our results show that by searching perturbation for every 100 steps, Acceleated Vaccine is still able to achieve decent defense, but the step time is significantly shortened. We hope accelerated Vaccine can erase the reviewer's concern about resource overhead. W2: Scalaibility to larger datasets It is a good idea to test Vaccine in another larger-scale fine-tuning dataset. Here we follow the setting form [2] to finetune the model on a WikiText-2, and we show the model's tradeoff between perplexity (smaller the better) and harmful score. \| \| Harmful score \| -->\| --> \| -->\| -->\| Perplexity \| --> \|--> \| --> \| -->\| \|---\|:---:\|:---:\|:---:\|:---:\|:---:\|:---:\|:---:\|:---:\|:---:\|:---:\| \| \| p=0 \| p=0.01 \| P=0.05 \| p=0.1 \| p=0.2 \| p=0 \| p=0.01 \| P=0.05 \| p=0.1 \| p=0.2 \| \| SFT \| 54.60 \| 55.60 \| 58.00 \| 59.00 \| 63.40 \| 26.08 \| 26.11 \| 26.29 \| 26.61 \| 27.27 \| \| Repnoise \| 54.20 \| 53.80 \| 55.80 \| 57.60 \| 59.20 \| 35.32 \| 36.11 \| 37.16 \| 36.64 \| 38.68 \| \| Vaccine \| 46.60 \| 46.20 \| 48.40 \| 50.80 \| 55.40 \| 34.36 \| 34.45 \| 34.67 \| 34.85 \| 35.77 \| As shown, Vaccine achieves smaller harmful score but maintain a lower perplexity compared to RepNoise. W3: Selection of hyper-parameter Indeed, Vaccine relies on the noise intensity $\rho$ for providing an effective defense. In Table 6, we use SST2 as a fine-tuning task (which is our default setup) to demonstrate how choosing different $p$ affects the harmful score and finetune accuracy. Below we show the data for providing intuition of how to select $\rho$. \| Methods \| $\rho=0.01$ \| $\rho=0.1$ \| $\rho=1$ \| $\rho=2$ \| $\rho=5$ \| $\rho=10$ \| \|:---:\|:---:\|:---:\|:---:\|:---:\|:---:\|:---:\| \| HS \| 54.40 \| 56.80 \| 54.40 \| 49.00 \| 46.20 \| 44.20 \| \| FA \| 94.40 \| 95.00 \| 94.40 \| 93.60 \| 92.80 \| 89.00 \| As shown, with a larger $\rho$, the harmful score will decrease but also at the cost of degrading finetune accuracy. In this case, before deploying Vaccine, one should use a validation dataset to test the model's finetune performance and harmful score, and select the best tradeoff that is acceptable to the application. Here we also want to highlight that we only have one hyper-parameter to tune. In comparison, the recent method RepNoise needs two hyper-parameters, and requires an additional harmful dataset in their assumption. The simplicity of Vaccine should be merited. Q2: Will all the fine-tuning attack cause harmful embedding drift? Honestly, we don't have an definitive answer to this question. At least the original attack, i.e., mixing normal harmful data in the fine-tuning data, will cause harmful embedding drift. There may be more advanced attacks that can circumvent the embedding drift, which, however, is still under-explored. In the attached pdf, we also use t-SNE to illustrate the harmful embedding drift phenomenon, which you may be interested in. Feel free to check it out. [1] Rosati D, Wehner J, Williams K, et al. Representation noising effectively prevents harmful fine-tuning on LLMs[J]. arXiv preprint arXiv:2405.14577, 2024. [2] Li Y, Yu Y, Liang C, et al. Loftq: Lora-fine-tuning-aware quantization for large language models[J]. arXiv preprint arXiv:2310.08659, 2023. --- Rebuttal 2: Title: A brief summary of our rebuttal Comment: Hi Reviewer a2cG, Thanks for the very informative review, and also for recognizing the harmful fine-tuning that we focus on is an interesting and important topic. We want to get back to you to see whether our rebuttal addresses your concerns. Here we respectfully summarize our efforts to address your main concerns. 1. To show the trade-offs between robustness and efficiency, we design an accelerated Vaccine method. It is shown in the table in the rebuttal that this method has a better tradeoff between robustness and efficiency -- harmful score is increased by a marginal 0.60 while the training time can be shortened by 50%, compared to the original Vaccine. 2. To show that our method can be generalized to larger datasets, we show additional experiments on WikiText-2 with a larger number of tokens. The results indicate that Vaccine can reduce harmful scores while maintaining the model's perplexity. 3. We show further guidelines on selecting hyper-parameter $\rho$ by conducting experiments. The general trend is that with a larger $\rho$, the harmful score will decrease but at the cost of degrading finetune accuracy. 4. We compare another alignment-stage solution RepNoise in terms of GPU memory and training time. Our experimental results reflect that Vaccine is superior to RepNoise regarding both GPU memory and training time. We hope that these efforts can fully address your concerns, and we are more than happy to discuss them with you! --- Rebuttal Comment 2.1: Title: Warm reminder of author-reviewer discussion deadline Comment: Hi Reviewer a2cG, We sincerely thank you for the insightful review comments. As the deadline for the author-reviewer discussion is approaching (less than 8 hours), could you please have a check on our rebuttal? It would be nice if the rating could be slightly adjusted if you found our rebuttal can address your concern. Per your initial review, it seems that your main concern lies in the resource overhead of Vaccine. In the rebuttal, we provide more information on the resource overhead of Vaccine on different size of models. Our results indicate that the method requires 2x training time and a marginal extra GPU memory overhead (0.039GB). To further shorten the training time, we propose accelerated Vaccine, whose idea is to search for the perturbation not for every step, but to do it for every $\tau$ step. In this way, we can significantly shorten the training time. Our results show that the accelerated Vaccine is approximately 2x faster than the original Vaccine (approximately the same training time with SFT), while still achieving decent defense (less than 0.6% loss of harmful score). We are more than happy to discuss this and other concerns with you!	Rebuttal 1: Rebuttal: We sincerely thank all the reviewers and the AC for their efforts in reviewing our Vaccine paper! Per the initial review, all the reviewers hold a positive view of our paper. The compliments are too many to be counted, e.g., "(Vaccine) focuses on an interesting problem and have diverse evaluations" (Reveiwer a2cG), "have comprehensive empirical evidence, well-explained problem and methodology" (Reveiwer Pifs), "can be a plug-and-play method in different application scenarios" (Reveiwer CLDA), and "a work that may be interesting to the broad community" (Reviewer M5RT). However, we do observe that there are two important weaknesses mentioned by the reviewers that our paper is lacking: ----------- 1) Resource/Training time overhead. All four reviewers have concerns on the deployment of Vaccine because it needs to double the training time for safety alignment. This extra overhead is required because Vaccine needs to do two forward-backward passes for each optimization step. We value this comment and we generally believe that this extra resource overhead may hinder the scalability of the algorithm. Fortunately, we can present new results to show that it is actually not necessary to do two forward-backward passes for each optimization step! We present the accelerated Vaccine in Algorithm 2 in the attachment. The evaluation result of accelerated Vaccine is given in Table 10 in the attachment. ----------- 2) Lack of intuitive visualization of harmful embedding drift. Reviewer CLDA mentions that the harmful embedding drift phenomenon, though interesting, lacks an intuitive visualization. Reviewer a2cG seems also to be curious about the phenomenon. In response to this concern, we follow Reviewer CLDA's suggestion to use t-SNE to visualize the embedding. The visualization is available in Figure 5 in the attachment. The results indicate that the embedding drift of Vaccine is less serious, explaining its success in maintaining the alignment knowledge and resisting harmful fine-tuning. ----------- Other relevant concerns are also addressed by answer to each individual reviewer. Pdf: /pdf/c5c19d10caab454fc121ca10beac35d6692f7c5b.pdf	NeurIPS_2024_submissions_huggingface	2,024	null	null	null	null	null	null	null	null
Black-Box Forgetting	Accept (poster)	Summary: This paper explores a selective memorization problem of classification models under a black-box setup. The proposed method selectively applies different learning objectives, cross-entropy minimization, and entropy maximization for classes to be memorized and classes to be unmemorized, respectively; they become guidance for a small amount of text embedding tokens (prompt) by a black-box optimization algorithm, CMA-ES. The authors evaluate the proposed method in terms of selective memorization capability. Strengths: - The problem setup "black-box forgetting" that the authors focus on is crucial yet unexplored. The authors timely provide a simple approach to address the problem. - The design of the proposed method is easy to follow and reasonable. Weaknesses: # 1. Lack of novelty * The proposed method can be viewed as a variant of BBT [1], which explicitly separates the seed latent context into shared parts and unique parts. Although this is a new application of BBT in selective memorization scenarios, if the authors want to claim the novelty of the modified BBT method, I think they should further provide analyses on efficiency and qualitative results on learned context to contrast with BBT. * Moreover, the proposed learning objective -a combination of cross-entropy minimization and entropy maximization- is also quite common [2], so it is hard to claim its originality from this work. # 2. Inappropriate validation setup * As the main goal of this study is selective memorization of classification models, I expected existing methods for selective memorization and unlearning to be the natural baselines. However, all the comparisons are conducted by training the black-box tuning model via the combination of cross-entropy and entropy. Comparing the proposed method with other available unlearning objectives would be necessary to measure its validity. * The scope of validation is also limited to three natural object recognition. It is important to validate the methods across a more diverse range of image domains to robustify the conclusion. --- > Reference * [1] Black-Box Tuning for Language-Model-as-a-Service, Sun et al. 2022 * [2] Mitigating Information Leakage in Image Representations: A Maximum Entropy Approach, Roy and Boddeti 2019 Technical Quality: 3 Clarity: 2 Questions for Authors: - Is the sign of $L_{uniform}$ term in the final objective a typo? To maximize entropy, I think we should minimize $-L_{uniform}$ to maximize entropy. - For selective memorization, I believe that not only the unmemorization for the selected classes but also the memorization for the remaining classes is crucial. However, in Table 1 CUB-200-2011 results indicate that the white-box method only improves $Acc_{mem}$ 0.2 % from zero-shot, which is a very trivial amount given that the strong few-shot learning capability of CoOp [3] and CoCoOp [4] on the similarly fine-grain visual recognition tasks. Furthermore, the proposed method even hurts the $Acc_{mem}$ from the zero-shot model. Could you explain why the proposed learning objective shows poor learning capability on the CUB dataset compared to the CIFAR datasets? --- > Reference * [3] Learning to Prompt for Vision-Language Models, Zhou et al. 2022 * [4] Conditional Prompt Learning for Vision-Language Models, Zhou et al. 2022 Confidence: 4 Soundness: 3 Presentation: 2 Contribution: 2 Limitations: The authors adequately stated the limitation in the last section. Flag For Ethics Review: ['No ethics review needed.'] Rating: 5 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: ### Q1. Lack of novelty. Let us recap the main novelties of this paper. 1. We proposed a novel task called Black-Box Forgetting, which aims to achieve selective class forgetting under the assumption that the parameters and gradients of the model are inaccessible. 2. Aiming at improving derivative-free prompt tuning for Black-Box Forgetting, we proposed a novel context parametrization method called Latent Context Sharing (LCS) that explicitly models shared and unique components over multiple latent contexts. To our knowledge, there is no literature addressing Black-Box Forgetting, nor has the idea of explicitly modeling the shared and unique components of learnable contexts in prompt tuning been explored. Moreover, in all the experiments reported in our paper, we compared our LCS with BBT and showed that ours was superior to BBT. We would be happy to discuss these novelties in more detail if the reviewer could provide specific examples of publications that contradict these novelties. ### Q2. Loss function not novel. Note that we are not claiming novelty in the loss function. We argue that our novelty lies in Latent Context Sharing (LCS), a novel parametrization of the learnable latent contexts in prompt tuning for Black-Box Forgetting. ### Q3. Baselines are black-box tuning methods, not forgetting methods. As far as we know, there is no existing method for Black-Box Forgetting, i.e., there is no existing forgetting / unlearning method that can be directly compared. Therefore, we compared our method with the existing black-box tuning methods, i.e., BBT and CBBT, under the same loss function. We would appreciate it if the reviewer could provide any specific existing methods that are directly comparable or existing loss functions to be applied so that we can have clearer discussions. ### Q4. Datasets limited to three natural image datasets. Due to the nature of the pre-trained CLIP that are naturally biased toward natural images, the performance of zero-shot CLIP has been reported to be significantly degraded in non-natural image domains (e.g., [a,b]). Since our task is selective forgetting of classes that can successfully be recognized by the pre-trained model, it is difficult to test its effectiveness on non-natural image datasets. On the other hand, we agree that conducting experiments on more diverse datasets is beneficial, so we conducted additional experiments using ImageNet30. The results can be found in Table G in the attached PDF. Our method is superior to the other methods in terms of $H$, which supports the effectiveness of our method further. [a] Shu et al., Test-Time Prompt Tuning for Zero-Shot Generalization in Vision-Language Models, in Proc. NeurIPS, 2022. [b] Gondal et al., Domain Aligned CLIP for Few-shot Classification, in Proc. WACV, 2024. ### Q4. Typo in $L_{uniform}$. We will fix it in the final version. Thanks! ### Q5. CUB-200-2011 in Table 1, white-box CoOp only improves $Acc_\text{mem}$ by 0.2%, and the proposed method even hurts $Acc_\text{mem}$. We respectfully disagree with the reviewer's premise. Our problem is to achieve forgetting only specified classes while maintaining the accuracy of the other classes to be memorized, i.e., to improve $Err_\text{for}$ while maintaining $Acc_\text{mem}$. Achieving both of these is more challenging than merely improving $Acc_\text{mem}$ alone. To verify our argument here, we report an additional quick analysis in Table H in the attached PDF. "CoOp (White-Box w/ only memorization)" shows the results when white-box tuning by CoOp is performed to minimize the cross-entropy loss over only the classes to be memorized, i.e., forgetting is not performed. We can see a significant improvement in $Acc_\text{mem}$, as expected by the reviewer. This result proves that, even in a white-box setting, achieving both improving $Acc_\text{mem}$ and $Err_\text{for}$ is more difficult than merely improving $Acc_\text{mem}$ alone. Since the black-box setting is generally more challenging than the white-box setting, it is not at all surprising that our method leads to a slight degradation in $Acc_\text{mem}$. --- Rebuttal 2: Comment: I appreciate the authors' comprehensive responses! Most of my concerns are well-addressed. About the novelty and contribution, I think if the authors want to claim the novelty of this work on the problem setup itself, i.e., Black-Box Forgetting, this problem should be challenging so that naive application of the existing method is struggled with. For example, similar work on the Black-Box optimization setup [Oh et al. 2023] clarifies the problems of naive extension of the white-box methods [Bahng et al. 2022] and addresses those problems with their new method. However, in Table 2. of the submitted draft, BBT with a different configuration of CMA-ES, which is a naive extension of the existing method, already achieves very strong performance. This raises doubt about the difficulty of the considered problem setup, and given that, I still lean towards the negative side because the proposed method, LCS, lacks the necessity. Could the authors give further comment and opinion on this? --- - Oh et al. 2023, BlackVIP: Black-Box Visual Prompting for Robust Transfer Learning - Bahng et al. 2022, Exploring Visual Prompts for Adapting Large-Scale Models --- Rebuttal Comment 2.1: Title: Further response to the question of the novelty of the Black-Box Forgetting problem Comment: Thank you for taking the time to read through our rebuttal so promptly! We are very pleased to hear that most of your questions have been properly addressed. Thank you also for clarifying your question about the novelty of our problem, Black-Box Forgetting. However, we would like to argue that novelty of a problem should not necessarily be judged based on the difference in performance between existing and proposed methods. In fact, looking back at the history of machine learning, even if the difference in accuracy between existing and proposed solutions is less than 5%, there are many studies that have brought about significant advances in community by introducing novel problems. For example, few-shot object detection [a], universal domain adaptation [b], and open-set semi-supervised learning [c], just to name a few (You would agree that it is not possible to determine a golden rule like "If and only if there is more than $x$% difference between the proposed and existing methods, the problem will be considered novel."). We also want to argue that our Black-Box Forgetting is challenging. Besides Table 2 on CIFAR-10, below we report results for three other more challenging datasets. \| \| \|CIFAR-100\| \| \|CUB-200-2011\| \| \|ImageNet30\| \| \|:-------\|:-------:\|:-------:\|:-------:\|:-------:\|:-------:\|:-------:\|:-------:\|:-------:\|:-------:\| \|Method\|$H\uparrow$\|$Err_{\mathrm{for}}\uparrow$\|$Acc_{\mathrm{mem}}\uparrow$\|$H\uparrow$\|$Err_{\mathrm{for}}\uparrow$\|$Acc_{\mathrm{mem}}\uparrow$\|$H\uparrow$\|$Err_{\mathrm{for}}\uparrow$\|$Acc_{\mathrm{mem}}\uparrow$\| \|BBT\|$79.38_{\pm 0.01}$\|$87.30_{\pm 0.01}$\|$71.09_{\pm 0.00}$\|$58.75_{\pm 0.01}$\|$88.98_{\pm 0.04}$\|$43.85_{\pm 0.01}$\|$94.22_{\pm 0.05}$\|$90.17_{\pm 0.08}$\|$99.06_{\pm 0.01}$\| \|BBT w/ Sep-CMA-ES\|$69.29_{\pm 0.01}$\|$67.83_{\pm 0.02}$\|$70.81_{\pm 0.01}$\|$53.74_{\pm 0.02}$\|$74.72_{\pm 0.02}$\|$41.96_{\pm 0.02}$\|$91.18_{\pm 0.02}$\|$84.44_{\pm 0.04}$\|$\\textbf{99.07}_{\pm 0.00}$\| \|BBT w/ VkD-CMA-ES\|$75.41_{\pm 0.02}$\|$79.56_{\pm 0.01}$\|$\\textbf{71.67}_{\pm 0.01}$\|$55.12_{0.02}$\|$81.49_{\pm 0.03}$\|$41.65_{\pm 0.02}$\|$91.25_{\pm 0.05}$\|$84.58_{\pm 0.09}$\|$99.06_{\pm 0.00}$\| \|Ours\|$\\textbf{80.99}_{\pm 0.01}$\|$\\textbf{93.37}_{\pm 0.02}$\|$71.52_{\pm 0.01}$\|$\\textbf{59.67}_{\pm 0.01}$\|$\\textbf{89.29}_{\pm 0.01}$\|$\\textbf{44.81}_{\pm 0.01}$\|$\\textbf{97.28}_{\pm 0.01}$\|$\\textbf{95.94}_{\pm 0.01}$\|$98.67_{\pm 0.01}$\| As we can see, the variants of BBT do not provide stable performance over different datasets. In particular, the three BBT variants show $Err_\text{for}$ more than 5-10% lower than our method on ImageNet30. Furthermore, the results on CUB-200-2011 highlight the difficulty of the Black-Box Forgetting problem, as even our method yielding the best results could achieve only 59.67% in $H$. Moreover, as shown in Tables 1 and G, all the existing methods failed to achieve 80% in $Err_\text{for}$ for all the datasets. The performance degradation in $H$ is also significant, with $H$ values for BBT and CBBT being up to about 10% lower than for Ours. Figure 3 suggests that the scalability of BBT is poor, which is an inherent limitation that can be resolved by our method. Based on these facts, we believe that our Black-Box Forgetting is not an insignificant problem, but one that is both novel and worthy of technical consideration. [a] Kang et al., Few-shot Object Detection via Feature Reweighting, in Proc. ICCV, 2019. [b] You et al., Universal Domain Adaptation, in Proc. CVPR, 2019. [c] Yu et al., Multi-Task Curriculum Framework for Open-Set Semi-Supervised Learning, in Proc. ECCV, 2020. --- Rebuttal 3: Comment: The authors' statement on the novelty of problem setup is somewhat convincing, and I want to express my huge gratitude for the extensive responses. However, I would regretfully say that I want to keep my score. This black-box forgetting problem has not been explored yet, and the authors address it with their new method. The proposed method, latent context sharing (LCS), is designed to reduce the number of learnable parameters for better black-box optimization. I still do not see the necessity of this proposal, and there is a missing link in the connectivity between LCS and the black-box forgetting problem. To be specific, the LCS method is not tailored for "forgetting," and it can be regarded as a method for improving parameter efficiency that will be applied to other PEFT scenarios. The weak connection between the LCS method and the problem setup is the main reason that I kept my score towards rejection. --- Rebuttal Comment 3.1: Title: Response to Reviewer SKod's comment Comment: Thank you for your response. We are very happy that now Reviewer SKod acknowledged the novelty of the Black-Box Forgetting problem. Let us clear up the reviewer's misunderstanding: LCS is designed for forgetting. Please see Figure 5 in our original paper. The results show that as the ratio of the number of classes to be forgotten $r_\text{for}$ increases, the forgetting performance $Err_\text{for}$ of the existing parametrization method, BBT, tends to decrease. The only way to improve the $Err_\text{for}$ of BBT is to increase the number of learnable parameters $m$ (as can be seen in Figure 3), which clearly suggests the limitation of the scalability of BBT. Based on this observation, we developed our LCS to fundamentally improve the scalability of parameter representation for improving $Err_\text{for}$, that is, for forgetting. Note that the reviewer's comment is never a disadvantage of our LCS, even though it may suggest that LCS can be applied to other problems, besides forgetting. --- Rebuttal 4: Comment: I am so sorry for my decision-invariance attitude, but I want to adhere to my rating. These long discussions might cause your mind severe damage, and I am regretful for keeping my statement due to the unsolved concern. I can not agree with the authors' claim that "LCS is designed for forgetting." Figure 5 is just an empirical result, where the authors provide post-experiment interpretation based on the results. To be specific, it is only a consequential interpretation, and the design of the methodology cannot be justified as it is for forgetting. That is why I think the connection between the problem definition and the derived methodology is weak, and I will try to stick to this position. However, since I found some good implications from this study and other reviewers are looking positively, I will not be upset even if this paper is accepted. Sorry again, Reviewer SKod --- Rebuttal Comment 4.1: Comment: We would like to thank Reviewer SKod for your patience and your consideration during this rebuttal period. But could you please allow us to provide one last response, because we feel that the reviewers' view is somewhat biased, saying like "good methods should not be inspired by empirical observations." Looking at the facts of the past, there are a number of good methods built on empirical observations (and are not designed specific to the target tasks). For example, Focal Loss was originally proposed for object detection, but it is actually a loss to mitigate class imbalance, and its connection to object detection is rather weak [a]. ArcFace [b] was proposed for face verification, but its core is a variant of max margin loss, which is not really relevant to faces. Our method is designed based on sharp observations in forgetting, which in itself should not be punished. Furthermore, even if the reviewer could not agree that the LCS is designed for forgetting, at least could agree that it is designed for the Black-Box setting, which is the other half of the focus of this paper. [a] Lin et al., Focal loss for dense object detection, in Proc. ICCV, 2017. [b] Deng et al., ArcFace: Additive Angular Margin Loss for Deep Face Recognition, in Proc. CVPR, 2019	Summary: This paper studies the black-box forgetting problem. To this end, the authors optimize the input prompts with a proposed latent context sharing scheme by CMA-ES optimization. The authors achieve the selective forgetting goal by minimizing the cross-entropy loss on memorized classes and maximizing entropy loss on forgetting classes. Experiments demonstrate the promising of the proposed method. Strengths: The model forgetting/unlearning problem for black-box models is both interesting and practical. (However, the specific forgetting setting in this work is still unclear to me.) Compared with BBT [Sun et al.], the proposed Latent Context Sharing (LCS) scheme effectively enhances the learning ability of CMA-ES, providing a promising technique for CMA-ES-based black-box model fine-tuning. Sensitivity analyses regarding hyper-parameters are sufficient. Weaknesses: The problem setting of this work does not convince me, possibly due to the presentation of the paper. I hope the authors can provide more justification regarding this. Experiments could be further enhanced, as my detailed comments below. Technical Quality: 2 Clarity: 2 Questions for Authors: Q1. “—Retaining the classes that do not need to be recognized may decrease overall classification accuracy, as well as cause operational disadvantages such as the waste of computational resources and the risk of information leakage” How do you define the overall accuracy? Is it the overall accuracy of all classes of the pre-trained model, or the overall accuracy of downstream tasks, such as CIFAR-10 in the experiments? This is quite confusing for me. If it is the latter, during the fine-tuning process on downstream tasks, it seems that tuning only the necessary classes (those to be recognized) can achieve the best performance. Why must we forget some classes? Moreover, the phrase ‘waste of computational resources’ is also confusing. Learning on all classes is conducted during the pre-training phase and cannot be modified by the proposed method. Then, during fine-tuning, why would forgetting to reduce computational resources? Aside from preventing information leakage, what is the rationale behind forgetting certain classes? i.e., When we fine-tune a CLIP model on CIFAR-10, what is the real and sound motivation for us to forget some classes of CIFAR-10? Q2. If the final goal is to forget 40% of classes and memorize 60% of classes on CIFAR-10, why must we fine-tune CLIP? Why not directly train a new model (with the same architecture as CLIP or a smaller ResNet or ViT) with only the 60% memorized classes? It would be better to include the results of this baseline in Table 1. I assume that the authors believe a model fine-tuned from CLIP could achieve better performance on the memorized classes than a model trained from scratch. If this is the case, experiments on CIFAR datasets alone are insufficient to demonstrate this, as these data are relatively easy. Experiments on more complex datasets/tasks, such as those mentioned in autonomous driving, are necessary to further help to justify the superiority of the proposed method. Q3. Can the proposed method be extended to more complex tasks such as segmentation or object detection? As the authors mentioned, in an autonomous driving system, it is sufficient to recognize a limited number of classes such as cars, pedestrians, and traffic signs. In autonomous driving, the task is often more complex than pure classification, such as 2D and 3D object detection tasks. Q4. How about the performance of Ours vs. Solely Fine-tune CLIP on memorized classes? It would be much better if this result is verified on more complex datasets. Q5. In the white-box setting, one can update more model parameters beyond just the input prompts. How would the performance change if we update all model parameters for CoOp (White-Box)? I would like to increase my score if the authors could convince me regarding the problem setting. Confidence: 3 Soundness: 2 Presentation: 2 Contribution: 2 Limitations: The proposed method is not fully black-box, as it still requires access to the embeddings and output logits, which are unavailable in a fully black-box setting, similar to the GPT-4 API. However, I understand that achieving a fully black-box setting is very challenging, and I do not consider this a reason for rejection. I mention this point merely as a slight limitation. Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: ### Q1. Why must we forget some classes? We assume that the reviewer already acknowledged the benefit of preventing information leakage through forgetting. In addition to this, we here would like to emphasize the potential benefits of exploring selective forgetting. 1. Toward addressing the "Right to be Forgotten": If a service provider is asked to remove information so that their model cannot recognize certain information, it might need to comply with the request. This can be accomplished by training the model from scratch by removing samples of that class from the training data. However, retraining a large-scale model consumes an enormous amount of energy, which should be avoided. Selective forgetting may provide an efficient solution to this problem. 2. Toward efficient large-scale pre-trained models: Improving the efficiency of large-scale pre-trained models is of current interest to many researchers. Various attempts have been made such as model compression and architecture optimization (e.g., https://sites.google.com/view/elvm/program). Meanwhile, as the "scaling law" indicates, the reasonable size of a model correlates with the amount of knowledge it memorizes. Therefore, if the number of classes (amount of knowledge) that can be recognized by the model is limited, the model can be scaled down accordingly, thereby improving the efficiency of the model. This contributes to expanding the applicability of large-scale pre-trained models. 3. Toward better control over text-to-image generation: While diffusion-based text-to-image generation can generate diverse types of high-quality images, controlling the content of images remains a challenge. Recent research has focused on "forgetting" visual concepts in order to avoid generating undesirable content [a-c]. These methods forget certain classes by directly fine-tuning the diffusion model, but tuning the diffusion model itself is often costly. In contrast, given that typical generative models use a text encoder of a pre-trained VLM (e.g., CLIP) for conditioning, our method may provide an efficient approach to class forgetting by fine-tuning only the prompts of the text encoder. We believe that our method will open new directions for these important problems of interest to the ML community, even if these are not immediately feasible with this paper alone. We would like to add the discussion as a broad impact of this work in the final version. [a] Heng et al., Selective Amnesia: A Continual Learning Approach to Forgetting in Deep Generative Models, In Proc. NeurIPS, 2023. [b] Lu et al., Mace: Mass Concept Erasure in Diffusion Models, In Proc. CVPR, 2024. [c] Zhang et al., Forget-me-not: Learning to Forget in Text-to-Image Diffusion Models, In Proc. CVPR, 2024. ### Q2. Forgetting vs. Learning from scratch. As requested by the reviewer, we show the accuracy of ResNet-18 and ViT-B/16 trained from scratch over only the classes to be memorized in Table D in the attached PDF. As can be seen, both models cause severe overfitting to the training data. That is, while these models achieve reasonable accuracy on the training data, they exhibit severely poor performance on the test data. We also tested both models when initialized with ImageNet pretrained weights. While the results improve somewhat for ResNet-18, these are still far behind our forgetting-based method. The reason for this is that, following the common protocol in context optimization [a], we conducted our experiments in few-shot scenarios as explained in Line 201, which overwhelmingly lacks the number of training samples to learn the weights for even ResNet-18 and Vit-B/16. [a] Zhou et al., Learning to Prompt for Vision-Language Models, IJCV, 2022. ### Q3. Is the method applicable to detection / segmentation? It would be possible by combining with some detectors / segmentors. For example, recent approaches in object detection rely on a two-stage framework [a,b], which first uses region proposal network to get object proposals and then applies zero-shot CLIP to each proposal to identify the object class. Replacing the zero-shot CLIP with one tuned by our method would prevent detecting classes that have been forgotten. [a] Zhao, et al. Exploiting unlabeled data with vision and language models for object detection, in Proc. ECCV, 2022. [b] Zhao, et al. Taming Self-Training for Open-Vocabulary Object Detection, in Proc. CVPR. 2024. ### Q4. Ours vs. Solely Fine-tune CLIP on memorized classes? We tested the performance of Solely Fine-tune CLIP on memorized classes on four datasets including ImageNet30. The results are shown in Table E in the attached PDF. As we can see, ours and Solely Fine-tune CLIP are comparable in $ACC_\text{mem}$ for the three datasets, except for CUB-200-2011. That is, even if we perform forgetting, it does not significantly impair accuracy on the memorized classes. ### Q5. In the white-box setting, one can update more model parameters beyond just the input prompts. Yes, but it does not work. We tested "CoOp (White-box) + Parameter Update", i.e., updating not only the learnable contexts in the prompt but also the model parameters. The results are shown in Table F in the attached PDF. We see that simultaneously updating the model parameters does not improve performance, but rather hurts it. This is not surprising, as it is known that straightforward fine-tuning of the zero-shot CLIP does not improve performance [a]. [a] Wortsman et al., Robust Fine-Tuning of Zero-Shot Models, in Proc. CVPR, 2022. ### Q6. The proposed method still requires access to the embeddings and output logits, I do not consider this a reason for rejection. Thank you for this fair consideration of the limitation described in our original paper (Sec. 6). In fact, existing black-box tuning methods such as BBT and CBBT assume the same setup, which we followed. On the other hand, the consideration of a complete black-boxing approach is an interesting challenge that we would like to pursue. --- Rebuttal Comment 1.1: Title: Response to authors Comment: Thanks for the authors’ thoughtful response. With newly provided results, the paper becomes much stronger on the empirical side. The clarifications regarding the significance of Selective Forgetting on Generative Models make the problem setting more practical. Most of my initial concerns have been addressed. I shall increase my score and please include the new results and discussions/clarifications in the revised paper. Moreover, for results in Table E, when the pre-trained model’s performance is not good enough, solely fine-tuning on memorized classes still achieves much better Acc_{mem} than the proposed method, i.e., results on CUB-200-2011. This could be further discussed in the revision. --- Reply to Comment 1.1.1: Title: Thank you, and our response to your additional question on Table E. Comment: We would like to thank you for your great effort in carefully reading our rebuttal and for your sincere consideration of it. We are very happy that our response has successfully resolved your questions. Of course, we will make sure to include our new experimental results, discussions, and explanations we provided in our rebuttal in the final version. Regarding your additional question about the results in Table E, our response to Q6 of Reviewer VErt should be highly relevant. That is, Table C in the attached PDF shows that $Err_\text{for}$ and $Acc_\text{mem}$ are in a trade-off relationship, i.e., Ours ($Acc$ prio.), which is optimized by prioritizing $Acc_\text{mem}$, successfully improves $Acc_\text{mem}$, but at the expense of $Err_\text{for}$. This suggests that the underlying reason why the pre-trained model in Table E achieved higher $Acc_\text{mem}$ than Ours is probably because Ours was forced to sacrifice $Acc_\text{mem}$ to increase $Err_\text{for}$. Nevertheless, our method achieves a better trade-off (i.e., higher $H$) than the existing methods (BBT and CBBT) in all of the experiments. Furthermore, despite being a black-box method, Ours ($Acc$ Prio.) achieves $Acc_\text{mem}$ comparable to CoOp, a white-box method, which emphasizes the effectiveness of our method as a black-box optimization method. We will add a clear discussion about this point in the final version as well. Thank you for your valuable suggestion again!	Summary: This paper addresses the problem of selective forgetting of specified classes, which involves tuning a pre-trained model to reduce the classification accuracy for only the specified classes without affecting the accuracy for the others. The proposed model introduces a novel method for Black-Box Forgetting based on the derivative-free optimization of a learnable text prompt. It introduces Latent Context Sharing (LCS), a novel parameterization method for contexts, which mitigates the difficulty of high-dimensional optimization using derivative-free optimization. Strengths: 1- The proposed method explores the selective forgetting problem for CLIP-based models (PTMs), where the task is to make the model unable to recognize only the specified classes while maintaining accuracy for the others. More importantly, it addresses a novel problem of selective forgetting for black-box models, termed Black-Box Forgetting, and proposes an approach to solve this problem. 2- It proposes Latent Context Sharing, which introduces common low-dimensional latent components among multiple tokens for the prompt to reduce complexity. 3- Instead of optimizing the model parameters, it proposes a novel method for Black-Box Forgetting based on derivative-free optimization of a learnable text prompt, which avoids the need for information about the pretrained model. Weaknesses: 1- The proposed approach for forgetting selective classes from the pre-trained model requires data samples for each elective class. Since it is applied in CLIP-based multi-modals, where the text encoder is also available, one might wonder why data samples are necessary for each forgetting class. Could the same result not be achieved using just the class name? 2- The proposed approach claims that it introduces Latent Context Sharing (LCS), a novel parameterization method for contexts, to mitigate the difficulty of high-dimensional optimization. However, considering that the embedding dimension in CLIP-based models is 512, which is not excessively large, and that this approach reduces the dimension to 100 for the CUB dataset, the reduction does not seem significantly lower than the original CLIP dimension. Therefore, in the case of larger datasets, this reduction might not have a substantial impact compared to the original CLIP dimension. 3- The proposed approach claims to be a black box due to the unavailability of pre-trained model information, such as architecture, parameters, and gradients, during training. However, my concern is that if you are given the pretrained model, extracting information about the model architecture and its parameters can still be quite challenging? 4- How is the proposed black-box forgetting different from machine unlearning? 5- I recommend conducting experiments on more diverse datasets, such as ImageNet-1K, to gain a better understanding of this approach. 6- Why is Acc_mem lower than the baseline approaches for each dataset? 7- Can the proposed approach be applied in a continual learning setup? Technical Quality: 3 Clarity: 3 Questions for Authors: Please address all my concerns raised in the weaknesses section. Confidence: 4 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: My concerns raised in the weaknesses section described the limitations of the proposed approach. Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: ### Q1. Could the same result not be achieved using just the class name? Good suggestion! We tried to tune the latent contexts by only using the class names (i.e., class embeddings). Specifically, let $z_c$ and $z$ denote the class embeddings before and after prompt tuning for the class to be forgotten, respectively. Only $z$ is trainable. We aim to tune $z$ by minimizing the following negative NT-Xent loss, $ \log \frac{\exp(z_c ^\top z / \tau)}{\sum_i \exp(z_i ^\top z / \tau)}$, where $z_i$ is the class embedding for the $i$-th class to be kept memorized. This loss requires $z$ to be orthogonal to $z_c$ as well as be similar to the embeddings of the other classes $\\{z_i\\}$. The results are reported in Table B in the attached PDF. We found that Ours is much better than the approach using only the class embeddings (C-Emb.), which proves that tuning with only the class embeddings does not provide satisfactory performance. ### Q2. 512-D should not be high-dimensional. Derivative-free optimization based on multi-point search, such as CMA-ES, is directly affected by the curse of dimensionality, so the number of dimensions exceeding, typically 10-D, is often considered high-dimensional [a]. This is why the existing black-box tuning method, BBT, optimizes low-dimensional (around 10-D) latent contexts instead of the raw 512-D embeddings. [a] Hansen et al., Reducing the Time Complexity of the Derandomized Evolution Strategy with Covariance Matrix Adaptation (CMA-ES). Evolutionary Computation, Vol. 11, No. 1, pp. 1-18, 2003. ### Q3. Given the pretrained model, extracting information of the model should not be challenging. Following the same setup as in previous black-box tuning (e.g., BBT and CBBT), we assume that no pre-trained model is given and can only access to the input/output of the pre-trained model (e.g., via an API). In such a case, we cannot know the internal information of the model and cannot extract information about the model architecture or its parameters. ### Q4. Machine Unlearning vs. Black-Box Forgetting. These two are closely related but different. Machine Unlearning typically aims to remove the influence of specified training samples on the training model, whereas Black-Box Forgetting aims to prevent the recognition of specified classes. Furthermore, we in this paper address the black-box setting, which has not yet been well-traveled in machine unlearning. ### Q5. Experiments on more diverse datasets, such as ImageNet-1K? Although we were not able to complete the experiment with ImageNet-1K in this short rebuttal period, we instead report the results of additional experiments with its subset, ImageNet30 [a]. The results shown in Table G in the attached PDF demonstrate that our method outperforms all the methods in $H$. We will add the results on ImageNet-1K in the final version. [a] Hendrycks et al., Using Self-Supervised Learning Can Improve Model Robustness and Uncertainty, in Proc. NeurIPS, 2019. ### Q6. Why is $Acc_\text{mem}$ lower than the baselines? This is because $Err_\text{for}$ and $Acc_\text{mem}$ are in a trade-off relationship, with $Acc_\text{mem}$ tending to decrease as $Err_\text{for}$ is increased. This is presumably because features between classes are not completely disentangled in the feature space, so forgetting one class may negatively affect other classes (just as dog and cat share come common features). To provide justification for this, we report the results of using a loss more prioritized (weighted) for $Acc_\text{mem}$ in our method in Table C as "Ours ($Acc$ prio.)" in the attached PDF. We can see that Ours ($Acc$ prio.) outperforms all the other methods in $Acc_\text{mem}$, with sacrificing $Err_\text{for}$. Notably, both Ours and Ours ($Acc$ prio.) outperform BBT and CBBT in $H$, indicating that our method achieves a better trade-off than BBT and CBBT. ### Q7. Can the proposed approach be applied in a continual learning setup? Thank you for your interesting suggestion! While application of our method to continual learning is not in the scope of this paper, theoretically we think it is possible. Our method is based on a context optimization as CoOp. So it can readily be coupled with some of standard continual learning methods, such as distillation-based [a]. [a] Li et al., Learning without Forgetting, IEEE TPAMI, Vol., 40, No.12, pp. 2935-2947, 2017.	Summary: The authors propose to apply selective forgetting to black-box pretrained models, instead of the usual white-box settings. Since it is a black-box method, there is no parameter update and instead the prompts are the ones being optimized to decrease the performance on the target class to be forgotten. This is done by using an existing covariance matrix adaptation, which is a derivative-free optimizer. The paper describes the issues of optimizing a large dimensional space while gradients are not available, and the problem of having samples from the target forgotten class being mostly pushed towards a close-by, non-forgotten class instead. The first issue is solved by splitting the latent context between unique and shared, and reparametrizing them individually to lower dimensions. The second issue is solved by excluding the label information from a classic CE loss, and adding it as regularization for maximizing the entropy of the overall confidence. Strengths: Selective forgetting and machine unlearning are of interest within our community, and thus this paper fits very well with the conference. The need to move from white-box to black-box methods is something natural, as it has happened before with topics like adversarial attacks, so it is relevant to explore how to tackle this more complex setting. In terms of originality, previous work is well referenced to establish the originality of the paper comes from the proposed strategy adapting to the needs of the proposed novel scenario. The introduction of the scenario and its needs, as well as most of the method explanations are well described and are easy to follow. The limitations are very well covered, and the significance of the paper is well motivated and clear. Weaknesses: Despite the main idea being clear and most of the method being well described, the writing could use some revision. CMA-ES is quite a key component of the proposed strategy, but is covered quite shortly. The LCS (ii) part is a bit over-complicated for such an easy concept. I think it could be explained a bit more elegantly and would highlight the usefulness and easiness of having this type of parametrization. Same goes for some of the subsections within the Analysis (4.3). I assume that the lack of space was the reason why some of the concepts from the compared methods are just mentioned and not explained in detail. However, in order to make it easier for the reader to understand what the analysis refers to and drive some of the conclusions, it might be better to just move one or two of them to the appendix. My suggestion would be moving 4.3.1 and 4.3.5 to the appendix, and using that space to improve the analysis of the others. Nomenclature in some cases is a bit confusing. In figure 1, the losses are called forget and memorize, while in equations (1) and (2missing) they are called uniform and CE. Same with z having subscripts or superscripts depending on BBT or LCS. Text in figures 3 and 4 is too small. Technical Quality: 3 Clarity: 2 Questions for Authors: 1. In the introduction, it is shortly mentioned, but I would like some more context on how close are some of the machine unlearning settings to the one proposed in this submission. 2. Footnote 2 is a bit weak. How are the results shown in section 4 enough reason to assume that there is a composition of ULC and SLC? 3. Why are the latent contexts optimized for the hyperparameters used? Are those decided under a validation set or overfitted to the test? Same for the number of latent contexts. 4. In appendix A.2, what is the reasoning for the groupings of classes to be forgotten? Just random? Or do they relate in some way that can support the use of equation (2missing) versus (1)? Confidence: 4 Soundness: 3 Presentation: 2 Contribution: 3 Limitations: Regarding the experimental statistical significance, the authors mention that they do not report error bars, and I would also consider that such set of experimental settings would require the need to run each method multiple times, specially when there are some results that are quite comparable and could lead to statistical significance showing that some methods is not significantly different than another under some metrics. This is such a strong point, in my opinion, that it is the reason why I rate the paper down from a 7 (accept) to a 6 (weak accept). Given that the paper relies quite heavily on its experimental analysis to showcase the interesting proposed scenario compared to current zero-shot or white-box approaches. Also in the appendix, there is this random sampling, which does not report how many times it is sampled either. Finlly, the point from question 2 is also important to showcase the limitation on the argumentation of the existence of ULC and SLC. Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: ### Q1. CMA-ES is covered quite shortly. We will expand the description of CMA-ES in the first paragraph of Line 105 in Sec. 3 as follows: > We employ CMA-ES, a widely used evolutionary algorithm for black-box optimization in continuous, because a textual prompt to be optimized is a continuous variable. CMA-ES is a multi-point search algorithm based on a multivariate normal distribution and proceeds the search by iterating (i) sampling candidate solutions, (ii) evaluating the loss values of the candidates, (iii) weighting the candidates based on the loss values, and (iv) updating the mean and covariance matrix of the distribution by using the weighted candidates. Due to the nature of multi-point search, the performance of CMA-ES degrades in high-dimensional problems, typically ten or more dimensions [a,b]. While several extensions have been proposed, e.g., [a,b], these methods require knowledge of independence among variables, which is not always known. In this paper, we propose a customized extension of CMA-ES to Black-Box Forgetting. [a] Ros and Hansen, A Simple Modification in CMA-ES Achieving Linear Time and Space Complexity, in Proc. PPSN, 2008. [b] Akimoto and Hansen, Projection-based Restricted Covariance Matrix Adaptation for High Dimension. in Proc. GECCO, 2016. ### Q2. LCS could be explained a bit more elegantly. To facilitate intuitive understanding of our LCS, we add at the beginning of the LCS section (Line 139) the inspiration behind our LCS and the justification for the composition of ULC and SLC. > Fig. 2c shows the overview of LCS. The key idea is to assume shared parameters among different latent contexts. This inspiration comes from successful word embedding methods; most word embedding methods are trained on the assumption that locally co-occurring words have semantic correlations between them (e.g., [a-c]). This inspires the idea of explicitly modeling semantic correlations between words in a prompt as shared components. [a] Mikolov, Efficient Estimation of Word Representations in Vector Space, arXiv pre-print 1301.3781, 2013. [b] Pennington, et al., GloVe: Global Vectors for Word Representation, in Proc. EMNLP, 2014. [c] Devlin et al., BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding, in Proc. NAACL-HLT, 2019. ### Q3. Sec. 4.3 could be reorganized. As suggested, we will move 4.3.1 and 4.3.5 to the appendix to put priority to the analyses on the effectiveness of the parametrization of our LCS. Thanks! ### Q4. Some nomenclature and texts in figures and equations could be improved. As suggested, we will revise the numbering of Eq. 2 and the name of the loss (i.e., change "CE" in Fig. 1 to "uniform") in the final version. The text in Fig. 1 will be enlarged. The superscript of $z$ is to distinguish shared components from unique components, so is essential for our LCS. ### Q5. Some more contexts on machine unlearning? We will add a review of machine unlearning in the Related Work section in the final version: > Machine unlearning aims to remove an arbitrary sample from a pre-trained model, i.e., obtaining a model that is identical to the one trained from scratch without that sample [a-g]. Many methods have been proposed, for example, to construct a forgettable model by transforming the learning algorithm into a sum of the training samples [a], to achieve forgetting by linear approximation of a nonlinear model [b], and to update the model to be closer to / farther from the original model in the retain / forget samples [e]. Methods specific to certain learning algorithms such as LDA [f] and SVM [g] have also been explored. Machine unlearning and Black-Box Forgetting are closely related but different; Machine unlearning aims to remove the influence of specified training samples on the training model, whereas Black-Box Forgetting aims to prevent the recognition of specified classes. Forgetting specified classes has attracted much attention recently in various contexts [h-l]. We in this paper address the black-box setting, which has not yet been explored. [a] Cao and Yang, Towards Making Systems Forget with Machine Unlearning. In Proc.IEEE Symp. Security and Privacy, 2015. [b] Golatkar et al., Mixed-Privacy Forgetting in Deep Networks, In Proc. CVPR, 2021. [c] Sekhari et al., Remember What You Want to Forget: Algorithms for Machine Unlearning, In Proc. NeurIPS, 2021. [d] Bourtoule et al., Machine Unlearning, In Proc. IEEE Symp. Security and Privacy, 2021. [e] Kurmanji et al., Towards Unbounded Machine Unlearning, In Proc. NeurIPS, 2024. [f] Guo et al., Certified Data Removal from Machine Learning Models, In Proc. ICML, 2020. [g] Chen et al., A Novel Online Incremental and Decremental Learning Algorithm based on Variable Support Vector Machine, Cluster Computing, 2019. [h] Heng et al., Selective Amnesia: A Continual Learning Approach to Forgetting in Deep Generative Models, In Proc. NeurIPS, 2023. [i] Lu et al., Mace: Mass Concept Erasure in Diffusion Models, In Proc. CVPR, 2024. [j] Zhang et al., Forget-me-not: Learning to Forget in Text-to-Image Diffusion Models, In Proc. CVPR, 2024. [k] Shibata et al., Learning with Selective Forgetting, in Proc. IJCAI, 2021. [l] Ye et al., Learning with Recoverable Forgetting, in Proc. ECCV, 2022. ### Q6: Hyperparameters tuned on validation sets or test sets? All the hyperparameters were tuned on validation sets. We will clarify this in the final version. ### Q7: Classes to be forgotten determined at random or intentionally? These were determined randomly for fairness. We will clarify this in the final version. ### Q8: No error bars reported. Good suggestion! All the results in our paper are averages of three runs with different random seeds. Check out Table A in the attached PDF, which is Table 1 of our submission with the standard deviations added. Ours clearly outperforms the compared methods. We will add standard deviations to all the results in the final version. --- Rebuttal Comment 1.1: Comment: Thanks for the author's response. I agree that answers to Q1-Q5 will help improve the manuscript. Specialy, the context on machine unlearning, and the differences from the proposed black-box forgetting scenario are relevant, and the proposed changes on literature review are well covered. However, reading through the other reviewer's and response, I would stress the need to highlight well those differences. (a) As further discussion, when saying "_Machine unlearning aims to remove the influence of specified training samples on the training model, whereas Black-Box Forgetting aims to prevent the recognition of specified classes._", it raises the question of how that affects the evaluation. If the shift between the two problems is going from specific samples to whole classes, how well do the metrics proposed evaluate the new scenario? The metrics used on the paper are based on the accuracy over specific samples, instead of measuring the shift between using the learned latent contexts and not. I understand that the metrics on single samples are aggregated, I am referring towards having a better metric to measure that the distribution of the specified class is forgotten. (b) Q6, Q7 and Q8 are satisfactory responses for the experimental setting and evaluation. I would support adding those details to the final version, since the scenario should be clearly stated for further reproducibility. I appreciate that the results are over more than one seed, and I think the low standard deviation is a positive sign for the conclusions of the experimental section and reinforce the conclusions. (c) My original question 2 is unaddressed. The paper states "_We assume that each latent context is composed of unique components (Unique Latent Contexts (ULC)) and common components (Shared Latent Context (SLC)) among multiple latent contexts_" with a footnote saying that the assumption is true due to the experiments in Sec. 4. However, this seems like a very weak argument. How did this assumption came to be, or what supports it, even before having an experimental support for it? (d) Reading through some reviews and rebuttals, the author's have also clarified some of the interesting questions raised by fellow reviewers, which in most cases are satisfying. Answer to Q1 of reviewer VErt is not very convincing. If the proposed scenario distinguishes itself from machine unlearning by moving away from specific samples, the idea of using the prompts becomes much more interesting, instead of keeping samples. The authors address on their response the performance improvements, but the interesting part of the question is not the numbers, but challenging the decision of using the given samples vs. sampling the prompts for the classes involved. --- Reply to Comment 1.1.1: Title: Our responses to your additional questions (b) and (c). Comment: Thank you for your detailed review not only of our responses, but also of the other reviewers' comments and our responses to them. We are very pleased to hear that most of our responses were satisfactory. Below we would like to respond to your additional questions, first to (b) and (c), followed by (a) and (d). To (b): Yes, we will make sure to add the results and discussions, as well as the detailed evaluation protocol in the final version. To (c): Sorry for the lack of a detailed explanation. The following statement, which is originally provided as our response to question 2, is the inspiration behind our design. Please let us know if it is still not convincing to you. > The key idea is to assume shared parameters among different latent contexts. This inspiration comes from successful word embedding methods; most word embedding methods are trained on the assumption that locally co-occurring words have semantic correlations between them (e.g., [a-c]). This inspires the idea of explicitly modeling semantic correlations between words in a prompt as shared components. [a] Mikolov et al., Efficient Estimation of Word Representations in Vector Space, arXiv pre-print 1301.3781, 2013. [b] Pennington et al., GloVe: Global Vectors for Word Representation, in Proc. EMNLP, 2014. [c] Devlin et al., BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding, in Proc. NAACL-HLT, 2019. --- Rebuttal 2: Comment: __following on (a) and (d)__: yes, that is what I meant. I also agree that it is a larger question than what the paper covers currently. I appreciate the results provided and the thorough explanation with the new results. I agree that the insights from the discussion are indeed of interest for further research! Overall, my questions have been answered. Thanks to the authors for their thorough responses. Some clarifications within the discussion would be needed for the final version, but I have no further questions. I think the questions from the other reviewers and the author responses to them are also satisfactory. Some of them highlight some of the limitations of the proposed setting and also should be included in a final version. The only comment from a reviewer that I also consider important to note, is how simple the datasets used are (Q4 from SKod), and agree with the observation. The results on ImageNet-30 by the authors are good step to solve that issue, although I would be cautious of results obtained during a rebuttal (not due to suspicion of malice, but just due to the fast implementation). However, I am satisfied with the experiments provided. Therefore, I would be in favour of accepting the paper. --- Rebuttal Comment 2.1: Title: Additional responses. Comment: Thank you again for your feedback. We are pleased to hear that we were able to properly address your questions and that you are in favor of accepting our paper. To (c): Thank you for drawing our attention to this interesting point. We will try to put more detailed discussions in the final version, but our preliminary idea in the case of pre-trained VLMs at present is as follows. Contrastive language-image pretraining learns to align image and text features in a common space using a massive dataset. Thus, the common and unique components of the text features are expected to be aligned with the image features as well, even if the modalities are different. Empirically, image-specific "levels" seem to be aligned with the text features, at least partially, but surprisingly very well. For example, it has been shown that the pre-trained CLIP can recognize an object in a local region of an image, suggesting that the text features are learned to absorb differences in scale [a]. CLIP guidance can produce an image that is faithful to the prompt by moving the image feature in the direction of the prompt embedding [b]. These results may support the earlier expectation and may also be a reason that effective forgetting is possible for pre-trained VLMs. To (a) and (d): We will make sure to add the results on ImageNet-1K in the final version. Thank you. [a] Miyai et al., LoCoOp: Few-Shot Out-of-Distribution Detection via Prompt Learning, in Proc. NeurIPS, 2023. [b] Nichol et al., GLIDE: Towards Photorealistic Image Generation and Editing with Text-Guided Diffusion Models, in Proc. ICML, 2022.	Rebuttal 1: Rebuttal: We thank all the reviewers for their thoughtful review and constructive feedback. We are happy to see that the reviewers acknowledged the major contributions of this paper. Namely, 1. We proposed a novel task called Black-Box Forgetting, which aims to achieve selective class forgetting under the assumption that the parameters and gradients of the model are inaccessible. 2. Aiming at improving derivative-free prompt tuning for Black-Box Forgetting, we proposed a novel context parametrization method called Latent Context Sharing (LCS) that explicitly models shared and unique components over multiple latent contexts. The main questions raised by the reviewers centered on 1) the difference between our task, i.e., Black-Box Forgetting, and typical machine learning, and 2) the requests for additional analyses to further support the validity of our method. In the rebuttal, we addressed all of the reviewers' questions with the support of thorough additional experimental results. We will implement all of these responses in our final version. For the additional experimental results, please see the PDF attached to this thread. ### Black-Box Forgetting vs. Machine Unlearning. These two are closely related but different. Machine Unlearning typically aims to remove the influence of specified training samples on the training model, whereas Black-Box Forgetting aims to prevent the recognition of specified classes. Furthermore, we in this paper address the black-box setting, which has not yet been well-traveled in machine unlearning. To explain this, we will add a brief review of Machine Unlearning in the Related Work section in the final version. ### Additional analyses to support the validity of our method We additionally conducted through experiments and report the new results. Namely, evaluatin of error bars, comparisons with fine-tuned CLIPs, comparisons with ResNet and ViT trained from scratch, and evaluations on an additional dataset (ImageNet30), and so on. For more information, please see Tables A-H in the attached PDF and our responses to each reviewer's questions. Pdf: /pdf/d3fe917e548585e7cb17add67d9885e011049782.pdf	NeurIPS_2024_submissions_huggingface	2,024	null	null	null	null	null	null	null	null
Be Confident in What You Know: Bayesian Parameter Efficient Fine-Tuning of Vision Foundation Models	Accept (poster)	Summary: This paper focuses on the parameter efficient fine-tuning (PEFT) for foundation models in the few-shot learning settings. The authors reveal that the adapted models lack accurate fine-grained uncertainty quantification capabilities. Specifically, the few-shot tuned models perform remarkable accuracy but with large expected calibration errors. To address this problem, they propose a lightweight Bayesian PEFT framework by designing two Bayesian components to improve models' confidence. Theoretical analysis and experiments on 4 visual benchmarks prove the effectiveness of the proposed method. Strengths: 1.The paper is well-written and the storyline is clear. 2.The proposed Bayesian-PEFT is simple-and-effective, and theoretical analysis in the paper can reveal the underlying mechanism of the method. 3.Extensive experiments on ViT in several few-shot settings show the effectiveness of Bayesian-PEFT. Weaknesses: 1.The title may be a bit inappropriate. If the Bayesian-PEFT is claimed to be used for foundation models, experiments on more types of foundation models are required. For example, I am curious whether it still works well on large language models (LLMs), e.g., LLaMA3. Also, it is unclear whether the under-confidence issue observed in the ViT also exists in LLMs. 2.Although it may be beyond the scope of the paper's research, I am still curious whether the under-confidence issue studied in the paper disappears when there is sufficient training data. 3.It would be better if a schematic diagram of the proposed Bayesian-PEFT could be provided. Technical Quality: 3 Clarity: 3 Questions for Authors: Does the proposed Bayesian-PEFT work well in LLMs, e.g., LLaMA3? Confidence: 4 Soundness: 3 Presentation: 3 Contribution: 2 Limitations: Yes Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thank you for taking the time to review our paper. We appreciate your constructive comments and suggestions. Here are the responses to the clarifying questions: Q1: LLM Clarification The proposed ideas in this work could potentially be extended to PEFT of LLM models for some downstream tasks. However, it is worth noting that the uncertainty quantification for LLM is challenging due to their generative nature, and is an open question. For instance, LLAMA3 is a generative model, and uncertainty quantification of the overall generated text response for an input query is difficult to interpret. In this work, we focus on PEFT of vision foundation models, where we identify the under-confidence issue for all 4 PEFT techniques considered. We introduce a novel evidence-guided method to address this challenge. We will update the title from foundation models to vision foundation models to make it more accurate. Q2: Under-confidence with sufficient data Please refer to the answer to Q1 in the general response for the results related to PEFT with sufficient data. Q3: Schematic Diagram of B-PEFT We present the schematic diagram of B-PEFT in Figure 3 of the attached pdf of the general response. We will include the schematic diagram in the revised draft. --- Rebuttal Comment 1.1: Title: Official Comment by Authors Comment: Dear Reviewer CbTK, We’d like to express our gratitude again towards the reviewer for evaluating the proposed work and providing constructive comments. In our rebuttal, - We analyze under-confidence behavior with varying sizes of data. - We provide the schematic diagram of B-PEFT method. - We discuss the potential extension to LLM models. We are happy to answer any further questions that you may have regarding our work.	Summary: This submission presented a lightweight Bayesian Parameter Efficient Fine-Tuning (Bayesian-PEFT) framework for large transformer-based foundation models. Experiments across diverse datasets demonstrated the improved calibration performance by Bayesian-PEFT on multiple other PEFT techniques. Strengths: * The paper is well written and the experiments are extensive. * The proposed transformation function $A_m$ is well motivated and the experiments supports authors' claim. Weaknesses: * The dependence on Evidential Deep Learning is quite significant. To some extent, the proposed method extends EDL, rather than being designed specifically for PEFT of foundation models. * Some editorial errors (e.g., "we we") exists, the submission may need further proofreading. Technical Quality: 3 Clarity: 3 Questions for Authors: * In Eq. (6), would there be a numerical stability problem when dividing e_{min}, as this value can be very small? Confidence: 2 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: Yes. Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thank you for taking the time to review our paper. We appreciate your constructive comments and suggestions. Here are the responses to the clarifying questions: Q1: The dependence on Evidential Deep Learning is quite significant. To some extent, the proposed method extends EDL, rather than being designed specifically for PEFT of foundation models. Please refer to the answer to Q2 in the general response, which makes it clear that the proposed approach is specifically designed to address the unique under-confidence behavior of PEFT methods. We extend evidential learning in novel ways by performing fine-grained uncertainty decomposition and base rate adjustment, which allows us to seamlessly integrate it into parameter efficient fine-tuning paradigm. Such an integration has the potential to significantly broaden the usage of foundation models in many critical domains by fixing their calibration issues. Q2: Typos and minor mistakes We will proofread the draft, and correct typos/minor mistakes in the updated draft. Q3: Numerical Stability Issue in Eqn. 6 Yes, since the evidence is in the range [0, infinity] (the evidence is obtained by the exponential of the logit outputs), the stability issue can arise in extreme cases when the logit output is extremely low (i.e. close to negative infinity). In our experiments, we did not observe the stability issue. Still, the issue can arise in some extreme cases, and to address this challenge, we could introduce a small delta in the denominator or bound the network’s logits to be greater than a small negative value. --- Rebuttal Comment 1.1: Title: Official Comment by Authors Comment: Dear Reviewer YJfL, We’d like to express our gratitude again towards the reviewer for evaluating the proposed work and providing constructive comments. In our rebuttal, - We emphasize the novelty of the proposed method to address the unique under-confidence behavior of the PEFT methods. - We clarify the numerical instability issue that can arise during experiments and how to tackle the issue. We are happy to answer any further questions that you may have regarding our work.	Summary: This paper studies the problem of parameter efficient fine-tuning. The authors pointed out mis-calibration issues caused by parameter efficient fine-tuning, which is, more specifically, under-confident estimation when fine-tuning data is limited. To solve this issue, the author proposed Bayesian PEFT, a Bayesian parameter efficient fine-tuning method that introduces two components, i.e., a base rate adjustment to strengthen the prior belief of pre-trained knowledge, and a evidential ensemble with diverse ensemble components. The authors provided both theoretical justification as well as extensive experiments across diverse datasets. Strengths: 1. This paper is well written and easy to follow. 2. The paper discusses an interesting observation on low confidence in few-shot setup for PEFT. 3. The proposed method seems to be reasonable and solve the identified problem. Weaknesses: 1. It would be interesting to see more in depth discussion and empirical analysis on the observation of uncertainty of PEFT, particularly, related to data size and number of unfrozen parameters. 2. More discussion related to OOD generalization might be needed. Particularly, how does the proposed method help OOD generalization in terms of prediction accuracy. Technical Quality: 3 Clarity: 3 Questions for Authors: 1. Is fully fine-tuned model also have low uncertainty compared to PEFT model? 2. Will the proposed method improve robustness such as OOD generalization in terms of model accuracy? 3. How does the proposed approach work on other modalities, such as audio and language? Confidence: 3 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: Yes. Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thank you for taking the time to review our paper. We appreciate your constructive comments and suggestions. Here are the responses to the clarifying questions: Q1: It would be interesting to see more in depth discussion and empirical analysis on the observation of uncertainty of PEFT, particularly, related to data size and number of unfrozen parameters. Please see the answer to Q1 in the overall response. Q2: Clarification on OOD performance The current work, being an instance of fine-grained uncertainty quantification works, can help in OOD detection. We further present the OOD performance of our model in the answer to Q3 of the overall response, that shows the potential of our model in handling OOD. Q3: Is fully fine-tuned model also have low uncertainty compared to PEFT model? Full fine-tuning in few-shot learning tasks leads to overconfident models (see Table 5/Figure 1(a) in the attached pdf) and, due to overfitting, also hurts the model's generalization performance. Q4: How does the proposed approach work on other modalities, such as audio and language? In this work, we focus on few-shot image classification task and extension to other modalities is beyond the scope of this work. It could be interesting follow-up to study the calibration performance of PEFT for audio/language foundation models. For instance, the ideas developed in this work could potentially be used in situations where the PEFT leads to miscalibrated models and the developed model requires trustworthy fine-grained uncertainty quantification capabilities. If these data modalities are also modeled using transformers and follow the parameter efficient fine-tuning paradigm in performing downstream tasks, we expect the proposed approach can benefit them in a similar way. --- Rebuttal Comment 1.1: Title: Official Comment by Authors Comment: Dear Reviewer DMeg, We’d like to express our gratitude again towards the reviewer for evaluating the proposed work and providing constructive comments. In our rebuttal, - We conduct an in-depth analysis on the under-confidence behavior of PEFT methods w.r.t number of classes, data size, and the number of unfrozen parameters. - We fully fine-tune the model to analyse the calibration behavior. - We discuss the possible extension of the proposed method to other modalities, such as audio and language. We are happy to answer any further questions that you may have regarding our work.	Summary: This paper advances the Bayesian Parameter Efficient Fine-Tuning (Bayesian-PEFT) approach: a strategy to fine-tune large pre-trained foundation models with well calibrated classifier and ability to gracefully deal with out-of-distribution (OOD) data thanks to uncertainty quantification from ensembling. First the authors study the behavior of a fine-tuned foundation model with several PEFT methods (bias fine-tuning, adapters, side-tuning, visual prompt-tuning) in the few-shot learning regime (defined here as reducing the number of training samples per class, but keeping the original number of classes) and show that even though they achieve good accuracy, they all display under-confident predictions. The authors attempt to trace back the source of this behavior with evidential learning tools and conclude that the models are accurate thanks to relatively greater evidence of the correct class compared to other classes, and that models are under-confident because all classes are being assigned very low evidence, including the correct ones. Inspired by this finding, the authors devise an evidential learning based PEFT strategy that adjusts the base rates of the evidential head such that prior belief from pre-training knowledge is strengthened while maintaining accuracy. Secondly the authors propose an evidential ensemble with a technique to induce diversity with different incorrect belief regularization strengths that penalize the model differently when assigning high beliefs to non-correct classes. The proposed Bayesian-PEFT is evaluated on 4 datasets (CIFAR10, CIFAR100, Food101, Flowers102) under different few-shot regimes, showing boosted accuracy and calibration. Strengths: ### Significance - this work deals with an important problem in the current landscape of machine learning: how to leverage existing foundation models to downstream tasks with few labels while ensuring their reliability ### Originality - I find the use of evidential learning for this setting to be original. The idea of preserving prior knowledge and increasing the evidence for all classes with evidential learning for this use-case is novel - the strategy of inducing diversity of ensembles with different strengths for incorrect evidence regularization is novel for the evidential learning literature and has the benefit of not needing communication between ensemble particles during training -> improved computational efficiency ### Clarity - the paper is overall clearly written conveying a well argued story about the behavior observed for different PEFT methods in the few-shot regime, then studying it with evidential learning lens and then proposing a few solutions to mitigate the encountered non-desirable behaviors - the code comes to support information potentially missing from the paper ### Quality - the authors conduct numerous experiments under several few-shot learning regimes, showing improved performances. They also conduct several ablations and sensitivity studies towards better understanding the impact of the building blocks of this method and the choices made for them - code is provided in the supplementary for reproducibility of results - the authors study their work on several PEFT methods from computer vision literature, mostly VPT in the main paper and the others in supplementary Weaknesses: ### Framing of few-shot learning task - the few-shot learning setting considered here is quite different from the one considered in most of the few-shot learning/meta-learning literature with many small sets of support images and classes and query images (sampled from miniImageNet, Omniglot, CIFAR, etc.), sometimes called _"episodes"_. In contrast the proposed few-shot learning setting proposed here keeps the original number of classes and reduces the number of samples to 1-20 samples per class. This ends up as a different setting as the model is finetuned on 200 to 4000 images for CIFAR100 for instance from the 1-shot to the 20-shot setting, which is not a big dataset but still not that few images to train on. - I'm wondering whether the proposed setup of having many classes, but few samples, may be the source of the under-confidence behavior of PEFT methods for the few-shot learning setup that is proposed here. In such cases, simple solutions from the few-shot learning literature that do not require training: e.g., cosine classifier [b], [c] (i.e., averaging features from the same class, do dot-product with query features, then softmax on similarities) - the authors could refer to this setting that they propose by a different name to avoid confusion with few-shot learning where models can be finetuned with very few samples and there are several Bayesian methods there already. A potential alternative could be: low-shot regime. ### Motivation for evidential learning approach and few baselines - the formalism based on evidential learning is interesting, but is not always trivial to implement and previous works [d],[e] reported difficulties in scaling evidential methods and the Dirichlet predictor to larger number of classes and encoders. - while the approach itself is clearly described, the motivation for using such a line of approach instead of others that are more simple, generic and with a wider adoption, e.g., last-layer methods like Laplace-Redux [f], Bayesian last layer [g]. Since the endeavor of this work is to look at more practical settings with few labeled samples and reliability constraints, it may make sense to compare this work with some simple baselines to show the benefits of the approach and its trade-offs. Other potentially relevant baselines would be the cosine classifier from few-shot learning, LoRA ensembles [h], linear probing with test-time augmentation [i] - It would be also interesting to expand the list of PEFT methods studied with more recent and potentially more parameter efficient ones like LoRA [j] or VeRA [k] that have been shown useful to large transformer networks and could expand the scope of this work beyond image classification. ### Missing related works - as mentioned above, the field of few-shot learning/meta-learning has seen several Bayesian methods with uncertainty quantification. Often these meta-learning methods fine-tune the network at test-time with just a few labeled samples (from classes not seen during pre-training). Here is a non-exhaustive list of relevant works from this area: PLATIPUS [m], VERSA [n], LLAMA [o], Bayesian MAML [p], Bayesian TAML [q], etc. - a discussion on how the current B-PEFT is different from them and what type of problems it addressed and with which benefits would be useful here. - the idea of converting a pre-trained network into a Bayesian Neural Network and training over a short period of a few epochs (not a low-shot setting though) has been recently proposed in ABNN [l]. The diversity-inducing mechanism with different incorrect belief regularization strengths can also be related with the random prior mechanism from ABNN [l] where an ensemble of networks was obtained from a pre-trained network and the diversity was induced by introducing a different class-level prior for each ensemble member. ABNN is pretty recent so it can be considered as concurrent work. ### Misc. - the OOD experiment could benefit from the inclusion of typical metrics used in literature (e.g., AUROC, FPR95, AUPR, etc.) to allow easier comparison and putting the performance in context of other works. ### Minor - Novelty - the evidential learning choice seems to be inspired to some extent by the multidimensional belief quantification work [37] that used uncertainty for vanilla few-shot learning / meta-learning. The authors however expand the idea further with belief strenghtening and incorrect belief regularization for ensembling. References: [a] Vinyals et al., Matching networks for one shot learning, NeurIPS 2016 [b] Gidaris et al., Dynamic Few-Shot Visual Learning without Forgetting, CVPR 2018 [c] Qi et al., Low-Shot Learning with Imprinted Weights, CVPR 2018 [d] Joo et al., Being Bayesian about Categorical Probability, ICML 2020 [e] Franchi et al., Encoding the latent posterior of Bayesian Neural Networks for uncertainty quantification, arXiv 2020 [f] Daxberger et al., Laplace Redux -- Effortless Bayesian Deep Learning, NeurIPS 2021 [g] Kristiadi et al., Being bayesian, even just a bit, fixes overconfidence in relu networks, ICML 2020 [h] Balabanov et al., Uncertainty quantification in fine-tuned LLMs using LoRA ensembles, arXiv 2024 [i] Ashukha et al. Pitfalls of In-Domain Uncertainty Estimation and Ensembling in Deep Learning, ICLR 2020 [j] Hu et al., Low-rank adaptation of large language model, ICLR 2022 [k] Kopiczko et al., VeRA: Vector-based Random Matrix Adaptation, ICLR 2024 [l] Franchi et al., Make Me a BNN: A Simple Strategy for Estimating Bayesian Uncertainty from Pre-trained Models, CVPR 2024 [m] Finn et al., Probabilistic Model-Agnostic Meta-Learning, NeurIPS 2018 [n] Gordon et al., Meta-Learning Probabilistic Inference For Prediction, ICLR 2019 [o] Grant et al., Recasting gradient-based meta-learning as hierarchical bayes, ICLR 2018 [p] Kim et al., Bayesian Model-Agnostic Meta-Learning, NeurIPS 2018 [q] Lee et al., Learning to Balance: Bayesian Meta-Learning for Imbalanced and Out-of-distribution Tasks, ICLR 2020 Technical Quality: 2 Clarity: 3 Questions for Authors: This paper takes an interesting direction of study: ensuring reliability of PEFT foundation models in the low-shot learning regime. I find the endeavor of the authors nice, with a good story and several PEFT methods. However I do have several concerns regarding the chosen few-shot learning setup (different from the common practices in the literature) and not necessarily realistic, the limited motivation brought for the use of evidential learning as well as the limited baseline. In addition, the relatively rich literature of Bayesian methods for few-shot learning is ignored. My current rating is leaning towards reject at this time (a bit on the fence), but I'm looking forward for the rebuttal. Here are a few questions and suggestions that could be potentially addressed in the rebuttal or in future versions of this work (please note that suggested experiments are not necessarily expected to be conducted for the rebuttal): 1. Please argue why using evidential learning should be used for this setting rather than last-layer or other Bayesian inspired methods 2. What is the performance of a cosine classifier in this setting? What about Laplace-Redux or LoRA ensembles? 3. How is this method working with LoRA-like methods, e.g., VeRA? 4. Why considering the proposed few-shot learning setup instead of the one from legacy few-shot learning / meta-learning literature 5. Can the authors discuss the difference and benefits of Bayesian-PEFT w.r.t. Bayesian few-shot learning methods? 6. Inclusion of OOD specific metrics in the OOD experiment. Confidence: 3 Soundness: 2 Presentation: 3 Contribution: 2 Limitations: The authors addressed some limitations in the supplementary. Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thank you for taking the time to review our paper. We appreciate your constructive comments, thorough review, and suggestions. Here are the responses to the clarifying questions: Q1: Framing of few-shot learning problem In this work, we consider few-shot classification problem, i.e., the $N$-way $K$-shot classification problem, where the $N$-class classifier has $K$ examples per class to learn from. For instance, 1-shot Cifar100 is a 100-way 1-shot classification problem where the support set has a total of 100 examples (i.e., 1 example per class), and the model is evaluated on the test set (identical to the query set in the meta-testing tasks). We clarify that the current work does not rely on the episodic learning paradigm of meta-learning (matching networks, MAML, PLATIPUS), where both meta-training and meta-testing are done on task level in an episodic fashion with a large number of N-way K-shot tasks. In fact, we consider more challenging few-shot learning tasks (e.g., 100-way 1-shot in Cifar100 and 102-way 1-shot in Flowers102) compared to the commonly used 5-way 1-shot meta-learning tasks. We leverage the power of the pre-trained foundation models eliminating the need of task based episodic meta-training. From a meta-learning perspective, the pre-training phase for the foundation model could be viewed as the meta-knowledge acquisition phase (i.e., meta-training). The pre-trained model can be seen as an expert equipped with meta-knowledge, and parameter-efficient fine-tuning technique performs quick adaptation to the downstream tasks, analogous to the support-set based adaptation done in meta-testing. To test our method in the standard 5-way 1-shot tasks, we apply our model to the mini-ImageNet dataset test tasks and compare it with representative meta-learning models. The results are summarized in Table 1 in the attached PDF. The PEFT-based model outperforms the episodic training meta-learning models demonstrating the potential of large vision foundation models for effective few-shot learning. Q2: Number of classes is the source of Under-confidence Please refer to the answer to Q1 in the overall response. Q3: Performance of simple solutions that do not require training As suggested by the reviewer, we carry out experiments with cosine classifier without training for the 100-way 1-shot task on Cifar100. The results are shown in Table 6 of the PDF file. The cosine classifier has comparable generalization performance (accuracy) in comparison to VPT based model. However, looking at the ECE, the miscalibration issue is even higher than VPT based model. Hence, the simple solution (cosine classifier) does not ensure calibrated predictions. Q4: Low-shot regime suggestion Thank you for the suggestion. We will add a detailed discussion of the few-shot learning setup, clearly explaining that we differ from the existing meta-learning works as we no longer rely on the meta-training phase and episodic learning paradigm. Q5: Scaling evidential methods and the Dirichlet predictor to larger number of classes and encoders. The lack of scalability of EDL has been observed and addressed by Ref [38] (Learn to Accumulate Evidence from All Training Samples: Theory and Practice, ICML 2023) that enables the evidential models to be scaled to ResNet for Cifar100, and we further extend the evidential framework to Vision Transformers in challenging 100-way, 101-way and 102-way few-shot tasks further demonstrating the scalability of evidential models. Q6: Comparison with baselines, e.g., last-layer methods like Laplace-Redux, Bayesian last layer, and linear probing with test-time augmentation We carry out experiments with the suggested models on 100-way 1-shot Cifar100 dataset. The results are presented in Table 6. We observe that these methods also suffer from the under-confidence issue when straightforwardly extended to the VPT. We resort to the Evidential Deep Learning framework to address the under-confidence issue of the VPT model. Similarly, we also fine-tune the ViT model using Lora as another parameter efficient fine-tuning method. The fine-tuning results in under-confidence issue as well. Further expanding the scope of the work to tasks beyond image classification with LoRA and VeRA based methods can be an interesting followup work. Thank you for the suggestion. Q7: Discuss the difference and benefits of Bayesian-PEFT w.r.t. Bayesian few-shot learning methods? Please refer to the answer to Q2 in the general response for a detailed discussion of the proposed method over Bayesian inspired models, which include all the Bayesian few-shot learning methods. As a result, these Bayesian models are inadequate to address the unique under-confidence behavior associated with the PEFT methods. Furthermore, benefiting from a large pre-trained vison foundation model, the proposed method is able to achieve much higher prediction performance as compared with the Bayesian few-shot learning methods trained through episodic meta-training with the specifically constructed few-shot tasks. In Table 1 of the attached PDF file, we show the results on 5-way 1-shot mini-ImageNet. As can be seen, our model outperforms the bayesian few-shot learning models in existing benchmark task by a large margin. Q8: ABNN [l] Discussion ABNN introduces bayesian normalization layers after training of deep learning model, and requires additional training of these layers in a post-hoc manner. In theory, an evidential deep learning model could be augmented with the bayesian normalization layers/ABNN idea as an alternative to belief-based diversity. We will include discussion of ABNN in the updated draft, and leave exploration of ABNN based diversity for bayesian evidential model as a potential future work. Q9: Inclusion of OOD specific metrics in the OOD experiment. Please see the answer to Q3 in the overall response. --- Rebuttal Comment 1.1: Title: Post-rebuttal comments Comment: I would like to thank the authors for the detailed and informative rebuttal. I imagine they invested a lot of effort into clarifying the concerns from the 4 reviewers and I'm confident they will improve this work. I've read the other reviews and the responses of the authors to them and skimmed through the paper again. The strong points I see in this rebuttal are: - several clarifications on the motivation of using evidential models for this tasks and observed under-confidence phenomenon, clarification on few-shot learning setting - new studies to better understand the under-confidence trend w.r.t. prompt size, number of tuned parameters, dataset size - new results comparing with simple baselines that don't need training at all (cosine classifier, test-time adaptation) or limited training (LoRA, Laplace Redux) highlighting the effectiveness of this method but also the presence of under-confidence or low calibration to them - results on typical few-shot learning/meta-learning protocols on mini-Imagenet - reporting of typical OOD metrics in OOD experiments A few minor things that could be improved: - for the OOD experiments (Table 4 in the submitted PDF), only the performance of the proposed model is shown. It would be useful to report the performance of the vanilla PEFT method - the description of the few-shot protocol for the CIFAR100 results could be added (number of sampled episodes in the evaluation) Overall, I think the authors did a nice job in the rebuttal and provided convincing responses. I will raise my rating to 6 - Weak Accept. --- Reply to Comment 1.1.1: Title: Official Comment by Authors Comment: Dear Reviewer 7WZR, Many thanks for carefully checking our rebuttal. We are happy that it has adequately addressed the main concerns and we will make sure to incorporate these changes into the revised paper. We also appreciate that the reviewer increases the rating to 6! We would like to take this opportunity to address the two additional inquiries from the reviewer. First, in the table below, we present the performance of the vanilla PEFT method for the OOD experiments, which is worse than ours. \| Setting \| Method \| AUROC \| AUPR \| FPR95 \| \| ------------- \| --------- \| --------- \| ------- \| -------- \| \| 100 way 1 shot \| PEFT \| 79.53 \| 80.09\| 70.58 \| \| 100 way 1 shot \| B-PEFT \| 81.24 \| 81.98\| 68.15 \| \| 100 way 5 shot \| PEFT \| 90.93 \| 90.75\| 39.82 \| \| 100 way 5 shot \| B-PEFT \| 92.58 \| 92.85\| 35.24 \| Second, for the few-shot protocol on CIFAR100, we have 3 settings. For 5-way 1-shot and 10-way 1-shot, we formulate a task by randomly selecting 5 and 10 classes, respectively. In each such task, the support set contains 1 sample and the query set contains 50 samples for each of the selected classes. For 100-way 1-shot, we formulate a task by randomly selecting 1 sample per class in the support set. The query set contains $100$ samples per class. The result for each setting is obtained by averaging the performance across 50 tasks.	Rebuttal 1: Rebuttal: ## Overall Response We thank all reviewers for their valuable feedback and constructive suggestions. We identify some important questions raised by multiple reviewers and answer them together in our general response below. Q1: Under-confidence behavior of PEFT methods w.r.t. number of classes (Reviewer 7WZR), data size (Reviewers 7WZR and CbTK), and number of unfrozen parameters (Reviewer 7WZR). In order to more thoroughly understand the under-confidence behavior of PEFT methods under diverse settings as suggested by the reviewers, we refer to some of already reported results in the paper while conducting additional experiments. The new results are presented in the attached PDF file. We summarize our main findings as follows: - Data size: In Table 1 of the paper, we varied the number of training samples per class from 1 to 20. We observe that the model's accuracy increases with more training samples. However, under-confidence remains, even with the increase in training samples. We conduct additional experiments on Cifar100 by increasing the training samples per class to 500 and observe the under-confidence issue despite the increase in the accuracy. The trend is summarized in Figure 2 (c-d) of the attached PDF. We see an increase in accuracy and a decrease in ECE. However, even with 500 samples per class, the under-confidence issue remains. Further, we report the accuracy and ECE of the fully fine-tuned model (fine-tuning of all the parameters) for 1 shot cifar100 in Table 5 of the attached pdf. We observe a decrease in accuracy. The miscalibration issue still remains. We observe the overconfidence behavior by looking at the reliability plot as presented in Figure 1 (a). The performance suggests the overfitting of the fully tuned model. - Number of classes: To study the relationship between model accuracy, uncertainty, and the number of classes, we formulate standard 5-way 1-shot MiniImagenet tasks and apply the PEFT to these tasks. The results are summarized in Table 2 of the attached PDF. As can be seen, the model achieves high overall accuracy of 89.78\%. However, the model remains under-confident with an ECE of 0.418, where under-confidence is shown by the the reliability diagram in Figure 1. We further formulate 5-way 1 shot, 10-way 1 shot, and 100-way 1 shot tasks in Cifar100. The results are presented in Table 3. As we decrease the number of shots from 100 to 5, we see an increase in accuracy and a decrease in ECE. However, the under-confidence remains. - Number of unfrozen parameters: We conduct additional experiments on Cifar100 via 100-way 1-shot tasks by varying the number of prompts for 1) shallow prompt: prompt added to the input only and 2) deep prompt: prompt added to all Transformer encoder layers' input as well. The accuracy and ECE trends are presented in Figure 2 (a)/(b) of the pdf. As can be seen, with the increase in the number of prompts for both shallow and deep prompts, there are fluctuations in accuracy and ECE performance. However, the under-confidence issue persists for all the cases. Q2: Justification of using evidential learning to achieve calibrated PEFT (Reviewers 7WZR and YJfL) As compared with the Bayesian-inspired models, evidential learning offers two key properties that allow us to formulate a principled solution to address the unique under-confident behavior of the PEFT methods. First, thanks to its evidence-based fine-grained uncertainty decomposition capability, we can separate two distinct sources of second-order uncertainty, including vacuity (line 172) and dissonance (line 176). Different from the commonly used first-order uncertainty (e.g., entropy), these two second-order uncertainty serve as a key tool to understand why PEFT methods are both accurate (with a low dissonance) while being under-confident (with a high vacuity). This key insight suggests that these methods systematically under-estimates the contribution from the prior knowledge to the downstream task. While the classical Bayes’ theorem offers a principal idea to address the issue, which is to strengthen the prior belief, there is a lack of practical way to achieve this. As the second key property, evidential learning allows us to leverage the base rate, which is rooted in the subjective logic theory as an effective vehicle to adjust the prior belief gained through pre-training. To this end, we propose a transformation function in Eq (6) to adjust the base rate that leads to the increase of the model confidence while maintaining the predictive accuracy of the model as guaranteed by our theoretical results in Lemma 2 and Theorem 3. We also conduct additional experiments and compare our approach with Bayesian inspired models for 100-way 1-shot Cifar100 tasks. The results are reported in Table 6. We use laplace approximation on last layer of the model using Kronecker Product and Diagonalization represented by KronLaplace and DiagLaplace in the table. As can be seen, the Bayesian-inspired models do not address the under-confidence issue, which still exhibits a fairly high ECE. This further confirms the effectiveness of the proposed evidential learning-based approach. Q3: OOD experiments and Clarification (Reviewers 7WZR and Dmeg) We present the OOD results of Cifar10 as in-distribution dataset and Cifar100 as out-of-distribution dataset with AUROC, FPR95, AUPR metrics for our model on $100$-way $1$-shot and $100$-way $5$-shot cifar100 tasks in Table 4. As seen, with more training data, the model's ood detection capabilities improve. Even with only 5 samples/class (i.e. $100$-way $5$-shot cifar100 task), the model can achieve AUROC of 92.58. Pdf: /pdf/f5135d4818f617d799f698be9d7c65c3692f86c8.pdf	NeurIPS_2024_submissions_huggingface	2,024	null	null	null	null	null	null	null	null
Diffusion Models With Learned Adaptive Noise	Accept (spotlight)	Summary: This paper introduces the Multivariate Learned Adaptive Noise (MULAN) model, a novel diffusion process that adapts the noise schedule to different regions of an image. The authors claim significant improvements in log-likelihood estimation and training efficiency, challenging the conventional assumption that the ELBO is invariant to the noise process, and demonstrate state-of-the-art performance in density estimation on CIFAR-10 and ImageNet, reducing training steps by 50%. I appreciate the paper because it is based on the presumption, which I share, that a significant improvement in generative diffusion approaches can be achieved by adjusting the forward process to the data. Specifically, the authors suggest adapting both drift and noise to the input data. My understanding is that the adaptivity of the MULAN model is achieved by utilizing more degrees of freedom associated with the forward process within the ELBO-style learning/training. However, the claim that the “optimal” schedule is spatially inhomogeneous is based purely on empirical evidence. I would like to see the authors turn this empirical finding into a more systematic exploration. For instance, can we draw conclusions about the spatio-temporal details of the optimal schedule, such as whether the earlier stages of the forward process are more spatially inhomogeneous than the later ones? Do we see more inhomogeneity on the periphery of the image? How does the inhomogeneity correlate with a spatially coarse-grained (filtered) version of the image? Is it more beneficial to have inhomogeneity in the drift (deterministic) or diffusion (stochastic) part of the forward process? The paper also discusses using labels and other high-level features of the image, but it is unclear how these affect or interplay with the spatial inhomogeneity. Overall, while the paper presents interesting ideas, I feel it is not yet ready for prime time. The empirical claims need to be backed by a more systematic and thorough analysis. Strengths: see summary Weaknesses: see summary Technical Quality: 2 Clarity: 3 Questions for Authors: see summary Confidence: 4 Soundness: 2 Presentation: 3 Contribution: 2 Limitations: see summary Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: # Response to uSJa We thank the reviewer for their detailed and thorough feedback. We address their concerns below. In the appendix, we have provided numerous experiments trying to explore how the noise schedule relates to different aspects of an image namely: 1. Frequency distribution. 2. Intensity. ## Concern 1: Finding Interpretable Spatio-Temporal Variation in the Noise Schedule We have provided a thorough analysis trying to relate the learnt noise schedule to different properties of the input image such as intensity and frequency in sec. 14 in the Appendix. We summarize three relevant experiments hereby: 1. CIFAR-10 Frequency Split To see if MULAN learns different noise schedules for pixels in an image with different frequencies, we modify the images in the CIFAR-10 dataset where we split an image into high frequency and low frequency components using gaussian blur. Observation We do notice two separate bands in the noise schedule (see fig. 11) however these bands don’t necessarily correspond to the low / high freq regions in the image. Upon comparing the schedules to that of the original CIFAR10-dataset, we do notice an increase in the inter-spatial variation of the noise schedules. 2. CIFAR-10 Masking We modify the CIFAR-10 dataset where we randomly mask (i.e. replace with 0s) the top of an image or the bottom half of an image with equal probability. Fig. 10a shows the training samples. MULAN was trained for 500K steps. The samples generated by MULAN is shown in Fig. 10b. Observation We don’t observe any interpretable patterns in the noise schedule for the masked pixels vs non-masked pixels. Upon comparing the schedules to that of the original CIFAR10-dataset, we do notice an increase in the inter-spatial variation of the noise schedules. 3. CIFAR-10 Intensity Split To see if MULAN learns different noise schedules for images with different intensities, we modify the images in the CIFAR-10 dataset where we randomly convert an image into a low intensity or a high intensity image with equal probability. Fig. 9a shows the training samples. MULAN was trained for 500K steps. The samples generated by MULAN is shown in Fig. 9b. Observation We don’t observe any interpretable patterns in the noise schedules for the high intensity images vs low intensity images. Upon comparing the schedules to that of the original CIFAR10-dataset, we do notice an increase in the inter-spatial variation of the noise schedules. Temporal Variation In fig. 11 (left) in the appendix, we visualize the noise schedules for all the pixels and we observe that the schedules are spatially identical in the beginning and the end of the diffusion processes i.e. at $t=0$ and $t=1$. Furthermore, the periphery of the image did not exhibit more inhomogeneity than the other parts of the image. ## Concern 2: Optimal schedule is spatially inhomogeneous is empirical We use principles of physics to motivate spatial inhomogeneity of an optimal noise schedule. In Section 3.5, we show that the diffusion loss is a line integral and can be interpreted as the amount of work done, $\int_{0}^{1} \mathbf{f}(\mathbf{r}(t)) \cdot \frac{d}{dt}\mathbf{r}(t) \, dt$, along the diffusion trajectory $\mathbf{r}(t) \in \mathbb{R}^n_{+}$ in the presence of a force field $\mathbf{f}(\mathbf{r}(t)) \in \mathbb{R}^n_+$. A line integrand is almost always path-dependent unless its integral corresponds to a conservative force field, which is rarely the case for a diffusion process [1]. The path $\mathbf{r}(t)$ is parameterized by the noise schedule, and thus by learning the noising process, we are effectively learning a path that incurs the least work done. Since this force field can be “arbitrary” as it is parameterized by the denoising neural network, it is highly unlikely that the path of least work done is a straight line joining $\mathbf{r(0)}$ and $\mathbf{r(1)}$. In other words, it is highly “unlikely” that the optimal noise schedule is scalar. Note that a straight line trajectory represents a scalar noise schedule as described in the appendix sec 11.5. ---- Reference: [1] Richard E. Spinney and Ian J. Ford. Fluctuation relations: a pedagogical overview, 2012. ## Concern 3: Inhomogeneity in the drift / diffusion term In this work we consider variance preserving diffusion process which means that we can’t possibly disentangle the inhomogeneity in the drift and the diffusion term. The forward process takes the form $\mathbf{x}_t = \mathbf{\alpha}\_{t\|s}(\mathbf{z}) \mathbf{x}_s + \sqrt{1 - \mathbf{\alpha}\_{t\|s}^2(\mathbf{z})} \epsilon$ with $s < t$. Removing the adaptivity in the drift or the diffusion term would mean that the marginals, $\mathbf{x}_t \sim q(\mathbf{x}_t \| \mathbf{x}_0)$ are intractable to compute and would require simulating the diffusion process to obtain intermediate latents $\mathbf{x}_t$ such as Diffusion Normalizing Flows [1]. ---- Reference: [1] Zhang, Q. and Chen, Y., 2021. Diffusion normalizing flow. Advances in neural information processing systems, 34, pp.16280-16291. --- Rebuttal Comment 1.1: Comment: Good response. My new score is 6. --- Reply to Comment 1.1.1: Title: Current rating doesn't reflect the updated score of 6 Comment: Dear Reviewer, Thank you very much for evaluating our manuscript and providing invaluable feedback. We would like to bring to your attention that the current rating still reflects the previous score of 4 instead of the updated score of 6. We wanted to mention this in case there was any confusion.	Summary: This paper proposes to extend diffusion models by learning a per-pixel noise schedule for the forward noising process, that can be conditional on a context or on an auxiliary variable. This leads to faster convergence and SOTA results on density estimation on the simple benchmark datasets CIFAR-10 and ImageNet-32. Strengths: I believe this is overall a good paper. The research question is relevant, the method is theoretically sound, well motivated, and explained clearly and in detail. The experiments are sufficiently extensive and the results are good, although only on 2 benchmark datasets that are relatively simple by today's standards. Weaknesses: - On tables 2, 7, 8 there are confidence intervals on the means, but it's unclear how many random seeds were used, and what's the impact of randomness (I guess weight initialization). Also, on the plots (I'm looking at those with ablations) there seem to be no CIs. - Sections 3.2.1 and 3.2.2 are a bit repetitive, since 3.2.2 extends 3.2.1 only in a minor way. Especially Eqs. (3) and (4) are exactly the same (except for the context dependency). - It would be helpful to have a clear self-contained definition for the forward and backward processes, after introducing everything, to summarize the overall setup. - In the forward process, the dependency on the context $\mathbf{c}$ or the latent variables $\mathbf{z}$ seems to be only through the element-wise noise schedule. In particular, the mentioned NDM and DiffEnc methods appear to be at least partially orthogonal and might lead to further improvements. As far as I can tell, this is not explicitly discussed. - Only up to 32x32 images are considered here. I think ImageNet-64 should really be included at this point. Since by now density estimation on these benchmarks works very well, it might be interesting and relevant to scale it up to even higher resolutions, but as a proof of concept I think something like the standard ImageNet-64 (used for example in the VDM paper, too) should be sufficient. - Typo: at line 54, $\sigma_t$ should be $\sigma_t^2$. Technical Quality: 3 Clarity: 3 Questions for Authors: Why are the results in Table 3 reported separately? I think it would be clearer to include them in Table 2 (maybe separated by a horizontal line from older methods). Confidence: 4 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: Yes. Not necessarily a limitation per se, but I would potentially mention that the forward diffusion process depends on the context or latent variables only through the pixel-wise noise schedule. There are orthogonal methods (which are already mentioned in the paper) that learn the mean of the forward process too. Flag For Ethics Review: ['No ethics review needed.'] Rating: 7 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: ## Response to 22Fq We thank the reviewer for their detailed and thorough feedback. We address their concerns below. ## Concern 1: Evaluating on larger scale experiments (ImageNet-64) Unfortunately, it is not feasible for us to train an ImageNet-64 model within our academic research group. Using the official VDM Jax codebase, it takes 8 days on 8xA100 to train a VDM model on ImageNet-32. We would need to train multiple models on the larger 64x64 dataset, and this is outside the scope of what out group has access to. However, the 32x32 CIFAR10 and ImageNet datasets on which we report our results are standard benchmarks in the field and we report state-of-the-art performance on these benchmarks. We anticipate our method to scale to larger datasets with similar computational complexity as existing diffusion models. Mulan requires an encoder that is much smaller than the denoising model, and requires 5-10% of additional memory. ## Concern 2: Comparisons with other learned diffusion algorithms Our method can be seen as a special case of a diffusion normalizing flow (DNF), which uses the following forward process: $\text{d}\mathbf{x} = \mathbf{f}_\theta(\mathbf{x}, t)\text{d}t + g(t)\text{d}\mathbf{w}$ . However, for such general processes, training requires backpropagating through sampling from the model, which is computationally expensive. In practice, this reduces scalability and produces worse results. However, there may be a class of processes between full normalizing flows and our simpler noise processes that will admit scalable training and improved performance. Our work shares similarity with the NDM and DiffEnc methods, which add an additive learned term to the noise process instead of an multiplicative one like in our paper. They still admit efficient sampling. These methods can potentially be combined with our work: in theory they are orthogonal, but we anticipate that getting them to work will require non-trivial engineering work around architecture tuning and optimization strategies. We do provide a brief discussion in lines 583-589. We’ll add discussions to the updated manuscript. ## Concern 3: Other Concerns We thank the reviewer for pointing out the typo. The reviewers comments on the presentation style such as consolidating sec 3.2.1 and 3.2.2 with self-contained definition for the forward and backward processes is also very useful. We’ll factor these comments into the next version of the manuscript. ## Question 1: Clarification regarding confidence Intervals To compute the confidence interval, we calculate the variance, $\sigma$, of the likelihood across all the data points in the validation/test set and report the `95% confidence interval` using the formula: $\pm 1.96 \frac{\sigma}{\sqrt{n}}$, where $n$ denotes the number of data points. Given the very small confidence interval of `0.001` on the likelihood, we did not further randomize the computation of likelihood using additional seeds. --- Rebuttal Comment 1.1: Comment: Thank you for addressing my questions and providing a thorough and insightful rebuttal to all reviewers. I am satisfied with your responses and recommend acceptance, though I will maintain my current score for now. --- Rebuttal 2: Title: Thank you for you invaluable feedback Comment: Dear Reviewer, We truly appreciate you providing invaluable feedback on our manuscript. Thanks, authors	Summary: This paper introduces an enhanced framework for variational diffusion models (VDM). Rather than using a uniform noise scheduler for all pixels, the new approach assigns different schedulers to individual pixels, adapting to the data distributions. The authors highlight the novelty of this extension, noting that the Evidence Lower Bound (ELBO) depends on the entire trajectory induced by the scheduler, not just the endpoints, as is the case in traditional VDM. Empirical studies demonstrate that their framework achieves state-of-the-art density estimation on CIFAR10 and ImageNet, reducing the number of training steps by 50%. Strengths: The ELBO perspective of diffusion models has drawn much attention since the diffusion models' introduction. This paper made a non-trivial extension to the current framework and showed that the well-known assumption is no longer held that noise schedule does not alter EBLO. This observation gives additional flexibility to increase ELBO, resulting in a better performance in density estimation. Weaknesses: 1. In section 3.5, the authors discussed why the generalization makes the ELBO rely on the entire trajectory; however, they state the fact without giving any intuitive explanations. The authors should discuss this issue more. In addition, providing some toy examples to clearly show how the extended framework is differentiated from the existing one could significantly improve the paper's presentation. 2. The FIDs of the proposed model are significantly worse than the existing diffusion models. 3. The FIDs reported for VDM seem inconsistent with the one reported in the original paper. Can authors clarify how the FIDs are obtained. Technical Quality: 3 Clarity: 2 Questions for Authors: Please refer to the Weakness section. Confidence: 4 Soundness: 3 Presentation: 2 Contribution: 3 Limitations: I am not aware of any potential negative societal impact of this work. Flag For Ethics Review: ['No ethics review needed.'] Rating: 7 Code Of Conduct: Yes	Rebuttal 1: Title: Response to reviewer QQVN Comment: ## Concern 1: Intuition behind ELBO being dependent on the diffusion trajectory Imagine you’re piloting a plane across a region where cyclones and strong winds are present. If you plot a straight line course directly through these adverse weather conditions, the journey requires more fuel and effort due to the resistance encountered. Instead, if you navigate around the cyclones and adverse winds, it takes less energy to reach your destination, even though the path might be longer. This intuition maps to mathematical and physical terms. The trajectory of the plane is denoted by $\mathbf{r}(t) \in \mathbb{R}^n_{+}$ and the forces acting on the plane are given by $\mathbf{f}(\mathbf{r}(t)) \in \mathbb{R}^n_+$. The work used to navigate the plane is $\int_{0}^{1} \mathbf{f}(\mathbf{r}(t)) \cdot \frac{d}{dt}\mathbf{r}(t) \, dt$. The work here is dependent of the trajectory because $\mathbf{f}(\mathbf{r}(t)) \in \mathbb{R}^n_+$ is not a conservative field. In Section 3.5, we argue that the same holds for the diffusion ELBO. The trajectory, $\mathbf{r}(t) \in \mathbb{R}^n_{+}$, is parameterized by the noise schedule which is influenced by complex forces, $\mathbf{f}(\mathbf{r}(t)) \in \mathbb{R}^n_+$ (akin to weather patterns). The diffusion loss can be interpreted as the amount of work done, $\int_{0}^{1} \mathbf{f}(\mathbf{r}(t)) \cdot \frac{d}{dt}\mathbf{r}(t) \, dt$, along the diffusion trajectory in the presence of these forces (the forces are modeled by dimension-wise reconstruction error of the denoising model). By carefully selecting (learning the noise schedule) our path—avoiding “high-resistance” areas where the loss accumulates quickly—we can minimize the overall “energy” expended, measured in terms of the NELBO. ## Concern 2: Worse FIDs than the existing diffusion models. Note that MuLAN does not incorporate many tricks that improve FID such as exponential moving averages, truncations, specialized learning schedules, etc.; our FID numbers can be improved in future work using these techniques. However, our method yields state-of-the-art log-likelihoods. We leave the extension to tuning schedules for FID to future work. Optimizing for the likelihood is directly motivated by applied problems such as data compression [2]. In that domain, arithmetic coding techniques can take a generative model and produce a compression algorithm that provably achieves a compression rate (in bits per dimension) that equals the model’s log-likelihood [3]. Other applications of log-likelihood estimation include adversarial example detection [4], semi-supervised learning [5], and others. ## Concern 3: Inconsistent FID numbers for VDM as reported in the original paper. Note that in their paper, VDM reported FIDs using the DDPM sampler with `T=1000`, whereas in our paper, we generate samples using an ODE solver that solves the reverse-time Probability Flow ODE, as shown in Eqn (76) for VDM and Eqn (81) for MuLAN. We employ the RK45 ODE solver [6] provided by `scipy.integrate.solve_ivp` with `atol=1e-5` and `rtol=1e-5`. These settings align with prior work by Song et al., 2021 [7]. One method to measure the complexity of the learned noise schedule is by comparing the Number of Function Evaluations (NFEs) required to generate samples using an ODE solver to solve the associated Probability Flow ODE. In Table 1, we aimed to study the difficulty of drawing samples from the diffusion model (measured by NFEs) and the quality of the samples (measured by FID). We observe that MuLAN does not degrade FIDs while improving log-likelihood estimation. --- References: [1] Theis, L., Oord, A. V. D., & Bethge, M. (2015). A note on the evaluation of generative models. arXiv preprint arXiv:1511.01844. [2] David JC MacKay. Information theory, inference and learning algorithms. Cambridge university press, 2003. [3] Thomas M Cover and Joy A Thomas. Data compression. Elements of Information Theory, pp. 103–158, 2005. [4] Yang Song, Taesup Kim, Sebastian Nowozin, Stefano Ermon, and Nate Kushman. Pixeldefend: Leveraging generative models to understand and defend against adversarial examples. arXiv preprint arXiv:1710.10766, 2017. [5] Zihang Dai, Zhilin Yang, Fan Yang, William W Cohen, and Russ R Salakhutdinov. Good semi-supervised learning that requires a bad gan. Advances in neural information processing systems, 30, 2017. [6] J.R. Dormand and P.J. Prince. A family of embedded runge-kutta formulae. Journal of Computational and Applied Mathematics, 6(1):19–26, 1980. ISSN 0377-0427. doi: https://doi. org/10.1016/0771-050X(80)90013-3. URL https://www.sciencedirect.com/science/ article/pii/0771050X80900133. [7] Yang Song, Conor Durkan, Iain Murray, and Stefano Ermon. Maximum likelihood training of score-based diffusion models. Advances in neural information processing systems, 34: 1415–1428, 2021. --- Rebuttal Comment 1.1: Title: Thank you for the detailed responses. Comment: I appreciate the authors' thorough responses. All my concerns have been addressed. However, I would appreciate it if the authors could incorporate discussions on the intuition behind ELBO into the final version. I have adjusted the score accordingly. Nice work! --- Reply to Comment 1.1.1: Title: Thank you for your feedback Comment: Dear Reviewer, We are immensely grateful to you for providing detailed and thorough feedback on our manuscript. We'll incorporate the discussion on the intuition on ELBO into the final version of the manuscript. Thanks, authors.	Summary: The authors propose MuLAN, a method to learn a multivariate (pixel-wise for images) noise injection schedule for diffusion models, leading to improved likelihood estimates compared to prior work. They provide extensive experimental results and an ablation study to demonstrate the efficiency of their method. Strengths: - The paper is well-written and easy to follow. - The authors present a theoretical motivation for making the noise schedule multivariate and learnable from a variational inference perspective, showing that it can impact the ELBO of the model unlike a univariate schedule. - The proposed method demonstrates state-of-the-art likelihood estimation results. - The authors conduct an extensive ablation study to thoroughly investigate the proposed method. Overall, I believe MuLAN is a well-motivated and logical extension of diffusion models that could be useful in practical applications. Weaknesses: - According to the experimental results provided, the only benefit of MuLAN is improved likelihood estimation. The experiments are conducted exclusively in the image domain, but MuLAN does not yield better FID scores compared to prior works. As discussed in Section 3.1, a model with better likelihood estimation may be useful for tasks like compression or adversarial example detection. However, the authors only provide likelihood estimation results and do not demonstrate improved practical outcomes in downstream tasks. It's not completely clear where MuLAN should be preferred over other models. Including results from downstream tasks, especially in non-image domains where uniform Gaussian noise may not be as effective, would enhance the paper. - The derivation of the multivariate noise schedule closely follows the derivations from VDM, substituting the original scalar $\gamma$ function with a multivariate (and conditional) one. This approach limits the novelty of the proposed method. - The significant part of the contribution seems to come from the introduction of the auxiliary latent variable $z$, as indicated by the ablation study in Section 4.4. I would appreciate a more detailed discussion of this technique and its efficacy. This step appears to add an extra hierarchical layer to the variational autoencoder. Since the vanilla diffusion model can be seen as a hierarchical VAE itself, it's not entirely clear what this addition offers. A more detailed discussion of this idea and its connection to other works, such as diffusion in the latent space of VAEs [1] or diffusion where the noising process does not necessarily match the prior at t=1 [2], would be beneficial. - The authors note that the learned noise injection schedule is challenging to interpret. However, including some visualizations of the learned schedule at different time steps would provide at least some intuition about its characteristics. [1] Vahdat, A., Kreis, K., & Kautz, J. (2021). Score-based generative modeling in latent space. [2] Lee, S., Kim, B., & Ye, J. C. (2023, July). Minimizing trajectory curvature of ode-based generative models. Technical Quality: 3 Clarity: 3 Questions for Authors: - You claim (e.g., on lines 190-193) that MuLAN leads to a tighter NLL-ELBO gap. Do you report the gap for MuLAN and compare it with prior works? In Table 2, to which you refer, you only provide likelihood estimation results, without an ELBO comparison. - At the end of the day, the polynomial parameterization of the $\gamma$ function involves just three parameters, which seems significantly less flexible than the parameterization from VDM. Am I correct in understanding that you claim that the main reason the polynomial parameterization performs better is because it is easier to optimize? - As you discuss, the polynomial parameterization may have two points with zero derivative. Do you restrict the polynomial parameterization to have zero derivative at t=0 and t=1? - Could you clarify how you calculate the likelihood? In regular diffusion, we may integrate the density starting from the unit Gaussian distribution and following the ODE dynamics. However, in MuLAN, you introduce auxiliary variables z, and it’s unclear how to incorporate them into the likelihood estimation pipeline. Do you simply sample z from its prior and then integrate the ODE starting from the data point x, conditionally on z? - In Section 4.3, where you discuss connections with other approaches that also learn the forward process, you mention that these papers do not provide experimental results on the same version of the ImageNet dataset. However, from what I see, NDM references the same version of ImageNet. I think it’s worth including this comparison in Table 3. While NDM reports a lower NLL score, as you discussed, your approach is significantly different, so I don’t believe it characterizes your work negatively. Confidence: 5 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: The authors have adequately addressed the limitations. Flag For Ethics Review: ['No ethics review needed.'] Rating: 7 Code Of Conduct: Yes	Rebuttal 1: Title: Response to reviewer 1B1k (1/3) Comment: We want to thank the reviewer for their constructive feedback. We address each concern below. ## Concern 1: Demonstrating the utility of density estimation on downstream tasks Optimizing for the likelihood is directly motivated by applied problems such as data compression [2]. In that domain, arithmetic coding techniques can take a generative model and produce a compression algorithm that provably achieves a compression rate (in bits per dimension) that equals the model’s log-likelihood [3]. Other applications of log-likelihood estimation include adversarial example detection [4], semi-supervised learning [5], and others. In this work, we take the first step of improving the likelihood, and we aim to explore major downstream applications in future work. We note that one immediate simple application of our work is faster training: we attain strong likelihoods and FIDs on ImageNet using half the training steps (Table 1). Note that MuLAN does not incorporate many tricks that improve FID such as exponential moving averages, truncations, specialized learning schedules, etc.; our FID numbers can be improved in future work using these techniques. However, our method yields state-of-the-art log-likelihoods. We leave the extension to tuning schedules for FID to future work. --- References: [1] Theis, L., Oord, A. V. D., & Bethge, M. (2015). A note on the evaluation of generative models. arXiv preprint arXiv:1511.01844. [2] David JC MacKay. Information theory, inference and learning algorithms. Cambridge university press, 2003. [3] Thomas M Cover and Joy A Thomas. Data compression. Elements of Information Theory, pp. 103–158, 2005. [4] Yang Song, Taesup Kim, Sebastian Nowozin, Stefano Ermon, and Nate Kushman. Pixeldefend: Leveraging generative models to understand and defend against adversarial examples. arXiv preprint arXiv:1710.10766, 2017. [5] Zihang Dai, Zhilin Yang, Fan Yang, William W Cohen, and Russ R Salakhutdinov. Good semi-supervised learning that requires a bad gan. Advances in neural information processing systems, 30, 2017. ## Concern 2: Discussion on auxiliary latent variables Diffusion models augmented with an auxiliary latent space [1, 2] have been used for representation learning. These auxiliary latents, $\mathbf{z} \in \mathbb{R}^m$, have a smaller dimensionality than the input $\mathbf{x}_0 \in \mathbb{R}^d$ and are used to encode a high-level semantic representation of $\mathbf{x}_0$. In this work we make an observation that an input conditioned multivariate noise schedule can indeed improve likelihoods unlike previously thought so. However, simply conditioning on the noise schedule is challenging because of the reasons mentioned in Sec 10.2 in the appendix. In this work we resolve these challenges by conditioning the noise schedule on the input via an auxiliary latent space. The key differences between these methods: \| \| Learned noise \| Multivariate noise \| Input Conditioned noise \| Auxiliary latents \| Noise parameterization \| \| --- \| --- \| --- \| --- \| --- \| --- \| \| InfoDiffusion [1] \| No \| No \| No \| In denoising process \| Cosine schedule \| \| MuLAN (ours) \| Yes \| Yes \| Yes \| In noising & denoising processes \| Polynomial (novel) \| These auxiliary latents, $\mathbf{z} \sim p_\theta(\mathbf{z})$ , differ from the noisy inputs, $\mathbf{x}_t$ in the sense that these latents are kept fixed throughout the reverse generation process; however, the $\mathbf{x}_t$ is progressively denoised to get clean $\mathbf{x}_0$. References: [1] Yingheng Wang, Yair Schiff, Aaron Gokaslan, Weishen Pan, Fei Wang, Christopher De Sa, and Volodymyr Kuleshov. Infodiffusion: Representation learning using information maximizing diffusion models. In International Conference on Machine Learning, pp. xxxx–xxxx. PMLR, 2023. [2] Ruihan Yang and Stephan Mandt. Lossy image compression with conditional diffusion models, 2023. --- Rebuttal 2: Title: Response to reviewer 1B1K (2/3) Comment: ## Concern 3: Visualizations of the learned schedule We have provided visualizations for the noise schedule for different pixels in Fig. 11 (left column) in the appendix for various datasets such as CIFAR-10 and ImageNet. Furthermore, we have numerous experiments trying to relate the learned noise schedule to different properties of the input image such as intensity and frequency in sec. 14 in the Appendix. We summarize three relevant experiments hereby: 1. CIFAR-10 Frequency Split To see if MULAN learns different noise schedules for pixels in an image with different frequencies, we modify the images in the CIFAR-10 dataset where we split an image into high frequency and low frequency components using gaussian blur. Observation We do notice two separate bands in the noise schedule (see fig. 11) however these bands don’t necessarily correspond to the low / high freq regions in the image. Upon comparing the schedules to that of the original CIFAR10-dataset, we do notice an increase in the inter-spatial variation of the noise schedules. 2. CIFAR-10 Masking We modify the CIFAR-10 dataset where we randomly mask (i.e. replace with 0s) the top of an image or the bottom half of an image with equal probability. Fig. 10a shows the training samples. MULAN was trained for 500K steps. The samples generated by MULAN is shown in Fig. 10b. Observation We don’t observe any interpretable patterns in the noise schedule for the masked pixels vs non-masked pixels. Upon comparing the schedules to that of the original CIFAR10-dataset, we do notice an increase in the inter-spatial variation of the noise schedules. 3. CIFAR-10 Intensity Split To see if MULAN learns different noise schedules for images with different intensities, we modify the images in the CIFAR-10 dataset where we randomly convert an image into a low intensity or a high intensity image with equal probability. Fig. 9a shows the training samples. MULAN was trained for 500K steps. The samples generated by MULAN is shown in Fig. 9b. Observation We don’t observe any interpretable patterns in the noise schedules for the high intensity images vs low intensity images. Upon comparing the schedules to that of the original CIFAR10-dataset, we do notice an increase in the inter-spatial variation of the noise schedules. Temporal Variation In fig. 11 (left) in the appendix, we visualize the noise schedules for all the pixels and we observe that the schedules are spatially identical in the beginning and the end of the diffusion processes i.e. at $t=0$ and $t=1$. ## Concern 4: Comparisons to other works ### Variational Diffusion Models We'd like to emphasize the fact that this work changes the understanding of noise schedules and how they affect the likelihood of a diffusion model. It discards the existing notion that the likelihood of a diffusion model is invariant to the choice of noise schedules. To achieve this we introduce the concept of adaptive noise schedules where the noise schedule is conditioned on the input via auxiliary latent variables. We include a comparison to VDM hereby that illustrates the key differences between these methods. \| \| Learned noise \| Multivariate noise \| Input Conditioned noise \| Auxiliary latents \| Noise parameterization \| \| --- \| --- \| --- \| --- \| --- \| --- \| \| VDM (Kingma et al., 2021) \| Yes \| No \| No \| No \| Monotonic neural network \| \| MuLAN (ours) \| Yes \| Yes \| Yes \| In noising & denoising processes \| Polynomial, sigmoid \| ### Score-based generative modeling in latent space. Latent Score-based Generative Model (LSGM) [1] defines a diffusion process in the latent space, rather than directly on the input data. In contrast, MuLAN defines the diffusion process on the input data but conditions the forward and reverse processes on the auxiliary latent space. Furthermore, we’d like to highlight that MuLAN achieves a Negative Log Likelihood (NLL) of 2.65 on CIFAR-10, which is significantly better than LSGM’s 2.87. ### Minimizing trajectory curvature of ode-based generative models. In Minimizing Trajectory Curvature (MTC) of ODE-based generative models, the primary goal is to design the forward diffusion process that is optimal for fast sampling; however, MuLAN strives to learn a more expressive forward process that optimizes for log-likelihood. In the appendix, we have provided a detailed explanation in sec 7.4 and Tab. 6 highlights the key differences between the methods. References: [1] Vahdat, A., Kreis, K., & Kautz, J. (2021). Score-based generative modeling in latent space. [2] Lee, S., Kim, B., & Ye, J. C. (2023, July). Minimizing trajectory curvature of ode-based generative models. --- Rebuttal 3: Title: Response to reviewer 1B1k (3/3) Comment: ## Concern 5: Other concerns > ### Polynomial Parameterization Although the Monotonic Neural Network parameterization proposed in VDM is intended to be more expressive, the design of the neural network restricts its expressivity. Specifically, the weights must always be positive, and it utilizes sigmoid activations to maintain the monotonicity of the noise schedule. We observed that in deeper layers, the activations could easily saturate, thus limiting the expressivity of the function. In contrast, the polynomial parameterization features only 3 free variables that can be set arbitrarily while still ensuring the noise schedule’s monotonicity. Furthermore, as the reviewer noted, the polynomial parameterization may have two zero derivatives between $t=0$ and $t=1$. However, we do not explicitly set the derivatives at the endpoints to zero. > ### Imagenet likelihood for NDM The reason we did not include the likelihood numbers for NDM on ImageNet is that they employed numerous tricks to achieve improved likelihood, which the baselines reported in Table 2 did not use, namely: 1. Data augmentation. NDM pre-processes their data using flipping which is a commonly used trick to improve the likelihood. For ex, on CIFAR-10, VDM w/o data augmentation achieves an NLL of 2.65 but with data augmentation, the NLL improves to 2.49; see [1]. 2. Parameter Count. NDM [2] uses an encoder model that is of the same size as that of the denoising model (see sec 4.1 in [2]) and hence they use almost 2x more parameters than ours. In contrast, MuLAN’s encoder is only ~10% size of the denoising model. References: [1] Diederik Kingma, Tim Salimans, Ben Poole, and Jonathan Ho. Variational diffusion models. Advances in neural information processing systems, 34:21696–21707, 2021. [2] Grigory Bartosh, Dmitry Vetrov, and Christian A Naesseth. Neural diffusion models. arXiv preprint arXiv:2310.08337, 2023. > ### Clarification on Likelihood computation To compute the likelihood of a datapoint $\mathbf{x}_0$ we need to evaluate the following equation (eqn. (84) in the paper): $- \mathbb{E}\_{q\_\phi(\mathbf{z}\|\mathbf{x}\_0)} \log p\_\theta(\mathbf{x}\_0 \| \mathbf{z}) + \mathbb{D}\_{\text{KL}} \left( q\_\phi(\mathbf{z}\|\mathbf{x}\_0) \parallel p_\theta(\mathbf{z}) \right)$ where $\mathbf{x}\_0$ denotes the input data, $\mathbf{z}$ denotes the auxiliary latent variable, $q_\phi(\mathbf{z}\|\mathbf{x}\_0)$ and $p_\theta(\mathbf{z})$ denote the approximate posterior and the prior distributions for the auxiliary latent variables respectively, $p_\theta(\mathbf{x}_0 \| \mathbf{z})$ is the likelihood of the diffusion model conditioned on the auxiliary latent. 1. To compute the term $\mathbb{E}\_{q\_\phi(\mathbf{z}\|\mathbf{x}\_0)} \log p_\theta(\mathbf{x}\_0 \| \mathbf{z})$, we draw a sample $\mathbf{z} \sim q_\phi(\mathbf{z} \| \mathbf{x}_0)$ and then use an ODE solver to compute the ODE dynamics in Eqn (83) in the paper. 2. The term $\mathbb{D}\_{\text{KL}} \left( q_\phi(\mathbf{z}\|\mathbf{x}\_0) \parallel p_\theta(\mathbf{z}) \right)$ is computed in closed form; see line 223 in the paper. --- Rebuttal Comment 3.1: Comment: I thank the authors for their comprehensive rebuttal. I will maintain my score. --- Reply to Comment 3.1.1: Title: Thank you for the detailed review Comment: Dear reviewer, We are immensely grateful to you for providing such detailed and thorough feedback on our paper. We will incorporate these discussions into the next version of the manuscript. Thanks, authors	null	NeurIPS_2024_submissions_huggingface	2,024	null	null	null	null	null	null	null	null
Human-Object Interaction Detection Collaborated with Large Relation-driven Diffusion Models	Accept (poster)	Summary: This paper introduces DIFFUSIONHOI, a novel Human-Object Interaction (HOI) detector that utilizes text-to-image diffusion models for HOI detection. It efficiently focuses on complex relationships between objects, providing a strong basis for HOI modeling. The relation-driven approach enhances image generation capabilities for HOI detection, enriching training samples for rare interactions. Additionally, it improves detector flexibility and accuracy, achieving good performance on HICO-DET and V-COCO benchmarks. Strengths: - The experiments are comprehensive. - The design of Inversion-Based HOI Modeling and Relation-Driven Sample Generation are intriguing.- - Achieving zero-shot generalization on SWiG-DET is impressive. Weaknesses: - Despite having fewer trainable parameters overall, DIFFUSIONHOI has a large total parameter count, which raises concerns about unfair comparisons with existing methods. - The authors claim to set a new state-of-the-art on HICO-DET, but the reviewer found that there are better models available, as also mentioned in the paper [79]. - Quantitative results and analysis of failure cases are needed. [79] Frederic Z Zhang, Yuhui Yuan, Dylan Campbell, Zhuoyao Zhong, and Stephen Gould. Exploring predicate visual context in detecting of human-object interactions. In ICCV, 2023 Technical Quality: 3 Clarity: 3 Questions for Authors: N/A Confidence: 3 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: Yes Flag For Ethics Review: ['No ethics review needed.'] Rating: 5 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Q1: Concerns about unfair comparisons as DIFFUSIONHOI has a larger total parameter count. A1: Though DIFFUSIONHOI contains a larger count of parameters, this does not incur heavy consumption in computation resources compared to existing work. First, the training time ($\textit{i.e.}$, 5.7 hours for relation embedding learning and 11.5 hours for HOI detection learning) is more efficient than most existing work, which can be evidenced by the comparison of training time below. Second, as the discussion provided in L321-330 and the results summarized in Table 4, the inference time of DIFFUSIONHOI is faster than the two-stage HOI detectors, and comparable to the one-stage HOI detectors without a significant gap. Based on the analysis above, we believe we render a fair comparison to existing work. \|Method\|Time (Hour)\| \|:-\|:-\| \|CDN[39]\|25.2\| \|HOTR[36]\|23.6\| \|UPT[76]\|17.9\| \|STIP[77]\|16.4\| \|GEN-VLKT[2]\|28.4\| \|HOICLIP[18]\|29.1\| \|CQL[80]\|29.7\| \|DIFFUSIONHOI\|$\textbf{5.7+11.5}$\| More importantly, as stated in L75-77, one valuable contribution of this work is the effective knowledge transfer from diffusion models without heavy fine-tuning. We address this challenge by learning task-oriented embeddings ($\textit{i.e.}$, relational embeddings) as textual prompts, and further transfer the knowledge from relation-driven sample generation and feature extraction two aspects. Considering the current trend towards leveraging large-scale pre-trained models to assist in various downstream perception tasks, we believe this work could offer valuable insights to the border community, and thus the larger count of parameters cannot be considered as the drawback of our work. --- Q2: Comparsion to [79]. A2: We provide the comparison between DIFFUSIONHOI with VQGAN and PViC with ResNet-50 as the backbone in Table 1. For PViC with Swin-L as the backbone, it leverages $\mathcal{H}$-Deform-DETR as the detector which achieves 48.7 mAP on MS COCO by running merely 12 epochs, significantly higher than DETR which achieves 36.2 mAP by running 50 epochs. For a fair comparison, we i) reimplement PCiV with ViT-L as the backbone and DETR as the detector, and ii) augment the two parallel decoders for instance and interaction detection of DIFFUSIONHOI with the technology used in $\mathcal{H}$-Deform-DETR, resulting in DIFFUSIONHOI with the decoder holds a similar detection ability to $\mathcal{H}$-Deform-DETR. The detailed results are summarized below. \|Model\| Backbone\| Detector\| Full\|Rare\|Non-rare\|AP$^{S1}_{role}$\|AP$^{S2}_{role}$\| \|:-\|:-\|:-\|:-\|:-\|:-\|:-\|:-\| \|PViC\|ResNet-50\|DETR\| 34.69 \|32.14 \|35.45\|62.8 \|67.8\| \|DIFFUSIONHOI\|VQGAN\|$\approx$DETR\|$\textbf{38.12}$\|$\textbf{38.93}$\|$\textbf{37.84}$\|$\textbf{66.8}$ \|$\textbf{70.9}$\| \|PViC\|ViT-L\|DETR\| 39.89\| 40.36\|39.98\|63.4\|68.9\| \|DIFFUSIONHOI\|ViT-L\|$\approx$DETR\|$\textbf{42.54}$\|$\textbf{42.95}$\|$\textbf{42.35}$\|$\textbf{67.1}$\|$\textbf{71.1}$\| \|PViC\|Swin-L\|$\mathcal{H}$-Deform-DETR\|44.32 \|44.61\| 44.24\|64.1\| 70.2\| \|DIFFUSIONHOI \|ViT-L\|$\approx$$\mathcal{H}$-Deform-DETR\|$\textbf{45.46}$\|$\textbf{45.78}$\|$\textbf{45.37}$\|$\textbf{67.9}$\|$\textbf{71.5}$\| As seen, for models with similar backbones, DIFFUSIONHOI achieves much better performance. For instance, 38.12 $\textit{vs.}$ 34.69, and 42.54 $\textit{vs.}$ 39.89. On this basis, as observed in the last two rows of the table above, though ViT-L is inferior to Swin-L in downstream tasks, DIFFUSIONHOI still surpasses PViC by 1.14 mAP on the Full set of HICO-DET and 3.8/1.3 mAP on the Scenario1/Scenario2 setup of V-COCO, respectively. Table 1 and $\S$4.2 will be updated accordingly. --- Q3: Quantitative results and analysis of failure cases. A3: Thank you for pointing this out. We will include the quantitative results regarding HOI detection in the Appendix for detailed analysis. Additional quantitative results for relation-driven sample generation have already been provided in Figure S1. Regarding failure case analysis, the visualized results are provided in the PDF attached with "global" response. After reviewing the predictions delivered by DIFFUSIONHOI, we found that failure cases primarily manifest in the following scenarios: i) scenes featuring only partial human bodies, such as arm or leg, which introduces challenges for person detection; and ii) chaotic scenes teeming with people, which causes occlusion and difficulties in identifying interactions. Despite these challenges, our algorithm has shown remarkable improvement over existing approaches. Additionally, the patterns of these failure cases provide valuable insights for future research. --- Thank you for your valuable time and constructive feedback. We will revise the main text with discussions and experimental results according to your review. It is welcome if there are any further questions or suggestions. --- Rebuttal Comment 1.1: Comment: Thanks for the authors' response. The rebuttal has addressed most of my concerns. I keep my positive score.	Summary: This paper introduces DIFFUSIONHOI, a new HOI detector leveraging text-to-image diffusion models. Unlike previous one-stage or two-stage models, diffusion models excel at discerning mid/low-level visual concepts as generative models and possess strong compositionality to handle novel concepts expressed in text inputs. To steer the focus of diffusion models from instance generation to the relationships between humans and objects, this paper exploits textual inversion and devises a human-object relation inversion strategy grounded in the disentanglement of HOI. Furthermore, to transfer extensive knowledge from large-scale diffusion models to assist in recognizing interactions, the paper leverages both text-prompted image generation and conditioned feature extraction capabilities of diffusion models. By embracing text-to-image diffusion models and facilitating relation-driven image generation and prompting, this method demonstrates superior performance. Strengths: This paper is well-written and easy to follow. It proposes a new diffusion-based solution, DiffusionHOI, for the human-object interaction (HOI) task. Unlike traditional one-stage and two-stage methods, DiffusionHOI benefits from controllable image generation and HOI knowledge transfer from diffusion models. Without making design changes to the HOI decoder, this method achieves superior performance compared to previous state-of-the-art models by deriving HOI-relevant features to assist in HOI detection. Weaknesses: 1. I am curious about the number of parameters in the VQGAN used as the backbone of DiffusionHOI. According to Table 1, the authors use ViT-L as their backbone, which appears to be significantly larger compared to previous works, especially those using the Res50 network. Could the authors provide more results using a smaller backbone, such as ViT-B, to offer a clearer and fairer comparison between DiffusionHOI and previous state-of-the-art models? 2. Another concern is that the improvement primarily stems from the additional knowledge provided by the powerful diffusion model. Could this be considered unfair when comparing it to previous works that do not have this significant benefit? In this case, how does this approach differ from simply adding more data for training? As I am familiar with HOI but not as familiar with diffusion models, could the authors provide more details about the data scale and parameter scale of the diffusion model used in this paper? Additionally, I look forward to feedback from other reviewers on this matter. Technical Quality: 3 Clarity: 3 Questions for Authors: Please refer to weakness. Confidence: 4 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: Please refer to weakness. Flag For Ethics Review: ['No ethics review needed.'] Rating: 7 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Q1.1: Number of parameters in the VQGAN. A1: We only utilize the encoder of VQGAN as the backbone which contains 39.2M parameters and is smaller than that of ResNet-101 (42.8M). To address your concern about unfair comparison with models using ResNet-50 as the backbone, we provide a detailed comparison with top-leading approaches using ResNet-101 below. \|Model\| Backbone\| Full\|AP$^{S1}_{role}$\|AP$^{S2}_{role}$\| \|:-\|:-\|:-\|:-\|:-\| \|QPIC\|ResNet-101\|29.90\|58.3\|60.7\| \|CDN\|ResNet-101\|32.07\|63.91\|65.89\| \|UDP\|ResNet-101\|32.62\|61.3\|67.1\| \|Iwin\|ResNet-101\|32.79\|60.9\|-\| \|GEN-VLKT\|ResNet-101\|34.95\|63.6\|66.0\| \|RmLR\|ResNet-101\|37.41\|64.2\|70.2\| \|DIFFUSIONHOI\|VQGAN\|$\textbf{38.12}$\|$\textbf{66.8}$ \|$\textbf{70.9}$\| As seen, our DIFFUSIONHOI achieves SOTA performance on both HICO-DET and V-COCO. This verifies the superiority of our proposed method. Table 1 will be updated accordingly. --- Q1.2: Comparison with large backbones. A2: The implementation of DIFFUSIONHOI relies on pre-trained Diffusion models that do not provide small backbones such as ViT-B. Therefore, we provide a detailed comparison with models using large backbones like ViT-L below to ensure a fair comparison. \|Model\| Backbone\| Full\|AP$^{S1}_{role}$\|AP$^{S2}_{role}$\| \|:-\|:-\|:-\|:-\|:-\| \|ADA-CM\|ViT-L\|38.40\|58.6\|64.0\| \|PViC\|ViT-L\|39.89\|63.4\|68.9\| \|DIFFUSIONHOI\|ViT-L\|$\textbf{42.54}$\|$\textbf{67.1}$\| $\textbf{71.1}$\| Currently, only a few models provide results on large backbones such as ViT-L, and DIFFUSIONHOI can achieve the best performance on two datasets. --- Q2.1: Unfair comparison to previous work without using additional knowledge. A2.1: Though prior work does not utilize diffusion models for knowledge transfer, they adopted other visual-linguistic models like CLIP ($\textit{e.g.}$, CTAN [45], SSRT [42], DOQ [44], GEN-VLK [2], HOICLIP [18], CQL [80], AGER[82], RmLR[83], and ADA-CM[84]). We provide a detailed comparison with approaches using pre-training models such as CLIP or MobileBERT in Table 1. It can be seen that when provided with additional knowledge, DIFFUSIONHOI can achieve a clear gap in performance ($\textit{e.g.}$, 38.12 mAP $\textit{vs.}$ 33.75 mAP for GEN-VLK[2], 35.35 mAP for CQL[80], 36.75 mAP for AGER[82], and 36.93 mAP for RmLR[83]), which highlights the effectiveness of our design. Q2.2: The difference from methods simply adding more data for training. A2.2: The primary distinction lies in three aspects: First, adding training data from other datasets is impractical in real-world applications, as it requires significant time and labor to gather annotated data, which include bounding boxes, object classes, and the interactions between humans and objects. This also imposes significant challenges in generating HOI samples using diffusion models due to the complexity of annotations. Second, the training data in our work is generated with respect to the relations between humans and objects, resulting in images that depict specific human-object interactions more accurately ($\S$3.3 Relation-Driven Sample Generation). As outlined in L178-180, for a text prompt such as "The man at bat readies to swing at the pitch," the action word "swing" is replaced with the relation embedding learned through relation-centric inversion. This steers diffusion models to focus on complex relationships rather than single objects, without heavy retraining. Third, as stated in L53-61, we achieve knowledge transfer from diffusion models in two aspects. In addition to data generation, features distinguished for each interaction can be derived with learned relational embeddings as prompts ($\S$3.4 Relation-Inspired HOI Detection). Therefore, we enhance HOI detection with diffusion models in both training data and hypothesis set ($\textit{i.e.}$, feature extraction), while adding more data can only affect the training set. Additional experiments on HICO-DET are given below to verify the effectiveness of our design. As seen, simply adding more data from SWiG-DET for training (row #2) delivers limited improvement compared to using additional data synthesized with relation-driven sample generation (row #3), or features extracted with relational embeddings as prompts (row #4). \|Method\|Full\|Rare\| Non-Rare\| \|:-\|:-\|:-\|:-\| \|BASELINE\|33.24\|30.25\|34.32\| \|+ data from SWiG-DET only \|33.89 \|31.13 \|34.72 \| \|+ Synthesized Data only\|35.49 \|36.27\| 35.02\| \|+ Relation Prompting only\|36.45\|35.78\|36.71\| \|DIFFUSIONHOI\|$\textbf{38.12}$\|$\textbf{38.93}$\|$\textbf{37.84}$\| Q2.3: More details about the data and parameter scale of the diffusion model used. A2.3: As specified in L254, DIFFUSIONHOI is built upon Stable Diffusion v1.5 pre-trained on laion2B-en, a dataset containing 2.3B samples, and subsequently fine-tuned on high-quality subsets of laion2B-en. The total parameter size is 1066M. Despite the large parameter size, as summarized in Table 4, the number of trainable parameters is significantly smaller compared to existing works ($\textit{e.g.}$, 27.6M $\textit{vs.}$ 50.4M for STIP [77], 41.9M for QPIC [37], and 42.8M for GEN-VLKT [2]), and the inference speed is comparable to leading counterparts ($\textit{e.g.}$, 9.49 FPS $\textit{vs.}$ 6.23 for iCAN [74], 7.12 for STIP [77], 3.24 for ADA-CM [84]). This indicates that the usage of diffusion models does not impose a heavy burden on our approach. Moreover, as stated in L76-77 and L365-366, this work presents an efficient way to transfer knowledge from diffusion models, with human-object relational embeddings learned via relation-centric inversion. This can offer valuable insights to the community, and further advance the exploration of large visual-linguistic pre-training models for downstream tasks. --- Thank you for providing such a comprehensive and detailed review. We are committed to revising the experiment section in accordance with your valuable suggestions. Please do not hesitate to post comments if there are any further questions we can help.	Summary: This paper tackles the human-object interaction (HOI) detection task. It aims to utilize the feature of generative models like diffusion models to help human-object interaction classification. More specifically, it utilize the inversion process in the diffusion model to learn the embedding for human-object interaction. The proposed method obtains obvious improvement over existing methods. Strengths: - The improvement on performance is significant compared with the baseline. - The figures are illustrative and helpful for understanding this paper. - The writing and overall demonstration is good. The problems to tackle are clearly explained, for example, in line 39-61. This does affect my rating. - The ablation study is though and well organized. Weaknesses: - The cost of training with the inversion process. As the inversion process is computationally costly and slow, it could affect the training speed. I would like to see more clarification or analysis over the cost of the proposed method. - The proposed model does not achieves SOTA as it claims. For example, it falls behind PViC (iccv23) with swin-L backbone by a large margin. As the proposed method have a ViT-L backbone, I think it is not fair to not mentioning PViC-swin-L and some other methods. Besides, PViC utilize no external knowledge like CLIP or SD while this method does. - The general technical design is complicated. This is not a big problem and does not affect my rating. Technical Quality: 2 Clarity: 3 Questions for Authors: - The analysis of the cost. I will look into the rebuttal and adjust my rating accordingly. Confidence: 4 Soundness: 2 Presentation: 3 Contribution: 2 Limitations: Discussion over the limitation is provided. Flag For Ethics Review: ['No ethics review needed.'] Rating: 5 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Q1: Analysis on the training cost. A1: For relation-centric inversion, unlike the original textual inversion technology that learns text embeddings within the image space, we optimize relation embeddings within the latent space by reconstructing interaction features. This results in reduced training costs. Consequently, the 117 relation embeddings in HICO-DET can be learned within 5.7 hours (23 minutes per relation embedding), using 8 Tesla A40 cards. \|Method\| Time (Minute)\| \|:-\|:-\| \|Textual Inversion\|32\| \|Relation-centric Inversion\|23\| For the main training of HOI detection on HICO-DET, as shown in Table 4, our method utilizes significantly fewer trainable parameters compared to existing work ($\textit{e.g.}$, 27.6M $\textit{vs.}$ 50.4M for STIP [77], 41.9M for QPIC [37], and 42.8M for GEN-VLKT [2]), and therefore the training process is completed in just 11.5 hours. The comparison of the whole training time with some representative work is summarized below, with all experiments conducted on 8 Tesla A40 cards. It can be observed that our method is more efficient. \|Method\|Time (Hour)\| \|:-\|:-\| \|CDN[39]\|25.2\| \|HOTR[36]\|23.6\| \|UPT[76]\|17.9\| \|STIP[77]\|16.4\| \|GEN-VLKT[2]\|28.4\| \|HOICLIP[18]\|29.1\| \|CQL[80]\|29.7\| \|DIFFUSIONHOI\|$\textbf{5.7+11.5}$\| The discussion and results above will be integrated into $\S$4.2. --- Q2: Comparison to SOTA work. A2: To solidly address your concern, we provide a comprehensive comparison with PViC below. Here PViC with Swin-L leverages $\mathcal{H}$-Deform-DETR as the detector which achieves 48.7 mAP on MS COCO by running merely 12 epochs, significantly higher than DETR which achieves 36.2 mAP by running 50 epochs. For a fair comparison, we i) reimplement PViC with ViT-L as the backbone and DETR as the detector, and ii) augment the two parallel decoders for instance and interaction detection of DIFFUSIONHOI with the technology used in $\mathcal{H}$-Deform-DETR, resulting in DIFFUSIONHOI with the decoder holds a similar detection ability to $\mathcal{H}$-Deform-DETR. The detailed results are summarized below. \|Model\| Backbone\| Detector\| Full\|Rare\|Non-rare\|AP$^{S1}_{role}$\|AP$^{S2}_{role}$\| \|:-\|:-\|:-\|:-\|:-\|:-\|:-\|:-\| \|PViC\|ResNet-50\|DETR\| 34.69 \|32.14 \|35.45\|62.8 \|67.8\| \|DIFFUSIONHOI\|VQGAN\|$\approx$DETR\|$\textbf{38.12}$\|$\textbf{38.93}$\|$\textbf{37.84}$\|$\textbf{66.8}$ \|$\textbf{70.9}$\| \|PViC\|ViT-L\|DETR\| 39.89\| 40.36\|39.98\|63.4\|68.9\| \|DIFFUSIONHOI\|ViT-L\|$\approx$DETR\|$\textbf{42.54}$\|$\textbf{42.95}$\|$\textbf{42.35}$\|$\textbf{67.1}$\|$\textbf{71.1}$\| \|PViC\|Swin-L\|$\mathcal{H}$-Deform-DETR\|44.32 \|44.61\| 44.24\|64.1\| 70.2\| \|DIFFUSIONHOI \|ViT-L\|$\approx$$\mathcal{H}$-Deform-DETR\|$\textbf{45.46}$\|$\textbf{45.78}$\|$\textbf{45.37}$\|$\textbf{67.9}$\|$\textbf{71.5}$\| As seen, with a similar backbone ($\textit{i.e.}$, ResNet-50 and VQGAN), our method surpasses PViC by 3.43 mAP on HICO-DET, 4.0 mAP on V-COCO. Next, after reimplementing PViC with DETR and ViT-L as the backbone, the performance is still inferior to DIFFUSIONHOI ($\textit{i.e.}$, 42.54 $\textit{vs.}$ 39.89 on HICO-DET). Finally, after augmenting DIFFUSIONHOI with tricks used in $\mathcal{H}$-Deform-DETR, the performance boosts to 45.46 mAP on HICO-DET. Please note that ViT-L (employed by DIFFUSIONHOI) is inferior to Swin-L (employed by PViC) for downstream tasks. We will provide a discussion in the experiment section, to highlight methods not using additional knowledge ($\textit{e.g.}$, PViC) during comparison. --- Q3: Complicated technical design. A3: This study represents the first attempt to utilize diffusion models for HOI detection, from the perspective of both data generation and conditioned feature extraction. This may bring a fundamental shift in architectural design and training strategy. However, we agree that simplicity in design is crucial for reimplementation and extension by the community. We will definitely consider this as our future work. --- Thank you so much for your careful review and suggestive comments. We will update our manuscript with the training time analysis and a more detailed comparison with the SOTA approaches. Please feel free to post your feedback if you have any further questions.	null	null	Rebuttal 1: Rebuttal: To all reviewers We express our sincere gratitude to all reviewers for their valuable time and thorough assessment of our manuscript. In response, we have carefully addressed each concern raised, and provided point-to-point clarifications which shall be integrated into the new version of our manuscript. We are gratified by the positive feedback from all reviewers, particularly regarding the significant improvement in performance, the quality of presentation, and the comprehensiveness of our experiments. Foremost among the concerns is the comparison to top-leading approaches. We provide additional experiments to ensure an all-inclusive comparison, of which our method can still achieve the best performance. Additionally, Reviewer MQdd raises concerns about the training efficiency of DIFFUSIONHOI. We appreciate this feedback and have included a detailed analysis regarding training costs to clarify this issue. Similarly, Reviewer UENt and Reviewer Xw3E highlight concerns about the parameter count of our approach. In response, we have provided additional experiments and discussions to clarify the motivation and contribution of using diffusion models for downstream tasks. Other questions, such as the complexity of the technical design of our approach, the differences from simply adding training data, more details about the diffusion model used in this work, and analysis on failure cases, have also been addressed accordingly. Pdf: /pdf/de5d221edc264c42e563e1793ef8e0fdd7370833.pdf	NeurIPS_2024_submissions_huggingface	2,024	null	null	null	null	null	null	null	null
Personalized Adapter for Large Meteorology Model on Devices: Towards Weather Foundation Models	Accept (poster)	Summary: This paper proposes an approach called LM-WEATHER that utilizes pre-trained language models (PLMs) as foundation models for on-device modeling of heterogeneous meteorological variables. LM-WEATHER enhances PLMs-equipped devices with local weather knowledge through a lightweight personalized adapter. Additionally, it leverages low-rank based transmission to fuse global knowledge among devices, enabling high-efficiency communication.This approach provides an effective solution for modeling real-world, heterogeneous, and continuously updated weather data, without requiring high resource demands. I believe this is a meaningful contribution to the field. Strengths: This paper extensively compares various time series models across different meteorological variables and weather station. The appendix section of the paper is comprehensive. Weaknesses: My first concern is that since the personalized adapter is already being used to adapt to the modeling of heterogeneous weather data collected from different weather station devices, what is the point of using the averaging operations during inter-device communication? This requires further explanation. The definitions and formulations of symbols and equations throughout the sections, from Section 3.1 Local Training to Section 4 Theorems, are disorganized and complex. Additionally, there are numerous instances of incorrect singular and plural word usage. It is recommended to rewrite these sections with increased clarity and accuracy, simplifying the content for better comprehension. Technical Quality: 3 Clarity: 2 Questions for Authors: This paper uses GPT-2 as the PLM, which is a very basic pre-trained language model. I am curious why more advanced PLMs like GPT-4 [1], Llama [2], and Vicuna [3] were not considered for better sequence modeling performance. Reference: [1] Achiam J, Adler S, Agarwal S, et al. Gpt-4 technical report[J]. arXiv preprint arXiv:2303.08774, 2023. [2] Touvron H, Lavril T, Izacard G, et al. Llama: Open and efficient foundation language models[J]. arXiv preprint arXiv:2302.13971, 2023. [3] Chiang W L, Li Z, Lin Z, et al. Vicuna: An open-source chatbot impressing gpt-4 with 90%* chatgpt quality, March 2023. Confidence: 3 Soundness: 3 Presentation: 2 Contribution: 3 Limitations: Yes Flag For Ethics Review: ['No ethics review needed.'] Rating: 7 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thank you for providing us with your valuable feedback. We have carefully considered your questions and would like to address them as below: --- Response to Weaknesses W1. Further explain the significance of using the averaging operation. We appreciate your concern and would like to clarify the importance of the averaging operation. Averaging is the most important operation in FL that knowledge interaction between clients, it serves two key purposes in our framework: * Knowledge sharing: By averaging learnable generic knowledge across devices, we prevent data silos and enable each device to benefit from the collective insights of the collaboration network, while maintaining local personalization and data privacy when sensitive data are involved. * Trade-off between personalization and generalization: The averaging operation allows devices to learn broader patterns from others, while preserving local adapter personalization. This interplay forms a robust framework for adapting to local characteristics and leveraging global knowledge, striking a balance between personalization and generalization. Our experiments (Table 1 below) show that LM-Weather with averaging outperforms LM-Weather-Local without it, highlighting the importance of averaging operation. We hope these explanations and additional experiments can address your concerns. Table 1. Results of LM-Weather and LM-Weather-Local (no averaging operation, i.e., no communication only updating the respective models locally at each client) on forecasting, $\downarrow$ is the average performance degradation relative to LM-Weather. \| ODW1T \| LM-Weather \| LM-Weather-Local \| ODW1V \| LM-Weather \| LM-Weather-Local \| \|:--:\|:--:\|:--:\|:--:\|:--:\|:--:\| \| 96 \| 42.3 \| 45.9 \| 96 \| 42.3 \| 44.5 \| \| 192 \| 44.4 \| 46.7 \| 192 \| 44.4 \| 46.1 \| \| 336 \| 45.8 \| 48.2 \| 336 \| 46.0 \| 48.8 \| \| 720 \| 49.2 \| 50.5 \| 720 \| 49.7 \| 53.2 \| \| Avg. \| 45.4 \| 47.8 \| Avg. \| 45.6 \| 48.2 \| \| Disparity \| - \| $\downarrow$ 5.29% \| Disparity \| - \| $\downarrow$ 5.70% \| W2. Recommended to rewrite Section 3.1-4 with increased clarity and accuracy, simplifying the content for better comprehension. We appreciate your careful review and constructive suggestions. We will thoroughly revise these content to streamline the presentation of symbol and equations and simplify the content for better clarity and readability. --- Response to Questions Q. This paper uses GPT-2 as the PLM, which is a very basic pre-trained language model. I am curious why more advanced PLMs like GPT-4, Llama, and Vicuna were not considered for better sequence modeling performance. We acknowledge the potential benefits of using more advanced PLM. However, Our primary goal is to establish a general framework for on-device weather sequence modeling via tuning PLMs. We chose GPT-2 for stability and cost-effectiveness, allowing for a basic yet effective demonstration of our approach. While more advanced PLMs may offer better performance, they require significant computational and storage resources, which can be a challenge for deployment on low-resource weather devices. To explore the LM-Weather performance with more advanced PLMs, we conducted additional evaluations using Llama-3B and Vicuna-7B, which show that more advanced PLMs can indeed improve performance (Tables 2-5 below). Although GPT-4 was not included due to the unavailability of its weights (we are fine-tuning based on weights), our results demonstrate the potential benefits of using more advanced PLMs. Nevertheless, the increased computational demands of these models must be carefully considered for practical deployment. We hope these explanations and additional experiments can address your concerns. The following results report the average forecasting (period [96, 192, 336, 720]) and imputation (50%) performance. Table 2. Performance of LM-Weather with different PLM backbone (ODW1T). \| PLM Backbone \| Forecasting \| Imputation \| \|--\|--\|--\| \| GPT2 (default)\| 45.4/74.6 \| 23.1/42.4 \| \| LLaMA-3b (4 layers)\| 47.2/77.6 \| 24.0/44.5 \| \| LLaMA-3b (6 layers) \| 45.0/74.3 \| 22.9/42.4 \| \| LLaMA-3b (8 layers) \| 44.0/73.5 \| 22.4/42.5 \| \| Vicuna-7b (4 Layers)\| 45.2/74.0\| 23.0/41.9 \| \| Vicuna-7b (6 Layers) \| 44.6/73.9 \| 22.7/41.8 \| \| Vicuna-7b (8 Layers) \| 43.3/72.8 \| 22.0/41.1 \| Table 3. Performance of LM-Weather with different PLM backbone (ODW1V). \| PLM Backbone \| Forecasting \| Imputation \| \|--\|--\|--\| \| GPT2 (default) \| 45.6/71.9 \| 43.7/63.8 \| \| LLaMA-3b (4 layers) \| 47.2/74.2 \| 45.7/67.2 \| \| LLaMA-3b (6 layers) \| 45.5/71.1 \| 43.2/63.5 \| \| LLaMA-3b (8 layers) \| 44.4/70.6 \| 42.1/62.0 \| \| Vicuna-7b (4 Layers) \| 45.0/71.5 \| 43.0/63.6 \| \| Vicuna-7b (6 Layers) \| 44.3/70.0 \| 41.9/61.4 \| \| Vicuna-7b (8 Layers) \| 42.5/69.7 \| 40.2/60.0 \| Table 4. Performance of LM-Weather with different PLM backbone (ODW2T). \| PLM Backbone \| Forecasting \| Imputation \| \|--\|--\|--\| \| GPT2 (default) \| 66.9/90.1 \| 38.8/61.7 \| \| LLaMA-3b (4 layers) \| 69.5/94.4 \| 40.5/65.0 \| \| LLaMA-3b (6 layers) \| 67.0/89.2 \| 38.2/61.3 \| \| LLaMA-3b (8 layers) \| 65.4/87.5 \| 36.7/59.9 \| \| Vicuna-7b (4 Layers) \| 66.1/89.2 \| 36.5/60.6 \| \| Vicuna-7b (6 Layers) \| 64.9/87.1 \| 35.4/59.2 \| \| Vicuna-7b (8 Layers) \| 62.9/85.0 \| 33.6/58.1 \| Table 5. Performance of LM-Weather with different PLM backbone (ODW2V). \| PLM Backbone \| Forecasting \| Imputation \| \|--\|--\|--\| \| GPT2 (default) \| 69.0/92.3 \| 31.1/47.0 \| \|LLaMA-3b (4 layers) \| 72.3/97.6 \| 32.4/49.8 \| \| LLaMA-3b (6 layers) \| 68.4/91.3 \| 30.9/46.2 \| \| LLaMA-3b (8 layers) \| 65.4/89.5 \| 29.6/45.1 \| \| Vicuna-7b (4 Layers) \| 68.8/90.4 \| 30.8/46.7 \| \| Vicuna-7b (6 Layers) \| 65.6/87.4 \| 29.4/45.3 \| \| Vicuna-7b (8 Layers) \| 63.2/85.5 \| 28.1/44.1 \| --- Rebuttal Comment 1.1: Comment: I appreciate the authors' response, which addresses most of my concerns. I am happy to raise my rating. --- Reply to Comment 1.1.1: Comment: Dear Reviewer C54F, Thank you for raising the rating of our paper. We are happy to have addressed your concerns. Best regards, Authors	Summary: This paper proposes LM-WEATHER, which builds upon previous work to explore further the powerful capabilities of PLMs in modelling meteorological variables. By learning sequence modelling capabilities from natural databases and applying them to on-device heterogeneity meteorological variable modelling, a lightweight adapter was developed to provide weather pattern awareness. Additionally, two real-world meteorological station datasets ODW1 & ODW2 support regional weather forecasting. Abundant experiments were designed to verify the feasibility of LM-WEATHER in modelling meteorological variables. The article work is more valuable to further validate the superiority of PLMs in modelling meteorological variables and the provision of weather station data that can be used as a baseline dataset for weather station forecasting. Strengths: 1.This work is more valuable to further validate the superiority of PLMs in modelling meteorological variables. 2.Two real-world meteorological station datasets ODW1 & ODW2, which were collected and compiled from four real-world versatile datasets for on-device meteorological variable benchmark. 3. Exploring the spatio-temporal sequence modeling of weather pattern specificity with high distributional similarity to provide a viable solution paradigm for sparse and heterogeneous weather station data. 4. From the experimental results, this method can further enhance the possibility of maintaining confidence in its future development in the presence of unconstrained data and PLM models. Weaknesses: 1. Figure 1 does not intuitively reflect the contributions of the paper. The Fig.1F mentioned in line 68 seems to be missing; additionally, the explanation for Figure C should be Task Adapter Generation, which is hard to understand. It is recommended to keep the labels a, b, … in the figure consistent with A, B, … in the text. 2. The paper discusses sequence modeling in the time dimension of stations. Can this method be validated for spatiotemporal modeling? 3. The experiments in the paper mainly focus on the accuracy of forecasting time. Can further validation be done on the timeliness and stability of the forecasts? 4. To better present the experimental data, some statistical charts can be drawn to visually demonstrate the advantages of LM-WEATHER in meteorological variable forecasting. Technical Quality: 4 Clarity: 3 Questions for Authors: 1. In the Zero-Shot Learning (Out of Distribution Modeling) experiments mentioned in line 282, the domain transfer experiments for regional forecasting on the OWD1 dataset showed that the transfer performance was not as good as the forecasting performance of GPT-4. Can you explain the reason for this? 2. In Theorem 4.2, how is it demonstrated that Low-Rank Matrices ensure privacy? Are there any related research papers on this topic? Can you provide a detailed explanation? 3. Currently, there are many sparse station forecasts. Can interpolation be performed based on sparse station data to further extend the model to spatial dimension forecasting and enhance its generalizability? 4. LM-WEATHER seems to outperform current methods. Can it maintain stability in long-term forecasting, and what is the maximum forecasting time it can achieve? 5. In Table 6, line 305, it appears that having more trainable parameters is better. If the comparison experiment groups had the same number of trainable parameters, what would the results be? Confidence: 4 Soundness: 4 Presentation: 3 Contribution: 3 Limitations: Regarding limitation 1, further validating the impact of increased data scale on model performance using large-scale ERA5 data would be beneficial. Utilizing ERA5 as pre-training data to model invariance in meteorological forecasting and learning fundamental meteorological variable patterns, followed by fine-tuning with real-world data, could be a highly valuable endeavor. Flag For Ethics Review: ['No ethics review needed.'] Rating: 7 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We greatly appreciate your insightful comments. We've carefully considered and addressed your concerns as follows *(additional results are in Global Rebuttal PDF file): --- Response to Weaknesses* W1. Some confusion about Figure 1. We will revise Fig. 1 to clarify the explanation of panel C and unify labels to improve overall clarity and readability, and correct the typo referencing Fig. 1F to Fig. 1a. W2. Effectiveness of the method in spatio-temporal modeling. To further demonstrate the effectiveness of LM-Weather for spatio-temporal modeling, we conducted additional experiments with a variant, LM-Weather-ST, which aggregates site data for centralized modeling. We compared its performance to the original LM-Weather and classical spatio-temporal forecasting baselines (STGCN, ASTGCN, TCGN, and CLCRN). The results *(Table 3, Global Author Rebuttal PDF file)* show that LM-Weather-ST achieves optimal performance, while the original LM-Weather still significantly outperforms the baselines despite being trained in a decentralized manner. This suggests that LM-Weather has strong potential for spatio-temporal modeling tasks. Future work will explore incorporating server-side graphs to model relationships between devices, further enhancing LM-Weather's benefits for spatio-temporal modeling. W3. Timeliness and stability of forecasts * Timeliness: We provide inference times for various forecasting periods (Table 4, Global Author Rebuttal PDF file). When combined with the communication time of 0.29 s, these results demonstrate that LM-Weather achieves efficient prediction times, ensuring timely delivery of forecasts. * Stability: All experiment have been conducted five times. We present the standard deviations of LM-Weather and the runner-up model *(Table 5, Global Author Rebuttal PDF file), which indicate that LM-Weather can maintain stable performance over multiple experiments. Furthermore, we assessed longer-term forecasting tasks (1080, 1440, 1800, and 2160 hours) with a fixed input length of 192 (Table 6, Global Author Rebuttal PDF file). LM-Weather not only achieves optimal performance but also exhibits relatively stable error variations compared to representative baselines, further validating its stability. W4. It is recommended that statistical charts be added.* We will add visualizations to each of the core experiment results in the revision. --- Response to Questions Q1. Why region transfer performance on ODW1 is not as good as that of GPT4. LM-Weather outperforms FL-GPT4TS in most zero-shot scenarios, but the specific case of OWD1T -> OWD2V highlights an interesting trade-off. The difference lies in the distinct local tuning and global communication strategies employed by each method. LM-Weather uses learnable low-rank parameters to balance generic sequence modeling with personalized regional insights, whereas FL-GPT4TS unfreezes LayerNorm layers for more direct adaptation of the PLM's knowledge. This allows FL-GPT4TS to update and communicate more parameters, potentially enabling greater knowledge sharing in modeling heterogeneous weather data, but at the cost of increased communication overhead. Additionally, significant non-iid between OWD1T and OWD2V contribute to this difference. Q2 How the low rank matrix of Theorem 4.2 ensures privacy. We provide additional description and proofs for this, which can be found in *G1, Global Author Rebuttal. Q3. Can interpolation based on sparse stations be used to extend the model to spatial dimension prediction and improve generalizability?* While spatial interpolation on sparse stations can increase the scale of training data, it is not an effective approach for extending the model's forecasting range and enhancing its generalizability due to: * Spatial interpolation of site data does not ensure the accuracy of the interpolated values, particularly when dealing with very sparse datasets like ODW2. * Even when interpolating between geographically proximate stations, the biases introduced by geographical differences cannot be fully compensated by interpolation alone, especially when multiple variables are involved. The complex interactions between variables and the influence of regional meteorological features render spatial interpolation ineffective for sparse station data. To further validate our statement, we conducted experiments on OWD1T and OWD2T using three spatial interpolation algorithms (Kriging, IDW, Spline). The results *(Tables 7&8, Global Author Rebuttal PDF file)* show that spatial interpolation performs poorly in sparse stations and fails to improve performance. Q4. Stability and maximum forecasting time. Regarding stability, please refer to our response for W3. As for the maximum forecasting time, while it is theoretically possible to extend the forecast horizon indefinitely by increasing computational resources, when the input data length remains fixed, longer forecast periods introduce greater uncertainty, leading to decreased performance. As discussed in W3, we observed a significant decline in performance when extending the forecast period from 720 to 2160 hours. This degradation highlights that, although technically feasible, very long forecasting times may not be practically useful due to the increased uncertainty and reduced accuracy. Q5. Results when parameters are the same in the ablation experiment. While it's true that increasing trainable parameters often boosts performance, LM-Weather-G/H's 4.5% gain comes at a steep cost: a five-fold increase in parameters. This trade-off is suboptimal. Our goal is to strike a balance between performance and efficiency. Notably, when parameter of the experiment group are similar, the original LM-Weather consistently outperforms its variants *(Table 9, Global Author Rebuttal PDF File)*, demonstrating its optimal trade-off between performance and efficiency.	Summary: The paper aims to develop weather foundation models by leveraging the on-device data on many distributed sensors. Specifically, a federated learning and low-rank adaption mechanism have been applied to a time-series-based foundation model training on Meteorology data. Strengths: 1) The paper proposes a new way to develop weather foundation models by leveraging many distributed devices with collected Meteorology data. It is worth noting that weather foundation models are an emerging research domain critically important to geoscience and tackling climate change. 2) The proposed solution is technique sounds. The procedure of time series processing, LoRA-based adaptation, and federated learning are seamlessly integrated to implement its goal. 3) The paper provides sufficient details with an appendix and open-sourced codes. It is an essential part to ensure the reproducibility of this work. 4) The paper conducted a comprehensive experiment for comparison and evaluation of the few/zero-shot learning scenarios on the foundation model. Weaknesses: 1) The paper needs a strong justification to describe the motivation for using a pre-trained language model as a basis for the weather foundation model. Are there any texts in the Meteorology dataset? 2) The paper’s contents organization could be improved. For example, the figure 1 is too complex to understand. In Eq 2, it would be better to write the function as F(theta, D). Because F(theta \| D) is easily confusing with the conditional distribution. Moreover, some technique details should not be placed in the appendix. It is a usual assumption that the main paper is self-contained so that the readers can understand the paper without reading the appendix. 3) In the contribution part (lines 73 - 94), the last two items look like the advantages of the proposed method rather than contributions. Technical Quality: 4 Clarity: 3 Questions for Authors: 1) Does the proposed method only focus on time series data? 2) Would you please highlight the main research question in plain language? It seems there are many components mixed in the paper. 3) In Section 3, what’s the federated aggregation mechanism? Is there any pseudo code to describe the algorithm? 4) Can you please explain the token embedding for time series data? Is the token to be described as a special pattern of a subsequence of the time series? How to define the vocabulary of the tokens? Confidence: 5 Soundness: 4 Presentation: 3 Contribution: 4 Limitations: N/A Flag For Ethics Review: ['No ethics review needed.'] Rating: 8 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thank you for providing us with your valuable feedback. We have carefully considered your questions and would like to address them as below: ------- Response to Weaknesses W1. Justification for using PLM and clarity on weather dataset. Although our dataset consists solely of time series data without any textual content, using a PLM as the basis of the weather foundation model offers substantial flexibility and an optimal trade-off between performance and costs. Our motivations can be summarized by: * Limitations of Existing Climate Foundation Models: Current climate foundation models, such as ClimaX [1] (73B parameters) and Pangu-Weather [2] (up to 100B parameters), demand vast computational resources and extensive reanalysis data volumes (over 100 TB) for pre-training, making them unsuitable for deployment on low-resource meteorological equipment due to prohibitive operational costs and computational demands. * Sequence Modeling Capabilities of PLMs: Despite being initially trained on text, their advanced sequence modeling capabilities, proven on time series analysis [3]. This eliminates the need for training models from scratch with high-quality weather observational data, which is often scarce and expensive to acquire as opposed to reanalysis data. Consequently, adopting a PLM not only employs its powerful sequence modeling abilities but also drastically cuts down on computational expenses and deployment costs when fine-tuning for weather sequence modeling. W2. The paper's organization and clarity can be improved, e.g., notation, figure, and placement of technical details. Thank you for your careful review and constructive comments. In our revision, we will recheck all notations in the paper to avoid confusion. Additionally, we will adjust the layout of the main text to incorporate important content from the appendix, enhancing readability. Your feedback will significantly improve the quality of our paper! W3. The last two items in contribution part appear to be advantages rather than contributions. We will adjust the contribution part to improve clarity of expression. -------- Response to Questions Q1: Does the proposed method only focus on time series data? Yes, LM-Weather currently specializes in processing weather time series. Utilizing a PLM can bring significant benefits in terms of flexibility and cost-effectiveness, as detailed in our response for W1. To further enhance the generality and applicability of LM-Weather, we plan to expand our dataset to include multi-modal data such as descriptive weather texts, satellite cloud imagery, and radar data, ultimately developing a comprehensive multi-modal weather sequence analysis framework. Q2. Highlight the main research question. The main research question can be summarized as: How can we adapt pre-trained language models to efficiently and effectively model weather sequences on local devices in a decentralized manner, while addressing the challenges of data diversity and limited computing resources? This question seeks to explore: * Adaptation: How can complex models originally designed for language processing be tailored to understand weather sequences? * Personalization: Can these models provide tailored weather predictions that suit the specific conditions and data available at individual locations and devices? * On-device efficiency: Can this approach operate effectively on local devices, ensuring privacy and minimizing resource use despite the constraints of limited computational power and storage? Q3. Federated aggregation mechanism and pseudo code. Our aggregation mechanism is based on FedAvg, where only low-rank parameters are shared and averaged, weighted by each client's sample size ratio. We provide the pseudo code below, which will be added to the revision for clarity. Algorithm 1. low-rank parameter-based aggregation mechanism (server-side, all clients participate in the training). \|---\| Input: Low-rank parameters from each client {$\theta_{l,1}, \theta_{l,2}, ..., \theta_{l,k}$}, number of samples per client {$n_1, n_2, ..., n_k$} Output: Aggregated low-rank parameters $\theta_{l}$ --- 1. Receiving low-rank parameters from clients 1. Calculate the total number of samples across clients $N = \sum_{i=1}^k n_i$ 2. Aggregating low-rank parameters: $\theta_{l}' = \sum_{i=1}^k \frac{n_i}{N} \theta_{l, i}$ 3. Broadcasting updated $\theta_{l}'$ to clients to continue next communication round. --- Q4. Explain the token embedding? Is the token to be described as a special pattern of a subsequence of the time series? How to define the vocabulary of the tokens? The token embedding module aims to transform subsequences of weather time series into a higher-dimensional space using a convolutional layer. Each transformed subsequence is represented as a continuous vector, known as a `token`. This process allows the model to capture localized patterns within a specific receptive field. Each token encapsulates patterns from a subsequence of the time series, capturing local dynamics. The kernel size determines the subsequence length, combining each point with its immediate neighbors to form a token. These tokens capture important temporal dependencies. Unlike traditional NLP, the `vocabulary` here is formed by the diverse outputs of the convolutional layer, representing a range of patterns in the input data. This vocabulary is continuously defined, shaped by the convolutional filters that learn to identify and abstract temporal features during training. --- Reference: 1. ClimaX: A foundation model for weather and climate. arXiv 2023. 2. Accurate medium-range global weather forecasting with 3D neural networks. Nature 2023. 3. "Large language models for forecasting and anomaly detection: A systematic literature review. arXiv 2024. --- Rebuttal Comment 1.1: Comment: Many thanks for your rebuttals. After carefully checking it, my questions have been answered and I will keep my score. --- Rebuttal 2: Comment: Thank you for your confirmation. We are pleased that our rebuttals have addressed your concerns and appreciate your continued support for our paper. Best regards, Authors	Summary: This paper introduces LM-WEATHER, a framework leveraging pre-trained language models (PLMs) for on-device meteorological variable modeling. The framework integrates personalized adapters into PLMs, enhancing their ability to handle heterogeneous weather data efficiently. Key contributions include superior performance in forecasting and imputation tasks, minimal communication overhead, and maintaining privacy during client-server interactions. Experiments on real-world datasets demonstrate LM-WEATHER's effectiveness across various scenarios, including few-shot and zero-shot learning. Strengths: S1. LM-WEATHER introduces a novel approach by integrating personalized adapters into PLMs, allowing for efficient and accurate on-device meteorological modeling. S2. The framework addresses heterogeneity in weather data and demonstrates robust performance in various tasks and scenarios. S3.The approach offers a promising solution for personalized weather modeling on resource-constrained devices, potentially benefiting various applications in meteorology and related fields. Weaknesses: W1.In the communication section, it is mentioned that only the low-rank matrix parameters are transmitted between the client and server, but specific details about the transmission mechanism and strategy are missing. This includes transmission frequency, bandwidth consumption, and potential issues in practical applications. These missing details may prevent readers from fully understanding the feasibility and effectiveness of the communication strategy. W2. In the parameter adapter section, it describes integrating the generated adapters with the original input through FFN to produce the final weather sequence predictions. However, the specific architecture of the FFN, how its parameters are configured, and why this particular architecture was chosen are not detailed. The lack of these key details might leave readers questioning the mechanism and effectiveness of the parameter adapter. W3. In Section 4, while Theorems 4.1 and 4.2 propose the rationality of time series decomposition and privacy assurance through low-rank matrix exchange. There is no introduction and description about the background and application of these Theorems. Also, the descriptions are too brief, making it difficult for readers to fully understand the theoretical basis. W4. The paper's experimental results are heavily supplemented by a 37-page appendix, which can hinder readability and accessibility of the core content. Important experimental details and results should be better integrated into the main body of the paper to facilitate easier comprehension and evaluation by the readers. Technical Quality: 3 Clarity: 2 Questions for Authors: Q1.Could you provide more detailed proofs and theoretical validation for the decomposition rationality and privacy assurances mentioned in the theorems? Q2.Could you clarify the potential confusion in the notations n¬i and nk in Section 2.2, and ensure the formulas are correctly presented? Q3.Could you provide detailed information on the architecture and parameter configuration of the FFN used in the parameter adapter, and explain why this architecture was chosen? Q4.Could you provide more detailed information about the datasets, particularly whether data from extreme climatic regions(e.g., tropical rainforests and deserts) are included? If not, how do you plan to expand these datasets to cover more global climatic conditions? Confidence: 4 Soundness: 3 Presentation: 2 Contribution: 3 Limitations: 1 The authors acknowledge the limitations in the diversity of the datasets and the potential challenges in modeling specific climatic conditions not covered in the current study. Future work should focus on expanding dataset coverage and providing more detailed theoretical and empirical validation. Additionally, the practical applications and limitations of few-shot and zero-shot learning capabilities should be more thoroughly explored and documented. 2 It would be much better integrate the essential experimental details and results into the main body of the paper to improve readability and comprehension? 3 I have questioned whether there will be a powerful foundation model for time series prediction. If the assumption is not true, the application of the proposed solution is limited. Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thank you for providing us with your valuable feedback. We have carefully considered your questions and would like to address them as below *(additional results are in Global Rebuttal PDF file): --- Response to Weaknesses and Question (& means joint response)* W1. Details about the transmission strategy. Our framework employs a synchronous federated learning strategy, where the client and server communicate at a fixed frequency (every 20 iterations). The process of transmission includes: * 1. client-side updating: each participant uploads low-rank parameters of their local model. * 2. server-side aggregation: the server aggregates uploaded parameters via federated averaging. * 3. low-rank parameter broadcasting: the server broadcasts aggregated parameters to each participant for the next round. By communicating only low-rank parameters (0.38M parameters), we minimize bandwidth consumption. We acknowledge potential issues in practical applications, including communication overhead, latency, and data privacy. Our method addresses these concerns by transmitting only low-rank parameters, thereby safeguarding privacy and minimizing communication overhead. We will clarify this key aspect of the transmission strategy in the revision to enhance understanding. W2 & Q3. Detailed information of the FFN in parameter adapter. Our parameter adapter is a straightforward FFN with a single linear layer, converting the PLM's output (input channels, e.g., 768 for GPT2, 4096 for Llama-3B) to align prediction horizons of [96, 192, 336, 720] (output channels). We chose this configuration to optimize the balance between performance and computational efficiency. Our additional experiments *(Table 1, Global Author Rebuttal PDF file)* show that adding more layers only marginally improves performance (0.21% with two or three layers) and even decreases with four layers, while significantly increasing computational demands. Therefore, the single-layer setup is the most cost-effective, providing an optimal trade-off between efficiency and performance. W3 & Q1. Introduction and description of backgournd, and more detailed proofs and theoretical validation for theorems. We provide additional description and proofs for theorems, which can be found in G1, Global Author Rebuttal. W4. Important details and results should be better integrated into the main body. Thank you for your constructive feedback. We will revise the layout to highlight the core results in the main body. Q2. Potential confusion in the notations $n_i$ and $n_k$. We apologize for the confusion caused by our typo. The correct Eq. 2 is: $n_i$, $F(\theta) := argmin \sum_{i=1}^{N} \frac{n_i}{n} F_i(\theta_i; \lbrace D_i \rbrace)$, where $n_i$ represents the $i$-th client. We will carefully revise and clarify these notations to enhance the readability. Q4. Does the dataset have detailed information on extreme climate regions, and if not, what are the plans for expanding the dataset? Including more details about specific extreme climatic regions indeed benefits the expansion of our dataset's potential application scenarios. Current version comprises real observations from various cities, with limited access to detailed topographic or regional information. Recognizing the potential benefits of expanding this coverage, we are actively working to include a broader range of climatic conditions. Specifically, we plan to introduce multimodal weather data, such as radar echoes and satellite cloud images, into our dataset's corresponding regions. Additionally, by collaborating with climate experts, we will provide textual descriptions of weather trends to create a comprehensive multimodal dataset. ----------- Response to Limitations L1: Future work: expanding the dataset, deepening theoretical and empirical validation, and exploring the applications and limitations of few/shot-learning. Thank you for your insightful comments, which have provided a clear direction for our future work. As the response for Q4, we plan to introduce additional data modalities to enhance the comprehensiveness and applicability of current datasets. In future work, we will conduct more detailed theoretical and empirical analyses based on the newly constructed multimodal dataset. Additionally, we will further explore the limitations of these observations in practical applications and few/zero-shot scenarios. L2: It would be much better integrate the essential experimental details and results into the main body. We will integrate essential details and results into the main body in the revision to enhance readability and adjust the layout accordingly. L3: Whether there will be a powerful foundation model for time series prediction. If the assumption is not true, the application of the proposed solution is limited. Our approach is indeed based on the assumption that there is already a good foundation model. We believe that this assumption is reasonable because there are already many well-established foundation models that can be used in time series forecasting tasks, e.g., MOMENT[1], Timer[2]. However, we also understand the your concern that without a powerful foundation model, our method may suffer. To address this concern, we used several PLMs that weak than GPT-2 (e.g., Bert [3], CTRL [4]) to evaluate the performance of LM-Weather, and the results *(Table 2, Global Author Rebuttal PDF File)* show that LM-Weather can achieve good performance when the assumption of a powerful foundation model is not hold, and significantly outperforms other regular model. --- Reference: 1. Moment: A family of open time-series foundation models. arXiv 2024. 2. Timer: Transformers for time series analysis at scale. arXiv 2024. 3. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv 2018. 4. Ctrl: A conditional transformer language model for controllable generation. arXiv 2019. --- Rebuttal Comment 1.1: Comment: Thanks a lot for your response and revisions. I have raised my score to reflect such improvement. --- Reply to Comment 1.1.1: Comment: Dear Reviewer Jjnz, Thank you for upgrading your score. We appreciate your constructive suggestions for improving our paper. Best, Authors	Rebuttal 1: Rebuttal: We thank the reviewers for their valuable feedback. *Additional experiments are included in the attached PDF file, indexed as follows: Reviewer Jjnz: Table 1, Table 2 * Reviewer mBbn: No additional experiments supplements. * Reviewer tT2F: Table 3 - Table 9 * Reviewer C54F: Additional experiments are provided in the box corresponding to Reviewer C54F. Below are our general replies to the general concerns from the reviewers: *G1. (Reviewer Jjnz, Reviewer tT2F)* Detailed introduction, description, and proof for Therorems.** * Introduction and Background for Therorem 4.1: Decomposing time series into trend and seasonal components helps model underlying patterns. This theorem states that non-orthogonal trend and seasonal components cannot be fully separated by orthogonal bases. Since self-attention layers learn orthogonal transformations similar to PCA, applying attention to raw time series is ineffective at separating non-orthogonal components, highlighting the need for manual decomposition. * Introduction and Background for Theorem 4.2: The theorem is built upon federated learning, where an external attacker can exploit client parameters to compromise sensitive information. We demonstrate that LM-Weather ensures privacy by limiting uploaded parameters to a small, low-rank subset, making it challenging for an attacker to reconstruct client data. This guarantees the protection of sensitive information. Additional Proof for Theorem 4.1 - Self-attention Layer is a Nonlinear PCA: The Self-attention can be viewed as a non-linear PCA, with the computation process represented as: $Attention(Q, K, V) = softmax(QK^T / \sqrt{d}) V$. This can be decomposed into: $Attention(Q, K, V) = \sum_i \alpha_i v_i$, where $\alpha_i$ is the weight coefficient, and $v_i$ is the $i^{th}$ component of the Value. The weight coefficient $\alpha_i$ can be computed as: $\alpha_i = \frac{\exp\left(\frac{QK_i^T}{\sqrt{d}}\right)}{\sum_j \exp\left(\frac{QK_j^T}{\sqrt{d}}\right)}$. This is similar to the SVD of PCA: $X = U\Sigma V^T = \sum_i \sigma_i u_i v_i^T$. where $X$ is the input data, $U$ is the left singular vector, $\Sigma$ is the singular value matrix, $V$ is the right singular vector, $\sigma_i$ is the singular value, and $u_i$ and $v_i$ are the $i^{th}$ components of the left and right singular vectors, respectively. Comparing the two formulas reveals the similarity between Self-Attention and PCA. Combined with Theorem 1 in [1], this suggests that self-attention layer in PLMs functions closely to PCA (only learn orthogonal transformation) and can't automatically decompose time series into trend and seasonal components (non-orthogonal) without manual intervention. Additional Proof for Theorem 4.2 - Model Indistinguishability: Consider two clients with different local models $\mathcal{M}_i$ and $\mathcal{M}_j$ parameterized by $\theta_i$ and $\theta_j$. Let $L(\mathcal{M})$ denote the low-rank matrix from model $\mathcal{M}$, where $B \in \mathbb{R}^{d\times r}$ and $A \in \mathbb{R}^{r\times d}$, with $r\ll d$. Then, for any polynomial-time attacker $\texttt{Adv}$, the following holds: $\|Pr[\texttt{Adv}(L(\mathcal{M}_i)) = 1] - Pr[\texttt{Adv}(L(\mathcal{M}_j)) = 1]\| \leq \epsilon$, where $\epsilon$ is a small positive number. This implies that the attacker cannot distinguish between client models from a shared low-rank update. Proof. LoRA aims to find the low-rank approximation: $ min \|\|\theta - \theta_0 - BA\|\|_F$, where $\theta_0$ is the initial weight. Consider two models $\mathcal{M}_i$ and $\mathcal{M}_j$ with weights differing by $\Delta\theta = \theta_i - \theta_j$ from different clients. The corresponding LoRA matrix is: $$L_i = B_iA_i \approx \Delta \theta_i = \theta_i - \theta_0, L_j = B_jA_j \approx \Delta \theta_j = \theta_j - \theta_0.$$ According to matrix approximation theory, for the best approximation with rank $r$, the upper bound on the error is: $\|\|\Delta \theta_i - L_i\|\|_ F \leq \sigma_{r+1}(\Delta \theta_i)$ where $\sigma_{r+1}(\Delta \theta_i)$ is the $(r+1)$-st singular value of $\Delta \theta_i$. By Johnson-Lindenstrauss Lemma [2], for any $\epsilon > 0$, there exists a mapping $f: \mathbb{R}^d \rightarrow \mathbb{R}^k$ with $k = O(log(n)/ \epsilon^2)$, such that for any $x,y \in \mathbb{R}^d$. LoRA can be regarded as such a mapping. Assuming $\|\|\Delta \theta_i - \Delta \theta_j\|\|_F \leq \delta$ and using the trigonometric inequality: $$\|\|L_i - L_j\|\|_F \leq \|\|L_i - \Delta \theta_i\|\|_F + \|\|\Delta \theta_i - \Delta \theta_j\|\|_F + \|\|\Delta \theta_j - L_j\|\|_F$$ $$ \leq \sigma_{r+1}(\Delta \theta_i) + \delta + \sigma_{r+1}(\Delta \theta_j).$$ Let $\varepsilon' = \sigma_{r+1}(\Delta \theta_i) + \sigma_{r+1}(\Delta \theta_j)$, we get: $$\|\|L_i - L_j\|\|_F \leq \delta + \varepsilon'$$ For any polynomial-time attacker $\texttt{Adv}$, its ability to distinguish between $L_i$ and $L_j$ is restricted to the difference in their Frobenius paradigms. We can define a function $f$ such that: $$\left\|Pr[\texttt{Adv}(L(\mathcal{M}_i)) = 1] - Pr[\texttt{Adv}(L(\mathcal{M}_j)) = 1] \right\| \leq f(\|\|L_i - L_j\|\|_F),$$ where $f$ is a monotonically increasing function representing the attacker's capability. Finally, we get: $\|Pr[\texttt{Adv}(L(\mathcal{M}_i)) = 1] - Pr[\texttt{Adv}(L(\mathcal{M}_j)) = 1]\| \leq \epsilon$. When $\delta$ and $\varepsilon'$ are small enough, the right-hand side is smaller than the intended $\varepsilon$, ensuring that an attacker cannot reverse-engineer local parameters and data, thus preserving privacy through low-rank parameters-only communication. We hope that the above additions will address the reviewers' concerns. We will add these additional introductions and proofs to the theorems in our paper during the revision to improve readability. --- Reference: 1. One fits all: Power general time series analysis by pretrained lm. NeruIPS 2023. 2. On variants of the Johnson–Lindenstrauss lemma. Random Structures & Algorithms 2008. Pdf: /pdf/5f2236920d93926f4dc7ba2d57f5759007db946f.pdf	NeurIPS_2024_submissions_huggingface	2,024	null	null	null	null	null	null	null	null
Compressing Large Language Models using Low Rank and Low Precision Decomposition	Accept (poster)	Summary: This paper proposes a framework that combines quantization and low-rank approximation. Given a neural net weight matrix $W\in \mathbb{R}^{n\times d}$, it considers the following decomposition: $W=Q+LR$, where $Q$ is a quantized sketch of $W$ with very few bits (2 or 4 in their experiments), and $L\in \mathbb{R}^{n\times k}, R\in \mathbb{R}^{k\times d}$ are low-rank factors that serve as a "correction" to the quantization sketch $Q$. They are usually quantized with more bits (4 or 16). To compute respective $Q, L, R$, the authors propose an alternating minimization framework: starts with $L, R=0$, update $Q$ as $Q_t=Q_Q(W-L_{t-1}R_{t-1})$, then solve for $L_t, R_t$ by solving the rank-constrained regression against $W-Q_t$ and $X\in \mathbb{R}^{m\times d}$, where $X$ is the calibration dataset. They further show their algorithm could be used in conjunction with LoRA and fine-tuning the randomized Hadamard transform. Extensive experiments show that their proposed algorithm gives low perplexities and high accuracies when compared against earlier algorithms. Strengths: This paper proposes a method that is a good mixture of quantization and low-rank approximation, as simply using only one of quantization sketch or low-rank approximation intuitively "misses" an important part. The algorithm proposed in this paper is a natural and simple alternating minimization framework. This paper also contains a theoretical analysis for approximation error under assumptions on the target rank. Empirically, it is also shown the proposed algorithm has good performance under small rank and bit precision. The paper is also well-written. Weaknesses: Overall, I don't see big weaknesses of this paper. One potential direction for improvement is theoretically, the alternating minimization framework could possibly be sped up given $k\ll n, d$, especially the rank-constrained regression, see [1]. On the other hand, it is not clear how to integrate these possible algorithmic pieces into post-processing a model weight, so I think the approach and runtime analysis of this paper are fair. [1] Gu, Song, Yin and Zhang. Low Rank Matrix Completion via Robust Alternating Minimization in Nearly Linear Time. ICLR'24. Technical Quality: 3 Clarity: 3 Questions for Authors: If we allow $L$ and $R$ to be computed to full bit-precision, could the statement / result of Theorem 4.1 be simplified? Confidence: 3 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: Yes. Flag For Ethics Review: ['No ethics review needed.'] Rating: 7 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: > If we allow $\mathbf{L}$ and $\mathbf{R}$ to be computed to full bit-precision, could the statement/result of Theorem 4.1 be simplified? Yes, in Thm. $4.1$, the error from quantizing the low-rank factors is captured in the additive $\epsilon$ term. When $\mathrm{B_L} \to \infty, \mathrm{B_R} \to \infty$, this error $\epsilon \to 0$, and we get $\frac{1}{nm}\mathbb{E}\left\lVert\mathbf{(Q + LR - W)X}\right\rVert_F^2 \lesssim \frac{4d\lambda_{max}\mathrm{R}^2}{\pi(2^{\mathrm{B_Q}}-1)^2}\left(1 - \frac{k}{2n}\right)^2$. Here, the only source of error is the finite precision of the $\mathbf{Q}$ matrix, determined by $\mathrm{B_Q}$. Furthermore, if the low-rank factors $\mathbf{L, R}$ are not required to be quantized, CALDERA simplifies to LoftQ -- the only difference being CALDERA uses LDLQ quantizer, whereas LoftQ uses NF4 quantizer. > One potential direction for improvement is theoretically, the alternating minimization framework could possibly be sped up given, especially the rank-constrained regression Thank you for pointing out this interesting work. We agree that solving rank-constrained regression can be sped up using sketching techniques from Gu et. al. [1]. A careful analysis is required to ensure that the error from sketching does not accumulate across the alternating iterations. A study of the efficacy of sketching in solving the non-convex optimization problem (1) is worth investigating. [1] Gu et. al., Low Rank Matrix Completion via Robust Alternating Minimization in Nearly Linear Time (ICLR, 2024). --- Rebuttal Comment 1.1: Comment: I thank the authors for answering my questions and including additional experiments. I'll keep my score as is.	Summary: This paper introduces CALDERA, a new post-training compression algorithm for large language models (LLMs). CALDERA uses the inherent low-rank structure of LLM weight matrices by approximating them via a low-rank, low-precision decomposition $W \approx Q + LR$, where $L$ and $R$ are low-rank factors with quantized entries. CALDERA provides an effective way to compress large language models by exploiting their low-rank structure, enabling efficient distribution and deployment of LLMs on resource-constrained hardware while maintaining strong performance. Strengths: 1. CALDERA introduces a new post-training compression technique for LLMs that exploits their inherent low-rank structure, setting it apart from existing methods. Also, the paper provides theoretical upper bounds on the approximation error of CALDERA using a rank-constrained regression framework, lending credibility to the approach. 2. Experimental results demonstrate that CALDERA outperforms existing post-training LLM compression techniques in the low-bit regime (less than 2.5 bits per parameter), highlighting its effectiveness. By enabling efficient distribution and deployment of LLMs on resource-constrained hardware, CALDERA can make LLMs more accessible and promote their broader adoption. Weaknesses: This paper is generally well-written. Here are some minor weaknesses: 1. The performance of CALDERA depends on the choice of target rank and quantization bit budget. The paper does not provide a systematic way to determine the optimal values for these hyperparameters. Is it possible to apply meta-learning here to decide the optimal values? 2. The experiments focus on compressing LLaMa models. It would be valuable to see how CALDERA performs on a wider range of LLMs to assess its generalizability. 3. As mentioned by the authors, the iterative nature of CALDERA's optimization process may require more computational resources compared to simpler compression methods, although this is a one-time cost. Technical Quality: 4 Clarity: 3 Questions for Authors: How does CALDERA perform compared to other compression approaches? Is it possible to combine CALDERA with these techniques? Confidence: 2 Soundness: 4 Presentation: 3 Contribution: 3 Limitations: Please see weaknesses. Flag For Ethics Review: ['No ethics review needed.'] Rating: 7 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: > wider range of LLMs We have performed quantization experiments on the Mistral 7B model, which can be found in Table 2 of the global response PDF. The PPLs obtained using CALDERA are consistently lower than QuIP# (no RHT fine-tuning) for comparable average bits. For future work, with additional compute, we intend to continue quantizing more such models. > other compression approaches ... possible to combine CALDERA Thank you for another interesting question. It is indeed possible to combine CALDERA with other existing model compression techniques. A primary motivation of CALDERA of is to jointly solve two popular approaches in LLM compression, namely, low-rank decomposition and quantization -- both of which are usually treated as independently in existing works. The same motivation can indeed be used if one wants to adapt CALDERA to incorporate other compression approaches, such as sparsity, quantization-aware training (QAT), knowledge distillation, etc. For instance, the quantized $\mathbf{Q}$ matrix can be additionally pruned, making it sparse in each CALDERA iteration, further compressing the model. Moreover, provided additional compute, the quantized entries of our CALDERA decomposition can also be subjected to QAT using a straight-through estimator. QAT can either be done for fine-tuning on downstream tasks. Alternatively, QAT can also be done by performing knowledge distillation with an uncompressed teacher model, similar to what is done in LLM-QAT (ref. [1] below). [1] Liu et. al., LLM-QAT: Data-Free Quantization Aware Training for Large Language Models (arXiV 2023). > systematic way to determine the optimal values for these hyperparameters ... apply meta-learning here Thank you for this very interesting question. Meta-learning can indeed be useful. For instance, each matrix decomposition problem along with a given calibration set, can be treated as a separate task. The objective would be to learn how to choose the target rank and bit budget for the decomposition. The input to the meta-learning algorithm could be input characteristics of the matrix, such as dimensions, dynamic range of the entries, spectrum of the matrix, etc. Furthermore, the measure of performance could be the Frobenius norm error, possibly regularized with some metrics for the computation time required for obtaining the decomposition, or inference time with the decomposed representation. This is a direction worth exploring! > iterative nature of CALDERA's optimization process may require more computational resources Yes, we acknowledge this additional computational cost of compressing models iteratively. We will release open-source our compressed models, so that this one-time compression cost need not be incurred for someone who wishes to align or fine-tune our compressed models for downstream evaluations. Compressing Llama-2 7B or Mistral-7B models (with rank-256) took approximately 34 GPU-hours, when done on NVIDIA A10G GPUs provisioned from a G5 instance on AWS. Additionally, compressing Llama-2 13B took 59 GPU-hours, when done on NVIDIA A6000 GPUs. As noted in the paper, CALDERA is expected to take more time that QuIP# (without RHT finetuning) as it is an iterative algorithm. However, as seen above, the wallclock times are not prohibitively large. Furthermore, LLaMa-2 70B can be quantized via CALDERA with rank-256 factors in about 90 GPU hours on an H100 cluster, which is on par with QuIP# (with RHT finetuning), which reports 100 GPU hours. CALDERA is also faster than some other state-of-the-art methods such as AQLM, which reports a range between 192 to 384 GPU hours, and LLM-QAT, which requires 960 GPU hours. It should be kept in mind that the cost of compressing a model is a one-time cost, and it can be reasonably afforded as long as it is not prohibitively huge. --- Rebuttal Comment 1.1: Comment: Thank you very much for your detailed response. I will keep my original score.	Summary: This paper introduces CALDERA, a novel post-training method that combines quantization and low-rank decomposition techniques for compressing large language models. The primary contribution lies in the design of a combined pipeline and the application of low-rank decomposition. Experimental results demonstrate that integrating the designed low-rank decomposition technique enhances compression performance compared to using quantization alone. Strengths: 1. This paper addresses a highly significant and widely studied research topic: compressing large language models. 2. The author not only conducts experiments but also provides numerical evidence to demonstrate the effectiveness of the designed compression method. 3. Experimental results indicate that the combined methods designed in this study effectively enhance the performance of large language models compressed using quantization-only methods. Weaknesses: 1. Missing Experimental Results: The experiments conducted in this paper are inadequate. Many critical experiments conducted in previous works have not been included. For instance: - Evaluation of throughput for LLaMA models compressed by CALDERA and QuIP# under the same average bits with the same batch size and sequence length. It's crucial to assess whether combining low-rank decomposition with quantization compromises efficiency compared to uncompressed LLMs. - Evaluation results on LLMs with different architectures from LLaMA and on generation datasets. - Evaluation of time consumption between CALDERA and QuIP# for compressing LLMs. - Evaluation of compression performance using different calibration sets and varying numbers of calibration data. 2. Lack of Novelty: The novelty of this paper is limited. The primary contributions are the design of a combination framework and a low-rank decomposition algorithm, both of which appear simplistic and resemble iterative weight compression methods with minor losses. The use of the existing quantization method LDLQ for combination and the effectiveness of the designed low-rank decomposition method in achieving minimal compression loss are questionable. Replacing it with established work like data whitening method in SVD-LLM[1] might yield better performance, given SVD-LLM's theoretical proof of achieving global optimality through data whitening. [1] Wang, X., Zheng, Y., Wan, Z., & Zhang, M. (2024). SVD-LLM: Truncation-aware Singular Value Decomposition for Large Language Model Compression. https://arxiv.org/abs/2403.07378 3. Intuitive Experimental Results: While it's intuitive that combining low-rank decomposition with quantization can improve performance over quantization-only methods, existing works such as SVD-LLM already demonstrate this. To add value to the proposed methods, it's crucial to include experimental results where the average bits are equal to or less than 1-bit, demonstrating capabilities that quantization-only methods cannot achieve. Technical Quality: 3 Clarity: 1 Questions for Authors: See above. Confidence: 5 Soundness: 3 Presentation: 1 Contribution: 1 Limitations: N.A. Flag For Ethics Review: ['No ethics review needed.'] Rating: 2 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: > Novelty We propose an idea which is simple in its execution; that does not mean it lacks novelty. CALDERA is the first work that combines quantization with a low-rank decomposition -- both of which are usually treated independently in existing works on LLM compression. Additionally, a significant contribution of this work lies in the novel theoretical guarantees bounding the expected approximation error with respect to the bit budget. The quantization of the low-rank factors, as well as the complexity of the LDLQ quantization method, make these bounds challenging to obtain. Thank you for pointing us out to the nice work, SVD-LLM. Specifically, our work is different in the following aspects: 1. For a given weight matrix $\bf A$, SVD-LLM considers solving the rank-constrained regression problem, ${\rm min}_{\bf L,R}\|\| {\bf (LR-Z)X} \|\|_F^2$ to global optimality. This is a known result (ref. [1] below). We use an equivalent whitening process in Lemma 4.2 of our paper. Additionally, our theoretical guarantees upper bounds the perturbation error to this optimal solution due to the joint quantization constraints on $\bf L$ and $\bf R$ -- something not considered in the SVD-LLM paper. Quantization of low-rank factors in SVD-LLM is done independent to the low-rank decomposition, (i.e., one after the other) in Sec. 4.4, which is suboptimal compared to the joint decomposition of our LPLRFactorize module. This is reflected in the numerical results, where SVD-LLM reports a perplexity of 13.29 (Table 8 for SVD-LLM + GPTQ-4bit), which is significantly higher that our reported PPLs in Table 1 (for example, 6.19 for CALDERA Rank-256, $B_L = B_R = 4$). 2. Moreover, CALDERA considers a $\bf Q + LR$ decomposition. This additive matrix $\bf Q$ (not considered in SVD-LLM), coarsely captures the effect of trailing singular values, which are entirely truncated in SVD-LLM. The $\bf Q$ matrix enables CALDERA to achieve PPLs close to the unquantized model in the extreme compression regime considered. [1] Xiang, et. al., Optimal exact least squares rank minimization (ACM SIGKDD, 2012) > Intuitive Experimental Results One of the core motivations of post-training quantization of LLMs is to ensure minimal loss in PPLs or accuracies compared to the uncompressed model. Current methods that compress an LLM in the regime of 2 to 2.5 bits per parameter, fail to close the gap to uncompressed performance. While contemporary works like BiLLM (Huang et. al, 2024) do consider the lower ~1 bit per parameter regime, the associated PPLs are relatively high compared to the uncompressed model. In contrast, our work is motivated with closing the gap with respect to uncompressed models in the 2 to 2.5 bits per parameter regime. While it is indeed intuitive that low-rank + quantization is expected to perform better than quantization only methods, our work rigorously formalizes this intuition with theoretical backing and experimental results. Although both SVD-LLM and our work explore the benefits of combining quantization with low-rank decomposition, SVD-LLM directly applies LLM quantization methods to the low-rank factors in an orthogonal fashion (which is suboptimal), whereas CALDERA additively combines a full-rank, aggressively-quantized matrix with a quantized low-rank correction (which is obtained by treating the quantization and low-rank constraints jointly). > Evaluation of throughput We agree that it is important to ensure CALDERA does not degrade generation throughput. We have performed some throughput experiments, which can be found in Table 1 of the global response PDF. CALDERA achieves a slightly lower throughput than QuIP#, as it requires dequantization of 3 matrices -- $\bf Q, L$ and $\bf R$ (instead of just one in the case of QuIP#), and perform additional matrix multiplications during inference. Despite this, it is worthwhile to note that the throughput is quite higher than the uncompressed model. Furthermore, it should be kept in mind that designing specialized CUDA kernels that are aware of the $\bf Q + LR$ decomposition can improve the throughput even further -- for example, by computing $\bf Qx$ and $\bf LRx$ in parallel. Writing custom kernels is left for future work. > Evaluation ... different calibration sets Our choice of sampling from RedPajama to get our calibration dataset is motivated from the fact that OpenLlama uses it to pre-train the model from scratch, and it is important to ensure that the calibration data is from the same distribution as the pre-training data. While ablations over calibration datasets is worthwhile to explore, we have decided to prioritize other experiments (cf. global response PDF) within our current computational budget. > different architectures We have performed additional experiments (specifically, we compressed Llama-2 13B and Mistral-7B), which can be found in Table 2 of the global response PDF. > Evaluation ... time consumption While a comprehensive evaluations would require more time, we report some numbers here. Compressing Llama-2 7B or Mistral-7B models (with rank-256) took approximately 34 GPU-hours, when done on NVIDIA A10G GPUs provisioned from a G5 instance on AWS. Additionally, compressing Llama-2 13B took 59 GPU-hours, when done on NVIDIA A6000 GPUs. CALDERA is expected to take more time that QuIP# (no RHT finetuning) as it is an iterative algorithm. However, as seen above, the wallclock times are not prohibitively large. Also, LLaMa-2 70B can be quantized via CALDERA with rank-256 factors in ~90 GPU hours on an H100 cluster, which is on par with QuIP# (with RHT finetuning), which reports 100 GPU hours. CALDERA is also faster than some other state-of-the-art methods such as AQLM, which reports a range between 192 to 384 GPU hours, and LLM-QAT, which requires 960 GPU hours. It should be kept in mind that the cost of compressing a model is a one-time cost, and it can be reasonably afforded as long as it is not prohibitively huge. --- Rebuttal Comment 1.1: Title: Further clarifications Comment: Dear Reviewer 32tN, Please let us know if your queries have been addressed satisfactorily. As mentioned in our response, we've thoroughly incorporated your feedback, along with suggestions from the other reviewers. We hope that our response has positively influenced your perception of our work. If you require further clarifications to potentially reconsider your score, we are enthusiastic about engaging in further discussion. Please do not hesitate to contact us. We highly value the generous contribution of your time to review our paper. --- Rebuttal 2: Title: More concerns Comment: Thank the authors for the detailed response and clarification. After reading the global response given by the authors. I have two more concerns. 1. The compression time cost of CALDERA is much higher than QuIP#. In Appendix F.7 of QuIP#'s paper, the authors claim that "All experiments were run on NVIDIA A100 GPUs ... We find that we can quantize Llama2 70B without fine-tuning in under 10 GPU-hours and with fine-tuning in around 100 GPU-hours." Therefore, the comparison of time costs in the global response is NOT fair. Based on the GPU hours for running CALDERA and QuIP#, I made the following table. The time cost of CALDERA without fine-tuning is already higher than QuIP# with fine-tuning. However, the perplexity (PPL) of the LLM compressed by CALDERA without fine-tuning is not as good as that compressed by QuIP# with fine-tuning, as reported in [1]. Therefore, given the same compute budget for compression, CALDERA is less competitive than QuIP#. Also, since QuIP# requires less time than CALDERA for compression, the explanation in the footnote of page 8 in the submission for using the LLM compressed by QuIP# but WITHOUT being fully fine-tuned is NOT convincing, and the all results reported in Section 5.2 are also NOT convincing. Therefore, it is still questionable whether the designed algorithm CALDERA is better than the baseline method QuIP#. \| \| w/o Fine-tuning \| w/ fine-tuning \| \|---------\|-------------------\|--------------------\| \| CALDERA \| 90 H100 GPU-hours \| Unknown \| \| QuIP# \| 10 A100 GPU-hours \| 100 A100 GPU-hours \| 2. The reported generation throughput is poor. In Table 1 shown in the global response, the throughput of LLM compressed by CALDERA with its best configuration (Rank=256, $B_L = B_R = 4$) is 45 tok/sec, which is much worse than that of QuIP#, which is 76 tok/sec. Therefore, if throughput is the primary target, CALDERA is still less competitive than QuIP#. Based on the concerns above, I have decided to temporarily adjust my score to reject and I hope more clarifications on these two concerns. Thank you. [1] Albert Tseng et al., QuIP#: Even Better LLM Quantization with Hadamard Incoherence and Lattice Codebooks. ICML, 2024 --- Rebuttal Comment 2.1: Title: Abnormal Results in Table 5 Comment: When comparing the accuracy of uncompressed LLaMA 2-7B and LLaMA 2-70B in Table 1 in the submission with LLaMA 2-7B compressed by CALDERA with LoRA fine-tuning in Table 5, I found that the accuracy of compressed LLaMA 2-7B is much higher than that of the uncompressed LLaMA 2-7B and even uncompressed LLaMA 2-70B, as shown below. \| \| Bits \| Wino \| RTE \| \|------------------------------------------\|------\|-------\|-------\| \| Uncompressed LLaMA 2-7B \| 16 \| 67.3 \| 63.2 \| \| Uncompressed LLaMA 2-70B \| 16 \| 77.0 \| 67.9 \| \| LLaMA 2-7B compressed by CALDERA with FT \| 2.4 \| 84.93 \| 86.28 \| This is abnormal, as the compressed LLM shows a significant improvement in accuracy despite an extensive compression ratio (the weight memory of LLaMA 2-70B 16bit is about 80 times larger than that of LLaMA 2-7B 2bit). The reasonable explanations I can think of are either the failure to use the optimal parameter configuration during inference, or the overfitting of the fine-tuned LLM on these three datasets. Therefore, the results in table 5 are NOT convincing. --- Rebuttal 3: Title: Response to more concerns and abnormal results Comment: Dear Reviewer 32tN, we believe you have made two major misunderstandings. Please allow us to clarify them. Firstly, we believe there has been a misunderstanding as to what is referred to as fine-tuning in our paper. In LoRA fine-tuning, the model is adapted or fine-tuned using a smaller, task-specific dataset. Please note that Randomized Hadamard Transform fine-tuning (RHT-FT) and LoRA fine-tuning are two separate components, and independent of each other. RHT-FT is a step of the quantization process in QuIP# -- it is done on the calibration set, and is meant to decrease the PPL (or increase the accuracy) across all tasks generally. In contrast, LoRA fine-tuning is task-specific, and is done on a specific dataset like Wino or RTE. For instance, an accuracy of 84.93\% for Wino and 86.28\% for RTE, is a consequence of fine-tuning the LLM using LoRA on Wino and RTE, respectively. The uncompressed models have NOT been fine-tuned to these task specific datasets, which is why CALDERA + Fine-tuning performs better than the uncompressed model. This is very natural to expect as a consequence of fine-tuning, and not abnormal at all. It is observed in other prior works as well -- for example, see Fig. 2 of LQ-LoRA [1]. Hence, it is not a fair comparison to compare CALDERA-quantized models fine-tuned for a specific downstream task with non fine-tuned uncompressed models. [1] Guo et. al, LQ-LoRA: Low-rank Plus Quantized Matrix Decomposition for Efficient Language Model Finetuning, ICLR 2024. Secondly, we believe there has been a misunderstanding on your end regarding performance of our compressed models and prior works like QuIP#. We agree that if the metric of performance is purely quantization time or generation throughput, QuIP# achieves higher performance with respect to those metrics. However, that has never been the primary thesis of our work, i.e., we do not claim that compressing LLMs with CALDERA is faster than QuIP#. A central thesis of our work is that CALDERA does perform better than QuIP\# in the sense that that it consistently achieves lower perplexities than QuIP# -- when comparing QuIP# without RHT-FT to CALDERA without RHT-FT, or QuIP# with RHT-FT to CALDERA with RHT-FT. As RHT-FT is an optimization that applies to both QuIP# and CALDERA, it is unfair to compare accuracies and PPLs achieved by QuIP# with RHT-FT to CALDERA without RHT-FT. For a fair comparison, please refer to Table 3 where we have compared CALDERA with RHT-FT and QuIP# with RHT-FT, for 7B models. It is evident from Table 5 that CALDERA consistently has lower PPL than QuIP#. Furthermore, if you look at Table 4 of QuIP#, they report PPLs of 6.19 (Wiki) and 8.16 (C4), which are still higher than 5.84 (Wiki) and 7.75 (C4) (highlighted numbers on Table 5). We have not done RHT-FT on the 70B model because of limited rebuttal duration. But our original submission does report extensive comparisons with RHT-FT for 7B models. Doing such extensive comparisons with the 70B model within the rebuttal window is infeasible. We report wallclock time for quantization, just to show that it not prohibitively high when compared to QuIP#. Additionally, doing RHT-FT on the CALDERA quantized 70B would take the same amount of time as the RHT-FT step in QuIP# does, which is once again, not prohibitively slow, given this is a one-time cost. Regarding throughput, our global response simply shows that it is not degraded much compared to QuIP# due to the low-rank component. It should be noted that it is still significantly higher than the unquantized model (even without custom CUDA kernels for the low-rank component). We reiterate that our work is motivated with closing the gap with respect to uncompressed models in the 2 to 2.5 bits per parameter regime. And, it is clear that we are able to get lower perplexities than QuIP# when the comparison is fair, i.e., either with or without RHT-FT in both cases. --- Rebuttal Comment 3.1: Title: Keep my rating Comment: Thanks for the authors' response. However, the response does not address my concerns: 1. The author is confusing the fine-tuning of LLMs with the fine-tuning of BERT. LLMs are designed to handle a wide range of tasks, including text generation, translation, summarization, and more. Fine-tuning an LLM often involves adapting the model to perform well on diverse tasks rather than focusing on a single, narrowly defined task like classification. In contrast, BERT is often fine-tuned for specific tasks such as classification, question answering, and named entity recognition. I highly recommend that the authors add evaluations on several generation tasks. As mentioned before, I deeply suspect that the LLM is overfitting on fine-tuned classification datasets such as RTE and will show much poorer generation ability. 2. The author does not follow the literature to fine-tune the LLM. To the best of our knowledge, none of the existing LLM compression works fine-tune their compressed LLM on classification datasets. Common datasets for fine-tuning include language modeling datasets such as WikiText-2 and C4, or instruction tuning datasets such as Alpaca. The paper [1] mentioned by the authors in their response also follows this trend. For instance, the authors fine-tune RoBERTa-Large on classification datasets but fine-tune models like LLaMA-2 on C4 and OpenAssistant datasets. We recommend the authors follow the literature to design the LoRA fine-tuning experiments. 3. The author does not address the poor time cost for compression and the poor throughput for inference. Even though CALDERA achieves slightly better PPL than QuIP#, its poor time cost for compression and inferior throughput for inference make it less attractive. For example, when the goal is to compress the LLM to reduce its inference latency on a server where memory is not the main concern, CALDERA is even less appealing than some 3-bit quantization algorithms, as these algorithms guarantee lower latency and similar or even better accuracy. 4. The author should use the default configuration to fully RHT-FT QuIP# and update the evaluation results on Table 3 and Table 4. It is insufficient to only use PPL to compare the accuracy of the compressed LLMs, especially when the PPL values from two compressed LLMs are so close. The authors should update the evaluation results of LLMs compressed by QuIP# with fully RHT-FT in both Table 3 and Table 4 in the submission. Therefore, I would still like to keep my rating.	Summary: The article introduces CALDERA, a post-training compression algorithm for large language models (LLMs) that leverages the low-rank structure of weight matrices to achieve significant compression. CALDERA approximates a weight matrix $W$ using a low-rank, low-precision decomposition $W \approx Q + LR$, where $Q$, $L$ and $R$ are quantized. The method aims to reduce the memory and computational footprint of LLMs, facilitating their deployment on memory-constrained edge devices. Strengths: Innovative Compression Technique: CALDERA combines low-rank approximation with low-precision quantization, addressing both the redundancy and precision issues in LLM weight matrices. Theoretical Foundations: The algorithm is backed by rigorous theoretical guarantees on the approximation error, which enhances its reliability. Performance: Empirical results show that CALDERA outperforms existing post-training compression techniques in terms of zero-shot performance, particularly when compressed to less than 2.5 bits per parameter. Adaptability: The method supports low-rank adaptation, which can further enhance the performance of the compressed models on specific tasks. Weaknesses: Complexity: The algorithm involves a nested optimization process, which may introduce significant computational overhead during the compression phase. Dependency on Calibration Data: The performance of CALDERA relies on the availability and quality of calibration data, which may not always be accessible or representative. Limited Empirical Comparisons: While the article claims superiority over existing methods, the empirical comparisons might be limited in scope, potentially omitting some relevant state-of-the-art techniques. Potential Issues Quantization Artifacts: Aggressive quantization, especially at very low bit budgets, might introduce artifacts that could degrade model performance in specific scenarios. Generalization: The effectiveness of the method across a diverse range of LLMs and tasks needs thorough validation, as it might not generalize well beyond the tested models and datasets. Optimization Stability: The iterative nature of the optimization process might lead to stability issues, particularly for large models with highly non-convex loss landscapes. Technical Quality: 2 Clarity: 3 Questions for Authors: See Weaknesses Confidence: 2 Soundness: 2 Presentation: 3 Contribution: 2 Limitations: Yes Flag For Ethics Review: ['No ethics review needed.'] Rating: 4 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: > Quantization Artifacts We agree that compression introduces artifacts. However, PPLs on language modeling datasets like Wikitext and C4 are generally considered to be reasonably good indicators of the performance of an LLM, as can be seen in the existing literature. A comprehensive evaluation over the complete spectrum of LLM evaluation tasks is beyond the scope of our computational budget. In addition, we have performed additional model quantization experiments on different sizes of LLaMa-2 model, as well as Mistral 7B, for which CALDERA generalizes well. These results can be found in Table 2 in the global response PDF. > Limited Empirical Comparisons Please refer to additional experiments in the global response PDF. > Optimization Stability While it is possible that the iterative nature of our algorithm may oscillate, it is not a cause of concern as Algs. 1 and 2 keep track of (and return) the best matrices $\mathbf{Q}$, $\mathbf{L}$ and $\mathbf{R}$ seen across all the iterations. Moreover, as can be seen from our Frobenius norm vs. iteration plots in Appendix G, the convergence of our algorithm is pretty well-behaved when decomposing LLM weight matrices. Furthermore, our theoretical guarantee in Thm 4.1 serves as an assurance that stability is not a cause of concern. > Complexity We have discussed the computational complexity of CALDERA in Appendix D, and acknowledge it as a price paid for the improved error guarantees (as stated in Thm 4.1, as well as our numerical evaluations). Additionally, we report (and compare) some numbers here. Compressing Llama-2 7B or Mistral-7B models (with rank-256) took approximately 34 GPU-hours, when done on NVIDIA A10G GPUs provisioned from a G5 instance on AWS. Additionally, compressing Llama-2 13B took 59 GPU-hours, when done on NVIDIA A6000 GPUs. As noted in the paper, CALDERA is expected to take more time that QuIP\# (without RHT finetuning) as it is an iterative algorithm. However, as seen above, the wallclock times are not prohibitively large. Furthermore, LLaMa-2 70B can be quantized via CALDERA with rank-256 factors in about 90 GPU hours on an H100 cluster, which is on par with QuIP# (with RHT finetuning), which reports 100 GPU hours. CALDERA is also faster than some other state-of-the-art methods such as AQLM, which reports a range between 192 to 384 GPU hours, and LLM-QAT, which requires 960 GPU hours. It should be kept in mind that the cost of compressing a model is a one-time cost, and it can be reasonably afforded as long as it is not prohibitively huge. > Dependency on Calibration Data While this is true, the availability of good quality data is a central premise for the success of any data-centric algorithm. It is possible to derive an entirely data-agnostic variant of CALDERA, by simply replacing the calibration data matrix $\mathbf{X}$ by an identity matrix, $\mathbf{I}$. However, using a data-aware version yields better results as validated by ours, as well as several prior works, including QuIP# and LQ-LORA. --- Rebuttal Comment 1.1: Title: Further clarifications Comment: Dear Reviewer xRtR, Please let us know if your queries have been addressed satisfactorily. As mentioned in our response, we've thoroughly incorporated your feedback, along with suggestions from the other reviewers. We hope that our response has positively influenced your perception of our work. If you require further clarifications to potentially reconsider your score, we are enthusiastic about engaging in further discussion. Please do not hesitate to contact us. We highly value the generous contribution of your time to review our paper.	Rebuttal 1: Rebuttal: Dear Reviewers, We are very grateful for the valuable time you spent in reading our paper and sharing your concerns, and greatly appreciate the voluntary nature of the review process. In this global response, we have have summarized the major points from our individual responses. Additional experiments compressing other models: We used CALDERA to compress some additional popular LLMs -- namely, Llama-2 13B, 70B, and Mistral 7B, which can be found in Table 2 of the global response PDF. We also recompressed Llama-2 70B using the Hessians provided by QuIP# for a fairer comparison. The perplexities obtained using CALDERA are consistently lower than QuIP# (without RHT finetuning). Similar trend can be seen for the zero-shot task accuracies as well. Additional experiments to evaluate throughput: We also did additional evaluations on the generation throughput of our compressed models, which can be found in Table 1 of the global response PDF. It is noteworthy that the throughput of our CALDERA-compressed model is quite higher than the uncompressed model. CALDERA does achieve a slightly lower throughput than QuIP# -- however, this is expected, as it requires dequantization of three matrices -- $\bf Q, L$ and $\bf R$. It should be kept in mind that designing specialized CUDA kernels that are aware of the $\bf Q + LR$ decomposition can improve the throughput even further. Computation cost of compressing using CALDERA: We conducted some more experiments to compare the wall-clock times of using CALDERA. Compressing Llama-2 7B or Mistral-7B models (with rank-256) took approximately 34 GPU-hours, when done on NVIDIA A10G GPUs provisioned from a G5 instance on AWS. Additionally, compressing Llama-2 13B took 59 GPU-hours, when done on NVIDIA A6000 GPUs. CALDERA is expected to take more time that QuIP\# (no RHT finetuning) as it is an iterative algorithm. However, as seen above, the wallclock times are not prohibitively large. Also, LLaMa-2 70B can be quantized via CALDERA with rank-256 factors in ~90 GPU hours on an H100 cluster, which is on par with QuIP\# (with RHT finetuning), which reports 100 GPU hours. CALDERA is also faster than some other state-of-the-art methods such as AQLM, which reports a range between 192 to 384 GPU hours, and LLM-QAT, which requires 960 GPU hours. It should be kept in mind that the cost of compressing a model is a one-time cost, and it can be reasonably afforded as long as it is not prohibitively huge. In summary, we reiterate that our work proposes and rigorously analyzes a simple optimization-centric point of view to the problem of obtaining quantized and low-rank decompositions of the weight matrices of an LLM. We evaluate the success of our $\bf Q + LR$ style of matrix decomposition algorithm for post-training LLM quantization in the challenging regime of 2 to 2.5 bits per parameter. Our LLM compression scheme reduces the model distribution and deployment costs, and the low-rank component $\bf LR$ provides good initializations for further fine-tuning using popular low-rank adaptation methods. We hope our responses clarify the reviewers' concerns. Pdf: /pdf/58eb86add3f16cca4a346bdfc0cb26f1a496f38e.pdf	NeurIPS_2024_submissions_huggingface	2,024	Summary: This work proposes a new decomposition procedure to achieve low-precision and low-rank compression of large language model (LLM) weight matrices. Such a new method could capture high singular components accurately while compressing less significant ones. An efficient algorithm is proposed to optimize the quantized backbone and low-rank factors, which can be fine-tuned to enhance the model performance. The work also provides a tighter approximation error bound and is validated by compressing the LlaMA LLMs to below 2.5 bits per parameter. Strengths: 1. The proposed CALDERA method can effectively compress LLMs without degrading the performance too much. And it is demonstrated through a series of experiments that the method is capable of model compression and low-rank adaptive fine-tuning at the same time. 2. The paper also provides a rigorous approximation error analysis of the proposed CALDERA method. Weaknesses: I think the presented experiment study is limited. I mainly have the following suggestions. 1. It would be better to supplement the experiments with LoftQ and LQ-LoRA, and add a set of experiments with full parameter fine-tuning as a baseline. 2. Please provide an experimental comparison of model size. 3. It would be better to add some experiments on evaluating the computation cost of the proposed CALDERA. Technical Quality: 3 Clarity: 3 Questions for Authors: 1. Since the CALDERA method requires many time-consuming operations like SVD and matrix inversion, can this be improved in future work? 2. Are there any efficient ways to improve the selection strategy of the hyperparameter rank? Confidence: 4 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: In my opinion, how to efficiently tune the hyperparameter rank is the main limitation of the proposed method. Flag For Ethics Review: ['No ethics review needed.'] Rating: 7 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: > experimental comparison of model size We have performed ablations over multiple sizes of the Llama-2 family. Please refer to Table 2 of the global response. The PPLs obtained using CALDERA are consistently better than QuIP# (without fine-tuning) for comparable average bits. > evaluating the computation cost Thank you for this suggestion. While a comprehensive evaluations would require more time, we report (and compare) some numbers here. Compressing Llama-2 7B or Mistral-7B models (with rank-256) took approximately 34 GPU-hours, when done on NVIDIA A10G GPUs provisioned from a G5 instance on AWS. Additionally, compressing Llama-2 13B took 59 GPU-hours, when done on NVIDIA A6000 GPUs. As noted in the paper, CALDERA is expected to take more time that QuIP# (without RHT finetuning) as it is an iterative algorithm. However, as seen above, the wallclock times are not prohibitively large. Furthermore, LLaMa-2 70B can be quantized via CALDERA with rank-256 factors in about 90 GPU hours on an H100 cluster, which is on par with QuIP# (with RHT finetuning), which reports 100 GPU hours. CALDERA is also faster than some other state-of-the-art methods such as AQLM, which reports 192 to 384 GPU hours, and LLM-QAT, which requires 960 GPU hours. Additionally, each layer is quantized independent of other layers -- hence, the wallclock time for CALDERA can be reduced by processing the layers in parallel (i.e., scaling horizontally by using more GPUs). Moreover, it should be kept in mind that the cost of compressing a model is a one-time cost, and it can be reasonably afforded as long as it is not prohibitively huge. > time-consuming operations like SVD and matrix inversion, can this be improved in future work? Thank you for raising this question. It is indeed possible to reduce the computational complexity of CALDERA. For instance, the SVD computation in the LPLRFactorize submodule can be replaced with randomized LPLR submodule from (ref [30] in the paper), which leverages Gaussian sketching matrices to reduce the complexity from $O(nd^2)$ to $O(ndm)$, where $m \ll \mathrm{min}\\{n,d\\}$ is the sketch size. Furthermore, we would like to clarify that matrix inversion is not necessary to obtain the left and right low-rank factors. These factors can be obtained by directly solving the corresponding least-squares minimization problem (in lines 8 and 9 of Alg. 2) using a conjugate gradient descent based solver, which will be significantly faster. The closed form expressions in our paper are used to facilitate analysis and derive Thm 4.1. More generally, the constrained optimization problem (1) is NP-hard, and designing efficient algorithms to solve it is an interesting research avenue, with applications much broader than LLM compression. > selection strategy of the hyperparameter rank? Thank you for this very interesting question! At a high level, it is possible to adaptively select the rank of each layer during fine-tuning using a strategy similar to AdaLoRa [1]. The initial rank can be chosen to be high, for instance 256. Subsequently, during fine-tuning for a downstream task, the singular values of the low-rank component can be analyzed such that smaller singular values can be truncated to adaptively reduce the rank of each layer. Finding efficient and optimal ways to do this warrants a deeper investigation. [1] Zhang et. al., Adaptive Budget Allocation for Parameter-Efficient Fine-Tuning (ICLR, 2023) > experiments with LoftQ and LQ-LoRA ... full parameter fine-tuning as a baseline. The results of LoftQ and LQ-LoRa are restated in Table 5 (copied from respective papers). Fine-tuning the low-rank factors obtained from CALDERA shows that we get lower PPLs for fewer average number of bits when compared to LoftQ or LQ-LoRA. LoftQ and LQ-LoRA report PPLs that are at par or better than vanilla LoRA on WikiText2 PPLs. Moreover, it is know that for several tasks, LoRA performs at par with full fine-tuning. Therefore, we have prioritized utilizing our limited computational budget for other ablation experiments on CALDERA (please refer to the global response). --- Rebuttal Comment 1.1: Title: Further clarifications Comment: Dear Reviewer ZG6L, Please let us know if your queries have been addressed satisfactorily. As mentioned in our response, we've thoroughly incorporated your feedback, along with suggestions from the other reviewers. We hope that our response has positively influenced your perception of our work. If you require further clarifications to potentially reconsider your score, we are enthusiastic about engaging in further discussion. Please do not hesitate to contact us. We highly value the generous contribution of your time to review our paper.	null	null	null	null	null	null
NeCGS: Neural Compression for 3D Geometry Sets	Reject	Summary: The manuscript introduces a neural compression paradigm for effectively compressing diverse sets of 3D geometry models. The authors propose a two-stage framework that first converts irregular mesh models into a regular 4D TSDF-Def volume representation and then employs a quantization-aware auto-decoder network to achieve redundancy elimination and compact representation. The method claims to compress a large number of 3D mesh models with high accuracy and preservation of geometric details, outperforming state-of-the-art methods both quantitatively and qualitatively. Strengths: - The paper presents a unique method for compressing 3D geometry sets by leveraging neural networks, which is a significant advancement in the field. NeCGS achieves an impressive compression ratio, which is a critical metric for 3D geometry data compression. - The method maintains high accuracy and preserves detailed geometric structures even at high compression ratios. The authors have conducted comprehensive experiments and ablation studies across various datasets, demonstrating the effectiveness of their approach. - The inclusion of source code in the supplemental material enhances the reproducibility and transparency of the research. - The paper is well-organized, with clear explanations of the methodology and results. Weaknesses: - The manuscript mentions that the optimization process for TSDF-Def volumes is time-consuming (over 15 hours), which could be a limitation for practical applications. The manuscript should address the long optimization time required for the TSDF-Def volumes. Future work could focus on accelerating this process to make the method more practical. - While the method performs well on tested datasets, it is unclear how well it generalizes to other, more complex, or varied 3D geometry sets, such as some geometry with thin structures or open boundaries (cloth). - The choice of an auto-decoder network is effective, but the paper could benefit from a more detailed explanation of why this architecture was chosen over others. - While the method outperforms existing techniques, a more thorough comparison in terms of trade-offs, especially related to computational resources, would be insightful. - The paper could provide more insights into how the method scales with the size and complexity of the 3D geometry sets. The paper should include scalability tests to understand how the method performs with larger and more complex datasets. Technical Quality: 3 Clarity: 3 Questions for Authors: The manuscript presents a contribution to the field of 3D geometry data compression with the introduction of NeCGS. The innovative approach of using a neural network for compression and the high compression ratios achieved is commendable. However, there are several areas where the manuscript could be improved: computation efficiency, scalability on various data, and other minor issues. In conclusion, the manuscript is well-written and presents a promising new direction for 3D geometry compression. Addressing the above points will significantly enhance the manuscript's contribution to the field. I am on the fence, and looking forward to the reply and other reviews. Confidence: 4 Soundness: 3 Presentation: 3 Contribution: 2 Limitations: NA Flag For Ethics Review: ['No ethics review needed.'] Rating: 5 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: ### Comment 1. The manuscript ... more practical. Response: * Thanks for the valuable suggestion. First, we clarify the optimization process of converting 3D models into TSDF-Def 4D volumes is efficient, as shown in the table below, while 15 hours refers to the time consumed by the whole compression process, including the TSDF-Def representation process and optimization process of both auto-decoder and features. * The optimization process of converting the 3D models of a dataset into TSDF-Def volumes can be parallelized using multiple GPUs. In our experiment, we utilized 8 NVIDIA RTX 3090 GPUs.The table below details the time consumed for processing the mixed dataset. \|Resolution\| Time per Shape (s)\|Total Time (h)\| \|-----\|-----\|-----\| \|32\|22.39 \| 0.46 \| \|48\| 24.34\| 0.50\| \|64\| 26.67\| 0.55 \| \|128\| 28.03\| 0.58\| \|256\|41.51 \|0.86 \| * Second, we want to emphasize that our NecGS is designed for the offline compression of 3D geometry datasets to save storage space, where the optimization/compression time should not be a key factor. Instead, we should concentrate more on the decompression/decoding speed because the users expect to obtain the 3D models timely when querying the dataset. As shown in Table 3, when the resolution is 128, the decompression time of our NeCGS is only 98.95 ms, satisfying the real-time requirement. * Moreover, there are various potential solutions to accelerate the optimization. At the software level, more efficient convolution (e.g., using multiple 1-D or 2-D convolution to approximate the 3-D convolution or using 3D sparse convolution to replace the traditional 3D convolution) can be used to speed up the process. At the hardware level, optimization programs can be run using multiple GPUs or other more efficient hardware. Like NeRF, the initial algorithms required several days for optimization. Subsequently, improved methods have been introduced to accelerate the optimization process to a few minutes or even seconds. ### Comment 2. While the method ... open boundaries (cloth). Response: * Our NeCGS can be adapted to newly added 3D models to the dataset. Given a new 3D model, we first represent it into a TSDF-Def volume through Algorithm 1. Then following the optimization progress in Sec. 3.2, we only optimize its corresponding embedded feature while keeping the trained decoder unchanged. The visual results are shown in Fig. R1 of the uploaded PDF, demonstrating its generalization ability to new geometry data. Moreover, in this situation, the optimization is significantly less because only embedded features are optimized. * The dataset we used indeed consists of complex shapes, as shown in Fig. 1, where the complex structures of the decompressed models remain. In Fig. R5 of the uploaded PDF files, we demonstrate additional complex shapes. * The signs in TSDF are ambiguous on models with open boundaries, our method cannot directly compress models of this category. However, an alternative and straightforward approach is to utilize UDF (unsigned distance field) rather than SDF to represent them in our method, which allows for the processing of models with open boundaries. ### Comment 3. The choice of an ... over others. Response: In the ablation study, we indeed compared our auto-decoder framework with auto-encoder, a widely used framework. The visual results shown in Fig. 8(a) demonstrate the superiority of our decoder-based structure. In the auto-encoder framework, the embedded features are adjusted by optimizing the encoder, which is less flexible than the auto-decoder, where the embedded features are optimized directly. In the final version, we will provide more explanations. ### Comment 4. While the ..., would be insightful. Response: We refer the reviewer to the 4th response to Reviewer qNSe for comparison of compression time. ### Comment 5. The paper ... complex datasets. * Thank you for the insightful comments. In the final version, we will complement more discussions about this point. In our experiments, the embedded features and the decoder are optimized over a fixed number of epochs with a constant batch size. Consequently, the overall optimization time and computational expenses scale proportionally with the amount of geometric shapes being compressed. * To validate this, in addition to the mixed dataset utilized in the experiments (600 shapes), we create two additional mixed datasets of different sizes by selecting 100 and 300 shapes from the remaining three datasets. This results in mixed datasets comprising 300 and 900 shapes, respectively. The table below shows the optimization times for various Mixed datasets. \|# Shapes\|Optimization Time (h)\| \|-----\|-----\| \|300\|8.25\| \|600\|16.32\| \|900\|24.37\| * More importantly, we also want to emphasize that our NeCGS our NecGS is designed for the offline compression of 3D geometry datasets to save storage space, where the optimization/compression time should not be a key factor. Instead, we should concentrate more on the decompression/decoding speed because the users expect to obtain the 3D models timely when querying the dataset. As shown in Table 3, when the resolution is 128, the decompression time of our NeCGS is only 98.95 ms, satisfying the real-time requirement. * Besides, there are various potential solutions to accelerate the optimization. At the software level, more efficient convolution (e.g., using multiple 1-D or 2-D convolution to approximate the 3-D convolution or using 3D sparse convolution to replace the traditional 3D convolution) can be used to speed up the process. At the hardware level, optimization programs can be run using multiple GPUs or other more efficient hardware. Like NeRF, the initial algorithms required several days for optimization. Subsequently, improved methods have been introduced to accelerate the optimization process to a few minutes or even seconds. --- Rebuttal Comment 1.1: Title: Reply Comment: Thanks for your great efforts! After reading the response, some major issues have been addressed well. I would keep my original score. Thanks! --- Reply to Comment 1.1.1: Comment: It is wonderful to get your further feedback mentioning that our initial responses have effectively tackled your concerns. We appreciate your recognition of our efforts. --- Rebuttal 2: Title: The authors are looking forward to your feedback. Thanks for your time. Comment: Dear Reviewer 18DA Thanks for your time and effort in reviewing our manuscript and the favorable recommendation. In our previous response, we addressed your remaining concerns directly and comprehensively. We very much look forward to your further feedback on our responses. Best regards, The authors	Summary: This paper proposes a neural compression algorithm, NeCGS to significantly compress geometry datasets. The algorithm mainly consists of 2 components, 1) regular geometry representation: This is an optimization algorithm to optimize the TSDF field such that the error between the original geometries and the geometries reconstructed by the deformable marching cube algorithm is minimized and 2) compact neural representation: regresses the optimized TSDF-def fields from compressed latent states, quantizes the latent states and compresses them further into bitstreams. The trained decoder can then be used to reconstruct the TSDF-def fields and the geometries can be reconstructed using the DMC algorithm. Strengths: The NeCGS algorithm can provide high compression ratios with impressive reconstruction capability of the geometries. Better geometry representations can be achieved using the proposed optimization algorithm. This is evident from the ability of the DMC method to accurately reconstruct surfaces. The DMC algorithm is also significant and seems to provide better reconstruction of detailed structure in the geometries. Overall, the developed compression method has high potential and the results presented in the paper are very impressive. Weaknesses: The biggest weakness of the proposed approach is the computational cost of the method. The exorbitantly large times required to compress the datasets reduce the value proposition. Additionally, it is not clear how much the computational cost scales with the size of the geometry dataset. Technical Quality: 3 Clarity: 3 Questions for Authors: - The reconstruction results from the GPCC method are very close to the NeCGS. Would it be possible to compare the compression times as well for all the base line methods? - In the ablation study section, the authors compare reconstruction accuracy for resolutions of 64, 128 and 256 and it seems the reconstruction quality does not vary by much. What happens if the resolution is reduced? It would certainly reduce the optimization costs. It would be interesting to find out how low of a resolution can be used that still out performs the baselines and the resolution at which the reconstruction accuracy significantly deteriorates? - What happens in the scenario where the geometry dataset needs to be modified or more geometries need to be added? Would the optimization cost be similar or significantly lesser? - It would be interesting to see how the optimization cost scales with the size of the geometry dataset? - Are the latent vectors of the auto decoder randomly sampled? Can more details be provided regarding that. - More details related to the DMC algorithm need to be provided. The workings of the algorithm are not entirely clear from the explanation in the paper. - In Fig. 4, is there an upper limit to the compression ratios achieved by the NeCGS? How do the results compare if you increase it? - Figure 6 is before 5 in the paper. Confidence: 4 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: NA Flag For Ethics Review: ['No ethics review needed.'] Rating: 5 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: ### Comment 1. The reconstruction results ... baseline methods? Response: * Actually, when zooming in Fig. 5 of the manuscript, the decompressed shapes by our NeCGS exhibit superior quality to those by GPCC, showcasing significantly smoother shapes decompressed by NeCGS. * We refer the reviewer to the 4th response to Reviewer qNSe for comparison of compression time. ### Comment 2. In the ablation ... significantly deteriorates? Response: * The 3D models reconstructed from their much low-resolution TSDF-Def volumes will exhibit substantial errors while leading to a notable increase in compression distortion. * In addition to the resolution examined in the ablation study, we also experimented with lower resolutions, specifically 32 and 48. The distortions are presented in the table below. It is evident that decreasing the resolution from 64 to 48 results in a significant increase in distortion. Visual representations can be found in Fig. R2 of the uploaded PDF file. \|Res.\|Size (MB)\|Com. Ratio\| CD (1e-3) \| NC\| F1-0.005\|F1-0.01\| \|-----\|-----\|-----\|-----\|-----\|-----\|-----\| \|32\|1.037\|364.898\|16.982\|0.872\|0.215\|0.542\| \|48\|1.269\|298.187\|12.240\|0.895\|0.314\|0.705\| \|64\|1.408\|268.75\|4.271\|0.927\|0.721\|0.966\| \|128\|1.493\|253.45\|3.436\|0.952\|0.842\|0.991\| \|256\|1.627\|232.58\|3.234\|0.962\|0.870\|0.995\| ### Comment 3. What happens ... or significantly lesser? Response: Our NeCGS can be adapted to newly added 3D models to the dataset. Given a new 3D model, we first represent it into a TSDF-Def volume through Algorithm 1. Then following the optimization progress in Sec. 3.2, we only optimize its corresponding embedded feature while keeping the trained decoder unchanged. The visual results are shown in Fig. R1 of the uploaded PDF, demonstrating its generalization ability to new geometry data. Moreover, it is worth noting that in this situation, the optimization is significantly less because only embedded features are optimized. ### Comment 4. It would be interesting ... geometry dataset? Response: * In the experiment, the embedded features and the decoder are optimized over a fixed number of epochs with a constant batch size. Consequently, the overall optimization time and computational expenses scale proportionally with the amount of geometric shapes being compressed. * To validate this, in addition to the mixed dataset utilized in the experiments (600 shapes), we create two additional mixed datasets of different sizes by selecting 100 and 300 shapes from the remaining three datasets. This results in mixed datasets comprising 300 and 900 shapes, respectively. The table below displays the optimization times for various Mixed datasets. \|# Shapes\|Optimization Time (h)\| \|-----\|-----\| \|300\|8.25\| \|600\|16.32\| \|900\|24.37\| ### Comment 5. Are the latent vectors ... regarding that. Response: Before the optimization, the latent vectors (embedded features) are initialized as random Gaussian noise with a mean of 0 and a standard deviation of 1/3. In the experiment, with the quantization limits set at 1 and -1, the 1/3 standard deviation ensures that nearly every value of the initialized latent vectors falls within the [-1, 1] range, aligning with the three-sigma rule of Gaussian distribution. ### Comment 6. More details ... in the paper. Response: Our DMC is modified from Marching Cubes and used to extract surfaces from TSDF-Def volumes. Besides TSDF, we assign a deformation for each corner of the cubes, making the cubes adjust detailed structures of the shapes, as shown in Fig. 3. The triangle extraction in each cube is the same as the original Marching Cubes, where the only difference is the coordinates of cube corners. We will add more detailed description of DMC in the future version. Algorithm 1 summarizes the whole optimization process to optimize the TSDF-Def volume from the given 3D model. Given 3D model, $\mathbf{S}$, we can optimize and obtain its corresponding TSDF-Def volume $\mathbf{V}$ through Algorithm 1. Initially, we distribute grid points $\mathbf{G}$ uniformly across space, serving as the corners of the cubes utilized in DMC. Before optimization, we initialize $\mathbf{V}[...,0]$ as the ground truth TSDF at the location of $\mathbf{G}$ and the deformation as $\mathbf{V}[...,1:3]=0$. During the optimization, the optimal TSDF-Def volume could be optimized by minimizing the difference between the reconstructed shape $\texttt{DMC}(\mathbf{V})$ and the original shape $\mathbf{S}$. In the future version, we will clarify the unclear descriptions to make the optimization process easier to understand. ### Comment 7. In Fig. 4, is ... increase it? Response: * The compression ratio is highly related to the reconstruction error, i.e., a larger compression ratio generally introduces more serious reconstruction errors. Only discussing the compression ratio does not make much sense. In practice, we need to balance compression ratio and reconstruction accuracy according to requirements.\\ To answer your question, We further increase the compression ratio, and the quantitative results on the Mixed dataset are shown in Fig. R4 of the uploaded PDF file. Obviously, when increasing the compression ratio, our NeCGS is still better than baseline methods. ### Comment 8.* Figure 6 is before 5 in the paper. Response: Thanks. We will carefully check the layout. --- Rebuttal 2: Title: The authors are looking forward to your futher feedback. Thanks for your time! Comment: Dear Reviewer jc1U Thanks for your time and effort in reviewing our manuscript and the favorable recommendation. In our previous response, we addressed your remaining concerns directly and comprehensively. We very much look forward to your further feedback on our responses. Best regards, The authors --- Rebuttal Comment 2.1: Title: Additional comments on rebuttal Comment: Thanks for addressing many of my concerns in the rebuttal. I wanted to follow up on comment 3 and the authors response to that because that seems to be the biggest weakness of the paper at the moment. I really appreciate the work that the authors have put in in such a short period of time to performed all the additional experiments. However, I have some additional comments that are important to address. As stated in the paper, the main objective is to propose a mechanism to compress large datasets and not to have this mechanism generalize to other datasets. Because the objective is to compress existing datasets the question of compression time is important. From the results it seems that to achieve a reasonable accuracy, a resolution of 128 or 256 is required and the optimization requires about 24hrs for a dataset containing just 600 geometries (roughly 400Mb). Now, if we consider datasets containing 60000 geometries where compression is truly required this method becomes computationally prohibitive and has no utility. The ability to add new samples to the existing dataset without spending much time on optimization would have been an important feature to prove the value of this method. However, from the results provided in Fig. R1 it seems like the reconstruction accuracy of the new geometries does not seem to be as good as the reconstruction accuracy of the training geometries even when the new geometries seem to be somewhat close to the training distribution. Why is that the case? This is a problem because this alludes to the fact that either the decoder is overfit to the training geometries or the decoder does not have the capacity to represent these geometries. Can you provide the error curves of this optimization to verify that these losses are actually decreasing and it is being able to find the correct embedded vectors representing this geometry? Would it be possible to add more diverse geometries to this experiment and report the reconstruction accuracy in each case? How does GPCC perform for the same geometries? Also, the authors state that the cost of optimization is significantly less, can the authors quantify that? In my experience the number of iterations required for this optimization to converge can be significantly larger. I think these details are important to improve the value proposition of the method. --- Rebuttal 3: Comment: It is great to receive your further feedback, showing that our initial responses have addressed many of your concerns. In the following, we will address your remaining questions. 1. From the results it seems that to achieve a reasonable accuracy, a resolution of 128 or 256 is required and the optimization requires about 24 hrs for a dataset containing just 600 geometries (roughly 400Mb). * As demonstrated in Table 3 of our manuscript, a resolution of 128 is sufficiently precise for the reconstructed meshes, requiring approximately 16 hours to finalize the optimization process. * In the response, we have gathered the compression times of different methods. It is worth noting that VPCC necessitates around 40 hours to finalize compression, which is much more time-consuming than our method. Additionally, the training process for method PCGCv2 is also time-consuming (PCGCv2 requires numerous hours for training. In the Table of the initial response, the time only counts the inference time without considering the training time.). Besides, we have explored multiple techniques to expedite the optimization procedure. 2. The ability to add new samples to the existing dataset without spending much ... training distribution. Can you provide the error curves of this optimization to verify that these losses are actually decreasing and it is being able to find the correct embedded vectors representing this geometry? Would it be possible to add more diverse geometries to this experiment and report the reconstruction accuracy in each case? How does GPCC perform for the same geometries? * The table below displays the reconstruction accuracy for unseen meshes during optimization. Through iterative processes, the precision of the reconstructed model is steadily enhanced. \|Epoch\|CD (1e-3)\|NC\|F1-0.005\|F1-0.01\| \|-----\|-----\|-----\|-----\|-----\| \|100\|6.397\|0.932\|0.506\|0.890\| \|200\|5.618\|0.942\|0.676\|0.944\| \|300\|4.699\|0.947\|0.709\|0.956\| \|400\|4.568\|0.948\|0.722\|0.959\| * Thingi10K comprises a total of 10,000 distinct meshes. Consequently, the unseen meshes sourced from the Thingi10K dataset exhibit greater diversity, with their reconstructed outcome depicted in Figure R1 of the provided PDF. * We evaluate the accuracy of the generalized new meshes, as illustrated in the table below. It is evident that the reconstructed new meshes exhibit greater errors compared to the training meshes. Nonetheless, our method excels in reconstructing the overall shapes and decompressing unseen meshes more accurately than GPCC. \|Data\|CD (1e-3)\|NC\|F1-0.005\|F1-0.01\|Opt. Time (min/per mesh)\| \|-----\|-----\|-----\|-----\|-----\|-----\| \|Seen\|3.436\|0.952\|0.842\|0.991\|1.60\| \|Unseen Ours\|4.568\|0.948\|0.722\|0.959\|1.01\| \|Unseen GPCC\|11.941\|0.912\|0.551\|0.854\|0.06\| 3. Now, if we consider datasets containing 60000 geometries where compression is truly required this method becomes computationally prohibitive and has no utility. * The simplest and most straightforward approach involves grouping the samples of the dataset, compressing each group separately in parallel, thereby cutting down on compression time. * By solely optimizing the embedded features and keeping the decoder weights fixed for new meshes, the average optimization time per new mesh significantly reduces, offering a fresh approach to compressing extensive geometric data. To achieve this, 1) Initially, a small subset of data is chosen for optimizing the embedded features and decoder weights. 2) Subsequently, the decoder weights are set, and only the embedded features are optimized for the remaining meshes, enabling rough reconstruction post-optimization. 3) The embedded features of all meshes are refined, along with the decoder weights, over several epochs. This three-stage optimization strategy, as opposed to directly optimizing all embedded features and decoder weights, results in considerable time savings. --- Rebuttal Comment 3.1: Comment: Dear Reviewer jc1U Thank you for dedicating your time and effort to reviewing our submission. We hope our thorough responses have effectively addressed your additional comments. We would greatly appreciate hearing from you before the impending discussion deadline. Best regards, The authors	Summary: This paper proposes a method to compress 3D geometry of diverse categories of objects. In the first step, the paper proposes a method to first convert an irregular mesh to a regular representation like a 4D TSDF-Def volume that implicitly describes the geometry. After this, an auto-decoder is trained that learns to reconstruct the 4D TSDF-Def volume from a compressed feature vector which is unique for each shape. Hence, with this design the model can summarize the similarity of local geometric structures within and across different 3D meshes resulting in a compact representation. Results on AMA, DT4D and Thingi10K datasets shows that the model can achieve compression of 3D models to a reasonable extent. Strengths: 1) Clarity: the paper is well written with each component of the method explained clearly which is easy to understand. 2) Reproducibility: All the details to replicate the results are provided along with the code and architecture details in the supplementary material. Weaknesses: 1. The intuition behind preferring TSDF-Def 4D volume over TSDF 3D volume is unclear, even though an ablation study shows better reconstruction for thin structures. The quantitative results in Table 2 only show marginal improvements. An brief intuitive explanation of the design choice is helpful. 2. There are lot of methods which try to compress a neural field. For e.g. Triplanes[1], HashGrid [2], Vector Quantization [3], TensoRF [4], Dictionary Fields [5]. It is not very clear why this method does not compare with all these techniques which can be used for compression? 3. Can this method generalize? Can I use the trained auto-decoder setting to compress a new 3D mesh on which the model is not trained on? How about other methods with which the method compares. 4. The paper does not do a relative comparison of the compression time with the baseline methods. Given the optimization time shown in Table 3, I have concerns about the practical usage of this method. [1] Peng, Songyou, et al. "Convolutional occupancy networks." ECCV, 2020. \ [2] Müller, T., Evans, A., Schied, C., & Keller, A. (2022). Instant neural graphics primitives with a multiresolution hash encoding. ACM TOG, 2022. \ [3] Takikawa, Towaki, et al. "Variable bitrate neural fields." ACM SIGGRAPH, 2022. \ [4] Chen, Anpei, et al. "Tensorf: Tensorial radiance fields." ECCV, 2022. \ [5] Chen, Anpei, et al. "Dictionary fields: Learning a neural basis decomposition." ACM TOG, 2023. Technical Quality: 3 Clarity: 3 Questions for Authors: Please refer to weakness above. Line 123, "mplicitly" should be "implicitly". Confidence: 5 Soundness: 3 Presentation: 3 Contribution: 2 Limitations: Limitations are adequately discussed. Flag For Ethics Review: ['No ethics review needed.'] Rating: 5 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: ### Comment 1. The intuition behind preferring TSDF-Def 4D volume over TSDF 3D volume is unclear, ... The quantitative results in Table 2 only show marginal improvements. An brief intuitive explanation of the design choice is helpful. Response: * For a geometry dataset usually containing 3D models with various simple and complex structures, TSDF can represent 3D models with simple structures but fails to accurately represent them with complex ones when the resolution is relatively limited. Differently, we propose TSDF-Def 4D volume by assigning additional offsets at each corner of the cubes, so that it can well represent 3D models with both simple and complex structures. * As Fig.7 shows, compared to TSDF volume, our TSDF-Def volume can handle thin structures of the shapes and has demonstrated much better performance. In Table 2, the quantitative advantage of our TSDF-Def over TSDF is marginal because the results are the average error of the entire dataset. * To address your concern, we divide the dataset into two parts when evaluating compression performance: (1) the shapes with thin and fine structures (20 shapes); and (2) those without detailed structures. * The quantitative results are shown in the following table, where it can be seen that our TSDF-Def volumes can significantly reduce the distortion for the models with thing structures. It is obvious that the advantage of our TSDF-Def 4D volume would be more significant than the TSDF 3D volume on the datasets with more complex and fine structures. Fig. R3 demonstrates the selected models. \|Representation\|Data\|CD(1E-3)\|NC\|F1-0.005\|F1-0.01\| \|----\|----\|----\|----\|----\|----\| \|TSDF\|All shapes\|5.015\|0.944\|0.662\|0.936\| \|TSDF\|Shapes w/ thin structures\|11.628\|0.454\|0.728\|0.861\| \|TSDF\|Shapes w/o thing structures\|4.786\|0.670\|0.943\|0.947\| \|TSDF-Def\|All shapes\|4.913\|0.947\|0.674\|0.943\| \|TSDF-Def\|Shapes w/ thin structures\|8.751\|0.506\|0.800\|0.874\| \|TSDF-Def\|Shapes w/o thing structures\|4.783\|0.680\|0.948\|0.950\| ### Comment 2. There are lot of methods which try to compress a neural field. For e.g. Triplanes[1], HashGrid [2], Vector Quantization [3], TensoRF [4], Dictionary Fields [5]. It is not very clear why this method does not compare with all these techniques which can be used for compression? Response: The methods in [1-5] are focused on the compact representation of feature volumes, where the continuous feature of any point in the space could be obtained through bilinear or trilinear interpolation. Although these methods could be used to compress geometry models, the required storage space can be as high as several MBs, as shown in Fig.7 of [2], Fig.5 of [5], and so on, which is much larger than our method (averaging only a few KBs per model). When reducing the feature dimension to make the feature to be KBs, these methods fail to recover shapes from their implicit fields. We believe that these algorithms can be combined with some compression techniques and additional efforts in the future to achieve more efficient compression. For these reasons, our current baseline methods do not include these methods. ### Comment 3. Can this method generalize? Can I use the trained auto-decoder setting to compress a new 3D mesh on which the model is not trained on? How about other methods with which the method compares. Response: * Thank you for your insightful comment! Yes, our NeCGS can be generalized to new 3D models. Specifically, given a new 3D model, we first represent it into a TSDF-Def 4D volume through Algorithm 1. Then following the optimization progress in Sec. 3.2, we only optimize its corresponding embedded feature while keeping the trained decoder unchanged. * The visual results are shown in Fig. R1 of the uploaded PDF, showing the ability to generalize. It is not surprising that our NeCGS can be generalized to new 3D models because after fitting 3D models with various structures, the decoder can learn the prior knowledge of various geometry data, and it can represent unseen models by only optimizing its embedded features. The methods under comparison can also be adapted to new 3D models. * GPCC, VPCC, and Draco are traditional compression methods that do not require training, enabling them to compress new meshes. PCGCv2 could utilizes the trained encoder to compress the new models directly. QuantDeepSDF operates as an auto-decoder framework, allowing it to generalize new models similarly to our NeCGS. ### Comment 4. The paper does not do a relative comparison of the compression time with the baseline methods. Given ..., I have concerns about the practical usage of this method. Response: * As shown in the table below, compared to compression methods for single shapes, i.e., GPCC, PCGCv2, and Draco, our method requires more time. While compared with VPCC and QuantDeepSDF which can compress the entire dataset at once, the compression process of our NeCGS is faster. We also want to note that according to the quantitative comparison in Fig. 4 and qualitative comparison in Fig. 5 of the manuscript, the compression performance of our method is significantly better than that of the baseline methods. \|Method\|Compression Time (h)\| \|-----\|----\| \|GPCC\| 0.625 \| \|VPCC\| 39.34 \| \|PCGCv2\| 1.76 \| \|Draco\| 0.03 \| \|QuantDeepSDF\| 18.91 \| \|Ours\|16.32 \| * Again, we want to emphasize that our NecGS is designed for the offline compression of 3D geometry datasets to save storage space, where the optimization/compression time should not be a key factor. Instead, we should concentrate more on the decompression/decoding speed because the users expect to obtain the 3D models timely when querying the dataset. As shown in Table 3, when the resolution is 128, the inference time of our NeCGS is only 98.95ms, satisfying the real-time requirement. ### Comment 5. Line 123, "mplicitly" should be "implicitly". Response: Thanks. We will correct this typo. --- Rebuttal 2: Title: The authors are waiting for your further feedback. Let's discuss. Comment: Dear Reviewer qNSe Thanks for your time and effort in reviewing our manuscript. In our previous response, we addressed your concerns directly and comprehensively. We very much look forward to your further feedback on our responses. Let us discuss. Best regards, The authors --- Rebuttal Comment 2.1: Comment: I thank the authors for the rebuttal. After carefully reading the rebuttal by the authors and comments by the reviewers, I can confirm that my concerns about generalization and compression time is resolved i.e. although the model has long training time, it can compress the entire dataset at once. Further, as mentioned and showed in the PDF the model can also generalize significantly faster to new shapes. This is beneficial. However, I am not fully convinced by the authors response on compressed neural fields (comment 2). The authors pointed out visual results in Fig 5 of [5]. However, I don't feel it is an apple to apple comparison as the geometry complexity of the model in Fig 5 of [5] is much higher than the complexity of models shown in this paper. Hence, I strongly suggest the authors to show a comparison with at least 1 method (the best one) for further insights. In addition to this, I also suggest the authors to benchmark the models generalization vs compression time trade off on a larger and complex pool of shapes to get a better picture. Having said this, most of my major concerns have been resolved and hence I am willing to increase my score to borderline accept! --- Rebuttal 3: Comment: It is wonderful to get your further feedback mentioning that our initial responses have effectively tackled your concerns. The authors appreciate your favorable recommendation with the highest confidence. Moving forward, we will tackle the remaining questions you have. Comment 1. However, I don't feel it is an apple to apple comparison as the geometry complexity of the model in Fig 5 of [5] is much higher than the complexity of models shown in this paper. Response As shown in our response, we have preliminarily conducted the experiment about [2] and [5] using their released codes on the Thingi10K dataset, when reducing the feature dimension to make the model/feature size to KBs, we cannot extract meshes from them. We will also explore them comprehensively and fairly in the final version. Comment 2. In addition to this, I also suggest the authors to benchmark the models generalization vs compression time trade off on a larger and complex pool of shapes to get a better picture. Response Thanks for your valuable comments! We will test the compression performance and generalization of our algorithm on larger and more complex shapes in the final version. Additionally, we will build a benchmark through these data to support the advancement of this field. Finally, we appreciate the valuable comments and timely feedback from the reviewers.	Summary: this paper looks at the problem of compressing 3d shapes (esp geometry). this paper proposes a two stage approach. the first stage is regular geometry representation. the second stage is compact neural compression. results show some improvements. Strengths: 1. compressing 3d shapes is important to many applications Weaknesses: 1. this paper over claims what it does. in L1-3, it says that they made the first attempt to tackle the problem of compressing 3D geometry sets containing diverse categories. this isn't true. there are at least two papers doing geometry compression of 3D geometry [a], [b]. [a] On the Effectiveness of Weight-Encoded Neural Implicit 3D Shapes https://arxiv.org/abs/2009.09808 [b] Neural Progressive Meshes https://arxiv.org/abs/2308.05741 2. [a] and [b] are very important references but they are not cited nor discussed. it's not necessary to compare the proposed method with [a] and [b], but at least the authors should acknowledge the existence of these two papers. 3. optimization time is too long 4. it is unclear whether the proposed method is reproducible 5. typo L43: Matching cubes -> Marching cubes Technical Quality: 2 Clarity: 2 Questions for Authors: see comments above Confidence: 5 Soundness: 2 Presentation: 2 Contribution: 2 Limitations: yes Flag For Ethics Review: ['No ethics review needed.'] Rating: 2 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: ### Comment 1. this paper over claims what it does. in L1-3, it says that they made the first attempt to tackle the problem of compressing 3D geometry sets containing diverse categories. this isn't true. there are at least two papers doing geometry compression of 3D geometry [a], [b]. Response: We strongly disagree with you, due to the following facts. * After a comprehensive survey, we confirm that all previous methods for 3D geometry compression, including the mentioned [a] and [b] for processing single 3D shapes, have primarily focused on either individual or sequential/dynamic 3D models. Different from previous methods, our NeCGS is designed for the offline compression of 3D geometry sets comprising various unrelated 3D shapes. * Technically, our NeCGS is significantly different from [a] and [b]. Specifically, [a] employs an individual MLP to regress the implicit field of a single model, using the MLP weights to represent the model. Meanwhile, [b] is designed for mesh simplification and restoration through triangle merge and divide, with the simplified meshes serving as representations of the original meshes. Differently, to well handle each model of a dataset usually containing 3D models with various structures, we propose innovative TSDF-Def volumes to represent 3D models as structured 4D tensors. TSDF-Def can not only represent 3D models with relatively simple structures like TSDF but also well preserve intricate details of 3D models with complex and fine structures. After optimizing each shape into a TSDF-Def volume, we design an auto-decoder structure to regress these tensors, where the embedded features and decoder weights are quantized during the optimization and encoded to bitstreams to represent the whole geometry set. ### Comment 2. [a] and [b] are very important references but they are not cited nor discussed. it's not necessary to compare the proposed method with [a] and [b], but at least the authors should acknowledge the existence of these two papers Response: * The authors confirm that [a] and [b] are not the most relevant references. But we are fine with citing these two papers in the final version. * In addition to the response to Comment 1 that has clearly elaborated the differences between ours and [a][b], we further clarify that both [a] and [b] emphasize geometric representation over compression techniques. Besides, there are so many methods for compressing single 3D models, and we have cited well-known methods; however, due to page limitations, it is impossible to cite all of them. ### Comment 3. optimization time is too long Response: * We must emphasize that our NeCGS is focused on the offline compression of geometry datasets to save storage space, where the optimization/compression/encoding time should not be a key factor. Instead, we should concentrate more on the decompression/decoding speed because the users expect to obtain the 3D models timely when querying the dataset. As shown in Table 3, the decompression process of our NeCGS is extremely fast, allowing for real-time invocation. * Moreover, there are various potential solutions to accelerate the optimization. At the software level, more efficient convolution (e.g., using multiple 1-D or 2-D convolution to approximate the 3-D convolution or using 3D sparse convolution to replace the traditional 3D convolution) can be used to speed up the process. At the hardware level, optimization programs can be run using multiple GPUs or other more efficient hardware. Like NeRF, the initial algorithms required several days for optimization. Subsequently, improved methods have been introduced to accelerate the optimization process to a few minutes or even seconds. ### Comment 4. it is unclear whether the proposed method is reproducible Response: * We have included the source code in the submitted supplementary material, as highlighted in the abstract, * We have provided sufficient implementation details in Sec. 4.1. Based on the information provided, we believe any qualified researcher can replicate the results in our paper. * Reviewer qNSe acknowledged the reproducibility aspect by stating, 'Reproducibility: All the details to replicate the results are provided along with the code and architecture details in the supplementary material'. And Reviewer 18DA also recognize the reproducibility by stating, 'The inclusion of source code in the supplemental material enhances the reproducibility and transparency of the research'. ### Comment 5. typo L43: Matching cubes $\to$ Marching cubes Response: Thanks for this valuable comment. We will correct the typo. --- Rebuttal 2: Title: The authors are looking forward to your further feedback. Let's discuss! Comment: Dear Reviewer tAuE Thanks for your time and effort in reviewing our manuscript. In our previous response, we addressed your concerns directly and comprehensively. We very much look forward to your further feedback on our responses. Let us discuss. Best regards, The authors --- Rebuttal 3: Comment: RE: Response to comment 1. I do not think the authors fully understand the references I pointed out. Both [a] and [b] can do offline compression and were tested on Thingi10K which consists of various unrelated 3D shapes. > We further clarify that both [a] and [b] emphasize geometric representation over compression techniques I do not see how representations and compression techniques are disentangled in this case. Choosing the right representation is part of the compression technique. RE: Response to comment 3. It doesn't really matter whether your method is online or offline. If one wants to apply your method to some new shapes and they don't want to lose accuracy, they'll have to go through the optimization process and this process takes a lot of time. It doesn't really matter if your method can uncompress fast. --- Rebuttal Comment 3.1: Comment: Dear Reviewer tAuE It is great to receive your further feedback so that we can discuss your misunderstandings comprehensively. 1) We believe we understood references [a] and [b] exactly, but you have seriously misunderstood them. We suggest you read the two papers carefully again. In your first post, you mentioned 'If you look at Figure 4 of [a], you'll see that their method is not one MLP for one shape.' Note it was posted first at 3:36 pm, but deleted at 3:42 pm. Thus, this sentence is not on the Openreview but appears in the email automatically sent to authors. It is true that [a] was applied to compress a set of 3D models (i.e., 10,000 3D models) from Thingi10K. [a] compresses the 3D models one by one, i.e., an MLP is regressed for a 3D model independently, so there are 10,000 MLPs regressed for the 10,000 3D models. Moreover, we refer you to the GitHub link of [a] (https://github.com/u2ni/ICML2021/tree/main/thingi10k-weightEncoded), where 10,000 files are presented to store the optimized parameters for each shape. [b] utilizes an encoder-decoder framework to achieve mesh simplification and restoration, as shown in Fig. 2 of [b], and it simplifies the meshes one by one, like the baseline method PCGCv2. By contrast, our method regresses a single auto-decoder (i.e., a 3D CNN) for all 3D models involved in a set. Such a shared network by all 3D models can explore the redundancy/local similarity among different 3D models to some extent. 2) As for a standard compression framework, the input is the raw data and the output is its bitstream, where technologies such as quantization and entropy coding are used to reduce the storage. [a] utilizes a tensorflow framework, where the network parameters are stored in h5 files (see L214 at https://github.com/u2ni/ICML2021/blob/main/neuralImplicitTools/src/model.py). The h5 file is different from the bitstream used in the compression field since it would include the names of variables and other attributes. [b] compresses/represents the raw data into simplified triangle meshes, which is significantly different from the bitstream. Thus we clarify that both [a] and [b] emphasize geometric representation over compression techniques. 3) More importantly, in your post, you mentioned '...Choosing the right representation is part of the compression technique', but you have completely ignored/overlooked one of our contributions, i.e., the proposed TSDF-Def representation converting any irregular 3D meshes into regular 4D volumes with fine structures well preserved. 4) ' ...If one wants to apply your method to some new shapes and they don't want to lose accuracy, they'll have to go through the optimization process and this process takes a lot of time.' We also disagree with this comment by you. We have conducted additional experiments to demonstrate the generalization of the trained decoder for new meshes. Specifically, given a new mesh, we only optimize the embedded features for the new mesh and keep the weights of the trained decoder fixed. Such an optimization process is very time-saving, and the decompressed models are accurate enough, as shown in Fig. R1. of the uploaded one-page PDF file. As discussed in our rebuttal, it is not surprising that our NeCGS can be generalized to new 3D models because after fitting 3D models with various structures, the decoder can learn the prior knowledge of various geometry data, and it can represent unseen models by only optimizing its embedded features. The methods under comparison can also be adapted to new 3D models. 5) 'It doesn't really matter if your method can uncompress fast.' We argue that the decompression speed DOES Matter. Imagine that given a set of 3D models that are stored in compressed bitstreams to save space, if you want to get the 3D models for downstream analysis or applications, the slow decompression process will seriously limit the efficiency of the downstream process. --- Rebuttal 4: Comment: Thanks for your quick action. Comment 1. If you could look at [b], they have one shared network for all 3D models. Response You still misunderstood [b]. Could you please take the time to read it carefully? As replied in our previous posts, [b] simplifies the meshes one by one, like the compared baseline method PCGCv2. More specifically, [b] trains a shared encoder to simplify the raw meshes one by one, like PCGCv2. [b] and our NeCGS have totally different working mechanisms. Our NeCGS utilizes an auto-decoder framework, where the embedded features and decoder weights are optimized to represent the whole dataset. Besides, our method concentrates on implicit representation, while [b] concentrates on explicit representation, i.e., triangle merge and divide. We believe your confusion is completely caused by your misunderstanding about these methods and lack of knowledge of this topic. Comment 2. Every reviewer raised different questions and it requires the authors to post such long responses to clarify things. Response We are wondering whether you have read and fully understood their comments and our responses. We answered all questions directly and comprehensively and provided additional necessary experiments. Comment 3. Not to mention, in this case, we already have a few back-and-forth discussions but there is still a lot of confusion. I strongly suggest the authors revise the way of presentation. Response We believe all confusion is due to your lack of relevant foundational knowledge and failure to carefully read the related papers. We believe our explanations have been sufficiently clear, and we strongly recommend you carefully read the paper, supplement the necessary foundational knowledge, and become a more professional reviewer, so that the entire community can progress more effectively.	Rebuttal 1: Rebuttal: We thank the reviewers for the time and effort in reviewing our work, as well as your recognition of the novelty of our work. We are grateful to the reviewers for acknowledging our NeCGS algorithm: 1. Reviewers qNSe, jc1U, and 18DA have all noted the significant compression performance of our NeCGS on various datasets, showcasing its effectiveness; 2. Reviewer 18DA recognizes our method as a notable advancement in the field of geometry compression; 3. Reviewers qNSe and 18DA have acknowledged the reproducibility of our NeCGS; 4. Reviewers qNSe and 18DA have praised the clarity and accessibility of our writing. In this work, we present a new compression framework, namely NeCGS. Different from previous compression methods for compressing either individual 3D models or sequential/dynamic 3D sequences, our NeCGS is the first method focused on the offline compression of 3D geometry datasets usually containing diverse and unrelated 3D models with various structures. Technically, to handle each model of a dataset well, we propose innovative TSDF-Def volumes to represent 3D models as structured 4D tensors. TSDF-Def excels in depicting not only straightforwardly structured 3D models like TSDF but also in accurately preserving intricate details of complex and finely structured 3D models. The extensive experiments demonstrate the significant superiority of our NeCGS over state-of-the-art ones. In the following, we will respond to the main comments that the reviewers are concerned about. ### Compression/Optimization/Encoding Time * We have to emphasize that our NeCGS is focused on the offline compression of geometry datasets to save storage space, where the compression (or encoding) time should not be considered as a key factor. Instead, we should concentrate more on the decompression speed because the users expect to obtain the 3D models timely when querying the dataset. As shown in Table 3, the decompression process of our NeCGS is extremely fast, allowing for real-time invocation. * Moreover, there are various potential solutions to accelerate the optimization. At the software level, more efficient convolution (e.g., using multiple 1-D or 2-D convolution to approximate the 3-D convolution or using 3D sparse convolution to replace the traditional 3D convolution) can be used to speed up the process. At the hardware level, optimization programs can be run using multiple GPUs or other more efficient hardware. Like NeRF, the initial algorithms required several days for optimization. Subsequently, improved methods have been introduced to accelerate the optimization process to a few minutes or even seconds. * As shown in the table below, compared to compression methods for single shapes, i.e., GPCC, PCGCv2, and Draco, our method requires more time. While compared with VPCC and QuantDeepSDF which can compress the entire dataset at once, the compression process of our NeCGS is faster. * Finally, We also want to note that according to the quantitative comparison in Fig. 4 and qualitative comparison in Fig. 5 of the manuscript, the compression performance of our method is significantly better than that of the baseline methods. Such advantages were also acknowledged by all reviewers. \|Method\|Compression Time (h)\| \|-----\|----\| \|GPCC\| 0.625 \| \|VPCC\| 39.34 \| \|PCGCv2\| 1.76 \| \|Draco\| 0.03 \| \|QuantDeepSDF\| 18.91 \| \|Ours\|16.32 \| ### Generalize Unseen Models * We confirm that our NeCGS can be generalized to new 3D models. Specifically, given a new 3D model, we first represent it into a TSDF-Def 4D volume through Algorithm 1 of our manuscript. Then following the optimization progress in Sec. 3.2, we only optimize its corresponding embedded feature while keeping the trained decoder unchanged. * In Fig. R1 of the uploaded PDF file, we also experimentally demonstrate this generalization ability. It is not surprising that our NeCGS can be generalized to new 3D models because after fitting 3D models with various structures, the decoder can learn the prior knowledge of various geometry data, and it can represent unseen models by only optimizing its embedded features. The methods under comparison can also be adapted to new 3D models. Last but not least, we will make the reviews and author discussion public regardless of the final decision, Besides, we will include the newly added experiments and analysis in the final manuscript/supplementary material. Thanks again for your time and effort in our submission. We appreciate any further questions and discussions Pdf: /pdf/f2372ae503852b7a52265761b3082da9a569f3c3.pdf	NeurIPS_2024_submissions_huggingface	2,024	null	null	null	null	null	null	null	null
FewViewGS: Gaussian Splatting with Few View Matching and Multi-stage Training	Accept (poster)	Summary: This paper presents FewViewGS to regularize sparse view 3DGS from unseen viewpoints without relying on pre-trained depth estimators. The main contribution is to reproject the pixels to an unseen view and calculate the losses at corresponding pixels found by image matching. Experiments on LLFF, DTU, and Blender datasets demonstrate the proposed method can achieve competitive or superior performance compared to existing SOTA methods. Strengths: - The method can exceed existing baselines without relying on extra depth estimators. - Results are good in visualization. Weaknesses: - The main contribution of this paper shares the same insight with SPARF [1], which also uses image matching to add supervision on unseen views. The main difference is that SPARF uses depth, while this work uses depth, color and semantic features, for which the novelty in this is limited. And also the multi-stage training is just a common composition of " initialization + regular training + refinement". - The proposed equations have too many manual factors and artificial designs that lack mathematical theory (e.g. Eq (7, 8,13)). These all seem to be engineering attempts but not technical contributions. - Experiment is insufficient and unclear. 1) Although a (Rand. Init.) is mentioned at one item in Table 1, I can not find any description of the used initialization for other settings, which may have a very huge influence on the performance. 2) And serving as a geometry reconstruction method, some geometry visualizations like depth are wished to report. 3) As the reprojection operation and semantic feature extraction may cost a lot of time, I'm wondering if the training time would significantly increase. - Performance improvements are somewhat limited compared to current methods. And although this work can be free from some pre-trained models, it has introduced other dependencies like VGG features and RoMa matching. - The paper has some errors, for example: 1) Eq (6) is unclear on the part (C, Z), which is a 2D->3D back projection I guess. 2) The citation information of SPARF [1] (numbered as [34] in the paper) is wrong. [1] Truong, Prune, et al. "Sparf: Neural radiance fields from sparse and noisy poses." Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2023. Technical Quality: 2 Clarity: 2 Questions for Authors: Please clarify the weaknesses, and make additional explanations of the unclear parts. Confidence: 5 Soundness: 2 Presentation: 2 Contribution: 2 Limitations: The authors have not discussed the limitations. This work seems to heavily rely on some manual hyperparameters, and some pre-trained models like VGG features and RoMa matching, which may limit the applicability to more real-world scenes. Flag For Ethics Review: ['No ethics review needed.'] Rating: 3 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We thank the reviewer for the insightful review. Below, we address the main concerns raised in this review. *Q: The main contribution of this paper shares the same insight with SPARF.* *A:* Although both our method and SPARF target solving the overfitting issue, there are key differences that contribute to our novelty and effectiveness: (1) Feature matching: Our method samples novel views and projects matched pixels to the novel view for supervision. However, SPARF only adds additional supervision on matched pixels for training views, without involving any novel views. (2) Supervision of novel views: Our method projects two training views to generate supervision for novel views, improving the reliability of supervision. This contrasts with SPARF's projection including one training view, which can suffer from noises in depth estimation. Our experiments in Table 4ix and 4xii show that our method with two-view projection for novel view supervision notably outperforms the single-view projection. We appreciate the reviewers' suggestions and will add these comparisons to our revised version. *Q: The multi-stage training is just a common composition of " initialization + regular training + refinement".* *A:* While we follow a common training composition, our 3-stage method aims to address overfitting to sparse training views and improve supervision reliability in novel views. Particularly, our approach introduces a novel method to integrate additional view information during the intermediate training stage. We apply feature matching to identify corresponding pixels, which are then projected onto novel views to generate pseudo labels. This approach further uses knowledge from training views, ensures multi-view consistency, enhances reliability during projection, and solves overfitting. Besides, our proposed locality-preserving regularization maintains the local smoothness of 3D Gaussians, enhancing the effectiveness of our method. *Q: The proposed equations have too many manual factors and artificial designs that lack mathematical theory (e.g. Eq (7, 8,13)).* *A:. (1) Hyperparameters: We apologize for missing the relevant ablation study. Table S2 - Table S6 of the uploaded file provide additional results, showing that our method performs consistently well across different settings. We will include these results in our revised appendix. (2) Mathematical theory: The proposed equations are designed based on the principles of multi-view projection (Eqs 7 and 8) and the theory of local smoothness in 3D space (Eq 13). Eq 7* is based on the principle that matched 2D pixels should map to the same 3D point. Thus, it measures the distance between 3D points to filter out noisy projections, which can solve depth inaccuracies in multi-view projection. Eq 8 tackles the problem of blurry results in texture-rich regions with large gradients during projection, which are sensitive to projection errors. Eq 13, incorporating local constraints using the KNN algorithm, is based on spatially local smoothness. *Q: 1) I can not find any description of the used initialization for other settings. 2) Some geometry visualizations like depth are wished to report. 3) I'm wondering if the training time would significantly increase.* *A:* We appreciate the reviewer's suggestions on experiments. We will incorporate additional details and results in our revised version. (1) Initialization: In Table 1, our method with random and SFM initialization is listed in the last two rows. In the ablation study, the random initialization was adopted. (2) Geometry visualization: In Figure S2 of the uploaded file, our method generates clearer and more accurate depths compared to the baseline 3DGS, demonstrating its effectiveness. (3) Training time: Table S7 of the uploaded file compares the training time of 3DGS-based methods. Our method has comparable training time and is faster than SparseGS (which uses a pre-trained diffusion model) and FSGS (which relies on a pre-trained depth network). *Q: Performance improvements are somewhat limited compared to current methods. And although this work can be free from some pre-trained models, it has introduced other dependencies like VGG features and RoMa matching.* *A:* (1) Performance: In Table 1, our method surpasses 3DGS-based methods over all metrics on both DTU and LLFF datasets. Specifically, on DTU, compared to SparseGS, our method improves PSNR by 4.5\% and SSIM by 22.6\%. These results demonstrate the effectiveness of our method. We will integrate additional relevant descriptions to our manuscript. (2) Pre-trained networks: Previous methods, such as FSGS and SparseGS, use pre-trained depth/diffusion models to create pseudo labels for novel views, which may be noisy. In contrast, our method takes a different approach by projecting the color data (ground truth) from training views to novel views to yield pseudo labels, which is more accurate than the labels produced by other networks. Specifically, the accuracy of this projection is maintained using RoMa matching, a technique that finds correspondence between training views. In addition to color supervision, our method also employs the VGG network to provide further semantic supervision for novel views. These techniques help maintain multi-view consistency and prevent overfitting. *Q: 1) Eq (6) is unclear on the part (C, Z), which is a 2D->3D back projection I guess. 2) The citation information of SPARF is wrong.* *A:* We apologize for the confusion caused. In Eq 6, as explained in Lines 159-164, C denotes 2D coordinates of pixels, while Z refers to their depth values. The process in Eq 6 involves a 2D-3D projection (i.e. $\pi^{-1}(C_i, Z_i)$) and a 3D-2D projection (i.e. $\pi(P_k \cdot)$). We will make it more clear in the next version. (2) Thank you for pointing that out. We will check the citations and ensure they are correct. --- Rebuttal Comment 1.1: Comment: Thanks for the reply from the authors. After reading the rebuttal, this work still has the following problems. 1. Lack of Novelty. Summarized by the authors in the paper, this work has three main contributions: an effective multi-stage training scheme, novel view consistency constraints, and a color smooth regularization loss. First, as in the review, the multi-stage training is just a common composition of " initialization + regular training + refinement", with no independent novelty on this point. And the color smooth regularization loss is just a trick to add local smoothness to reduce some artifacts, which only improved PSNR but not SSIM and LPIPS, demonstrating its limited effect on the overall quality. If these are declared to be the most important novelties in the paper, I can only believe there are no other valuable things to show. And for the novel view consistency constraints, it is still the incremental technique of SPARF even according to the rebuttal. The contribution lies only in some loss designs with no interesting insight, moreover, has too many manual hyperparameters in eq(7,9,10,11) and replaceable mapping functions like $exp()$ in eq(8, 13). Besides, L_sem is less effect according to Table 4 (vi, ix, xiv) and Figure 4. Especially, some details became worse after adding L_sem in Figure 4. It is also sensitive to the type of feature. I do not think these novelties can match the bar at NeurIPS. 2) Details of "SFM initialization" is still not clear enough. Please explain how many views are involved in the SfM process to get the camera poses and point cloud, how to align the excluded test views into the camera coordinate system, and whether is there any post-processing of the point cloud. As far as I know, original 3DGS can not recover such precise scenes shown in Figure S2 when just use 3-view SfM point cloud as initialization, and COLMAP will fail in some DTU scenes when using only three training views. I have doubts about these. 3) The authors claimed introducing depth or diffusion models will "make training time longer and highly dependent on the quality of the pre-trained models" (lines 40-41), which is the stem of this work. However, why introducing pre-trained RoMa and VGG can escape from these weaknesses? The logic is weird. I'll temporarily keep my rating. --- Rebuttal 2: Comment: We sincerely appreciate your time and efforts in reviewing our responses. We hope that our responses below adequately address your concerns. *Q1: Clarification of our novelty* *A1:* In this response, we would like to clarify our novelty and design details. (1) Novel view consistency constraints: Our method offers a new solution to address overfitting and ensure multi-view consistency in sparse view synthesis by using correspondence priors to generate labels for novel views. Since accurate depth prediction is crucial for reliable view projection, our method projects matched pairs from training views to the novel view and filters out pairs with significant depth discrepancies in the novel view (see Eq 7). This design enhances label reliability over SPARF's single-view projection. The table below shows the results of SPARF's multi-view correspondence loss (MV-Corr) and depth consistency loss (DCons) on 3DGS, which is much worse than our method (PSNR 19.13, SSIM 0.792, LPIPS 0.186). Furthermore, DCons in SPARF shows limited improvement, see Row i and Row ii. In contrast, our depth supervision for novel views in Table 4iii (PSNR 18.17, SSIM 0.736, and LPIPS 0.198) greatly improves the performance of 3DGS. \|\| Method \| PSNR \| SSIM \|LPIPS \| \| --- \| --- \| --- \| ---\| --- \| \| i\| 3DGS \| 15.04 \| 0.676 \| 0.246 \| \| ii \| + DCons \| 15.17 \| 0.699 \| 0.232 \| \| iii\| + MV-Corr \| 16.25 \| 0.722 \| 0.219 \| \| iv\| + MV-Corr + DCons \| 16.37 \| 0.729\| 0.214 \| (2) Multi-stage training: Common multi-stage training typically involves an additional teacher network or large dataset. In contrast, our approach tailors the multi-stage training to sparse view synthesis, eliminating the need for extra networks or datasets. It alternates between training and novel views, expanding the view information and solving overfitting. As shown in Table 4ix and 4xi, this multi-stage approach improves the PSNR by 4.4\%. (3) Locality preserving regularization: Our locality regularization explicitly constrains the color attributes of 3D Gaussians to preserve local smoothness in 3D space. Table 2 demonstrates reduced artifacts and improved accuracy, and Table 4i and 4ii highlight gains in all metrics. (4) Loss design: Our proposed loss functions integrate additional constraints directly, which is intuitive and effective to address issues in sparse view synthesis, without increasing inference time. (5) $L_{sem}$: Our work introduces the semantic constraint to integrate local constraints and reduce artifacts. The comparisons between Table 4i and 4v show a 3.2\% increase in PSNR and a 4.5\% reduction in LPIPS. Indeed, as discussed in Lines 272-277, the local constraint needs to balance the benefits of local semantics with the risk of missing details. (6) Other details: We apply consistent values for $\theta$ and $\theta_{grad}$ in Eqs (9, 10, 11), to avoid too many human factors. Fine-tuning these parameters and exploring better mapping functions could improve results. *Q2: Details of SFM initialization* *A2:* The code for SFM process will be released, providing all the details. (1) SfM initialization: We follow the SfM process provided in FSGS ('tools/colmap\_llff.py' in FSGS code) to generate camera poses and 3D points, using only training views without any post-processing. The choice of training/test views is consistent with FSGS on LLFF and DNGaussian on DTU. Novel views are randomly interpolated between training views. (2) Visualizations: Figure S2 shows 3DGS results generated with SfM initialization, a 3-view setting, and size_threshold=None (as used in FSGS and DNGaussian). We will release the model for further clarification. (3) COLMAP in DTU: Following FSGS, we use the dense reconstruction in COLMAP to generate 3D points, with only training views, which successfully yields point clouds on all scenes except 'scan110'. They are used to initialize 3D Gaussians. 'Scan110' is initialized randomly. *Q3: Comparison with other works based on pre-trained depth/diffusion models.* *A3:* We try to clarify the confusion below. (1) Training time: We pre-compute and store VGG features and RoMa-matched pairs during data pre-processing, avoiding extra training time and redundant computations. However, pre-trained depth/diffusion models generate results online, making them more computationally expensive than pre-trained Roma and VGG models. (2) Quality: Networks using pre-trained depth/diffusion models generate labels for the novel view often ignore the priors in training views and lack multi-view consistency. In contrast, our method projects labels from training views to novel views, ensuring higher accuracy and multi-view consistency. Using a pre-trained depth network for novel views on 3DGS yielded PSNR 16.74, SSIM 0.734, LPIPS 0.214. Our method with depth supervision alone in Table 4iii (PSNR 18.17, SSIM 0.736, and LPIPS 0.198) outperforms this. Title: Reply to Reviewer oLUf --- Rebuttal Comment 2.1: Comment: I'll respond to some important problems, but not mean that other parts are good: 1. For novel view consistency constraints: According to the response, the authors agree that their work lacks novelty, since the novelty claimed by the authors, eq (7), is just a simple filter mask supported by a manual hyperparameter, while the other parts in "novel view consistency constraints“ is strongly similar to SPARF. on the part of the effect, the provided reproduction experiment of SPARF on 3DGS is not convincing for me, as there is not any explanation about why the performance gap happens. BTW, according to its paper, SPARF can achieve the performance of PSNR 21.01, SSIM 0.87, LPIPS 0.10 on NeRF backbone in the 3-view DTU setting, which can already fully validate the effect of their strategies. Considering the working principle of this work is extremely close to SPARF without significant innovation, it can be considered as an incremental work to transfer SPARF to 3DGS backbone. 2. For semantic loss: According to Table 4 vi and ix, the improvement is very marginal. What's worse, when applying an unsatisfactory type of feature, the performance will drop significantly. On the other hand, Despite that the authors try to prove the effect through "The comparisons between Table 4i and 4v show a 3.2% increase in PSNR and a 4.5% reduction in LPIPS", the analysis is partial. It only compares the situations with raw 3DGS. According to the comparisons between Table 4 vi, vii, viii and ix, its effect significantly shrinks to none after other constraints are applied, showing its redundancy. 3. For the so-called "SFM initialization", it shows that the authors lack the basic knowledge in multi-view geometry and the ability of discernment. Following the description and code process of FSGS, the author declared they use "the SfM process ... to generate camera poses and 3D points, using only training views without any post-processing". However, first, there is a significant mistake in FSGS that it calls the process they used as SfM but actually is an MVS method. Second, this process does not estimate any camera poses. So, the initialization method the authors used is actually MVS, rather than SfM they announced in the paper and rebuttal. This problem shows that the authors do not even know what they are actually doing. Considering this problem is to be revealed after the reviewer's twice asks, I'm worrying about this work's quality. 4. For introducing pre-trained models: the authors did not directly reply to my questions. First, are RoMa and VGG pre-trained models? Then, can these pre-trained models escape from the problems that "make training time longer and highly dependent on the quality of the pre-trained models"? In other words, is there not any extra time cost in order to make the proposed method as fast as raw 3DGS, especially when needing to use VGG to generate the semantic embedding for novel view k in eq (11) online? Or can it also work well when using models with poor pre-training quality? If not, how can the authors claim that their method is different from previous works in this part? --- Reply to Comment 2.1.1: Title: Reply to Reviewer oLUf Comment: We sincerely appreciate your time and efforts in reviewing our responses. We hope that our responses below adequately address your concerns. *Q1: Clarification of our novelty* *A1:* We strongly disagree with the reviewer's statement about novelty. This assertion is not only inaccurate but also a misrepresentation of our previous responses. Our method aims to address the overfitting and maintain the multi-view consistency, as detailed in Section 3.2. While SPARF also aims to mitigate overfitting, a challenge common to sparse view synthesis, the similarity with our method ends there. Our method utilizes the correspondence priors to guide multi-view projection, filtering out outliers and generating reliable pseudo-labels, which have been verified to be crucial. Subsequently, we compute the appearance/geometry/semantic losses with gradient weighting for the novel view. This not only broadens the available view information but also maintains multi-view consistency. Notably, during loss computation, we again utilize correspondence priors to select the minimal loss among each matched pair, which can reduce the effect of large noises. Table 4ix and 4x prove the effectiveness of the minimal operation. Our method diverges significantly from SPARF in both approach and outcomes, delivering superior performance on 3DGS. The depth consistency loss (DCons) employed in SPARF is fundamentally limited by its reliance on single-view projections, where depth inaccuracies lead to flawed projections. The DCons in SPARF's original paper improves PSNR by 0.20 (from 20.81 to 21.01) on NeRF—a result that aligns with our own reproduction on 3DGS. This proves that the single-view projection on SPARF is insufficient to address the overfitting issue on both NeRF and 3DGS. In contrast, our depth supervision for novel views in Table 4iii improves the PNSR by 3.13. *Q2: Analysis for semantic loss* *A2:* As explained in Lines 271-277 and Lines 441-443, DINOv2 and CLIP down-sample features early in the encoder with a large stride, e.g. 16. It is widely recognized that features extracted with a large stride often result in a loss of detail, which is the reason why the semantic loss with DINOv2 and CLIP obtains poorer performance. Additionally, Figure 3 further illustrates that using DINOv2 and CLIP as feature extractors generates worse results in some boundary and detailed regions. In contrast, our semantic loss using VGG can reduce artifacts and maintain details. *Q3: Details of SFM initialization* *A3:* We have already explained in our previous response that we initialize the 3D Gaussians with dense reconstruction in COLMAP. This process involves two parts, SfM followed by MVS, and does estimate camera extrinsics during SfM, therefore, the statement that "this process does not estimate any camera poses" is simply incorrect. Training and test views are both served as input for SfM to obtain their poses, then only the training views are used during MVS to get the fused point clouds. The technically sound initialization method name for the shorthand "SfM initialization" would be "Initialization based on the poses estimated by SfM for all views, and the fused point clouds from MVS on training views only", which will need to be shortened when referring to. Thus the initialization method we used is actually SfM+MVS, rather than just MVS as the reviewer suggested. The integrity of our initialization is by no means affected by the name, we ensure the dense point clouds are only from the training views, making it fair and comparable with the baselines. *Q4: Comparison of training time and results* *A4:* Our method, which incorporates RoMa and VGG, achieves significantly reduced training time while delivering superior performance in sparse view synthesis compared to networks relying on pre-trained depth/diffusion models, such as FSGS and SparseGS. As noted in our previous responses, only the VGG model introduces additional training time. However, as detailed in Table S7 of the uploaded file, our method requires just 5.82 minutes of training—substantially less than the 51.78 minutes needed by SparseGS (which uses a pre-trained diffusion model) and the 10.90 minutes required by FSGS (which relies on a pre-trained depth network). Furthermore, our approach not only enhances efficiency but also surpasses SparseGS and FSGS in performance, as demonstrated in Table 1. While incorporating VGG does involve some additional training time compared to raw 3DGS, the efficiency and speed trade-off is outweighed by the significant improvements in performance. Our method offers considerable gains over both raw 3DGS and methods reliant on pre-trained depth or diffusion models. For example, it improves PSNR by 4.09 dB on DTU and 3.32 dB on Blender compared to 3DGS. In summary, the integration of RoMa and VGG enhances the overall speed and effectiveness of our method, making it a notable advancement over previous approaches.	Summary: This paper introduces a novel few-shot Gaussian Splatting method for synthesizing novel views. Unlike conventional approaches that rely on pre-trained monocular depth estimation or diffusion methods, the proposed method leverages the matches of available training views to generate novel sample views between the training frames. It employs color, depth, and image feature losses. Additionally, a novel regularization loss is introduced to preserve the local structure of the object. Experimental results demonstrate that the proposed method achieves significant performance improvements in both real and synthetic datasets. Strengths: The strengths of the paper are as follows: 1. The proposed few-shot Gaussian splatting method for novel view synthesis, which does not rely on pre-trained depth estimation or diffusion models while achieving state-of-the-art performance. It is important to note that pre-trained depth estimation and diffusion models often require large parameters, which can lead to longer training times. 2. A multi-stage training scheme consisting of pre-training, intermediate, and tuning stages. This scheme optimizes the scene representation gradually. The pre-training process aims to obtain a basic representation of the scene using known views. The intermediate process transfers knowledge from known views to novel views while preserving consistency in overlapping regions between known training views. Finally, the tuning process removes artifacts that occur in few-shot scenarios. 3. The consistency loss function, which maintains the similarity of color, depth, and image features between pixels in novel views projected from known training views. This ensures that novel views have similar semantic, color, and depth information. It is worth noting that the loss functions are adaptively weighted to minimize the impact of errors in texture-rich regions. 4. The locality loss function, which maintains color similarity between the Gaussian and its neighboring regions. Accurate rendering results are more likely to occur when color values are smooth between neighborhoods. 5. The paper provides an exhaustive ablation study for each network design decision, leading to a convincing algorithm design. Weaknesses: The weaknesses of the paper are as follows: 1. The authors claim that relying on depth estimation or diffusion priors requires longer training times, but there is no comparison or ablation study provided to justify this claim. It would be beneficial to perform a comparison between state-of-the-art methods and the proposed method in terms of training and inference time. 2. The paper does not provide a justification for why the proposed method, which relies on matches between training views, performs better than methods that rely on pre-trained depth estimation or diffusion methods. A more in-depth analysis is needed to demonstrate how the proposed method outperforms these approaches. Technical Quality: 3 Clarity: 4 Questions for Authors: Based on the weaknesses, the following questions arise: 1. How do the training and inference times of the proposed method compare to those of state-of-the-art methods? 2. How well does the performance of the proposed method generalize to unseen data? Given that state-of-the-art methods typically employ large pre-trained models for data reconstruction, the proposed method, which relies solely on matches between known views, may encounter difficulties when dealing with unseen regions. Confidence: 4 Soundness: 3 Presentation: 4 Contribution: 3 Limitations: The proposed method may face challenges when dealing with texture-rich regions that are not visible from the input views. Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We thank the reviewer for the insightful and thorough review. We will integrate the additional results and analysis in the next version. In the following, we address the main concerns raised in this review. *Q: How do the training and inference times of the proposed method compare to those of state-of-the-art methods?* *A:* Thank you for your valuable feedback on the efficiency evaluation. Table S7 of the uploaded file compares the training time of 3DGS-based methods. Our method demonstrates comparable training time and is faster than SparseGS [41] (which uses a pre-trained diffusion model) and FSGS [49] (which relies on a pre-trained depth network). For inference time, since all methods utilize the same rendering process as 3DGS, their inference time is similar across these methods, at around 300 FPS. *Q: The paper does not provide a justification for why the proposed method, which relies on matches between training views, performs better than methods that rely on pre-trained depth estimation or diffusion methods. A more in-depth analysis is needed to demonstrate how the proposed method outperforms these approaches.* *A:* We appreciate the reviewer for the valuable suggestions and will incorporate the following analysis in the updated version. Our method offers several advantages over existing approaches, contributing to its improved performance: (1) Address overfitting: Methods that incorporate pre-trained depth estimation networks, such as DNGaussian [16] and DRGS [5], add depth supervision only on known view, which does not solve the issue of overfitting to sparse training view. In contrast, our method designs extra appearance/geometry/semantics constraints on novel views, which can address overfitting and enhance performance. (2) Provide accurate supervision: The methods, such as FSGS [49] and SparseGS [41], rely on a pre-trained depth network or diffusion model to generate pseudo labels for supervising the novel view. This may generate inaccurate pseudo labels, causing noises and encountering scale-ambiguity issues with the generated depth. Besides, the diffusion model used in SparseGS increases the training time significantly (see Table S7 of the uploaded file). Contrastively, with feature matching, our method projects the pixels from two training views to the novel view to provide supervision. This can leverage the ground truth of color from the training view, which is more accurate than pseudo labels generated by other networks, and ensure consistency across multiple views, thereby enhancing overall performance. *Q: How well does the performance of the proposed method generalize to unseen data?* *A:* In the orange boxes of Figure S1 (see the uploaded file), the visualization of unseen regions demonstrates that our method yields fewer artifacts and superior performance than FSGS [49] (which uses a pre-trained depth network). This can be attributed to that our multi-view consistency constraints can provide accurate supervision for the seen regions, and the proposed locality preserving regularization helps maintain local smoothness and reduce artifacts.	Summary: This paper proposes a new method for sparse-view novel view synthesis. It proposes a multi-stage training scheme including pre-training, intermediate, and tuning stages. It introduces pre-trained dense matching models to find pixel correspondences between different-view images and encourage consistency. A Locality Preserving Regularization is proposed to encourage local smoothness. Strengths: 1. The proposed method achieves SOTA performances on the 3-view LLFF dataset, 3-view DTU dataset and 8-view Blender dataset. 2. The proposed method suits both random points and mvs points as shown in Table 1. 3. The paper is well-organized. Weaknesses: 1. The comparisons are not enough. It lacks comparisons on more input views, such as the 6-view and 9-view settings used in FreeNeRF, which is also important to evaluate the proposed method in sparse-view settings. 2. The ablation studies are not enough. There are lots of hyperparameters listed in the implementation details, however, it is unclear how these hyperparameters are selected and how they impact the performance, which is important to evaluate the robustness of the proposed method. 3. The proposed method introduces ore-trained dense matching networks, thus I think it is still similar to those introduced pre-trained depth estimation networks. Technical Quality: 3 Clarity: 3 Questions for Authors: 1. I wonder is the proposed method still work with more input views such as 6 views and 9 views. Confidence: 5 Soundness: 3 Presentation: 3 Contribution: 2 Limitations: The authors have discussed the limitations. Flag For Ethics Review: ['No ethics review needed.'] Rating: 5 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We thank the reviewer for the insightful and thorough review. We will integrate the additional results and analysis in the revised version. In the following, we address the main concerns raised in this review. *Q: It lacks comparisons on more input views, such as the 6-view and 9-view settings used in FreeNeRF.* *A:* We thank the reviewer for the constructive suggestions on experiments. The results for 6-view and 9-view settings are given in Table S1 of the uploaded file. Our method shows improved performance over the baseline 3DGS with both SFM and random initialization. Notably, with SFM initialization, our method surpasses FreeNeRF on both datasets. We will include these results in the revised version. *Q: It is unclear how these hyperparameters are selected and how they impact the performance.* *A:* We appreciate the reviewer's suggestions on the ablation study. We have given evaluations for hyperparameters $\alpha, \beta, \gamma, \eta$, and the multi-stage settings in Table S2 - Table S6 of the uploaded file on DTU with 3 training views. (1) $\alpha$ in Eq 12: we set it to a low weight due to the lack of ground truth for the depth. It can be seen that, in Table S3, larger $\alpha$ (e.g. 0.2) gets poorer results, primarily due to noise in predicted depth values. (2) $\beta$ in Eq 12: due to the importance of color supervision for the novel view synthesis, a large weight is utilized in the experiment. In Table S2, a value of 0.5 achieves the best performance. (3) $\gamma$ in Eq 12: The results in Table S5 demonstrate that our method is robust to different $\gamma$ values, with $\gamma=0.001$ being optimal. (4) $\eta$ in Eq 15: The pre-training loss in the intermediate stage is employed to ensure sufficient supervision for novel views. A lower $\eta$ (0.01) does not solve this very well, while higher $\eta$ (0.1 and 0.5) fails to prevent overfitting to the training views. (5) Multi-stage settings: we evaluate the influence of iterations for each stage in our 3-stage training in Table S6. The last two rows show that a fewer/more iteration in the first stage leads to worse results, due to underfitting/overfitting to training views. The 1st-4th rows demonstrate that the fine-tuning stage enhances performance and achieves optimal results with 500 iterations. We will integrate these results and analysis to our revised version. *Q: The proposed method introduces pre-trained dense matching networks, thus I think it is still similar to those introduced pre-trained depth estimation networks.* *A:* Our method offers several advantages over existing approaches with a pre-trained depth estimation network, contributing to its improved performance: (1) Address overfitting: Methods that incorporate pre-trained depth estimation networks, such as DNGaussian [16] and DRGS [5], include depth supervision only on known view, which does not solve the issue of overfitting to sparse training view. In contrast, our method designs extra appearance/geometry/semantics constraints on novel views, which can avoid overfitting and enhance performance. (2) Provide accurate supervision: The methods, such as FSGS [49], rely on a pre-trained depth network to generate pseudo labels for supervising the novel view. This may generate inaccurate pseudo labels, encountering scale-ambiguity issues with the generated depth. Contrastively, with feature matching, our method projects the pixels from two training views to the novel view to provide supervision. This can leverage the ground truth of color from the training view, which is more accurate than pseudo labels generated by other networks, and ensure consistency across multiple views, thereby enhancing overall performance.	Summary: This paper tackles the problem of few view (or sparse view) 3DGS multistage training with correspondence-driven losses that enforce projected colors, depths, and semantic features (extracted by a pre-trained VGG) are consistent. Contributions are straightforward and geometrically inspired. In addition, the authors propose a locality preservation loss, which enforces color smoothness in neighbouring gaussians. Strengths: 1. Simple yet effective geometrically inspired regularization for few view 3DGS. 2. Strong qualitative and quantitative results. 3. Extensive ablation studies. Weaknesses: 1. Intuitions behind a 3-stage training strategy are not well established. If intermediate training already includes L_pre_training, why is further fine-tuning with L_pre_training only needed? Technical Quality: 3 Clarity: 3 Questions for Authors: See weaknesses Confidence: 4 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: Ok Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We appreciate the reviewer’s insightful and thorough review. We will further integrate the explanation in the revised version. Below, we address the main concerns raised in this review. *Q: Intuitions behind a 3-stage training strategy are not well established. If intermediate training already includes L_pre_training, why is further fine-tuning with L_pre_training only needed?* *A:* The design of 3 stages aims to address overfitting to sparse training views and improve supervision reliability in novel views. Specifically, (1) the pre-training stage aims to obtain a basic representation of the 3D scene. (2) The intermediate stage mainly focuses on integrating additional view information into the network. We found that the sparse supervision and noises of pseudo labels in the novel views cause model collapse in some scenes. Thus, $L_{pre-training}$ is applied in this stage to solve this issue, employing a lower weight to prevent overfitting to the training view. (3) The fine-tuning stage is designed to refine the network with ground truth, add supervision for unmatched pixels during the intermediate stage, and minimize the impact of noise in the novel view. The results in the 1st-4th row of Table S6 (see the uploaded file) also indicate that incorporating this extra fine-tuning stage improves performance. We will include these explanations to our revised version. --- Rebuttal Comment 1.1: Title: Thanks for your reply Comment: Thanks for supplementing your results. It is now clear the 3rd stage provides additional performance improvements. Is is possible to show those "unmatched pixels during the intermediate stage" to support your claims? I am also wondering why you did not include an experiment with 1st: 2000 2nd: 7500 Third: 0, for a more direct comparison. --- Rebuttal 2: Title: Reply to Reviewer Eqzk Comment: We sincerely appreciate your precious time and efforts in reviewing our paper and responses. We hope that the response provided below adequately addresses your concerns. *Q1: Is it possible to show those "unmatched pixels during the intermediate stage" to support your claims?* *A1:* We appreciate the review's suggestion on visualizations to better support our claim. We provided visualizations for the matched results in Figure 7. The white regions in Figure 7c and the areas without red lines in Figure 7d represent unmatched pixels, which lack effective supervision during the intermediate stage. We will include detailed descriptions in our revised version. *Q2: I am also wondering why you did not include an experiment with 1st: 2000 2nd: 7500 Third: 0, for a more direct comparison.* *A2:* We thank the reviewer for the suggestion on the experiment. The setting with '1st: 2000 2nd: 7500 Third: 0' achieves a performance of PSNR 18.37, SSIM 0.782, and LPIPS 0.198. We will integrate these results into our revised version.	Rebuttal 1: Rebuttal: We appreciate the constructive feedback and positive comments from the reviewers. We are encouraged that the reviewers found that: - Our paper is well-organized (Reviewer 8CAh). - Our method is effective and convincing (Reviewer Eqzk, Reviewer ZPEG). - Our method achieves SOTA performance, with strong qualitative and quantitative results, as well as extensive ablation studies (Reviewer Eqzk, Reviewer 8CAh, Reviewer ZPEG, Reviewer oLUf). * We have uploaded a PDF file that includes two figures and several tables to address the reviewers' concerns: - Figure S1: Visualizations of unseen regions generated by our method and FSGS. - Figure S2: Qualitative depth comparison between the baseline and our method. - Table S1: Results for 6-view and 9-view settings. - Table S2 - Table S6: Evaluation of the influence of hyperparameters. - Table S7: Comparison of the training time for 3DGS-based methods in few-shot novel view synthesis. * We address the reviewers' questions below, and will incorporate the suggested results and description into our revised version. We will actively participate in the Author-Reviewer discussion session. Please feel free to tell us if anything remains unclear. *** We sincerely appreciate PCs, ACs, and reviewers for their time and effort in evaluating our submission. Pdf: /pdf/6a5dc9b1a2c410287f455a7fa30693a335fb6f75.pdf	NeurIPS_2024_submissions_huggingface	2,024	null	null	null	null	null	null	null	null
Maximum Entropy Reinforcement Learning via Energy-Based Normalizing Flow	Accept (poster)	Summary: This paper proposes a new framework for Maximum Entropy Reinforcement Learning using Energy-Based Normalizing Flows (EBFlow). The authors argue that this framework offers three main advantages: Firstly, the soft value function can be obtained without estimation. Second, the framework combines policy evaluation and policy improvement in one iteration. Finally, it does not use Monte Carlo methods and interacts with the environment effectively. To achieve this goal, Learnable Reward Shifting (LRS) technique which involves reward shifting for the value function using a reward shifting network and Shifting-Based Clipped Double Q-Learning (SCDQ) technique that employs two reward shifting networks for the value function and utilizes the minimum value function similar to double Q-learning have been proposed. Strengths: 1. The paper is well-structured, and the overall content is easy to follow and intuitive. The detailed background explanations allow readers to progress through the main text with a solid understanding of the prior work. 2. The experimental domains are diverse, clearly demonstrating the strengths of the proposed algorithm in an easy-to-understand manner. The experiments are well-structured, and the detailed explanations make it easier for readers to follow the setup. 3. The proposed algorithm does not use Monte Carlo-based methods, allowing it to select the proper action for training during each step without waiting for the end of an episode. This enables the policy to interact with the environment more effectively 4. Existing SAC-based algorithms often rely on value function estimation, which can introduce errors. The proposed algorithm can obtain the soft value function without estimation, thus avoiding such errors. Weaknesses: 1. Despite the fact that the proposed method is the first to apply EBFlow to reinforcement learning (RL), the novelty of the algorithm may appear weak, as there are several existing RL methods based on flow-based models. Besides applying EBFlow to the RL structure, were there are any additional considerations or advantages that were taken into account to enhance its performance or applicability in the context of RL? If such points exist, highlighting them would help to emphasize the contributions of the authors' proposed technique. 2. Introducing new frameworks generally necessitates providing sufficient mathematical proof to validate their efficacy, even for basic techniques or simple conjectures. It would be beneficial if the paper included mathematical proofs demonstrating that the proposed framework converges to the optimal policy. For example, the techniques presented in the paper, namely Learnable Reward Shifting (LRS) and Shifting-Based Clipped Double Q-Learning (SCDQ), have a significant impact on performance. However, the paper does not provide adequate mathematical justification for these techniques. Including such proofs would greatly enhance the confidence in the proposed method. 3. It would be beneficial for the paper to compare its methods with more recent state-of-the-art algorithms. The comparison with SAC, published in 2018, feels a bit outdated. Given that reinforcement learning is a rapidly evolving field, including comparisons with more recent methods would provide a clearer understanding of where this framework stands in the current landscape. Such comparisons would greatly enhance the paper by placing the proposed method in the context of the latest advancements. Technical Quality: 3 Clarity: 3 Questions for Authors: 1. Please answer the questions in the weakness section 2. According to the appendix, the proposed method is said to be 2.3 times slower than SAC. Given this difference in speed, a time-wise comparison of the algorithm's performance would be very helpful to understand its efficiency relative to SAC. 3. In Figure 3, SAC and MEow show similar performance in the Mujoco domain, but there is a noticeable difference in performance in the Isaac Gym domain in Figure 4. Could you provide a further analysis on the reasons for the occurrence of this difference? It would provide valuable insight into the strengths and limitations of the proposed method. Minor Issues\ Typo in line 149: "Jocobian" should be "Jacobian."\ Misreference in section 4.1: It should refer to Fig. 2, not Fig. 1. Confidence: 3 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: They have handled it well. Flag For Ethics Review: ['No ethics review needed.'] Rating: 5 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We thank the reviewer’s time and effort spent on the review, and would like to respond to the reviewer’s questions as follows. --- Comments C1. Despite the fact that the proposed method is the first to apply EBFlow to reinforcement learning (RL), the novelty of the algorithm may appear weak, as there are several existing RL methods based on flow-based models. ... If such points exist, highlighting them would help to emphasize the contributions of the authors' proposed technique. Response: As described in Lines 41-53 of Section 1 and Lines 166-174 of Section 3 of our paper, the proposed framework possesses the following two unique features. To the best of our knowledge, none of the existing MaxEnt RL frameworks, including those with flow-based models (e.g., [1-3]), possess the following characteristics: - Our framework enables the exact calculation of the soft value function, which improves the accuracy of the optimization process. - Our framework integrates the policy evaluation step and policy improvement step into a single training step, which streamlines the overall training process. Previous actor-critic frameworks without these features may suffer from inaccurate soft value estimation (i.e., Eqs. (6) and (7)) and estimation error when minimizing the policy improvement loss (i.e., Eq. (5)). We substantiate our claim through an extensive set of experiments conducted on the MuJoCo and Omniverse Isaac Gym benchmarks as depicted in Fig. 3 and 4 of the main manuscript. [1] Haarnoja et al. Latent Space Policies for Hierarchical Reinforcement Learning, ICML 2018.\ [2] Mazoure et al. Leveraging exploration in off-policy algorithms via normalizing flows, CoRL 2019.\ [3] Ward et al. Improving Exploration in Soft-Actor-Critic with Normalizing Flows Policies, ICML Workshop 2019. --- C2. Introducing new frameworks generally necessitates providing sufficient mathematical proof to validate their efficacy, even for basic techniques or simple conjectures. ... However, the paper does not provide adequate mathematical justification for these techniques. Including such proofs would greatly enhance the confidence in the proposed method. Response: We respectfully remind the reviewer that our method is mathematically supported by multiple theories. Our main methodology is theoretically supported by Proposition 3.1 in Lines 181-183, which verifies the validity of our selections of $Q_\theta$ and $V_\theta$. According to Theorem 3 in [1], which is described in Section 2.1, the policy of MEow converges to the optimal policy when the soft Bellman error in Eq. (4) is minimized. The theoretical justification for learnable reward shifting (LRS) is provided in Lines 212 to 219. Lines 214 and 215 show that the soft value function can be calculated exactly (Proposition A.3). In Lines 217 and 218, we mathematically demonstrate that the action distribution remains unchanged after applying LRS. The newly defined $Q_\theta^b$ and $V_\theta^b$ follow Theorem 3 in [1] and converge to the optimal policy when minimizing the soft Bellman error in Eq. (4). The theoretical justification for SCDQ is given in Lines 226 to 233. SCDQ allows for the implementation of clipped double Q-learning [2] without duplicating the policy in MEow. In Eq. (13) and Lines 228-230, we mathematically show how SCDQ prevents the introduction of additional soft value functions. The mechanisms by which clipped double Q-learning reduces training variances and overestimation are analyzed in TD3’s original paper [2]. We would also like to emphasize that other techniques and claims in this work are theoretically supported with proofs. The deterministic inference technique is supported by Proposition A.4 in Lines 605-609. The theoretical analysis of the soft value estimation methods is offered in Proposition A.1 and Remark A.2 in Lines 565-568. These technical results support the validity of our method and provide motivation for formulating our approach. [1] Haarnoja et al. Reinforcement Learning with Deep Energy-Based Policies. ICML 2017.\ [2] Fujimoto et al. Addressing Function Approximation Error in Actor-Critic Methods. ICML 2018. --- C3. It would be beneficial for the paper to compare its methods with more recent state-of-the-art algorithms. ... Such comparisons would greatly enhance the paper by placing the proposed method in the context of the latest advancements. Response: A performance comparison between MEow and two latest online RL frameworks, DIPO [1] published in 2023 and S2AC [2] published in 2024, is presented in Table 1 in the [global comment](https://openreview.net/forum?id=lhlIUxD5eE&noteId=Vapk6NMOyp). The results show that MEow outperforms DIPO and S2AC with noticeable margins in five MuJoCo environments. [1] Yang et al. Policy representation via diffusion probability model for reinforcement learning, 2023. \ [2] Messaoud et al. S2AC: Energy-Based Reinforcement Learning with Stein Soft Actor-Critic, ICLR 2024. --- Questions Q1. According to the appendix, the proposed method is said to be 2.3 times slower than SAC. Given this difference in speed, a time-wise comparison of the algorithm's performance would be very helpful to understand its efficiency relative to SAC. Response: In response to the reviewer's request, we extended the training time of SAC and provided a timewise comparison between MEow and SAC evaluated on the Hopper-v4 environment in Fig. 3 of the attached PDF file in the [global comment](https://openreview.net/forum?id=lhlIUxD5eE&noteId=Vapk6NMOyp). The results show that MEow converges to a policy that achieves a better average return and exhibits greater stability. --- The response to Q2 is provided in the [global comment](https://openreview.net/forum?id=lhlIUxD5eE&noteId=Vapk6NMOyp). --- Rebuttal Comment 1.1: Comment: Thank you for providing the additional experiments and for addressing the questions with such a detailed rebuttal. Your response, along with the referenced paper, has more clarified the aspects I inquired about. I appreciate your efforts in improving the clarity of the work, and I will be increase my score accordingly. --- Reply to Comment 1.1.1: Comment: We sincerely appreciate the reviewer’s response and valuable feedback. Thank you again for your thoughtful review.	Summary: This paper proposes to use the energy-based normalizing flow (EBFlow) to represent the policy and value functions in maximum entropy (MaxEnt) RL. Specifically, the paper builds the connection between EBFlow and MaxEnt RL by linking the conditional unnormalized density that depends on the input to the action value function Q and the constant that is independent of the input to the value function V. As a consequence, the Boltzmann policy defined by Q and V can be expressed exactly using a mapping in EBFlow. By updating the current Q estimate to approximate the action value of the Boltzmann policy, the algorithm would spontaneously update the Boltzmann policy. Hypothetically, the algorithm should converge to the optimal policy under entropy regularization. Using an existing implementation of EBFlow in the relevant literature, the paper proposed MEow based on this formulation. Experiments on four MuJoCo environments and six Issac Gym environments demonstrate the learning performance of MEow. Strengths: The strengths of this paper are as follows: 1. The connection between EBFlow and MaxEnt RL revealed in this paper is very interesting and, to my knowledge, novel. Given that MaxEnt is a popular approach in the RL literature, the issue addressed in this paper, learning the action value function and the Boltzmann policy, as well as the proposed approach, will be interesting to a broad set of audiences in the RL community. 2. The paper includes abundant comparative and ablation experiments. Notably, the ablation study covers the effect of the two components in the practical implementation, the difference between the deterministic and stochastic evaluation, etc. 3. The paper is well-written and mostly clear. Weaknesses: 1. There is a potential concern about the empirical investigation: The proposed method is tuned over the target smoothing parameter, while the baselines are not tuned. Why is the target smoothing parameter tuned per environment? What would happen if the proposed method also used the same, fixed smoothing parameter? 2. This is a minor point. The organization of the paper can be improved. Specifically, the background section is too long, which takes up the space for more interesting discussions presented in the appendix. If space allows, the main text should include the network architecture in Figure A5, as it may be unfamiliar to a lot of readers. In addition, the limitation, especially the computational aspect of the proposed approach, should be discussed in the main paper. Other suggestions that do not affect the evaluation: * It would be great if there is at least some qualitative comparison between the proposed method and baselines using MaxEnt RL in the multi-goal environment. * Typos between Lines 254 and 276: The reference to Figure 1 should be to Figure 2. Technical Quality: 3 Clarity: 2 Questions for Authors: Some questions that may affect the evaluation: 1. Are deterministic policies used for other algorithms in the evaluation in Figures 3 and 4? Confidence: 3 Soundness: 3 Presentation: 2 Contribution: 3 Limitations: The limitations of the work are discussed in the appendix, which should be moved to the main text. Another unmentioned potential limitation is the sensitivity to the target network smoothing parameter. Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We appreciate the reviewer’s valuable feedback and effort spent on the review and would like to respond to the reviewer’s questions as follows. --- Comments C1. (1) There is a potential concern about the empirical investigation: The proposed method is tuned over the target smoothing parameter, while the baselines are not tuned. (2) Why is the target smoothing parameter tuned per environment? What would happen if the proposed method also used the same, fixed smoothing parameter? Response: (1) Our baselines use the refined hyperparameters of Stable Baseline 3 (SB3) as mentioned in Line 285. Further modifications to the hyperparameters may not result in performance improvement. For a fair evaluation, we provide the results of SAC (the best-performing baseline) tuned with different target smoothing factors used in MEow (i.e., $\tau=0.005$, $0.003$, $0.0005$, and $0.0001$) in Fig. 1 of the attached PDF file in the [global comment](https://openreview.net/forum?id=lhlIUxD5eE&noteId=Vapk6NMOyp). The results show that SB3’s original hyperparameter ($\tau=0.005$) is the most effective. (2) We appreciate the reviewer's question regarding the target smoothing parameter $\tau$. According to Fig. 1 in the [global comment](https://openreview.net/forum?id=lhlIUxD5eE&noteId=Vapk6NMOyp), our current approach requires different values of $\tau$ to achieve good performance. However, we acknowledge the importance of determining a more generalized parameter setting. As part of our future work, we plan to expand our experimental framework to explore a broader range of $\tau$ values to identify a generally applicable $\tau$ across diverse environments. --- C2. This is a minor point. The organization of the paper can be improved. Specifically, the background section is too long, which takes up the space for more interesting discussions presented in the appendix. If space allows, the main text should include the network architecture in Figure A5, as it may be unfamiliar to a lot of readers. In addition, the limitation, especially the computational aspect of the proposed approach, should be discussed in the main paper. Response: We appreciate the reviewer's suggestion and concur that a more concise Section 2, coupled with the inclusion of discussions on MEow's model architecture and limitations in the main manuscript, will enhance the paper's quality. These modifications will be incorporated in the final revision. --- C3. Other suggestions. (1) It would be great if there is at least some qualitative comparison between the proposed method and baselines using MaxEnt RL in the multi-goal environment. (2) Typos between Lines 254 and 276: The reference to Figure 1 should be to Figure 2. Response: (1) We appreciate the reviewer's suggestion. A qualitative comparison among MEow, SAC, and SQL in the multi-goal environment is presented in Fig. 2 of the attached PDF file in the [global comment](https://openreview.net/forum?id=lhlIUxD5eE&noteId=Vapk6NMOyp). (2) We appreciate the reviewer’s attention to detail. The typo and figure reference will be corrected in the final revision. --- C4. (1) The limitations of the work are discussed in the appendix, which should be moved to the main text. (2) Another unmentioned potential limitation is the sensitivity to the target network smoothing parameter. Response: (1) We express our gratitude to the reviewer for the suggestions. A discussion of this work's limitations will be incorporated into the main manuscript in the final revision. (2) An examination of the sensitivity to the target network smoothing parameter is presented in Fig. 1 of the attachment to the [global comment](https://openreview.net/forum?id=lhlIUxD5eE&noteId=Vapk6NMOyp). This analysis will be included in the paper's final revision. --- Questions Q1. Are deterministic policies used for other algorithms in the evaluation in Figures 3 and 4? Response: During the evaluation phase, MEow, SAC, TD3, DDPG, and PPO employ deterministic policies. SQL, however, does not support deterministic inference and thus utilizes a stochastic policy for inference. --- Rebuttal Comment 1.1: Title: Follow-up questions on hyperparameter tuning Comment: Thank you for the rebuttal and additional experiment results on SAC’s sensitivity on the target smoothing factor, $\tau$. I think the difference in the sensitivity to $\tau$ of SAC and MEow is very interesting and worth further investigation. In addition, I wonder what the hyperparameter tuning process is in the Issac gym environments. Are both SAC and MEow tuned (for both $\tau$ and $\alpha$)? If yes, how? --- Reply to Comment 1.1.1: Comment: We sincerely appreciate the reviewer’s response and valuable feedback. In the Isaac Gym experiments, both SAC and MEow were tuned using the same search space for $\tau$ and $\alpha$ to ensure a fair comparison. Specifically, a grid search was conducted with $\tau$ values ranging from 0.1 to 0.00025 and $\alpha$ values from 0.8 to 0.0005 for both algorithms. The setups with the highest average return were selected for each environment. We agree with the reviewer that the differences in the hyperparameters present an interesting direction for further investigation. Thank you again for your thoughtful review.	Summary: The paper presents a new framework for Maximum-Entropy (MaxEnt) Reinforcement Learning (RL) using Energy-Based Normalizing Flows (EBFlow). Traditional MaxEnt RL methods, particularly for continuous action spaces, typically utilize actor-critic frameworks and alternate between policy evaluation and policy improvement steps. The proposed EBFlow framework integrates these steps into a single objective, eliminating the need for Monte Carlo approximation while enabling efficient modeling of multi-modal action distributions. The framework was experimentally validated on the MuJoCo benchmark and high-dimensional robotic tasks in Omniverse Isaac Gym, demonstrating superior performance compared to existing baseline methods. Strengths: - By combining policy evaluation and policy improvement into a single training process, the proposed method simplifies the optimization process and avoids potential optimization errors inherent in alternating updates. - The use of normalizing flows facilitates efficient sampling and exact calculation of probability density functions, which addresses the inefficiencies associated with Monte Carlo methods in traditional MaxEnt RL approaches. - Experimental results indicate that the EBFlow framework outperforms widely-adopted baselines on both the MuJoCo benchmark and complex robotic tasks, showcasing its effectiveness and robustness. Weaknesses: - There are some unreasonable aspects in the organization of the content. For example, Section 3.2 appears somewhat abrupt; it is suggested to move the content of A.4 into the main text. The discussion about inference in Section A.3 is also very important and should be included in the main text. - The experimental section does not sufficiently highlight the advantages of the MEow method. The experiments in Section 4.1 on toy examples are not enough. Can comparisons be made from the perspectives of inference time, training time, representation capability, etc.? - While the integration of policy evaluation and improvement steps simplifies the training process conceptually, the implementation details of the EBFlow framework may introduce additional complexity that needs to be managed. Technical Quality: 3 Clarity: 3 Questions for Authors: - GFlowNet points out that the MaxEnt RL paradigm has deficiencies in terms of diversity. Specifically, MaxEnt RL tends to achieve target states that are "further" from the initial state. Does MEow also inherit this defect of the MaxEnt RL paradigm? - Energy-based normalizing flow is a generative model, and its original loss is to maximize log-likelihood. In Algorithm 1, this loss does not seem to be present. In my view, MEow uses the network structure of energy-based normalizing flow, but the loss still follows the MaxEnt RL paradigm. In this case, could the title of this paper be potentially misleading? - Figure A5 shows that the MEow network uses a hypernetwork module. Why was this module introduced? Is this a common architecture for energy-based normalizing flow? - What are the learning dynamics observed during the training process? Are there any notable differences in the convergence behavior compared to traditional actor-critic methods? - How does the proposed framework scale with increasing complexity of the environment or the state-action space? Are there any specific limitations or challenges observed during experimentation? - How robust is the EBFlow framework to different hyperparameters? Confidence: 3 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: The author has already discussed the limitations ofthe method and its potential impacts in the paper, and provided possible solutions. Flag For Ethics Review: ['No ethics review needed.'] Rating: 7 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We appreciate the reviewer’s valuable feedback and effort spent on the review, and would like to respond to the reviewer’s questions as follows. --- Comments C1. There are some unreasonable aspects in the organization of the content. For example, Section 3.2 appears somewhat abrupt; it is suggested to move the content of A.4 into the main text. The discussion about inference in Section A.3 is also very important and should be included in the main text. Response: We appreciate the reviewer's suggestion and would be glad to incorporate the contents of Sections A.3 and A.4 into the main manuscript to improve the paper's quality. We will implement this arrangement in the final revision. --- C2. The experimental section does not sufficiently highlight the advantages of the MEow method. The experiments in Section 4.1 on toy examples are not enough. Can comparisons be made from the perspectives of inference time, training time, representation capability, etc.? Response: We appreciate the reviewer’s suggestion and have expanded the experiments in Section 4.1 to include a comparison of inference and training times in the Humanoid environment (MuJoCo benchmark), as shown in the following table. The results demonstrate that performing Monte Carlo estimation and MCMC sampling at each step imposes a computational burden during both the training and inference phases. \|\|Training time (sec.)\|Inference time (sec.)\| \|-\|-\|-\| \|MEow\|0.022 $\pm$ 0.000\|0.004 $\pm$ 0.000\| \|MC-based ($M=10$)\|0.110 $\pm$ 0.000\|0.040 $\pm$ 0.004\| \|MC-based ($M=100$)\|1.064 $\pm$ 0.001\|0.388 $\pm$ 0.004\| \|MC-based ($M=200$)\|3.181 $\pm$ 0.012\|0.760 $\pm$ 0.004\| Table. Training and inference time comparison. $M$ denotes the number of samples used for soft value estimation and the number of steps in the MCMC sampling process. The results are evaluated on NVIDIA V100 GPUs. Regarding the representation capability of our method, we have provided an analysis in Section A.5.4. The results indicate that our method possesses the ability to model multi-modal action distributions. --- C3. While the integration of policy evaluation and improvement steps simplifies the training process conceptually, the implementation details of the EBFlow framework may introduce additional complexity that needs to be managed. Response: The implementation of EBFlow can be simplified by integrating existing normalizing flow model libraries (e.g., [1,2]). Since log determinant calculation is already supported by many normalizing flow packages, the only required modification is the addition of an if-else condition during the forward pass to separate the energy function and the normalizing constant. We have made our code implementation available in an anonymous repository, with the corresponding link provided in Lines 699-700 of Section A.6. [1] Stimper et al. normflows: A PyTorch Package for Normalizing Flows, Journal of Open Source Software 2023.\ [2] Durkan et al. nflows: normalizing flows in PyTorch. 2020. --- Questions Q1. GFlowNet points out that the MaxEnt RL paradigm has deficiencies in terms of diversity. Specifically, MaxEnt RL tends to achieve target states that are "further" from the initial state. Does MEow also inherit this defect of the MaxEnt RL paradigm? Response: We would like to highlight that the primary focus of this paper is on investigating the parameterization and its implications on MaxEnt RL. Comparing different RL frameworks (e.g., GFlowNet versus MaxEnt RL) falls outside the scope of this work. We acknowledge that further investigation into the topic of diversity within this context is an interesting direction for future research, and welcome any additional references the reviewer may provide to enrich this potential line of inquiry. --- Q2. Energy-based normalizing flow is a generative model, and its original loss is to maximize log-likelihood. ... but the loss still follows the MaxEnt RL paradigm. In this case, could the title of this paper be potentially misleading? Response: We believe that our paper title accurately reflects the key concept of the proposed methodology. As discussed in EBFlow’s original paper [1], a normalizing flow can be represented as an energy-based model and optimized using energy-based training objectives. Specifically, the authors in [1] employ score matching (which minimizes Fisher divergence) as an alternative to maximum likelihood training (which minimizes KL divergence) to optimize EBFlow. MEow can be viewed as an extension of EBFlow to the MaxEnt RL domain, where a policy is also represented as an energy-based model [2]. By synthesizing these concepts, we parameterize the policy using an EBFlow and demonstrate that the policy can be optimized by minimizing the soft Bellman errors. [1] Chao et al. Training Energy-Based Normalizing Flow with Score-Matching Objectives. NeurIPS 2023.\ [2] Haarnoja et al. Reinforcement Learning with Deep Energy-Based Policies. ICML 2017. --- Q3. Figure A5 shows that the MEow network uses a hypernetwork module. Why was this module introduced? Is this a common architecture for energy-based normalizing flow? Response: Hypernetworks are commonly employed in conditional normalizing flow models for modeling the weights in the transformations (see [1] for more details). In our approach, the policy is modeled as a state-conditioned normalizing flow, with hypernetworks utilized to encode state information. As for implementation, we adopt the existing implementation of conditional normalizing flow [2], which incorporates a hypernetwork architecture. [1] Kobyzev et al. Normalizing Flows: An Introduction and Review of Current Methods. TPAMI 2019.\ [2] Stimper et al. normflows: A PyTorch Package for Normalizing Flows, Journal of Open Source Software 2023. --- The responses to Q4-Q6 are provided in the [global comment](https://openreview.net/forum?id=lhlIUxD5eE&noteId=Vapk6NMOyp). --- Rebuttal Comment 1.1: Comment: Many thanks to the authors for thoroughly supplementing the key experiments and discussions. I have raised the score to 7. --- Reply to Comment 1.1.1: Comment: We sincerely appreciate the reviewer’s response and valuable feedback. Thank you again for your thoughtful review.	Summary: The paper introduces a new MaxEnt RL framework called Meow based on Energy-based Normalizing Flow, which integrates the policy evaluation steps and the policy improvement steps and results in a single objective training process. Besides Meow enables the calculation of the soft value function used in the policy evaluation target without Monte Carlo approximation and supports the modeling of multi-modal action distributions. Strengths: 1. The proposed method is innovative and interesting, which utilizes the special property of Energy-based Normalizing Flow, and thus enables the consistency between the actor and critic and a single objective training process. 2. The paper is overall well-written. The background and related works are introduced appropriately. 3. The experimental results are comprehensive, involving the mujoco and Issac gym benchmarks. Weaknesses: 1. The calculation of the derterminant is usually time-consuming. So the reviewer wonders how the training time of Meow is compared with that of the classical off-policy RL methods like SAC. 2. In the experiments, Meow is just compared with several classical RL methods for continuous control and lacks the comparison with advanced RL methods [1, 2] with other generative models like diffusion model and consistency model, which also supports the modeling of multi-modal action distribution. [1] Yang L, Huang Z, Lei F, et al. Policy representation via diffusion probability model for reinforcement learning[J]. arXiv preprint arXiv:2305.13122, 2023. [2] Yue Y, Kang B, Ma X, et al. Boosting offline reinforcement learning via data rebalancing[J]. arXiv preprint arXiv:2210.09241, 2022. Technical Quality: 4 Clarity: 4 Questions for Authors: 1. How is the training time of Meow compared with other methods? 2. How is Meow compared with other advanced RL methods with the utilization of generative-model-based policy? 3. Could the authors explain why Meow can learn a good reward shifting rather than a trivial shifting, like for any $s_t$, $b(s_t)=0$? Confidence: 4 Soundness: 4 Presentation: 4 Contribution: 3 Limitations: The authors have clearly presented the limitations in the paper. Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We appreciate the reviewer’s valuable feedback and effort spent on the review and would like to respond to the reviewer’s questions as follows. --- Comments C1. (1) The calculation of the determinant is usually time-consuming. (2) So the reviewer wonders how the training time of MEow is compared with that of the classical off-policy RL methods like SAC. Response: (1) Several existing normalizing flow architectures (e.g., [1] used in this work) support efficient calculation of Jacobian determinants, as discussed in Section 2.2 (Lines 138-140). In general, the Jacobian determinant of these architectures can be calculated with time complexity $O(D^2L)$, where $D$ represents the dimension of actions and $L$ is the number of layers in $g_\theta$ (defined in Line 132). This complexity is the same as forward passing a normalizing flow model. (2) A discussion regarding the training time of MEow and SAC is offered in Section A.7. Although the training time of our method is longer than that of SAC, SAC can not model multi-modal action distribution. For a fair comparison, we compared the training time of SAC-Flow (i.e., [2]) and MEow. We found that the training time of SAC-Flow is 1.13x slower than MEow due to the additional policy improvement updates (i.e., Eq. (5)). [1] Dinh et al. NICE: Non-linear Independent Components Estimation, ICLR Workshop 2015.\ [2] Haarnoja et al. Latent Space Policies for Hierarchical Reinforcement Learning, ICML 2018. --- C2. In the experiments, MEow is just compared with several classical RL methods for continuous control and lacks the comparison with advanced RL methods [1, 2] with other generative models like diffusion model and consistency model, which also supports the modeling of multi-modal action distribution. Response: We appreciate the reviewer’s suggestion and have included a performance comparison between MEow and DIPO [1] in Table 1 in the [global comment](https://openreview.net/forum?id=lhlIUxD5eE&noteId=Vapk6NMOyp). The results show that MEow outperforms DIPO, even when DIPO uses 100 sequential forward passes during action sampling. To the best of our knowledge, consistency models have not been applied to online RL setups. In addition, combining consistency models with online MaxEnt RL may be challenging due to the intractability of the entropy calculation. On the other hand, reference [2] addresses offline RL tasks, which differ from the online RL setting discussed in this paper. The primary distinction is that in offline RL, agents cannot interact with the environment during training. Therefore, reference [2] represents a separate research direction and is distinct from the main focus of this paper. Finally, we would like to highlight that several online MaxEnt RL methods developed for modeling multi-modal policies are discussed in Section 2.2 and compared in Section 4.2. We use a number of representative MaxEnt RL methods, such as SQL, as our baselines in Fig. 3. We also compare our method with more advanced variants like SAC-Flow (i.e., [3]), denoted as ECFA in Fig. A3. Furthermore, Table 1 in the [global comment](https://openreview.net/forum?id=lhlIUxD5eE&noteId=Vapk6NMOyp) demonstrates that our method outperforms the latest MaxEnt online RL framework (i.e., S2AC [4]), where S2AC also supports modeling multi-modal action distributions. [1] Yang et al. Policy representation via diffusion probability model for reinforcement learning, 2023. \ [2] Yue et al. Boosting offline reinforcement learning via data rebalancing, 2022.\ [3] Haarnoja et al. Latent Space Policies for Hierarchical Reinforcement Learning, ICML 2018.\ [4] Messaoud et al. S2AC: Energy-Based Reinforcement Learning with Stein Soft Actor-Critic, ICLR 2024. --- Questions Q1. How is the training time of MEow compared with other methods? Response: The training time comparison is discussed in the response of C1. --- Q2. How is MEow compared with other advanced RL methods with the utilization of generative-model-based policy? Response: Please refer to the responses to C2 and Table 1 in the [global comment](https://openreview.net/forum?id=lhlIUxD5eE&noteId=Vapk6NMOyp). --- Q3. Could the authors explain why MEow can learn a good reward shifting rather than a trivial shifting, like for any $s_t$, $b(s_t)=0$? Response: We appreciate the reviewer's question regarding MEow's ability to learn a good reward shifting rather than a trivial one. The inclusion of a learnable reward shifting term serves a crucial purpose: it prevents numerical errors in the Jacobian determinant calculations, as discussed in Section 3.2 and analyzed in Appendix A4. The efficacy of this approach is demonstrated in Fig. 1 of the main manuscript, where we observe the successful resolution of these numerical issues. Furthermore, Fig. 6 in the manuscript illustrates that this method contributes to improved overall performance. These results collectively indicate that MEow learns a meaningful and beneficial reward shifting, rather than a trivial one. In addition, according to our observation, $b_\theta$ does not learn to be a constant. Its value changes according to different visited states. To verify this, we sample a trajectory of states and plot the corresponding values of $b_\theta$ for the Hopper-v4 environment during the inference phase in Fig. 5 of the attached PDF file in the [global comment](https://openreview.net/forum?id=lhlIUxD5eE&noteId=Vapk6NMOyp). --- Rebuttal Comment 1.1: Comment: Thanks for your responses and the efforts on the additional experiments. However, as shown in Table 2 of "Boosting offline reinforcement learning via data rebalancing", the experiments in online RL setting are also considered. In that case, the reviewer thinks this paper cannot be viewed as separate research work from yours. However, the contribution and novelty of this paper should be acknowledged. Hence, I am sticking to my score. --- Reply to Comment 1.1.1: Comment: We sincerely appreciate the valuable feedback and recognition of our work's contribution and novelty. Thank you once again for your thorough review.	Rebuttal 1: Rebuttal: This global comment includes additional experimental results and extended discussions addressing the questions raised by reviewers UfJ6 and S3sW. --- ### Additional Results The attached PDF file contains five figures, denoted as Figs. 1-5, which encompass the following content: - Fig. 1: A performance comparison of MEow and SAC trained with different target smoothing values ($\tau$). - Fig. 2: The soft value functions and the trajectories generated using MEow, SAC, and SQL on the multi-goal environment. - Fig. 3: The return versus training time comparison of MEow and SAC. - Fig. 4: $V_\theta$ and $b_\theta$ of MEow evaluated on the multi-goal environment. - Fig. 5: The learnable reward shifting values evaluated along a trajectory of states in the Hopper-v4 environment. The following table compares the performance of MEow and a number of recent online RL methods: \|\|Hopper\|HalfCheetah\|Walker\|Ant\|Humanoid\| \|-\|-\|-\|-\|-\|-\| \|MEow\|3332.99 $\pm$ 521.63\|10981.47$\pm$1812.97\|5526.66 $\pm$ 276.99\|6586.33$\pm$188.73\|6923.22$\pm$125.93\| \|DIPO [1] ($K$=100)\|3123.14$\pm$ 636.23\|10472.31$\pm$654.96\|4409.18$\pm$469.06\|5622.30$\pm$487.09\|4878.41$\pm$822.03\| \|DIPO [1] ($K$=50)\|3214.83$\pm$491.15\|9198.20$\pm$1738.25\|4199.34$\pm$ 1062.31\|4877.41 $\pm$ 1010.35\|4513.39$\pm$ 1075.94\| \|DIPO [1] ($K$=20)\|2511.63 $\pm$ 837.03\|9343.69 $\pm$ 986.82\|4467.20 $\pm$ 368.13\|5288.77 $\pm$ 970.35\|4294.79 $\pm$ 1583.48\| \|S2AC [2]\|<3100\|<10000\|<3500\|<3000\|<3500\| Table 1. Comparison of the average return between MEow, DIPO [1], and S2AC [2]. The results for DIPO [1] are obtained directly from Table 1 of their original paper. $K$ represents the number of diffusion steps used in the sampling process of [1]. The results for S2AC [2] are derived from Fig. 8 of their original paper. [1] Yang et al. Policy representation via diffusion probability model for reinforcement learning, 2023.\ [2] Messaoud et al. S2AC: Energy-Based Reinforcement Learning with Stein Soft Actor-Critic, ICLR 2024. --- ### Reviewer UfJ6 (Cont'd) Q4. What are the learning dynamics observed during the training process? Are there any notable differences in the convergence behavior compared to traditional actor-critic methods? Response: In Section A.5.3 of the Appendix, we compare MEow with four different types of actor-critic frameworks. Our observations indicate that MEow converges to policies that achieve higher returns in most of the environments. In particular, the comparison between MEow and FCFA (i.e., an actor-critic framework modeled using two normalizing flows) suggests that the integration of policy improvement and evaluation steps enhances training stability and overall performance. This finding substantiates the primary claim of our paper. --- Q5. How does the proposed framework scale with increasing complexity of the environment or the state-action space? Are there any specific limitations or challenges observed during experimentation? Response: Both the loss function calculation and the inference process of MEow are scalable with respect to the state-action space. This scalability is attributable to EBFlow's compatibility with existing normalizing flow architectures that support efficient inverse transformation and Jacobian determinant calculation for high-dimensional inputs. These architectures can be utilized in MEow to facilitate efficient training and inference processes with a large action space. For environments with a large state space, states can be encoded using hypernetworks (e.g., MLP in our implementation) to reduce their dimensionality, thus ensuring scalability. Our experiments in high-dimensional environments, such as Humanoid (state dim.: 108; action dim.: 21) and AllegroHand (state dim.: 72; action dim.: 16), support this claim. --- Q6. How robust is the EBFlow framework to different hyperparameters? Response: We appreciate the reviewer’s suggestion. In response, we conducted a performance comparison under different values of $\tau$ (i.e., smoothing target factor in Section 3.3) on five MuJoCo environments. The results of this analysis are presented in Fig. 1 in the attachment of this global comment. --- ### Reviewer S3sW (Cont'd) Q2. In Figure 3, SAC and MEow show similar performance in the Mujoco domain, but there is a noticeable difference in performance in the Isaac Gym domain in Figure 4. Could you provide a further analysis on the reasons for the occurrence of this difference? It would provide valuable insight into the strengths and limitations of the proposed method. Response: The difference in performance between SAC and MEow in the MuJoCo and Isaac Gym environments can be attributed to the varying complexity and nature of tasks in these environments. In the MuJoCo domain, the five tasks exhibit relative homogeneity, primarily focusing on locomotion. These tasks involve straightforward action outputs, typically torque. On the other hand, the Isaac Gym domain presents a more diverse and complex set of tasks. For instance, the ANYmal robot, unlike HalfCheetah, must navigate in multiple directions and follow specific target velocities. Its action space is different as well, involving outputs related to joint position targets rather than torque. Dexterous manipulations, such as those required in AllegroHand tasks, are widely recognized as challenging in the realm of RL. These tasks entail intricate interactions and contacts between the hand and the objects it manipulates. While SAC performs comparably to MEow in the relatively simple MuJoCo tasks, MEow demonstrates greater robustness and generalizability in the more diverse and complex tasks of the Isaac Gym domain. This highlights the strengths of MEow in handling a wider range of challenging scenarios. Pdf: /pdf/ff64c09c5b0d23c96aadf0469d8bcece3af2de43.pdf	NeurIPS_2024_submissions_huggingface	2,024	Summary: The paper presents a new method for MaxEnt RL based on the recently proposed energy-based normalizing flows (EBFlow). The adoption of EBFlow allows us to overcome two major issues with training MaxEnt RL algorithms: (I) sampling from an energy-based model (of the policy) and (ii) approximating the soft value function (with MC methods). Strengths: * Contribution: I think the contribution of this work is significant and novel. The issues discussed with MaxEnt RL are important. The idea proposed is novel, it naturally emerges from the EBFlow framework and it's seamlessly applied to the MaxEnt RL domain. The approach solves the problems mentioned and it demonstrates superior performance compared to other standard approaches. * Presentation: the presentation is clear and well-organised, and the method is described in detail. Weaknesses: * New and old challenges: while the approach resolves some issues with the current MaxEnt RL methods, it still suffers from typical deep RL issues (e.g. overestimation, for which the authors proposed SCDQ) and it introduces additional challenges (e.g. exploding Jacobian determinants, for which the authors introduce LRS). Normalizing flows also introduce additional constraints on the network's architecture, a problem which I believe is strongly mitigated by the introduced LRS (?) * Required clarification on LRS: see Questions Technical Quality: 3 Clarity: 3 Questions for Authors: * Typo at line 149, "Jocobian" * I am not sure if the role of the Learnable Reward Shifting (LRS) is clear to me. The authors say that they were inspired by the reward-shifting literature. However, I don't see the LRS term working the same way as classical reward shaping, as a state-dependent non-linear function ($b_\theta$), without any particular constraints (I don't see any stated) wouldn't preserve the properties of the original functions. I think the role of the learnable reward shifting is to have some form of "baseline" in the Q and V functions that simplifies the estimation process (since it's using a more flexible learning architecture), and so it avoids the explosion of the Jacobian. Could the authors provide clarification on this? * I would be interested in seeing a comparison between the learned reward shifting term and the value function (V) in a plot (it would suffice to see this in the Appendix) Confidence: 3 Soundness: 3 Presentation: 3 Contribution: 4 Limitations: Limitations are addressed in Appendix. I think the authors should mention this in the main text. Flag For Ethics Review: ['No ethics review needed.'] Rating: 7 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We appreciate the reviewer’s valuable feedback and effort spent on the review and would like to respond to the reviewer’s questions as follows. --- Questions Q1. Typo at line 149, "Jocobian". Response: Thank you for pointing this out. We will correct the typo in the final revision. --- Q2. I am not sure if the role of the Learnable Reward Shifting (LRS) is clear to me. The authors say that they were inspired by the reward-shifting literature. However, I don't see the LRS term working the same way as classical reward shaping, as a state-dependent non-linear function ($b_\theta$), without any particular constraints (I don't see any stated) wouldn't preserve the properties of the original functions. I think the role of the learnable reward shifting is to have some form of "baseline" in the Q and V functions that simplifies the estimation process (since it's using a more flexible learning architecture), and so it avoids the explosion of the Jacobian. Could the authors provide clarification on this? Response: Based on our formulation, the original $Q_\theta$ and $V_\theta$ functions are the residuals between their respective targets and $b_\theta$. We agree with the reviewer that this property of $b_\theta$ helps $Q_\theta$ and $V_\theta$ in predicting their targets during the optimization process. Another important aspect of LRS is that it reduces the magnitude of the soft value function $V_\theta$, which makes our model less susceptible to numerical calculation errors. An analysis of this issue is provided in Section A4. This problem arises due to limited numerical precision. Consider a case where FP32 precision is in use, `MEow (Vanilla)’ could fail to learn a target $V_\{\theta^{\}\} (s_t)>38, \forall s_t$, since $\exp(-V_\{\theta^{\}\}(s_t))=\Pi_\{i \in S_l\} \\| \det (J_\{g_\{\theta^{\}\}^i\}(s_t)) \|<2^{-126}$ cannot be represented using FP32 precision. Therefore, without shifting the reward function, the loss sometimes becomes undefined values, and can lead to ineffective training (e.g., the green lines in Fig. 6). The reward shifting term can be designed as a (state-conditioned) function or a (non-state-conditioned) value. It can also be learnable or non-learnable. All of these designs (i.e., $b_\{\theta\} (s_t)$, $b$s_t$$, $b_\{\theta\}$, and $b$) can be directly applied to MEow since none of them influences the action distribution. Based on our preliminary experiments, we identified that a learnable state-conditioned reward shifting worked the best and proposed this technique. --- Q3.* I would be interested in seeing a comparison between the learned reward shifting term and the value function (V) in a plot. Response: We appreciate the reviewer's suggestion. The plot is offered in Fig. 4 of the attached PDF file in the [global comment](https://openreview.net/forum?id=lhlIUxD5eE&noteId=Vapk6NMOyp). --- Rebuttal Comment 1.1: Comment: I am satisfied with the authors' response, which clarifies my main concerns with the work. I think that a more appropriate description of how LRS is useful should be included in the paper, to improve clarity about the authors' contribution. A description similar to the one provided here would suffice. The additional results presented also seem to reflect my original understanding and the authors' response about how LRS works and its useful in their method. I increased my score to 7 and will recommend acceptance of the work. --- Reply to Comment 1.1.1: Comment: We sincerely appreciate the reviewer’s response and valuable feedback. We would be glad to incorporate the additional explanation of LRS into the manuscript in the final revision. Thank you again for your thoughtful review.	null	null	null	null	null	null
Clustering with Non-adaptive Subset Queries	Accept (poster)	Summary: [Setting] This paper studies the problem of clustering $n$ items into $k$ clusters using an oracle that can tell how many ground-truth clusters are represented in any given subset $S$ of items. The goal is to develop non-adaptive algorithms (where all queries are chosen before the oracle answers anything) and study their sample complexity. [Contributions] 1. Randomized non-adaptive algorithms for constant k that make: 1. $O(n k \log \log n \log k)$ queries when subsets of any size can be queried 2. $O(n k \log n \log\log n)$ queries when $\|S\| = O(\sqrt{n})$ 2. Randomized non-adaptive algorithms for general k that make: 1. $O(n \log n \log k (n/s + \log s))$ queries of size at most $s$ 3. An $\Omega(\min(n^2 / s^2, n))$ lower bound for any non-adaptive algorithm that is allowed to query subsets of size at most $s$. The developed algorithms run in polynomial time. Additionally, the paper also presents results for the case when : 1. Clusters are roughly balanced (i.e., their sizes are within constant factors of each other). Here, the algorithm makes $O(n \log k)$ queries when $k = O(n/\log^3 n)$ and $O(n \log^2 k)$ queries for any $k$. 2. An adaptive round is allowed. This improves dependency on logarithmic factors. For these results, the details are only in the appendices. Strengths: 1. The paper is fairly comprehensive and presents query complexity results in a wide range of settings. The proposed algorithm are optimal up to logarithmic factors. 2. The algorithms are well motivated. I especially appreciate the high-level intuition because of the clarity it adds to the paper. 3. While I have some concerns about the problem setting (see Weaknesses), I believe the presented results are a very good first step towards filling an interesting research gap - non-adaptive algorithm for clustering with subset queries. Weaknesses: 1. My primary concern is with the feedback model. The oracle returns the number of clusters in the subset $S$, which implies that it knows the correct grouping in $S$. Why then would it only return the number of clusters in $S$ in practice and not the clusters themselves? Perhaps there are bandwidth constraints on response in some applications? But wouldn't one require way fewer queries if the oracle returns the clusters in $S$, which alleviates the bandwidth concerns, especially in the non-adaptive setting? 2. This is a relatively minor concern, but the paper defers a lot of details to the appendix without even including high-level ideas. I recommend including some details in the main paper, perhaps at the cost of a few technical details from Section 2. For example, 1. What changes to the algorithm allow the query complexity to improve for the roughly balanced case? 2. What is the "extra round of adaptivity"? Technical Quality: 4 Clarity: 3 Questions for Authors: 1. Please address point 1 under weaknesses. 2. It seems to me that both Algorithm 1 and 2 work for any value of k. Can the authors elaborate on what they mean by constant $k$ and general $k$? Do you have two algorithms because one of them has a better dependence on $k$ ($O(\log k)$ instead of $O(k \log k)$)? On a similar note, how should one go about choosing between Algorithm 1 and 2 in practice? 3. Your lower bound is based on the idea of converting subset queries to pairwise queries. Unfortunately, this hides the dependence on $k$. Are you aware of any lower bounds in the pairwise case that depend on $k$ as well? Confidence: 2 Soundness: 4 Presentation: 3 Contribution: 4 Limitations: While it is true that all theorems have their assumptions listed, it would help the readers if the authors include a dedicated paragraph pointing out some of the research gaps that are left open by their work. Flag For Ethics Review: ['No ethics review needed.'] Rating: 7 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thanks for this constructive evaluation. Please see the global response for the question related to motivation. The rest of the responses are provided below. We will also add some high level ideas from the appendix to the main paper as suggested. Can the authors elaborate on what they mean by constant $k$ and general $k$ You are correct that both algorithms work for any value of $k$. We have two algorithms because one of them has better query complexity when $k$ is small and the other has better query complexity when $k$ is large. Specifically, the algorithm of Theorem 1.2 makes $O(n \log \log n \cdot k \log k)$ queries while Theorem 1.1 gives $O(n \log^2 n \log k)$. Thus, to be precise, Theorem 1.1 is better when $k \leq O(\frac{\log^2 n}{\log \log n})$ and Theorem 1.2 is better otherwise. In practice, the algorithm should be chosen based on where $k$ falls with respect to this threshold. We wrote "constant $k$" because we wanted to emphasize that Theorem 1.2 gives query complexity $O(n \log \log n)$ when $k$ is a constant and "general $k$" to emphasize that Theorem 1.1 gives query complexity $O(n \log^2 n \log k)$ for all $k$. We will add a few lines in the write up to make this more clear. Are you aware of any lower bounds in the pairwise case that depend on $k$ as well? For non-adaptive pairwise query algorithms there is an $\Omega(n^2)$ lower bound for $k=3$ which is as strong as possible (for any $k\geq 3$) since $O(n^2)$ is also a trivial upper bound for any $k$. For $k = 2$, $O(n)$ is possible non-adaptively and this is optimal since the algorithm will need at least one pairwise query involving every point. Thus, this is the complete picture for non-adaptive pairwise query algorithms. --- Rebuttal Comment 1.1: Comment: Thank you taking time to respond to my questions. I appreciate the added motivation for subset queries and the clarification regarding constant-vs-general k. All the best for your submission !	Summary: The problem of cluster recovery via membership queries asks to assign each point in a dataset to its $k$ ground-truth cluster by using as few queries as possible to an oracle. A well-studied oracle is the same-cluster oracle that answers, given two points, whether the points are in the same or in different clusters. The complexity of adaptive algorithms for this problem is $\Theta(nk)$, but non-adaptive algorithms require $\Omega(n^2)$ for $k \geq 3$. The authors of this paper consider subset queries to surpass this barrier. Given a set of points $S$, a subset query returns the number of clusters the points in $S$ belong to. It is known that the complexity of an adaptive algorithm is $O(n)$. The authors consider non-adaptive algorithm and obtain the following results. For general subset queries, they obtain an algorithm that makes $O(n \log^2 n \log k)$ queries for unbalanced clusters, and $O(n \log k)$ queries for balanced clusters. For constant $k$, they obtain an algorithm that makes $O(n \log \log n)$ queries. If the allowed subset size of the queries is bounded by $s \leq \sqrt{n}$, their algorithms need an essentially optimal $\tilde{O}(n^2 / s^2)$ number of queries for constant $k$, and $\tilde{O}(n^2 / s)$ for arbitrary $k$. Finally, they show that 2-round adaptivity can improve on the log factors in the complexity. The algorithms follow two algorithmic approaches. For constant $k$, the authors observe that one can use subset queries to identify an isolated point representing its cluster in a subset and explore the cluster from this point, or one can use subset queries that contain two points from the same cluster to sample intra-cluster edges. The two strategies have a reverse complexity with respect to the cluster size, and a combination of them give the final complexity trade-off. For general $k$, the authors related subset queries to Boolean OR queries, and use these to recover clusters one by one. Strengths: The paper provides a rather rich collection of results on foundational variants of the non-adaptive cluster recovery problem with near-linear complexity for unbounded subsets. Weaknesses: Unbounded subset size may be a rather strong assumption, and the results for bounded subset size of non-adaptive algorithms are not near-linear and almost trivial for constant subset size. One may be interested in few-round adaptive algorithms with bounded subset size. Technical Quality: 3 Clarity: 3 Questions for Authors: / Confidence: 3 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: / Flag For Ethics Review: ['No ethics review needed.'] Rating: 7 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thank you for the very thoughtful review. Unbounded subset size We agree that unbounded subset size is a strong assumption and we would like to better understand the query complexity for bounded size in follow-up work. We have shown that $O(\frac{n^2}{s^2} k\log n)$ is possible non-adaptively with $s \ll \sqrt{n}$ and this is essentially optimal in terms of $n$ and $s$. It would be interesting to know if near-linear query complexity is possible with query sizes that are even smaller than $\sqrt{n}$, say ${\rm poly}\log(n)$. We know this is not possible non-adaptively, but your idea of studying bounded subset size algorithms with few rounds of adaptivity is a great idea to look at in follow-up work. --- Rebuttal Comment 1.1: Comment: Thank you for your response! If I understand correctly, the details you provided agree with my understanding at large, and in particular, the bound is essentially optimal for "the largest possible" $s \in o(\sqrt{n})$ only (and not all $1 \leq s \ll \sqrt{n}$). If that's not the case, please follow up and feel free to let me know. --- Rebuttal 2: Comment: Thank you for seeking clarification. We apologize if there were some ambiguities in the previous response. Our result is optimal for all values of $s \leq \sqrt{n}$ within a $\log n \log \log n$ factor. In particular, our algorithm in Theorem A.1 (or the simplified version, 1.4) achieves $O(\frac{n^2}{s^2} k \log n \log \log n)$ non-adaptive queries for any $s \leq \sqrt{n}$. Thus for constant k (which is the regime for Theorem A.1 and 1.4), our bounds are optimal for all values of $s \leq \sqrt{n}$ within the $O(\log n \log \log n)$ factor since $\frac{n^2}{s^2}$ is a lower bound for non-adaptive queries. So to summarize, it is not correct that the bound is optimal only for "the largest possible" $s \in o(\sqrt{n})$. Instead, the bound is optimal for all values of $s \leq \sqrt{n}$. --- Rebuttal Comment 2.1: Comment: Sorry for the confusion, that was my bad, and thank you for the clarification! Edit note for review: Taking the whole rebuttal into account, I think this paper should be accepted (change score 6 -> 7).	Summary: The paper studies the clustering problem with non-adaptive subset queries. The problem formulation is as follows. Suppose we are given an oracle $q: \mathcal{V}\rightarrow \mathbb{R}$ such that for any query on a subset $S\subseteq V$ of vertices, the oracle returns how many clusters are in $S$ in the optimal clustering. The goal is to recover the optimal clustering with subset queries that are as small as possible. The problem setting is similar to the clustering with same-cluster query problem, where the oracle returns whether two vertices $(u,v)$ are in the same cluster in the optimal clustering. Both problems are motivated by scenarios where we could use semi-supervised learning to obtain such types of oracles. However, in that setting, the number of necessary queries becomes $\Omega(n^2)$ when $k$ is at least $3$. As such, to circumvent this strong barrier, it is natural to look into the new model with extra power for the queries. In the new model, we can easily obtain an algorithm $O(n\log{k})$ queries by binary search. However, such an algorithm requires at least $\Theta(\log{k})$ rounds of adaptivity. The paper focused on the non-adaptive setting, where the queries have to be made independent of the results of previous queries. The main results include $i).$ an algorithm that makes $O(n \log\log{n})$ queries when $k=O(1)$ (linear dependency on $k$), and; $ii).$ an algorithm that makes $O(n\log^{2}{n} \log{k})$ queries for general $k$. The paper also studied the case when the maximum size of $S$ is bounded and obtained tight bounds when ${\|S\|}\leq \sqrt{n}$. Techniques. The two main algorithms used different ideas. For the algorithm with constant $k$, the key observation is that after recovering the biggest cluster(s) in the queried set, there are non-trivial probability that the remaining of the queried set has exactly one or two points. For the two cases, we can recover clusters via subset queries for some specific sizes. As such, we can repeat enough times, and use a ‘valid’ set to recursively recover the clusters in the reconstruction phase. For the algorithm with general $k$, the paper reduces the problem to solving the OR queries of boolean vectors. In this way, the paper provides some new results for OR queries and obtains the desired bound. Strengths: I have a positive view of the paper. In my opinion, the problem setting is well-motivated, and it is a natural extension of the well-studied same-cluster query model. The paper is well-written, and the authors did a nice job explaining the ideas in the high-level overview. Therefore, although the techniques are non-trivial, I managed to follow the ideas well. Given the page limits of the conference, the paper still managed to do fairly well in terms of organization. Weaknesses: I don’t see any major issue in the paper. Some minor issues to address: - It might be good to further clarify what ‘non-adaptive’ means in your paper. I think in your algorithm, you could still adaptively ‘add on’ to clusters after making the queries. For a moment I thought your algorithm could conduct $\log {n}$ rounds in parallel to recover all clusters, which will be even stronger in the ‘non-adaptive’ notion. - The paper doesn’t contain experiments. I’m personally OK with it, but I think for NeurIPS you might want to justify why this is not an issue. I think AKB [NeurIPS’16] (‘clustering with same-cluster queries’) gives a nice motivation for the same-cluster oracle model, and maybe this paper should provide some similar justifications. - The dependence on $k$ should be mentioned after theorem 1.2. I think $O(nk)$ is really not that bad – I thought it was something like $2^k$. Technical Quality: 4 Clarity: 4 Questions for Authors: Most of the questions are asked in the ‘weakness’ part. Two MISC questions: - Line 165: should it be $p\leq \log\log{n}$? - If we just want the clustering of an $\alpha$ fraction of the vertices, can we use an even smaller number of queries? Confidence: 4 Soundness: 4 Presentation: 4 Contribution: 4 Limitations: I do not see any potential negative societal impacts. Flag For Ethics Review: ['No ethics review needed.'] Rating: 8 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thanks for this very encouraging message. Please see the responses to your evaluation below. It might be good to further clarify what ‘non-adaptive’ means in your paper Thank you, that is a good point. Our notion of non-adaptivity is only that the queries are made in one round. I.e. queries are not made based on responses from previous queries. Once the query responses are obtained there is no adaptivity-esque restriction on how the algorithm recovers the clusters. We agree that this should be clarified, and will do so in the future version. Experiments Being a theoretical paper, we thought simulations will not add a lot of value to the paper; however we can certainly add simulation of our algorithm. The dependence on $k$ Thank you, that is a great point. We will add clarification regarding the dependence on $k$ directly after the theorem statement. Line 165: should it be $p \leq \log \log n$? No, here $p \leq \log n$ is correct since $\delta$ could be as small as $1/n$ and we need $2^p$ to approximate $1/\delta$ within a constant factor for some $p$. Note that in this paragraph we are only describing how this strategy leads to a $O(n \log n)$ query algorithm and the $O(n \log \log n)$ comes by combining the two strategies as described in the subsequent paragraphs. If we just want the clustering of an $\alpha$ fraction of the vertices, can we use an even smaller number of queries? This is a great question. Identifying the cluster containing any one point requires $\Omega(\log k)$ bits and so determining the clustering of $\alpha n$ points requires $\Omega(\alpha n \log k)$ bits. Since a query gives $O(\log k)$ bits of information we require $\Omega(\alpha n)$ queries to cluster an $\alpha$-fraction of the points. In particular, for any constant $\alpha$, the $\Omega(n)$ lower bound still holds. To answer your question directly, we do think there is a simple way to modify our algorithm to give improved query complexity in this case. For instance in "Strategy 1" described in line 161, we only need to handle the case of $\delta \geq \alpha/2$ since we can ignore clusters that are smaller than $\frac{\alpha}{2}n$ (we are talking specifically about the $k=3$ case, but this could be generalized). Thus we only need to iterate over $p \leq \log \frac{2}{\alpha}$ and this naturally leads to a $O(n \log \frac{1}{\alpha})$ query algorithm, which is $O(n)$ for constant $\alpha$. We will add this to the paper. It would be interesting to see if it is possible to get close to $O(\alpha n)$ for general $\alpha$, and we don't immediately see how to achieve this using our current ideas. Thank you for the great question. --- Rebuttal Comment 1.1: Comment: Thanks for the clarifications and the response. I do not have any further questions.	Summary: The paper gives results on clustering a dataset using "subset queries." This is a generalization of "same-cluster queries." The same cluster query asks whether two given elements belong to the same optimal cluster. The subset query asks how many different clusters the elements of a given subset span. There has been recent interest in clustering using same-cluster queries. This paper is an extension of this line of work. Clearly, O(n^2) queries are sufficient to cluster all the points. We would want to explore the possibility of making a much smaller number of queries. There are two sub-models to consider when discussing subset queries: 1. Adaptive queries 2. Non-adaptive queries. Adaptive queries allow the query algorithm to adapt new queries based on the response for the earlier queries, whereas, in the non-adaptive model, all subsets should be predefined before making any queries. Adaptive queries are more powerful since O(nk) adaptive same-cluster queries are sufficient, whereas O(n^2) are required in the non-adaptive setting. This paper gives results in the non-adaptive setting: - An O(n log^2{n} log{k}) query algorithm for general values of k. This is using a reduction to a known problem in group testing. This is improved to O(n log{k}) for the balanced case (when cluster sizes are within a constant factor of each other). - An O(n log log{n}) algorithm for constant values of k. This is using a simple randomized method. Strengths: - From a purely theoretical point of view, the results are interesting. The discussion is elaborate, with various special cases discussed. Except for logarithmic factors, the bounds are tight. - The writing is good. The intuition is clearly stated within the main write-up, and the proofs are in the appendix. Weaknesses: - The motivation for considering subset queries is not immediately clear. The motivation for same-cluster queries was easier to understand since one can set up a crowdsource platform and ask users to check whether two objects (say images) belong to the same class. Setting up subset queries for large subsets may be tricky since answering such queries is not simple. Technical Quality: 4 Clarity: 3 Questions for Authors: - Subset queries may not be easy to answer accurately, specifically when the subsets are large. However, answering such queries approximately (within some margin of error) may be possible. It may be interesting to explore whether some interesting clustering can be obtained using a small number of such "inaccurate approximate queries." Confidence: 4 Soundness: 4 Presentation: 3 Contribution: 3 Limitations: This work is theoretical, and there are no potential negative societal impact. Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thanks for the thoughtful evaluation of the paper. Please see the global response for the motivation of the model. Subset queries may not be easy to answer accurately This is a fantastic point. Clustering under noisy subset queries is a direction we're interested in exploring in follow-up work. Positive results under a meaningful/realistic noise model would create a more convincing justification for studying subset queries. We believe that our work is an important first step towards this. --- Rebuttal Comment 1.1: Title: Acknowledgement Comment: Thanks for your response. I have no further questions. I will maintain my current score.	Rebuttal 1: Rebuttal: We sincerely thank all of the reviewers for their very thoughtful reviews and comments. Motivation for Subset Queries. As reviewer SHzL has pointed out, clustering with subset queries is a natural extension of the well-studied same-cluster queries. While with same-cluster queries we would require $O(n^2)$ non-adaptive queries to recover the clusters even for small $k$, here we show that just near-linear number of queries suffice with subset queries. To address the questions related to the practical motivation for the model, we provide a few justifications below. Rest of the questions have been addressed with a response to each reviewer separately. - In the paper, we have considered non-adaptivity where queries are asked in one round as a major practical motivation. Several prior works have shown it is extremely time consuming to get query answers adaptively from a crowd. Subset queries is the only way to get good query complexity while considering non-adaptivity. - As a reviewer pointed out, bounded size queries are more practical than unbounded size queries. We have studied the effect of query size for both of our algorithms. Our strategy 1 gives optimal dependency on query size. - Our work serves as a stepping stone for future work to consider few rounds of adaptivity (partial answers have been provided in Appendix F), and where errors are allowed in query answers. - As reviewer hP14 brings up, if there is a bandwidth constraint, then returning just a number is much easier. Suppose the algorithm sends its queries $S_1,\ldots,S_q$ off to $q$ different entities. Suppose these entities have a bandwidth constraint in that they do not want to send back much information. Answering a subset query requires them to send only $O(\log k)$ bits (just a number), instead of $O(\|S_i\| \log k)$ needed for describing the whole clustering on $S_i$. The total number of bits sent is $O(q \log k) = \widetilde{O}(n)$ and the max number of bits sent by any entity is $O(\log k)$. - If the model was instead that the query returned the entire clustering on $S$, then the problem is trivially solvable with $1$ query when unbounded size is allowed. On the other hand, for bounded size queries, it does not change the information theoretic lower bound. - Consider the problem of counting the number of clusters represented in $S$ vs computing the clustering in $S$ from the perspective of an entity answering a query. The counting problem is easier, because one does not have to resolve ambiguities in assignments. Therefore, even if assignments can be erroneous, counts are likely to be correct (hence more robust). Again, thank you to all of the reviewers. We greatly appreciate the time and effort you've invested to carefully read and evaluate our work.	NeurIPS_2024_submissions_huggingface	2,024	null	null	null	null	null	null	null	null
In-Context Learning with Transformers: Softmax Attention Adapts to Function Lipschitzness	Accept (spotlight)	Summary: This study examines the impact of softmax attention on ICL regression, advancing beyond the typical linear treatment of the topic. In particular, the authors find that 1) softmax attention leads the model to adapt to the target function's Lipschitz constant and 2) enables the model to recover low dimensional structure in the target. Strengths: The paper is very well-written and easy to read. It's great to see work that pushes beyond the usual linear treatment of attention, and explicitly considers the role of the softmax nonlinearity. The connection to kernel smoothing is very neat, and the resulting interpretation of softmax attention as an adaptive process sensitive to the target's Lipschitz constant is very insightful. I was particularly surprised that Lipschitz adaptation is both necessary and sufficient to illustrate the role of softmax attention in ICL regression, and that swapping function classes did not matter so long as the Lipschitz constant remained the same. Very cool work overall. Well done! Weaknesses: I found no substantive weaknesses in the analysis, but do have some follow-up question listed below. As with any theory paper, there's always more that can be done, but I think the level of work demonstrated in this paper is more than sufficient to merit a strong accept. Technical Quality: 4 Clarity: 4 Questions for Authors: - Could your results transfer in any way to an ICL classification setting, for instance like the one in [Reddy 2024](https://arxiv.org/abs/2312.03002) among others? Your Lipschitz adaptation argument, particularly as it controls a "window size" over inputs, sounds obliquely related to the way softmax attention is sometimes discussed as implementing "selection" or "copy" operations over context examples in ICL classification. I'm curious if there could be a deeper connection here? - When you train models in practice, do you find that the QKV matrices of the final model become like the M-parameterization of the matrices you highlight in eq. 4? Confidence: 3 Soundness: 4 Presentation: 4 Contribution: 4 Limitations: The authors adequately address their limitations. Flag For Ethics Review: ['No ethics review needed.'] Rating: 8 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We thank the reviewer for their valuable feedback. Please find the response below, and a global response above. 1. Do trained models have the same decomposition for $\mathbf{W}_K^\top \mathbf{W}_Q$ as in Equation (4)? We indeed observe this decomposition empirically, please see the attached pdf in the global rebuttal for examples of $\mathbf{W}_K^\top \mathbf{W}_Q$ trained on ReLU functions of varying ambient dimensions. 2. Classification problems and selection. Please see our global response for some discussion of how our results may be extended to classification. Regarding the selection idea: One framework that is used to interpret ICL in language is that of the ``induction head", which allows a transformer to implement an algorithm that searches for a previous occurrence of a token and repeating its following token. Concretely, if the context provided is $\{... \texttt{[A]}, \texttt{[B]}, ... \texttt{[A']}\}$, the network outputs $\texttt{[B']}$ such that $\texttt{[A]}$ is to $\texttt{[B]}$ analogously what $\texttt{[A']}$ is to $\texttt{[B']}$. While this is outside the scope of our work, one of the intuitions from our work is that the soft-max attention can help inform this analogy in two ways: (1) by informing a notion of "distance" that constitutes a significance of match between $[\texttt{A}]$ and $[\texttt{A'}]$, and (2) by selecting a subspace within which to calculate attention (in case such structure exists). --- Rebuttal Comment 1.1: Comment: Thank you for your rebuttal and clarifying details. Well done overall! I continue to recommend acceptance, and maintain my current score.	Summary: This paper analyzes how softmax attention learns to perform in-context learning (ICL) through pretraining. The authors show that softmax attention adapts its "attention window" based on the Lipschitzness and noise characteristics of the pretraining tasks. They provide theoretical analysis for affine and ReLU-based function classes, demonstrating that softmax attention learns an optimal trade-off between bias and variance. The paper also explores how softmax attention can recover low-dimensional structure in the input space. Experiments are conducted to validate the theoretical findings. Strengths: - The paper addresses an important and timely topic in understanding the mechanisms behind in-context learning in transformer models. This is a crucial area of research given the widespread adoption of large language models. The paper makes a novel connection between softmax attention and nearest neighbor regression, providing an intuitive interpretation of how ICL works in this setting. - The analysis is clean and the main message is clear: softmax activation plays a critical role in enabling effective in-context learning. For example, Theorem 3.4 provides concrete bounds on how the attention window scales with task Lipschitzness, noise level, and context size. Also, I like Theorem 3.5, whose test data are any arbitrary L−Lipschitz task, which is pretty general ICL setting. - The experimental results generally support the theoretical analysis. Figure 3 nicely illustrates how the spectral norm of M (representing the inverse of the attention window size) increases with the Lipschitz constant L, aligning with the theory. Weaknesses: - There's a potential gap between the paper's main message and practical ICL scenarios. The analysis is based on a nearest-neighbor interpretation, but in real-world applications, LLMs often perform well even when there are no "close" examples in the context. - Some of the experimental writing and presentation could be improved. Figures 4 and 5 are particularly hard to follow without careful reading of the appendix. It would be helpful to have more detailed captions and clearer explanations in the main text. - The theoretical analysis is limited to relatively simple function classes (affine and ReLU-based). While this provides valuable insights, it's not clear how well these results generalize to the more complex functions learned by large language models in practice. Technical Quality: 3 Clarity: 2 Questions for Authors: - How do you think your analysis might extend to more complex, hierarchical function classes that might better represent the capabilities of large language models? How can it be generally applied when there is no neighbor data? - Have you considered how your findings might relate to or explain the emergent capabilities observed in large language models as they scale? Confidence: 3 Soundness: 3 Presentation: 2 Contribution: 3 Limitations: The authors acknowledge some limitations in Section 5, including that their model only considers the output of a single layer of attention and that establishing a mathematical framework for priming effects in LLMs remains an open challenge. These are important limitations to note, as they highlight the gap between this analysis and the full complexity of modern language models. Flag For Ethics Review: ['No ethics review needed.'] Rating: 7 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We thank the reviewer for their valuable feedback. Please find the response below, and a global response above. 1. No ``close" examples, simple function classes. Please see the global rebuttal (Points 1 and 2). To summarize, it is possible that the intermediate layers of the model can learn the notion of closeness that it needs, but this is outside the scope of our analysis of a single layer. 2. Writing improvements. We thank the reviewer for this feedback. Should the paper be accepted, we will make these revisions in the camera-ready version with the extra space provided. 3. How can our results be extended to hierarchical function classes? Our analysis in Section 4 begins to answer this question. When the ground-truth functions are a hierarchy consisting of a linear projection followed by a linear link function, the attention weights learn the direction-wise Lipschitzness of the class, that is, they learn this projection. Based on our intuition and experiments in Figure 10 in Appendix J, we suspect that when the functions may have nonlinear link functions (which are still preceded by a shared linear projection), the attention weights again learn the projection and implement a nearest neighbors regressor in the range of the projection with appropriate neighborhood size based on the Lipschitzness of the link functions and noise level. More complex hierarchies likely require additional layers to learn; this is an interesting direction for future work. --- Rebuttal Comment 1.1: Title: Thanks. Update score from 6 to 7 Comment: I have checked all the rebuttals. They address all my concerns well, and I have updated my score from 6 to 7.	Summary: This paper explores how softmax attention in transformer models enables in-context learning (ICL), where a model can adapt to solve new tasks using only a few input examples without additional training. The authors focus specifically on regression tasks, where the model must predict a continuous value given some input features. They show that during pretraining on a distribution of ICL regression tasks, softmax attention learns to implement a nearest neighbors predictor that is adapted to properties of the pretraining task distribution. The key insight is that softmax attention learns an "attention window" - a neighborhood around each input query point that determines which other input points influence the prediction. The size and shape of this attention window adapts based on properties of the pretraining tasks, specifically their Lipschitzness (how quickly function values can change) and the amount of label noise. Importantly, the authors demonstrate that learning this adapted attention window is crucial for generalization. The authors also prove that softmax attention can learn to project inputs onto a relevant low-dimensional subspace when the pretraining tasks depend only on projections of the inputs onto this subspace. To validate their theoretical results, the authors conduct experiments on synthetic regression tasks with varying Lipschitzness, noise levels, and input dimensionality. Empirically, the authors demonstrate that softmax attention learns appropriate attention window scales across a range of nonlinear function classes, including ReLU networks and trigonometric functions. Strengths: A first strength is that the paper is not only well written but also excellently presented, which helps in understanding its mathematical content and putting it into a better light. The clarity of exposition is therefore a first great point. In terms of originality, it starts off with the fairly known/commonplace insight that there exists a connection between self-attention and Nadaraya-Watson kernel regression, thus establishing that learning the bandwidth of that estimator across multiple tasks is a necessity for ICL. In this sense and if it stopped there, the contribution wouldn't be particularly novel, as this is intuitive (if thinking of self-attention as learning a summary of the autocovariance function of data and therefore of its characteristic length) and fairly well understood in the literature already (see the cited works of Tosatto et al). One can argue - as in the paper - that Tosatto et al only give an upper bound on the bias, and this work provides a lower bound as well. However where the paper takes off in my opinion is when these arguments move away from purely the 'frequency cutoff / Lipschitzness' length-related realm, to then move into the directional, via using concentration arguments on the hypersphere. This is Theorem 4.4 which formalizes the intuition that ICL in transformers identifies low-dimensional subspaces shared by training tasks, an argument more typically found in the analysis of (non-contrastive) self supervised learning. The mathematical method of proof is elegant as well. The authors derive novel concentration inequalities for functionals of points uniformly distributed on high-dimensional spheres, and use a careful symmetry argument to show that any non-zero component in the orthogonal subspace increases the loss. Overall this represents a standout technical contribution well worthy of publication in my view. Weaknesses: - A small weakness in presentation is I believe that Theorem 4.4 should be emphasized, as to my knowledge this is the more novel part of the contribution. - Similarly, the theoretical guarantees are provided for relatively simple function classes (affine and ReLU-based), which may not represent the full range of tasks where ICL is effective. - The limited scope of experiments is understandable given theoretical assumptions (single-layer) but also important enough that it becomes a weakness, IMHO. - Finally, the paper could benefit from a more extensive comparison to other theoretical frameworks for understanding ICL. Technical Quality: 4 Clarity: 4 Questions for Authors: Would there be (not extremely involved) ways of moving towards more realistic settings ? i.e. moving to infinite width settings ? Would classification tasks with a cross-entropy loss be somewhat tractable if making the right Gaussian (concentration) assumptions ? Confidence: 5 Soundness: 4 Presentation: 4 Contribution: 3 Limitations: There are some obvious, if hard to tackle, limitations in this work: a. Simplicity of setting: The analysis focuses on single-layer models and synthetic tasks, which may limit its immediate applicability to more complex real-world scenarios. b. Gap to large language models: While insightful, the work doesn't fully explain emergent ICL in large language models trained on natural data. Flag For Ethics Review: ['No ethics review needed.'] Rating: 7 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We thank the reviewer for their valuable feedback. Please find the response below, and a global response above. 1. Simple function class. Please see Point 1 of the global rebuttal. 2. Extensions Please see Point 3 of the global rebuttal. 3. Emphasis on Theorem 4.4. We thank the reviewer for appreciating Theorem 4.4 as a novel, significant and technically impressive contribution. If the paper is accepted, we will allocate some of the additional space provided in the camera-ready version towards further discussing Theorem 4.4's importance in the introduction. 4. Comparison to other frameworks. We will add a section in the appendix dedicated to induction heads and the framework comparing ICL to gradient based "mesa" optimizers. Please see Appendix A for an extended discussion of related works. We will refine this discussion and include it in the main body given the additional space if accepted. --- Rebuttal Comment 1.1: Comment: Thanks for your reply and rebuttal. I will maintain my score towards acceptance. Nice work !	Summary: The paper studies in-context learning (ICL) of one-layer attention-only transformers in a regression task. The paper argues that the product of query and key projection matrix is associated with the Liptchitzness of input data. The notion of attention windows is introduced based on this. The paper shows that attention windows will adapt to the Lipchitness and the noise level of data and perform dimensional reduction. Experiments are conducted to support theoretical claims. Strengths: The paper provides a rigorous approach to the problem both with proof and empirical simulations. Weaknesses: - Although explaining ICL is important, the setting seems to be less realistic (see question 1). - The approach of understanding ICL through the lens of function Lipschitzness is somewhat limiting. For example, Figure 5 demonstrates a decline in generalization error, which is primarily attributed to the function's structure. While the Lipschitz property of a function is indeed a valuable aspect in studying ICL, it doesn't provide a comprehensive explanation or solution for ICL as a whole. Technical Quality: 2 Clarity: 3 Questions for Authors: - The assumption in Equation 4 can be understood as the optimal weights of values $W_V$ will ignore covariates $x$ while the attention will only consider covariates to decide where to attend in the sequence while disregarding predictor $f$. I wonder if this makes sense for hidden states of LLMs. In particular, there is no separation between covariates and predictors in LLMS. Also, why is the optimal $W_V$ in the proof of Lemma B.1 assumed to have first columns to be zeros (in between Line 577-578)? - I agree that $x^\top M y$ can be understood as the distance between $x$ and $y$. It might not straightforward to change $x^\top M y$ into $\|\|x - y\|\|$ like used in Appendix. Can you clarify this more on how to obtain Equation (8) in Appendix from Equation (ICL) in the main text. Confidence: 3 Soundness: 2 Presentation: 3 Contribution: 2 Limitations: Please see above. Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We thank the reviewer for their valuable feedback. Please find the response below, and a global response above. 1. Closeness of the setting to reality. The data model we consider in the paper is widely studied (Ahn et al. 2023, Akyürek et al. 2023, Mahankali et al. 2023, Garg et al. 2022, von Oswald et al. 2023, etc.) as a setting that demonstrates a form of in-context learning. In that sense, our work brings a common framework for understanding ICL closer to practice using nonlinear function classes and softmax activations. We address the reviewer's specific points related to the closeness of the setting we study to reality below. * Assumption that the optimal $\mathbf{W}_V$ ignores the covariates. At a high level, it makes sense that the optimal $\mathbf{W}_V$ ignores the covariates because the value embeddings should live in label space rather than input (covariate) space. For example, suppose each task $t$ entails translating a word in Language $A_t$ to Language $B_t$. A natural prediction of the translation of the query word is some mixture of the example translated words in Language $B_t$, weighted by the closeness of the corresponding word in Language $A_t$ to the query. Languages $A_t$ and $B_t$ might be very different, so the prediction should not include any mixture of the example words in Language $A_t$ (as the prediction must be in Language $B_t$). Returning to our setting, we prove that the optimal $\mathbf{W}_V$ leads to a prediction that ignores the covariates; this is not an assumption. First, the estimator does not depend on the first $d$ rows of $\mathbf{W}_V$ (not columns, as we mistakenly said in line 135), since these only affect the top $d$ entries of the output token which does not affect the loss, which is the standard loss studied by e.g. Ahn et al. 2023, Zhang et al. 2023a, and Mahankali et al. 2023. Second, and more importantly, we prove in Lemma B.1 that the optimal value of the first $d$ elements in the $(d+1)$-th row of $\mathbf{W}_V$ is zero. Concretely, suppose $\textbf{W}_V = \begin{bmatrix} \mathbf{A}&\mathbf{b} \\\\ \mathbf{v}^\top & c \end{bmatrix}$ and let the attention weight on the $i$-th token be $\beta_i$ (as in the paper). Then the estimator, which is the $(d+1)$-st coordinate of the output token, is $\sum_i \left(cf(\mathbf{x}_i)\beta_i + \mathbf{v}^\top \mathbf{x}_i\right)$. As mentioned above, this does not depend on $\mathbf{A}$ or $\mathbf{b}$ (the first $d$ rows of $\mathbf{W}_V$ only affect the first $d$ rows of the output), and we show in Lemma B.1 that the optimal value of $\mathbf{v}$ is $\mathbf{0}$. Please see the proof of Lemma B.1 for additional technical details. * Assumption that the optimal attention parameters $\mathbf{W}_K^\top \mathbf{W}_Q$ ignore the predictor $f$. Again using the translation task example, it makes sense that the attention weights should only depend on the similarities of the query word in Language $A_t$ with the example inputs in Language $A_t$, since the query lives in a different space than the example labels (which are in Language $B_t$) and thus cannot be compared with them. In the model we use, we follow prior works by setting the last rows of $\mathbf{W}_K$ and $\mathbf{W}_Q$ to zero; for example please see Equation (2) in Ahn et al. 2023 and Equation (11) in Mahankali et al. 2023. * Clarification on how to derive $\Vert \mathbf{x}-\mathbf{y}\Vert$ from $\mathbf{x}^\top \mathbf{My}$. We would like to clarify that we show $e^{\mathbf{x}^\top \mathbf{My}}$ simplifies to $e^{-w\|\|\mathbf{x}-\mathbf{y}\|\|^2}$. To do this, we show all optimal $\mathbf{M}$ must be of the form $\mathbf{M}=w\mathbf{I}_d$. Then, we complete the square, and use the fact that all covariates are on the unit sphere to cancel the squared terms in both numerator and denominator, leaving us only with exponents raised to the cross terms. Please see lines 593, 594, Equation (7), and Lemma B.5. * Understanding in-context learning via Lipschitzness. In the model that we consider, we show how a transformer can exploit a shared Lipschitzness in pretraining for ICL. Figure 5 shows that this is necessary and sufficient for generalization: the attention layer fails to generalize to tasks that would be considered as sharing a common structure to those it sees in pretraining. A model that is pretrained on Affine, ReLU and Cosine functions performs equally well for inference on any particular one of those tasks despite not sharing any common structure other than lipshitzness (Right), meanwhile, pretraining on tasks with a different lipshitzness leads to poor test performance. While this is one important aspect, we agree that there could be other aspects of the data and training that contribute to ICL. For instance, in Section 4 we show that attention can adapt to a ``low rank" structure, essentially picking out specific directions along which to consider distance. --- Rebuttal Comment 1.1: Comment: Thank you for clarification. I have raised the score accordingly.	Rebuttal 1: Rebuttal: We would like to thank all reviewers for their thoughtful and detailed feedback. Some reviewers raised common concerns and questions; we will address these next here. 1. Restricted function class (Reviewers BJxR, ftLf). We expect that our results will hold for quite general function classes, in particular, those that satisfy Assumption B.4. To instantiate specific bounds in the paper we work with specific function classes, which nevertheless comprises the main technical difficulty. Specifically, we have Lemmas C.3 and C.7 that show that our classes satisfy Assumption B.4. The upper bounds are shown in Lemmas C.5 and C.9. It is necessary that the bias of the estimator should increase with an increase in the size of the attention window (to counteract the decrease in variance), and we establish this for these two function classes specifically in Lemmas C.5 and C.9. The lemmas in Appendix C relate the correlation between the function values of a sample of neighbouring points and the bias of the estimator that is built from such a random sample of neighbours, and to derive these correlation we need to work with a specific class. 2. Interpreting ``closeness" (Reviewers BJxR, ftLf). Our paper is a study of what a single layer of attention does; a formal exploration of the notion of attending to "close" points similar to a Nadaraya-Watson (NW) kernel (Nadaraya 1964) (which is a consequence of the particular model we study, but also an intuition that many practitioners carry). We hypothesize that with more layers the hidden token embeddings at intermediate layers reflect some learned metric that this type of estimator can exploit for a notion of closeness beyond closeness of the input tokens. The groundwork for this is laid in Section 4 which shows that in the presence of a low rank structure, the kernel projects out the invariant components of the tokens. Suppose for instance that there are two layers. Then the first layer could learn to project onto the relevant subspace by placing all of its attention on tokens in that subspace, and for the second layer, distances are computed only within this subspace. 3. How can our results be extended to more realistic settings (Reviewer hHKc) and/or settings that capture emergent properties of transformers as they scale (Reviewer BJxR)? (cc. Reviewer 3PLG) The regression model for in-context learning that we consider has been studied extensively (Garg et al. 2022, Akyurek et al. 2023, Ahn et al. 2023, Zhang et al. 2023a, von Oswald et al. 2023a, Raventos et al. 2023, Fu et al. 2023, etc.) as a setting that demonstrates the in-context learning capabilities of transformer models. Our work brings this framework closer to practice due to its consideration of softmax-activated attention and nonlinear and low-dimensional target functions. Below are possible avenues to extend our results closer to practice (though out of scope of the current paper). * Classification. The nature of the attention estimate will not change based on the loss function, so it will again be a nearest neighbors estimate in the classification setting. Thus, we expect that the optimal attention weights will again scale with the label noise level and inversely with the Lipschitzness of the labels, for an analogous notion of Lipschitzness defined for the classification setting (e.g. inverse of minimum KL divergence between cluster distributions, or using Fourier coefficients for functions on the hypercube). Technical details need to be resolved regarding analyzing the cross entropy loss, but these should not be prohibitive. * More layers. Additional layers introduce the following difficulty: the covariates themselves change across layers, so a deeper transformer does not simply implement a kernel estimator with a different kernel. We believe the post-attention multi-layer perceptron (MLP) block and/or layer normalization has to be incorporated into the model and analysis to prevent the tokens from collapsing to a single point similar to what happens in a power iteration (a version of this phenomenon is noted in (Geshkovski et al. 2024)). With that being said, it is conceivable that the first $l-1$ layers of an $l$-layer transformer can be interpreted as mapping the input tokens into an appropriate space in which to do nearest-neighbors regression in the $l$-th layer. This is an interesting direction to explore in future work. While plausible, these extensions present significant technical challenges that we believe can be topics for separate works. Regarding infinite width (suggested by Reviewer hHKc), increasing the inner dimension of $\mathbf{M}_K^\top \mathbf{M}_Q$ will not change the optimal value of $\mathbf{M}=\mathbf{M}_K^\top \mathbf{M}_Q$, so our results will not change. References Geshkovski, Borjan, et al. "The emergence of clusters in self-attention dynamics." Advances in Neural Information Processing Systems 36 (2024). Pdf: /pdf/3f768889178bf29fa7f21bad9fa713d32b678bb2.pdf	NeurIPS_2024_submissions_huggingface	2,024	null	null	null	null	null	null	null	null
(FL)$^2$: Overcoming Few Labels in Federated Semi-Supervised Learning	Accept (poster)	Summary: This work addresses a very practical challenge against successful FL deployments, which is of unlabeled data at FL clients. Furthermore, the problem is set in the regime of low count of labeled samples at server. The proposed solution to train a model in semi-supervised manner includes having an adaptive confidence threshold for each client to get pseudo-labels of more samples in the initial stages of training, then updating the model by perturbing the weights and training them on high-confidence pseudo labels; and lastly by aggregating the model weights through a learning status aware hyperparameter. The results show superiority of the proposed method against existing state-of-the-art methods for federated self-supervised learning, and shows why naive application of centralized self-supervised methods are not cut for the disjoint nature of server and clients in FL. Strengths: 1. The paper is written very well. The flow of logic is mostly clear. 2. The issue this paper is tackling is very important (especially low labeled sample count at the server), and the solution proposed is elegant (clearly states why and how the existing centralized semi-supervised methods are not enough, which brings a unique solution for the FL setting). 3. Strong results with comprehensive experiments. Weaknesses: 1. As expanded in Questions, some parts of the methodology is unclear. As an example, why do we need Eq 10 (unsupervised training objective) at the stage of adaptive thresholding? I thought we are just getting the confidence threshold for pseudo-labels and sending it to the server. Another example where the methodology was slightly fuzzy is around line 193, "While we use client-specific adaptive threshold, we use a high fixed threshold to get high-confidence data samples". What if that fixed threshold is not set correctly? The goal of dynamic client threshold was to avoid the sub-optimal results of a fixed threshold, wouldn't the same issues arise for this adversarial perturbations? 2. The authors have tried two datasets: CIFAR10 and SVHN. I wonder how this translates to harder classification problems with more classes (where a few classes might not have representative samples at server and many clients just do not have samples related to those classes at all) or predicting classes or next words on natural language datasets. Technical Quality: 4 Clarity: 4 Questions for Authors: 1. In Line 68, "Clients with lower learning status (e.g., whose models are less certain about their predictions) receive higher aggregation weights, ensuring their updates are more significantly reflected in the global model." It's unclear why do we need the above? Shouldn't low confidence show bad generalizability of the trained model? If yes, I didn't catch how the authors are preventing the model from learning on "wrong" input-output pairs. 2. Related to Question #1, in what cases would the learning status be low? And what was the intuition behind giving high importance to those low status clients during the aggregation? What if this leads to some other clients getting low learning status then? 3. Why does $(FL)^2$ need strongly augmented samples for SACR? What happens if we just use weakly-augmented samples instead? 4. A minor suggestion: Figure 2 can benefit from numbers to show the flow of what happens after what. It was difficult to figure out where to start reading the diagram from. Confidence: 4 Soundness: 4 Presentation: 4 Contribution: 3 Limitations: I do not see a limitations section. Although the authors mention that the limitations are mentioned in Conclusion, I am not sure what they are referring to (not having theoretical analysis does not sound like a limitation of the proposed method, more like potential future work). Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We greatly appreciate your thoughtful comments and feedback. In our revised manuscript, we intend to address these points as follows: ### Clarification of methodology About Eq10 - At each communication round, the selected clients individually calculate their adaptive thresholds based on their own unlabeled data using Eq 9. Once the adaptive threshold is determined, each client trains its local model according to Eq 10, which incorporates pseudo-labeling and consistency regularization using strongly augmented samples. About fixed threshold in SACR - In Section 5.3, the ablation study reveals that applying an adaptive threshold for SACR can actually degrade performance, as seen in the CAT+SACR (All data) scenario. This occurs because applying SACR to wrongly pseudo-labeled samples leads to the generalization of erroneous samples. To mitigate this, we opted for a high fixed threshold to ensure that only high-confidence data samples, which are more likely to be correct, are utilized. Since we already employ an adaptive thresholding scheme (CAT) to address the limitations of a fixed threshold, we believe it is safe to use a fixed threshold specifically for SACR. ### Motivation of LSSA method (Related to questions 1, 2) - In a centralized SSL, methods like FlexMatch and FreeMatch introduced the use of different thresholds for each class within a dataset and showed their effectiveness. The rationale is that different classes pose varying levels of learning difficulty, so lower thresholds are assigned to more challenging classes. - In the context of FSSL, the learning difficulty can vary across clients. This variation arises for two main reasons. First, since the server has access to only a small labeled dataset, clients whose data closely resembles the server’s data will face lower learning difficulty, while those with more distinct data will encounter higher difficulty. Second, due to the non-iid distribution of data across clients, the learning difficulty naturally differs among them. - We propose LSAA to take account of different learning difficulties of clients, where LSAA assigns higher aggregation weights to clients with higher learning difficulty, enabling the global model to learn more effectively from these clients. Figure 1 in the global response PDF highlights the effectiveness of LSAA. Not only does it achieve higher test accuracy compared to the fixed aggregation weights (CAT + SACR), but it also demonstrates higher pseudo-label accuracy. Additionally, LSAA consistently achieves the highest correct label ratio, indicating the percentage of correct pseudo labels among all unlabeled data, throughout the experiment. After 600 training rounds, LSAA also records the lowest wrong label ratio, representing the percentage of incorrect pseudo labels among all unlabeled data. These findings suggest that LSAA effectively reduces incorrect pseudo labels while increasing correct ones, thereby mitigating confirmation bias [1]. ### Questions Why does $(FL)^2$ need strongly augmented samples for SACR? What happens if we just use weakly-augmented samples instead? - Thank you for your question regarding SACR. Building on previous work in centralized SSL and SemiFL, our method trains the client's local model using pseudo-labeling combined with consistency regularization (as detailed in Eq 10). In this approach, the pseudo-label, which is generated from the confidence of weakly augmented data, guides the training of its strongly augmented counterpart. This ensures the model produces consistent predictions across various perturbations of the same unlabeled data. Following this scheme, we aim to provide additional consistency regularization with SACR , thus we used strongly augmented samples to calculate the loss. A minor suggestion: Figure 2 can benefit from numbers to show the flow of what happens after what. It was difficult to figure out where to start reading the diagram from. - Thank you for your valuable suggestion to enhance the readability of our manuscript. We will make sure to update Figure 2 accordingly in our camera-ready version ### Limitations of $(FL)^2$: Thank you for the feedback on the limitation. Our limitations are as follows, and we will address them in the camera-ready version. - CAT generates more pseudo-labels compared to fixed threshold methods, which in turn increases the time required for client training. - In SACR, additional computation is required due to the need for inference on unlabeled samples using perturbed models. - Our methodology relies on strong data augmentation, which may not be feasible for datasets such as sensor data. [1] Arazo, Eric, et al. "Pseudo-labeling and confirmation bias in deep semi-supervised learning." 2020 International joint conference on neural networks (IJCNN). IEEE, 2020. --- Rebuttal Comment 1.1: Comment: Thank you for your detailed answers. I would like to maintain my score. --- Reply to Comment 1.1.1: Title: Thank you for your response! Comment: Thank you so much for your valuable feedbacks and opinions on our work. We will further strengthen our final manuscript based on your suggestion. If you have any further concern, please do not hesitate to leave a comment for us!	Summary: This paper focuses on the federated semi-supervised learning (FSSL) scenario, which is a more challenging problem in FL. There are two different scenarios in FL, labels-at-server and labels-at-clients and this paper tackles the former issues. The author found the gap between SSL and FSSL is confirmation bias. To diminish this bias, this paper proposes client-specific adaptive threshold, modified SAM, and learning status-aware aggregation for a new aggregation scheme. The experiments of $(FL)^2$ demonstrate improvement on two benchmark datasets. Strengths: This paper is well-organized and easy to follow with clear contribution. Weaknesses: - Lack of motivation and insights. It is not clear how the performance is influenced by confirmation bias Moreover, while the technologies discussed in this paper are not new, the argument that their particular combination can mitigate confirmation bias lacks persuasiveness. - The related work about labels-at-clients is outdated. Some new works should be included, e.g., [R1][R2]. - The experiments of this paper are insufficient. E.g., How $(FL)^2$ tackles the confirmation bias is not mentioned. This paper should add more ablation studies to prove the bias is decreased rather than only the performance. There are many reasons to get the improvement. Reference:\ [R1] Li, Ming, et al. Class balanced adaptive pseudo labeling for federated semi-supervised learning. In CVPR, 2023.\ [R2] Zhang, Yonggang, et al. Robust Training of Federated Models with Extremely Label Deficiency. In ICLR, 2024. Technical Quality: 2 Clarity: 3 Questions for Authors: - I go through the source code and find that implementing FedMatch may lack some components, e.g., parameter decomposition for disjoint learning and K-Dimensional Tree for helper selection. So the incomplete baseline is not convincing to compare with the proposed methods. - Why the performance of SemiFL under SVHN dataset with Balanced IID, 250 labeled samples is lower than 40 samples which is counterintuitive. - The performance of SemiFL under CIFAR10 dataset with Unbalanced Non-IID, and Balanced IID is 10.0. This seems to be a training problem. Confidence: 4 Soundness: 2 Presentation: 3 Contribution: 2 Limitations: - The dataset is limited, there are also CIFAR-100 and FMNIST which are most used in FL research. - The limitations of this work are not well-discussed. Flag For Ethics Review: ['No ethics review needed.'] Rating: 5 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We greatly appreciate your thoughtful comments and feedback. In our revised manuscript, we intend to address these points as follows: ### Effect of $(FL)^2$ on confirmation bias Thank you for your feedback. Since the wrong pseudo-labels usually lead to confirmation bias[1], we evaluated pseudo-label accuracy, label ratio, correct label ratio, wrong label ratio and C/W ratio in addition to test accuracy. We compared $(FL)^2$ against baseline methods using the SVHN dataset with 40 labels in a balanced IID setting, as illustrated in Figure 1 of the global response PDF. A high pseudo-label accuracy indicates that the method produces reliable pseudo labels. A high correct label ratio suggests that the method supplies the model with a higher number of accurate labels. Conversely, a low wrong label ratio indicates that the model encounters fewer incorrect labels, which is crucial for minimizing confirmation bias [1]. Lastly, a high C/W ratio signifies that the model is exposed to more correct labels than incorrect ones, further helping to reduce confirmation bias. We observed that $(FL)^2$ consistently outperforms the baseline SemiFL across all metrics. Notably, while SemiFL generates more incorrect labels (C/W ratio < 1), $(FL)^2$ produces twice as many correct labels compared to incorrect ones. Additionally, the wrong label ratio for $(FL)^2$ is approximately 30%, significantly lower than SemiFL’s 45%. These results suggest that $(FL)^2$ effectively reduces the number of incorrect pseudo-labels while increasing the number of correct ones, thereby mitigating confirmation bias. Furthermore, we can observe the effectiveness of each component of $(FL)^2$, which are CAT, SACR, LSAA. Using CAT and SACR alone delivers better performance compared to the baseline in terms of all metrics. If we use CAT + SACR, pseudo label accuracy increases, correct label ratio increases, and wrong label ratio decreases, which means we can reduce the confirmation bias. When LSAA is added, which is $(FL)^2$, it achieves the best performance across all metrics. This suggests that the synergistic effect of CAT, SACR, and LSAA is reducing confirmation bias effectively. ### Novelty of $(FL)^2$ We appreciate your question regarding the technical novelty of $(FL)^2$. We would like to clarify $(FL)^2$’s novelty as follows: Use of client-specific threshold (CAT) unlike FreeMatch - $(FL)^2$ can deal with non-iid settings by calculating specific adaptive thresholds for each client, while FreeMatch calculates a single global learning status. Because clients’ data distributions vary in non-iid settings, each client’s learning status can be different, making it necessary to estimate client-specific thresholds. - $(FL)^2$ mitigates the risk of overfitting by evaluating the learning status based on the entire dataset at a fixed point in time. Since the local model repeatedly encounters the same data across multiple local epochs, using a running batch as FreeMatch for estimating learning status might reinforce incorrect labels. Introducing Learning Status-Aware Aggregation (LSAA) - We newly introduced LSAA, which adjusts aggregation weights of client local models. While CAT gives lower thresholds to the clients with low learning status to learn more from abundant unlabeled data, such information is not effectively reflected in the global model because of using fixed model aggregation weights. LSAA gives more weight to the client with low learning status, thus such information can be reflected more in the global model. Utilization of SAM under FSSL - We developed a novel SAM objective specific to FSSL setting. We selectively apply the SAM objective to a small subset of high-confidence pseudo-labeled data, while using an adaptive threshold to incorporate a larger portion of unlabeled data. ### Questions: - Thank you for your thorough review. We have re-implemented the missing components of FedMatch and conducted an evaluation on the updated implementation. The updated performance results, presented in Table 1 of the global response PDF, demonstrate that $(FL)^2$ consistently outperforms FedMatch across all settings. - Thank you for your thorough review of the experimental results. Our observations indicate that SemiFL is sensitive to the labeled dataset when the number of labeled data is small. For instance, in the case of SVHN with 250 labels, two out of three runs failed after training 700 rounds, resulting in a low accuracy of around 20%. However, the one successful run achieved an accuracy of 90.6%. When given 40 labels on SVHN, we could not observe the training failure and the average accuracy was 53.4%. Similarly, when testing CIFAR10 with just 10 labels, all runs of SemiFL failed completely, yielding an accuracy of only 10%. We also attempted to replicate these experiments using the official SemiFL repository with 10 labels and observed the same outcomes. - We truly appreciate your thoughtful suggestions. We will certainly update the related work on labels-at-clients in our future manuscript. ### Limitations of $(FL)^2$: - CAT generates more pseudo-labels compared to fixed threshold methods, which in turn increases the time required for client training. - In SACR, additional computation is required due to the need for inference on unlabeled samples using perturbed models. Our methodology relies on strong data augmentation, which may not be feasible for datasets such as sensor data. - Our methodology relies on strong data augmentation, which may not be feasible for datasets such as sensor data. [1] Arazo, Eric, et al. "Pseudo-labeling and confirmation bias in deep semi-supervised learning." 2020 International joint conference on neural networks (IJCNN). IEEE, 2020. --- Rebuttal Comment 1.1: Comment: Thanks for the efforts and additional experiments. My concerns have been addressed, I would like to raise my score to borderline accept --- Rebuttal 2: Title: Thank you for your response Comment: Thank you for your thoughtful review of our rebuttal. We are pleased that our responses have addressed your concerns. We will further enhance our final manuscript based on your valuable feedback. Thank you once more for your helpful comments and feedback to improve our work. Best, Authors.	Summary: This paper studies the federated semi-supervised learning (FSSL) problem. A significant gap between the centralized semi-supervised learning and FSSL is found due to the confirmation bias. To address this issue, the current paper proposes a new FSSL algorithm, by incorporating three new ideas, namely client-specific adaptive threshold, sharpness-aware consistency regularization, and learning status-aware aggregation. Experimental results show that the proposed method significantly improves the performance of existing FSSL algorithms. Strengths: + This paper is very well-written. The proposed method and the underlying idea are clearly explained. + The idea of using client-specific adaptive thresholds is very neat and well-motivated by mitigating the confirmation bias. + The paper proposes the sharpness-aware consistency regularization to address the issue of generalizing to wrongly labeled data points when using sharpness-aware minimization. + The numerical experiments show that the proposed method outperforms the existing methods by a significant margin consistently. + The paper also numerically assesses the contribution of each component and studies the impact of incorrect pseudo-labels, providing further insight into the success of the proposed method. Weaknesses: + The proposed method is based on heuristic reasoning and lacks theoretical justification. It would be nice if the authors could provide some theoretical justification for some components of the proposed method. + The numerical studies are not that extensive; only two public datasets are used for benchmarking. It would be more convincing if the authors could provide more thorough comparison studies over different datasets across different settings. Technical Quality: 3 Clarity: 4 Questions for Authors: + How does the proposed method perform in Figure 1? Does it perform comparably to the centralized SSL method? Confidence: 4 Soundness: 3 Presentation: 4 Contribution: 3 Limitations: The paper only briefly mentions the lack of theoretical formulation as one limitation of the work. Flag For Ethics Review: ['No ethics review needed.'] Rating: 7 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We greatly appreciate your thoughtful comments and feedback. In our revised manuscript, we intend to address these points as follows: ### Theoretical justification Thank you for your feedback. We added more experiments to show that our method is robust in different settings. We plan to prove our methodology in terms of theoratical formulation in future works. ### More experiments Thank you for suggesting additional experiments on other public datasets and settings. In response, we conducted further experiments using the CIFAR-100, Fashion-MNIST, and AGNews datasets, and also introduced non-iid-0.1 settings for the CIFAR-10 and SVHN datasets. For detailed information and experimental results, please refer to the global response. ### Questions: Thank you for your feedback. Our accuracies are 38.9%, 81.5%, 83.2%, and 92.2% for 10, 40, 250, and 4000 labels, respectively. While the centralized FreeMatch method achieves 91.9% accuracy, our results, though still limited, represent a significant improvement. Specifically, our method outperforms the previous best-performing approach by a factor of 2.3X. This demonstrates that our approach narrows the gap between centralized and federated settings. ### Limitations of $(FL)^2$: Thank you for the feedback on the limitation. Our limitations are as follows, and we will address them in the camera-ready version. - CAT generates more pseudo-labels compared to fixed threshold methods, which in turn increases the time required for client training. - In SACR, additional computation is required due to the need for inference on unlabeled samples using perturbed models. - Our methodology relies on strong data augmentation, which may not be feasible for datasets such as sensor data. --- Rebuttal Comment 1.1: Title: Response to Author Rebuttal Comment: Thank you very much for your detailed responses. Including theoretical justification would strengthen the paper. I would maintain my score. --- Reply to Comment 1.1.1: Title: Thank you for your response! Comment: Thank you again for the valuable suggestions and comments. If you have any remaining concerns, please let us know!	Summary: The paper proposes a new method for federated semi-supervised learning tasks where only the server has a small amount of labeled data. The paper combines 3 different methods to tackle the problem and claims to reduce the confirmation bias issue with the method proposed. Strengths: The paper is well-written and gives established motivations. Weaknesses: - The method seems to be just a combination of FreeMatch and FlatMatch under the FL case. - It is better to give more motivation to the LSSA method - Only on 2 datasets while most of the existing SSL /FSSL methods use three datasets. Based on a lot of existing semi-supervised learning literature, it is quite common to at least also include the CIFAR100 dataset. It will be better to also incorporate other type of dataset (not only image) to show the robustness of the method - Authors claim that the new method can effectively reduce the confirmation bias, so it would be better to have experiments specially design to show this part. It would be better to show that the method can successfully label the hard data compared to the baseline. Technical Quality: 3 Clarity: 3 Questions for Authors: - Section 5.3: I find it hard to understand why to compare with ‘correctly pseudo-labeled data’. First, under the semi-supervised case, we won’t know the label, hence there is no point to compare test only on the correct pseudo-label data. This can only induce bias. Second, if we compare with CAT and CAT+SACR(all data), CAT only performs consistently better. This negate the idea of SACR, isn’t it? Confidence: 4 Soundness: 3 Presentation: 3 Contribution: 2 Limitations: Same as above Flag For Ethics Review: ['No ethics review needed.'] Rating: 4 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We greatly appreciate your thoughtful comments and feedback. In our revised manuscript, we intend to address these points as follows: ### Novelty of $(FL)^2$ We appreciate your question regarding the difference between $(FL)^2$ and related works (FreeMatch and FlatMatch). We would like to clarify $(FL)^2$’s novelty as follows: Different motivation with FlatMatch - The primary goal of FlatMatch is to bridge the gap between labeled and unlabeled data by addressing the differences in their loss landscapes. However, since $(FL)^2$ is designed for the labels-at-server scenario, we cannot utilize both unlabeled and labeled data at the same moment, FlatMatch’s methodology cannot be applied. Use of client-specific threshold (CAT) unlike FreeMatch - $(FL)^2$ can deal with non-iid settings by calculating specific adaptive thresholds for each client, while FreeMatch calculates a single global learning status. Because clients’ data distributions vary in non-iid settings, each client’s learning status can be different, making it necessary to estimate client-specific thresholds. - $(FL)^2$ mitigates the risk of overfitting by evaluating the learning status based on the entire dataset at a fixed point in time. Since the local model repeatedly encounters the same data across multiple local epochs, using a running batch as FreeMatch for estimating learning status might reinforce incorrect labels. Introducing Learning Status-Aware Aggregation (LSAA) - We newly introduced LSAA, which adjusts aggregation weights of client local models. Utilization of SAM under FSSL - We developed a novel SAM objective specific to FSSL setting. We selectively apply the SAM objective to a small subset of high-confidence pseudo-labeled data, while using an adaptive threshold to incorporate a larger portion of unlabeled data. ### Motivation of LSSA method - The learning difficulty can vary across clients. First, since the server has access to only a small labeled dataset, clients whose data closely resembles the server’s data will face lower learning difficulty, while those with more distinct data will encounter higher difficulty. Second, due to the non-iid distribution of data across clients. - We propose LSAA where LSAA assigns higher aggregation weights to clients with higher learning difficulty, enabling the global model to learn more effectively from these clients. Figure 1 in the global response PDF highlights the effectiveness of LSAA. Not only does it achieve higher test accuracy compared to the fixed aggregation weights (CAT + SACR), but it also demonstrates higher pseudo-label accuracy. Additionally, LSAA consistently achieves the highest correct label ratio, indicating the percentage of correct pseudo labels among all unlabeled data, throughout the experiment. After 600 training rounds, LSAA also records the lowest wrong label ratio, representing the percentage of incorrect pseudo labels among all unlabeled data. These findings suggest that LSAA effectively reduces incorrect pseudo labels while increasing correct ones, thereby mitigating confirmation bias [1]. ### Effect of $(FL)^2$ on confirmation bias Thank you for your feedback. Since the wrong pseudo-labels usually lead to confirmation bias [1], we evaluated pseudo-label accuracy, label ratio, correct label ratio, wrong label ratio and C/W ratio in addition to test accuracy. We compared $(FL)^2$ against baseline methods using the SVHN dataset with 40 labels in a balanced IID setting, as illustrated in Figure 1 of the global response PDF. A high pseudo-label accuracy indicates that the method produces reliable pseudo labels. A high correct label ratio suggests that the method supplies the model with a higher number of accurate labels. Conversely, a low wrong label ratio indicates that the model encounters fewer incorrect labels, which is crucial for minimizing confirmation bias [1]. Lastly, a high C/W ratio signifies that the model is exposed to more correct labels than incorrect ones, further helping to reduce confirmation bias. We observed that $(FL)^2$ consistently outperforms the baseline SemiFL across all metrics. Notably, while SemiFL generates more incorrect labels (C/W ratio < 1), $(FL)^2$ produces twice as many correct labels compared to incorrect ones. Additionally, the wrong label ratio for $(FL)^2$ is approximately 30%, significantly lower than SemiFL’s 45%. These results suggest that $(FL)^2$ effectively reduces the number of incorrect pseudo-labels while increasing the number of correct ones, thereby mitigating confirmation bias. Furthermore, we can observe the effectiveness of each component of $(FL)^2$, which are CAT, SACR, LSAA. Using CAT and SACR alone delivers better performance compared to the baseline in terms of all metrics. If we use CAT + SACR, pseudo label accuracy increases, correct label ratio increases, and wrong label ratio decreases, which means we can reduce the confirmation bias. When LSAA is added, which is $(FL)^2$, it achieves the best performance across all metrics. This suggests that the synergistic effect of CAT, SACR, and LSAA is reducing confirmation bias effectively. ### Questions: We want to clarify terms used in the manuscript. - CAT + SACR in Table 3 (Section 5.2): Our proposed method in Section 4.2. We apply SACR to only high-confidence samples. - CAT + SACR (all data) in Figure 3 (Section 5.3): SACR is applied to all of the pseudo labels generated by CAT. This is an ablation study to show why SACR should not be applied to all of the pseudo labels by CAT. - CAT + SACR (only correct pseudo-label) in Figure 3 (Section 5.3): SACR is only applied to correctly pseudo-labeled samples. This is for testing upper bound of our proposed method assuming we know the ground truth labels. [1] Arazo, Eric, et al. "Pseudo-labeling and confirmation bias in deep semi-supervised learning." 2020 International joint conference on neural networks (IJCNN). IEEE, 2020. --- Rebuttal Comment 1.1: Comment: Thanks for more experiments and results. I still have some concerns regarding innovation and motivations. Also, I'm not entirely convinced of the authors' answer regarding the confirmation bias. The authors mentioned about the pseudo label accuracy and correct label ratio, which seems to be the same thing (the accuracy of pseudo labeling). However, the definition of the confirmation bias proposed by the reference is about whether or not the model is overfitting to the wrong pseudo-labels. Therefore, how the proposed methods can mitigate the confirmation bias still lacks persuasiveness. Therefore, I would maintain my score for now. --- Rebuttal 2: Title: Thank you for your response Comment: Thank you for your response to our rebuttal. We sincerely appreciate your effort and time in going through our rebuttals and raising your concern. We want to be more explicit in addressing your remaining concerns as follows: ### Regarding the confirmation bias As highlighted in previous work on Semi-Supervised Learning and others that mentioned confirmation bias [2, 3, 4, 5, 6], one approach to reducing confirmation bias is to filter out noisy pseudo-labels and retain only the high-quality ones. For instance, Fixmatch [2] claimed that it effectively reduced confirmation bias by generating higher-quality pseudo-labels. Similarly, Mean Teacher [3] suggested that improving target quality can help mitigate confirmation bias. Additionally, the Softmatch [4] paper highlighted that incorrect pseudo-labels often contribute to the occurrence of confirmation bias. Based on these previous works, we believe that demonstrating a higher correct label ratio effectively shows how well our method filters out incorrect pseudo-labeled samples. Consequently, since $(FL)^2$ has a significantly lower rate of incorrect pseudo-labels, the influence of these on the unsupervised loss is much less pronounced compared to SemiFL. This suggests that the training signals propagated by these pseudo-labels are more reliable in $(FL)^2$ due to lower contribution of noisy pseudo-labels - our unsupervised loss is dominated mainly by correct pseudo-labeled samples compared to SemiFL. Nevertheless, to further evaluate this, we designed and conducted an experiment to compare the level of overfitting of $(FL)^2$ and SemiFL to wrongly pseudo-labeled data in a scenario with extremely scarce labeled data. We measured this by inspecting Rsoft, RLoss, and RAccuracy. Rsoft is the average confidence of the wrongly pseudo-labeled data, similar metric measured in [1]. RLoss and RAccuracy means the training loss and accuracy on the wrongly pseudo-labeled data, respectively. High Rsoft and RAccuracy, and low Rloss indicate that the model is overfitting to the incorrectly pseudo-labeled data. The experiment used the CIFAR-10 dataset under a balanced IID setting with only 10 labeled samples at the server, keeping all hyperparameters the same as in the main experiments, except for the number of communication rounds, which we set to 500. \|Method\|Label Ratio (%) \|Rsoft (After softmax) \|RLoss\|RAccuracy (%) \| \|:---:\|:---:\|:---:\|:---:\|:---:\| \|SemiFL\|100\|99.85\|0.0015\|100\| \|$(FL)^2$\|86.86\|84.53\|0.4008\|87.09\| The results clearly show that SemiFL overfits significantly to the wrongly pseudo-labeled data compared to $(FL)^2$. This demonstrates $(FL)^2$’s robustness against incorrect pseudo-labels, particularly in scenarios with extremely scarce labeled data at the server, outperforming SemiFL. Thus, we conclude that $(FL)^2$ is more effective in reducing confirmation bias in low-label scenarios. Thank you again for your insightful suggestion to explore overfitting with incorrectly pseudo-labeled data. We will ensure that the results of this experiment are included in the camera-ready version. - [1] Arazo, Eric, et al. "Pseudo-labeling and confirmation bias in deep semi-supervised learning." 2020 International joint conference on neural networks (IJCNN). IEEE, 2020. - [2] Sohn, Kihyuk, et al. "Fixmatch: Simplifying semi-supervised learning with consistency and confidence." Advances in neural information processing systems 33 (2020): 596-608. - [3] Tarvainen, Antti, and Harri Valpola. "Mean teachers are better role models: Weight-averaged consistency targets improve semi-supervised deep learning results." Advances in neural information processing systems 30 (2017). - [4] Chen, Hao, et al. "SoftMatch: Addressing the Quantity-Quality Tradeoff in Semi-supervised Learning." The Eleventh International Conference on Learning Representations. - [5] Nassar, Islam, et al. "All labels are not created equal: Enhancing semi-supervision via label grouping and co-training." Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2021. - [6] Zhang, Bowen, et al. "Flexmatch: Boosting semi-supervised learning with curriculum pseudo labeling." Advances in Neural Information Processing Systems 34 (2021): 18408-18419. --- Rebuttal 3: Title: Thank you for your response Comment: Thank you for your comment regarding motivation and innovation. We would like to explain further about $(FL)^2$'s motivation and novelty. ### Novelty of $(FL)^2$ We would like to further explain the novelty of $(FL)^2$ in detail. Use of client-specific threshold (CAT) unlike FreeMatch - $(FL)^2$ provides a precise measure of the client's learning status rather than relying on an EMA-based estimation like FreeMatch, since $(FL)^2$ directly calculates learning status using all unlabeled data. - $(FL)^2$ is computationally efficient. Instead of continuously updating the learning status every batch like FreeMatch, $(FL)^2$ only requires one interaction of learning status calculation over unlabeled samples per communication round. Additionally, since we utilize the global pseudo-labeling scheme proposed in SemiFL, we must go through all the unlabeled data anyway to pseudo-label them. Additional computation for CAT fits naturally into the existing workflow. Utilization of SAM under FSSL - We developed a novel SAM objective specific to FSSL setting. We first applied the Sharpness-Aware Minimization (SAM) objective in a FSSL setting and discovered that applying it to all pseudo-labeled data degrades performance even though SAM shows strong generalization ability across different tasks. Instead, we introduced a novel approach that selectively applies the SAM objective to a small subset of high-confidence pseudo-labeled data, while using an adaptive threshold to incorporate a larger portion of unlabeled data. Our ablation study (Table 2 of the paper), which emphasizes the importance of each component, demonstrates that this combination of techniques effectively reduces confirmation bias and achieves high performance. ### Motivation of LSAA We would like to further clarify the motivation of LSAA. - In a centralized SSL, methods like FlexMatch and FreeMatch introduced the use of different thresholds for each class within a dataset and showed their effectiveness. The rationale is that different classes pose varying levels of learning difficulty, so lower thresholds are assigned to more challenging classes—those with a lower learning status—to facilitate more effective learning from these harder classes. - In the context of FSSL, the learning difficulty can vary across clients. This variation arises for two main reasons. First, since the server has access to only a small labeled dataset, clients whose data closely resembles the server’s data will face lower learning difficulty, while those with more distinct data will encounter higher difficulty. Second, due to the non-iid distribution of data across clients, the learning difficulty naturally differs among them. - We propose LSAA to take account of different learning difficulties of clients, where LSAA - assigns higher aggregation weights to clients with higher learning difficulty, enabling the global model to learn more effectively from these clients. In contrast, previous FSSL approaches did not account for these variations in learning difficulty and instead relied on fixed aggregation weights. --- Rebuttal 4: Comment: Thank you again for your critical and constructive feedbacks to our papers and rebuttals. If there is still anything that is not clear about our justification regarding the confirmation bias and the novelty/motivation of our work, we would be happy to provide more clarification. If you have any further concerns or questions, please do not hesitate to let us know as it will be extremely important for our revised manuscript. Thank you very much!	Rebuttal 1: Rebuttal: ## More experiments We conducted additional experiments using the CIFAR-100, Fashion-MNIST, and AGNews datasets, and introduced a non-iid-0.1 setting for the SVHN and CIFAR-10 datasets. Additionally, we performed an ablation study to determine the optimal rho value for Adaptive Sharpness-Aware Minimization (ASAM [1]). Through grid search, we identified rho=0.1 as best performing for both the SVHN and CIFAR-10 datasets, leading to revisions in the main table (Table 1 in the global response PDF). Thanks to reviewer YhMN's feedback, we discovered and addressed missing parts in the FedMatch implementation, re-implementing it and re-conducted the experiments. The results for CIFAR-100 are presented in Table 1, while the results for Fashion-MNIST and AGNews are detailed in Table 2, both of which can be found in the PDF file of the global response. CIFAR-100 For CIFAR-100, we used WideResNet-28x8 as in previous works [2,3]. $(FL)^2$ consistently outperforms the baselines across most configurations. Notably, in the 400-label, non-iid-0.3 setting, $(FL)^2$ achieved a 6.4% improvement over the baselines. We will add experiments on a non-iid-0.1 setting in our camera-ready version. Fashion-MNIST For Fashion-MNIST, we used WideResNet-28x2 as in SVHN and CIFAR-10 datasets. We compared our method with SemiFL, the previous state-of-the-art. Because of the time constraint, we only compare with SemiFL, but we will add other baselines for the camera-ready version. With only 40 labeled samples, SemiFL failed in 3 out of 3 runs in iid setting and 2 out of 3 runs in non-iid-0.3 setting, resulting in ~10% accuracy. In one unfailed run in a non-iid-0.3 setting, SemiFL achieved 18.4% accuracy. In contrast, $(FL)^2$ successfully trains all 3 runs in non-iid-0.3 setting and 2 out of 3 runs in iid setting. In one failed run in iid setting, it yields ~10% accuracy. However, in the remaining iid runs, $(FL)^2$ achieved accuracies of 69% and 70.4%. In the non-iid-0.3 setting, $(FL)^2$ achieved an average accuracy of 63.2%. The results demonsrate that $(FL)^2$ is robust and effective even with minimal labeled data, particularly in a few-labels setting. AGNews We randomly sampled 12,500 training samples per class (out of total 50,000 samples) and applied back-translation for strong augmentation, following the methodology outlined in the SoftMatch [4] paper. We used bert-base-uncased as a backbone model. We froze the bert parameters and trained only the linear classifier's parameters, running the training for 20 epochs. Since the mixup loss cannot be applied to an NLP dataset, we conducted our comparison using SemiFL without the mixup loss. $(FL)^2$ significantly outperforms the baseline, achieving a 39.6% accuracy improvement in the iid setting and a 14.5% improvement in the non-iid setting. With only 20 labels provided, SemiFL exhibited considerable performance variance across experiments, where STD is 14.3 and 13.7 for iid and non-iid-0.3 setting, respectively. In contrast, $(FL)^2$ consistently delivered stable results, where STD is 0.6 for iid and 3.7 for non-iid-0.3. More experiments on previously used datasets (CIFAR-10 and SVHN) We also introduced a new non-IID-0.1 setting to our previous experiments on the CIFAR-10 and SVHN datasets. In this non-IID-0.1 setting, $(FL)^2$ consistently outperformed all baseline methods across the datasets, showing 10% higher accuracy over the best-performing baseline on CIFAR-10 with 40 labels. [1] Kwon, Jungmin, et al. "Asam: Adaptive sharpness-aware minimization for scale-invariant learning of deep neural networks." International Conference on Machine Learning. PMLR, 2021. [2] Wang, Yidong, et al. "FreeMatch: Self-adaptive Thresholding for Semi-supervised Learning." The Eleventh International Conference on Learning Representations. 2022. [3] Diao, Enmao, Jie Ding, and Vahid Tarokh. "Semifl: Semi-supervised federated learning for unlabeled clients with alternate training." Advances in Neural Information Processing Systems 35 (2022): 17871-17884. [4] Chen, Hao, et al. "SoftMatch: Addressing the Quantity-Quality Tradeoff in Semi-supervised Learning." The Eleventh International Conference on Learning Representations. Pdf: /pdf/37baec4787b37ce85640d253574cf4d33fe20858.pdf	NeurIPS_2024_submissions_huggingface	2,024	null	null	null	null	null	null	null	null
Shuffling Gradient-Based Methods for Nonconvex-Concave Minimax Optimization	Accept (poster)	Summary: This work propose shuffling gradient-based methods for nonconvex-linear and nonconvex-strongly convex minimax optimization and obtain complexity results. Strengths: The presentation is clear. Based on my knowledge on both shuffling and minimax optimization, I found that the algorithms and complexity results are reasonable. The experiments provide sufficient details for reproducibility. Weaknesses: The contribution is in general combinational, with slightly more techniques such as smoothing and shuffling-based function evaluations. Also, comparison with previous stochastic algorithms for minimax optimization and shuffling for nonconvex optimization is lacking, as shown in question 1 below. Technical Quality: 3 Clarity: 3 Questions for Authors: (1) You could compare your complexity results (Theorem 2) with previous stochastic gradient-based minimax optimization works (e.g. [a]) and shuffling for nonconvex optimization (e.g. [b]). [a] Lin, T., Jin, C., \& Jordan, M. (2020, November). On gradient descent ascent for nonconvex-concave minimax problems. In International Conference on Machine Learning (pp. 6083-6093). PMLR. [b] Nguyen, L. M., Tran-Dinh, Q., Phan, D. T., Nguyen, P. H., \& Van Dijk, M. (2021). A unified convergence analysis for shuffling-type gradient methods. Journal of Machine Learning Research, 22(207), 1-44. (2) At the bottom of page 1, for problem (NL), it looks possible to remove matrix $K$ and use $F_i:\mathbb{R}^p\to \mathbb{R}^q$, since $\langle F_i(w),Ku\rangle=\langle K^{\top}F_i(w),u\rangle$ and we could replace $K^{\top}F_i(w)\in \mathbb{R}^q$ with $F_i(w)\in \mathbb{R}^q$. (3) In Assumption 2, do you mean for any $v\in {\rm dom}(h)$, $\|\|v\|\|_2\le M_h$? (4) In "Condition (6) can be relaxed to the form $\frac{1}{n} \sum_{i=1}^n\|\|\nabla F_i(w)-\nabla F(w)\|\|^2 \leq \sigma_J^2+\Theta_J\|\|\nabla \Phi_0(w)\|\|^2$", should $\Phi_0$ be $F$? (5) After Assumption 5, ``If $f=0$, then $\mathcal{G} _ {\eta}(w)\equiv\nabla\Phi_{\gamma}(w)$''. Should it be $\nabla\Phi_0(w)$ instead of $\nabla\Phi_{\gamma}(w)$, since the definition of $\mathcal{G} _ {\eta}(w)$ in eq. (18) involves $\nabla\Phi_0(w)$ not $\nabla\Phi _ {\gamma}(w)$? Or you might consider defining $\mathcal{G} _ {\eta,\gamma}(w)$ with $\nabla\Phi_0(w)$ replaced by $\nabla\Phi _ {\gamma}(w)$, such that $\epsilon$-stationary point of (3) and that of (10) in Lemma 2 correspond to $\mathcal{G} _ {\eta,0}(w)$ and $\mathcal{G} _ {\eta,\gamma}(w)$ respectively. (6) I found that Eq. (21) (Option 1) needs to save {$F_{\pi^{(t)}(j)(w_0^{(t)})}$}$_{j=1}^n$ to avoid repeated evaluation of these $n$ values. Yes? (7) Line 255: "fix number" $\rightarrow$ "fixed number". (8) At the end of Theorem 2, do you mean to change one of the $\nabla_w\mathcal{H}_i$ into $\nabla_u\mathcal{H}_i$? It seems that the complexity of semi-shuffling is bettter than full-shuffling. Then what's the advantage of full-shuffling? Confidence: 3 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: The checklist mentions the limitation that this work only focuses on nonconvex-linear and nonconvex-strongly convex minimax optimization, and says "We do not yet know if our paper has an immediate broader impact. However, since our problems and our algorithms are sufficiently general, we hope they will create broader impacts." I agree with both of them. Flag For Ethics Review: ['No ethics review needed.'] Rating: 7 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We thank you for your appreciation of our strengths and your positive evaluations. Please see our general responses along with the individual reply to you below. The contribution ... below We thank you for your comments. While we agree that some individual techniques in our paper are not new, we believe that our algorithms and problem setting is new and has not been addressed in prior works. Our shuffling approximation for the "hyper" gradient in the (NL) case is new. To our best knowledge, our methods for (NL) appear to be the first to develop shuffling-type methods for the class of nonconvex-linear minimax problems which has various applications in distributionally robust optimization and other learning scenarios (like risk-averse portfolio, model-agnostic meta learning). For the nonconvex-strongly concave setting (NC), we have two strategies, and our model is more general than that of (Cho & Yun, 2022) due to $f$ and $h$. Our assumptions and methods are also different from (Cho & Yun, 2022). Comparisons of complexity Thank you for your suggestion. We have a detailed comparison in the general response, which included the reference (Lin et. al., 2020): For (Nonconvex-SC/PL) setting ($f=h=0$), the complexity $\mathcal{O}(\sqrt{n} \epsilon^{-3})$ of our Algorithm 2 with a random reshuffling strategy and (Cho & Yun, 2022) is better than the complexity $\mathcal{O}(\epsilon^{-4})$ of non-shuffling iid scheme (Lin et. al., 2020, Theorem 4.5). About the paper (Nguyen et. el. 2021), although the setting is different, we will still compare with that reference in our paper. Remove K Yes, it is possible to remove $K$ as you suggested. Though, we note that $K$ plays an additional role as a linear transformation and matches the dimension between $u$ and $F(w)$ if they are not the same. In Assumption 2 ... Yes. That is correct. In "Condition (6) ...? It is $\nabla \Phi_0$ since our convergence guarantee is eventually on $\nabla \Phi_0$. After Assumption 5 ...respectively. Thank you for your suggestion. We will do as you suggested. I found ... Yes? That is correct. (21) saves the evaluation of $F_{\pi^{(t)}(j)}$, while (22) uses the most updated information of $F_{\pi^{(t)}(j)}$. Line 255 ... It will be fixed. Thank you. At the end ... shuffling? Yes, sorry for the typo, we meant that the complexity of $\nabla_w \mathcal H_i $ and $\nabla_u \mathcal H_i $ is $O ( \sqrt{n} \epsilon^{-3} )$. For the full shuffling, we believe that we can improve the complexity of $\nabla_u \mathcal H_i $ to match with the complexity of the semi-shuffling case, as discussed with reviewer jEDb. This would only be a minor change to the theoretical contributions of our paper. We hope our response answers all of your questions. If you have any additional comments and suggestions, please discuss with us and we are happy to clarify further. --- Rebuttal Comment 1.1: Title: Reviewer QEyv is satisfied with the authors' response and will increase rating to 7 Comment: Reviewer QEyv is satisfied with the authors' response and will increase rating to 7. --- Reply to Comment 1.1.1: Title: Thank you for your support! Comment: Dear Reviewer QEyv, We are glad that you are satisfied with our response and increase the rating score. This is very encouraging! Thank you very much for your support! Best regards, Authors	Summary: This paper focuses on nonconvex-concave, finite-sum (stochastic) minimax problems with possibly nonsmooth regularization. Aiming to find $\epsilon$-stationary points, the paper proposes shuffling-based proximal gradient descent-ascent algorithms and verifies gradient computation complexity upper bounds via a bilevel optimization perspective. The results consider various settings, such as (1) the nonconvex-linear (NL) case with either strongly convex or convex $h$, each requiring $\mathcal{O}(n \epsilon^{-3})$ and $\mathcal{O}(n \epsilon^{-7/2})$ gradient evaluations, respectively, and (2) the nonconvex-strongly-concave (NS) case with smoothness assumptions, where semi-shuffling requires $\mathcal{O}(n \epsilon^{-3})$ while full-shuffling requires (naively) $\mathcal{O}(n \epsilon^{-4})$ gradient evaluations. The paper contains proofs of the convergence theorems and supporting numerical experiments. Strengths: * The theoretical results and the problem-specific proof techniques could be a solid contribution. To the best of my knowledge, this is one of the very few works that discuss shuffling-type algorithms for minimax problems apart from Das et al., 2022 [1] and Cho & Yun, 2023 [2]. I think the results under the relaxed setting, especially the linear, non-strongly-convex case of (NL), and for the bilevel-type shuffling algorithms are quite novel. The paper also includes experimental results that align with the theoretical results. Weaknesses: * Despite that the novelty of this work would be in the fact that the results apply for broader settings than previous literature rather than obtaining faster convergence under similar settings, the paper still should have included a fair, more detailed comparison with previous work and illustrated how the gradient evaluation complexity becomes different from settings with stronger assumptions. The current draft does not contain any meaningful quantitative comparison with previous work. * In particular, I don’t think that the settings in [2] are different from the results for the (NS) setting. Apart from the fact that the objective function does not contain $f$ and $h$, the assumptions used in Theorem 1 of [1] (Assumptions 1-4) and the NS setting seem to be nearly identical, and the goal of finding $\epsilon$-stationary points also seem to be equivalent. (In fact, for the variable $u$ the PŁ assumption is weaker than strong concavity.) In terms of the convergence rates, Theorem 1 of [1] requires only $\mathcal{O} (n \epsilon^{-2} + n^{1/2} \epsilon^{-3})$ gradient evaluations, which is smaller than $\mathcal{O} (n \epsilon^{-3})$. The authors should have at least explained how introducing $f, h$ changes the difficulty of the problem, or given any other reasons why one could prefer the bilevel optimization framework over the simple SGDA algorithm as in [1]. Technical Quality: 3 Clarity: 2 Questions for Authors: - Why is it necessary to use bilevel optimization, i.e., update $n$ steps of $w$ and then $n$ steps of $v$, instead of iteration-wise updates? Also, am I right that the results for the (NS) settings are weaker than that of [2] at least when $f, h = 0$? Is there a particular reason why the weaker convergence rates are inevitable if we consider $f, h$? - Some technical parts of the algorithms are brought from previous work, such as the use of $\Phi_{\gamma}$ inspired by Yang et al., 2020 [3]. Some other parts look new, such as the placement of the proximal steps considering $f$ and $h$ and the variants of algorithms like choosing between (21), (22) in Algorithm 1 and (26), (27) in Algorithm 2. I am curious about the motivations of the new components of the algorithms, especially considering the different types of shuffling techniques I mentioned last—whether these have clear motivations or are more like artifacts of the proof techniques, if there were similar approaches in minimization problems, etc. - Light question: Would it be able to relax the strongly concave case to weaker PŁ-type assumptions, or consider alternating type updates as in [2]? Moreover, is it possible to compare the presented results for minimax problems with nonconvex, nonsmooth minimization as in Cutkosky et al., 2023 [4]? - Light question: Are there examples of when the nonconvex-linear (NL) setting could be of interest in practice? - Minor: Why are there two bibliographies, and why are the contents different? It seems that the one at the end of the appendix has the correct numbers. Confidence: 3 Soundness: 3 Presentation: 2 Contribution: 3 Limitations: There seems to be no potential negative societal impact. Flag For Ethics Review: ['No ethics review needed.'] Rating: 5 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We thank you for your appreciation of our contributions and your detailed reviews. Please see our general responses along with the individual reply to you below. Comparisons of complexity Thank you. As you mentioned above, we found the three references [1] [2] and (Emmanouilidis et. al., 2024) that are most related to our paper. In our general response we compared them with our work in details (we do not repeat due to space limit). In particular ... in [1] We believe that you meant [2] (Cho & Yun, 2022). Thank you very much for this excellent point. We agree with you that the assumptions of Theorem 1 in [2] are similar to our assumptions for (NC) when $f=0$ and $h=0$ (the difference is Assumptions 3 + 4 of [2] and our strong concavity of $\mathcal{H}$). We believe that our analysis can be relaxed with Assumptions 3+4 in [2] by means of Eq. (21) in [2]). Let us explain in detail below. If we set S=1 in (27), then Alg. 2 performs only one shuffling loop to solve $\max_u$ subproblem. Alg. 2 is different from [2] due to epoch-wise, not update-wise as [2], but both have the same cost per epoch. When S is a constant and independent of $\eta$, we would have similar complexity as Theorem 1 of [2] if we use a random shuffling strategy. This was noted at the end of Theorem 4: a $\sqrt{n}$ improvement is achieved if random-reshuffling is used. Therefore, our complexity is $O (\sqrt{n} \epsilon^{-3} )$ as in [2]. In our paper, the complexity of $\nabla_w \mathcal H_i $ could be $O ( \sqrt{n} \epsilon^{-3} )$, but the complexity of $\nabla_u \mathcal H_i $ is worse than than [2] by a factor of $1/\epsilon$. This is due to an overestimate of $S$ in Theorem 7, which depends on $1/\eta$, leading to a $1/\epsilon$ factor worse than [2]. This factor can be removed by using a similar Lyapunov func. $V_{\lambda}$ in [2] and modify our proof using Eq. (21) in [2] and Theorem 1 in [Nguyen et al, 2021] instead of our (56). We will update it in our revision. We believe that adding nonsmooth terms $f$ and $h$ could change the algorithm as one has to decide when to apply the prox operator, every single update or at the end of each shuffling loop. We are not sure how these terms could change the analysis of [2], including their assumptions (like our Ass 5). We also consider the strong-concavity either of $\mathcal{H}$ or of $-h$ separately (resp., our Thm 3 or 4). We think that the bilevel opt. approach has more flexibility to tackle the $\max$ and the $\min$ independently, while the SGDA tackles them simultaneously. In general [not only specific to [2], since [2] looks like a mixed approach], the bilevel optimization approach also allows one to explore techniques from standard optimization, while Nash’s game approach (simultaneously solve the min and the max) such as SGDA often requires theory from montone/nonmonotone operators. Apart from the above points, our work also considers the nonconvex-linear case (NL), overlapped with (NC) in Algorithm 1 (a new shuffling rule like (22)), and a semi-shuffling strategy, which could also be useful in practice. Why is ... $f$ and $h$? For the choice of bilevel optimization approach, we have explained above. It is not necessary, but just a different approach. Note that Algorithm 2 only proposed two options (26) and (27), but other options (like accelerated or momentum gradient methods) may be possible to use to solve the strongly concave subproblem $\max_u$. For $f, h=0$, our complexity of $\nabla_w\mathcal{H}_i$ is the same as in Th. 1 of [2], but the complexity of $\nabla_u\mathcal{H}_i$ is currently weaker than Th. 1 of [2]. As we explained above, we can improve the latter one to $O(\sqrt{n}\epsilon^{-3})$. Some ... etc. Yes, we have a clear motivation. For (NL), it is overlapped with (NC), but one is not included in the other. This setting is important as it can be reformulated equivalently to the compositional minimization (CO), and has many applications, including distributionally robust optimization (see Q4 below). The choice of (21) saves evaluations of $F_i$, while (22) uses the most updated information of $F_i$. For (NC), we are flexible to use different methods to solve the strongly concave subproblem $\max_u$. We gave two options: (26) and (27), but other choices can be used. When the maximization is "simple", it may be better to use a semi-shuffling strategy (26), where we can solve the max problem by deterministic methods. Note that one can replace (26) with other deterministic methods, but we need to carry out different analyses. Furthermore, we believe that the strong convexity either in the smooth term $\mathcal{H}$ or in the nonsmooth term $h$ also makes a difference. So, we cannot incorporate $f$ and/or $h$ in the smooth term $\mathcal{H}$ if they are nonsmooth. For the motivation of the algorithm, we aim to design a natural application of shuffling data schemes in solving the stated minimax problems. The analyses of our algorithms follow after we have tried several different proof techniques. Light question: Would ... [4]? As mentioned above, our analysis still works with Assumptions 3+4 in [2] by means of Eq. (21) in [2], at least when $f=0$ and $h=0$. This expression can be written as $L_u \Vert \tilde{u}_t - u^{} ( \tilde{w} _{t-1} ) \Vert^2 \leq 2\kappa_2 [ \Phi_0 ( \tilde{w} _{t-1} ) - \mathcal{H} (\tilde{w} _{t-1}, \tilde u_t ) ] $ in our context. We just need to adapt our proof lightly to accommodate it. Light question: Are ... practice?* Yes, there are so many applications, including distributionally robust optimization with finite discrete probability distribution, risk-averse portfolio, model-agnostic meta learning [Finn et al (2017)], etc. Minor: Why ... numbers? We are sorry about this technical issue. We will fix it. We hope our response answers all of your concerns and questions. If you have any additional comments, please discuss with us and we are happy to clarify further. --- Rebuttal Comment 1.1: Title: Follow up on the rebuttal Comment: Dear Reviewer jEDb, We hope our responses answer all your questions! In case you need any remaining clarifications, we would be more than happy to reply. Please let us know your thoughts as soon as you can (within this discussion period). If your questions are all properly addressed, we really hope that you consider increasing your score to support our work. Regards, Authors	Summary: This paper proposes shuffling gradient-based methods for addressing two classes of minimax optimization problems: nonconvex-linear and nonconvex-strongly concave settings. The first algorithm focuses on the nonconvex-linear minimax setting and the second algorithm, consisting of semi-shuffling and full-shuffling schemes, focuses on the nonconvex-strongly concave minimax setting. The authors establish oracle complexity bounds for these algorithms under some standard assumptions. Numerical experiments are conducted. Strengths: The authors establish oracle complexity bounds for their algorithms under standard assumptions. These bounds provide theoretical guarantees on the performance of the proposed methods. Weaknesses: The proposed Algorithm 1 requires solving a maximization problem at each iteration, which can be computationally expensive. The numerical experiment section is quite weak. First, only Algorithm 1 is tested, with no results related to Algorithm 2. Second, only one minimax model is tested. More complex and popular minimax models should be included, and state-of-the-art competitors should be used to evaluate the performance of the proposed methods. Typo in Line 187 and Line 459: $M_f \rightarrow M_f^2$ Technical Quality: 3 Clarity: 3 Questions for Authors: What are the advantages of the proposed methods in this paper, given the introduction of the semi-shuffling and full-shuffling schemes? Do they have a better convergence rate or sample complexity compared to existing methods? Confidence: 2 Soundness: 3 Presentation: 3 Contribution: 2 Limitations: yes Flag For Ethics Review: ['No ethics review needed.'] Rating: 5 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We thank you for your appreciation of our strengths and soundness. Please see our responses to your concerns and questions below: The proposed ... expansive We can choose a simple $b$ in (9) so that $u^_{\gamma} ( \cdot )$ in Step 9 of Algorithm 1 has a closed-form solution. For instance, if $b(u) = \frac{1}{2}\Vert u - u^0\Vert^2$ for a fixed $u^0$, then computing $u^_{\gamma}(\cdot)$ reduces to computing the proximal operator of $h$ plus a linear term, i.e. $u^_{\gamma}( w ) := \mathrm{prox}_{h/\gamma}(u^0 - K^{\top}F(w)/\gamma)$. For many $h$ (e.g., $\ell_1$-norm, the indicator of simple sets), this proximal operator has a closed form. The numerical ... methods* We appreciate your point. We can add more numerical examples in the Supp. Doc. as you suggested. Apart from this point, we believe that our algorithms and theoretical analysis deserve to be reconsidered since we have 4 theorems corresponding to 5 settings (2 settings for (NL) and 3 ones for (NC)). To our best knowledge, our methods for (NL) appear to be the first to develop shuffling-type methods for the class of nonconvex-linear minimax problems which has various applications in distributionally robust optimization and other learning scenarios (like risk-averse portfolio, model-agnostic meta learning). The shuffling strategy we use in Algorithm 1 is also new. For the nonconvex-strongly concave setting (NC), we have two strategies, and our model is more general than that of [9] due to $f$ and $h$. Our assumptions and methods are also different from [9]. Typo Thank you. We will fix it. What are the advantages In our general response, we showed the complexity comparisons for our paper, which highlight the advantages of our shuffling method and settings compared to other prior work. A comparison between shuffling and SGD was done in (Cho & Yun, 2022), showing improvement of complexity under concrete scenarios. As discussed in the first paragraph of the introduction, the motivation to study shuffling methods is that they are widely implemented in well-established packages like TensorFlow and Pytorch for optimization and deep learning. Empirically, it was observed, see e.g. [(Bottou, 2009, 2012; Hiroyuki, 2018)] that shuffling schemes decrease the loss faster than SGD and work well in practice, and in many cases. Such aspects motivate us to propose shuffling methods for minimax problems. We are not aware of a rigorously empirical study for minimax problems since such methods have not widely been developed yet. However, many researchers have used SGD to train GANs through standard libraries like TensorFlow or Pytorch, we believe that they could use shuffling SGD implemented in these packates. Our paper studies theoretical convergence guarantees for such strategies in minimax algorithms. We hope our response answers all of your concerns and questions. If you have any additional comments and suggestions, please discuss with us and we are happy to clarify further. --- Rebuttal Comment 1.1: Title: Follow up on the rebuttal Comment: Dear Reviewer czC8, We hope our responses answer all your questions! In case you need any remaining clarifications, we would be more than happy to reply. Please let us know your thoughts as soon as you can (within this discussion period). If your questions are all properly addressed, we really hope that you consider increasing your score to support our work. Regards, Authors --- Reply to Comment 1.1.1: Title: Follow up Comment: Dear Reviewer czC8, Since the Author-Reviewer discussion phase will end in two days, we would like to follow up and discuss with you. Please do not hesitate to contact us if there are additional answers or explanations that we can make to clarify our paper within this discussion period. We appreciate your timely response, as it would provide us with an opportunity to address any remaining questions. If your concerns are all properly addressed, we really hope that the reviewer positively re-evaluates our work. We appreciate your inputs and we thank you for your time spent reviewing this paper. Best regards, Authors	Summary: The paper presents new shuffling gradient-based methods for solving two classes of nonconvex-concave minimax optimization problems: nonconvex-linear and nonconvex-strongly concave settings. The first algorithm is designed for the nonconvex-linear setting and achieves state-of-the-art oracle complexity, employing a new shuffling estimator for the hyper-gradient. The second method introduces semi-shuffling and full-shuffling schemes for the nonconvex-strongly concave setting, establishing their oracle complexity bounds for the first time. Numerical examples are provided to demonstrate the performance of the proposed algorithms. Strengths: 1. The paper is well-written, offering a clear exposition of the current state of the art, along with explicit assumptions and conditions for the proposed methods. 2. The authors introduce novel smoothing techniques similar to Nesterov's smoothing approach for lower-level maximization problems, and establish oracle complexity bounds for both nonconvex-linear and nonconvex-strongly concave settings. This introduces a novel perspective on addressing nonconvex optimization problems. 3. The paper presents two distinct algorithms, each carefully tailored to different problem settings, showcasing a profound comprehension of diverse problem landscapes. Weaknesses: 1. Complexity Comparison with Existing Works. The authors are encouraged to present a comparative table that delineates the oracle complexity of their proposed methods alongside that of existing works, including standard nonshuffling methods. This side-by-side comparison should aim to explicitly demonstrate the theoretical advantages of their methods. 2. Shuffling Methods vs. Nonshuffling Methods. The paper should discuss the distinctions between shuffling methods and traditional nonshuffling approaches. Specically, it is better to clarify whether shuffling methods consistently yield superior performance over nonshuffling methods theoretically. 3. Numerical Experiments Expansion. While the experiments have been conducted on synthetic datasets and small-scale datasets, which is customary for theoretical papers, the scope of empirical validation could be enhanced. It would be much better if the authors could extend their experiments to include large-scale datasets. Technical Quality: 3 Clarity: 3 Questions for Authors: See above. Confidence: 3 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: Yes Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We thank you for your appreciation of our strengths and your positive evaluations. We addressed your comments in the general response. We repeat it below for your convenience: 1. Comparison of complexity with other shuffling methods Thank you for your suggestions. Among the stochastic shuffling gradient methods for the minimax problem, we have found three references (Das et. al., 2022; Cho & Yun, 2022; Emmanouilidis et. al., 2024) that are most related to our paper. However, the two references (Das et. al., 2022; Emmanouilidis et. al., 2024) are different from us that they consider two-sided PL condition for $H(w,u)$, or strongly monotone variational inequality. These settings are stronger than our general nonconvexity of $H_i$ in the first variable $w$. Thus our results cannot obtain the strongly convex (or PL) rate of $\mathcal{O}(\frac{1}{nK^2})$ as in these references. In comparison with the reference (Cho & Yun, 2022), our setting is broader than the problem in (Cho & Yun, 2022) that we consider functions $f$ and $h$. When $f = h = 0$, the semi-shuffling variant of Algorithm 2 with a random reshuffling strategy obtains the complexity of $\mathcal{O}(\sqrt{n} \epsilon^{-3})$ (noted in line 300, page 9), which is comparable with the complexity of Theorem 1 (Nonconvex-PL) in (Cho & Yun, 2022). Note that our Algorithm 2 is different from (Cho & Yun, 2022) even when $f=h=0$ since we do alternating epoch-wise update, while (Cho & Yun, 2022) uses component-wise update (alternating or simultaneous). When $f$ and/or $h$ are present, it remains unclear how to incorporate the prox operators in (Cho & Yun, 2022) as we have done. In addition, our analysis can be relaxed to the settings of the PL Assumptions 3+4 in (Cho & Yun, 2022) and obtain the same complexity as the strongly convex case. 2. Comparison of complexity with other non-shuffling methods A comparison between shuffling and other methods, including non-shuffling schemes was done in (Cho & Yun, 2022). For (Nonconvex-SC/PL) setting, the complexity $\mathcal{O}(\sqrt{n} \epsilon^{-3})$ of our Algorithm 2 with a random reshuffling strategy and (Cho & Yun, 2022) is better than the complexity $\mathcal{O}(\epsilon^{-4})$ of non-shuffling iid scheme (Lin et. al., 2020, Theorem 4.5). Not only for minimax optimization, but it has also been observed that shuffling gradient methods often have faster convergence than non-shuffling versions for stochastic optimization (Nguyen et. al. 2021). Moreover, in practice, these methods have shown improved performance over iid algorithms (Bottou, 2009, 2012; Hiroyuki, 2018). This is the motivation for us to work on the broad setting of our paper. All the discussions above will be added to our latest revision, and we will also add complexity comparison table. 3. Experiments We thank you for your suggestions. We will add more examples in the revision. We tested our method on a relatively large dataset url from LibSVM with n=2,396,130 and p=3,231,951 for our binary classification. The results are in the appendix. However, we would like to emphasize that the main contributions of our paper are theoretical. We have 4 theorems corresponding to 5 settings (2 settings for (NL) and 3 ones for (NC)). To our best knowledge, our methods for (NL) appear to be the first to develop shuffling-type methods for the class of nonconvex-linear minimax problems which has various applications in distributionally robust optimization and other learning scenarios (like risk-averse portfolio, model-agnostic meta learning). The shuffling strategy we use in Algorithm 1 is also new. For the nonconvex-strongly concave setting (NC), we have two strategies, and our model is more general than that of (Cho & Yun, 2022) due to $f$ and $h$. Our assumptions and methods are also different from (Cho & Yun, 2022). Overall, our paper is significant different from prior works except for (Cho & Yun, 2022) which we discussed above. We have other main contributions for the (NL) case, and our semi-shuffling scheme, apart from handling the nonsmooth terms $f$ and/or $h$. We hope our response answers all of your questions. If you have any additional comments and suggestions, please discuss with us and we are happy to clarify further. Again, thank you for your efforts in reviewing this paper!	Rebuttal 1: Rebuttal: Dear Reviewers, Thank you all so much for reviewing our submission. We valued your comments and appreciation of our strengths and contributions. We have addressed each comment individually for each reviewer. In this general response, we highlight some common responses to all the reviewers. 1. Comparison of complexity with other shuffling methods Thank you for your suggestions. Among the stochastic shuffling gradient methods for the minimax problem, we have found three references (Das et. al., 2022; Cho & Yun, 2022; Emmanouilidis et. al., 2024) that are most related to our paper. However, the two references (Das et. al., 2022; Emmanouilidis et. al., 2024) are different from us that they consider two-sided PL condition for $H(w,u)$, or strongly monotone variational inequality. These settings are stronger than our general nonconvexity of $H_i$ in the first variable $w$. Thus our results cannot obtain the strongly convex (or PL) rate of $\mathcal{O}(\frac{1}{nK^2})$ as in these references. In comparison with the reference (Cho & Yun, 2022), our setting is broader than the problem in (Cho & Yun, 2022) that we consider functions $f$ and $h$. When $f = h = 0$, the semi-shuffling variant of Algorithm 2 with a random reshuffling strategy obtains the complexity of $\mathcal{O}(\sqrt{n} \epsilon^{-3})$ (noted in line 300, page 9), which is comparable with the complexity of Theorem 1 (Nonconvex-PL) in (Cho & Yun, 2022). Note that our Algorithm 2 is different from (Cho & Yun, 2022) even when $f=h=0$ since we do alternating epoch-wise update, while (Cho & Yun, 2022) uses component-wise update (alternating or simultaneous). When $f$ and/or $h$ are present, it remains unclear how to incorporate the prox operators in (Cho & Yun, 2022) as we have done. In addition, our analysis can be relaxed to the settings of the PL Assumptions 3+4 in (Cho & Yun, 2022) and obtain the same complexity as the strongly convex case. 2. Comparison of complexity with other non-shuffling methods A comparison between shuffling and other methods, including non-shuffling schemes was done in (Cho & Yun, 2022). For (Nonconvex-SC/PL) setting, the complexity $\mathcal{O}(\sqrt{n} \epsilon^{-3})$ of our Algorithm 2 with a random reshuffling strategy and (Cho & Yun, 2022) is better than the complexity $\mathcal{O}(\epsilon^{-4})$ of non-shuffling iid scheme (Lin et. al., 2020, Theorem 4.5). Not only for minimax optimization, but it has also been observed that shuffling gradient methods often have faster convergence than non-shuffling versions for stochastic optimization (Nguyen et. al. 2021). Moreover, in practice, these methods have shown improved performance over iid algorithms (Bottou, 2009, 2012; Hiroyuki, 2018). This is the motivation for us to work on the broad setting of our paper. All the discussions above will be added to our latest revision, and we will also add complexity comparison table. 3. Motivation (Reviewer #jEDb). As we will explaine below, we have clear motivation when combining different components. Some are due to different assumptions (e.g., strongly concave vs merely concave), and some are optional (e.g. semi-shuffling vs full-shuffling). In addition, our shuffling strategy is quite general, which can cover both deterministic (e.g., incremental methods) and randomized variants. 4. Experiments We thank you for your suggestions. We will add more examples in the revision. We tested our method on a relatively large dataset url from LibSVM with n=2,396,130 and p=3,231,951 for our binary classification. The results are in the appendix. However, we would like to emphasize that the main contributions of our paper are theoretical. We have 4 theorems corresponding to 5 settings (2 settings for (NL) and 3 ones for (NC)). To our best knowledge, our methods for (NL) appear to be the first to develop shuffling-type methods for the class of nonconvex-linear minimax problems which has various applications in distributionally robust optimization and other learning scenarios (like risk-averse portfolio, model-agnostic meta learning). The shuffling strategy we use in Algorithm 1 is also new. For the nonconvex-strongly concave setting (NC), we have two strategies, and our model is more general than that of (Cho & Yun, 2022) due to $f$ and $h$. Our assumptions and methods are also different from (Cho & Yun, 2022). Overall, our paper is significant different from prior works except for (Cho & Yun, 2022) which we discussed above. We have other main contributions for the (NL) case, and our semi-shuffling scheme, apart from handling the nonsmooth terms $f$ and/or $h$. Finally, We hope our response answers all of your concerns and questions. If you have any additional comments and suggestions, please discuss with us and we are happy to clarify further. Again, thank you for your efforts in reviewing this paper! Added References: Tianyi Lin, Chi Jin, and Michael Jordan. On gradient descent ascent for nonconvex-concave minimax problems. In International Conference on Machine Learning, pp. 6083–6093. PMLR, 2020 L. Bottou. Curiously Fast Convergence of Some Stochastic Gradient Descent Algorithms. In Proceedings of the Symposium on Learning and Data Science, Paris, volume 8, pages 2624–2633, 2009 L. Bottou. Stochastic Gradient Descent Tricks. In Neural networks: Tricks of the trade, pages 421–436. Springer, 2012. K. Hiroyuki. SGDLibrary: A MATLAB Library for Stochastic Optimization Algorithms. Journal of Machine Learning Research, 18(215):1–5, 2018.	NeurIPS_2024_submissions_huggingface	2,024	null	null	null	null	null	null	null	null
OPEL: Optimal Transport Guided ProcedurE Learning	Accept (poster)	Summary: The paper presents OPEL, a novel framework for procedure learning from videos that leverages optimal transport (OT) to align key steps across different video instances. OPEL treats video frames as samples from an unknown distribution and formulates the distance calculation between them as an optimal transport problem, allowing for more flexible alignment compared to frame-to-frame mappings. The authors introduce two regularization terms to improve the OT formulation and experiments significantly outperforms state-of-the-art methods on benchmark datasets. Strengths: 1. This paper proposes a novel optimal transport-based procedure learning framework that aligns frames with similar semantics together in an embedding space. 2. The paper enhances the OT formulation with two regularization terms that address temporal and semantic relationships, contributing to better alignment and learning. 3. OPEL demonstrates effective performance improvements over previous state-of-the-art methods on benchmark datasets. Weaknesses: 1. I believe the authors can compare with recent action segmentation methods to further strengthen the experiments. 2. I believe the authors can also compare with methods that use temporal alignment methods like dynamic time warping to strengthen their contributions. Technical Quality: 3 Clarity: 3 Questions for Authors: I do not have specific questions on the model design. I only have some experiment suggestions which are the same as I listed in "weaknesses". Confidence: 4 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: N/A Flag For Ethics Review: ['No ethics review needed.'] Rating: 5 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thanks for your encouraging and insightful feedback. Please find the answers to your specific comments: Comparison with AS methods: Procedure learning (PL) and action segmentation (AS) are related but not the same. PL, when applied to a set of instructional videos depicting the same task, involves two primary steps: (i) assigning each video frame to one of the k key-steps (including background elements), and (ii) determining the logical sequence of these key-steps necessary to complete the task. As illustrated in Fig. 1 of the main paper, PL addresses multiple videos of a given task, enabling the identification of repetitive key-steps across these videos [i, j]. In contrast, AS [k] focuses on a single video, thereby lacking the ability to discern repetitive key-steps across different videos. Despite the differences between PL and AS, as per the reviewer's suggestion, we compare our approach against existing SOTA unsupervised AS models [a-f] and present the results in Table X below. Our model demonstrates a significant performance improvement compared to these works. In [b, f], authors report a high recall score for CrossTask as it assigns majority of the frames to a single key-step - a phenomenon also reported by [j]. While achieving high recall is important for ensuring that most positive instances are correctly identified, it can result in a greater number of false positives, which in turn lowers precision and leads to undesirable results. Therefore, it is crucial to balance recall with precision to develop an effective model. This balance is reflected in the superior performance of our model, as evidenced by the F1-score results across various benchmarks. Note in the Table X, our approach is compared with SOTA Unsupervised AS methods for only third-person datasets, as these do not report any result on egocentric datasets. Table X: Comparison of our approach with SOTA Unsupervised Action Segmentation methods. Note '-' denotes authors have not provided any data on those metrics. \|Action Segmentation Papers\| \|ProceL \| \| \| \|CrossTask\| \| \|:----\|:---:\|:---:\|:---:\|:-:\|:---:\|:---:\|:---:\| \| \|P\|R\|F1\|\|P\|R\|F1\| \|JointSeqFL (2019) [a]\|-\|-\|29.8\|\|-\|-\|-\| \|Elhamifar et al. (2020) [b]\|9.5\|26.7\|14.0\|\|10.1\|41.6\|16.3\| \|Fried et al. (2020) [c]\|-\|-\|-\|\|-\|28.8\|-\| \|Shen et al. (2021) [d]\|16.5\|31.8\|21.1\|\|15.2\|35.5\|21.0\| \|Dvornik et al. (2022) [e]\|-\|-\|-\|\|-\|-\|25.3\| \|StepFormer (2023) [f]\|18.3\|28.1\|21.9\|\|22.1\|42\|28.3\| \|OPEL\|33.6\|36.3\|34.9\|\|35.6\|34.8\|35.1\| Temporal Alignment methods: According to the reviewer's suggestion, we have included additional comparisons with temporal alignment techniques in Table Y. Specifically, we have compared our model with methods like TCC [g] and LAV [h], which incorporate temporal cycle consistency (TCC) and dynamic time warping (DTW), respectively. Other methods like CnC [i] uses TC3I (TCC + contrastive inverse difference moment, C-IDM) as the loss function while GPL [j] uses graph-based representation for temporal alignment. Our results clearly demonstrate the efficacy of our approach compared to these established methods, further validating the effectiveness of our model in maintaining temporal alignment while delivering superior PL performance. Note, in Table Y, our approach is compared with SOTA Unsupervised Temporal alignment methods for only egocentric (first-person) datasets, as these do not report any result on third-person datasets. Table Y: Comparison of our approach with existing Unsupervised Temporal Alignment methods. \|Temporal Alignment Methods \| CMU-MMAC\| \| \| \| MECCANO\| \| \| \|EPIC-Tent\| \| \| \|EGTEA-GAZE+\| \| \| \|PC Assembly\| \| \| \|PC Disassembly\| \| \| \|:----\|:---:\|:---:\|:---:\|:---:\|:---:\|:---:\|:---:\|:---:\|:---:\|:---:\|:---:\|:---:\|:---:\|:---:\|:---:\|:---:\|:---:\|:---:\|:---:\|:---:\|:---:\|:---:\|:---:\| \| \|Precision\|F1\|IoU\| \|Precision\|F1\|IoU\| \|Precision\|F1\|IoU\| \|Precision\|F1\|IoU\| \|Precision\|F1\|IoU\| \|Precision\|F1\|IoU\| \|TCC [g]\|18.5\|19.7\|9.5\| \|15.1\|17.9\|8.7\| \|14.2\|14.9\|7.8\| \|17.5\|19.7\|8.8\| \|19.9\|21.7\|11.6\| \|22.0\|23.4\|12.2\| \|LAV (DTW) [h]\|20.6\|21.1\|9.4\| \|14.6\|17.4\|7.1\| \|15.2\|15.8\|8.3\| \|17.4\|19.1\|8.0\| \|21.5\|22.7\|11.7\| \|26.4\|26.5\|12.9\| \|LAV + TCC\|18.8\|19.7\|9.0\| \|13.4\|15.6\|7.3\| \|16.0\|16.7\|8.5\| \|16.4\|18.6\|7.5\| \|21.6\|21.1\|10.8\| \|21.0\|24.3\|12.3\| \|CnC(TC3I) [i]\|21.6\|22.7\|11.1\| \|15.5\|18.1\|7.8\| \|17.1\|17.2\|8.3\| \|19.6\|21.7\|9.5\| \|25.0\|25.1\|12.8\| \|28.4\|27.0\|14.8\| \|GPL [j]\|30.3\|31.7\|17.9\| \|18.8\|20.7\|10.0\| \|17.9\|19.8\|9.1\| \|23.8\|27.1\|16.0\| \|27.1\|27.5\|15.2\| \|28.1\|26.7\|15.2\| \|OPEL\|32.8\|36.5\|18.8\| \|28.9\|39.2\|20.2\| \|18.8\|20.7\|10.6\| \|24.3\|29.5\|13.2\| \|32.5\|33.7\|17.9\| \|29.6\|32.2\|16.7\| Thanks again for the suggestions to strengthen our paper, looking forward to your kind consideration. Refs.: [a] Elhamifar et al. Unsupervised Procedure Learning via Joint Dynamic Summarization, ICCV 2019. [b] Elhamifar et al. Self-supervised multi-task procedure learning from instructional videos, ECCV 2020. [c] Fried et al. Learning to segment actions from observation and narration, ACL 2020. [d] Shen et al. Learning to segment actions from visual and language instructions via differentiable weak sequence alignment, CVPR 2021. [e] Dvornik et al. Flow Graph to Video Grounding for Weakly-Supervised Multi-step Localization, ECCV 2022. [f] Dvornik et al. StepFormer: Self-supervised Step Discovery and Localization in Instructional Videos, CVPR 2023. [g] Dwibedi et al. Temporal cycle-consistency learning, CVPR 2019. [h] Haresh et al. Learning by aligning videos in time, CVPR 2021. [i] Bansal et al. My view is the best view: Procedure learning from egocentric videos. ECCV 2022. [j] Bansal et al. United we stand, divided we fall: Unitygraph for unsupervised procedure learning from video. WACV 2024. [k] Kumar et al. Unsupervised Action Segmentation by Joint Representation Learning and Online Clustering. CVPR, 2022. --- Rebuttal Comment 1.1: Title: Discussion period Comment: Dear Reviewer, thanks again for your feedback, hope you consider our responses. Please let us know if you have any further queries, looking forward to the discussion.	Summary: The authors propose a novel approach for procedure learning leveraging optimal transport, OPEL. OPEL integrates optimality and temporal priors, and incorporates a novel inter-video contrastive loss. OPEL achieves significant improvements on egocentric and third-person benchmarks. Strengths: 1. An interesting and relevant topic regarding the treatment of procedure learning as an optimal transport problem. 2. The authors thoroughly consider deviations in real-world sequences, including background frames, redundant frames, and non-monotonic frames. 3. OPEL achieves excellent results across several benchmarks. Weaknesses: 1. The presentations of the paper needs improvement, especially in Section 3, where an excessive number of formulas obstruct readability. 2. Some figures are unclear, such as Figure 1. Are V1, V2, V3, V4 from four different videos, or do they appear to be from the same video? Additionally, what distinguishes V3 from V4, where one represents background frames and the other represents redundant frames? Technical Quality: 3 Clarity: 2 Questions for Authors: 1. I still have some doubts regarding the difference between procedure learning and action segmentation. I believe their final output forms are same. From related work, I believe that action segmentation involves classifying each frame, while procedure learning only needs to identify key steps. Is this correct? Could you please explain in detail? 2.I also feel that the paper doesn't clearly present the problem formulation, including the input and output, which makes it somewhat confusing to understand. Does the output video contain a fixed number of K key steps? According to the paper, setting K to 7 is optimal. However, based on the dataset, the average number of steps for each task may be greater than or less than K. When it exceeds K, are some steps merged? And what happens when it's less than K? Is it reasonable to pre-define K? 3. Could you provide more results of ablation training on K for additional tasks? Confidence: 3 Soundness: 3 Presentation: 2 Contribution: 3 Limitations: Yes Flag For Ethics Review: ['No ethics review needed.'] Rating: 5 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thanks for your encouraging feedback. Please find answers to your specific comments: Presentation: We agree that the equations are too dense in Section 3, and the formulations can be simplified. We will update accordingly as per the reviewer’s suggestion during revision. The V1 to V4 in Fig. 1 are from four different videos, while the frames from the same video are shown in x-axis. We apologize for the ambiguity of background frames vs redundant frames; the Fig.1 has been updated as Fig. R1 (rebuttal pdf). Both refer to frames not directly related to any key-step (hence, do not belong to any cluster) and thus are clustered together as a background cluster (Fig. R1, right). So, we remove the term redundant frames for clarity. PL vs AS: Procedure learning (PL) and action segmentation (AS) are related but not the same. PL, when applied to a set of videos of the same task, involves 2 primary steps: (a) assigning each frame to 1 of the k key-steps (including background elements), (b) determining the logical sequence of the key-steps to complete the task. As illustrated in Fig. 1, PL addresses multiple videos of a given task, enabling the identification of repetitive key-steps across these videos [a, b]. In contrast, AS [c] focuses on a single video, thereby lacking the ability to discern repetitive key-steps across videos. Additionally, AS does not consider the sequence of individual events, which is often crucial for accurately identifying procedures. For instance, AS fails to capture the variations in the order of key-steps, such as those observed between V1 and V2 in Fig. 1. In our work, the OPEL loss is used to train a representation learning algorithm to obtain an embedding space where similar frames are located close by. Then, as described in the ‘Clustering and Key-step Ordering’ part of page 6 of the main paper, these key-steps are clustered, and their sequential order is determined. A PyTorch function (with a toy example) is provided in codeblock R1 of the rebuttal pdf to illustrate this process of finding the sequential order (part b from the definition of PL given above) of the frame-wise key-step predictions (output from part a from the above PL definition). Problem formulation: Given a set of videos of the same task (e.g. making brownies), PL aims to find the constituent key-steps (e.g. break egg, mix contents, add oil, mix egg) and their sequential order (break egg --> mix egg --> add water --> add oil --> mix contents). During training: Inputs are untrimmed and unlabeled videos. We train with pairs of videos at a time (formulation detailed in Section 3 of the main paper). Output is the frames clustered into different key steps and their sequential order. For evaluation metrics, we use framewise cluster predictions, following other SOTA works [a,b,d]. During inference: Input: a single video, $P = \left[p_1,\ p_2,\ .\ .\ .\ ,\ p_N\right]$ Output: Assign each frame to a phase using trained model $f_{\theta}∶ f_{\theta}(p_i)= l_i$, where $l_i ∈ {c}^k_{c=1}$, represents the key-step corresponding to the phase of each frame $p_i$ in the video P. From all $l_i$, we determine the sequential order of the sub-tasks (eg: 0,3,2,4,1...). Note, the cluster number is arbitrary, each of them denoting a key-step. We provide a PyTorch code to illustrate this process of finding the sequential order in codeblock R1 of the rebuttal pdf. *Choice of k:* In unsupervised PL, k is not set during training (learning of the embeddings), rather k is used during inference only to fix the no. of output clusters, and each frame of the video being inferred is mapped to one of those k clusters. Now, an output video may not contain that exact k number of key-steps (e.g. in making a sandwich, steps such as adding jelly, butter etc. may be present in some videos, while absent in others). We obtain best results with k=7 (Table 7), consistent with other SOTA methods on the same datasets [a,b,d]. We agree that based on the dataset, the average number of steps for each task may be > or < k. If it exceeds k, some steps get merged, e.g., for PC Disassembly, though the ground-truth number of steps is 9, 3 are quite similar (remove hard disk, remove motherboard, remove RAM); effectively making them quite close in the feature space. As a result, choosing k=7 results in these 3 key-steps getting merged. When the average number of key-steps becomes < k, each subtask gets split into multiple smaller clusters with very similar embeddings. This phenomenon is illustrated in Fig. R2 (rebuttal pdf), where a larger k might result in several split clusters of small windows (blue and red boxes in Fig. R2). Note, this demarcation of subtasks (no. of clusters) varies from task to task, dataset to dataset and is subjective as some may consider semantically similar tasks (e.g. pouring oil vs water) to be same subtask, while others may consider it different. So, for best performance, k might be adjusted task-wise (using methods such as the elbow method, AIC criterion). However, optimizing for task-wise k is not the goal of this work; our contribution is on the OT-based learning and the clustering is only used as an inference post-processing step. So for fair comparison, we experiment with some reasonable values of k (following prior art [a,b,d]) and consistently outperform them across all k and all datasets with similar trends. We report additional ablation on varying k in the rebuttal pdf, Table R1. We find that the performance degrades as k is increased from $7\to10\to12\to15$, thus overall k=7 provides the best performance. Note, our contribution is on the OT-based learning, the clustering is only used as an inference post-processing step. Overall, we achieve similar trends for varying k w.r.t. SOTA works [a,b,d]. [a] Bansal et al. ECCV 2022 [b] Bansal et al. WACV 2024 [c] Kumar et al. CVPR, 2022 [d] Shah et al. ICCV, 2023 --- Rebuttal Comment 1.1: Title: Discussion period Comment: Dear Reviewer, thanks again for your feedback, hope you consider our responses based on your comments. Please let us know if you have any further queries, looking forward to the discussion.	Summary: The paper proposes an unsupervised method for procedure learning that identifies the key steps and their orders in several videos of the same task. The paper formulates the distribution of video frames as an optimal transport (OT) problem to compute the distances between the key steps. To handle the variations of the videos, the paper introduces a regularization with prior distribution. The proposed method achieves SOTA on third-person and first-person video benchmarks. Strengths: - Different from prior methods having ordering constraints assumption, the paper relaxes this by formulating OT. - The paper addresses the variation between videos and then introduces a regularization with priors to enhance the OT. - The paper shows SOTA results in both first-person and third-person evaluations. Weaknesses: - The explanation of using priors to mitigate the variation of the videos, i.e., action speeds, non-monotonic sequences, or starting of actions, is not clear. - Besides the ablation studies of different prior distributions, the paper should explain why Laplacian outperforms other distributions. Technical Quality: 3 Clarity: 2 Questions for Authors: The authors should give a clear explanation of the regularization usage as mentioned in the weakness. Confidence: 3 Soundness: 3 Presentation: 2 Contribution: 3 Limitations: The authors have mentioned the limitations. Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thanks for your encouraging feedback. Please find the answers to your specific comments: Explanation of priors: The concept of the Optimality Prior is crucial when dealing with video alignment, especially in challenging scenarios. When two videos are perfectly aligned (case 1 of Fig. 2B), the $\hat{T}$ becomes strictly diagonal. However, the real value of the optimality prior emerges in more complex situations, such as speed variation, non-monotonicity, etc. Speed variation: Consider the scenario in Fig. 2B, case 3: video 1, shown at the top, takes two frames to complete task A, while video 2, displayed on the left, completes task A in one frame. Here, both frames 1 and 2 from video 1 should be aligned with frame 1 of video 2. The optimality prior enables this alignment, as depicted by the blue curve in Fig. 2C and explained by Eqn. 2. The procedure involves finding, for each frame $i$ from video 2, the closest matching frame from video 1 based on feature similarity, even if the corresponding frame $j$ from video 1 has a different index. In Fig. 2C, for example, frame $j$ from video 1 has the highest likelihood of aligning with frame $i$ from video 2, with the assignment likelihood decaying exponentially for frames further away from $j$. The optimality prior also addresses non-monotonic cases (Fig. 2B, case 4). Here, though the first two frames from video 1 align with those in video 2, the third frame of video 1 reverts to an earlier task A. To correctly align this frame, it must be matched with its closest counterpart in video 2, which is frame 1 in this case. The optimality prior facilitates this by aligning non-monotonic frames that share a higher feature-wise match (similar to Fig. 2C). However, relying solely on the optimality prior is insufficient because it overlooks the temporal ordering inherent in videos, which is crucial for maintaining temporal coherence. As discussed in the regularization section on page 4 of our paper and illustrated in Fig. R4 (rebuttal pdf), optimizing only for optimality can result in temporal incoherence. Proper alignment should account for the temporal relationship between frames, ensuring that corresponding frames in one sequence align closely with adjacent frames in the other. This necessity leads to a second critical factor: the Temporal Prior. Unlike the optimality prior, which seeks the best feature match regardless of temporal distance, the temporal prior promotes alignment between frames that are temporally adjacent, thereby preserving the overall temporal coherence. Similar coherence based concepts have been utilized in other temporal alignment works, such as TCC [a], and contrastive regularization [b]. This temporal prior encourages the alignment matrix to exhibit peak values along the diagonal, with diminishing values away from the diagonal; we model this with a Laplace distribution (Eqn. 3 and the red curve in Fig. 2C). Essentially, there are 2 factors in play- (i) optimality which tries to find the best match between frames irrespective of their temporal distance (which may result in temporally incoherent alignment), (ii) the temporal factor which promotes transport between nearby frames without considering their feature matching. We hypothesize (and later validate with results) that the optimal solution requires a balance between both, and therefore propose to merge these two priors addressing the above factors, as expressed in Eqn. 4. This combined prior ensures accurate alignment between videos considering all the factors, as further illustrated in Fig. 2D and supported by the results in Table 5 of the main paper. When the starting of actions is different (case 2, Fig. 2B), video 2 has a background frame to start when video 1 starts doing task A. As mentioned in page 5 of our paper, we introduce a \'virtual frame\' to handle such cases. Even in this case, the combined prior comes into play as we assign any frame to the virtual frame if the likelihood of that frame aligning with any other task-related frames from the other video falls below a predefined threshold. This phenomenon is also illustrated in Fig. 4(B) of the main paper. Laplace distribution: As shown in appendix Table A6 (main paper), Laplace as a prior outperforms other distributions. To analyze this, we plot the distributions in Fig. R3 (rebuttal pdf). Note, we use the same distribution for optimality as well as temporal priors. For the optimality prior, the x-axis is the difference of frames in the feature space (1-d representation for illustration purposes), and the y-axis denotes the corresponding probability of alignment. We want the point representing the most likely alignment (as per $\hat{T}$) to have the highest likelihood, while the assignment probability should exponentially decay further away. The graph clearly shows that the Laplace distribution captures this behavior suitably compared to Uniform and Gaussian. Similarly, for the temporal prior, the x-axis denotes the temporal distance between the frames, and the y-axis denotes the corresponding probability of alignment. The graph shows that Laplace distribution facilitates alignment of the frames when they are temporally aligned (close to center), and its long tail distribution enables better correlation of non-monotonic frames compared to Gaussian or Uniform. As a result, as shown in Fig. R3, even at far away locations from the center (temporally distant frames), alignment is possible if indeed these frames have a high match feature-wise. In this case, the Laplace temporal prior provides non-zero probability to that far away frame (due to its long tail) unlike other distributions and the optimality prior gives a large score (due to feature match), resulting in improved handling of non-monotonicity. Refs: [a] Dwibedi et al. Temporal cycle-consistency learning. CVPR 2019. [b] Haresh et al. Learning by aligning videos in time. CVPR 2021. --- Rebuttal Comment 1.1: Title: Discussion period Comment: Dear Reviewer, thanks again for your feedback. We hope you consider our responses based on your comments, looking forward to the discussion.	Summary: OPEL is a novel technique for Procedure Learning. Procedure learning is the task of finding key steps in an action (such as cooking brownie) and aligning the videos based on these key steps. OPEL proposes to use the optimal transport distance between the two videos as the similarity metric rather than direct frame by frame mapping or assuming a strict monotone mapping. This work also proposes two regularizers to incorporate priors about monotone mapping and about increasing the correspondence between nearby frames. The results on both egocentric and 3rd person videos show significant improvement upon the previous SOTA tasks. The ablation studies analyze the effect of each of the losses and regularizers. It seems having the temporal prior and the inter-intra cross entropy are the most important factors in terms of f1 accuracy. Strengths: This is a sound and novel technique for procedure learning. Using optimal transport between frames significantly improves the results and add leniency when the frames are not exact matches temporally. This is a common happenstance in real world scenarios since not all the steps of a procedure need to happen in chronological order. Two sub procedures can be interchangeable and using optimal transport with only a regularizer on temporal monotony is can account for this. The experiments are sufficiently evaluated on several datasets. Over all the datasets this method out performs the prior work. The manuscript is written in a detailed fashion and the full information is provided for sake of replicability. Also the ablation study provided improves the clarity of what is important in this method. Weaknesses: There are many factors added together to make this method work. Although many of them are common in other methods as well, but the method and also the ablation seem to suggest that some of the losses may be redundant. It is not clear if the difference between some of the lines in table. 5 is statistically significant. Page 6. is written in such a compact way that it's not easy to read and follow. The indices such as i,j can be probably dropped. Words such as temperature in the formula can be probably replaced with tau. Technical Quality: 3 Clarity: 3 Questions for Authors: Why do you think the number of steps makes such a large difference in table 7? going from 7 to 10 drops the performance significantly. another anomaly is 12 being worse than both 10 and 15. Confidence: 3 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: not explicitly. Flag For Ethics Review: ['No ethics review needed.'] Rating: 7 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thanks for your encouraging and insightful feedback. Please find the answers to your specific comments. Also please refer to our overall rebuttal response to all reviewers and the corresponding pdf containing additional figures and tables to support our claims. Method factors: The reviewer correctly points out that all the components are not equally critical for OPEL’s performance. We show the contributions of each factor individually to analyze their effect on the overall result in Table 5. Comparing row 3 with row 9, we observe, the priors jointly play a critical role; as without them (row 3), the F1 and IoU scores drop by \~5 points. Specifically, the optimality prior has a significant impact (\~2 point), while the temporal prior affects the score by \~1 point. Similar to the combined priors, the intra and inter-video contrastive losses together (see row 6 vs row 9) have a significant effect (\~3.5 points) on the overall performance. The individual effect of virtual frame is negligible as it only plays a role in case of excessive background frames - a scenario that is not prevalent in most datasets. Furthermore, due to the IDM structure of $M\left(\hat{T}\right)$, the $\hat{T}$ and $\hat{Q}$ are similar by formulation. This results in $KL\left(\hat{T}\ \parallel\hat{Q}\right)$ to be already small. As a consequence, adding the KL divergence as a standalone loss component in the proposed pipeline has a minimal impact. Overall, while some loss components may have a smaller individual impact, they do contribute to performance improvements, even if incrementally. Therefore, our proposed approach incorporates all of them to achieve the best possible results. We are encouraged with the reviewer's note - the provided ablation study improves the clarity of what is important in this method. Writing: The equations are an integral part in explaining the proposed approach and the different loss terms. But we agree that the page 6 is too dense, and the formulations can be simplified to improve readability. We will update the notations and improve the overall writing as per the reviewer’s suggestion during revision. Effect of k: Note, we obtain best results with k=7 (Table 7), and performance drops sharply as k goes from 7 to 10. This observation is consistent with all the other SOTA methods on the same datasets [a, b, c]. We hypothesize that k=7 works best as it is the optimal number of clusters considering the average number of distinct key-steps (subtasks) of the datasets. For example, for PC Disassembly, although the ground-truth (GT) number of steps is 9, 3 steps are quite similar (remove hard disk, remove motherboard, remove RAM), effectively making them quite close in the feature space. This results in k=7 being a better estimation of the cluster number with distinct steps. Note, this demarcation of subtasks (hence, number of clusters) is subjective and varies from dataset to dataset as well as from task to task; as some may consider semantically similar tasks (e.g. pouring oil vs water) to be one subtask, while others may consider it different. As k becomes larger than the actual distinctive number of clusters, each subtask gets split into multiple clusters with very similar embeddings, which upon comparison with GT leads to inferior results. This phenomenon is illustrated in Fig. R2 (see rebuttal pdf), where a larger k might result in several erroneous clusters with very small windows (blue and red boxes in Fig. R2). This leads to large fluctuations (jittery predictions) within a single GT phase, thus deteriorating the overall performance. Secondly, the anomalous trend (k = 12 being slightly worse than 10 and 15) might be a dataset (PC Assembly) specific issue and not unique to our approach, as a similar trend for this dataset has been reported in [a, b]. However, for PC Disassembly and other datasets (reported as additional results in the rebuttal pdf, Table R1), we consistently find that the performance degrades as k is increased from 10 $\to$ 12 $\to$ 15. Note, that our contribution is on the optimal transport (OT)-based representation learning, the clustering is only used as an inference post-processing step. Overall, we achieve similar trends with respect to the SOTA works but consistently outperform them across all k and datasets. Refs.: [a] Bansal et al. My view is the best view: Procedure learning from egocentric videos. ECCV 2022. [b] Bansal et al. United we stand, divided we fall: Unitygraph for unsupervised procedure learning from video. WACV 2024. [c] Shah et al. Steps: Self-supervised key step extraction and localization from unlabeled procedural videos. ICCV, 2023. --- Rebuttal Comment 1.1: Title: Acknowledged Comment: Thank you for your response. In terms of statistical significance, what is the standard deviation for a set of 5 runs for example? --- Reply to Comment 1.1.1: Title: Consistent results across multiple runs Comment: Thanks for your response. In our paper, we reported just the mean values obtained over multiple runs and not the standard deviations as the results did not vary significantly. Also, previous SOTA works [a, b, c] do not report the standard deviation across runs either. However, as per the reviewer's suggestion, we report the results of the mean ± standard deviation (SD) over 5 separate runs in Table A below. Note, in all cases, we find that the SD is quite low, and we get consistent results over the multiple runs, further demonstrating the statistical significance of the results. Table A: Results showing mean ± SD over 5 runs for all the datasets \| Dataset \| F1 (mean ± SD) \| IoU (mean ± SD) \| \|:----------------\|:-----------------:\|:------------------:\| \| CMU-MMAC \| 36.5 ± 0.138 \| 18.8 ± 0.106 \| \| EGTEA-GAZE+ \| 29.5 ± 0.147 \| 13.2 ± 0.145 \| \| MECCANO \| 39.2 ± 0.319 \| 20.2 ± 0.258 \| \| EPIC-Tents \| 20.7 ± 0.165 \| 10.6 ± 0.101 \| \| PC Assembly \| 33.7 ± 0.311 \| 17.9 ± 0.184 \| \| PC Disassembly \| 32.2 ± 0.317 \| 16.9 ± 0.203 \| \| ProceL \| 34.9 ± 0.095 \| 21.3 ± 0.037 \| \| CrossTask \| 35.1 ± 0.142 \| 21.5 ± 0.111 \| Please let us know if you have any further queries. Thanks again and hope you reconsider your score based on our responses. Refs.: [a] Bansal et al. My view is the best view: Procedure learning from egocentric videos. ECCV 2022. [b] Bansal et al. United we stand, divided we fall: Unitygraph for unsupervised procedure learning from video. WACV 2024. [c] Shah et al. Steps: Self-supervised key step extraction and localization from unlabeled procedural videos. ICCV, 2023.	Rebuttal 1: Rebuttal: We thank all the reviewers for their insightful comments and feedback. We are encouraged that the reviewers like the soundness and novelty of our approach for procedure learning along with comprehensive evaluation (Reviewer XDhK), and enhancement of results over current state-of-the-art (SOTA) works (Reviewer pvnX). Additionally, Reviewer Yy5N finds our work interesting and relevant as we thoroughly consider deviations in real-world sequences, including background and non-monotonic frames. Lastly, Reviewer 3Xtn appreciates the novelty of our work and its SOTA performance. The reviewers have also raised some concerns that we have addressed in their individual rebuttal responses. Furthermore, as per the reviewers' suggestions, we have added additional relevant figures and results to support our claims in the attached 1 page pdf and referred to them in our responses. Specifically, 1. Figure R1 is a modified version of Figure 1 of the main paper. 2. Figure R2 illustrates the qualitative effect of varying k for clustering the key-steps. 3. Figure R3 shows the reasoning for the choice of Laplace distribution as a prior over other distributions. 4. Figure R4 emphasizes of the importance of both the optimality and temporal priors to combat real-world non-idealities like non-monotonic frames, speed variation, etc., while maintaining temporal coherence. 5. Table R1 shows additional ablation study on the effect of k on MECCANO and EPIC-Tents datasets. 6. Codeblock R1 depicts a Pytorch function to determine the sequential ordering of tasks from frame-wise key-step predictions. In general, we agree that the formulations in Section 3 can be simplified to improve readability. We will update accordingly as per the reviewers' suggestions during the camera-ready version, if accepted. Please note, in our rebuttal responses, Fig./Table X (e.g. Fig. 2/Table 5) refers to the main submitted manuscript, while Fig./Table RX (e.g. Fig. R2/Table R1) refers to the rebuttal pdf attached herewith. Once again, we sincerely appreciate your time and consideration. Please let us know if you have any further queries. We look forward to your responses. Pdf: /pdf/0e5d1a8445e5954832a4f120b4aebca6a69f126d.pdf	NeurIPS_2024_submissions_huggingface	2,024	null	null	null	null	null	null	null	null
SearchLVLMs: A Plug-and-Play Framework for Augmenting Large Vision-Language Models by Searching Up-to-Date Internet Knowledge	Accept (poster)	Summary: The paper presents UDKAG, a novel framework designed to enhance the capabilities of Large Vision-Language Models (LVLMs) by integrating up-to-date knowledge retrieved from the internet during the inference phase. The authors have developed a hierarchical filtering model to identify pertinent content from search engine results and a UDK-VQA dataset to evaluate the model's performance on VQA tasks. Strengths: 1. The paper introduces an novel internet-augmented generation framework to update newest knowledge. 2. The creation of the UDK-VQA dataset is a valuable contribution. 3. Significant performance improvement. Weaknesses: I'm not familiar with this field and from my perspective this paper may not have obvious weakness. Technical Quality: 3 Clarity: 3 Questions for Authors: I'm curious about the generalization ability after integrating up-to-date knowledge. Will the model forget some important information and perform worse on other data? Confidence: 1 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: Yes Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thank you for your time and effort in reviewing our paper and for your recommended acceptance. We sincerely appreciate your valuable comments and feedback. --- > Question: I'm curious about the generalization ability after integrating up-to-date knowledge. Will the model forget some important information and perform worse on other data? Answer: The performance of large vision-language models (LVLMs) with our framework remains robust even on datasets that do not require up-to-date knowledge. We have conducted experiments on more existing datasets, including GQA [1], INFOSEEK [2] and A-OKVQA [3]. GQA is designed to evaluate the reasoning abilities of LVLMs and does Not Rely on external Knowledge (NRK). INFOSEEK and A-OKVQA Require external Commonsense Knowledge (RCK), such as "the Eiffel Tower is in Paris" and "the important function of a folding chair is portability", rather than up-to-date knowledge. In contrast, our UDKAG dataset Require Up-to-Date Knowledge (RUDK). Our experimental results, presented in the table below, demonstrate that our framework consistently improves various baselines across these three datasets, underscoring the generalizability of our approach. --- \| Baseline \| Variant \| Local Data (clear GT) \| Internet Data (unclear GT) \| GQA [1] (NRK) \| INFOSEEK [2] (RCK) \| A-OKVQA [3] (RCK) \| UDKAG (RUDK) \| \|---------------\|-----------\|-----------------------\|----------------------------\|---------------\|--------------------\|-------------------\|--------------\| \| CFR [5] \| - \| - \| - \| 72.10 \| - \| - \| - \| \| Oracle→FID [2]\| - \| ✓ \| - \| - \| 45.60 \| - \| - \| \| Omni-SMoLA [4]\| - \| ✓ \| - \| - \| - \| 84.10 \| - \| \| LLaVA-1.6 \| Raw \| - \| - \| 61.66 \| 37.86 \| 75.53 \| 31.8 \| \| \| Ours \| - \| ✓ \| 62.33 \| 41.25 \| 76.22 \| 90.2 \| \| InternVL-1.5 \| Raw \| - \| - \| 74.03 \| 51.13 \| 84.53 \| 42.6 \| \| \| Ours \| - \| ✓ \| 74.41 \| 53.10 \| 84.59 \| 92.9 \| --- [1] Hudson et al. Gqa: A new dataset for real-world visual reasoning and compositional question answering. In CVPR 2019. [2] Chen et al. Can pre-trained vision and language models answer visual information-seeking questions? In EMNLP 2023. [3] Schwenk et al. A-okvqa: A benchmark for visual question answering using world knowledge. In ECCV 2022. [4] Wu et al. Omni-SMoLA: Boosting generalist multimodal models with soft mixture of low-rank experts. In CVPR 2024. [5] Nguyen et al. Coarse-to-fine reasoning for visual question answering. In CVPRW 2022. --- Rebuttal Comment 1.1: Comment: I appreciate the author's commitment and valuable time in addressing my concerns. After reading through all the reviews and rebuttals, I believe this work deserves a broader discussion. I tend to maintain the 'weak accept' recommendation. --- Reply to Comment 1.1.1: Title: Thank you for your review! Comment: Thank you for your prompt reply and valuable feedback on our manuscript. In the revised version, we will include experimental results on more datasets to better demonstrate the generalizability of our framework.	Summary: This paper investigates a novel approach to augment large vision-language models with up-to-date knowledge from the internet. The author proposes to extensively leverage existing search engines (e.g., Bing and Google) and foundation models like ChatGPT for web information searching and parsing. A hierarchical filtering model has been proposed to filter the irrelevant information from the retrieved websites. The method can augment existing LVLM for augmented generations and achieve the state-of-the-art (SOTA) performance on a newly proposed VQA dataset, i.e., UDK-VQA, where the proposed plug-and-play method can improve SOTA LLM/VLMs like Gemini 1.5 Pro, GPT-4v, GPT-4o and LLaVA 1.6. Extensive ablation studies have been conducted to analyze the effectiveness of each component of the proposed framework. Strengths: 1. The paper is very well written and strongly motivated, given the necessity of augmenting the LVLM with up-to-date knowledge. To the best of my knowledge, this should be the first paper on augmenting the LVLM with internet information in a comprehensive manner. 2. The proposed method is novel, given its design choice and intelligent combination of search engine and foundation model to enable precise information retrieval from the internet. The reviewer believes that such a new systematic system will inspire future exploration of IAG for multi-modality foundation models. 3. The proposed method can be plug-and-play into many existing foundation models without trial and error, demonstrating its universality and significance. Moreover, the foundation models augmented by the proposed method can be improved significantly, highlighting the effectiveness of the proposed method. 4. Extensive analysis has been conducted on the proposed method to verify further the effectiveness of the proposed framework. Weaknesses: Overall, this is a strong paper. The primary concern is the cost of leveraging the search engine and LLM like ChatGPT. The author should elaborate more on this aspect to show whether such an IAG is cost-efficient. Moreover, the reviewer is also curious about whether the code of the proposed framework will be open-source to serve as the cornerstone for future research. Technical Quality: 4 Clarity: 4 Questions for Authors: Please refer to the Weaknesses section for more details. Confidence: 4 Soundness: 4 Presentation: 4 Contribution: 3 Limitations: N/A. Flag For Ethics Review: ['No ethics review needed.'] Rating: 7 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thank you for recognizing the significance of my work and for your constructive comments, which will help enhance the quality of my paper. --- > Weakness1: The primary concern is the cost of leveraging the search engine and LLM like ChatGPT. The author should elaborate more on this aspect to show whether such an IAG is cost-efficient. Response: Thanks for your suggestions. Here is a cost calculation for using our framework for inference. The main costs involved are generating queries with ChatGPT, calling Bing Text Search, and calling Bing Visual Search. The cost of ChatGPT is `$1.50` per 1M input tokens and `$3.00` per 1M output tokens. The cost of Bing Text Search and Bing Visual Search are `$0.025` and `$0.015` per search, respectively. For our framework, testing a VQA sample involves approximately 50/20 input/output tokens for ChatGPT, 3 Bing Text Searches and 1 Bing Visual Search. Therefore, the total cost for testing a VQA sample using our framework is approximately `$0.09`. --- > Weakness2: Moreover, the reviewer is also curious about whether the code of the proposed framework will be open-source to serve as the cornerstone for future research. Response: We will release the dataset and the code once the paper is accepted. --- Rebuttal 2: Comment: Thanks the author for the detailed response! The analysis of the cost for the search engine further increase the significance of the paper. I tend to keep my score as "Accept". --- Rebuttal Comment 2.1: Title: Thank you for your review! Comment: Thank you for your positive feedback and for noting the significance of the cost analysis in my paper. I appreciate your time and thoughtful evaluation.	Summary: The paper proposes a framework that enhances large vision-language models (LVLMs) to handle visual question answering (VQA) tasks involving up-to-date knowledge. The system utilizes a search engine to retrieve relevant websites and parses their content for more comprehensive information. To manage the large volume of data, a hierarchical filtering model is introduced, consisting of a website filter and a content filter. These filters score and select the most relevant information, which is then used to prompt LVLMs for improved performance on VQA tasks. The framework also includes a diversity selection mechanism to ensure varied and unbiased content selection. Strengths: 1. Clear Motivation: The paper is well-motivated and addresses a very practical problem. 2. Hierarchical Filtering Model: The two-step filtering process effectively manages large data volumes, enhancing the performance of LVLMs by ensuring that only the most relevant content is used. 3. Comprehensive Evaluation: The introduction of the UDK-VQA dataset and the experimental comparisons with LVLMs seem to to provide evidence of the framework’s effectiveness. Weaknesses: 1. "First open-source framework which seamlessly incorporates existing LVLMs to equip them with up-to-date knowledge during inference." I am concerned about this statement. There have been many frameworks utilizing external knowledge. Many of them can seamlessly apply to LVLMs. I am wondering what makes the proposed method different from them. 2. Lack of Comparison against Existing SOTA Methods Utilizing External Knowledge: I encourage the author to conduct experiments to include more SOTA methods directly utilizing retrieved external knowledge for clear comparison. 2. The author mentions the construction of the dataset as one of the contributions. However, I am concerned there is limited novelty in the construction pipeline of UDK-VQA as the steps are very engineering and practical. 3. The paper claims UDK-VQA to be the first test set for evaluating the ability of LVLMs in handling VQA about up-to-date knowledge. I have concerns about this. There have been plenty of VQA benchmarks requiring external knowledge in the field. I am wondering what really makes UDK-VQA different from the existing ones. Are the up-to-date knowledge? However, this terminology is very blurry. What is the definition of up-to-date knowledge? How do you define the timestamp threshold? 4. Broader Evaluation on Other Benchmarks Requiring External Knowledge: Extend the evaluation to include more diverse datasets and LVLMs to verify the generalizability of the framework across different benchmarks requiring external knowledge. 5. Performance Dependency: The framework’s performance is heavily dependent on the quality of the hierarchical filtering model, which may vary across different LVLMs and datasets. Technical Quality: 2 Clarity: 3 Questions for Authors: Please see above. Confidence: 4 Soundness: 2 Presentation: 3 Contribution: 2 Limitations: The authors have made a commendable effort to address the limitations of their work, including 1) Separate Training for Filters and 2) LVLMs and Content Snippet Limitations. Suggestions for Improvement: 1) Detailed Impact of Separate Training: The authors could provide more insights into how the separate training of filters and LVLMs specifically impacts performance and whether integrating these components could lead to substantial improvements. 2) Addressing Incomplete Snippets: While the paper acknowledges the limitation of incomplete snippets, it would be beneficial to discuss potential strategies to enhance snippet completeness or alternative methods to initially extract more comprehensive information. Flag For Ethics Review: ['No ethics review needed.'] Rating: 4 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thank you for taking the time to review our paper. We value your suggestions. Here, we would like to mention that our work focuses on multimodal Internet-Augmented Generation (IAG), a type of method to enhance LVLMs in handling multimodal prompts based on Internet knowledge. IAG is an emerging topic in the field of Retrieval-Augmented Generation (RAG), offering significant potential and application value. For example, OpenAI's recently revealed SearchGPT employs IAG as one of its main technologies. We have carefully addressed your comments, but due to the length, some responses have been placed in the comments section. --- > Weakness1: Differences from existing frameworks. Response: Our proposed framework significantly differs from existing works. Existing frameworks for LVLMs utilizing external knowledge can be categorized into two types: (1) Frameworks that retrieve from local knowledge bases. (2) Frameworks based on external tools. The first type of framework [2][6] often retrieves external commonsense knowledge, such as "the Eiffel Tower is in Paris" and "the important function of a folding chair is portability". This type of knowledge is not frequently updated and can be easily embedded in LVLMs during training. Differently, we focus on the up-to-date knowledge, which is frequently updated over time, such as "how many gold medals did the USA win in the Olympics on August 6, 2024.". LVLMs cannot acquire this type of knowledge during the training phase. The second type of framework [7][8] uses search engines to obtain the up-to-date knowledge knowledge, but they focus on the orchestration of different tools. Differently, we focus on how to retrieve and use the obtained knowledge to augment LVLMs, which is orthogonal to the second type of framework and our method can be embedded into their framework. --- > Weakness2: Lack of comparison against existing SOTA methods utilizing external knowledge. Response: Thanks for your suggestions. We immediately conducted experiments using more methods: GENREAD [6], Chameleon [7], and CLIP→FID [2]. GENREAD uses LLMs for knowledge retrieval and cannot access the up-to-date knowledge. Chameleon is an external tool-based framework that uses snippets returned by search engines for augmentation. CLIP→FID is a framework based on local database retrieval, and we integrated it with some components of our framework to allow it to access the up-to-date knowledge. The experimental results are shown in the table below, where QG denotes Query Generator and HFM denotes Hierarchical Filtering Model. We can observe that our framework achieves the best performance, proving the effectiveness of our approach. If you have any suggestions for frameworks to compare, please let us know. We look forward to discussing them with you. \| Baseline \| Variant \| QG (Q) \| QG (V) \| HFM \| UDKAG \| \|------------\|---------------\|----------\|----------\|-------\|--------\| \| LLaVA 1.6 \| Raw \| - \| - \| - \| 31.8 \| \| \| GENREAD [6] \| - \| - \| - \| 31.3 \| \| \| Chameleon [7] \| - \| - \| - \| 62.3 \| \| \| CLIP→FID [2] \| ✓ \| ✓ \| - \| 54.7 \| \| \| Ours \| ✓ \| ✓ \| ✓ \| 90.2 \| \| InternVL 1.5 \| Raw \| - \| - \| - \| 42.6 \| \| \| GENREAD [6] \| - \| - \| - \| 27.6 \| \| \| Chameleon [7] \| - \| - \| - \| 59.3 \| \| \| CLIP→FID [2] \| ✓ \| ✓ \| - \| 57.7 \| \| \| Ours \| ✓ \| ✓ \| ✓ \| 92.9 \| \| Qwen-VL \| Raw \| - \| - \| - \| 35.2 \| \| \| GENREAD [6] \| - \| - \| - \| 23.5 \| \| \| Chameleon [7] \| - \| - \| - \| 43.8 \| \| \| CLIP→FID [2] \| ✓ \| ✓ \| - \| 36.7 \| \| \| Ours \| ✓ \| ✓ \| ✓ \| 84.8 \| \| LLaVA 1.5 \| Raw \| - \| - \| - \| 41.2 \| \| \| GENREAD [6] \| - \| - \| - \| 31.9 \| \| \| Chameleon [7] \| - \| - \| - \| 60.7 \| \| \| CLIP→FID [2] \| ✓ \| ✓ \| - \| 58.5 \| \| \| Ours \| ✓ \| ✓ \| ✓ \| 88.9 \| --- > Weakness3: Limited novelty in the construction pipeline of UDK-VQA. Response: (1) Automatically generating high-quality internet VQA samples is challenging. We employed several strategies to improve the quality of the generated data, resulting in VQA samples that can be used for training. The strategies include using an entity replacement strategy to avoid generating samples with meaningless images, and using consistency answering (as described in Section 4.2, second paragraph) and image clustering to automatically filter out incorrect samples. (2) With our pipeline, it is easy to collect a large number of VQA samples for training and regularly generate test data that is free from data contamination issues. --- Rebuttal 2: Title: Response to Reviewer qhJo Comment: > Weakness4: Differences between UDK-VQA and existing datasets. Definition of up-to-date knowledge and timestamp threshold. Response: The difference between our UDK-VQA dataset and existing datasets lies in the timeliness of the test data. Existing RAG datasets rely on external knowledge that is usually commonsense knowledge, which is not time-sensitive and does not update frequently. Such knowledge can be learned by LVLMs during pre-training and supervised fine-tuning. In contrast, the knowledge our dataset relies on is time-sensitive and belongs to the up-to-date knowledge, which refers to the most current and relevant information, data, and understanding available at a given point in time. Specifically, this refers to information that emerged after the LVLMs completed their training, such as news generated after a certain date. We set the timestamps for data collection to be after the latest release date of all LVLMs we tested, ensuring that the knowledge required by the test set was not accessible to LVLMs during their training. This work is a long-term project, and we will continue to update the test set automatically using our framework to ensure the data is new for LVLMs. > Weakness5: Broader evaluation on other benchmarks requiring external knowledge. Response: Thanks for your suggestions. We have evaluated different LVLMs on a broader benchmark, including GQA [1], INFOSEEK [2], A-OKVQA [3] and the proposed UDKAG. The experimental results are shown in the table below, where GQA does Not Rely on external Knowledge (NRK), INFOSEEK and A-OKVQA Rely on Commonsense Knowledge (RCK), and UDKAG Relies on the Up-to-Date Knowledge (RUDK). From the table, we can observe that our framework improves the performance of different LVLMs across various datasets. The improvements on these three datasets are not as significant as on our UDKAG dataset for the following reasons: (1) GQA does not rely on external knowledge and is used to evaluate the reasoning ability of LVLMs, which is beyond the scope of our framework. (2) Our framework focuses on retrieving the up-to-date knowledge, whereas INFOSEEK and A-OKVQA rely on commonsense knowledge, much of which has already been used in the training data of LVLMs. \| Baseline \| Variant \| Local Data (clear GT) \| Internet Data (unclear GT) \| GQA [1] (NRK) \| INFOSEEK [2] (RCK) \| A-OKVQA [3] (RCK) \| UDKAG (RUDK) \| \|---------------\|-----------\|-----------------------\|----------------------------\|---------------\|--------------------\|-------------------\|--------------\| \| CFR [5] \| - \| - \| - \| 72.10 \| - \| - \| - \| \| Oracle→FID [2]\| - \| ✓ \| - \| - \| 45.60 \| - \| - \| \| Omni-SMoLA [4]\| - \| ✓ \| - \| - \| - \| 84.10 \| - \| \| LLaVA-1.6 \| Raw \| - \| - \| 61.66 \| 37.86 \| 75.53 \| 31.8 \| \| \| Ours \| - \| ✓ \| 62.33 \| 41.25 \| 76.22 \| 90.2 \| \| InternVL-1.5 \| Raw \| - \| - \| 74.03 \| 51.13 \| 84.53 \| 42.6 \| \| \| Ours \| - \| ✓ \| 74.41 \| 53.10 \| 84.59 \| 92.9 \| --- > Weakness6: Performance dependency on the hierarchical filtering model. Response: Our experimental results have proven that: (1) The hierarchical filtering model has transferability and can be applied to different LVLMs and datasets without fine-tuning. (2) The quality of the hierarchical filtering model is not significantly related to the backbone it uses. As shown in Table 1 and Figure 4 of the main manuscript, we tested our hierarchical filtering model on 13 different LVLMs, all of which showed significant improvement. As described in our response to weakness 5, our hierarchical filtering model can also bring some improvement when directly transferred to other datasets, whether they are based on additional knowledge or not. Furthermore, as shown in Table 2 of the main manuscript, using LLaVA and Qwen as the backbone for the hierarchical filtering model can achieve significant performance improvements. --- Rebuttal 3: Title: Response to Reviewer qhJo Comment: > Limitation1: Detailed impact of separate training. Response: Separate training is necessary as joint training results in performance degradation. As shown in the table below, whether using Qwen-VL or LLaVA-1.5 as the baseline model, separate training brings more significant performance improvement. The main reasons are: (1) Our training data uses pseudo-labeling instead of high-quality human annotation. Training based on such data may cause LVLMs to lose their original semantic understanding capabilities. (2) Our training and testing sets are generated from news from different time periods, involving different entities and having different distributions. Training LVLMs on our training set easily leads to overfitting, resulting in lower generalization on the test set. \| Baseline \| Variant \| Training Strategy \| UDKAG \| \|-----------\|--------------\|--------------------\|-------\| \| QWen-VL \| Raw \| - \| 35.2 \| \| \| Ours \| Joint Training \| 68.5 \| \| \| Ours \| Separate Training \| 84.8 \| \| LLaVA 1.5 \| Raw \| - \| 41.2 \| \| \| Ours \| Joint Training \| 68.0 \| \| \| Ours \| Separate Training \| 88.9 \| --- > Limitation2: Addressing incomplete snippets. Response: Thanks for your suggestions. The completeness of snippets has little impact on accuracy because their purpose is to allow the website filter to perform an initial screening of web pages, reducing the input the content filter needs to process, thereby improving the efficiency of our framework. Directly using incomplete snippets with existing search engines is already widely practiced [7][8]. There are two intuitive strategies to enhance the completeness of snippets: (1) Using large language models (LLMs) to continue writing the snippets. (2) Crawling the complete content of websites to fill in the snippets. The first method may result in incorrect snippets because LLMs lack the up-to-date knowledge and cannot accurately continue the writing. The second method will incur additional time consumption during inference because it requires crawling the content of all websites, regardless of whether the content filter needs to process that content. This contradicts our original intention of improving efficiency with the website filter. We are implementing the second method to verify whether complete snippets will bring performance improvements, and the experimental results will be provided in the discussion phase. --- [1] Hudson et al. Gqa: A new dataset for real-world visual reasoning and compositional question answering. In CVPR 2019. [2] Chen et al. Can pre-trained vision and language models answer visual information-seeking questions? In EMNLP 2023. [3] Schwenk et al. A-okvqa: A benchmark for visual question answering using world knowledge. In ECCV 2022. [4] Wu et al. Omni-SMoLA: Boosting generalist multimodal models with soft mixture of low-rank experts. In CVPR 2024. [5] Nguyen et al. Coarse-to-fine reasoning for visual question answering. In CVPRW 2022. [6] Yu et al. Generate rather than retrieve: Large language models are strong context generators. In ICLR 2023. [7] Lu et al. Chameleon: Plug-and-play compositional reasoning with large language models. In NeurIPS 2023. [8] Yang et al. Mm-react: Prompting chatgpt for multimodal reasoning and action. In arXiv 2023. --- Rebuttal 4: Title: Message to Reviewer qhJo Comment: Dear Reviewer qhJo, We hope this message finds you well. We sincerely appreciate the time and effort you are dedicating to evaluating our submission. As the conference timeline approaches important deadlines, we would be grateful for any updates you could provide regarding the review status of our paper. If there is anything further we can provide to assist with the review process, please don’t hesitate to let us know. Additionally, we present the experimental results of snippet completeness on performance here, where $\theta$ represents the percentage as mentioned in Section 5.6 of the main manuscript. For each website snippet, we attempted to locate the full sentence corresponding to the snippet by crawling the website's content. However, the content of many websites could not be crawled. For such websites, we experimented with two strategies: (1) Discarding these websites during training and testing. (2) Using the incomplete snippets. The experimental results are shown in the table below, where "Raw" represents all snippets without completion, "Discard" represents strategy (1), and "Mixture" represents strategy (2). The experimental results show that directly discarding the websites leads to a significant performance loss, as discarding reduces the number of usable websites by approximately half, thereby limiting the performance of the website filter. As $\theta$ increases, the performance of strategy (2) becomes increasingly close to that of all snippets without completion, which validates our rebuttal argument that the completeness of the snippets has little impact on accuracy. We greatly appreciate your attention to this matter and look forward to any feedback you may have. Best regards, Authors --- \| Baseline \| Variant \| $\theta$=10\% \| $\theta$=25\% \| $\theta$=40\% \| $\theta$=55\% \| $\theta$=70\% \| $\theta$=85\% \| $\theta$=100\% \| \|-----------\|----------\|-------------\|-------------\|-------------\|-------------\|-------------\|-------------\|--------------\| \| LLaVA-1.6 \| Raw \| 80.8 \| 86.1 \| 88.4 \| 89.4 \| 89.7 \| 90.0 \| 90.2 \| \| \| Discard \| 76.6 \| 80.2 \| 80.7 \| 82.1 \| 81.9 \| 81.3 \| 81.6 \| \| \| Mixture \| 83.6 \| 87.9 \| 89.4 \| 89.6 \| 89.8 \| 90.4 \| 90.2 \| \| InternVL-1.5 \| Raw \| 84.6 \| 89.2 \| 91.1 \| 92.9 \| 92.4 \| 92.7 \| 92.9 \| \| \| Discard \| 79.7 \| 82.2 \| 82.9 \| 83.8 \| 84.0 \| 82.8 \| 83.2 \| \| \| Mixture \| 86.0 \| 89.5 \| 92.2 \| 92.3 \| 92.2 \| 92.5 \| 92.9 \| --- Rebuttal 5: Title: Hoping for further discussion with you. Comment: Dear Reviewer qhJo, We express gratitude for your time spent on reviewing and your valuable comments. We have addressed your concerns by providing relevant responses and results. We look forward to engaging in further discussion to confirm whether or not your concerns have been addressed. Best regards, Authors --- Rebuttal 6: Title: Looking Forward to Your Reply. Comment: Dear Reviewer qhJo, I hope this message finds you well. We are truly grateful for the effort you have put into reviewing our manuscript and for the insightful feedback you have provided. We understand that you may have a busy schedule, but we kindly ask if you could review our responses at your earliest convenience. Your feedback is essential for us to move forward, and we would greatly value any additional insights you may have. Thank you once again for your time and consideration. Best regards, Authors	null	null	Rebuttal 1: Rebuttal: ### General response We would like to thank the Area Chairs and the Reviewers for carefully reading our paper and providing valuable comments. We are pleased to hear that the majority of the reviewers found our paper "well-written " (jo4p) and "strongly motivated" (qhJo, jo4p). We also appreciate the reviewers' recognition of the "novelty" (jo4p, PaHW) and the "universality and significance" (jo4p) of our work. Additionally, we are glad to receive positive feedback on our experiments, which "provide evidence of the framework's effectiveness" (qhJo), "verify further the effectiveness of the proposed framework" (jo4p), and show "significant performance improvement" (PaHW). We look forward to the upcoming discussion sessions and to making further contributions to the vision-and-language community.	NeurIPS_2024_submissions_huggingface	2,024	null	null	null	null	null	null	null	null
Pedestrian-Centric 3D Pre-collision Pose and Shape Estimation from Dashcam Perspective	Accept (poster)	Summary: This paper proposed a dataset (PVCP) followed by a solution (PPSENeT), which is the first one that focuses on the 3D pre-collision pose of pedestrians. The dataset is collected by dashcam that contains various poses with both static and dynamic backgrounds. Various types of annotations are provided, including 2D, 3D, and human mesh, leading to diverse evaluation protocols. Following the dataset, the method (PPSENet) is a top-down method, that first crop the human subjects with a bounding box and then predict the 2D poses. The 2D pose will be lifted to a 3D mesh with the help of iterative regression and a novel pose class loss, following MotionBERT backbone. All combined, the method achieves SOTA performance on the proposed PVCP dataset. Strengths: 1. The proposed dataset contains various types of annotations, including 2D, 3D, and human mesh. This contains abundant ways of evaluation for every method. 2. The topic is in general interesting. It is a very good intention to expand the scope of the daily human pose to the pose in a specific setting. Especially, in such a setting related to the safety of daily life. 3. The paper is generally good-written and easy to follow. Weaknesses: The major problem addressed by the paper is somewhat trivial to me. Typically, there are two logical holes that the author fails to fulfill before they propose this dataset and solutions. (1). Are the pre-collision poses really so different from normal or daily poses? This is intuitively not correct in my perspective. The pre-collision pose is a subset of the daily pose. In this spirit, the model trained on a large human pose set (such as COCO), should work decently on the pre-collision pose. The pose such as running or avoiding, are covered in COCO (as COCO also contains various). COCO has even more diverse scenes (indoor and outdoor), while PVCP presumably only contains outdoor scenes (as it is for pre-collision poses). In this case, authors should give solid evidence to support their claim. Moreover, in the author's Fig1(c), the proposed dataset PVCP, seems even more general than COCO, which is contradictory to my intuition. (2). Are the pre-collision poses really an important factor in terms of avoiding collision? There are many factors that could influence the rate of avoiding collision, such as the speed of vehicles or the light conditions as the authors also mentioned in the appendix. Are there any experimental results to support having the pre-collision poses will significantly reduce the rate of accidents? If the above two questions are not answered properly, then it seems this problem is not really a significant problem to be investigated, which will reduce the necessity of the proposed dataset and method. The proposed method, on the other hand, follows a top-down paradigm, which is known as not compatible with real-time applications, not to mention that architecture is used to produce 3D mesh. However, the pre-collision pose estimation definitely requires real-time processing. Technical Quality: 3 Clarity: 2 Questions for Authors: 1. What is the processing time (inference time) of the proposed method? Is it a real-time method? 2. What is the performance of the method only trained on COCO or another large human pose dataset without finetuning on the PVCP? For the other two questions, please see the weaknesses. These questions are the key to changing the rating. Confidence: 4 Soundness: 3 Presentation: 2 Contribution: 2 Limitations: No negative societal impact is found. Flag For Ethics Review: ['No ethics review needed.'] Rating: 5 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thank you very much for your positive feedback and valuable comments on our dataset and paper writing. Thank you very much for your careful review, we will reply to your opinions on Questions and Limitations respectively. ------ Questions Reply: Reply1. We fully understand your concern about differences in data sets. Indeed, in a broad sense, the pre-collision pose is a subset of everyday poses, since similar poses can always be found in everyday life. However, PVCP is significantly different from other pose datasets (such as COCO) in the aspects of scene specificity, action space difference and time series continuity. Here's a breakdown of these differences: Scene specificity: PVCP focuses on the dashcam view, capturing pedestrians' emergency behaviors during vehicle movement, such as sudden road crossings or running into the street, against a dynamic traffic background and changing observation field. This is crucial for road safety. COCO, while including some road scenes, covers a wide range of everyday human movements that aren't specifically related to potential vehicle collisions. Action space difference: Pedestrian pre-collisions in the PVCP's dashcam involve rapid, sudden motions like accelerating, swerving, or jumping to avoid a vehicle. These actions are highly dynamic and uncertain in time and space, making detection accuracy crucial. The PVCP dataset categorizes these pre-collision poses into four types, and PPSENet uses pose category loss to learn their spatial differences. In contrast, COCO's common poses cover broader, more static motions. While some include movement, they aren't specific to emergency vehicle collision situations, making their spatial distribution and dynamics more stable and predictable. Time series continuity: In traffic recordings, a pedestrian's pre-collision poses are closely tied to the vehicle's time series. These poses occur in quick succession as the vehicle approaches and is about to collide, showing strong time sensitivity and continuity. Algorithms must capture and analyze these pose variations from a continuous context to provide timely pre-pose estimates and warnings. In contrast, the COCO dataset's poses do not emphasize time series continuity. Static poses in single images cannot reflect time-sensitive changes in human movements, especially those related to emergency vehicle collisions, making COCO poses relatively loose and independent in time series. The attached PDF compares the visual differences between PVCP and other datasets (see Figure R1). ------ Reply2. Indeed, there are many factors affecting pedestrian collision injury: pre-collision pose, pedestrian position, vehicle speed, vehicle front end shape and so on. Each factor has a lot of work to do to analyze their impact on injury in accidents. Our work focuses on estimating the most uncertain pedestrian precollision posture to help in accident reconstruction and pedestrian injury assessment. Many studies have demonstrated the impact of pedestrian's initial posture on vehicle-vehicle collision. Paper [1] showed "Head impact locations on the vehicle are predictable but their severity varies with pre-impact stance by up to 30%”. Paper[2] showed higher sensitivities to the pedestrian posture and its relative position with respect to the vehicle than to the vehicle speed for the chosen design space. Paper [3] shows the pedestrian’s rotation is highly influenced by leg and arm posture, which makes predictions of head impact risk in secondary impact difficult. Paper [4] investigate the effects of preimpact pedestrian posture on the loading and the kinematics of the lower extremity when struck laterally by vehicle. The experiment showed that "the walking posture increased the injury risk of soft connection tissue about 20-30% and reduced the internal force.” paper[5] showed that pedestrian injuries differ in different gait serials. So do different gaits in the same gait serial. Paper [6] states that "it has been demonstrated that resulting impact kinematics and dynamics from vehicle–cyclist/pedestrian collisions are influenced significantly by initially specified human posture and motion”. This might be briefly described in the "Related Work" section, we will expand on this in the final version. References > [1] Effects of pre-impact pedestrian position and motion on kinematics and injuries from vehicle and ground contact. > > [2] Crash reconstruction of pedestrian accidents using optimization techniques. > > [3] Vehicle related influence of post-car impact pedestrian kinematics on secondary impact. > > [4] Influence of pre-impact pedestrian posture on lower extremity kinematics in vehicle collisions. > > [5] Pedestrian gaits observed from actual pedestrian-vehicle collisions. > > [6] Forward dynamics computational modelling of a cyclist fall with the inclusion of protective response using deep learning-based human pose estimation. ------ Limitations Reply: Reply1. Our method is not real-time at present, because our input is Image and pre-selected Bbox sequence of collision pedestrian targets, the purpose of our work is to provide pose data support for the study of vehicle active and passive protection system, so as to facilitate subsequent accident reconstruction, pedestrian injury assessment and vehicle structural design. Reply2. We trained on the COCO and PW3D dataset respectively and validated on the PVCP testset. Because these datasets do not have pose category annotations, we set them uniformly to a single category. The results, recorded in the submitted PDF, show that training only on another large human pose dataset without fine-tuning for PVCP does not perform well on PVCP this particular pose dataset (see Table R1) . ------ Thanks again for your comments. We will address all suggestions in the final revision and supplemental material. --- Rebuttal Comment 1.1: Comment: Thank you for your detailed response, which has effectively addressed most of my concerns. However, I have an additional question regarding the emphasis on scene-specific factors. I disagree that scene specificity is a critical factor justifying the necessity of this dataset. In the context of human pose estimation as an example, the primary factors should be the visual features of the human body, such as the shape of the hand or other general anatomical structures. Logically, these features should be independent of scene information. For example, if a person running in an open field (as seen in many COCO images) can be accurately detected by a method trained on COCO, it stands to reason that the same method should also successfully detect the person when they are running in front of a car, provided the pose remains the same or similar. I would appreciate further clarification on this point. Regarding the final rating, I recognize and commend the efforts made in collecting this dataset, as it is a valuable contribution. Therefore, I am inclined to support the acceptance of this paper. However, I may not rate it as highly as reviewer 9csR, as I believe the overall quality does not fully meet the threshold for acceptance or higher. --- Rebuttal 2: Comment: Thank you very much for your insightful feedback and invaluable support regarding our paper, especially your perspective on scene-specific factors in human pose estimation. We fully appreciate your viewpoint that, logically, pose features should be independent of scene information. However, upon delving deeper into the literature, we have encountered studies that present a contrasting view, highlighting the extent to which pose features can be constrained by scene information, particularly given the limitations inherent in datasets like COCO (as exemplified in paper [3] below). Our explanations are as follows: ------ Our PVCP dataset has two core characteristics: (1) dynamic perspective of the dashcam; (2) pedestrian emergency posture in pedestrian-vehicle collision scene. In Paper[1], within the section titled `I. INTRODUCTION -> B. Overview of Deep Learning Framework for MHPE`, it is stated that "The human body is nonrigid and flexible for high degreeof-freedom poses, therefore, predicting human pose estimation from a monocular camera faces many challenges, such as complex or strange posture, person-object/person-person interaction or occlusion, and crowded scenes, etc. Different camera views and complex scenes will also introduce problems of truncation, image blur, low resolution, and small target persons". (Our PVCP dataset encompasses a wide range of scenarios that are not present in similar datasets like COCO, including pedestrians occluded by vehicles, dynamically approaching pedestrian targets from far to near, and varying resolutions.) In paper[2], within the section titled `7.3.2 Benchmark Leaderboards -> Beyond Fully Supervised Learning` , it is stated that "Building 3D human mesh datasets is time-consuming and of high cost. A MoCap system needs to be set up beforehand. After capturing, the cleaning and annotation process of raw 3D data is highly demanding. Besides, 3D datasets lack diversity in human motion and background, but 2D datasets...". (This means that background information on the different scene pose datasets is important.) In paper[3], within the section titled `3.3 2D HPE Summary` , it is stated that "Another challenge lies in the limited data for rare poses. Although the size of current datasets for 2D HPE is large enough (e.g., COCO dataset [108]) for the normal pose estimation (e.g., standing, walking, running), these datasets have limited training data for unusual poses, e.g., falling. The data imbalance may cause model bias, resulting in poor performance on those poses. It would be useful to develop effective data generation or augmentation techniques to generate extra pose data for training more robust models." (This section clearly demonstrates the shortcomings of the COCO dataset on rare gestures such as falls.) In paper[4], within the section titled `2.2 Dataset Structure -> Additional Mixed Reality Test Data` , it is stated that "Real images may show people in complex poses, but the diverse backgrounds as well as the scene illumination and the occlusions can vary independently and represent important nuisance factors the vision systems should be robust against." (This indicates that background and illumination in different scenes affect the ability of the network to estimate the pose.) In paper [5], within the section titled `8.Conclusions and Discussions`, it is stated that "Deep learning based HPE methods rely heavily on labelled data with specific characteristics. For example, the MHP dataset covers daily activities, while the LSP dataset focuses on the sports scene. The model trained on one dataset may perform badly on another dataset ." (This comparison is similar to COCO's mostly everyday pose scene and PVCP's mostly pre-collision pose scene.) ------ References > [1] Liu, W., Bao, Q., Sun, Y., & Mei, T. (2022). Recent advances of monocular 2d and 3d human pose estimation: A deep learning perspective. ACM Computing Surveys, 55(4), 1-41. > > [2] Tian, Y., Zhang, H., Liu, Y., & Wang, L. (2023). Recovering 3d human mesh from monocular images: A survey. IEEE transactions on pattern analysis and machine intelligence. > > [3] Zheng, C., Wu, W., Chen, C., Yang, T., Zhu, S., Shen, J., ... & Shah, M. (2023). Deep learning-based human pose estimation: A survey. ACM Computing Surveys, 56(1), 1-37. > > [4] Ionescu, C., Papava, D., Olaru, V., & Sminchisescu, C. (2013). Human3. 6m: Large scale datasets and predictive methods for 3d human sensing in natural environments. IEEE transactions on pattern analysis and machine intelligence, 36(7), 1325-1339. > > [5] Zhang, F., Zhu, X., & Wang, C. (2021). Single person pose estimation: a survey. arXiv preprint arXiv:2109.10056. ------ Thanks again for your comments. We hope that our reply can solve your questions, if you have new questions about our reply, you are welcome to put forward further comments, we will actively reply and try to correct.	Summary: The paper conducts the first Pedestrian-Vehicle Collision Pose dataset(PVCP) including the pedestrian-vehicle collision pose from the dashcam perspective and pose annotations. Moreover, the paper presents the method(PPSENet) to estimate the collision pose and shape based on PVCP. Strengths: 1. The first Pedestrian-Vehicle Collision Pose dataset, annotated by algorithm and manual method. 2. The paper proposes the framework(PPSENet) to estimate the 2D pose and lift it to the 3D pose. The method achieves SOTA in PVCP Weaknesses: Because the dataset annotation largely depends on manual methods. The paper may need more discussions or a more convincing evaluation, such as more visualization of the dataset or a video. Technical Quality: 3 Clarity: 3 Questions for Authors: Will the SMPL annotation tool be made publicly available? Confidence: 3 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: The author addresses the limitations in the apendix. Flag For Ethics Review: ['Ethics review needed: Research involving human subjects', 'Ethics review needed: Data privacy, copyright, and consent'] Rating: 5 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thank you very much for your recognition of our work, especially the PVCP dataset and annotation tools. Thank you very much for your careful review. ------ Indeed, our annotation tool has already been made accessible to the public, but due to anonymity requirements, we have withheld the full GitHub link. We will include the complete link in the final version of our paper, so that more researchers can utilize and improve upon our work. In this rebuttal, we have submitted an anonymous link to the AC, which provides access to all source code, datasets, annotation tools, and additional visualization videos (including dataset videos and experimental visualization results). We will also make the papers publicly available once they are fortunate enough to be accepted for publication. The PVCP dataset is a novel pedestrian pre-collision pose dataset, and we believe it holds significant value for studying pedestrian behavior and enhancing pedestrian safety. Given the constraints of traffic scenarios and video sources, we had to rely on manual annotation for human pose data. Fortunately, we were able to leverage the support of many previous excellent open-source efforts to accomplish this task. We will continue to refine and enhance the PVCP dataset and our algorithms in order to achieve even greater progress. ------ Thanks again for your comments. We will address all suggestions in the final revision, and some additional notes will also be extended to the newly submitted supplemental material.	Summary: This paper collects a new dataset that includes pedestrian-vehicle collision poses from the dashcam perspective and presents a two-stage framework for estimating human pose and shape parameters. The dataset has the potential to benefit the study of pedestrian injuries. Given the challenge of obtaining ground truth pose and shape parameters in the wild, the authors employ the two-stage framework to generate initial results. These are then refined through manual adjustment using their annotation tool. Strengths: This paper makes a significant contribution by collecting real-world pedestrian pose and shape parameters before collisions. The datasets include useful modalities such as bounding boxes and tracking information. Since it is difficult to obtain human poses in uncontrolled environments, the authors collect online dashcam videos and introduce a straightforward framework to estimate initial results. This framework iteratively regresses the pose and shape parameters. Weaknesses: Since the dataset uses videos collected from the Internet, it is difficult to obtain the ground truth pose and shape parameters. The accuracy of a dataset is important. The paper could elaborate more on the procedures to ensure accuracy such as how many annotators are recruited, how they are trained before the annotation, whether there is a post-processing stage, etc. Technical Quality: 3 Clarity: 3 Questions for Authors: 1. How does the author ensure the motion smoothness of results during the manual annotation refinement stage? 2. How many annotators are assigned to each video clip? 3. How annotators are recruited and trained to ensure accuracy? Confidence: 4 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: The dataset is relatively small in size and no camera parameters are provided. Accurate vehicle speed and the global position of pedestrians could also be helpful for collision damage analysis. Flag For Ethics Review: ['No ethics review needed.'] Rating: 7 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thank you very much for your recognition of our work in pedestrian injury research. Thank you very much for your careful review, and we will reply to your opinions on Questions and Limitations respectively. ------ ##### Questions Reply: Reply1. Unlike regular and periodic postures, pre-collision postures exhibit strong specificity and abrupt changes. After manual annotation, we performed simple smoothing operations (such as the moving average method), but for pre-collision postures with particularly large variations, the smoothed postures significantly differed from the GT (ground truth) postures. Therefore, to ensure the accuracy of joint positions, we retained the manually adjusted annotations. In a separately submitted PDF, we show an example of a 2D posture sequence with smoothing operations to support our claim (see Figure R2.) ------ Reply2. In order to obtain relatively accurate pose annotation data while controlling the cost of dataset production, after clipping the videos into independent segments, we arrange two annotators for each video and conduct effective annotation through the following three steps: - 1） Firstly, we assign one annotator to perform the initial annotation for each video, ensuring that the SMPL human pose aligns as closely as possible with the task contour of the background image; - 2） After the initial annotation is completed, the second annotator reviews the annotations, checking for omissions, errors, or incomplete alignments; - 3） Finally, any discrepant labels are sent back to the first annotator for re-annotation to ensure correctness. This process is repeated until there are no more discrepancies in all annotations. In this way, we obtain visually consistent SMPL pose annotations that closely match the images. Subsequently, the annotated pose data is sent for subsequent post-processing. ------ Reply3. Our annotation tool is designed for ease of use, requiring minimal to no training. Simply adjust the sliders for pose and shape parameters of the corresponding joints based on the interface prompts, aligning the SMPL pose with the background pedestrian silhouette. The annotated files are automatically saved, and we will subsequently conduct unified data post-processing. To aid in comprehension, we have included an anonymous link in this Rebuttal submission to the AC, showcasing a simple annotation process video. ------ Limitations Reply: Indeed, our dataset lacks camera parameters due to the fact that real-world vehicle-pedestrian collisions are accidental events that are difficult to systematically record. Most of our data is sourced from online platforms, which is one of the challenges and limitations of related datasets. Additionally, accurate vehicle speed and the global position of pedestrians could be valuable for collision damage analysis. In our appendix (A2.2), we have proposed a simple solution to annotate this information: > A.2.2 Vehicle Annotations > > Distance between the pedestrian and the vehicle, as well as the vehicle's speed, are crucial factors in collision accidents. In real traffic scenarios, these parameters can often be acquired through the sensors equipped on the vehicle. However, in our collected accident videos, obtaining the distance and speed through sensors is not feasible (as these data are not provided). Therefore, we refer to RootNet [67] to obtain the distance between the pedestrian and the vehicle as our dataset's distance annotation. Similarly, vehicles are often fitted with additional sensors to get accurate speed, but for our crash video dataset it's not perfect and the speed information is missing. Therefore, we use the estimated pedestrian-vehicle distance and video frame rate to obtain an approximate speed label. ------ Thanks again for your comments. We will address all suggestions in the final revision, and some additional notes will also be extended to the newly submitted supplemental material. --- Rebuttal Comment 1.1: Title: Reply from Reviewer Comment: I appreciate the authors' comprehensive rebuttal. However, the link to the annotation process video appears to be invalid. Could the authors please provide a new link? --- Rebuttal 2: Title: New video link. Comment: We are sorry that you cannot view the link we provided. We used the website [Anonymous GitHub] to upload our code and videos. If you can't watch them, you can go to [Anonymous GitHub] website through the link (https://anonymous.4open.science/r/SMPL_Tools-0C7A/README.assets/smpl_tools_tutor.mp4) provided by us and click `Download file` in the upper right corner to download and watch it locally. If none of the above work, we temporarily registered a Youtube account without personal information to upload the video, you can click the link below to watch the Annotation process video we provide. (Note: you can turn down the volume of your device, as the video may be a bit noisy) Link: https://youtu.be/EBOuW21z0v0 If none of the above methods can browse the videos we provide, you are welcome to communicate with us in time, we will return to a more efficient way. Thank you for your review. --- Rebuttal Comment 2.1: Title: Reply from Reviewer Comment: Thank you for preparing a thorough rebuttal that addresses my questions. The provided material offers sufficient proof of the smoothing process and includes a user guide for the SMPL annotation tool. While my concerns have been addressed, I believe this paper has the potential to make a significant impact on downstream tasks such as autonomous driving. I appreciate the authors' efforts to contribute to this field. I have also updated the score accordingly. --- Reply to Comment 2.1.1: Comment: Thank you for carefully reviewing our work and providing valuable feedback to make our work more complete. We are pleased to know that our response to your question adequately addressed your concerns and that you approved the material we provided. We are also very grateful for your recognition of our work, and we are pleased that you see this paper as having the potential to have an important impact in downstream tasks such as autonomous driving. We will continue to work hard to make more contributions to this field. Thank you again for your time and recognition of our work.	Summary: This work focuses on the pre-collision posture of pedestrians in real traffic scenarios. The authors semi-automatically constructed the first pedestrian-vehicle collision pose dataset PVCP by collecting pedestrian-vehicle collision poses from the perspective of dashcams, which includes more than 40,000 accident frames and more than 20,000 pedestrian pre-collision pose annotations. Furthermore, this work constructs a pedestrian pre-collision pose estimation method(PPSENet) to estimate the collision pose and morphological sequence of pedestrians from pedestrian-vehicle accident videos. PPSENet first estimates the 2D pose from the image (image to pose, ITP), and then promotes the 2D pose to a 3D mesh (pose to mesh, PTM). Experiments show the relative advantages of PPSENet. Strengths: 1. I consider the dataset collected in this work is meaningful. It can provide a well guide for many autonomous driving methods. Although the pose before the collision may not even last for a second, the human pose data at this moment is indeed worth studying. 2. The author designed an SMPL annotation tool to align the initial results with the image contour. The PVCP dataset finally obtained about 40,000 accident images and more than 20,000 2D and 3D annotations of pedestrian emergency poses. Weaknesses: 1. I consider the PPSET method is the weak point of this work. From a methodological perspective, except for the robust 2D pose estimation method in the first stage, the second stage is almost similar to the backbone network of MotionBERT. There is almost no innovation. From an experimental perspective, the results of 2D pose sequences in Table 2 are really poor. Although PVCP is a brand new dataset for pre-collision human poses, the network input is sequences, and the overall 2D joint index, MPJPE is greater than 100. In the results of Table 5, Pose2Mesh performs better than MotionBERT without being trained on PVCP. I highly doubt this experimental result. However, PARE, a one-stage method that has not been retrained, can outperform PPSET in the first four indicators. We know that PARE is from the work in 2021. From this result, I think PPSET is completely at a disadvantage. 2. Although PVCP has collected a relatively large number of pedestrian-vehicle collision pose datasets, it is not the first one. [1] As early as 2020, there was a description of the relevant dataset, and the video in it is the same as PVCP. Although I saw the relevant reference, it was glossed over. The statement in line 110 of Related Work should be modified. 3. Although I think the annotation application of SMPL is an innovation, the author also cleverly hid the relevant github content. However, the relevant GUI interface had similar content as early as 2020 [2]. Typo: 1. The two time spans in Year in Table 1 should be based on "2017" instead of "1017" 2. The use of "impractical" in line 46 is easy to cause ambiguity. Because PVCP is doing such content. Technical Quality: 2 Clarity: 3 Questions for Authors: I hope the author will emphasize the breakthrough and innovation brought by the proposed method PPSET. I think the starting point of the dataset is very good, and the content of the dataset is also very good. But I think the corresponding method should be equally good to meet the standards of NeurIPS. Confidence: 4 Soundness: 2 Presentation: 3 Contribution: 3 Limitations: See Weakness. Flag For Ethics Review: ['Ethics review needed: Data privacy, copyright, and consent'] Rating: 5 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thank you very much for your recognition of our work and especially the significance of the PVCP dataset. And thank you very much for your review. We will address your comments point by point. ------ Reply1. Our PPSENet employs a two-stage strategy. In the first stage (ITP), we train using the PVCP dataset for 2D pose estimation. In the second stage, we select MotionBert as the baseline for our second phase (PTM) and make improvements based on it. We chose MotionBERT as baseline for the following reasons: - SOTA for Temporal Pose Estimation: At the start of our work, this was the SOTA method for temporal human pose estimation and could be easily transferred to downstream tasks. - Pose Prior for Large-Scale Data: Pedestrian-vehicle collision events are rare, making our PVCP dataset relatively small compared to other large datasets such as Human3.6M. MotionBERT’s pre-trained model on large-scale datasets provides a good pose prior. However, we innovatively added an auxiliary loss for collision pose classification, tailored to the pre-collision poses in our dataset, and employed a simple iterative regression approach. Under identical conditions, our PPSENet significantly outperformed MotionBERT on the PVCP dataset (see paper Table 2), demonstrating the effectiveness of our improvements. We have submitted the source code to the AC with the paper version and make them public subsequently. Indeed, while our PPSENet outperforms the original MotionBERT on PVCP dataset, the quantitative results are not exceptionally high. We believe the reasons are as follows: - Annotation Challenges: 3D pre-collision pose annotations in traffic scenes cannot be obtained through motion capture (Mocap) but are aligned using our annotation tools with the SMPL foreground and image background contours. - Complexity of Pre-Collision Poses: Pre-collision poses differ from daily human poses. As a specific type of pose, the inherent complexity of the pedestrian structure makes it difficult to achieve the same quantitative results as other pose datasets. - Dataset Size: Although our PVCP dataset contains over 20K annotations of pedestrian pre-collision poses, its size is still relatively small compared to the training results of large-scale datasets such as MSCOCO (＞1000K) and Human3.6M (＞500K) . These factors contribute to the relatively poor quantitative performance of our results. We will strive to improve our dataset and algorithm, hoping to achieve better results in future research. Our PPSENet is a two-stage method. In Table 5, we compare it with both single-stage and two-stage algorithms. Indeed, the single-stage PARE achieves relatively good results. In this section, we tested the PVCP’s testset. Before procrustes-aligned, MPJPE and MPVE achieved better results, but after procrustes-aligned, our PPSENet outperformed PARE in PA-MPJPE and PA-MPVE. We hypothesize that due to data collection constraints, we could not obtain camera parameters and the world position of the pedestrian. When considering pedestrian orientation and position, we only used annotation tools for alignment. Ignoring pedestrian orientation, position, and camera parameters, our method outperforms PARE in human pose metrics after procrustes-aligned. ------ Reply2. Thank you for your correction. Regarding our description of "the first relevant dataset," it was indeed inappropriate. We will replace "first" with "a" to make the expression more precise. As follows: > Using a semi-automatic method, we collected dashcam videos of collisions to create ~~the first~~ (a) pedestrian-vehicle collision pose dataset,… ------ Reply3. We apologize for hiding the complete GitHub link of the annotation tool due to the requirement of anonymity. We have submitted the complete and anonymous code for this submission to AC for review. As mentioned in the supplemental material, our SMPL Annotation Tool is not a completely new development. Instead, we improved upon the tool mentioned in [Reference 66 of paper], which was originally designed for merely viewing various SMPL parameters. We have added many interactive modules to make it capable of annotating SMPL, including functionalities such as importing background images, automatically saving annotations, adjusting the weights of foreground and background to meet our annotation requirements, and annotating our dataset based on these functionalities. We hope that this annotation tool will be helpful to the work related to SMPL datasets. ------ Reply4. I apologize for making such a basic mistake. Thank you for your careful review. We will correct the year to the correct one in the final revision. ------ Reply5. Thanks for your correction, our expression is really easy to cause confusion, so the following changes have been made to line 46: > Original: Dashcams or public surveillance devices are the only sources of data, ~~and constructing a large-scale dataset of pedestrian pre-collision poses for network training is similarly impractical.~~ > Revised: Dashcams or public surveillance devices are the only sources of data , further constrains the approaches to dataset creation, thereby enhancing the difficulty and complexity of producing such datasets. ------ Thanks again for your comments. We will address all suggestions in the final revision, and some additional notes will also be extended to the newly submitted supplemental material. --- Rebuttal Comment 1.1: Comment: Thanks to the author for the detailed reply. --- - It can be seen that the author has done a lot of work and made sufficient preparations to make the community accept PVCP and PPSENet. - The most obvious flaw of PVCP is that the data volume is too small. I can understand that the human collision poses are displayed for a very short time, resulting in a small overall data volume. But for a dataset, it is still too small. And I don’t quite understand what the downstream tasks of PVCP are. If the vehicle has collided with the human body, what is the meaning of posture estimation? If it is to predict human body movements, how much time can be gained for the autonomous driving system? --- These questions of mine are fundamental. It is difficult for the author to give me a clear answer directly from the rebuttal. But it cannot be denied that "Pedestrian-Centric 3D Pre-collision Pose and Shape Estimation from Dashcam Perspective" is indeed a standard data set + method paper. --- Rebuttal 2: Title: Reply Comment: Thank you very much for your recognition of our work and our preparation. We value your feedback and have carefully considered every question you have raised. Here are our detailed responses to your questions: ------ 1. The problem with data volume is small. You are correct in pointing out that the limited data volume is indeed a constraint of our dataset (PVCP). Indeed, due to the extreme rarity and brevity of pedestrian collision poses in real-world traffic scenarios, collecting a large amount of high-quality data poses significant challenges. Although we preliminarily verified The validity of our method and dataset, the current scale of our dataset is relatively small. Nevertheless, we have established a comprehensive dataset creation process and designed and published corresponding SMPL pose annotation tools. This means that any researcher can conduct similar research using the same methods and tools, allowing the relevant dataset to gradually expand as more researchers contribute. We have made every effort to ensure the diversity and accuracy of the data through multi-source data collection and meticulous annotation. Furthermore, we are exploring data augmentation techniques to further expand the dataset and improve the generalization capabilities of the models. 2. Regarding downstream tasks of PVCP. The PVCP dataset is a pedestrian pre-collision pose dataset captured from a dashcam perspective. In our paper, we primarily focused on a fundamental computer vision task: Human Pose Estimation (HPE). However, our dataset also holds potential for applications in pedestrian collision warning and emergency response systems for autonomous vehicles, encompassing but not limited to pedestrian emergency pose prediction, behavior recognition, and more. For instance, in predicting human movements, even though the time window may be relatively short, even a few milliseconds of advance notice can significantly impact the decision-making of autonomous driving systems. Specifically, by predicting pedestrian poses and shapes prior to a collision, autonomous driving systems can identify potential collision risks earlier, allowing them to take proactive measures such as decelerating or avoiding the pedestrian, thereby reducing the likelihood of a collision occurring or mitigating the severity of the collision if it does occur. Naturally, these tasks may require additional efforts in enhancing the real-time performance and generalization capabilities of the models, which also represents the room for subsequent improvements in our methodology and dataset. In addition, the pre-collision pose of pedestrians plays a crucial role in the entire automotive active and passive safety system, beyond merely a visual task. We wrote in the section `Related work -> Pedestrian Pre-collision Pose` that "Rapid and accurate acquisition of pedestrian pre-collision poses supports research on collision damage and active safety protection (lines 86-88)" . Although the pose estimation itself may no longer directly affect the collision outcome once a vehicle collides with a pedestrian, it is significant for understanding the pedestrian dynamics before the collision occurs, assessing the severity of the collision, and optimizing subsequent emergency response measures. Some efforts have been made to analyze the collision process between pedestrians and vehicles by reconstructing the collision scene afterward, where the authentic initial posture of pedestrians is a vital consideration `[paper 1,2,3,4,5,6,7,8; due to character restrictions, see the next comment for references]`. Reconstructing the authentic pedestrian posture from videos can contribute to more precise accident reconstruction and analysis, further enabling the analysis of the dynamic and kinematic characteristics of pedestrians during collisions, and ultimately optimizing the shape and structural design of the vehicle's front end to reduce human-vehicle collision injuries. The PVCP dataset provides the ability to accurately estimate pedestrian collision poses for the analysis of post-accident events, which may not be achieved by other pose datasets. ------ References > Due to character restrictions, see the next comment for references. ------ We also believe that the PVCP dataset and its proposed method have significant application value in the field of autonomous driving, especially for tasks such as pedestrian collision pose estimation, behavior recognition, vehicle collision avoidance decision-making, and post-accident analysis, and can contribute to enhancing road traffic safety. Thank you again for your attention and support to our work, and we look forward to continuing to communicate with you to further improve our research. --- Rebuttal 3: Title: Reply (continued) Comment: This comment is a continuation of the previous one, and here are some references that support the downstream tasks of our PVCP dataset. References > [1] K. Gildea, D. Hall, C. R. Cherry, and C. Simms, “Forward dynamics computationalmodelling of a cyclist fall with the inclusion of protective response using deeplearning-based human pose estimation,” JOURNAL OF BIOMECHANICS, vol. 163, JAN2024. [Online]. Available: https://doi.org/10.1016/j.jbiomech.2024.111959 > > [2] J. Wang, Z. Li, D. Zou, Y. Chen et al., “Reconstruction of a real-worldcar-to-pedestrian collision using geomatics techniques and numerical simulations,”Journal of forensic and legal medicine, vol. 91, p. 102433, 2022. [Online]. Available:https://doi.org/10.1016/j.jflm.2022.102433 > > [3] T. Zou, A. Zha, Q. Liu, and C. Simms, “Pedestrian gaits observed from actualpedestrian-vehicle collisions,” International Journal of Crashworthiness, vol. 27,no. 1, pp. 1–23, 2022. [Online]. Available: https://doi.org/10.1080/13588265.2020. > 1769455 > > [4] A. Schubert, N. Erlinger, C. Leo, J. Iraeus, J. John, and C. Klug, “Developmentof a 50th percentile female femur model,” in 2021 IRCOBI conference proceedings. International Research Council on the Biomechanics of Injury (IRCOBI . . . , 2021, pp. > 308–32. [Online]. Available: https://www.ircobi.org/wordpress/downloads/irc21/pdf-files/2138.pdf > > [5] S. Shang, C. Masson, M. Llari, M. Py, Q. Ferrand, P.-J. Arnoux, andC. Simms, “The predictive capacity of the madymo ellipsoid pedestrian modelfor pedestrian ground contact kinematics and injury evaluation,” AccidentAnalysis amp; Prevention, vol. 149, p. 105803, Jan 2021. [Online]. Available:http://dx.doi.org/10.1016/j.aap.2020.105803 > > [6] Y. Huang, Q. Zhou, C. Koelper, Q. Li, and B. Nie, “Are riders of electrictwo-wheelers safer than bicyclists in collisions with motor vehicles?” AccidentAnalysis & Prevention, vol. 134, p. 105336, 2020. [Online]. Available: https://doi.org/10.1016/j.aap.2019.105336 > > [7] M. Lalwala, A. Chawla, P. Thomas, and S. Mukherjee, “Finite elementreconstruction of real-world pedestrian accidents using thums pedestrian model,”International journal of crashworthiness, 2019. [Online]. Available: https://doi.org/10.1080/13588265.2019.1594547 > > [8] L. Shi, Y. Han, H. Huang, Q. Li, B. Wang, and K. Mizuno, “Analysis of pedestrian-to-ground impact injury risk in vehicle-to-pedestrian collisions based on rotation angles,” Journal of Safety Research, vol. 64, pp. 37–47, 2018. [Online]. > Available: https://doi.org/10.1016/j.jsr.2017.12.004	Rebuttal 1: Rebuttal: We appreciate the thorough review and insightful comments from AC and all reviewers. The reviewer's comprehensive feedback is as follows： - Reviewer SmGM acknowledged the importance and quality of our PVCP dataset for autonomous driving research and sought further opinions on the innovativeness of PPSENet and related datasets. - Reviewer 9csR affirmed that "This paper makes a significant contribution by collecting real-world pedestrian pose and shape parameters before collisions." And reviewer inquired about the annotation process, post-processing, and requested more visualizations of the dataset. - Reviewer 3M7p praised our PVCP dataset and PPSENet algorithm, and inquired about the open-sourcing of the dataset annotation tool. - Reviewer j3Vv found the topic generally interesting, commending the intention to expand the scope of daily human pose to poses in specific settings, particularly those related to daily life safety. Reviewer remarked that "The paper is generally well-written and easy to follow." Further, reviewer questioned the differences, uniqueness, and necessity of our PVCP dataset compared to others, as well as the implementability of the real-time performance and its performance when trained on other large datasets. ------ We have thoroughly addressed all the comments and, during this period, accomplished the following tasks: 1. We have provided detailed responses and revision plans to each reviewer's comments in a separate Rebuttal document. 2. We have supplemented the relevant experiments and visualization results in a separate PDF submission, including: - Visual comparisons of PVCP with other pose datasets (MSCOCO, Human3.6M, PW3D, PedX). `(For all Reviewers)` - Smoothing post-processing of the dataset. `(For Reviewer 9csR, Questions Reply: Reply1)` - The performance of the method when trained solely on another large human pose dataset (COCO, PW3D) without fine-tuning on PVCP. ` (For Reviewer j3Vv, Limitations Reply: Reply2)` 3. We have submitted anonymous source code links for the SMPL annotation tool, PVCP dataset description, and the method's source code to the AC. These links facilitate efficient review of our work by the AC and reviewers. The links include: - SMPL Annotation Tools Code` (For Reviewer 3M7p)` - Annotation process video` (For Reviewer 9csR, Reply3)` - PVCP dataset and PPSENT Readme.md `(For all Reviewers)` If you have any further questions, please feel free to discuss with us, and we will actively revise and improve our paper. Thank you again for your review. Pdf: /pdf/223ee074dce52d5eaa365dcee8be6fe15f21be2a.pdf	NeurIPS_2024_submissions_huggingface	2,024	null	null	null	null	null	null	null	null
Optimal Aggregation of Prediction Intervals under Unsupervised Domain Shift	Accept (poster)	Summary: It is crucial to assess and quantify uncertainties associated with distribution shifts in the context of unsupervised domain shift. The paper propose methodologies for aggregating prediction intervals, which capture the range of likely outcomes for a given prediction. The authors support these methodologies theoretically by considering a bounded density ratio and a measure-preserving transformation between source and target domains. This includes finite sample bounds for the coverage and width of their prediction intervals. Strengths: 1.The method is innovative in aggregating various prediction intervals on the target domain under the domain-aligning assumption within an unsupervised domain adaptation framework. 2.The paper includes sufficient theoretical discussions, presenting finite sample concentration bounds to demonstrate that their method achieves adequate coverage with a small width. Weaknesses: 1.Apart from theoretical results, there is a lack of intuitive understanding as to why aggregation can achieve minimal-width prediction intervals. 2.The paper does not include comparisons with previous methods of aggregated prediction intervals (Hosen et al., 2014). 3.Limited experiments and comparative methods make it difficult to affirm the effectiveness of the algorithm. [1] Hosen, A., Khosravi, A., Nahavandi, S., and Creighton, D. Improving the Quality of Prediction Intervals Through Optimal Aggregation. IEEE Transactions on Industrial Electronics. 2014 Technical Quality: 3 Clarity: 3 Questions for Authors: 1.What does $f \in \mathcal{F}$ signify in equation (2.1)? There is no $f$ explicitly used in the main part of this equation. 2.What does $m_0$ represent in line 118? Is it the same as $m_0$ mentioned in line 137? What is the relationship between $m_0$ and $f_0$? 3.In line 164, it states, "However, any such estimator $ \hat{\omega}(x) $ can be non-zero for $ x $ where $ \omega_0(x)=0 $ due to estimation error. Consequently, $ \hat{\omega} $ may not be efficient in selecting informative source samples." Given this observation, why not apply the hinge function not to $ \hat{\omega} $, but to $ (Y - m_0(x))^2 - f(x) $ in Equation (3.5)? 4.The notation $B_{\mathcal{F}}$ first appears in Theorem 3.2 (please correct me if I am mistaken), and there is no explanation provided for it. 5.The paper introduces two algorithms, but Section 5 includes experiments for only one of them. Which algorithm is it, and why are there no experiments conducted on the other one? Confidence: 2 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: N/A Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thanks a lot for reading our paper and for your insightful comments. 1. Lack of intuitive understanding: Thank you for your concern. To achieve a minimal-width prediction interval, it is crucial to accurately capture the shape of the interval. However, a single predictor may fail to do so due to its specific structural limitations. For example, the correct shape might be a polynomial function, but the predictor could output a linear function, resulting in model misspecification. By aggregating several predictors and considering a convex hull of these base predictors, we reduce model misspecification. This approach leads to an ensemble predictor that more accurately captures the interval's shape, resulting in a smaller width. The second step (shrinkage) is equivalent to conformal prediction, which maintains the coverage guarantee. 2. Comparison with Hosen et al. (2014): Thank you for pointing out this interesting paper. We will include a comparison with the previous methods from Hosen et al. (2014) in our revised version. Hosen et al. (2014) focus on ensemble prediction intervals obtained using neural networks through ranking and weighted averaging. While they consider both coverage and width, they combine these metrics into a single optimization objective (coverage width criterion). In contrast, our method focuses on finding the interval with the smallest width among those that provide sufficient coverage. Additionally, Hosen et al. (2014) do not provide theoretical guarantees for coverage and width, nor do they address domain shift scenarios. Our work, on the other hand, emphasizes providing a theoretically robust methodology specifically designed to handle domain shift challenges. 3. Limited experiments: Thank you for raising this point. We have now conducted three additional real-data experiments to evaluate our method. We have added one more baseline, the weighted quantile conformal prediction method. Please see the global response for details on these experiments. Last but not least, we have also conducted a simulation experiment quantifying the robustness of our method against conformal prediction. Please see our response to Reviewer 6ToN for details of that experiment. 4. What does $f \in \\mathcal{F}$ signify?: Sorry for the typo, it should be $\min_{l, u}$. 5. What does $m_0$ represent?: $m_0$ is the conditional mean function, as defined in line 137. We will clarify it in our revision. 6. Why not apply hinge to $\hat w(x)$?: The purpose of using a hinge function is to ensure the convexity of the problem. Specifically, we replace the indicator function for $(Y - m_0(x))^2 - f(x)$ with a convex surrogate (hinge function) to facilitate the use of standard convex optimization techniques. 7. Notation $\\mathcal{B}_{\mathcal{F}}$: $\\mathcal{B}\_{\mathcal{F}}$ is defined to be an upper bound of $\sup_{f\in\mathcal{F}}\\|f\\|_{\infty}$ (as shown in line 181 in Theorem 3.2). We will make it more clear in our revision. 8. Experiment on Algorithm 2: We have conducted an experiment on the Airfoil data using our second algorithm, which requires estimating the optimal transport map. Please see our global response for the details. --- Rebuttal Comment 1.1: Comment: Thank you for your detailed explanations. I appreciate the way you’ve clarified the intuitive understanding and your commitment to comparing it with previous approaches. However, in the quesiton 3, it states that "$\hat{\omega}$ may not be efficient in selecting informative source samples." Given this, would applying the hinge function to $(Y - m_0(x))^2 - f(x)$ help address this issue? Besides theoretical analysis convenience, are there any other benefits to this approach? --- Reply to Comment 1.1.1: Comment: Thanks for your further comments. In Step 1 (Shape estimation), the ideal optimization problem we aim to solve is given by equation (3.2) in our paper, which can be expressed as: $$ \\min\_{f \in \\mathcal{F}} \ \mathbb{E}\_{n, T}[f(X)]\, \ \ {\rm s.t.}\ \\mathbb{E}\_{n,S}[w_0(X)\\mathbf{1}_{\{f(X)<(Y-m_0(X))^2\}}]=0 \.\tag{1} $$ If we had perfect knowledge of the source samples for which $ w_0 > 0 $, we could directly solve equation (3.2). However, in practice, $w_0$ is unknown, and we must estimate it using $ \hat{w} $. This introduces the possibility of estimation error, where $ w_0(X_i)=0 $ but $ \hat{w}(X_i)> 0 $ for some source samples $X_i $. Such errors lead to additional constraints—namely, requiring $ f(X_i)\geq(Y_i-m_0(X_i))^2 $ for all $X_i $ where $ \hat{w}(X_i)> 0 $—that are not necessary if $w_0(X_i)=0 $. To address this, instead of enforcing the right-hand side of the constraint in (1) to be exactly 0, we relax it to $ \epsilon $ to accommodate the estimation error in $\hat{w} $. This adjustment leads to the following optimization problem: $$ \min_{f \in \\mathcal{F}} \ \ \\mathbb{E}_{n, T}[f(X)]\, \ \ \text{s.t. }\ \\mathbb{E}\_{n,S}\left[\hat{w}(X)\\mathbf{1}\_{\{f(X)<(Y-m_0(X))^2\}}\right]\leq \epsilon \.\tag{2} $$ However, there is a caveat: the presence of the indicator function in (2) makes the constraint non-convex. To address this, we replace the indicator function with a convex surrogate, specifically the hinge function $ h_{\delta}\left((Y-m_0(X))^2-f(X)\right) $, which restores convexity to the problem. In summary, we start with equation (3.2) from our paper, substitute $w_0 $ with $ \hat{w} $, and then replace the indicator function with a hinge function to ensure the problem remains convex. Thus, the use of the hinge function is driven by computational considerations rather than merely for theoretical analysis.	Summary: The authors study the problem of how to construct prediction intervals on a target domain under both covariate shift and domain shift assumptions (i.e., the source and target domains are related either via a bounded density ratio, or a measure-preserving transformation), designed to ensure adequate coverage while minimizing interval width. They provide theoretical guarantees, with finite sample bounds, regarding the prediction interval coverage and width. They apply their method on the airfoil dataset, comparing the performance with weighted split conformal prediction. Strengths: - The paper is well-written overall, the authors definitely seem knowledgeable. - The paper studies an important and interesting problem. - The proposed method outperforms weighted split conformal prediction on one dataset. Weaknesses: - The experimental evaluation is limited to a single low-dimensional tabular dataset, with a synthetic/simulated distribution shift. It is not possible to determine if the proposed method actually would have significant real-world utility/impact. Summary: - Important problem and a well-written paper overall, but the experimental evaluation is too limited. I think this could be a solid paper, but I don't think it can be accepted in its current form. Technical Quality: 2 Clarity: 3 Questions for Authors: - Could you extend the experimental evaluation with more datasets and baseline methods? - Why is the figure on page 9 missing a figure number and caption? Confidence: 3 Soundness: 2 Presentation: 3 Contribution: 2 Limitations: Yes. Flag For Ethics Review: ['No ethics review needed.'] Rating: 5 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thanks a lot for reading our paper and for your insightful comments. 1. Extension of empirical evaluation: Thank you for your comment, please see our global response for the details of the additional experiments. (i). We now have two more real-data experiments on our first method, Algorithm 1(which relies on the estimation of the density ratio of the source and the target covariates), one using the real estate valuation data and the other using the energy efficiency data. Furthermore, we have an additional experiment on our second method, Algorithm 2 (which relies on estimating the optimal transport map), using the Airfoil data. (ii). We have now added one more baseline, namely the weighted quantile-adjusted conformal method, i.e. we now compare our against both variance-adjusted weighted conformal and weighted quantile-adjusted conformal method. (iii). We have also conducted a simulation experiment to showcase the robustness of our algorithm in comparison to weighted conformal prediction. Kindly see our response to Reviewer 6ToN for the details of this experiment. 2. Figure 9: Thank you for pointing out this problem. We will take care of it in the revised version of the paper. --- Rebuttal Comment 1.1: Comment: Thank you for your response. I have read the other reviews and all rebuttals. The other reviews are all borderline, but slightly more positive than mine. The authors add some new experiments, and respond well overall to the reviews. My main concern _"The experimental evaluation is limited to a single low-dimensional tabular dataset, with a synthetic/simulated distribution shift. It is not possible to determine if the proposed method actually would have significant real-world utility/impact"_ is not fully addressed. The evaluation is still limited to low-dimensional tabular datasets with synthetic/simulated distribution shifts, and I thus still find it difficult to know if the method actually would have significant real-world utility. - Could you evaluate on some dataset with a real-world distribution/domain shift? - Could you evaluate on some non-tabular dataset, on an image-based dataset? - Also, is it not a fairly major concern that the coverage of your method falls below 0.95 in the "Airfoil data with Algorithm 2" experiment? Yes, the interval length is roughly half that of WVAC and WQC, but is this (big) improvement worth it if the method fails to achieve valid coverage? I am open to increasing my score at least to "5: Borderline accept", but would still like to see at least some more convincing experimental results. --- Reply to Comment 1.1.1: Comment: Thanks for your further comments. 1.Could you evaluate on some dataset with a real-world distribution/domain shift?: We applied our method to the ETDataset (ETT-small) from [1], which contains hourly-level data from two electricity transformers at two different stations, including measurements of load and oil temperature. Each data point consists of 8 features, including the date of the point, the predictive value "oil temperature", and 6 different types of external power load features. For our experiment, we used the data from one transformer during the period from July 1, 2016, to November 2, 2016, as our source data, and data from the same time period from the other transformer as our target data. As our source data and the target data are from different locations, we have a geographical covariate shift; see [1] for more details. Our results are as follows: Table: Experimental results for the ETDataset \| Outcome \| Our Method \| WVAC \| WQC \| \|-----------\|------------\|-------\|-------\| \| Coverage \| 0.976 \| 0.982 \| 0.842 \| \| Bandwidth \| 41.525 \| 57.9 \| 54.981\| [1] Haoyi Zhou, etc. Informer: Beyond Efficient Transformer for Long Sequence Time-Series Forecasting. 2. Could you evaluate on some non-tabular dataset, on an image-based dataset? The widely used non-tabular data in the field of distribution shift are various types of image data (e.g., MNIST-USPS digit data, Waterbird data etc.), which are generally used for classification tasks. However, our current method for constructing prediction intervals relies on the continuous response variable. Therefore, it cannot accommodate non-continuous responses. We acknowledge this limitation of our method and believe that this extension will be an interesting future research direction. 3. Also, is it not a fairly major concern that the coverage of your method falls below 0.95 in the "Airfoil data with Algorithm 2" experiment? Yes, the interval length is roughly half that of WVAC and WQC, but is this (big) improvement worth it if the method fails to achieve valid coverage? We believe the reason our method fails to achieve coverage is a small sample effect rather than the limitation of our methodology. As is evident from other real data experiments, our method achieves the nominal level of coverage. Therefore, our method does not sacrifice coverage to reduce width; our theoretical results indicate that asymptotically, our method achieves adequate coverage while having minimal width. --- Rebuttal 2: Comment: _(I will quickly write a reply now, to try to give you some time to respond before the deadline, but I will also think more carefully about this later before making my final recommendation. I will still consider raising my score)_ * * > (notably, the AssetWealth dataset you referenced is 13GB) - There are 8 different datasets in the TMLR paper i pointed to, I thought that maybe it could be easy to apply to at least one of those? > ...suggesting its significant potential impact across diverse real-world scenarios - The method has only been applied to a single dataset with a real-world distribution shift though, right? Are there some other tabular datasets with realistic distribution shifts? > These experiments consistently demonstrate the advantages of our method over WVAC and WQC - Expect for "Airfoil data with Algorithm 2" where the coverage drops below 0.95? Also, is this the only evaluation of Algorithm 2? Why? (could perhaps be made more clear in general what the two different algorithms should be utilized for, how do I know which one to use in a real-world application?) > ...performing dimensionality reduction. Once these preprocessing steps are completed and we obtain vector representations of the images, our method can be implemented efficiently and executed rapidly - Why is dimensionality reduction a required step? Isn't this a quite significant limitation of your method then, which would limit the real-world utility also for non-image data? No dataset in the experiments has more than 8(?) features, what's the maximum number your method can handle? --- Rebuttal Comment 2.1: Comment: Thanks for your further comments. 1. The method has only been applied to a single dataset with a real-world distribution shift though, right? Are there some other tabular datasets with realistic distribution shifts? Following your concern, we managed to run an additional experiment within this short timeframe. The new dataset is the Appliances Energy Prediction Dataset, which is freely available from the UCI repository. This dataset is a time series data with 28 covariates and one response variable. We used data from 2016-01-11 to 2016-02-15 (5000 samples) as our training set and data from 2016-05-13 to 2016-05-27 (2000 samples) as our testing set. Since the source and target data are from different time periods, this experiment involves a non-synthetic real-world time shift. The results based on our Algorithm 2 are presented below: Table: Experimental results for the Appliances Energy Prediction Dataset \| Outcome \| Our Method \| WVAC \| WQC \| \|-----------\|------------\|----------\|---------\| \| Coverage \| 0.95 \| 1.00 \| 1.00 \| \| Bandwidth \| 461.69 \| 6809.87 \| 2032.12 \| 2. Why is dimensionality reduction a required step? We would like to clarify that our experiment involves two main steps: first, training the component functions, and second, efficiently aggregating these components. Our methodology specifically focuses on the aggregation step, where we combine pre-trained models to obtain a prediction interval with adequate coverage and minimal width. The key point is that our method assumes these component functions are already trained, and our contribution lies in the efficient aggregation of these components. When we say our method is easy to implement, we are referring to this aggregation step, which is independent of the data's dimensionality. In many real-world applications, these components are typically pre-trained and provided to us. However, we acknowledge that if the pre-trained components are not available, they must first be trained from the data. This training process can indeed be time-consuming, especially as the models' complexity increases. The dimension reduction step may be required for training these component functions, but is not required for aggregation. To summarize, when we refer to our method, we mean the aggregation (shape estimation) and shrinkage steps, which are straightforward to implement and do not depend on the data’s dimensionality. The more challenging aspect is the initial training of the components if they are not already available. Our method can indeed handle large datasets, with the main consideration being the time required to compute the component functions. To illustrate, we conducted an additional experiment on Appliances Energy Prediction Datase, using data with 28 covariates. It's important to note that the aggregation step involves combining multiple component functions, each of which is univariate. Therefore, the dimension involved in the aggregation step is the number of component functions, not the dimension of the data; the data's dimensionality only affects the training of the component functions. 3. Expect for "Airfoil data with Algorithm 2" where the coverage drops below 0.95? Also, is this the only evaluation of Algorithm 2? Why? As we mentioned, the drop in coverage below 0.95 is due to the small sample effect rather than a limitation of our methodology. We evaluated Algorithm 2 on both the Airfoil data and the ETDataset. The choice between Algorithm 1 and Algorithm 2 depends on prior knowledge about the type of distribution shift. If the shift is due to covariate changes, and the support of the test covariates lies within the support of the source covariates, then Algorithm 1 is appropriate. However, Algorithm 2 is more suitable when the shift is caused by transformations (e.g., shifts, rotations). If the support of the source and target covariates is likely to be (almost) disjoint, density ratio estimation will be poor, making Algorithm 2 the better option in such cases. 4. There are 8 different datasets in the TMLR paper I pointed to, I thought that maybe it could be easy to apply to at least one of those? All eight datasets referenced in the TMLR paper are large-scale datasets. Even the smallest among them, SkinLesionPixels, is still 2GB. Additionally, accessing some of these datasets requires account registration, which further complicates the process.	Summary: Building on the work of Fan et al. (2023), this paper addresses the challenge of computing prediction intervals in an unsupervised domain adaptation setting, where labeled samples are available from a related source domain, and unlabeled covariates are available for the target domain. The primary objective is to generate prediction intervals for the target domain that provide both an asymptotic coverage guarantee and minimal width. The authors also explore the aggregation of prediction intervals on the target domain under both covariate shift and domain shift scenarios. Theoretical guarantees, including finite sample concentration bounds on approximation errors and coverage guarantees, are presented. The proposed methods are illustrated and compared with the weighted split conformal prediction method using a real-world dataset. Strengths: - The paper tackles the important and challenging problem of uncertainty quantification in the unsupervised domain adaptation setting. - The authors propose algorithms to generate prediction intervals with both coverage and minimal width guarantees. - Experimental results show that the proposed method generates prediction intervals with comparable coverage but smaller average width compared to an existing method. Weaknesses: - The paper assumes familiarity with Fan et al. (2023), making it difficult for readers who have not read that work to understand some notations and concepts. Section 2.1 on problem formulation is particularly hard to read without prior knowledge. - There is a lack of flow in the paper, with some explanations and definitions missing or unclear. For instance, terms like "valid prediction interval", and expressions such as U(x) and l(x) are not clearly explained. - The authors write "This translates into solving the following optimization problem:" without any explanations. - Another example is Line 234, "adding a slight delta to ensure coverage even when F is complex.". This is not clear. It is discussed in Fan et al (2023) in (2.4). - Even if simple, I suggest adding important details about different expressions in the appendix to help the reader. - The title mentions "optimal aggregation," but model aggregation is not the central focus of the paper. It is first discussed in Remark 4.2 and briefly mentioned in line 272. - The paper should highlight the similarities and differences with Fan et al. (2023) in terms of contributions, both theoretical and empirical. Also, the authors use different quantile levels for estimators compared to Fan et al. (2023) without explaining the reason for these differences. In Fan et al (2023), the following quantile levels are considered: 0.05, 0.35, 0.65, and 0.95, while the authors use 0.85, 0.95, 0.9, and 0.9. - The statement that "the shape of the prediction band does not change much if we change the level of coverage" lacks clarity on scenarios where this might not be true, such as multimodal distributions (?). - The discussion on lambda <=1 is confusing and needs improvement. The paper inconsistently explains the shrinkage factor and its implications. For example, the authors write "Therefore, it is not immediate whether the shrinkage factor is smaller than 1, i.e., whether we are indeed shrinking the confidence interval (lambda > 1 is undesirable, as it will widen finit , increasing the width of the prediction band)." and "Although ideally lambda <= 1, it is not immediately guaranteed as we use separate data for shrinking.". The authors also discuss it in Lemma 3.3. and Lemma 4.4. - The term "shape of the prediction band" with prediction intervals can be confusing as it can have other meanings (see functional data). I suggest the authors clarify this in line 147. - The paper ignores a significant amount of research on prediction interval estimation. I suggest discussing important related work. - The limitations of the paper are not discussed. For example, focusing solely on intervals might have implications under multimodal distributions, where High-Density Regions (HDR) would be the smallest regions. - The relationship between the proposed approach and conformal prediction is not clearly explained. The paper should discuss the guarantees provided by the proposed approach compared to conformal prediction, including the relevant theoretical guarantees and calibration aspects. For example, the weighted conformal approach also splits the data and estimates a density ratio. What about the (weaker?) guarantees you provide? The following paper could also be useful: "On the expected size of conformal prediction sets". Also, in conformal prediction, in addition to finite-sample marginal guarantees, the authors often provide asymptotic conditional guarantees. What can you say about this with your approach? - The comparison with the conformal approach in the experiments might not be fair, as conformal prediction does not optimize for the smallest width. Additionally, reporting the interval score, which is a proper scoring rule for intervals, would be useful for evaluating the methods. - Typos and clarifications: - line 137: "remains same." - line 154: "via by solving" - line 453 "a.s. on source". On the source domain? - line 470, "we with have" - It is not "Leduox-Talagrand" but Ledoux-Talagrand". - line 481, a minus sign is missing - Line 490 to 492. It is not clear why tilde tau <= alpha. - Line 182, what does \mathcal{F} - f* mean? - About Theorem 3.4, we expect the performance (or convergence rates) to also depend on alpha. Where does this appear in the theoretical guarantees? Technical Quality: 3 Clarity: 2 Questions for Authors: See weaknesses. Confidence: 4 Soundness: 3 Presentation: 2 Contribution: 3 Limitations: The limitations are not discussed in the paper. Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thanks a lot for reading our paper and for your insightful comments. 1. The paper assumes familiarity with Fan et al. (2023) Thank you for raising this concern! We will carefully proofread the revised version of the manuscript according to your suggestions. 2. Mention optimal aggregation Thank you for pointing this out. Indeed, our current approach presents the general problem and then casts prediction interval aggregation as a special case. We will highlight the aggregation perspective more in the revised version of the manuscript. 3. Similarities and dissimilarities to Fan et al. (2023) Please see our global response. We selected the quantile levels of 0.85, 0.95, 0.9, and 0.9 because they are close to the desired coverage and we believe they effectively capture the correct shape of the interval. It is important to note that our method is not restricted to these specific quantiles; we can also apply the quantile levels used in Fan (2023) to our approach if needed. 4. Change of shape of prediction band Thank you for your feedback. We will clarify this in line 147 in the revision. Specifically, when we refer to the "the shape of the prediction band does not change much", we mean that the form of the minimal average width prediction band covering the data at a certain coverage level should not change much if we slightly change the coverage level, i.e., the form is stable with respect to the perturbation of the coverage level. For instance, if the optimal prediction band that covers $98\\%$ of the data is quadratic in $X$, then the minimal-width interval that covers $95\\%$ of the data is also a quadratic function of $X$, rather than abruptly changing to a linear function. Under this consistency assumption, a constant shrinkage factor $\lambda$ in the second step (i.e., (3.3) and (4.2)) is sufficient to adjust from $100\\%$ to $95\\%$ coverage. However, if this is not the case, we can modify our method to allow the shrinkage level $\lambda$ to depend on the input $X$, ensuring more accurate adaptation to the data distribution. For the multimodal distributions, we acknowledge that the optimal region is not necessarily a prediction band (i.e., $Y\|X=x$ is not necessarily a single interval) and will include this in the limitation of the work. 5. Discussion on $\lambda \le 1$ Thank you for the comments. We will clarify this discussion in the revised version of the paper. In the first step, we typically choose $\epsilon$ to be small; therefore, the average width of $\hat f_{\rm init}$ can be larger than necessary, and coverage is also typically larger than $95\\%$. Therefore, in the second step, we aim to shrink the interval (i.e., ensure $\lambda \leq 1$). However, it is not immediate that $\hat \lambda$, obtained by solving equation (3.3) is indeed $\le 1$. To ensure this, we prove in Lemma 3.3 and Lemma 3.4 that with high probability, we indeed have $\hat \lambda(\alpha) \le 1$. 6. Research on prediction interval estimation Thank you for pointing this out. In our revision, we will add a detailed related work section that covers significant research on prediction interval estimation, including methods applicable to both shifted and non-shifted data. We will discuss approaches such as conformal prediction, quantile regression, and RKHS-based methods, highlighting their contributions and relevance to our work. 7. Limitations of the paper are not discussed Thank you very much for highlighting this limitation! It is indeed true that if the conditional distribution of $Y$ given $X$ is multimodal, the prediction interval may not be the best thing to do. We will mention this in the limitation. 8. Comparison with conformal (1) Methodology relation and difference: Step 2 of Algorithm 1, i.e., shrinkage (equation (3.3) and (4.2)) in our method is equivalent to (weighted) conformal prediction, using the score function $(y - \hat{m}(x))/\hat{f}(x)$, where $\hat{f}$ is obtained from the first step of our method. By efficiently aggregating various predictors, our $\hat{f}$ accurately captures the shape of the interval, resulting in a smaller bandwidth. In contrast, variance-adjusted conformal prediction selects $f$ as the variance function, which may not necessarily capture an optimal shape for a small bandwidth. Additionally, our method explicitly aims to minimize the bandwidth in the first step. In contrast, conformal prediction does not explicitly minimize bandwidth; it implicitly addresses this by setting the cutoff at the $1-\alpha$ quantile of the score function. (2)Robustness: Our method is more robust compared to variance-adjusted conformal prediction. Please see our response to Reviewer 6ToN for the details of our experiment regarding robustness. (3)Theoretical guarantees: We provide finite sample guarantees for both the width and coverage of our prediction intervals. In contrast, while typical conformal literature offers finite-sample marginal or asymptotic conditional guarantees for coverage, they often do not include finite-sample guarantees for width, as our paper does. For distribution shift [1] provides coverage guarantee assuming the density ratio $w(\cdot)$ is known, whereas our Theorem 3.4 explicitly highlights the effect of the estimation error of $w(\cdot)$. We acknowledge that our method is particularly focused on addressing marginal coverage, and therefore, its pointwise guarantee is not immediately clear. We will include this as a limitation in our discussion and consider it a topic for future research. [1] Tibshirani, R., et al., Conformal prediction under covariate shift. 9. Typo Thank you very much for carefully reading our paper and pointing out the typos! We will take care of it in the revised version of the manuscript. --- Rebuttal Comment 1.1: Comment: Thank you for answering my questions and concerns. I have raised my score to 6 (weak accept). --- Reply to Comment 1.1.1: Comment: Thank you very much for both your valuable reviews and the increased score. We will carefully revise our manuscript based on your suggestions.	Summary: This work proposes a method to construct prediction intervals in distribution shift problems, where unlabeled data from the target domain is available. The method is inspired by Fan et al (2023) in the i.i.d. setting and assumes we can estimate a transform (defined by reweighting or a transport map) that maps the joint distribution from the source domain to the target. Strengths: - The problem studied is relevant. - The method appears reasonable and sound. Weaknesses: - My main concern is about the novelty of this work: given the assumed condition on domain shift (namely a good approximation to the reweighting or transport map is readily available) it is a natural application of the method in Fan et al (2023). The proof will require some additional effort but given the assumptions I do not expect them to be technically challenging. - While the paper employed assumptions on domain shift are common in literature, I am uncertain about their relevance, especially in high-dimensional problems. It is well-known that the bounded density ratio assumption in Section 3 is unrealistic in high dimensions. And while I have come across past works on the transport map assumption, I do not know of realistic examples where the assumption actually holds so that the coverage guarantees established in this work could be relevant: it is much easier to think of scenarios where the assumption does not necessarily hold, but heuristic methods based on this assumption achieve better predictive performance; the latter is more often the goal in previous works. - Finally, I have doubts about the novelty/utility of Fan et al (2023), which seems quite similar to conformal prediction. It could be alternatively summarized as applying quantile estimation to the "conformity score" of $(y - \hat m(x)) / f(x)$, and compared with conformal prediction the main difference is the lack of sample splitting (in split conformal) or leave-one-out operations (as in full conformal). I wonder if the removal of sample splitting is really a wise choice: the coverage guarantees in Fan et al (2023) and this work relies on the Rademacher complexity of the function class for $f$. If the function class is defined by general ML models the guarantees would be rather unreliable in high-dimensional problems, in sharp contrast to the distribution-free guarantees in (split/full) conformal prediction. Otherwise, we would need further sample splitting to construct a finite set of candidates for $f$, in which case we could just do split conformal prediction. Technical Quality: 3 Clarity: 3 Questions for Authors: Given the above questions on the theoretical motivation, novelty or significance, I believe there should be extensive empirical evaluation for the proposed method. In light of question 3 above, I also wonder if the observed performance difference between the proposed method and conformal prediction could be attributable to the choice of different conformity scores; how does the method compare to an implementation of split conformal prediction using the same $(y - \hat m(x)) / f(x)$ as the conformity socre? Confidence: 4 Soundness: 3 Presentation: 3 Contribution: 2 Limitations: Yes. Flag For Ethics Review: ['No ethics review needed.'] Rating: 5 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thanks a lot for reading our paper and for your insightful comments. We answer your questions and concerns in the following. 1. Novelty compared to Fan et al. (2023): Please see our global response. 2. Assumptions on the covariate shift: Covariate shift is a common assumption, even for high-dimensional data, e.g., see [1] or [2]. The impact of domain shift in our analysis is primarily through the estimation error of the density ratio and the optimal transport map. We acknowledge that high-dimensional settings can pose challenges, such as the curse of dimensionality. However, these issues can often be mitigated by incorporating more structure; without any structural assumption, we do not expect to gain any theoretical guarantee under covariate shift in high-dimension. An example of a structured distribution shift is the exponential tilt assumption (the one we used in our experiments), or that only a subset of variables undergoes the shift (sparsity type assumption). Similarly, we can assume some structure in the optimal transport map to avoid the curse of dimensionality. While our analysis is quite general, we believe that with specific structural assumptions, our method and analysis can be appropriately adjusted to incorporate these structures. Therefore, we agree that we need some structural assumptions in high-dimension, but without any assumption, the curse of dimensionality is unavoidable. Once we have that, our method is applicable to obtain a valid prediction band with a small width. [1] Tripuraneni, N., et al., Covariate shift in high-dimensional random feature regression. [2] Tripuraneni, N., et al., Overparameterization improves robustness to covariate shift in high dimensions. 3. Similarity and dissimilarity with conformal prediction: We agree with you that there is a similarity between our method and the conformal prediction in the second stage, i.e., the shrinkage (i.e., (3.3) and (4.2)), which is equivalent to (weighted) conformal prediction as both of these methods aim to derive weighted quantile of the sample score $s(x, y) = (y - \hat{m}(x))/\hat{f}(x)$. However, the key difference lies in the construction of $\hat f$; by efficiently aggregating various predictors, $\hat f$ accurately captures the shape of the interval, resulting in a smaller bandwidth. # Robustness check Another advantage of our method is its robustness. Our method aggregates various predictor intervals, including an estimator for the conditional variance function (Estimator 5). In contrast, the weighted variance-adjusted conformal (WVAC) prediction interval [3] heavily relies on accurately estimating this conditional variance, while conditional quantile regression relies on accurate estimation of the conditional quantile function. We did a small robustness check using the following model: \begin{align} & X \sim{\sf Unif}([-1, 1]), \ \ \xi \sim{\sf Unif}([-1, 1]),\ \ X,\xi \text{ independent}, \\ & Y = \sqrt{1 + 25 X^4} \xi . \end{align} We estimated the conditional variance using a random forest with varying depths \{3, 5, 7, 15\}. We generated n = 2500 samples, keeping $75\\%$ as source data and resampling the remaining 25\% with weighted samples proportional to $w(x) \propto (1 + \exp{(-2x)})^{-1}$. As depth increases, overfitting leads to poor out-of-sample variance predictions. The following table is the result of 100 Monte Carlo experiments. The number inside the parenthesis is the median of coverage over these Monte Carlo iterations. \| Max depth \| Avg. width--Our \| Avg. width--WVAC \| \|-----------\|------------------\|------------------\| \| 3 \| 2.07 (0.975) \| 3.08 (0.9712) \| \| 5 \| 2.07 (0.95) \| 3.28 (0.9664) \| \| 7 \| 2.068 (0.94) \| 3.33 (0.97) \| \| 15 \| 2.08 (0.97) \| 5.00 (0.97) \| The above table implies our method is more robust to the misspecification of some model components and remains stable. In a nutshell, our method has a similar coverage guarantee as the conformal prediction, which does not require many assumptions; our coverage guarantee (Theorem 3.4) requires a reasonable estimation of $w(\cdot)$, which is necessary for both our method and that of [3]; [3] even assumed $w(\cdot)$ is known, whereas our bounds quantify the estimation error of $w(\cdot)$. When we require a guarantee on the width, we require some assumptions on the predictor class. [3] Tibshirani, R., et al., Conformal prediction under covariate shift. 4. Extensive empirical evaluation Please see our global response. --- Rebuttal Comment 1.1: Comment: Thank you for your response. I appreciate the clarifications regarding the theoretical contributions and the new experiments. I will update my score accordingly. --- Reply to Comment 1.1.1: Comment: Thank you very much for reading our rebuttal and for increasing the score. We will put the new experiments in our revision. --- Rebuttal 2: Comment: As the rebuttal period is coming to a close, we wanted to check if there are any remaining concerns that might be preventing you from adjusting your score. If there is any additional clarification we can provide, please let us know. We would be more than happy to address any questions you might have. Thank you again for your time and effort in reviewing our paper.	Rebuttal 1: Rebuttal: # Global response We thank all the reviewers for their insightful comments. We address a few concerns that were raised by multiple reviewers. 1. Comparison with Fan et al. (2023): The key contribution of our paper lies in adapting the methodology from Fan (2023) (which only addresses no-shift scenarios), to tackle domain shift challenges. We deal with unsupervised domain adaptation, i.e., i) the distribution of the covariates of the target domain is different from the source domain, and ii) we do not observe any label from the target domain. Therefore, we are required to change the methodology; to adapt to these challenges, the density ratio or the optimal transport map between the covariates must be estimated. Furthermore, as pointed out in Section 3, the shift may cause the optimization problem non-convex, for which we need to introduce a convex surrogate (e.g., the hinge function). As a consequence, we have various additional theoretical challenges to establish that our method indeed produces a prediction band on the target domain with adequate coverage and minimal width. For example, our theory highlights the precise effect of the estimation error of the density ratio or the optimal transport map in the bound on the coverage and the width. Furthermore, when we replace the indicator function with its convex surrogate hinge function (see (3.5)), we introduce two additional parameters $(\epsilon, \delta)$, effectively measuring the closeness between the original NP-hard problem (involving indicator function) and its convex surrogate. Our theoretical bounds also highlight the effect of these hyperparameters in our finite sample bounds. 2. Additional Experiments: We conducted three more real-data experiments. In the following, WVAC is weighted variance adjusted conformal prediction interval, where the score function is $s(x, y) = \|y - \hat m(x)\|/\hat \sigma(x)$ with $\hat m$ and $\sigma$ being the estimate of conditional mean and variance function respectively. WQC is the weighted quantile conformal prediction interval with $s(x, y) = \max \\{\hat q_{\alpha/2}(x) - y, y - \hat q_{(1-\alpha)/2}(x)\\}$, where $\hat q$ is the estimated quantile function. (a). Real Estate Dataset The Real Estate Valuation dataset consists of 414 instances. The goal is to predict the house price per unit area based on the 6 other features. The data is available at the UCI ML repository. The construction of shifted data (with $\beta = (-1, 0, -1, 0, 1, 1)$) and implementation procedure are the same as our paper. The following table and Figure 1 in PDF present the results over 200 Monte Carlo iterations. It is evident from the table that our method produced a small average width in comparison to the other methods while maintaining the coverage guarantee. \| Outcome \| Our Method \| WVAC \| WQC \| \|-------------------------------------------------\|------------\|---------\|---------\| \| Coverage (Median) \| 0.98 \| 0.962 \| 0.971 \| \| Coverage (IQR) \| 0.058 \| 0.048 \| 0.048 \| \| Bandwidth (Median) \| 36.889 \| 46.392 \| 46.858 \| \| Bandwidth (IQR) \| 15.001 \| 19.027 \| 14.189 \| \| Bandwidth (Median for Coverage > 95%) \| 40.774 \| 50.703 \| 51.031 \| (b). Energy efficiency data The goal of the Energy Efficiency dataset is to predict the heating load based on 8 other covariates. The construction of shifted data (with $\beta = (-1, 0, 1, 0, -1, 0, 0, -1)$) and implementation procedure are the same as our paper. This data is also available at the UCI ML repository. The following table and Figure 2 in PDF present the results over 200 Monte Carlo iterations. It shows that our method produced a smaller bandwidth than WVAC. While WQC has a smaller median bandwidth, it sacrifices coverage. The last row indicates that for experiments with coverage $\geq 95\\%$, WQC's median average width is significantly larger than ours. Thus, whenever WQC provides adequate coverage, its bandwidth is much larger than ours. \| Outcome \| Our Method \| WVAC \| WQC \| \|-------------------------------------------------\|------------\|---------\|---------\| \| Coverage (Median) \| 0.995 \| 0.969 \| 0.973 \| \| Coverage (IQR) \| 0.047 \| 0.036 \| 0.05 \| \| Bandwidth (Median) \| 4.332 \| 5.045 \| 2.842 \| \| Bandwidth (IQR) \| 1.358 \| 3.269 \| 2.551 \| \| Bandwidth (Median for Coverage > 95%) \| 4.373 \| 5.681 \| 4.94 \| (c). Airfoil data with Algorithm 2 We implement our second method (based on OT) using the airfoil data. Here, we shift 25% of the data by a linear transformation: $x \mapsto Ax + b$, where $A$ = diagonal(1.5, 1.2, 1.6, 2, 1.8) and $b = (1, 0, 0, 1, 0)$. Our method produces a small width in comparison to other methods. See Figure 3 in PDF and the following table. \| Outcome \| Our Method \| Our Method (without OT) \| WVAC \| WQC \| \|-------------------------------------------------\|------------\|-------------------------\|---------\|---------\| \| Coverage (Median) \| 0.928 \| 0.749 \| 0.984 \| 0.952 \| \| Coverage (IQR) \| 0.035 \| 0.22 \| 0.024 \| 0.077 \| \| Bandwidth (Median) \| 15.075 \| 18.512 \| 36.298 \| 32.143 \| \| Bandwidth (IQR) \| 1.638 \| 3.089 \| 10.619 \| 8.364 \| \| Bandwidth (Median for Coverage > 95%) \| 16.429 \| 25.268 \| 37.783 \| 36.433 \| (d).For the robustness experiment, please see our response to Reviewer 6ToN. Pdf: /pdf/7ce7c270de4e869bcafa3a8150a620c0e9ec673a.pdf	NeurIPS_2024_submissions_huggingface	2,024	null	null	null	null	null	null	null	null
Coherent 3D Scene Diffusion From a Single RGB Image	Accept (poster)	Summary: The task is to perform 3D scene reconstruction from a single RGB views, such that they can output both scene object poses and geometries. Prior work cannot jointly output both poses and geometries, one design choice that is validated by this work to be effective. Method-wise, extending on SALAD, this work introduces a diffusion model that learns a generative scene prior that can capture contextual information across scene elements. To achieve joint learning, they introduce a alignment loss such that they only need to single view supervision. The author conducts extensive experiments to justify their design choices. Strengths: 1. The introduction of the surface alignment loss (Sec 3.6) provides a nice solution to jointly improve object poses and geometry estimation. This kind of supervision only needs image level supervision (no need for 3D supervision). 2. The experiments on comparison and ablation is informative, which reasons well the key design decision made in this paper. 3. The qualitative demo is multifaceted, which makes it easier to evaluate their pose and geometry quality. Weaknesses: 1. Minor concern is that the paper over-claim a bit on the contribution. Despite phrasing as 3D scene reconstruction (as is mentioned in teaser, abstract and intro) from single view, the proposed framework in the method actually only reconstruct poses and geometries of foreground objects in the image without the background part, e.g., ground or wall. 2. Global shape prior cannot always fit the observed object perfectly, with different local details, e.g., Fig.3 shelf. This could accumulate error in the downstream alignment. 3. Furniture kinds of 3D objects, e.g., those in ShapeNet, are simple, which usually is easier to get good shape reconstruction. How would the proposed framework work on more intricate 3D objects, e.g., plants, trees? Technical Quality: 3 Clarity: 3 Questions for Authors: 1. Could you clarify on how the proposed framework resolve with scale ambiguity given the single view setup? 2. Writing needs more polish. For example, Equation 1 seems incorrect (feel free to clarify in the rebuttal). Based on the description in Line 121, RGB image I seems to be given, which indicates it is one item to be conditioned on. And in Line 150, missing parenthesis for (See Sec. 4.2 3. Could you clarify on how the gradient being back-propagated through point sampling (Line 190)? "We directly sample m = 1000 points p(j,l) ∼ N (μj , Σj ) per Gaussian Gˆi,j resulting in a shape point cloud Pi = {p(j,l)\|j ∈ {1,...,g},l ∈ {1,...,m}}." In SALAD, it has another cascaded diffusion Transformer that condition on extrinsic and intrinsic parameters to get the part-level point cloud from Transformer. But I am not sure if the process is the same in this paper. 4. Could you clarify how the texture of the 3D objects are learned? The qualitative results show objects with color but there is no supervision on the color. Confidence: 4 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: Yes. Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: 1. Reconstructing background objects Indeed, structural elements can be considered part of holistic scene reconstruction. Including structural elements in the scene prior can certainly be beneficial and can improve the coherence of the scene. We will extend our formulation with the structural elements of the scene and train a layout estimator to reconstruct them. We will report our findings in the final paper - thanks a lot for this suggestion! 2. Global shape prior Thanks for the comment. While our model produces clean shape modes even under heavy occlusions compared to blurry or incomplete shapes of prior work, our learned shape prior does not necessarily match the input image perfectly. In our alignment loss formulation, we sample points from the 3D Gaussians, which represent the coarse shape of the object. Incorporating the per-Gaussian intrinsics, which encode local details, into the loss formulation is an interesting experiment. 3. Reconstruction of natural objects, e.g. plants, trees Thanks for the interesting question. In Fig. 3, 5 and 9, we show that our geometry-based model is able to reconstruct certain fine details, such as individual braces at the back of a wooden chair or thin table legs. However, for natural, intricate objects such as plants, geometric approaches generally struggle to reconstruct very fine details. To study the limits of our approach, we will include a comparison with an increased number of 3D Gaussians in the shape model. To further mitigate this issue, our approach can reconstruct the geometry of the object, and additional fine details, such as individual leaves, can be further represented by local textures, as has been done in computer games, or with 3D Gaussian splats. 4. Resolving scale ambiguity Since our model learns object poses and scales together with a strong scene prior, our model naturally reasons about common scene scales and hence is less prone to ambiguities of individual objects given the scene context. 5. RGB image input Thanks for the comment. We will go through the manuscript thoroughly and improve the writing and fix minor typos. Eq. 1 defines the overall problem setup of reconstructing multiple objects from a single RGB image ($\mathbf{I}$, L121). Instead of conditioning on the input image directly, we use an image feature extractor $\Theta_i$ and instance segmentation on the input image to extract per-detection image features, which are then used to condition the diffusion model ($\text{feat}_i$, L135, Eq. 6). 6. Gradient backpropagation through point sampling For sampling we use `torch.distributions.MultivariateNormal` function with the estimated per-Gaussian centers $\mu$ and covariance matrices $\Sigma$. While the sampling itself is not differentiable, we can compute the gradients on the sampled points. During joint training with the proposed efficient alignment loss, we omit the additional denoising of the intrinsic vectors and the costly rejection of point samples using the shape decoder. Hence, our point sampling formulation is different from the shape decoding of SPAGHETTI [1] or SALAD [2]. 7. Textures of objects Our model does not predict object colors or textures, however, this is an interesting future direction. For visual clarity, we color each object according to its semantic class - we will indicate this in the captions of the figures. References - [1] Hertz et al. - “SPAGHETTI: Editing Implicit Shapes Through Part Aware Generation”, SIGGRAPH (TOG) 2022 - [2] Koo et al. - “SALAD: Part-Level Latent Diffusion for 3D Shape Generation and Manipulation”, ICCV 2023 --- Rebuttal Comment 1.1: Comment: Thanks for the detailed clarification. I retain my rating. Nice work!	Summary: This paper proposes a novel framework for 3D scene reconstruction from single images based on diffusion models. This framework jointly predicts 3D object poses and shapes with seperate diffusion models and captures global scene context with a standard mutli-head attention module. To train this model, the paper proposes a surface alignment loss based on one-sided Chamfer Distance, which makes their method able to utilize depth maps as supervision. The experimental results on Sun RGB-D, ScanNet, and Pix3D and numerical ablation studies demonstrate the efficacy of proposed method. Strengths: This paper shows its strengths from several aspects: 1. It proposes an overall solid method for 3D scene reconstruction from single images and deals with the ill-posed nature of this problem with powerful generative prior from diffusion models. 2. The proposed alignment loss makes the method able to utilize partial supervision from depth maps for model training. 3. This paper conducted comprehensive experiments in both scene and object reconstruction and achieve good results comparing to previous methods. 4. The presentation and writing are overall good and clear. Weaknesses: My concerns are mainly about the description of the method: 1. The model architecture is not clear enough in Section 3.7. It seems that each object pose, each object shape, and each Gaussian, are denoised using an independent diffusion model? Where is ISA added and how is it integrated with other modules? Are the inputs of ISA from Equation (7)? 2. In Figure 1, are the noisy shape Gaussians as in shape with guassian noise? Or are they initial estimation directly from image features? It's a bit confusing. 3. Are the diffusion models for shape and pose trained with full 3D supervision before joint training with alignment loss? If yes, what if there are only ground truth depth maps in training data? Besides, the overall idea of 3D generation from features as in Equation (6) combined with global context via attention has been explored in previous related works, such as part-based 3D object reconstruction, e.g., He, Qian, et al. "Single Image 3D Object Estimation with Primitive Graph Networks." Proceedings of the 29th ACM International Conference on Multimedia. 2021. This should be included in related works. Technical Quality: 3 Clarity: 3 Questions for Authors: 1. According to the alignment loss, the predicted Gaussians should be converged to centered on the surface of 3D objects? Otheriwse, the direct sampling won't be around the surface? 2. How do you deal with varying number of objects within a single image? How do you deal with false positives and false negatives from detection? 3. Have you tried other normalization methods besides Equation (12-14)? Does it affect the learning efficiency? 4. Does the disentanglement of pose and shape helps with learning convergence? 5. How is the proposed method comparing to other SOTAs in single object reconstruction? Can this method be applied to part-level object reconstruction? Confidence: 4 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: The authors adequately addressed the limitations and I do not have more to add here. Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: 1. Clarification of model architecture Thank you for the feedback regarding the architecture description of our model in Section 3.7. We will improve the clarity of this section in the final paper. In the rebuttal PDF, we have included an architecture figure of the shape model (Fig. 4). In particular, the pose diffusion model is conditioned on all objects’ image features of a scene simultaneously (Eq. 6 & 7). This setup allows the model to learn a scene prior via ISA by exchanging relational information between the objects and to denoise the individual object poses. During joint training, the entire model takes as input all scene object features (Eq. 6 & 7), estimates the scene context implicitly via ISA, predicts the object poses and the intermediate shape Gaussians of each object. These Gaussians are transformed into world space using the specific object pose to compute the Alignment Loss (Fig. 2, right). 2. Clarification of “Noisy Shape Gaussians” in Figure 1 The “Noisy Shape Gaussians” in Figure 1 depict the anisotropic 3D Gaussians of 3 objects at the beginning of the denoising process. During the backward diffusion process, these 3D Gaussians get denoised to form the intermediate, scaffolding representation of each shape. We will clarify it in the figure and the caption. 3. 3D supervision Yes, the pose and shape models are first trained with respective 3D supervision before joint fine-tuning. We follow the training scheme of previous works, Total3D [3], Im3D [4], and train the poses on SUN RGB-D [5] and shapes on Pix3D [6]. Hence, our approach learns marginal distributions for layout and shape first. We then fine-tune for the joint distribution by using the provided depth map from SUN RGB-D and the proposed alignment loss. We agree that training on depth data only poses an interesting and challenging problem setting. In order to directly learn the joint distribution, image-based representation methods, such as EG3D [7], could be employed to learn meaningful shape priors from RGB-(D) data. These could be combined with depth and pose estimation like NOCs [8]. We will address this in the limitation section. 4. Context-learning via attention Thanks for pointing us to this relevant work, we will include it in the related work section. 5. Surface alignment of Gaussians In our experiments, we observed that the intermediate 3D Gaussians usually align well with the object’s parts (see Fig. 1 and Fig. 2), leading to sampled points that are close to the surface. For future work, it would be interesting to integrate surfel-based representations [9, 10] as promising alternatives for binding Gaussians to the surface of the object (2DGS [11]). 6. Varying number of scene objects Our model can in fact handle a varying number of objects in the scene, because the dot-product attention in our ISA formulation is agnostic to the number of objects during train and inference time. In our experiments, we have not experienced negative impacts of false positives/negatives from the image detection backbone, however it is possible to perform NMS on the reconstructed 3D shapes to further filter out false detections. 7. Object pose normalizations During our experiments, we did not try different normalization factors for the object pose parameters. We will include a comparison with different, e.g., unnormalized, object poses in the final paper. 8. Convergence of disentangled of pose and shape Thank you for the interesting question. Investigating the convergence properties of the presented disentangled pose and shape formulation is indeed a meaningful addition. We will perform a comparison in a controlled environment, e.g., using synthetic data with full 3D ground truth such as 3D-FRONT [12], and include it in the final paper. 9. Comparison with SOTA single-object methods We show comparisons for single object reconstruction with InstPIFU [13] in Figure 8. 10. Can it be applied to part-level object reconstruction In our experiments, the scaffolding 3D Gaussians align well with individual objects parts. For a part-level reconstruction, each vertex of the output shape mesh can be assigned to its nearest 3D Gaussian to get a part-level segmentation of the object. We will include a visualization of that in the final paper. References - [1] Hertz et al. - “SPAGHETTI: Editing Implicit Shapes Through Part Aware Generation”, SIGGRAPH (TOG) 2022 - [2] Koo et al. - “SALAD: Part-Level Latent Diffusion for 3D Shape Generation and Manipulation”, ICCV 2023 - [3] Nie et al. - “Total3DUnderstanding: Joint Layout, Object Pose and Mesh Reconstruction for Indoor Scenes from a Single Image”, CVPR 2020 - [4] Zhang et al. - “Holistic 3D Scene Understanding from a Single Image with Implicit Representation”, CVPR 2021 - [5] Sun et al. - “Pix3D: Dataset and Methods for Single-Image 3D Shape Modeling”, CVPR 2018 - [6] Song et al. - “SUN RGB-D: A RGB-D Scene Understanding Benchmark Suite”, CVPR 2015 - [7] Chan et al. - “EG3D: Efficient Geometry-aware 3D Generative Adversarial Networks”, CVPR 2022 - [8] Wang et al. - “Normalized Object Coordinate Space for Category-Level 6D Object Pose and Size Estimation”, CVPR 2019 - [9] Pfister et al. - “Surfels: Surface elements as rendering primitives”, SIGGRAPH 2000 - [10] Zwicker et al. - “EWA volume splatting”, Visualization 2001 - [11] Huang et al. - “2DGS: 2D Gaussian Splatting for Geometrically Accurate Radiance Fields”, SIGGRAPH 2024 - [12] Fu et al. - “3D-FRONT: 3D Furnished Rooms with layOuts and semaNTics”, ICCV 2021 - [13] Liu et al. - “Towards High-Fidelity Single-view Holistic Reconstruction of Indoor Scenes”, ECCV 2022 --- Rebuttal Comment 1.1: Comment: Thank you for the response. The architecture diagram is very helpful and most of my concerns are resolved. I do believe this is a good work and would like keep my rating.	Summary: In this work, the authors propose a diffusion-based method for generating coherent 3D scenes from a single input image. They use a diffusion model to jointly denoise the 3D poses and geometries of all objects. To address the incomplete ground truth of existing datasets, they also introduce a surface alignment loss to better learn the 3D geometry. Experimental results show that the proposed approach significantly outperforms state-of-the-art methods. Strengths: - I like the formulation of this method. I think anisotropic 3D Gaussians are a good representation for 3D reconstruction. We could treat the 3D reconstruction in a coarse-to-fine manner and use the 3D Gaussians as the 3D proxy to guide the fine reconstruction. - The results look promising. With partial ground truth for supervision, the proposed method learns complete geometry from a single RGB image with high occlusion. Weaknesses: - I am unclear on how to learn the shape information: (i) For the shape code sigma_i (16 Gaussians x 16 dimensions per Gaussian), do you have the ground truth for it? If not, how do you apply the shape diffusion loss mentioned in Equation 10? (ii) To decode the shape code from Gaussians to an occupancy grid, you apply a shape decoder diffusion model (mentioned in Line 213). How do you train this model? Do you include any details about this model? - I am interested in how the proposed method learns the intermediate scene representation (i.e., anisotropic 3D Gaussians). You use the surface alignment loss as supervision. Are there other indirect supervisions? How do you ensure that the Gaussians do not intersect each other or become trivial (e.g., the same as each other)? - If I understand correctly, you add inter-object attentions to the intra-scene attention to allow the pose and shape model to run in parallel. If so, this operator is quite common. - I find the author may miss the citation and discussion with an important literature [a]. [a] CoReNet: Coherent 3D scene reconstruction from a single RGB image. ECCV 2020. Technical Quality: 3 Clarity: 2 Questions for Authors: - It is suggested to add notation for cls in Equation 6. - It is suggested to add some figures for the ablation study. - I find [b] to be quite related to this work. It is suggested to discuss this paper in the related work section. Combining some 3D architecture with the proposed method could be a nice extension. [b] XCube: Large-Scale 3D Generative Modeling using Sparse Voxel Hierarchies. CVPR 2024. Confidence: 2 Soundness: 3 Presentation: 2 Contribution: 2 Limitations: I think the authors properly discuss the limitations of this work. Flag For Ethics Review: ['No ethics review needed.'] Rating: 5 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: 1. Clarification of shape learning To support our latent diffusion approach for modeling shapes, we represent 3D shapes following the disentangled formulation of SPAGHETTI [1]. SPAGHETTI learns the disentanglement into 16 Gaussians and per-Gaussian “intrinsics” in a self-supervised way through augmentation with rigid transformations of the shapes. Since the official code only provides the inference part and checkpoints for chairs and airplanes, we re-implement the method according to the paper and train it across all categories on Pix3D [2]. From this trained model, we extract the shape codes $\sigma_i$ for supervision of the shape diffusion model. The shape decoder mentioned in L213 refers to the one of SPAGHETTI. We will add a more detailed description of our re-implementation and training in the paper as well as release the code and checkpoints together with the final paper. 2. Intermediate scene representation The intermediate scene presentation consists of the per-shape 3D Gaussians from the shape diffusion model and the object pose from the object pose diffusion model. The 3D Gaussians are transformed to world space by the predicted object poses. Following baseline protocols and Sec. 4.2, we train our model in stages and use the respective supervision: The scene prior and object poses are trained on SUN RGB-D [3]; Shape reconstruction is trained on Pix3D [2] using the preprocessed 3D shape codes $\sigma_i$ (see above). The final model is then jointly trained on SUN RGB-D. We do not constrain the 3D Gaussians to be non-overlapping or non-trivial, however by restricting the number of Gaussians (16 in our experiments), this encourages the shape model to distribute the Gaussians effectively in 3D space. 3. Intra-Scene Attention Since we condition the model on all scene objects simultaneously, the intra-scene attention allows our model to learn a joint scene prior, e.g., properties such as object-object relationships and scene contexts. This is different from methods that individually regress the objects’ pose/shape (MeshRCNN [5], ROCA [6]) or refine the object poses in a second step (Im3D [4]). 4. Notion of cls in Eq. 6 ‘cls’ in Eq. 6 refers to the semantic class in L135. We will clarify this in the paper. 5. Figures for ablation studies We included two additional figures (Fig. 1 & Fig. 2) for qualitative results of our ablation studies in the rebuttal PDF. 6. Additional related works Thanks for pointing us to additional related works, we are happy to include and discuss them in the paper. References - [1] Hertz et al. - “SPAGHETTI: Editing Implicit Shapes Through Part Aware Generation”, SIGGRAPH (TOG) 2022 - [2] Sun et al. - “Pix3D: Dataset and Methods for Single-Image 3D Shape Modeling”, CVPR 2018 - [3] Song et al. - “SUN RGB-D: A RGB-D Scene Understanding Benchmark Suite”, CVPR 2015 - [4] Zhang et al. - “Holistic 3D Scene Understanding from a Single Image with Implicit Representation”, CVPR 2021 - [5] Gkioxari et al. - “Mesh R-CNN”, ICCV 2019 - [6] Gümeli et al. - "ROCA: Robust cad model retrieval and alignment from a single image.", CVPR 2022 --- Rebuttal Comment 1.1: Comment: Thanks for your rebuttal. I would like to keep my rate.	Summary: This paper proposes an approach to reconstruct the 3D surfaces of multiple objects from a single RGB image. Given an instance-segmented RGB image as input, the poses and shapes of objects are jointly predicted and denoised using a diffusion model conditioned on the input image. A one-sided chamfer distance (from available ground truth surface point to sampled points in camera space) is used to provide supervision on incomplete surfaces. The output is geometrically evaluated against Total3D (3D detection-based composition), Im3D (implicit surface), and InstPIFu (pixel-aligned implicit surface) and shows improved scene coherence and shape quality. Strengths: + Efficient training + Appropriate use of category-specific object shape priors + Clean ablation results Weaknesses: In order for the simple loss (one-sided chamfer distance) to work well, you need strong category-specific priors on generated shapes; it requires the estimated shape gaussians to be stable in the first place. The joint learning of poses is also inherently category-specific and might not generalize well to unseen or uncommon object categories. While this limitation is mentioned in the paper, I think this is an inherent limitation that is not easy to mitigate. Limits of one-sided chamfer distance: In qualitative results containing office chair images for example, I noticed that many of them have extra leg parts hallucinated. I think this is an example of something that the one-sided chamfer loss won't be able to fix since the error isn't reflected in the loss. I'm surprised the method worked well despite this; this isn't necessarily a weakness, but tells us something about the problem we're trying to solve and evaluation methods (maybe "alignment is all you need"?). I think a retrieval-based baseline (with alignment) would have been interesting for the reasons mentioned. I tend to think the improvements are mostly due to architectural advancements such as attention models (that can process arbitrary unordered variable-length objects) and better off-the-shelf shape models available. The paper appropriately applied those techniques and that is a good thing, but maybe it does not fundamentally improve our understanding of the problem. Perhaps this isn't a fair criticism considering the limitations and quality of other papers accepted in similar venues. I currently feel borderline about the paper. The holistic nature of the problem is difficult enough and the paper proposes one simple solution that I think can be used by other researchers still. Technical Quality: 3 Clarity: 2 Questions for Authors: (please see weaknesses section) Confidence: 3 Soundness: 3 Presentation: 2 Contribution: 2 Limitations: I think they are addressed well. Flag For Ethics Review: ['No ethics review needed.'] Rating: 5 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: 1. Generalization to unseen/uncommon object categories Our shape prior is learned across all categories and in our experiments, we have seen that the shape Gaussians are indeed quite stable and align well with geometric parts (Figs. 1 & 2). However, we do rely on strong priors for seen categories, making generalization to unseen categories challenging. This limitation could be mitigated by training on larger shape datasets such as Objaverse [1]. 2. Comparison to a retrieval baseline We include qualitative results of the CAD retrieval method ROCA [2] in the rebuttal PDF (Fig. 3). While the retrieved CAD models are complete by definition, ROCA struggles to capture the correct mode of instances and produces misaligned and intersecting objects. We will include a detailed evaluation and comparison of this baseline in the final paper. 3. Improvements In contrast to the regression-based baselines, such as Total3D [3], Im3D [4] and Mesh R-CNN [4], we introduce a fully probabilistic approach to model the joint distribution of object arrangements and shapes. As our quantitative and qualitative results show, our formulation leads to more accurate shapes and object arrangements. In particular, for ambiguous and challenging cases, our diffusion prior still yields plausible results where the baselines tend to produce inaccurate geometries and unrealistic poses. References - [1] Deitke et al. "Objaverse: A universe of annotated 3d objects.", CVPR 2023. - [2] Gümeli et al. - "ROCA: Robust cad model retrieval and alignment from a single image.", CVPR 2022 - [3] Nie et al. - “Total3DUnderstanding: Joint Layout, Object Pose and Mesh Reconstruction for Indoor Scenes from a Single Image”, CVPR 2020 - [4] Zhang et al. - “Holistic 3D Scene Understanding from a Single Image with Implicit Representation”, CVPR 2021 - [5] Gkioxari et al. - “Mesh R-CNN”, ICCV 2019 --- Rebuttal 2: Comment: Thank you for the rebuttal. My "borderline accept" rating remains the same. I agree with other reviewers that the paper is above threshold.	Rebuttal 1: Rebuttal: We would like to thank the reviewers for their constructive and valuable feedback. We are pleased that all reviewers recognized the soundness of our approach for the challenging problem of single-image 3D reconstruction. Our method was highlighted as solid (scCV) and efficient (W8nX) by appropriately employing strong diffusion priors (W8nX, scCV). The proposed surface alignment loss was well-received (scCV, U6My, BXMX), as was the use of anisotropic Gaussians for shape representation (U6My). These contributions led to results described as “multi-faceted” (BXMX), “good” (scCV) and “promising” (U6My). Additionally, we are glad that our experimental setup and ablation studies were received as “comprehensive” (scCV), informative (BXMX) and “clean” (W8nX). Pdf: /pdf/60cb162fb4378209e7597f031024b3f0cc3c2328.pdf	NeurIPS_2024_submissions_huggingface	2,024	null	null	null	null	null	null	null	null
Towards Universal Mesh Movement Networks	Accept (spotlight)	Summary: The authors propose a method for general mesh movement. They design a network that takes the original mesh and the monitor values as input to predict the mapping for the adapted mesh. This makes it able to handle any mesh without the need of training PDE-type-specific models. Strengths: 1. The proposed method can handle mesh with different geometries and PDE with different type using a single model. The model only needs to be trained once. 2. Another advantage is its time cost is much lower than the traditional non-learned Monge-Ampère without the sacrifice of the movement accuracy. 3. Compared with prior works, ie, MA and M2N, it is more robust and easier to converge. 4. The paper is well written and easy to follow. Weaknesses: The idea of this paper is interesting to me, and the results, especially those shown in the supplementary video, are pretty nice. I believe the proposed method can have many potential application in mesh formation and the modeling of physical phenomena. I didn't see any obvious weaknesses from this paper. However, I have to admit that I'm not the expert in this field. Technical Quality: 3 Clarity: 3 Questions for Authors: 1. The network takes the original mesh and the monitor values $m$ as input. $m$ is the values for the original mesh or the target mesh? If it is the value of the original mesh, then how does the network know the target? If it is the value of the target mesh, then the symbol of $m_\xi$ is a bit misleading, since $\xi$ denotes the original mesh. 2. When using chamfer distance for the optimization of mesh matching, the result is easy to be stuck at the local minimum. Is it the case for the training of the proposed method? Confidence: 3 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: In the section of Limitation, the authors mentioned that a low-quality original mesh can lead to poor prediction. A visualization of it will be nice. Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We greatly appreciate your insightful comments that precisely recognize the strengths of our work! We are also very happy the Reviewer Jzyt checked our supplementary video and found it nice and useful. Below, we provide our point-to-point responses to the comments, denoted by [W] for weaknesses and [Q] for questions. Response to Q1: The monitor values 𝑚 correspond to the original mesh. Given a time-dependent PDE, the monitor values $𝑚$ at the first time step are computed based on the initial condition on the original mesh. Then the model takes in the monitor values $𝑚$ as input (with other features) and outputs a moved mesh. The PDE is solved based on this moved mesh and the solution obtained will be used for computing monitor values for next time-step. The target meshes only come in during the training process: the moved mesh is compared to the target mesh with element volume loss and Chamfer distance. The networks get to know the target during the training. In other words, the monitor values are part of input features and are defined in the nodes of the input meshes and not those of the target meshes. This indeed explains the notation $m_ξ$ with the monitor values being associated with the input mesh. The training process can be explained to teach the model to learn to solve an optimal transport problem according to equidistribution theory given the monitor values and other related features. Response to Q2: Thanks for raising this question. We acknowledge that Chamfer distance-based optimization for mesh matching or point cloud registration can potentially lead to convergence at local minima. However, in our case, we did not observe significant issues of being trapped in poor local minima during the training process, though we cannot guarantee that a global minimum was reached. We believe there are two primary reasons for this. First, the initial alignment between the original uniform meshes and the target meshes in our training dataset is not significantly different. This likely increases the chances of finding well-matched vertices when searching for nearest points using Chamfer distance. Second, in addition to the Chamfer distance, we employed an element volume loss, which provides additional information beneficial for vertex matching. Response to limitation: Thanks for the suggestions. We include an additional experiment about poor quality initial mesh as shown in Figure R4. The input mesh has highly anisotropic elements and several vertices with valence 6. This can certainly be considered a “low quality” mesh. Consider the monitor function to be the l2-norm of the recovered gradient of the solution of an anisotropic Helmholtz problem. With this monitor function, we tried applying conventional MA solvers and found that both the quasi-Newton and relaxation approaches failed to converge and/or resulted in tangled meshes, even though the elements are already aligned with the anisotropy of the Helmholtz solution. The UM2N approach, however, was able to successfully apply mesh movement without tangling. We would also like to clarify that our work targeting mesh movement that dynamically adapts to a PDE solution, which is not a replacement for mesh generation in general, so we would always expect a reasonable input mesh when applying our model. --- Rebuttal Comment 1.1: Comment: Thanks for your clarification for my questions. Please include the new results in the revised paper. --- Reply to Comment 1.1.1: Title: Thank you! Comment: Dear Jzyt, We sincerely thank you for your insightful comments with us! We commit to incorporate the new results in the revised paper. Best regards, Authors	Summary: This paper introduces a learnable model for mesh movement - which is a method for improving efficiency of a PDE solver by moving mesh nodes, while keeping the topology fixed. The proposed model itself is a two-stage neural architecture comprising a transformer encoder - taking a representation of the input mesh - and a graph attention network as a decoder outputting an updated mesh. The resulting model is independent of specific PDE type and thus can be applied to novel problem types in zero-shot manner. Quantitative evaluation is conducted on multiple synthetic and one realistic (tsunami prediction) tasks. Strengths: - Proposed method provides a way to improve PDE solvers agnostic to the underlying PDE type. A zero-shot model like this could be extremely useful in a lot of engineering fields. - Quantitative results indicate that the method is outperforming existing method for mesh movement and leads to improved surrogate modeling. Weaknesses: - Although overall the architecture makes sense, there is no ablation study conducted on the individual parts of the architecture. - Similarly, there are no baselines to off-the-shelf graph convolutional or transformer methods (any recent method e.g. for semantic mesh segmentation would work here?). - The model is only concerned with vertex movement of the model, which means there is no guarantee that there won't be any degraded triangles, which can influence the quality of the dowstream solutions. This could be a fundamental flaw of purely data-driven methods since there is not simple way to enforce constraints at test time. - The scale of the dataset seems to be very small for the model to be directly applicable in realistic scenarios - both in terms of number of samples and number of vertices. - Model relies on a transformer encoder, which will not scale to meshes with large number of vertices. Technical Quality: 3 Clarity: 3 Questions for Authors: - In figure 6, you demonstrate that um2n is better at handling complex boundaries. From what I can see there are no guarantess for your purely data-driven method to produce non-tangled meshes? Do you have an intuition to why this is the case? - Would this model work for larger scale problems with >1M? Confidence: 2 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: Limitations are discussed. Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We thank Reviewer jEN5 for acknowledging the contributions, soundness, and presentation quality of our paper. And greatly appreciate Reviewer jEN5 for proposing these insightful questions. Below, we provide our point-to-point responses to the comments, denoted by [W] for weaknesses and [Q] for questions. Response to W1: We perform an additional ablation study of our network architecture on the Swirl case. We construct two models, which replace the GAT Decoder with one layer GAT denoted as UM2N-w/o-Decoder and remove the graph transformer denoted as UM2N-w/o-GT. The results are shown in the Table R1.It can be observed that both variants give worse results compared to the full UM2N. A visualization is also shown in Figure R2. It can be observed that without the Decoder, the model distorts the shape of the ring, i.e., missing relative information between vertices; without the graph transformer, the model fails to capture the details of the ring shape. Response to W2: To the best of our knowledge, there are not a lot of works addressing the end-to-end learning-based r-adaptation problem. Many existing works focus on end-to-end surrogate neural PDE solvers. The models proposed in these works try to predict the PDE solutions directly, therefore it is hard to make a fair comparison as our UM2N predicts the mesh i.e., PDE discretization. To compare with them qualitatively, these methods have advantages in efficiency. On the other hand, our approach has the advantage that PDE solutions obtained are physically plausible based on guarantees from classical numerical analysis regarding the PDE discretization, whereas such guarantees are absent or enforced in a weaker sense in directly learned PDE solutions. For the network architecture, we apply the graph transformer as encoder for its powerful expressivity, and apply the graph attention network as decoder because it can naturally help alleviate mesh tangling. There are existing works such as efficient transformers from natural language processing or computer vision community, which provide promising future directions to improve the efficiency of our model. The semantic mesh segmentation mentioned by the Reviewer jEN5, although not directly applicable to our problem, would also be an interesting direction to explore. Response to W3/Q1: We agree that there are no guarantees for our method to produce non-tangled meshes as we have also noted in the limitations section of our paper. In Figure 6, we claim that UM2N is better at handling complex boundaries compared to the MA method. The MA method requires smooth boundaries and convex domains. Its behaviour can be observed from our additional experiments shown in Figure R1. It produces good results for the rectangle domain (shown in the first row) but outputs a highly tangled mesh given a non-convex domain (shown in the following rows). The intuitions behind why our data-driven UM2N can handle complex boundaries can be divided into two aspects. The first is that we use Graph Attention Network (GAT) to build the mesh deformer. Therefore, the coordinate of each vertex is updated by the weighted sum of its neighbors, the weights (or coefficients) of which are determined by the GAT module. This guarantees each vertex to only move within the convex hull of its neighboring vertices, hence effectively alleviating mesh tangling issues. In addition, we utilize element volume loss penalizing negative element volumes (i.e., inverted elements), which helps reduce mesh tangling. Response to W4/W5/Q2: As for the scale of the dataset, we consider the test case shown in the paper - Cylinder (~5k vertices, ~10k triangles), Tsunami ( ~8k vertices, ~16k triangles) - to have moderate degrees of freedom. The tsunami case is already a real-world scenario application in the ocean modelling community. To show a case with more degree of freedom, we further apply our trained model on a flow past cylinder case with ~11k vertices and ~22k triangles as shown in Figure R5. It can be shown that our UM2N works well in this case. As for the large case such as >1 million vertices/triangles mentioned by the reviewer, the model can be naturally applied to larger scale problems as there is no limitation for the input length of graph transformer and graph neural network but the GPU memory is the limit. We performed stress tests using a time-independent Helmholtz case on our RTX 3090 24 GB GPU and observe that the maximum scale of the problem that can be run on this GPU has ~50k number of elements. The inference time is ~760 ms, the MA method on the same problem requires ~37800 ms with a residual threshold 1e-4. Therefore, given our current limited computational resource, it is hard for us to construct a case with 1 > million vertices. Considering this limitation, applying the memory efficient transformer (linear-attention transformer for further improvement of the inference efficiency) is targeted as future work. --- Rebuttal Comment 1.1: Comment: Thanks for the detailed response. This addresses my questions/concerns so I am keeping my original score. --- Reply to Comment 1.1.1: Title: Thank you Comment: Thanks for the reviewer's acknowledgment of our efforts to address the concerns!	Summary: The paper introduces the Universal Mesh Movement Network (UM2N), a deep-learning model that enhances solving Partial Differential Equations (PDEs) through adaptive mesh movement. UM2N, with a Graph Transformer encoder and a Graph Attention Network (GAT) based decoder, is trained on a PDE-independent dataset, allowing zero-shot application to various PDEs and geometries without retraining. The model uses element volume loss to improve robustness and reduce mesh tangling. Strengths: 1/ Originality: The paper presents a novel approach that advances the field of adaptive mesh movement for solving Partial Differential Equations (PDEs). By combining a Graph Transformer encoder with a Graph Attention Network (GAT) decoder, the authors develop a unique method that can be applied zero-shot to various PDE types and boundary geometries without retraining. This approach overcomes the limitations of traditional and existing learning-based methods, which often require costly retraining and struggle with complex geometries. 2/ Quality: The quality of the paper is good, demonstrated through comprehensive evaluations on different PDE types, including advection, Navier-Stokes, and a real-world tsunami simulation. The experiments clearly show that UM2N outperforms existing methods in terms of accuracy, efficiency, and robustness. The use of a PDE-independent dataset for training and the adoption of element volume loss for reducing mesh tangling further highlight the thoroughness and rigor of the approach. 3/ Clarity: The methodology is clearly explained, and the results are presented in a way that effectively supports the claims made. 4/ Significance: The significance of this work is substantial, as it addresses a fundamental challenge in scientific and engineering simulations: solving PDEs efficiently and accurately. By enabling zero-shot application to various PDEs and geometries, UM2N offers a versatile tool that can be integrated into a wide range of numerical solvers, potentially benefiting numerous applications in geophysics, renewable energy, climate modeling, and aerodynamics. The improvements in mesh movement accuracy and reduction in computational costs have broad implications for advancing simulation technologies and their applications. 5/ Learning-based mesh generation and adaptation: One of the primary advancements is the use of a PDE-independent dataset for training. This allows UM2N to generalize across various PDE types and boundary geometries without the need for retraining. This is a significant improvement over existing methods that typically require retraining for new scenarios, making UM2N more versatile and practical for real-world applications. Another key strength is the incorporation of element volume loss in the training process. Unlike coordinate loss used in previous methods, element volume loss not only supervises the mesh movement but also penalizes negative volumes. This effectively reduces mesh tangling and enhances robustness. The architectural choice of combining a Graph Transformer encoder with a Graph Attention Network (GAT) decoder is particularly well-suited for handling the irregular structures of unstructured meshes. This graph-based architecture allows UM2N to capture both local and global information more effectively, leading to more accurate and efficient mesh movements. The attention mechanisms within these networks help in managing the complex dependencies and interactions within the mesh elements. Moreover, UM2N's ability to perform zero-shot generalization is another significant strength. The model can be applied to new, unseen problems without the need for additional training. This capability is particularly beneficial in practical applications where retraining models is computationally expensive and time-consuming. By using a learning-based approach for mesh movement, UM2N significantly improves the accuracy of numerical solutions while also reducing computational costs. This balance between accuracy and efficiency is crucial for advancing simulation technologies in various scientific and engineering domains. The method's ability to enhance accuracy without incurring additional computational overhead makes it an important tool for high-fidelity simulations. 6/ Monge-Ampère PDE-solver point of view: From the perspective of Monge-Ampère, UM2N enhances practical applicability by addressing computational and robustness challenges through a novel learning-based approach. Traditionally, Monge-Ampère-based methods offer attractive theoretical properties such as equidistribution and optimal transport but are computationally expensive and struggle with complex geometries. UM2N mitigates these limitations by decoupling the underlying PDE solve from the mesh movement process, focusing on learning the auxiliary Monge-Ampère equation. The integration of a Graph Transformer and GAT-based architecture allows UM2N to handle the complexities of mesh movement efficiently. The Graph Transformer captures both local and global information from the mesh, which is essential for accurately solving the Monge-Ampère equation. It processes the mesh as a graph, taking into account the positional relationships and interactions between mesh elements. The GAT-based decoder then uses this rich feature set to move the mesh nodes in a way that aligns with the desired mesh density distribution. UM2N's use of the element volume loss function is a significant technical advancement. This loss function is inspired by the equidistribution principle of the Monge-Ampère method, which aims to distribute the monitor function values uniformly across the mesh. By focusing on the volumes of the mesh elements rather than their coordinates, the element volume loss function ensures that the mesh adapts according to the desired density distribution. This approach also penalizes negative volumes, which helps prevent mesh tangling a common issue in traditional Monge-Ampère-based methods. Another important detail is UM2N's training on a PDE-independent dataset, which allows it to generalize well to different PDEs and boundary conditions without retraining. This is a significant improvement over traditional Monge-Ampère methods, which often require problem-specific adjustments. UM2N's ability to perform zero-shot generalization makes it highly versatile and practical for a wide range of applications. UM2N's robustness is further demonstrated through its successful application to scenarios with complex boundary geometries, such as the Tohoku tsunami simulation. Traditional Monge-Ampère methods often fail in such scenarios due to the challenges in solving the non-linear Monge-Ampère equation under these conditions. UM2N, however, effectively manages these complexities, maintaining high-quality mesh adaptation even in challenging geometries. Weaknesses: 1/ Comparison with State-of-the-art methods: While UM2N is compared with a few existing methods, the comparison could be broadened to include more recent state-of-the-art approaches. This would provide a clearer context for UM2N's performance and highlight its relative advantages and limitations more comprehensively. Expanding the experimental section to include comparisons with more recent and relevant learning-based mesh movement methods would be valuable. If new experiments are not feasible during the rebuttal, a detailed qualitative comparison discussing potential strengths and weaknesses based on the literature can be provided. 2/ Ablation Studies: The ablation studies are somewhat limited. A more comprehensive analysis of different components of UM2N, such as the impact of various hyperparameters, the contribution of each network component (Graph Transformer vs. GAT), and the sensitivity to different types of monitor functions, would provide deeper insights into the model's workings. The authors could add more ablation studies focusing on hyperparameter tuning, the role of different network components, and the sensitivity analysis to different monitor functions. 3/ Handling extreme geometries and mesh qualities: The current paper does not extensively explore how UM2N handles extremely poor-quality initial meshes or highly irregular and extreme geometries. These scenarios are crucial for practical applications, as real-world problems often involve complex and challenging geometrical domains. Poor-quality initial meshes, characterized by elements with highly skewed aspect ratios, large variations in element sizes, or even degenerate elements, can significantly complicate the mesh movement process. For instance, elements with aspect ratios far from unity can lead to numerical instabilities and inaccurate solutions due to poor approximation properties. Large variations in element sizes can cause localized errors to propagate or lead to uneven error distribution across the mesh, while degenerate elements with nearly zero area or volume can invert or fold during movement, leading to invalid meshes. Similarly, highly irregular geometries, such as those with sharp corners, narrow channels, intricate boundary shapes, or complex topological features like holes and multiple connected components, can test the robustness and adaptability of mesh movement algorithms. Sharp corners and narrow channels require fine mesh resolution to capture the geometric details accurately, while avoiding mesh element inversion or tangling. Intricate boundary shapes and complex topological features increase the difficulty of maintaining mesh quality and conformity to the domain boundaries during adaptation. A. Aspect Ratio and skewness: Maintaining aspect ratios close to one is crucial to minimize numerical errors. High aspect ratios can lead to poor interpolation properties, making it challenging to preserve or improve aspect ratios during significant mesh adjustments. High skewness causes poor approximation and solver instability. Controlling skewness during mesh movement to prevent highly distorted elements is critical. Suggestion: implement a loss term that penalizes high aspect ratios and skewness during training to encourage well-shaped elements. B. Element size variation: Uniformity and adaptivity: Balancing element sizes to ensure smooth transitions between refined and coarse regions is complex. Small elements can lead to computational overhead, while large elements might miss critical solution features. Dynamic adaptation: Efficiently refining in regions of high gradients and coarsening elsewhere requires dynamic strategies, balancing computational cost and accuracy. Suggestion: Use adaptive mesh refinement algorithms that dynamically adjust element sizes based on local error estimates and monitor function gradients. C. Mesh tangling: Inversion and Overlap: Preventing elements from inverting or overlapping during movement is essential. This requires strategies like regularization terms in the loss function to penalize negative volumes and ensure coherent adaptation. Local and Global Coherence: Ensuring local coherence in element displacement to prevent tangling, especially in regions with high gradient monitor functions, is challenging. Suggestion: Introduce a regularization term in the loss function to penalize negative volumes and use local displacement constraints to maintain element integrity. D. Boundary conformance: Complex boundaries: Accurately adapting the mesh to complex geometries with sharp features and varying topologies without introducing distortions requires careful refinement near boundaries. Boundary preservation: Techniques such as boundary snapping and boundary layer refinement are necessary to maintain accurate geometric fidelity. Suggestion: For instance applying boundary snapping techniques and boundary layer refinement to ensure elements conform accurately to the physical boundaries. E .Adaptive refinement and coarsening: Dynamic adaptation: Adapting the mesh based on evolving solution features, such as moving fronts, involves adding or removing elements dynamically. Smooth transitions: Ensuring smooth transitions between different resolution regions without creating poorly shaped elements is crucial. Hierarchical refinement and mesh smoothing techniques are often employed. 4/ Theoretical guarantees and limitations: While the paper mentions that there are no theoretical guarantees that mesh tangling will be completely prevented, it would be beneficial to delve deeper into this limitation. Understanding the theoretical bounds of UM2N's performance and the conditions under which it might fail is important for setting realistic expectations. 5/ Real-world application scenarios: Although the tsunami simulation is an excellent example, including more diverse real-world application scenarios would further demonstrate UM2N's versatility and robustness. For instance, applications in aerodynamics, biomechanics, or climate modeling could be explored. Suggestion: Add more case studies or at least detailed discussions of how UM2N could be applied to other real-world problems, emphasizing its practical benefits and potential challenges in these domains. 6/ Computational efficiency and scalability: The paper discusses the reduction in computational costs compared to traditional methods but does not provide detailed benchmarks or discussions on the scalability of UM2N for very large meshes or real-time applications. Suggestion: Include more detailed benchmarks of computational efficiency and scalability, comparing UM2N's performance on large-scale problems with other methods. Discuss strategies for optimizing performance and potential bottlenecks in the current implementation. 7/ Standpoint from Multi-Scale phenomena: The paper does not address the specific challenges posed by multi-scale phenomena, where different parts of the domain may require vastly different resolutions to capture fine-scale features accurately while maintaining computational efficiency. Multi-scale phenomena often involve a wide range of spatial and temporal scales, making it difficult to balance resolution and computational cost effectively. Suggestion: Introduce experiments that specifically target multi-scale phenomena, demonstrating UM2N's ability to adapt the mesh dynamically to capture fine-scale features in high-gradient regions while coarsening in less critical areas. This could involve benchmarks on problems known for their multi-scale nature, such as turbulent flows or geophysical simulations. Additionally, discuss potential enhancements to the model, such as multi-grid techniques or hybrid approaches that combine UM2N with other multi-scale modeling strategies, to better handle these complex scenarios. 8/ Analysis from mesh continuity, local and global deformation viewpoints: The paper does not thoroughly address how UM2N manages local and global deformations of the mesh, which are essential for accurately capturing complex physical phenomena. Effective mesh adaptation requires the ability to handle fine-scale local deformations to capture detailed features and large-scale global deformations to adapt to overall changes in the domain. Additionally, ensuring mesh continuity during these deformations is crucial to maintain the fidelity and stability of numerical simulations. Inadequate handling of these deformations can lead to poor resolution of critical areas, excessive computational costs, and potential discontinuities that degrade the accuracy of the simulation. A/ Coupling local and global deformations: Balancing local and global deformations is crucial for maintaining overall mesh quality and continuity. This involves ensuring that local refinements do not introduce excessive computational overhead and that global adaptations do not degrade the resolution of critical regions. Achieving this balance is technically challenging and requires sophisticated mesh adaptation algorithms that can handle both fine-scale and large-scale changes seamlessly. B/ Mesh quality and continuity maintenance: During both local and global deformations, maintaining high-quality elements (in terms of aspect ratio, skewness, and smoothness) is essential to avoid numerical instability and ensure accurate simulations. Poorly shaped elements can significantly degrade the performance of numerical solvers. Additionally, ensuring continuity in the mesh, where element shapes and sizes transition smoothly, is vital for maintaining numerical stability and accuracy. 9/ Analyze the optima of the learned meshes: The paper does not address the issue of ensuring that the learned mesh configuration is globally optimal, nor does it consider the potential existence of multiple local optima in the optimization landscape. This oversight can result in suboptimal mesh configurations, which may not fully capture the desired features of the simulation domain or may introduce unnecessary computational overhead. In complex adaptive meshing scenarios, the risk of converging to local optima rather than the global optimum can significantly impact the accuracy and efficiency of the numerical solutions. The optimization process for adaptive mesh movement often involves a highly non-convex loss landscape with many local minima. A single global optimum represents the best possible configuration of the mesh in terms of accuracy and computational efficiency. However, due to the complex nature of the problem, the optimization algorithm may converge to multiple local optima, each providing a suboptimal solution that fails to maximize the potential benefits of the adaptive meshing process. The presence of multiple local optima can lead to inconsistent mesh configurations across different runs or simulations. This variability can make it difficult to ensure that the mesh is optimally adapted to the specific features of the simulation domain. Inconsistent convergence can also lead to variability in simulation results, reducing the reliability and robustness of the numerical methods. Suggestion: Benchmark the learned mesh against known optimal solutions or high-resolution reference meshes. This benchmarking can help assess how close the learned meshes are to the ideal configuration and identify specific areas where the model falls short. Technical Quality: 4 Clarity: 3 Questions for Authors: 1/ How does UM2N handle extremely poor-quality initial meshes and highly irregular geometries? Suggestion: Conduct additional experiments with initial meshes characterized by highly skewed aspect ratios, large variations in element sizes, and complex boundary geometries. Include quantitative metrics on mesh quality before and after adaptation to highlight the robustness of UM2N. 2/ How does UM2N ensure mesh continuity and high-quality element shapes when dealing with both local and global mesh deformations? 3/ How does UM2N's learned mesh configuration compare to known optimal solutions or high-resolution reference meshes? Are there benchmarks or validation cases included to demonstrate the accuracy and efficiency of the mesh adaptation? Suggestion: Benchmark the learned mesh configurations against known optimal solutions or high-resolution reference meshes. Include comparisons using quantitative metrics such as error norms and gradient capture to highlight the accuracy and efficiency of UM2N in approaching the optimal mesh configuration. 4/ How does UM2N perform in resolving boundary layers and turbulent flows? Could you provide experimental results or benchmarks that validate its effectiveness in these critical fluid dynamics applications? Suggestion: Conduct targeted experiments on canonical boundary layer cases (e.g., flow over a flat plate) and turbulent flow benchmarks (e.g., turbulent channel flow). Use quantitative metrics like wall shear stress and turbulence intensity to evaluate performance and compare with high-fidelity simulations or experimental data. 5/ Can you elaborate on the rationale behind selecting Graph Attention Networks (GAT) over other Graph Neural Network (GNN) methods for mesh movement? Additionally, have you considered or benchmarked against hierarchical GNNs, Octree-based methods, or point cloud-based methods, such as those utilizing multi-scale architectures like U-Net? Exploring these alternatives could provide insights into potential improvements in handling mesh adaptation across different scales and complex geometries. if the authors address the weaknesses and questions outlined above, I would be happy to increase my score. Doing so will significantly enhance the robustness and applicability of the proposed work. Confidence: 3 Soundness: 4 Presentation: 3 Contribution: 3 Limitations: Yes. Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We thank Reviewer yaVX providing a detailed summary of our strengths and valuable suggestions. Below, we present our detailed responses to the comments, indicating [W] for weaknesses and [Q] for questions. Response to W1: As the reviewer correctly identifies we do compare our method with existing methods, focussing on the M2N method. The reason for the limited number of comparisons is because there really has not been a lot of research so far on learning-based mesh-movement methods, most of the literature has focused on acceleration of other mesh adaptation strategies. We do cite the very recent flow2mesh paper by Jian Ju et al. (ref. [32] in our paper) who confirm our view: “However, these works all focused on mesh refinement rather than the mesh movement" Response to W2: We perform an additional ablation study of our network architecture on the Swirl case, analysing the contribution of each network component. We construct two models, which replace the GAT Decoder with one layer GAT denoted as UM2N-w/o-Decoder and remove the graph transformer denoted as UM2N-w/o-GT. The results are shown in Table R1. Both variants give worse results compared to the full UM2N. A visualization is also shown in Figure R2. It can be observed that without Decoder, the model distorts the shape of the ring, i.e., missing relative information between vertices; without the GT, the model fails to capture the details of the ring shape. Response to W3/W4/Q1: Thanks for raising this question. We include an additional experiment dealing with poor quality initial meshes as shown in Figure R4 and an additional experiment for irregular geometries. Also note the real-world Tsunami case shown in Figure 5 of our paper is also a good example for highly irregular geometries. And more cases with complex geometries are shown in our supplementary video. Figure R4 shows a low-quality input mesh with highly anisotropic elements. Using a monitor function based on the l2-norm of the gradient of the solution of an anisotropic Helmholtz problem, we tried applying conventional MA solvers and found that both the quasi-Newton and relaxation approaches failed to converge and/or resulted in tangled meshes, even though the elements are already aligned with the anisotropy of the Helmholtz solution. The UM2N approach, however, was able to successfully apply mesh movement without tangling. We would also like to clarify that our work targets mesh movement that dynamically adapts to a PDE solution, which is not a replacement for mesh generation in general, so we would always expect a reasonable input mesh when applying our model. In Figure R1, we show a stress test for both MA and UM2N. It can be observed that MA soon fails in the presence of non-convexity in the geometry while UM2N only gets tangled in the most challenging right angle case. It indicates that our method allows a relaxation of the geometric limitations compared to MA, although we admit that there is no theoretical guarantee for no mesh tangling as we mentioned in the limitation. Response to W5/W6: We appreciate that the reviewer acknowledges that our Tsunami case is an excellent real-world scenario application. We also kindly remind that there are also more applications shown in our supplementary video. We do agree more examples for different areas would better demonstrate the versatility of UM2N. Due to the time limit in the rebuttal period, we would like to leave this to our revisions. For some limited scalability, we further apply our trained model on a flow past cylinder case with ~11k vertices and ~22k triangles as shown in Figure R5. It is shown that our UM2N works well on this case with more degrees of freedom compared to the one in the paper. For an even larger case, we perform stress tests using a time-independent Helmholtz case on our RTX 3090 24 GB GPU and observe that the maximum scale of the problem that can be run on this GPU has ~50k number of elements. The inference time is ~760 ms, the MA method on the same problem requires ~37800 ms with a residual threshold 1e-4. Considering this limitation, applying the memory efficient transformer (linear-attention transformer for further improvement of the inference efficiency) is targeted as future work. Response to Q2/W8: There are two types of mesh continuity to consider: continuity in space (i.e. mesh smoothness) and continuity in time (avoiding large changes between timesteps). The latter is promoted by training with samples based on Monge Ampere solutions. Based on optimal transport theory, these represent the unique minimal change transformation that satisfies the equidistribution principle provided certain smoothness and convexity criteria are satisfied. Given that this is already the unique optimal mesh that with these properties, quality of elements in terms of their aspect ratio is not explicitly enforced, but the minimal change property does mean a high quality input mesh is transformed to a close-in-quality output mesh as long as the required transformation in terms of the monitor function is not too demanding (e.g. non-smooth). It should be noted it is not the case that equal-aspect triangles always provide the highest quality mesh. This in fact depends on the anisotropy of the PDE solution, and in cases where more control over the aspect ratio is required, other mesh movement methods allow for a tensorial input metric, but this necessarily means the minimal change (continuity) property is no longer satisfied in general. There is thus always a trade-off between different desired properties. Likewise, when asking for a considerable global redistribution of degrees of freedom this may come at the cost of the local mesh quality. Any mesh movement method will therefore be a compromise with UM2N focusing on maintaining minimal movement and maintaining desired output cells areas. Please refer to Supplement Response to yaVX in general response for more responses Q3/W9, Q4, Q5. --- Rebuttal Comment 1.1: Title: Response to the rebuttal and follow-up Comment: Thank you for providing detailed responses to my comments and questions. I appreciate your efforts to address the points raised and for conducting additional experiments to clarify the strengths and limitations of the Universal Mesh Movement Network (UM2N). Here are some additional thoughts. Comparison with State-of-the-art methods: I understand the challenge in comparing UM2N with a broader set of state-of-the-art methods due to the limited number of learning-based mesh movement approaches. Your focus on the M2N method and the confirmation from the Flow2Mesh paper about the scarcity of learning-based mesh movement methods is noted. However, as the field evolves, it might be beneficial to continually update your comparisons with new emerging methods. Exploring the literature for recent advancements and including these in future versions of the paper would be valuable. Ablation studies: Thank you for conducting additional ablation studies. The results provided in Table R1 and Figure R2 help elucidate the contributions of each network component. It would be interesting to see a more detailed analysis of how different hyperparameters impact the performance of UM2N, which could be explored in future work. Handling poor-quality meshes and irregular geometries: Your additional experiments on poor-quality initial meshes and irregular geometries, including the real-world tsunami case and stress tests, effectively demonstrate UM2N's robustness. The clarification that UM2N targets mesh movement rather than mesh generation is helpful. It’s commendable that UM2N performs well even with challenging input meshes. However, it would be beneficial to include detailed quantitative metrics in future revisions to further emphasize UM2N's capabilities. Computational efficiency and scalability: The scalability experiments and the comparison with the MA method are informative. It’s promising to hear about future work involving memory-efficient transformers for improved inference efficiency. Continued focus on optimizing computational performance and discussing potential bottlenecks will be crucial as UM2N scales to even larger and more complex problems. Mesh continuity and quality: Your explanation of the trade-offs between mesh quality, continuity, and the minimal movement property is insightful. It provides a clear rationale for UM2N's design choices. Future work could explore strategies for balancing these trade-offs, especially in scenarios requiring significant global redistributions. Optimal mesh configurations: I appreciate your discussion on the challenges of proving mesh optimality and the use of PDE solution error reduction rate as a metric. The references to high-resolution meshes and your detailed error analysis are compelling. Benchmarking against known optimal solutions or employing more sophisticated error metrics could further strengthen these comparisons in future work. Performance in boundary layers and turbulent flows: The experiments on flow past parallel plates provide valuable insights into UM2N’s capability to resolve boundary layers. Continued exploration of turbulent flow scenarios, along with detailed performance metrics, would further solidify UM2N’s applicability in critical fluid dynamics applications. --- Rebuttal 2: Title: Summary Comment: The authors have provided a comprehensive and thoughtful rebuttal, effectively addressing the primary concerns and questions raised about the UM2N architecture. They clarified their focus on mesh movement rather than generation and highlighted the challenges of comparing their work with the limited number of learning-based mesh movement methods available. The additional experiments conducted, especially on poor-quality initial meshes and irregular geometries, demonstrate UM2N’s robustness and versatility. Their detailed responses on the integration of Graph Transformer and Graph Attention Network (GAT) components, the role of element volume loss, and handling mesh deformations provide valuable insights into the technical strengths of their approach. However, as the field evolves, ongoing comparisons with new methods, further exploration of hyperparameter sensitivities, and expansion of real-world applications will strengthen the work’s impact and applicability. Overall, the authors have made a strong case for the effectiveness and innovation of UM2N, particularly through their emphasis on zero-shot generalization and computational efficiency. The clear explanations and additional experiments help illustrate the model's capabilities in diverse scenarios. I appreciate the thoroughness and clarity of the paper, as well as the authors' proactive engagement with the feedback. The improvements and insightful discussions provided in the rebuttal demonstrate the paper's potential to advance simulation technologies through adaptive mesh movement.	Summary: In the present work, authors tackle the challenging task of accelerating PDE-solving process with deep learning, focusing on mesh movement problem. They suggest a new architecture that combines graph transformer encoder and graph attention decoder to significantly accelerate the PDE-solving. Authors propose a way to decouple the underlying PDE solving from the mesh movement process itself. It makes the approach universal with regard to solvers as well as to boundary geometries and mesh structures. They demonstrate the superiority of their method over conventional solver, and the ability on various benchmark tasks. Strengths: The work is well written and concise. The work focuses specifically on the r-adaptation, i.e. mesh movement problem. Authors carefully describe what the complete problem is and which domain their method attributes to. This might look excessive but overall leaves good impression and facilitates understanding. The approach itself of generating the dataset of random fields and train the model to learn an auxiliary PDE sounds compelling and indeed universal. It allows them to effectively decouple the monitor values from the PDE solved, which is one of the main strengths of the work. The use of data, the architecture, the losses used and other deep learning-related details are well explained; the losses and the input varioations are justified in the convincing ablation study. Authors claim that they would like to provide the code for reproducibility, which is always for the better. Weaknesses: Authors do not demonstrate a single experiment, where their model provides mesh tangling. Authors explain such effect by the use of the volume loss (line 326) calling it "tangling-aware". In general, it might be good, since one can be sure that any use of UM2N is reliable. Yet, without proper evaluation (or at least, clarification for such perfection), It cannot be regarded as a virtue. It implies either that the experiment setups are chosen to be weak or the model has the inductive bias that might limit its universality. In other words, there might exist some setups that the model won't be able to estimate correctly, because of oversmoothness. As for example, one can imagine the domain with a tight bottleneck, the flow might not propagate to the other side at all. Another side of the same issue can be seen in the flow-past-cylinder experiment. Both MA and M2N provide some tangling and hence considered as "Failed", so entire discussion in this experiment is in favor of the proposed UM2N. Authors call the setup "classic and challenging", yet there is not a single baseline that provides at least some solution. Either another baseline that actually gives some number should be provided or the setup should be simplified, e.g. choosing the object that provides less turbulence, such as the airfoil. Technical Quality: 2 Clarity: 4 Questions for Authors: I would appreciate if authors discuss some of the issues I raised above in the Weaknesses. Mainly: non-tangled results of other methods on the flow-past-cylinder; tangled results of the UM2N model itself Confidence: 4 Soundness: 2 Presentation: 4 Contribution: 4 Limitations: - Flag For Ethics Review: ['No ethics review needed.'] Rating: 5 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We like to thank the reviewer for their kind words on the presentation and level of contribution of our paper. Below, we present our detailed responses to the comments, indicating [W] for weaknesses and [Q] for questions. We appreciate the feedback regarding the amount of explanation we were able to provide regarding limitations of our approach and validation of the flow past a cylinder case. As the reviewer correctly identifies our paper focuses specifically on r-adaptation which has benefits over other mesh adaptation methods - maintaining mesh topology, and, in the Monge-Ampere (MA) approach, minimising the amount of change - but also drawbacks: solving the MA equations is very costly, and it puts limitations on the physical geometry, demanding smooth boundaries and convex domains. In our paper, we demonstrate a significant reduction in the cost through the UM2N acceleration. We also demonstrate a relaxation to some extent of the geometric limitations, but this is indeed a bit harder to outline and quantify. Response to W1/Q1:‘tangled results of the UM2N model’: There are in fact still cases where UM2N might fail, we perform additional experiments as shown in Figure R1. Starting from a rectangle, we gradually distort the geometry to a more non-convex shape i.e., the case is more and more challenging from top to bottom. The UM2N finally fails in the case surrounding sudden jumps in the boundary accompanied with a large variation in the required resolution as shown in the bottom row of Figure R1: a sudden constriction of the flow leads to tangling with UM2N in the two left corners of the channel. In a more smoothed version UM2N does produce a valid mesh, where the original MA method still fails (second to bottom rows in Figure R1). We agree that this warrants a more extensive discussion in our paper in the revision. Response to W2/Q2: ‘non-tangled results of other methods on the flow-past-cylinder’: We additionally perform a non-tangled flow-past-cylinder shown in Figure.R6 as mentioned by the reviewer xMNp. We simplified the setting: lower the Rrenynolds number by reducing the inflow velocity and set the top and bottom slip boundary, which finally makes it a laminar flow. As shown in Figure.R6, both the UM2N and M2N give non-tangled meshes and it can be observed that UM2N better captures the dynamics i.e., adapts to the PDE solution of stable state compared to the High Res solution. The quantitative results shown in the Tab.R3 also indicate that UM2N perform better than M2N regarding to error reduction. The aim of the flow past a cylinder case was to demonstrate that mesh movement can now readily be applied, which is not feasible with a classical approach like MA for several reasons: the non-convexity of the domain means the equations are fundamentally ill-posed, and, even if that could be overcome, e.g. by applying mesh movement only in the right half of the domain, the prohibitive costs to apply this every timestep. This case is based on the so called DFG-benchmark 2D-2 [1], which has been modelled with a variety of other frameworks, including those that incorporate other mesh adaptation approaches [see e.g. 2-5] The results in our paper were validated by comparison with a high resolution uniform-mesh model, which corresponds well with the reference value in [2,3]. Note that convergence to the reference depends on various discretisation choices, finite-element pair, type of boundary condition, etc., which are not the focus of our paper. We therefore decided to only compare with a high resolution case, using the same discretisation details, to demonstrate the improvement in accuracy through mesh movement alone. We would like to clarify this however in our paper, and make better reference to and comparison with published results for this case. [1] https://wwwold.mathematik.tu-dortmund.de/~featflow/en/benchmarks/cfdbenchmarking/flow/dfg_benchmark2_re100.html [2] D. Capatina et al. ‘22 https://doi.org/10.1016/j.cma.2019.112775 [3] V. John ‘04 https://onlinelibrary.wiley.com/doi/abs/10.1002/fld.679 [4] T. Coupez et al. ‘13 https://doi.org/10.1016/j.jcp.2012.12.010 [5] Hachem et al ‘13 https://doi.org/10.1002/nme.4481 --- Rebuttal Comment 1.1: Title: reply to Rebuttal Comment: I thank authors for their clear and instructive reply to all the issues I raised. Authors added new experiments and showed where their method fails and provides mesh tangling. Also, they showed non-tangled results for flow-over-cylinder for the MA method. I believe this was not very informative when the proposed method seems to perform fine, while others completely fail (even if this fail is inherent). Authors found where this fail boundary lies and illustrated a more compelling comparison; this definitely improves the overall picture. --- Reply to Comment 1.1.1: Title: Thank you! Comment: Dear Reviewer xMNp, Thanks again for the time and effort you've invested in reviewing our manuscript. We also greatly appreciate your acknowledgment of our efforts to address your concerns. Your insights have been instrumental in improving our work, and we are committed to incorporating your suggestions into the revision. We would also greatly appreciate your consideration in possibly raising the original rating if you find our responses satisfactory. If you believe there are any remaining areas where additional clarifications/responses could contribute to such a higher rating, please don't hesitate to inform us. Your guidance is always crucial to improve the quality of our submission. Thank you very much, Authors	Rebuttal 1: Rebuttal: We are glad to receive valuable and constructive comments from all the reviewers. We have made a substantial effort to clarify reviewers' doubts and enrich our experiments in the rebuttal phase. In our responses, Tab. Rxx or Fig. Rxx refers to the new Rebuttal results in the attached PDF. Below is a summary of our responses: Reviewer xMNp: 1. We perform additional experiments to demonstrate a failing case for UM2N and to explore the limits to its ability to avoid mesh tangling (Figure. R1) 2. We demonstrate how UM2N provides a relaxation to some extent of the geometric limitations compared to the Monge-Ampère (MA) based method. (Figure. R1) 3. We conduct additional experiments to show non-tangled results of other methods on the flow-past-cylinder (Figure. R6), and provide an explanation on why MA inherently fails on flow-past-cylinder. Reviewer yaVX: 1. We perform an additional ablation study on network components (Table.R1/Figure.R2) to demonstrate the effectiveness of both Graph Transformer based encoder and Graph Attention Network based deformer. 2. We conduct an additional experiment dealing with poor quality initial meshes (Figure R4). 3. We provide a detailed discussion about mesh continuity in mesh movement. 4. We provide a clarification for optimality of mesh movement and explain our decision on the metric we use for evaluating mesh movement. We also provide measures between model output mesh and target mesh. (Tab. R2) 5. We perform an additional boundary layer experiment. (Figure.R3) 6. We explain the rationale for selecting Graph Transformer and Graph Attention Network and discuss the other related networks as potential future work. Supplement Response to yaVX: Response to Q3/W9: Here we would like to discuss both the comparisons to optimal mesh and reference PDE solutions. It is generally hard to strictly prove the optimality of a mesh in cases with more than one dimension, and, as argued previously, there are various definitions of optimal that can be considered, and, in a sense, the meshes obtained from the MA method can be regarded as optimal meshes already. In Table R2, we provide the coordinate loss, element volume loss and Chamfer distance during training. Note that these metrics for measuring the difference between model output mesh and MA reference mesh are intermediates. The ultimate metric we care about is the PDE solution error reduction rate (ER), which is why we focus on ER in our paper. We provided reference solutions for both Swirl and Flow past cylinder cases in our paper which are computed on high-resolution meshes (~10k for swirl, ~260k for flow-past-cylinder), indicated as ‘High Res’. As shown in Figures 2 and 3 of the paper, the solution obtained on a UM2N moved mesh has a smaller L2 error (plots in right part of Figure 2) and more accurate drag coefficient (bottom part of Figure 3) compared to that obtained on an original uniform mesh. These errors are computed against the reference, high-res PDE solution. Response to Q4: In our flow past the cylinder case, at the top and bottom we imposed non-slip boundary conditions which already promotes a boundary layer similar to the suggested flow over a flat plate. For a clearer illustration, we have now additionally conducted a flow past two parallel plates experiment. We observe that the velocity parallel to the boundary ($u_y$) obtained from UM2N moved mesh better aligns with the high resolution results compared to the results on uniform mesh as shown in the Figure R3. Response to Q5: Thank you for this inspiring question. Graphs are a natural way to represent unstructured meshes. We apply the graph transformer as encoder for its powerful expressivity, and apply the graph attention network as decoder because it can naturally help alleviate mesh tangling with some extra design. More specifically, we update the coordinate of each vertex with the combination of its neighbours, the coefficients of which are determined by the GAT module. This guarantees each vertex to only move within the convex hull of its neighbouring vertices, hence effectively alleviating mesh tangling issues. We thank the reviewer for their suggestions for other benchmark methods. We agree these methods have their strengths but not all are equally relevant to our fully unstructured mesh based approach. For example, octree-based methods can be only applied to octree-based hierarchical meshes; point-cloud-based models ignore the topological relationship between vertices. Hierarchical models or multi-scale architectures such as U-Net is a very interesting idea, as the input and output of the model share the same structure, and capturing and sharing local and global information with hierarchical model structures is appealing, hence is definitely worth further investigation. In general, we agree with the reviewer that exploring these methods for mesh adaptation are interesting directions for future work. Reviewer jEN5: 1. We perform an additional ablation study on network components (Table.R1/Figure.R2) to demonstrate the effectiveness of both Graph Transformer based encoder and Graph Attention Network based deformer. 2. We compare our method to related work qualitative and discuss the potential future direction. 3. We discuss the intuition and limitation of our model handing boundary with complex geometries. 4. We conduct additional experiment on a larger case (Figure.R5) and investigate the scalability based on our computational resources. Reviewer Jzyt: 1. We provided clarification about the computation of monitor values as well as our pipeline. 2. We provide explanation about the chamfer distance issue the reviewer concerns. 3. We conduct an additional experiment dealing with poor quality initial meshes and visualize the results. (Figure R4). Pdf: /pdf/da392d47b266c26a98bd4b7432427d3ff8a605cc.pdf	NeurIPS_2024_submissions_huggingface	2,024	null	null	null	null	null	null	null	null
Automated Multi-Task Learning for Joint Disease Prediction on Electronic Health Records	Accept (poster)	Summary: The paper introduces AutoDP, a novel AutoML framework that performs combined task grouping and architecture search with a novel surrogate model. The proposed model achieved higher performance in multiple tasks on MIMIC VI dataset compared to common single-task, multi-task, and AutoML approaches. Strengths: - Novel AutoML approach for EHR tasks - Superior performance when compared to the traditional approach - Evaluated with the adequate dataset - Good range of baselines - Ablation study included - Open access data and code - Clear methodology presentation - Results contribute to AutoML and EHR fields Weaknesses: - It claims "feasible search cost", but there is no runtime comparison or evaluation Technical Quality: 4 Clarity: 3 Questions for Authors: - How much longer does it take to run compared to DARTS and other baselines? Confidence: 4 Soundness: 4 Presentation: 3 Contribution: 3 Limitations: The traditional limitation in AutoML work is the computational budget. The paper does not discuss the runtime of the proposed model and alternatives. Ethics limitations are discussed in the supplementary material. Flag For Ethics Review: ['No ethics review needed.'] Rating: 8 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thank you very much for providing such positive feedback to our work. For the runtime concern, we include the specific GPU hours in Table 4. MTG+DARTS has the same runtime as our method. We maintain this for a fair comparison with the baselines. We have shown that the computational cost is feasible for the problem setting in our work. Also, as the surrogate modelling method is a generally efficient search algorithm, it has the potential to scale to larger scenarios. --- Rebuttal 2: Comment: Dear Reviewer azWt, I am a NeurIPS 2024 Area Chair of the paper that you reviewed. This is a reminder that authors left rebuttals for your review. We need your follow up answers on that. Please leave comment for any un-answered questions you had, or how you think about the author's rebuttal. The author-reviewer discussion is closed on Aug 13 11:59pm AoE. Best regards, AC	Summary: The paper proposes an automated multi-task learning framework, AutoDP, for disease prediction using EHR data. It optimizes task grouping and model architecture to enhance prediction performance. AutoDP efficiently searches a vast space of task combinations and architectures by employing a surrogate model-based optimization approach. Experiments show it outperforms existing methods. Strengths: 1. The proposed AutoDP automates the search for optimal task groupings and model architectures simultaneously, reducing the reliance on human experts and improving the efficiency of designing multi-task learning frameworks. 2. The experimental results demonstrate that AutoDP outperforms existing hand-crafted and automated methods on real-world EHR datasets, achieving higher multi-task learning gains. Weaknesses: 1. Generalization Concern: The paper shows AutoDP's effectiveness beyond disease prediction on EHR data. However, its effectiveness and applicability in other domains remain uncertain. 2. Computational Scalability: The proposed method involves a vast search space and significant computational costs. While feasible within the scope of the study, the proposed approach may hinder scalability in larger real-world scenarios. Technical Quality: 3 Clarity: 3 Questions for Authors: 1. Mapping Assumption Validity: The paper assumes a learnable mapping for multi-task gains, which needs more explanation. 2. Comments on Disease-Based Grouping: As mentioned in Appendix D, Grouping similar diseases might make it hard to discriminate between them, reducing MTL effectiveness. 3. Reporting individual task gains is necessary to show the effectiveness of the proposed method on MTL. 4. Are those gains robust? Will the proposed method result in overfitting? Confidence: 4 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: 1. The paper lacks a discussion on model interpretability, which is a critical aspect for trust and clinical decision-making in healthcare. Flag For Ethics Review: ['No ethics review needed.'] Rating: 5 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Weaknesses: 1. We focus on the disease prediction problem in this paper. Our contention is that this problem is a very important problem with potential for high impact such that the method should be publicized for practitioners and users of ML in the field. However, we believe but have not shown that our method has the potential to generalize to other domains, as the framework design can be easily adapted to other problems. We will leave this to future work. 2. Our method has involved several techniques to address the efficiency issue, such as surrogate modelling, efficient sampling method, and efficient search method. While we do not test the framework to larger scenarios, it has the potential to scale up in larger search spaces. Our computational needs are in the same order of complexity as without doing AutoDP. If in a larger setting, there are no resources to run our method, most probably there will be not enough resources to run the baseline methods themselves due to the size of the problem. In such cases, actually, one can use our method because we use a sampling-based method and while we have a smaller sample the results may not be as good, it is actually quite efficient and effective because of the smart sampling methods we have designed. As long as the problem we are solving has some low-dimensional patterns for the search space, the surrogate model has the potential to learn the mapping with a relatively small number of samples. Thus, in most cases, the proposed framework can maintain the computational cost within a feasible range. Questions: 1. Every pair of task combinations and architecture has unique values of multi-task gains, which satisfies the definition of a function. As neural networks are well known to be good at fitting black-box functions, we can assume that it can learn the mapping to multi-task gains. 2. We perform this ablation study to show the necessity of using search algorithms to find better task groupings. The searched configuration does not necessarily follow the similar disease grouping, but it can achieve higher performance gains, which can maximize the MTL effectiveness. In other words, disease-based grouping does not perform as well as our method as shown in our ablation studies. It does not obviate the need for MTL. 3. We report the individual task gains at Appendix B. We will also include the specific values in the camera-ready version. 4. We conduct multiple rounds of experiments and report the means and variances of performance gains, which shows that our method is robust and consistently performs better. Also, our method does not overfit since we are performing a search algorithm and we use intelligent sampling essentially. The evaluation of any samples from the search space is independent of each other. Also, during search, we use the validation set to compute multi-task gains for each sample, so essentially we are choosing the configurations that have the best generalization ability. Limitations: 1.We include two case studies at Appendix E for showing the searched configurations under the settings of Task @ 10 and Task @ 25. By analyzing the searched configuration, we can interpret how AutoDP groups tasks together and what kinds of architectures are actually effective to modelling EHR time series. Admittedly, if we had space, we could address all of these issues to a greater extent. However, given these questions have been raised, we will address all of these at least to the extent we can fit them in the camera-ready paper and have fuller discussions in a full version on arXiv, which we will point to since we cannot fit answers to all of these in a conference paper. We regret the inconvenience. --- Rebuttal Comment 1.1: Title: Reply to authors Comment: Thank you for your response and the additional information provided. After carefully considering the paper and the rebuttal content, I have decided to maintain my initial evaluation, based on a comprehensive assessment of the work's contribution to the field. Additional clarity to my point Q2: I want to suggest that in the disease prediction domain, it is less likely to provide many benefits (in terms of performance) to group similar diseases into one learning task, as typically differential diagnosis is even harder. Therefore, claiming the proposed method performs better than such a method is less attractive. Thank you for your efforts and good luck.	Summary: N/A Strengths: N/A Weaknesses: The authors' identity can be easily inferred by googling the linux username ("sxc6192") found in the .idea/deployment.xml file in the supplementary material. Technical Quality: 1 Clarity: 1 Questions for Authors: N/A Confidence: 5 Soundness: 1 Presentation: 1 Contribution: 1 Limitations: N/A Flag For Ethics Review: ['No ethics review needed.'] Rating: 1 Code Of Conduct: Yes	null	Summary: The paper discusses an automated approach for multi-task learning on electronic health records, called AutoDP. This approach aims to improve the design of task grouping and model architectures by reducing human intervention. Specifically, AutoDP searches for the optimal configuration of task grouping and architectures simultaneously. The document mentions that existing MTL frameworks for EHR data suffer from limitations such as requiring human experts to select tasks and design model architectures. AutoDP proposes a surrogate model-based optimization approach to address these limitations. The document also details related work in multi-task learning with EHR data, multi-task grouping, and multi-task neural architecture search Strengths: The paper introduces a novel approach, AutoDP, that addresses limitations in current MTL frameworks for EHR data by automating task grouping and model architecture design. The idea of using a surrogate model-based optimization for joint search is innovative and holds promise for efficient exploration of the search space. The paper is well-written and the overall presentation is clear. Weaknesses: The paper could benefit from a more detailed explanation of the architecture search space, particularly the types of operations allowed in the directed acyclic graph (DAG). Additional details regarding the progressive sampling strategy for the surrogate model would be helpful to understand how it balances exploration and exploitation during search. The paper mentions experiments on real-world EHR data but lacks specific results or comparisons with other MTL approaches to demonstrate the effectiveness of AutoDP. Technical Quality: 2 Clarity: 2 Questions for Authors: Can the authors elaborate on the specific operations allowed in the DAG-based architecture search space? Are there any constraints on the number or types of operations that can be included? The paper mentions a progressive sampling strategy for the surrogate model. Can the authors provide more details on how this strategy works, particularly how it balances selecting informative data points for the surrogate model and exploring new areas of the search space? While the paper mentions using AutoDP on real-world EHR data, it would be beneficial to see concrete results and comparisons with other MTL approaches to assess the performance gains achieved by AutoDP. The use of EHR data raises ethical concerns about patient privacy. The authors should ensure proper anonymization techniques are used throughout the development and application of AutoDP. Confidence: 3 Soundness: 2 Presentation: 2 Contribution: 3 Limitations: Limited Evaluation: The experimental validation of AutoDP is confined to a single dataset (MIMIC-IV). Evaluating the framework on additional EHR datasets would bolster the generalizability of its findings. Lack of Interpretability: The paper could provide deeper insights into the specific task groupings and architectures discovered by AutoDP. Visualizations or case studies could help elucidate the underlying reasons for the observed performance improvements. Privacy Concerns: The paper does not explicitly address privacy concerns related to EHR data. Extending AutoDP with data processing pipelines for automatic feature engineering could offer enhanced privacy safeguards. Other concerns: This paper did not discuss the data imbalance issues as some diseases might have large and well-annotated data, while others might have small and many missing data. For disease prediction, there are similarities among some diseases, and also training on more disease can help to predict other similar diseases. Flag For Ethics Review: ['No ethics review needed.'] Rating: 4 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: 1. Architecture search space: We introduce our search space at Appendix A. We will try to move parts of it into the main paper to make the full paper clearer. We define the candidate operation set as {Identity, Zero, FFN, RNN , Attention}, which includes widely used operations for processing EHR time series. Identity means that we keep the input the same as the output. Zero means we output all zero tensors with the same shape as the input. FFN means we use feed-forward layers to process the input time series (same as defined in Transformer[1]). RNN means we use recurrent neural networks to process the input feature (we use LSTM in our experiments). Attention means that we use a self-attention layer to handle the input (same as defined in Transformer). [1] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Łukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Advances in neural information processing systems, pages 5998–6008, 2017. 2. Details regarding progressive sampling: We introduce the detailed procedure in Algorithm 1. To balance the exploitation and exploration, we estimate the mean and variance of the sampled task combinations. And then, we use upper confidence bound as the acquisition function to balance the exploitation and exploration via a predefined parameter $\lambda$. 3. Performance comparison: We conduct comprehensive experiments to compare the proposed method with existing works. All the results are included in Table 1. Our method can outperform both hand-crafted MTL models and automatically searched models. There are no other clear alternatives that we could compare to. If there are any other methods that the reviewer thinks can fit into our problem, we would be happy to include more baselines. 4. Privacy concern: We do not specifically address this issue in this paper. The data we use has already been anonymized. We assume that the input data has already been preprocessed and does not have privacy issues. Also, the workflow of our framework does not require the patient demographic features (only rely on the time series signals), which further reduces the risk of privacy issues. We will add a note about the privacy issues and the need to make sure the data is anonymized and no quasi-identifiers are present, etc. before our system is used. 5. Limited Evaluation: MIMIC-IV is the most widely used benchmark in the EHR domain. So we choose this dataset for evaluating our method. At least in the EHR domain, MIMIC-IV is enough to show the effectiveness of our method. Other public EHR datasets normally have lower quality than MIMIC-IV. We will try to include more datasets from other domains & scenarios in future work for investigating the extensibility of AutoDP. 6. Interpretability: We include two case studies at Appendix E for showing the searched configurations under the settings of Task @ 10 and Task @ 25, which can provide some insights of what kinds of task groupings and architectures are effective to MTL on EHR data. Also the searched configuration might provide some guidance to the MTL framework design in similar problems. 7. Other concerns: Data imbalance issues often happen across different datasets. In our setting, we are predicting multiple diseases for the same patient, so all diseases are well annotated. Also, we conducted an ablation study (disease based grouping) to show that solely training on similar diseases might not bring the best performance gain. So it is necessary to use a search algorithm to discover the optimal task grouping. We will add to the camera-ready paper that our dataset was imbalanced and discuss a bit more about similar diseases, etc. --- Rebuttal Comment 1.1: Comment: Thanks for your rebuttal. After carefully reading your comments, i still want to maintain my original rating. Using only one public available EHR data is limited for the generalization purpose. --- Rebuttal 2: Comment: Dear Reviewer Zp9P, I am a NeurIPS 2024 Area Chair of the paper that you reviewed. This is a reminder that authors left rebuttals for your review. We need your follow up answers on that. Please leave comment for any un-answered questions you had, or how you think about the author's rebuttal. The author-reviewer discussion is closed on Aug 13 11:59pm AoE. Best regards, AC	null	NeurIPS_2024_submissions_huggingface	2,024	Summary: The research work presents an exploration into the potential of multi-task learning (MTL) within the context of electronic health record (EHR) data analysis and clinical prediction tasks. This study innovatively addresses the critical challenges of task grouping and model architecture design, which are essential for optimizing MTL frameworks. Strengths: S1. By proposing an automated approach, AutoDP, the paper contributes to reducing human intervention in the configuration process, thereby streamlining the identification of optimal task combinations and neural architectures tailored for EHR data. S2. The employment of surrogate model-based optimization and a progressive sampling strategy demonstrates a novel and efficient methodology for navigating the vast search space, leading to improved performance with a feasible computational cost. S3. The experimental results on the real-world EHR dataset, MIMIC-IV, validate the efficacy of the proposed AutoDP framework, showcasing substantial performance improvements over both hand-crafted and existing automated methods. S4. This work not only advances the state-of-the-art in automated machine learning for healthcare but also provides valuable insights and tools for the broader research community working on multi-task learning problems. Weaknesses: Q1. Figure Clarity and Comprehensibility: The overview of the proposed AutoDP in Figure 1 lacks self-evidency, making it challenging for readers, particularly those who are not specialists in electronic health record (EHR) modeling and analysis, to grasp the main insights at a glance. The figure is dense with symbols and elements that are not immediately interpretable without additional context. It is recommended that the authors enhance the figure's presentation to improve readability and comprehension. This could involve breaking down complex elements, using clearer labeling, and providing a more detailed legend or accompanying text that guides the reader through the visualization. Q2. Evaluation Metric Consideration: For many clinical prediction tasks on datasets like MIMIC, where label balance may not be achievable, the authors might consider the use of AUPRC (Area Under the Precision-Recall Curve) as an alternative or complementary evaluation metric to AUC-ROC (Area Under the Receiver Operating Characteristic Curve). AUPRC can be a more informative measure when dealing with imbalanced classes, as it evaluates the model's ability to rank samples correctly, independent of the classification threshold. Precision-recall metrics are less sensitive to the choice of threshold compared to accuracy or precision alone, which is a critical consideration in clinical settings where the cost of false positives and negatives can vary significantly. Q3. Generalizability and Extensibility: The current experiments are limited to the MIMIC dataset, and the baseline models have been fine-tuned for this specific dataset in the level of model-design. While the results on MIMIC are promising, readers may be interested in the general applicability of the proposed method to other datasets and clinical scenarios. It would be beneficial for the authors to provide experiments on additional datasets to demonstrate the robustness and generalizability of their approach. Furthermore, an exploration of how the method performs when integrated with different models or when adapted to various clinical prediction tasks would strengthen the paper's contribution and credibility. Given the focus on EHR data, it would be insightful if the authors could discuss the transferability of their model to other EHR datasets, potentially from different geographical regions or healthcare systems. This discussion could include challenges related to data heterogeneity, variations in clinical practices, and how these factors might affect the performance of the proposed AutoDP framework. Technical Quality: 3 Clarity: 2 Questions for Authors: - Confidence: 2 Soundness: 3 Presentation: 2 Contribution: 3 Limitations: - Flag For Ethics Review: ['No ethics review needed.'] Rating: 5 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Q1: Thank you for pointing this out. We will further improve Figure 1 in the camera-ready version for better clarity and comprehensibility. Q2: We have included Averaged Precision for considering the class imbalance. Averaged Precision and AUPRC are essentially the same thing in this setting. Please see the definition of Averaged Precision via the link below. https://en.wikipedia.org/w/index.php?title=Information_retrieval&oldid=793358396#Average_precision Q3: It would be interesting to evaluate our method to different settings such as more datasets, more clinical tasks & scenarios, transferability, and etc. However, we will leave these to future work as they diverge from the major claim of this paper and it is difficult to fit in due to lack of any further space to explain everything. We focus on the automation of the MTL framework design for joint disease prediction for now. --- Rebuttal 2: Comment: Dear Reviewer S7e4, I am a NeurIPS 2024 Area Chair of the paper that you reviewed. This is a reminder that authors left rebuttals for your review. We need your follow up answers on that. Please leave comment for any un-answered questions you had, or how you think about the author's rebuttal. The author-reviewer discussion is closed on Aug 13 11:59pm AoE. Best regards, AC	null	null	null	null	null	null
VISA: Variational Inference with Sequential Sample-Average Approximations	Accept (poster)	Summary: This paper presents a method for approximate inference in expensive to evaluate models where the gradients may not be available. The presented approach proceeds in a trust-region-optimization fashion where a series of deterministic surrogates objectives are optimized. Each objective is optimized until a trust region condition is met, indicating that the objective needs to refreshed. The proposed approach shows promising results when compared to other competing approaches. Strengths: The paper presents a simple procedure that is clearly laid out. It is well written for most parts and seems to be promising direction for future research. Weaknesses: - I think the use of ESS for determining the trust region is rather arbitrary. I did not find a sound explanation for why this should be the goto way for determining the trust region. - Also, the analysis around the choice of alpha seems rather experimental. The performance seems extremely critical to the choice of this parameter and there doesn't seem to be any good way to select this parameter. - Moreover, rerunning the optimization to select this parameter can result in a lot of evaluations. Since we wanted to avoid this in the first place, would a user not be better of using the IWFVI approach? - Also, what exactly is the difference between the IWFVI approach and the standard Re-weighted wake sleep? Why rename an already named algorithm? - I did not find enough details about the exact optimization procedure used to optimize the VISA approach. Which optimizer was used? - I am also not sure how are the number of evaluations calculated for the Figure 2. Are we only counting the evaluations during the refresh step? If so, should we not see step decays in the metrics for VISA approaches? Technical Quality: 3 Clarity: 3 Questions for Authors: Please, see the Weaknesses section. Confidence: 3 Soundness: 3 Presentation: 3 Contribution: 2 Limitations: The scalability of the proposed approach is rather limited as pointed out by the authors. The proposed method seems extremely sensitive to the choice of learning rate and the alpha. It is unclear how to chose them properly. Flag For Ethics Review: ['No ethics review needed.'] Rating: 5 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Motivation for ESS as trust-region criteria. See general response. Choice of ESS threshold. See general response. IWFVI vs Reweighted Wake-Sleep (RWS). RWS is used in the context of amortized variational inference and optimizes both the parameters of the model and the parameters of the (amortized) variational approximation. Moreover standard RWS optimizes different objectives to update the variational approximation for the sleep- and wake-phases of the algorithm, which use simulated and real data for training respectively. IWFVI only optimize the variational distribution and does not use simulated data from the model (sleep-phase in RWS). Hence, IWFVI updates correspond to the wake-phase updates to the variational parameters in RWS only; both are based on optimizing an importance-sampling based estimate of the forward KL-divergence. Details on exact optimization procedure. We use Adam with standard parameters (as implemented in optax) for all experiments. We will clarify this in the final version of the manuscript. Number of evaluations. It is correct that only freshly drawn samples are counted towards the number of evaluations, hence the number of evaluations only increases after a “refresh step”. The reason why there is no visible staircase pattern in the corresponding plots is that VISA refreshes after a relatively small amount of evaluations (for conservatively chosen ess thresholds) and that the plots are averaged over multiple runs. For a single VISA run with a low enough ess threshold we would indeed see a visible staircase pattern. --- Rebuttal 2: Title: A few more clarifications. Comment: Thank you for the response. I had a few more clarifying questions. 1. Can you provide numbers on how many times the samples are refreshed? At what stages are samples refreshed more? Is it during the start of the optimization or towards the end? Also, can you give a version of the plots with number of optimization iterations on the x-axis? I understand the main motivation is to reduce the number of model evaluations; however, I still think answer to these might improve my understanding of the method. 2. Is N the same across the experiments for both VISA and IWFVI? How is the N value selected? I wonder what are the trade-offs when selecting the N. It is likely that for a given setting of N, IWFVI suffers because it uses more samples at every update iteration. 3. From what I understand, the combined choice of $\alpha$ and learning-rate is tricky. Things work fine in the simple Gaussian examples; however, on real problems, it sounds like as a practitioner I will have to play around with different settings and see what works. This makes me squeamish--the ultimate aim of the setting of this paper is to reduce the number of model evaluations. As such, the method seems a bit understudied to me. Can the authors offer some comments on this thought process? Thank you. --- Rebuttal Comment 2.1: Title: Re: A few more clarifications. Comment: Thank you for your questions and your interest in our methodology! We hope our response provides further clarification on the points you have raised. Number of model evaluations. We have prepared the requested plots illustrating inference performance over the number of optimization iterations, as well as additional plots showing the relationship between optimization iterations and model evaluations. However, after reviewing the guidelines for the discussion process, we realized that we are not permitted to include anonymous links to external sources. Consequently, while we will include the plots in the final manuscript, we must address the questions in this response without the aid of visual representations. In the plots in our experiment section, the number of times VISA refreshes samples corresponds directly to the number of evaluations. For instance, in the experiments with the full-covariance Gaussian, VISA with $\alpha \leq 0.9$ converges after approximately 1,000 model evaluations when using a learning rate of 1e-3. However, at this point, VISA has already performed around 30,000 gradient steps (optimization iterations). In contrast, for the same scenario, IWFKL also converges after about 30,000 gradient steps but requires 30,000 model evaluations, as it draws fresh samples at each iteration. Our results also show, across experiments, that VISA resamples at relatively constant intervals and uses slightly less model evaluations early in training. This makes sense, considering that the posterior approximation, and consequently the proposal distribution, is initialized relatively broad and the ESS criteria is not very sensitive to perturbations of the variational approximation within the coverage of the proposal. As training progresses, the variational approximation, and eventually the proposal, become more tightly peaked, causing the ESS criteria to become more sensitive to perturbations of the variational distribution, which might move it outside the coverage of the proposal. As a result VISA refreshes samples more frequently after the early phases of training. Number of samples per model evaluation. In our experiments, both VISA and IWFVI use the same number of samples per (batch) model evaluation, denoted $N$, which we heuristically selected to be sufficiently high to ensure stable convergence for IWFVI. In our experience, this approach works well as long as the ESS threshold is chosen conservatively. If $N$ were significantly increased, we would indeed expect to observe a greater difference in the overall number of model evaluations between IWFVI and VISA. Moreover, with a larger sample size, it would likely be possible to select a lower ESS threshold for VISA without compromising training stability, potentially leading to faster convergence. Conversely, if we chose $N$ to be the (unknown) minimal number of samples required for IWFVI to achieve stable convergence, then VISA, for $\alpha < 1$, would potentially suffer from instabilities and might not be able to reduce the number of model evaluations substantially. However, in practice, the minimal number of samples required is difficult to identify and might vary for different phases of training. In practical scenarios we expect VISA to be able to reduce the amount of model evaluations. Tuning hyperparameters. We acknowledge that introducing an additional hyperparameter can make practitioners hesitant, and often for good reason. However, it's important to note that IWFVI is a special case of VISA, for $\alpha=1$. Therefore, it is reasonable to base the hyperparameter selection for VISA on those used for IWFVI, as long as we choose a conservative ESS threshold, for which we expect VISA to behave similar to IWFVI. In our experiments, we applied this method for both the initial selection of the learning rate and the number of samples per model evaluation. While practitioners may choose to invest computational resources in optimizing $\alpha$ further, in many cases, VISA already outperforms IWFVI with the same hyperparameters and a conservative ESS threshold, e.g. $\alpha=0.99$. We believe VISA is particularly valuable for practitioners who currently use IWFVI for inference on models that are costly to evaluate. In such scenarios, VISA can serve as a drop-in replacement that requires minimal tuning. However, even when hyperparameters cannot be directly "bootstrapped" from previous experiments, we argue that finding appropriate hyperparameters for VISA is no more challenging than it is for IWFVI.	Summary: This paper introduces a new method for variational inference called VISA, which stands for Variational Inference Using Sequential Sample-Average Approximations. VISA is based on importance-weighted forward-KL variational inference and allows for reusing model evaluations across multiple gradient steps. This makes it suitable for non-differentiable and computationally intensive models. The authors conducted experiments using simulated and real data to validate VISA, showing that it can achieve similar approximation accuracy to standard IWFVI while requiring significantly fewer samples. Strengths: 1. The paper is well-written and organized, making complex concepts easy to understand through a logical flow of information. 2. VISA effectively addresses the challenge of minimizing the number of model evaluations, thereby achieving significant computational efficiency. 3. The experiments conducted provide robust evidence that VISA converges faster, requiring fewer model evaluations compared to existing methods. Weaknesses: 1. While the concept of trust regions based on Effective Sample Size (ESS) is intuitive, the paper would benefit from a more formal theoretical justification for the specific choice of trust region. 2. The paper currently lacks a discussion on the robustness of the proposed method to various optimization methods. While it touches on the use of L-BGFS, including an analysis of how VISA performs with other optimization techniques such as classical stochastic gradient descent (SGD) and adaptive methods like RMSProp and Adam would enhance the comprehensiveness of the study. Technical Quality: 3 Clarity: 3 Questions for Authors: 1. Could the algorithm's performance be enhanced by incorporating adaptive adjustments for the ESS threshold $\alpha$ and the learning rate? 2. It appears there may be a typo in the caption for Figure 3 when referring to the methods VISA and IWFVI. Could you please verify and correct this if necessary to ensure clarity and accuracy in the presentation of your results? Confidence: 3 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: The authors have adequately discussed the limitations of their approach. Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Motivation for ESS as trust-region criteria. See general response. Robustness to optimization methods. We indeed use Adam (optax implementation) for all experiments and will clarify this in the final version of the manuscript. Based on your suggestion, we conducted additional experiments for the Gaussian target densities to demonstrate that VISA can be used with SGD, Adam, and RMSProp for appropriately selected stepsizes. Our evaluations shows that plain SGD (without momentum or Nesterov acceleration) might run into local optima more easily as Adam and RMSProp for small step sizes, while RMSProp might not be as robust to larger step sizes as Adam. Overall, the experiment suggests that Adam is a good standard choice for VISA. We have included the corresponding convergence plots in the PDF file attached to the general rebuttal and will add a polished version to the final manuscript. Adaptive ESS threshold. While we did not test adaptive strategie, we experimented with using a schedule for the ess threshold. However, we did not find consistent improvement over using a fixed threshold and hence opted for the simplest possible presentation of the algorithm. That being said, we expect that problem specific scheduling and adaptation strategies are able to outperform fixed ess threshold. However, determining a performant schedule or adaption hyperparameters might require running inference multiple times, which contributes to the overall number of evaluations and complicates a fair comparison. Typo in Figure 3. There was indeed a typo, thanks for pointing it out! The caption should read: "For smaller step sizes (0.001, 0.005) VISA achieves comparable forward KL-divergence to IWFVI while requiring significantly less model evaluations to converge (see vertical lines). For larger step sizes (0.01) VISA only converges with a high ess threshold (0.99) and requires more evaluations than IWFVI and VISA with a smaller step size (0.005)." --- Rebuttal Comment 1.1: Comment: Thank you for your response, which clarified my concerns. I will maintain my current score.	Summary: The paper proposes to use a sample-average approximation for variational inference. In contrast to previous works, the method uses the forward-KL which does not require differentiability of the joint likelihood. In order to sample from an approximate posterior instead of the exact posterior, a sequential trust-region approach is introduced. The method is evaluated on small-scale problems and some benefits of the forward-KL SAA approach compared to BBVI or reverse-KL SAA are demonstrated. Strengths: - The derivatinos and mathematical formulation are clear, the paper is easy to read. - The proposed trust-region approach is rigorously evaluated on small-scale examples with various ablation studies. Weaknesses: - The proposed VISA method (and other baselines as well) seem quite sensitive to the learning rate and other tuning parameters. Therefore it is hard to judge whether the proposed approach has significant benefits over other methods, or whether the benefits are due to a tuning of the learning rate. - The relevance of the considered applications to the wider machine-learning audience seems unclear. The work is motivated by non-differentiability but then only toy examples are shown. Some real-world applications (for example, reinforcement learning) would make the paper stronger. Technical Quality: 3 Clarity: 3 Questions for Authors: - When compared to other SAA methods (Giordano et al 2024), (Burroni et al 2023), the novelty of the proposed method seems to be in the use of forward KL rather than reverse KL. I can see the computational advantages of using forward-KL, but does the mode-covering property of forward-KL pose problems sometimes? For example, when the approximate posterior is simple but the true posterior is multimodal with high loss inbetween? - Natural-gradient VI methods can also be used in black-box settings -- see for instance "Variational Adaptive-Newton Method for Explorative Learning", Khan et al., 2017. Related natural evolution strategies (such as CMA-ES) have shown state-of-the-art performance in optimizing loss functions with only gradient evaluations. Have you considered comparing to these methods, or would you expect them to perform similary as BBVI-SF? Confidence: 3 Soundness: 3 Presentation: 3 Contribution: 2 Limitations: Limitations are adequately discussed in the last section of the paper. Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Sensitivity to the learning rate. See general response. No real-world experiments. See general response. Failure cases of the Forward KL-divergence. While we consider the mode-covering behavior of the forward KL-divergence as a positive, there can be instances where a mode of the variational approximation covers multiple modes of the target density and as a result assigns not insignificant mass to a low density area. To prevent these scenarios it is important to choose an appropriate variational approximation that allows to model the expected degree of multimodality. Moreover, in practice, optimizing an importance-weighted approximation to the forward KL-divergence often results in an under approximation of posterior variance similar to optimizing a reverse KL-divergence, especially when targeting well separated modes in the small sample regime. Adding more specialized baselines. We chose to compare VISA to IWFV and BBVI-SF because these methods are equally back-box, in the sense that they do not assume differentiability of the model and are not restrictive in the variational families that can be used. However, in settings where more specialized methods can be applied, we expect these methods to outperform our baselines and potentially VISA. One instance of this can be seen in the first experiment where we included a BBVI-RP baseline, which is generally considered a black-box method but leverages the differentiability of the model. The suggested baseline, VAN and CMA-ES, specifically use a multivariate Gaussian as a variational approximation and VAN-D makes use of a reparameterization trick to approximate the diagonal of the Hessian, which requires a differentiable model. That being said, we agree that it would be insightful to compare to and potentially combine VISA with a black-box (in the sense described above) stochastic second-order method. We will work to include an appropriate method in the final manuscript. --- Rebuttal Comment 1.1: Title: Thanks for the clarifications! Comment: I have carefully read the rebuttal, and some of my concerns were addressed. Other issues still stand (e.g., problems with forward KL), but I have concluded that this may just be an inherent limitation of the method which is not fixable. I will increase my score accordingly. Regarding the real-word experiments, I am still not fully convinced that the method will be that useful in practice and unsure how it compares to other baselines (e.g., natural-gradient methods). But as the method is fully "black-box", releasing the algorithm as a software package to the community an interesting real world use case for it may emerge.	Summary: The authors propose to reduce the number of model evaluations during the optimization of the variational lower bound. To this end, they integrate sample avarage approximations (SAAs) into IWFVI framework, which allows updates of variational parameters while keeping approximate samples generated from a variational distribution obtained a few steps before during optimization. The method was evaluated on a few synthetic experiments. Strengths: 1. The method is effective in the experiments considered (it requires less model evaluations to converge). 2. The clarity of the paper is good and is easy to follow. Weaknesses: 1. The method seems to be a combination of a few engineering tricks. It is unclear what the effect of inconsistency between approximate samples ($\tilde{\phi}$) and variational distribution ($\phi$) during optimization is. 2. Only synthetic experiments are considered. 3. In the paper, the model is assumed to have no tubable parameters, while many applications requires tunable model as well (e.g. VAE). Technical Quality: 3 Clarity: 2 Questions for Authors: See weakness. Confidence: 3 Soundness: 3 Presentation: 2 Contribution: 2 Limitations: The authors thinks there is no social impact. However, no justifications are given. Flag For Ethics Review: ['No ethics review needed.'] Rating: 4 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Clarification regarding our methodology. VISA does not optimize the reverse KL-divergence or corresponding variational lower bound but a forward KL-divergence or corresponding variational upper bound. This is a crucial difference to previous related work, which studies SAAs in the context of the reparameterized reverse KL-divergence. Presentation. Given the clarity of our paper was highlighted as a strength, we are slightly confused by the low score for the presentation. Please let us know in which ways we can further improve our presentation. Inconsistency between proposal and variational distribution. The discrepancy between the proposal and variational distribution is captured by the importance weight $q(z \mid \phi) / q(z \mid \tilde \phi)$, which contributes to the overall importance weight used in the sample average approximation and corrects for the error introduced by not sampling directly from $q(z \mid \phi)$. In general, the bias and variance of the gradient estimates will be higher if the proposal has not been updated recently and the discrepancy to the variational distribution is large. To quantify the effect that this discrepancy has on the overall performance of the algorithm numerically we conducted experiments with various ESS thresholds $\alpha$, which are a way to formalize an upper bound on the maximum allowed discrepancy between the proposal and variational distribution. Only synthetic experiments. See general response. No tunable model parameters. VAEs are trained by jointly optimizing a generative model and a variational posterior approximation by maximizing an ELBO in the context of amortized variational inference. In contrast, our work considers posterior inference for a given model, a setting that is frequently encountered when working with simulation-based models, and treats parameter estimation of that model as an orthogonal problem. --- Rebuttal Comment 1.1: Title: Request for clarification on remaining concerns Comment: Dear Reviewer, We noticed that you provided a very low score, and we would like to know if our rebuttal has sufficiently addressed your concerns. If any issues remain unresolved, we would greatly appreciate it if you could share them with us at your earliest convenience, so we can address them in time. Thank you for your time and consideration! --- Rebuttal Comment 1.2: Comment: Thanks for your rebuttal, which clarifies part of my concerns (regarding inconsistency between proposal & variational distribution; no tunable model parameters), and I am willing to raise my score to 4 to reflect that. However, the lack of real-world experiments makes me hesitate to recommend this work for acceptance.	Rebuttal 1: Rebuttal: We would like to thank the reviewers for their detailed reviews! We are delighted to see the clarity of our manuscript and rigor of our experiments listed among the strengths of our work. We address the point raised by multiple reviewers below and respond to the individual concerns and questions in dedicated rebuttals under the corresponding reviews. Sensitivity to learning rate and choice of ESS threshold (kZyU, sEay). We believe that one of the strengths of VISA is its robustness to the choices of learning rate for conservatively chosen thresholds $\alpha$. As seen in the first two experiments, the rate of convergence of VISA is relatively stable across all but the largest presented learning rates, which are chosen too large on purpose. In contrast, the convergence rates of IWFVI and BBVI are more sensitive to the learning rate and only competitive when chosen close to the upper limit in terms of training stability. In general, while finding the optimal training hyperparameters for either method might require careful tuning, we found that a simple recipe for selecting the learning rate and ess-threshold for VISA is often enough to outperform IWFVI: 1) Choose the same learning rate as for IWFVI and a conservative ess-threshold, e.g. $\alpha=0.99$. 2) If convergence is stable, one might choose to reduce $\alpha$ to achieve even faster convergence, otherwise decrease the learning rate slightly. Motivation for ESS as a trust-region criteria (Bkjs, sEay). There is in fact a good reason to use the ESS to define the trust region, and we will clarify this in the manuscript. The Kish effective sample size is directly related to the variance of the importance ratio between the current variational distribution and the trust region distribution, which can in turn be expressed as a Chi-Square divergence between these distributions. $ \chi^2(q(\cdot; \phi) \mid q(\cdot; \tilde{\phi})) = \mathrm{Var}_{q(\cdot; \tilde{\phi})} \left[\frac{q(z; \phi)}{q(z; \tilde{\phi})}\right] \approx \frac{n}{\mathrm{ess}} - 1 $ As a result, the ESS-based trust region can be motivated from (1) a divergence perspective. i.e. we want the sampling distribution close to the variational distribution as measured by the chi-square divergence, or from (2) the perspective of the variance of the corresponding importance weight (or ESS as a proxy measure), which we want to be low (or high ESS respectively). The second perspective also motivates the use of the ESS to assess the quality of importance samplers or adaptive resampling strategies based on an ESS criteria. We will expand upon the motivation of the effective sample size as a trust-region criteria in the manuscript and will add deviations for the above equalities to the appendix. Only synthetic experiments (snbt, kZyU). We see the contribution of this work in proposing VISA and studying its behavior by exploring a wide spectrum of hyperparameters. While this setting makes it somewhat challenging to evaluate VISA on truly expensive real-world models, we want to highlight that the Lotka-Volterra model and Pickover attractor capture the properties of many real-world scientific simulation models. The Pickover attractor model in particular can exhibit extremely chaotic behavior, is not fully differentiable, and is fairly expensive to evaluate. To be able to track the attractor state we run SMC with 500 particles for 100 time steps for each of the 100 samples that make up a single batch evaluation of the model. VISA runs with different optimizers (Bkjs). Please find convergence plots in the attached PDF file. Pdf: /pdf/5798fcdba1444d3a2bba4c631b4f94278a9bcdf8.pdf	NeurIPS_2024_submissions_huggingface	2,024	null	null	null	null	null	null	null	null
Cascade Speculative Drafting for Even Faster LLM Inference	Accept (poster)	Summary: This paper concentrates on the inference efficiency of LLMs and thinks that the autoregressive generation contained in drafting process of speculative decoding leads to the suboptimal performance of speculative decoding. It introduces a novel speculative execution algorithm, Cascade Speculative Drafting (CS Drafting), which containes Vertical Cascade and Horizontal Cascade to achieve better speedup, while preserving the same output distribution as the target model. Strengths: 1. The proposed model introduces a new framework that fully considers the acceleration at different granularities and varying capabilities of the model to complete decoding. 2. The method proposed in this paper is model-agnostic and does not require additional training of the model. 3. Experimental results indicate that compared to the vanilla model, the proposed model can achieve better speedups. Weaknesses: 1. The model employs a heuristic approach, involving a large number of hyperparameters, especially the $k$-matrix, whose quantity is related to the involved drift models. 2. The experimental results are not sufficient. On one hand, it does not compare with the latest speculative decoding methods. On the other hand, the model does not explore performance with different number of candidate tokens. It also lacks ablation studies and does not individually investigate the different impacts of the two cascade methods. 3. There are writing issues in the article, such as the first line of Algorithm 1, line 178, and line 292. Additionally, many variables in Algorithm 1 are not explained. Technical Quality: 3 Clarity: 2 Questions for Authors: 1. The description of the core algorithm, Algorithm 1, in this paper is not clear. It would be helpful to include an explanation of the overall logic of the algorithm and add descriptions of some variables. 2. The paper lacks relevant ablation experiments to illustrate the relationship between the two cascade methods and lacks analytical experiments to demonstrate the robustness of the method with different numbers of candidate tokens. Confidence: 4 Soundness: 3 Presentation: 2 Contribution: 3 Limitations: The authors adequately addressed the limitations and, if applicable, potential negative societal impact of their work. Flag For Ethics Review: ['No ethics review needed.'] Rating: 5 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We thank the reviewer for highlighting the advantages of our method. We appreciate the insightful and detailed feedback, and would like to address each of your concerns. > W1: The model employs a heuristic approach, involving a large number of hyperparameters, especially the K-matrix, whose quantity is related to the involved drift models. We understand your concern about the number of hyperparameters. However, having additional hyperparameters is an intrinsic problem of algorithms utilizing speculative execution as hyperparameters are usually necessary to be adjusted based on the probability of acceptance. As a remedy to the additional hyperparameters, during our experiments with Vicuna-7B, we used the same hyperparameter across different datasets for each model. This represents the performance of CS Drafting without meticulous hyperparameter tuning. We also include the experiment results with different $K_{00}$ (the entry limiting the generation length of the largest draft model. This is the hyperparameter that has the largest effect on our experiments) on GSM8K using Vicuna-7B: \| $K_{00}$ \| Walltime (tokens/s) \| \| -------- \| ------- \| \| 1 \| 56.16 \| \| 2 \| 55.51 \| \| 3 \| 53.73 \| As shown in our experimental results, the performance of CS Drafting only decreases slightly when the hyperparameter is sub-optimal. Therefore, the end user who does not require maximum performance can use a simple setup to achieve near-optimal performance with CS Drafting. > W2: The experimental results are not sufficient. On one hand, it does not compare with the latest speculative decoding methods. On the other hand, the model does not explore performance with different number of candidate tokens. It also lacks ablation studies and does not individually investigate the different impacts of the two cascade methods. Thank you for your comment! We have included extensive results for different settings in Table 2 and have conducted a different set of experiments shown in Table 3. Following your suggestion, we include additional experimental results with different $ K_{00} $ (refer to response to W1 above). As an ablation study, we removed the horizontal cascade from CS Drafting. This results in a performance decrease from 56.16 to 53.55 when $ K_{00} = 1$. This demonstrates the effectiveness of the horizontal cascade. For the vertical cascade, the reviewer may refer to our results in Table 2. The performance usually decreases when we remove a draft model, indicating the effectiveness of the vertical cascade. Additionally, our theoretical analysis in Section 4 also supports the improvement of both cascades when isolating one of them. Regarding the comparison with other speculative decoding methods, we have compared our results with both Medusa and speculative decoding. While we are aware of newer methods such as Eagle or Hydra, we would like to remind you that our experiments were performed before their introduction and our preprints were available months earlier than theirs. > W3: There are writing issues in the article, such as the first line of Algorithm 1, line 178, and line 292. Additionally, many variables in Algorithm 1 are not explained. Thank you for the detailed feedback on writing. We will do a more careful proofread to fix typos and add explanations of the variables. > Q1: The description of the core algorithm, Algorithm 1, in this paper is not clear. It would be helpful to include an explanation of the overall logic of the algorithm and add descriptions of some variables. We would like to adopt the suggestion to improve the clarity. The following is the overview of our algorithm. The heart of the CS Drafting algorithm involves using smaller models as drafters for larger draft models (vertical cascade) as well as allocating smaller draft models to continue drafting less important tokens after larger draft models (horizontal cascade). Algorithm 1 implements these two cascades by using a for loop and recursion. The for loop is responsible for the horizontal cascade by gradually limiting the usable draft models to smaller models. The recursion calls smaller draft models to perform drafting for the target model or large draft models. We will also ensure that all variables in the algorithm are properly defined and are easy to understand to the readers. > Q2: The paper lacks relevant ablation experiments to illustrate the relationship between the two cascade methods and lacks analytical experiments to demonstrate the robustness of the method with different numbers of candidate tokens. Thank you for your suggestion! Please refer to our response to W1 and W2. Thank you again for reviewing our paper! We hope our response addresses your questions. Please let us know your thoughts, and we are more than happy to answer any further questions. --- Rebuttal Comment 1.1: Title: Response to Authors Comment: Thank you very much for your response. After reading your reply, I am very pleased with the improvements in the score. I hope you can incorporate our suggestions into the final version. --- Reply to Comment 1.1.1: Comment: Thank you for your feedback! We are glad that our response addressed your concerns. We will incorporate your suggestions into the final version.	Summary: The paper proposes to use multiple draft model for speculative decoding. Specifically, the smallest model could be the statistic language model which has negligible latency therefore reducing the cost of autoregressive regression. Experiments show the proposed method works better than baselines. Strengths: (1) The idea is very interesting and supported by analysis such as Figure 2. (2) Experiments show the proposed method works well. Weaknesses: (1) Algorithm 1 is heavy and should be simplified for the reader. (2) The whole pipeline seems to introduce lots of engineering work. For instance, we need to set hyperparameters for how to choose model size for each token position. Lenience is also used which is good but increases the complexity of the pipeline. Technical Quality: 3 Clarity: 3 Questions for Authors: N/A Confidence: 3 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: N/A Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We thank the reviewer for highlighting the idea, analysis, and experiments of our paper. We appreciate the feedback provided and would like to address the weaknesses mentioned. > Q1: Algorithm 1 is heavy and should be simplified for the reader. While we made numerous attempts to simplify the algorithm, we always found that simplified versions missed essential details needed by end users. To better help readers understand our algorithm, we would like to add the following summary as a caption to the algorithm: The heart of the CS Drafting algorithm involves using smaller models as drafters for larger draft models (vertical cascade) as well as allocating smaller draft models to continue drafting less important tokens after larger draft models (horizontal cascade). Algorithm 1 implements these two cascades by using a for loop and recursion. The for loop is responsible for the horizontal cascade by gradually limiting the usable draft models to smaller models. The recursion calls smaller draft models to perform drafting for the target model or large draft models. We will also add more explanations of the variables to make them easier for the readers to understand. > Q2: The whole pipeline seems to introduce lots of engineering work. For instance, we need to set hyperparameters for how to choose model size for each token position. Lenience is also used which is good but increases the complexity of the pipeline. We understand your concern about the number of hyperparameters. However, having additional hyperparameters is an intrinsic problem of algorithms utilizing speculative execution as hyperparameters are usually necessary to be adjusted based on the probability of acceptance. As a remedy to the additional hyperparameters, during our experiments with Vicuna-7B, we used the same hyperparameter across different datasets for each model. This represents the performance of CS Drafting without meticulous hyperparameter tuning. We also include the experiment results with different $K_{00}$ (the entry limiting the generation length of the largest draft model. This is the hyperparameter that has the largest effect on our experiments) on GSM8K using Vicuna-7B: \| $K_{00}$ \| Walltime (tokens/s) \| \| -------- \| ------- \| \| 1 \| 56.16 \| \| 2 \| 55.51 \| \| 3 \| 53.73 \| As shown in our experimental results, the performance of CS Drafting only decreases slightly when the hyperparameter is sub-optimal. Therefore, the end user who does not require maximum performance can use a simple setup to achieve near-optimal performance with CS Drafting. --- Rebuttal Comment 1.1: Comment: Dear Reviewer FGcM, thank you again for reviewing our paper! We hope our response addresses your questions. Please let us know your thoughts, and we are more than happy to answer any further questions.	Summary: This study introduces a novel method to accelerate large language model (LLM) decoding by integrating speculative decoding with two types of model cascades: vertical and horizontal. The horizontal cascade utilizes larger draft models for generating initial tokens, while smaller models assist in producing subsequent tokens. The vertical cascade implements a series of verification steps using a stack of cascaded models of varying sizes. Through the integration of these approaches, the authors report notable improvements in decoding speed across benchmarks. Strengths: Originality This paper demonstrates originality by expanding on speculative decoding. It introduces two novel techniques - horizontal and vertical cascades - that effectively factorize the draft-and-verification steps of speculative decoding across multiple models. Quality and Clarity The authors base their approach in intuitive assumptions, like the complexity of first token generation. The speed improvements over vanilla speculative decoding illustrate the effectiveness of the proposed methods. Significance The paper makes a considerable contribution to the community. Although CS drafting alone underperforms compared to Medusa (the multiple decoding heads method), the authors successfully combined their technique with Medusa, achieving superior performance. This suggests that their technique could be useful when integrated with other decoding methods. Weaknesses: Algorithmic Complexity: The paper discusses the use of horizontal and vertical cascades, but each additional cascade increases the algorithmic complexity of the decoding process. This complexity can become particularly challenging with larger models. The paper does not address how to balance this increased complexity with the potential speedup gains. Important questions such as the optimal number of cascades and the comparative usefulness of horizontal versus vertical cascades remain unanswered. Ablation Study: The max-gram drafting technique appears to be quite efficient for generating many tokens without needing to invoke other medium-sized models. However, the paper lacks an ablation study to analyze the impact of different model combinations on decoding performance. Including such a study would provide valuable insights into the efficiency and effectiveness of the proposed method. Memory and Compute Constraints The use of multiple models in complex scenarios poses challenges related to memory and computational limits. With limited memory and bandwidth, adding more cascades might not lead to faster end-to-end latency, especially as the target models grow in size. The paper does not consider this factor or provide a discussion on its implications. Technical Quality: 3 Clarity: 2 Questions for Authors: Please see the "Weaknesses" section. Confidence: 3 Soundness: 3 Presentation: 2 Contribution: 3 Limitations: Ok. Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We thank the reviewer for recognizing the significance of our contribution. We appreciate the insightful feedback provided and would like to address each of your questions. > Q1: Algorithmic Complexity To make an additional smaller draft model meaningful in the CS Drafting algorithm, it needs to be much smaller than the smallest used draft model, typically by a factor of at least 10. This represents an exponential relationship between the number of draft models and the size of the target model, so the total number of draft models should ideally be capped at 6, even for trillion-parameter models. Unfortunately, we do not have sufficient GPU capacity to conduct inference with models large enough to establish the relationship between the number of parameters and the ideal number of draft models. While it’s unclear how much additional performance gain can be achieved by further increasing the number of draft models, we expect a similar performance boost by our algorithm on larger models while keeping the same level of system complexity. > Q2: Ablation Study We have provided a thorough analysis on different combinations in Table 2. We are also happy to share the performance of various combinations of draft models for Vicuna-7B on GSM8K below: \| Draft model \| Walltime (Tokens/s) \| \| --- \| --- \| \| MaG \| 52.89 \| \| Vicuna-68m \| 44.31 \| \| Vicuna-68m & MaG \| 56.72 \| We observe the effectiveness of MaG as it exceeds the performance of Vicuna-68m due to its nearly zero latency. The optimal performance in our experiment is achieved by using both MaG and Vicuna-68m as draft models. > Q3: Memory and Compute Constraints We conducted an experiment to monitor memory usage across different systems accelerating Vicuna-7B on GSM8K: \| Method \| GPU Memory(MiB) \| \| -------- \| ------- \| \| Huggingface Generation \| 14,016 \| \| CS Drafting \| 16,154 \| \| Speculative Decoding \| 16,118 \| While both Speculative Decoding and CS Drafting utilize a moderate amount of memory compared to Huggingface generation, there is little difference between the performance of CS Drafting and Speculative decoding due to the additional draft models used by CS Drafting being much smaller than the first draft model. --- Rebuttal Comment 1.1: Comment: Dear Reviewer LRu7, thank you again for reviewing our paper! We hope our response addresses your questions. Please let us know your thoughts, and we are more than happy to answer any further questions. --- Rebuttal Comment 1.2: Comment: Thank you to the authors for their response. It was helpful in gaining a better understanding of the paper. I am maintaining my positive score.	null	null	null	NeurIPS_2024_submissions_huggingface	2,024	null	null	null	null	null	null	null	null
Deep Learning in Medical Image Registration: Magic or Mirage?	Accept (poster)	Summary: The paper delves into the comparison between classical optimization and learning-based medical image registration methods. Some valuable insights are proposed and the authors propose a general recipe to choose the best paradigm for a given registration problem. Strengths: 1. The motivation behind the work is clear. The observation is in detail. 2. Rich experiments discover the regular and make people rethink the development of learning-based registration methods. 3. The illustrations are clear and easy to understand. 4. Clear conclusions are obtained for registration paradigms in different conditions. Weaknesses: 1. New solutions or deeper insights are missed. 2. Findings in Section 4,5,6 are somewhat valuable, but the relevancy between different phenomena is not very clear. The author failed to point out what should we do or what we may further research for unsupervised or supervised DLIR methods. Technical Quality: 3 Clarity: 3 Questions for Authors: What are the potential directions of future work based on this paper? Confidence: 3 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: N/A Flag For Ethics Review: ['No ethics review needed.'] Rating: 5 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We thank the reviewer for their positive and helpful feedback, we are glad to find they found the paper having a clear motivation, detailed observations, clear illustrations and easy-to-read paper. We believe we have addressed all the remaining concerns and hope the reviewer increases their score and advocates for the acceptance of our paper. ”New solutions or deeper insights are missed.”, “Findings in Section 4,5,6 are somewhat valuable, but the relevancy between different phenomena is not very clear” Our work serves as a bedrock for contextualizing the performance of current state-of-the-art deep learning and classical registration methods, and identifying the critical confounding variables responsible for the performance gaps observed in DLIR and classical methods. We show empirically the two confounding variables – instrumentation bias of classical methods, and label supervision are majorly responsible for the superior DLIR performance of in-distribution data. Our work provides an independent, fair and unbiased evaluation of both DLIR and classical methods, and show that in the unsupervised case, classical baselines outperform DLIR methods without having any in-distribution data – making it useful in the low-data regime. Furthermore, since anatomical labels are an intrinsic property of the anatomy (hence being domain-invariant), one may expect label supervision to add some robustness to domain shift. However, this is not the case (Sec 6). These experiments naturally imply a recipe to use for a given instance of registration (mentioned in Sec 7). Moreover, these results also allow us to reconsider deploying expensive data collection pipelines, especially when domain shift may be expected at test time (as a consequence of the finding in Sec 6). Moreover, observations in Sec.4. and Sec.6. serve as a problem statement for DLIR-based methods to alleviate these limitations. We urge the reviewer to also read more details about this in the common rebuttal. What are the potential directions of future work based on this paper? We appreciate the reviewer's interest in future research directions. Our limitations section outlines several avenues for further investigation. This study can serve as a foundational basis for examining and disentangling the variables responsible for robust performance in various contexts, such as: Multimodal Registration: Extending our analysis to multimodal registration scenarios to explore the applicability of our findings across different imaging modalities. Non-Brain Anatomy: Investigating the generalizability of our results to other anatomical regions beyond brain MRI. Diverse Registration Settings: As suggested by Reviewer 23gn, conducting experiments on inter-subject, intra-subject, and atlas-based registration to provide a comprehensive understanding of performance across different registration paradigms. We have also incorporated these future work suggestions more explicitly in the revised paper to guide readers on how our findings can be extended and applied in broader contexts. --- Rebuttal Comment 1.1: Comment: Thank you for addressing my concerns. I have no further questions and currently maintain my original rating. --- Reply to Comment 1.1.1: Title: Revising rating Comment: We thank you for reading the rebuttal. We would request you to increase your score if you think our responses are satisfactory. If not, we would love an opportunity to discuss your concerns.	Summary: The work benchmarks the traditional methods and deep learning-based methods for medical image registration and gives a general recipe to choose registration methods. Strengths: Comprehensive experiments: The authors implement several classical variational and deep learning-based registration models on four public datasets of brain CT/MR images. Weaknesses: - The training set of deep learning models may not be sufficient because the number of image pairs is obviously not enough, which potentially makes the comparison with classical models unfair. - Medical image registration contains a lot of content, it cannot and should not be simply represented by monomodal registration on four brain image datasets. The authors need to implement more experiments on images of more organs and tissues with monomodal and multimodal registration tasks. - The subtitle “DLIR methods do not generalize across datasets” is almost common sense that networks trained on specific datasets cannot easily generalize to other datasets. Technical Quality: 3 Clarity: 3 Questions for Authors: - Please explicitly clarify the advantages compared to the well-known lean2reg challenge. - Does the conclusion still hold on large datasets? AMOS dataset provides lots of CT and MR images can be used for model training. - Why not submit this work to the dataset and benchmark track? Confidence: 5 Soundness: 3 Presentation: 3 Contribution: 1 Limitations: yes Flag For Ethics Review: ['No ethics review needed.'] Rating: 4 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We thank the reviewer for their insightful and helpful feedback, which has significantly contributed to improving the quality of our paper. We believe we have addressed all the remaining concerns and hope the reviewer will consider increasing their score and advocating for the acceptance of our paper. The training set of deep learning models may not be sufficient All methods are trained on the modified OASIS training split of 364 images, yielding approximately 132k inter-subject training pairs (364 * 363). This is a substantial number of image pairs, and in practice, most DLIR methods converge within 2-3 epochs. Furthermore, most baselines like LapIRN train on an even smaller subset of OASIS, approximately 62k training pairs (250 * 249) (although we train on 364 images for consistency). Thus, the dataset size used is sufficient for a fair comparison with classical models. The authors need to implement more experiments on images of more organs and tissues with monomodal and multimodal registration tasks We have already acknowledged this in the Limitations section in our paper, which forms the basis for future research. We chose the inter-subject MRI brain registration problem due to its widespread adoption by the biomedical ML community, ensuring adequate reproducibility and unbiased evaluation. While we recognize the importance of exploring other organs and tissues, none of our analysis, baselines, or evaluation protocols use domain-specific knowledge, suggesting that our results are not strictly limited to brain MRI. Expanding to more variations is a priority for future work. The subtitle “DLIR methods do not generalize across datasets” is almost common sense We agree, this is almost trivial in hindsight! However, seeing the results in Sec.4. and Sec.5., one might expect the domain-shift performance to fall between the results of in-distribution unsupervised and supervised training, with supervised DLIR methods outperforming classical methods to some extent. However, this is not the case (please see common rebuttal for more details), showing that large amounts of labeled data would not necessarily help on datasets with even the same modality and resolution but slightly different intensity statistics. This raises a question for large-scale, annotated data collection efforts, and addressing current limitations of DLIR methods to be robust to domain shift. Please explicitly clarify the advantages compared to the well-known lean2reg challenge. Learn2reg challenge which focuses on methods to maximize registration performance (measured by label overlap, landmark distance, determinant of Jacobian) on a particular dataset with data/modality specific challenges. In contrast, we perform a meta-analysis and unbiased evaluation on the factors that lead to better performance of aforementioned registration algorithms. We find that DLIR methods attribute the performance improvement to architecture, loss functions, and training specifications. However, we show that most of the improvement is attributed to being able to learn from supervised label map objectives during training. Moreover, we show that training with label maps (which are intrinsic to modality and therefore invariant to the downstream modality – see more in common rebuttal) does not necessarily imply robustness in performance to domain shift. We have modified the “Contributions” subsection to emphasize these distinctions. Does the conclusion still hold on large datasets? AMOS dataset provides lots of CT and MR images can be used for model training The AMOS dataset, with 500 CT and 100 MR images, is comparable to the OASIS dataset in size (414 images). We selected the OASIS dataset for reproducibility and fairness considerations, as numerous DLIR methods provide pretrained models and training recipes for OASIS. The active community around OASIS and its long-standing presence has led to detailed analyses of dataset-specific challenges (for example, registration of smaller subcortical structures). Although the AMOS dataset is relatively new, it holds significant potential for future work, and we plan to explore it in subsequent studies. Why not submit this work to the dataset and benchmark track? Our work goes beyond merely reproducing or benchmarking existing methods. We provide a theoretical and empirical analysis of the factors responsible for improved performance in registration tasks, particularly the access to label maps during training, and their implications for robustness across similar datasets. This paper is a meta-analysis and unbiased evaluation, similar in spirit to [7]. --- Rebuttal Comment 1.1: Comment: Thanks for the response. The claimed new advantages compared to learn2reg are not convincing to me because these conclusions can also be derived from participants' algorithms. Accordingly, I slightly raise my score.	Summary: This paper discussed about an explicit correspondence between mutual information of the distribution and performance of classification registration methods. The authors argued that this correlation will not be affected by the learning-based methods. They validated this hypothesis on both classical and learning-based registration models and found that the weakly supervised learning-based methods performed high-fidelity intensity and label registration. Besides, they showed that these high-fidelity feature learning methods did not translate to invariance to domain shift, ending with proposing a general approach to select the best paradigm for a registration task. Strengths: S1. The authors experimented on the feature learning approach and their translation to invariancy to domain shift, which is an important aspect in deformable image registration. Also, they covered both classical and learning-based image registration to validate their hypothesis. S2. The authors considered out-of-distribution datasets for testing and evaluating performances, which showed promising results regarding generalization. Weaknesses: W1. Experimental findings. 1. I appreciate the authors' finding on getting improved performance for classical methods over learning-based methods. However, I would argue that this is typically not always true considering the reported findings in the learning-based papers, e.g., VoxelMorph (VM), TransMorph (TM), etc. This registration results highly depending on the image pairs that are being considered. For example, image pairs with large age variations won't perform well in the learning-based registration tasks (please check VM, Deepflash, and related papers where they restricted the subject ages to deal with large deformations). Besides, inter-patient registration has a higher probability of getting similar results compared to the atlas-based (or pre-selected template) registration, thoroughly discussed in VM and TM papers. I found this important information missing from their experimental evaluation to support their first hypothesis (Sec. 4). I would encourage the authors to consider experimenting and reporting performances for within-patient, inter-patient, and atlas-to-patient registration results. Also, comparing the anatomical dice scores further justifies the registration models as in the existing SOTA it has been discussed about getting larger variations for critical anatomical structures such as ventricles, pallidum, cerebellum, etc. 2. I found some of the mode's performance very inconsistent with the reported version in the original papers. For example - LKU-Net is achieving more than $~0.925\pm0.025$ dice score where the authors' reported best within subject accuracy is $0.8861\pm0.01$ with an average dice score of $0.7758\pm0.0390$. The same goes for the LapIRN model's performance. There's a large difference between the reported dice in this paper vs the original paper's dice. I understand there might be different biases that might be there in terms of implementation. However, the current reported performance of these models raises questions about the credibility of the adapted implementations, preparation of the pairwise images, hyperparameter setup, etc. I would suggest the authors kindly shed some light on this part. 3. I found the supporting experiments to validate the hypothesis presented in Sec. 6 missing some important information. Did the authors perform an affine transformation on all the datasets and what are the pre-processing steps? Are they considering all 3D volumes for their experiments? W2. Missing large deformation diffeomorphic and related baselines. I appreciate the authors for trying to carry out thorough experiments on different classical and learning-based registrations. However, I believe the current hypothesis would be more understandable and justifiable if the authors could initiate some experiments considering large deformation diffeomorphic-based registration methods such as LDDMM [1,2,3], where this kind of method is structured upon time-varying velocity fields, that is proven to be more robust in various deformation-based image analysis tasks. W3. Overclaimed hypothesis. Statements in L332-L335 seem overstated. The authors need to discuss and experiment on within-subject/class, inter-subject/class, and atlas/template-based registration to validate the stated hypothesis in that line, which seems to be missing in the current version. Besides, the selection of the DLIR/classical registration method depends on the image analysis tasks that are being performed over image registration. Without verifying the studied registration on different image analysis tasks, it is inappropriate to come to a conclusion stated in L332. W4. (Minor) Technical writing. The authors tried to aggregate all their potential findings in a structured way. But reading the whole paper kind of messed me up in understanding what are the actual contribution of this paper compared to the other survey/review papers in this domain other than performing experiments on OOD data. Overall, the presentation is kind of above the borderline but I guess if the authors tried to focus on their storyline and make their findings more clearer that would be great for the readers. For example, I found implementation details in most sections which is kind of redundant. Overall, I appreciate the authors for working on this paper which is very relevant as well as important in the medical imaging domain, specifically in medical image registration. However, the current version of the manuscript lacks some important experimental justification and further experiments. With that being said, the current version of the manuscript is under the threshold of acceptance. However, I am open to reconsidering the initial rating if the above concerns are adequately justified. References ---------------- [1] Yang, Xiao, Roland Kwitt, and Marc Niethammer. "Fast predictive image registration." Deep Learning and Data Labeling for Medical Applications: First International Workshop, LABELS 2016, and Second International Workshop, DLMIA 2016, Held in Conjunction with MICCAI 2016, Athens, Greece, October 21, 2016, Proceedings 1. Springer International Publishing, 2016. [2] Shen, Zhengyang, François-Xavier Vialard, and Marc Niethammer. "Region-specific diffeomorphic metric mapping." Advances in Neural Information Processing Systems 32 (2019). [3] Niethammer, Marc, Roland Kwitt, and Francois-Xavier Vialard. "Metric learning for image registration." Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2019. [4] Wang, Jian, and Miaomiao Zhang. "Deepflash: An efficient network for learning-based medical image registration." Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. 2020. Technical Quality: 3 Clarity: 2 Questions for Authors: Please see the weaknesses section. I tried to summarize all the findings, concerns, and questions there. Confidence: 4 Soundness: 3 Presentation: 2 Contribution: 2 Limitations: Limitations have been discussed in the paper. Flag For Ethics Review: ['No ethics review needed.'] Rating: 4 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We thank the reviewer for their highly detailed and insightful feedback, this has immensely helped in improving the quality of our work. We believe we have addressed all major concerns and incorporated changes in the paper. “I would argue that this is typically not always true considering the reported findings in the learning-based papers, e.g., VoxelMorph (VM), TransMorph (TM), etc” We appreciate this perspective and have addressed this concern in Sec4 (Instrumentation Bias) and Fig3. VM reports a Dice score of 0.749 for ANTs using suboptimal parameters ($\sigma$ of 9px followed by 0.4px), while we observe a Dice score of 0.787 with default parameters, highlighting the need for unbiased evaluation. Our evaluation of VM uses their pretrained model and scripts, yielding consistent results w/ their paper. Other studies on lung [5] and histology [6] registration also show that classical DIR methods far outperform DLIR methods. An unbiased and unified evaluation of algorithms is necessary to fairly compare the performance gaps of said algorithms. Our paper finds that TM is the only DL method beating classical methods (p<0.01) on the OASIS dataset with and without Dice loss (Fig 2 and 4). TM also demonstrates good generalization among DLIR methods competing with SynthMorph, although still underperforming compared to classical methods. We urge the reviewer to review these figures and results. We are happy to clarify any remaining discrepancies in the discussion. “Performance is different for inter-subject and atlas-based registration, image pairs with large age variations won't perform well in the learning-based registration tasks” In general, a vast majority of methods focus on inter-subject registration, and atlas-building and atlas-based registration are considered separate problems altogether. Therefore, we keep inter-subject registration datasets a consistent theme in the paper to corroborate existing findings and avoid mixing up results from different problem statements. Re large age variations: A good registration model in general should perform well across large variations in age, subject demographics, and acquisition configurations. Classical methods performing well for age variations further strengthen our findings in Sec.4 and Sec.6, i.e. labeled data becomes necessary in these settings for DLIR methods to perform better. We have added these observations into the paper further motivating our study. “consider experimenting and reporting performances for within-patient, inter-patient, and atlas-to-patient registration results” These are interesting experiments indeed and form the scope for future work, along with multimodal registration, and registration of different anatomy like chest CT or abdomen. Due to space constraints, fairness and reproducibility considerations, we choose OASIS dataset for training since a large number of SOTA DLIR methods provide training recipes for this dataset, and IBSR/CUMC/MGH/LPBA datasets that have been extensively used in unbiased community-standard neuroimaging evaluations [7,9]. “model performance is very inconsistent with the reported version in the original papers, LKU-Net is achieving 0.77 average Dice… same with LapIRN model's performance” We’re not sure where these numbers come from. In Fig.3 we show that the LKU paper reports a 0.88 Dice score, and our evaluation shows a 0.904 Dice score – Fig.4. (top) shows that LKU-Net is the top performer with Dice supervision. Similarly, we show in Fig2. that LapIRN reports 0.808 Dice but we report 0.788 – a reasonable deviation. In fact, LapIRN performance improves significantly when Dice supervision is added (Fig2 v/s Fig4), the latter of which the original paper did not implement or report. Our work takes great care to accurately reproduce and not misrepresent the performance of DLIR and classical methods. We ask the reviewer to review the subsection on Instrumentation bias (Sec 4) – we’re confident they will find DLIR methods represented accurately. DLIR methods may misrepresent classical baselines by choosing suboptimal hyperparameters; we work towards a more fair, unbiased and unified evaluation of classical and DLIR methods under the same umbrella. “hypothesis presented in Sec 6 missing some important information” For all datasets, we follow the preprocessing steps followed by Klein et.al. [7]. We have added this detail in the paper. “W2. Missing large deformation diffeomorphic and related baselines” Classical methods like ANTs, Greedy, FireANTs build on the LDDMM framework to avoid explicitly storing the full 4D velocity fields by integrating the infinitesimal velocities using gradient descent, & are shown to model very large deformations by independent evaluations [7,8]. Moreover, these are widely used classical baselines used by the broader biomedical community, and should not be seen as a reason to invalidate the results in the paper. “W3. Overclaimed hypothesis” We have changed L332-335 to reflect the conclusions for inter-subject registration, which is a defacto standard for registration. Since none of the analysis, baselines, and evaluation make any domain/subject-specific modeling assumptions, and the datasets are community-standard benchmarks, the results are valuable, both within the neuroimaging and the biomedical communities at large. “W4. (Minor) Technical writing” The contribution of the paper is to recalibrate claims about DLIR methods' performance by isolating label supervision as the primary contributor to their superiority. This motivates the question if label supervision can be robust to domain shift and finds it is not, prompting a reconsideration of data collection pipelines to improve registration performance, and rethinking DLIR design decisions to be robust to domain shift. We have made these changes in the contributions and discussion sections. --- Rebuttal Comment 1.1: Title: Official Comment by Reviewer 23gn Comment: I thank the reviewers for their response. After reading the rebuttal, I have the following standings — - Some of my concerns regarding LDDMM and the performances of different existing models have been adequately addressed. - The paper's contribution is limited considering the hypothesis that the authors evaluated in this paper. - I agree with my other reviewer that this paper is more aligned with the Dataset and Benchmarking track. Besides, I agree that experiments on datasets other than Brain could reinforce the contribution of their standings. More experiments related to intra-subject, atlas-based registration could further rectify their claims. After reading all the reviewers' comments and the rebuttal, I think the paper is still on the Borderline (keeping my score as it is), considering the technical contribution, carried out experimentations, overall writing, and the submission track. I suggest that the authors to address the findings from all reviewers in their revised version.	Summary: The manuscript investigates the characteristics of two types of registration approach based on traditional variational optimization and deep learning. Experiments revealed a correlation between the mutual information of the distribution of per-pixel intensity and labels, and the performance of classical registration methods. Then the manuscript argues unsupervised deep learning does not improve label matching performance compared to traditional methods, whereas supervised learning methods show improved label matching. Lastly they show learning methods do not generalize well. Strengths: 1. This study reflects original thinking, trying to draw insights on a challenging topic. 2. Experiments design is well motivated to study the hypothesis. 3. The results challenge a claim made by a number of existing studies "learning methods can provide improve label matching when optimized in an unsupervised fashion" Weaknesses: 1. Two of the three claims discussed herein seem trivial. It is not surprised that "Supervised DLIR methods demonstrate enhanced label matching" and "DLIR methods do not generalize across datasets". 2. I feel that the abstract and intro set up a high expectation by saying "we propose a general recipe to choose the best paradigm for a given registration problem, based on these observations." but then again it is a one-sentence recipe in the end that is not surprising to readers "a practitioner should choose DLIR methods only if they have access to a large labeled dataset, and their application is limited to the same dataset distribution. In all other cases, classical optimization-based methods are the more accurate and reliable choice" Technical Quality: 4 Clarity: 3 Questions for Authors: In Fig. 1, shouldn't there also be a correlation within each individual dataset? Confidence: 3 Soundness: 4 Presentation: 3 Contribution: 3 Limitations: While the intro and abstract suggest a study on general registration problems, all experiments are based on brain MRI. It is not clear whether the problem/hypothesis/conclusion is specific to brain MRI. Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We thank the reviewer for their time and insightful feedback, and are pleased to note the recognition of the originality of our work, the well-motivated design of our experiments, and the challenge posed to existing claims in the DLIR literature. We address some of their concerns below, and hope that their review and discussion can help convince all reviewers of the importance of our work, and champion our paper. Two of the three claims seem trivial. We agree that the claims seem straightforward in hindsight. However, we contextualize this conclusion from the claims made by existing DLIR literature, i.e. a carefully selected intensity loss function, network architecture, pretraining schema and large datasets lead to SOTA performance on alignment of anatomical regions. A variety of studies cited in our paper report results with and without label supervision, which we show is the confounding variable that leads to a significant improvement in performance. Therefore, performance of DLIR methods with and without Dice score must be compared to justify the factors involved in improving performance of DLIR methods. Our analysis in Section 3 elucidates that unsupervised methods do not benefit from advancements in network architecture, a point further verified in Section 4. Section 5 is included to underscore the fact that DLIR methods’ superior performance indeed comes from incorporating label supervision during training, a capability absent in classical methods. Section 6 seems trivial in the context of our paper, but for the biomedical community at large, pretrained models and large datasets are built to train models that generalize outside the training domain (see common rebuttal). Our findings in Section 6 suggest a need to reassess huge labeled data collection efforts and to formulate DLIR advancements that exhibit robustness against minor domain shifts instead. This reevaluation is crucial for leveraging large amounts of annotated data effectively. We have added this both in the Motivation and Discussion sections. “I feel that the abstract and intro set up a high expectation..” We acknowledge that the phrasing of our contribution may have set an overly ambitious expectation. We have revised the statement in the paper to: “We reassess and recalibrate performance expectations from classical and DLIR methods under access to label supervision and training time, and its generalization capabilities under minor domain shifts.” This modification more accurately reflects the scope and implications of our study. “shouldn't there also be a correlation within each individual dataset” We find that since the labeled regions are fixed within a dataset, the variation between MI and registration score is dominated by other intra-dataset factors. The correlation between MI and registration performance occurs at the granularity of the dataset, which is shown in Fig.3. This granularity highlights the broader applicability of our findings across different datasets rather than within a single dataset. “are the hypothesis/conclusion specific to brain MRI”: We assert that the methods and evaluation metrics employed in our study do not incorporate domain-specific information, making our observations broadly applicable. Brain MRI was chosen due to the extensive availability of solutions and open-access resources provided by the neuroimaging and machine learning communities. This choice ensures that our analysis is unhindered by fairness or reproducibility concerns. --- Rebuttal Comment 1.1: Comment: Thanks for clarification. I'm still positive about the paper after reading all responses and therefore remain my rating.	Rebuttal 1: Rebuttal: We thank all reviewers for their insightful feedback and for taking time to improve the quality of our work. We are glad that reviewers found our work reflecting original thinking and drawing useful insights [tuv2], well motivated problem [Aeaf], consideration of both classical and DLIR methods [23gn], well motivated and well designed experiments [tuv2, Xf1e, Aeaf], clear writing [Aeaf]. We have addressed all questions in the individual comments. We summarize and clarify some common concerns: Some of the claims are trivial We agree that these claims look trivial in hindsight. However, these claims must be contextualized from the perspective of claims made by existing DLIR literature. Specifically, most DLIR methods propose either a new architecture, training recipe (multi-scale, cascaded, or pre-training), or loss functions, and claim that these methodologies lead to superior performance over classical registration methods. However, the argument in Section 3 / Fig2. indicates that these contributions do not add any information on the existing correlation between mutual information and dice score. Hence, the primary driver of performance improvement is anatomical label supervision. Our work provides a unified and fair evaluation of classical and DLIR methods without label supervision (Sec 4), confirming that the conclusions from Section 3/Figure 2 hold true. An ablation study over label supervision (Sec 5) further verifies that label supervision is indeed the confounding variable leading to superior performance. Thus, Sections 4 and 5 are pivotal in isolating the real confounding variable, both theoretically and empirically. Importantly, anatomical labels are intrinsic to the anatomy and not inherently modality-dependent, making them domain-invariant. This motivation for using synthetic data with consistent label maps is often used to train domain-invariant models [1][2]. Our key non-trivial finding, presented in Section 6, reveals that while DLIR models trained with label supervision may perform worse under domain shift (DS) compared to in-distribution (ID) scenarios, the expected performance hierarchy does not hold: Expected: (DLIR Sup ID) > (DLIR Sup DS) > (Classical) Observed: (DLIR Sup ID) > (Classical) > (DLIR Sup DS) This significant finding highlights that DLIR methods do not maintain their performance advantage over classical methods under domain shift. This result has profound implications for annotated data collection and challenges the notion that large labeled datasets ensure robust generalization. For example, Learn2Reg 2023 [3] mentions “By providing easy-to-use solutions applicable to a variety of medical registration problems, we hope to strengthen both generalization and comparability between them.”. However, our findings underscore that the generalization of DLIR methods to domain shifts is heavily understudied and generally negative. This necessitates a reassessment of expensive labeled data collection efforts and emphasizes the need for developing DLIR methods, like those proposed in [1], that are more robust to domain shifts. We have added a subsection in Section 3 summarizing these motivations and the broader implications of our findings. We hope this better contextualizes and emphasizes the significance of our results. Are the findings applicable to brain MRI only? We assert that the methods and evaluation metrics employed in our study do not incorporate domain-specific information, making our observations broadly applicable. Brain MRI was chosen due to the extensive availability of solutions and open-access resources provided by the neuroimaging and machine learning communities, also somewhat inspired by [7]. This choice ensures that our analysis is unhindered by fairness or reproducibility concerns. What about other modalities / anatomy / intra-subject / atlas-based settings? We have already acknowledged this in the Limitations section in our paper, forming the basis for future research. We chose the inter-subject MRI brain registration setup due to its widespread adoption by the biomedical ML community, ensuring adequate reproducibility with unbiased and fair evaluation. Registration literature in other anatomy and modalities – hippocampus - Table 1 in [4], lung images [5], and histology [6] also seem to suggest similar results (classical methods outperform DLIR in unsupervised settings), but a more comprehensive study with unified training and inference setup is necessary and forms the motivation for future work. [1] Hoffmann, Malte, et al. "SynthMorph: learning contrast-invariant registration without acquired images." IEEE transactions on medical imaging 41.3 (2021): 543-558. [2] Dey, Neel, et al. "AnyStar: Domain randomized universal star-convex 3D instance segmentation." Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. 2024. [3] https://learn2reg.grand-challenge.org/learn2reg-2023/ [4] W. Zhu, Y. Huang, D. Xu, Z. Qian, W. Fan and X. Xie, "Test-Time Training for Deformable Multi-Scale Image Registration," 2021 IEEE International Conference on Robotics and Automation (ICRA) [5] Fu, Yabo, et al. "LungRegNet: an unsupervised deformable image registration method for 4D‐CT lung." Medical physics 47.4 (2020): 1763-1774. [6] Borovec, Jiří, et al. "ANHIR: automatic non-rigid histological image registration challenge." IEEE transactions on medical imaging 39.10 (2020): 3042-3052. [7] A. Klein, et al. Evaluation of nonlinear deformation algorithms applied to human brain MRI registration. NeuroImage, 46(3):786–802, July 2009. [8] Murphy, Keelin, et al. "Evaluation of registration methods on thoracic CT: the EMPIRE10 challenge." IEEE transactions on medical imaging 30.11 (2011): 1901-1920. [9] Mok, Tony CW, and Albert Chung. "Affine medical image registration with coarse-to-fine vision transformer." Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2022.	NeurIPS_2024_submissions_huggingface	2,024	null	null	null	null	null	null	null	null
EMR-Merging: Tuning-Free High-Performance Model Merging	Accept (spotlight)	Summary: The paper presents a method called EMR-MERGING (ELECT, MASK & RESCALE-MERGING) for merging models finetuned on different tasks into a single model with multi-task capabilities without the need for additional tuning or training. This method addresses the limitations of existing model merging techniques, which often suffer from performance degradation or require additional data and tuning. Strengths: - Theoretical and Empirical Analysis: The paper provides a solid theoretical foundation for the effectiveness of the proposed method, including detailed proofs and empirical analyses. This adds robustness to the claims made by the authors. - Tuning-Free: One of the major advantages of this method is that it does not require any additional data, tuning, or training, making it highly practical for real-world applications where data availability is limited or privacy concerns restrict access. Weaknesses: - Additional Computational Overhead: The proposed method introduces extra computational overhead during inference due to the use of masks and rescalers. This aspect should be discussed in the paper, particularly regarding its impact on the overall efficiency and performance in real-world applications. - Storage Considerations for PEFT Models: Pre-trained models finetuned on downstream tasks using Parameter-Efficient Finetuning (PEFT) methods typically require only the storage of the pre-trained model weights and additional adapter weights. These adapter weights consume significantly less storage space compared to the masks utilized in EMR-MERGING. The authors should address this comparison and discuss the storage implications in detail. - Impact of Model Quantization: With the trend towards quantization of models, reducing their precision from 32-bit to 8-bit, 2-bit, or even 1-bit, the storage overhead of masks becomes increasingly significant. Quantization is an essential step in the development of large models, and as such, the relative storage cost of the masks in EMR-MERGING will be more pronounced in the future. The authors should consider discussing the impact of model quantization on their method and the potential challenges it poses. - Limited Dataset Size in Experiments: The datasets used in the experiments are relatively small, which undermines the persuasive power of the results. For example, in the case of visual tasks, large models like CLIP's ViT-L exhibit strong generalization capabilities and can achieve high zero-shot classification performance on these small datasets. As a result, the performance differences among various model merging methods might not be significant. The authors should consider using larger, more diverse datasets to better demonstrate the effectiveness and robustness of their approach. Technical Quality: 3 Clarity: 2 Questions for Authors: NA Confidence: 4 Soundness: 3 Presentation: 2 Contribution: 3 Limitations: NA Flag For Ethics Review: ['No ethics review needed.'] Rating: 5 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thank you for your hard work and constructive comments. --- ## Weakness 1: Additional Computational Overhead. Ans: The unified task vector, masks and rescalers are computed during the merging process, which is before the inference process. During the inference process, before we evaluate the performance on a single task, we simply apply the specific mask and rescaler to the unified task vector. This only causes little additional computational costs. --- ## Weakness 2: Storage Considerations for PEFT Models Ans: When merging PEFT models, the task vectors are only generated by the adapters. EMR-Merging only needs to store the masks applied to these adapters because the other weights in the model are fixed. Therefore, compared to storing $N$ adapters, using EMR-Merging to store a unified one and $N$ masks for the adapters can be a strong alternative. --- ## Weakness 3: Impact of Model Quantization Ans: + The focuses of model merging and Quantization overlap but are not the same. Model merging focuses on reducing the deployment and storage costs by giving a model multi-task capabilities without training while quantization focuses on reducing computational and storage costs of a single model on a specific task by reducing parameter precision. + Barriers to combining the model merging and low-bitwidth quantization. Though combining model merging and low-bitwidth quantization can further reduce the storage costs and computational costs, all the existing model merging methods cannot be directly applied to low-bitwidth quantized models including INT8, INT4, etc. This is due to that model merging needs to subtract or average the weights from different models but after quantization, the low-bitwidth weights cannot be directly operated because these weights may have gone through different quantization functions due to different models' specific weight distribution. Combining model merging with low-bitwidth quantization has broad application prospects and is a potential future work. + Merging FP-16 models. In order to explore the applicability of EMR-Merging to FP16 models, we conduct experiments on RoBERTa models loaded in FP16 and the results are shown in Table R12 of the Rebuttal PDF. EMR-Merging still shows promising performance, demonstrating its potential to be applied to quantized models. + Quantizing the unified task vector. Although the quantized model cannot be directly merged, we find that the unified task vector can be simply quantized. We layer-wisely convert the elements in the unified task vector into low-bitwidth ones linearly. We use this method to merge RoBERTa models and the results are shown in Table R13. We find that EMR-Merging 's performance is not affected when the unified task vector is quantized to 4-bit and is still promising when the unified task vector is quantized to 2-bit. This can further reduce the storage cost of EMR-Merging. + The increasingly significant storage overhead of masks. A potential limitation of EMR-Merging is that, with the trend towards quantization of models, the storage of the masks can be increasingly significant. We will add this to the Limiataion section. However, compared to storing multiple individual non-1-bit quantized models, the storage costs by the masks (1-bit) and rescalers (a single value) are still fewer. When applied to merge multiple models, EMR-Merging may still tend to be a competitive alternative. We will add this discussion to the Future Work section. --- ## Weakness 4: Limited Dataset Size in Experiments Ans: + The good performance of EMR-Merging is not merely due to the strong zero-shot capabilities of CLIP models because other merging methods normally suffer from significant performance degradation compared to individual models while EMR-Merging can achieve the performance close to individual models. + We have conducted some experiments on large-scale datasets including COCO, ImageNet, and VQA v2 when merging multiple BEiT3 models and EMR-Merging demonstrates its good performance on all these large-scale datasets. + Your opinion that using a larger and more diverse dataset may better demonstrate the effectiveness of EMR-Merging is reasonable. Based on merging eight vision tasks using CLIP ViT-B/32, we add a task, i.e., ImageNet-1K. We downloaded a CLIP ViT-B/32 model finetuned on the task in timm and merged the nine models. The results are in General Response, Additional Result #2. EMR-Merging shows a much more significant improvement compared to existing merging methods (up to 20%). We will release this benchmark as a new model merging benchmark. --- Rebuttal 2: Comment: The author seems to have misunderstood my points regarding Weaknesses 2 and 3. I would like to clarify them as follows: - Weakness 2: A pre-trained model with N adapters can also perform multiple tasks. In this case, I don't quite understand why model fusion is necessary. What advantages does the proposed EMR-MERGING method have over the PEFT approach? PEFT allows for fine-tuning downstream tasks while reducing storage space requirements and achieving satisfactory results. Furthermore, PEFT consumes significantly less storage space compared to storing a mask. - Weakness 3: In cases of highly quantized models (e.g., 8-bit), EMR-MERGING would occupy 1/8 of the original model’s size, which significantly limits the effectiveness of the method. In contrast, PEFT does not encounter this issue. --- Rebuttal 3: Title: Rebuttal by Authors Comment: Thank you so much for your review and feedback. We further address your comments regarding Weakness 2 and Weakness 3 as follows: Answer to Weakness 2: 1) Model finetuning and model merging are fundamentally different tasks with different targets. For model fine-tuning, it consists of full-parameter fine-tuning that pursues fine-tuning performance and parameter-efficient fine-tuning (PEFT) that pursues fine-tuning efficiency, both of which have the same target of customizing models for specific domains, necessitating labeled datasets and sufficient computational resources for supervised training. For model merging, under the background that there exists an exponentialy increasing number of pre-trained or finetuned model weights, it targets to take advantage of these existing weights and obtain a single model with multi-task abilities, without the need for labeled data or supervised training. 2) Model merging can be combined with kinds of model fine-tuning methods. More specifically, the proposed EMR-Merging can be applied to not only fully finetuned models but also PEFT models, as shown in Tab. 6 of our paper. We only need the fine-tuned adapters on each task, and EMR-Merging can merge these adapters into a single adapter with a few masks and rescalers, whose overhead is less than storing all the adapters while achieving performance close to that of multiple individual adapters. 3) Model finetuning and model merging are both significant techniques. In this paper, our EMR-Merging focuses on realizing tuning-free and high-performance model merging instead of reducing storage space requirements as much as possible like PEFT. Besides, model merging and PEFT can be easily combined to achieve further parameter reduction, which is a potential area for future work. For example, there have been some studies focusing on merging PEFT models [1,2]. Answer to Weakness 3: 1) When compared to multiple individual 8-bit models, EMR-Merging that consists of one 8-bit model and several masks and scalars can still significantly reduce the parameter numbers. 2) When compared to other merging methods, EMR-Merging shows significantly better performance requring no tuning, demonstrating its greater applicability. 3) A potential solution to your concern is using PEFT techniques to train the quantilized models and merge the PEFT modules using EMR-Merging. This can minimize the number of parameters on multiple tasks, which will be included in our future work. ---- References [1] Parameter efficient multi-task model fusion with partial linearization, ICLR 2024. [2] Composing Parameter-Efficient Modules with Arithmetic Operations, NeurIPS 2023. --- Rebuttal Comment 3.1: Comment: Thank you for your responses, I appreciate the effort and detail you have provided in addressing the concerns raised. I have no additional questions at this time and will determine my final score after conferring with the Area Chair and the other reviewers. --- Reply to Comment 3.1.1: Title: Thanks for your comments Comment: We appreciate your comments, which can help improve our paper's quality greatly and inspire our future work. The discussion about the impact of model quantization and PEFT will be included in the revision. Thank you for your efforts. You are welcome to discuss with us if you have any other questions.	Summary: The authors identify two issues in current model merging methods: significant performance degradation from the multi-task counterpart and requirement of additional data and training. To tackle those issues, they propose EMR merging, which first creates a unified task vector, then selects masks based on each task-specific model’s correspondence with the unified task vector, finally rescales it. EMR merging shows strong empirical performance among several benchmarks. Strengths: 1. The paper is well written and the algorithm is coupled with adequate amount of analysis. 2. The proposed algorithm is computationally lightweight, yet obtains significant performance improvement. Weaknesses: 1. The novelty is limited. The three steps are essentially a combination of TIES and DARE, where only the masking step is slightly different from TIES. Due to this high similarity, it is unclear why the proposed method performs so much better than TIES and DARE. It would be better to provide a more disentangled analysis for each of the three steps through ablation studies, so that the readers can have a better understanding where exactly the performance improvement comes from. 2. Several related works are missing [1,2] (both on ArXiv in Feb 2024). It is important to do a thorough comparison, especially when the paper basically claims SoTA results. 3. Not all methods are compared on all benchmarks. For instance, AdaMerging is not compared with on NLP benchmarks. Although the original paper of AdaMerging does not provide results in NLP, I do not see any algorithmic obstacle that prevents it from being implemented in the NLP setting. [1] Representation Surgery for Multi-Task Model Merging. Yang et al. ICML 2024. [2] Merging Multi-Task Models via Weight-Ensembling Mixture of Experts. Tang et al. ICML 2024. Technical Quality: 3 Clarity: 3 Questions for Authors: 1. What is the sparsity of the masks? I know that it is kind of task/model dependent, but it would be nice to show those results. 2. Do EMR still need the hyperparameter tuning on the scaling factors? Or the one from rescalder is final? Because in DARE, they tune the scaling factors even after they compute the rescaling factors. Confidence: 5 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: The authors adequately addressed the limitations. Flag For Ethics Review: ['No ethics review needed.'] Rating: 5 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thank you for your hard work and helpful comments. --- ## Weakness 1: limited novelty. The three steps are a combination of TIES and DARE. Ans: The proposed method is totally different from DARE+TIES-Merging. + The motivation for EMR-Merging is different from existing methods. We first decouple model merging into unified parts and task-specific parts. Existing model merging methods can be formulated as: $W_M = \mathcal{M}\left(\left[ W_1..W_N \right]\right)$, while our EMR-Merging can be formulated as: $W_{uni}, \left[ E_1..E_N \right] = \mathcal{M'}\left(\left[ W_1..W_N \right]\right)$. We elect the unified task vector to reserve the most significant elements and use the masks and rescalers to align the signs and the magnitudes of the merged model, thus bringing about significant improvements. As Reviewer Lqs2 noted, the idea of electing a unified model and identifying different modulators for different tasks is interesting because directly merging models can lead to function conflicts. Differently, TIES and DARE are both proposed to reduce interference but a single model may not optimally simulate the task-specific model weights on all the tasks. + The process of EMR-Merging is different from TIES and DARE. TIES and DARE drop partial elements based on magnitude or randomly. The goal of dropping is to reduce interference while our EMR-Merging can avoid sign conflicts by pre-computed masks. DARE's rescaling is to compensate for the dropping process while ours is to align each task-specific model. Additionally, the resclaers for DARE are normally larger than 1 while the rescalers for EMR-Merging are normally smaller than 1 as shown in Table R8 of the Rebuttal PDF. This reflects the difference in the functions of the rescalers of EMR-Merging and DARE. + EMR-Merging shows tuning-free and plug-and-play properties. TIES+DARE introduces multiple hyper-parameters while EMR-Merging requires no hyper-parameter tuning, making it highly practical for real-world applications (Reviewer b5vt). EMR-Merging also shows plug-and-play capabilities and the performance of EMR-Merging has been verified under vision, NLP (including PEFT and FFT), and multi-modal settings. + Experimental comparisons. We additionally compare the performance of EMR-Merging and TIES+DARE in Table R11 of the Rebuttal PDF. It can be seen that the performance TIES+DARE is sensitive to hyper-parameters and our tuning-free EMR-Merging achieves better performance. + The reason for the improvement. As to the reason for the performance improvement, we believe that it is because EMR-Merging uses masks to avoid sign conflicts and rescalers to lower the L2 distance between the merged model and task-specific ones. Please check Appendix B for more information. Additionally, we have conducted ablation studies on all three steps of EMR-Merging in Tab. 8 and 9 of our paper. The results show that: 1) exiting merging methods can obtain performance improvement through Mask and Rescale. 2) Both masking and rescaling can boost the performance of the elected vector. --- ## Weakness 2: comparison to the WEMoE [R3] and Representation Surgery [R4]. Ans: It should be noted that we did not compare EMR-Merging to WEMoE [R3] and Representation Surgery [R4] in our paper because they had not been officially published by ICML when we submitted the paper. We compare the parameter numbers and prerequisites of WEMoE, Representation Surgery, and EMR-Merging in Table R1 of the Rebuttal PDF. The proposed EMR-Merging needs no data, training, or tuning while WEMoE and Surgery require unlabeled test data to train partial parameters in their frameworks. Meanwhile, EMR-Merging achieves promising performance. The results of merging ViTs on the eight vision tasks are shown in Table R2 of the Rebuttal PDF. EMR-Merging achieves the SOTA performance on the benchmark of merging ViT-L/14 models. Moreover, EMR-Merging can be easily extended to other settings including NLP and multi-modal. --- ## Weakness 3: not all methods are compared on all benchmarks. Ans: AdaMerging needs to train the merging coefficients. Shifting it to NLP settings requires re-building the framework and re-training. Since AdaMerging is only implemented in vision settings officially and its implementation in other settings is not yet realized by some newly-released toolkits and benchmarks including MergeKit [R1] and FusionBench [R2], we only compare the performance of AdaMerging under vision settings. We argue that the rebuttal time is very limited and re-building and re-training all the methods including AdaMerging under all the benchmarks is beyond the scope of this paper. However, we are glad to release the benchmarks, checkpoints, and the code of EMR-Merging under all the experimental settings. We have released all the checkpoints and datasets used in our experiments in the anonymous github repository of the manuscript. We have released the code under vision settings and we will release the code for all the other experimental settings soon. --- ## Question 1: the sparsity of the masks. Ans: We show some results of the sparsity of the masks under different settings in Table R5, R6, and R7 of the Rebuttal PDF. --- ## Question 2: does EMR-Merging still need the hyperparameter tuning on the scaling factors? Ans: We promise that EMR-Merging requires no tuning on any hyper-parameter under any experimental setting. Actually, EMR-Merging has no hyper-parameters. The unified task vector and the task-specific modulators are all calculated using model weights. We have released the anonymous code under vision settings and we welcome you to reproduce our experiments. --- ## References [R1] Arcee's MergeKit: A Toolkit for Merging Large Language Models. [R2] FusionBench: A Comprehensive Benchmark of Deep Model Fusion. [R3] Merging Multi-Task Models via Weight-Ensembling Mixture of Experts. [R4] Representation Surgery for Multi-Task Model Merging.	Summary: Model merging directly fuses multiple independently trained models at the weight level to obtain a single comprehensive model, which is a current research hotspot. This paper proposes a new model merging method EMR-MERGING to reduce the performance gap between the merged model and the independent models. Strengths: - The proposed EMR-MERGING method does not require additional training, data, and tuning, which enhances the applicable scenarios of the method. - This paper conducts extensive experiments, including visual models, NLP models, and multi-modal models. In particular, the visual model is extended to 30 tasks. - This paper has a clear structure and the proposed methods are easy to implement. Weaknesses: - Loss of parallel ability: Existing model merging methods obtain a unified model for all tasks, so samples from all tasks can be included in one inference (i.e., a batch of data). However, EMR-MERGING needs to configure an independent model for each task during inference, thus losing the ability of parallel inference. In real scenarios, a batch of data may come from different tasks rather than a single task. - Important baseline missing: It is unclear what the benefits and advantages of EMR-MERGING are compared to "sparse independent models". Ties-Merging shows that removing 90% of the neurons in each independent model has almost no change in performance. So why not just use the sparse version of the task vector instead of using EMR-MERGING? - Symbol or spelling error: - The symbols are not uniform, and the uniform vector is sometimes $T_F$ and sometimes $T_{uni}$. - Table 11: “huper-parameter settings” -> “hyper-parameter settings” - Sec 4.2.2: “using 1eleven datasets” -> “using eleven datasets” - Finally, there is a lack of datasets and checkpoints for CV (30 tasks), NLP, and multimodal in the anonymous link. If the authors can release them, it will help the community research. Technical Quality: 3 Clarity: 3 Questions for Authors: - Is it possible to select only the maximum value in the Electing operation? Is it also possible to select the average value? - In Figure 4(c), regarding Cos similarity, TV > Ties-Merging > AdaMerging, but in terms of accuracy, TV < Ties-Merging < AdaMerging. Figure 4(b) is similar. These phenomena seem to be contradictory. How does the author explain this? Confidence: 4 Soundness: 3 Presentation: 3 Contribution: 2 Limitations: The authors discuss some limitations of the method: for example, it consumes additional memory compared to other model merging methods, and all rely on the same pre-trained model. However, as in Weaknesses, this method may lose the ability to parallelize inference on multiple tasks, and it is unclear how much benefit it has compared to sparse independent models. Flag For Ethics Review: ['No ethics review needed.'] Rating: 5 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thank you for your hard work and kind comments. --- ## Weakness 1: parallel ability. Ans: 1) Most multi-task model merging methods cannot handle the situation where multi-task samples are included in one inference because only one classification head can be applied during one inference. All the existing merging methods do not merge classification heads [R1-R3] because the classification heads may not share the same architecture. The usual practice of merging methods is applying the proper classification head manually to the merged model before testing on a specific task. 2) We follow this setting and additionally apply a mask and a rescaler to the unified model before testing on a specific task. Once the problem of manual classification head selection is handled, the modulators can also be selected automatically, thus realizing parallel inference. 3) A potential solution is using a gating network to sample-wisely determine which task an input comes from before each inference [R4]. --- ## Weakness 2: comparison to multiple Sparse Models Ans: Here we show a detailed comparison. + The performance of EMR-Merging is more promising. We compare the performance of Sparse Models and EMR-Merging on merging (2-8) ViT-B/32 models and 30 ViT-B/16 models in Table R3 of the Rebuttal PDF. The $K$ for the Sparse Models refers to keeping the top $K\%$ elements following TIES-Merging [R2]. It can be seen that EMR-Merging performs better than Sparse Models when merging vision models. Additionally, we compare the performance of Sparse Models and EMR-Merging on merging BEiT3 models in Table R4 of the Rebuttal PDF. The performance of Sparse Models severely decreases on some tasks while EMR-Merging shows promising performance on all tasks. + EMR-Merging leads to smaller storage costs when merging multiple models. The comparison of parameter numbers is shown in Figure R1 of the Rebuttal PDF. EMR-Merging shows both fewer parameter numbers and better performance when compared to Sparse Models (K=10). Though Sparse Models (K=5) can realize fewer parameters, their performance is severely reduced. It should be noted that the number of parameters in EMR-Merging increases more slowly with the number of tasks due to the lightweight task-specific modules. Therefore, when the number of tasks further increases, the parameter numbers of EMR-Merging tend to be fewer than Sparse Models (K=5). + The elected unified task vector of EMR-Merging can also be sparse. During EMR-Merging, after we elect $\tau_{uni}$ and before calculating the rescalers, we can make $\tau_{uni}$ sparse by keeping the top $10\%$ ones. This can further reduce the storage cost of EMR-Merging. We identify this method by EMR-Merging (Sparse). The performance and parameter numbers of EMR-Merging (Sparse) can be found in Table R2 and Figure R1 of the Rebuttal PDF. It can be seen that EMR-Merging (Sparse) sharply reduces the parameter numbers while maintaining high performance, which is still better than Sparse Models while showing the fewest parameter numbers. --- ## Weakness 3: typos. Ans: In the method section, the $\tau_{uni}$ was mistakenly written as $\tau_{F}$. Sorry for that. However, in Theoretical Analysis, we use $\tau_{F}$ to refer to the collection of the merged task vector by existing merging methods and the elected task vector $\tau_{uni}$. This is to demonstrate that regardless of which method is applied for merging (or electing), masking and rescaling can lower the distance between the merged model and individual models, thus explaining why other methods can be combined with Mask & Rescale and get improved in Tab. 8. To avoid misunderstanding, we will uniformly use $\tau_{uni}$ in the revision. We will also fix the other typos and thanks for pointing them out. --- ## Weakness 4: datasets and checkpoints. Ans: We have released the checkpoints and datasets in the anonymous link of the manuscript. Hopefully, the released checkpoints and newly established benchmarks can help community research. --- ## Question 1: Is the maximum or the average strategy possible? Ans: Additional ablation studies on the electing process are shown in Table R9 and R10 of the Rebuttal PDF. Applying the maximum strategy may result in better performance on partial tasks including vision tasks and partial multi-modal tasks and worse performance on the rest tasks. In contrast, the average strategy results in worse performance on most tasks. Note that the main contribution of EMR-Merging is the idea of decoupling model merging into unified parts and task-specific parts. The unified vector can be obtained by not only electing but also existing merging methods, which is shown in Tab. 8. EMR-Merging shows tuning-free, data-free, and plug-and-play properties. The effectiveness of EMR-Merging is verified through theoretical and empirical analysis. --- ## Question 2: the contradictory phenomena. Ans: 1) The proposed EMR-Merging first considered using masks and rescalers to align the signs and magnitudes. Therefore, the sign conflicts can be avoided and the L2 distances can be maximally reduced, as we illustrated in Theoretical Analysis. This explains why EMR-Merging shows the lowest sign conflicts, L2 distance, and the highest cos-sim in Fig. 4. 2) In Fig. 7, we show the sign conflicts, L2 distance, and cos-sim of each step of EMR-Merging. Combined with Tab. 9, it can be seen that during the 3 steps of EMR-Merging, we reduce sign conflicts and L2 distance and increase cos-sim, and the performance is correspondingly increased. 3) Existing methods may not focus on reducing sign conflicts and L2 distances, thus their performance may not correspond with sign conflicts and L2 distances. --- ## References [R1] Editing models with task arithmetic [R2] Ties-merging: Resolving interference when merging models [R3] Adamerging: Adaptive model merging for multi-task learning [R4] Merging Vision Transformers from Different Tasks and Domains	Summary: This paper aims to improve the performance of model merging in the multi-task learning domain. By electing a unified model, identifying a mask and a rescaling factor for each task, the proposed method EMR-Merging is able to significantly improve the merging performance over the previous state-of-the-art, such as Adamerging, Ties-Merging, Task Arithmetic, etc. Besides the CV model merging, EMR-Merging also effectively merges the large language models, approaching the individual fine-tuned performance. Strengths: 1. I like the idea that we need to elect a unified model first, and then identify different masks for different tasks. Directly merging models can lead to function conflicts. 2. The experiments are extensive, including VIT model, language models, and multi-modal models, with considerable good performance. 3. The ablation study is also provided. Weaknesses: 1. I am curious why the proposed way works for the unified model. One specific case is that, if the the average of the task vector is close to zero, but not exactly equal to zero, e.g., 0.0001, which will give us a positive sign. If the magnitude of this weight is large, a positive sign multiplied by this large magnitude weight can lead to a large error. Can the authors elaborate more the intuition behind the unified model, mask, and rescaler? 2. Can you explain the process of obtaining the mask? I could not find the definition of $\tau_F$ after reading the paper in detail. 3. What is the configuration of Figure 4? More details are needed here. 4. How to interpret the t-SNE visualization? 5. The distance between figures is over-reduced. It can be confusing and misleading. For instance, figure 5 and figure 4. 6. How expensive to extend EMR-Merging to LLMs, such as LLaMa3? 7. How does the proposed method perform compared to SOTA https://arxiv.org/pdf/2402.00433? Technical Quality: 3 Clarity: 3 Questions for Authors: Please see the above weaknesses. Confidence: 4 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: Provided. Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thank you for your hard work and admitting the value of our contribution. --- ## Weakness 1: why the proposed way works, what if the average of the task vector is close to zero, and the intuition of the method. Ans: When we elect the sign vector, we first element-wisely add the task vectors and obtain the elected sign, i.e., $\gamma_{uni} = sgn(\sum_{t=1}^{N} \tau_t)$ when merging $N$ task vectors. Then we element-wisely choose the maximum absolute value of each parameter with the sign consistent with $\gamma_{uni}$ and obtain the elected unified task vector $\tau_{uni}$ . It contains significant elements from all the task vectors. Next, we calculate the masks by: $M_i= \left( \tau_i \odot \tau_{uni} > 0 \right)$ for task $i$. Finally, we calculate rescalers by $\lambda_i = \frac{sum(abs(\tau_{i}))}{sum(abs(M_i \odot \tau_{uni}))}$. We apply absolute values in both the numerator and denominator, avoiding the situation that the average of a task vector is close to zero. Additionally, the $\tau_{uni}$ contains the maximum values of the elected direction. This will also prevent the denominator from approaching zero. We observe that the rescalers are normally not larger than 1. We provide the values of the rescalers under various settings in Table R8 of the Rebuttal PDF. When designing EMR-Merging, we hope to reserve the shared and significant features from all the task vectors in the unified task vector $\tau_{uni}$. We reserve the shared features among task vectors by electing the sign vector and reserve the significant features by electing the elements of the largest absolute values with the sign consistent with the sign vector. Ties-Merging shows that sign conflicts cause significant performance degradation. DropOut [R3] and DARE [R1] shows that simply dropping without magnitude alignment can also effect the performance. Inspired by this, we use masks to avoid sign conflicts between the unified task vector and task-specific ones and use rescalers to align the magnitude. --- ## Weakness 2: the process of obtaining the masks; the definition of $\tau_{F}$. Ans: The mask for task $i$ is calculated by: $M_i= \left( \tau_i \odot \tau_{uni} > 0 \right)$. In the method section, the $\tau_{uni}$ was mistakenly written as $\tau_{F}$ when introducing the process of obtaining the masks. Sorry for the mistake. However, in the theoretical analysis, we use $\tau_{F}$ to refer to collection of the merged task vector by existing methods (e.g., Ties-Merging) or the elected task vector $\tau_{uni}$. The definition of $\tau_{F}$ is to demonstrate the effectiveness of the masks and rescalers regardless of the electing process. This can explain why other methods can be combined with Mask & Rescale and obtain performance improvement in Table 8. We find that this may lead to misunderstanding and we will uniformly replace $\tau_{F}$ by $\tau_{uni}$ in the revision. --- ## Weakness 3: the config of Fig. 4. Ans: In Fig. 4, we hope to compare the sign conflicts, L2 distance, and cosine similarity of the merged model weights and individual model weights. To calculate the sign conflicts, we element-wisely compare the merged model weights to each individual model weights and record the ratio of the elements whose signs conflit. We report the average value of the sign conflits between the merged model and each individual model. To calculate the L2 distance (cosine similarity), we first flatten the merged model weights and each individual model weights as 1-dimension vectors. Then we calculate the L2 distance (cosine similarity) between the merged model and each individual model and report the average value. We will add the configuration in the revision. --- ## Weakness 4: interpret the t-SNE. Ans: t-SNE is introduced to visualize high-dimensional data. In our paper, we visualize the outputs of the last transformer block (before the classification head). The results can represent the feature extracting capabilities by the separation among the classes. In Fig. 5, EMR-Merging shows the clearest separation among all the merging methods and is close to individual models. This is correspondent with the experimental results. --- ## Weakness 5: the small distance between figures. Ans: Thanks for your advice and we will increase the distance between figures in the revision. --- ## Weakness 6: how expensive to extend to LLMs? Ans: The proposed EMR-Merging can easily be adapted to various settings including merging LLMs. + In our paper, we show the experimental results when applying EMR-Merging to T0-3B models finetuned by $(IA)^3$, demonstrating the applicability of EMR-Merging to both LLMs and PEFT models. + To further demonstrate the high performance and broad applicability of EMR-Merging, we apply EMR-Merging to a new benchmark, which merges GPT-2 models (fully-finetuned) on 7 NLP tasks and it achieves the SOTA performance. The performance improvement is up to more than 10%. The results are in General Response. --- ## Weakness 7: comparison to WEMoE [R2] Ans: We did not compare EMR-Merging to WEMoE [R2] in our paper because WEMoE had not been officially published by ICML when we submitted the paper. We compare the parameter numbers and prerequisites of WEMoE and EMR-Merging in Table R1 of the Rebuttal PDF. The proposed EMR-Merging needs no data, training, or tuning while WEMoE requires unlabeled test data to train partial parameters from the framework. Meanwhile, EMR-Merging achieves promising performance. The results of merging ViTs on the eight vision tasks are shown in Table R2 of the Rebuttal PDF. EMR-Merging achieves the SOTA performance on the benchmark of merging ViT-L/14 models. Moreover, EMR-Merging can be easily extended to other settings. --- ## References [R1] Language models are super mario [R2] Merging Multi-Task Models via Weight-Ensembling Mixture of Experts [R3] Dropout: a simple way to prevent neural networks from overfitting	Rebuttal 1: Rebuttal: # General Response: We thank all the reviwers for their time and constructive comments. We appreciate the reviewers' praise of the strengths of our paper including: - the idea that decoupling model merging into unified parts and task-specific parts to avoid conflicts (reviewer Lqs2). - extensive expreiments and considerable good performance (reviewer Lqs2, S3RQ). - adequate amount of theoretical and empirical analysis (reviewer b5vt, Qs8J). - tuning-free and data-free properties (reviewer b5vt, S3RQ) ----- In the responses, we show additional experimental results including: 1. Merging GPT-2 models on seven NLP tasks. We merge GPT-2 models finetuned on seven NLP tasks from GLUE benchmark following the settings from FusionBenchmark. EMR-Merging improves the best performance of compared model merging methods by up to 10%. The results are in General Response, Additional Results #1. 2. Merging ViT-B/32 models on nine vision tasks (ImageNet-1K added). We additionally add ImageNet-1K to the original eight tasks. We merge nine CLIP ViT-B/32 models and EMR-Merging shows a much more significant improvement compared to existing merging methods (up to 20%). The results are in General Response, Additional Results #2. 3. Detailed comparison with Sparse Models. We compare the parameter number and performance of EMR-Merging and Sparse Models. In addition, we find that the elected unified task vector of our method can be sparsed to further reduce parameter numbers without affecting performance much. The results are in Table R3, R4, and Figure R1 in the Rebuttal PDF. 4. More ablation studies on the Electing process. We conduct additional ablation studies on the electing process of EMR-Merging. In addition, we find that the elected unified task vector can be quantilized to 4-bit or even 2-bit to further reduce parameter numbers without affecting performance much. The results are in Table R13 in the Rebuttal PDF. 5. Detailed comparison with WEMoE and Representation Surgery. We show detailed comparison of EMR-Merging with WEMoE [R1] and Representation Surgery [R2] in parameter numbers, prerequisites, and performance. EMR-Merging shows the SOTA performance on merging ViT-L/14 models and requires no data or training. The results are in Table R1 and R2 of the Rebuttal PDF. ---- # Some Additional Results ## 1. Merging GPT-2 Models on seven NLP tasks Here we provide the results of EMR-Merging on a new benchmark [R3], merging GPT-2 models on seven NLP tasks from GLUE. All the tasks are evaluate by accuracy. Our EMR-Merging improves the performance of model merging by up to 10%. \| Merging GPT2 Models \| CoLA \| MNLI \| MRPC \| QNLI \| QQP \| RTE \| SST-2 \| Avg \| \| :--------------------- \| :------- \| :------- \| :------- \| :------- \| :------- \| :------- \| :------- \| -------- \| \| Individual \| 76.8 \| 82.1 \| 80.4 \| 88.3 \| 89.6 \| 65.3 \| 91.2 \| 82.0 \| \| Simple Average \| 55.0 \| 55.1 \| 51.0 \| 57.6 \| 76.7 \| 44.8 \| 52.5 \| 56.1 \| \| Fisher Merging \| 54.8 \| 58.0 \| 39.5 \| 63.3 \| 81.5 \| 49.1 \| 64.7 \| 58.7 \| \| RegMean \| 61.7 \| 70.4 \| 65.4 \| 69.7 \| 78.8 \| 56.0 \| 79.7 \| 68.8 \| \| Task Arithmetic \| 68.7 \| 68.6 \| 69.6 \| 70.5 \| 81.8 \| 47.3 \| 83.6 \| 70.0 \| \| Ties-Merging \| 68.4 \| 71.4 \| 68.4 \| 69.6 \| 82.4 \| 47.7 \| 81.8 \| 70.0 \| \| EMR-Merging (Ours) \| 72.8 \| 81.1 \| 79.2 \| 84.8 \| 88.1 \| 66.5 \| 90.3 \| 80.4 \| ## 2. Merging ViTs on nine vision tasks On the basis of merging ViTs on eight vision tasks, we additionally add ImageNet-1K to the these tasks. The finetuned model on ImageNet-1k is released by timm. We merge nine CLIP ViT-B/32 models and EMR-Merging shows a much more significant improvement compared to existing merging methods (up to 20%). \| ViT-B/32 on nine vision tasks \| SUN397 \| Cars \| RESISC45 \| EuroSAT \| SVHN \| GTSRB \| MNIST \| DTD \| IN-1k \| Avg \| \| ----------------------------- \| -------- \| -------- \| -------- \| -------- \| -------- \| -------- \| -------- \| -------- \| -------- \| -------- \| \| Individual \| 75.3 \| 77.7 \| 96.1 \| 99.7 \| 97.5 \| 98.7 \| 99.7 \| 79.4 \| 82.0 \| 89.6 \| \| Averaging \| 61.8 \| 56.4 \| 65.9 \| 66.2 \| 62.7 \| 44.5 \| 81.8 \| 49.0 \| 61.5 \| 61.1 \| \| Task Arithmetic \| 51.8 \| 30.9 \| 55.8 \| 64.3 \| 69.0 \| 42.2 \| 92.7 \| 46.8 \| 66.6 \| 57.8 \| \| Ties-Merging \| 53.3 \| 34.1 \| 57.0 \| 55.8 \| 72.3 \| 43.2 \| 90.5 \| 46.5 \| 68.9 \| 58.0 \| \| EMR-Merging (Ours) \| 77.0 \| 75.2 \| 92.9 \| 92.7 \| 79.7 \| 90.2 \| 97.6 \| 76.2 \| 79.8 \| 84.6 \| The impressive performance of our EMR-Merging on various benchmarks demonstrates its good performance and applicability. ---- # References [R1] Merging Multi-Task Models via Weight-Ensembling Mixture of Experts. ICML 2024. [R2] Representation Surgery for Multi-Task Model Merging. ICML 2024. [R3] FusionBench: A Comprehensive Benchmark of Deep Model Fusion. arXiv:2406.03280. Pdf: /pdf/b7ff0d3d13eb082a4936aa7e770347fef5a6df50.pdf	NeurIPS_2024_submissions_huggingface	2,024	null	null	null	null	null	null	null	null
Fine-grained Analysis of In-context Linear Estimation: Data, Architecture, and Beyond	Accept (poster)	Summary: The paper studies the in-context learning capabilities of linear attention (ATT) and linear state space layers (SSM) on a linear regression task. It shows that for both ATT and SSM there exists a parametrization such that they perform as well as one step of preconditioned gradient descent (PGD) with optimal preconditioning in expectation, and that they cannot perform better. The authors then propose a model for retrieval-augmented generation where in-context examples and the final query are correlated. For PGD, they provide an approximate expression for the optimal preconditioning weights and the related loss and relate it to the optimal weights in the noiseless, i.i.d. case, showing how the correlations between in-context examples and query translate into an increased effective sample size. Next, the authors give an upper bound on the optimal expected loss (for 1-step PGD <=> ATT <=> SSM) in a low-rank adapter setting. Experiments validate the theoretical results. A final experiment shows some initial results hinting that the SSM model studied (H3) performs slightly better than linear attention in an online setting and with a non-stationary task. Strengths: In terms of results, the main original contribution is the analysis of the model with correlated training examples and query points. However, the analysis of the SSM model in the appendix seems non-trivial and should therefore also be considered as a significant element of originality. The section on related work is well-written and covers relevant work. The problem setup is very clean, making it very easy to read and understand the results. All results are clearly stated with the necessary assumptions and accompanied by clear explanations. The questions studied are highly relevant to contemporary machine learning research and applications of AI. Weaknesses: * The main result (12) is not stated as a theorem, and the derivation in the appendix is not completely rigorous. In particular, it is not clear under which conditions the approximation in l.558 is justified. Unless the conditions for the approximations to be valid are stated more clearly, it is difficult to extrapolate from the experimental results to the full claim (A2). * The result on LoRA is not interpreted, and it is not immediately clear what the implications are from the statement of the result. I am not sure if the result implies claim (A3). * The additional experimental results (Fig. 3) seem somewhat preliminary and detached from the rest of the paper. Based on these results, it is not clear that the part of claim (A1) claiming an advantage of H3 is fully supported by the evidence. Technical Quality: 2 Clarity: 3 Questions for Authors: * Under which assumptions are the approximation in (12) valid? What is the nature of the approximations made in the proof of the statement (high probability, asymptotic, ...)? * Can you elaborate on how claim (A3) follow from (14)? * Define $\alpha$-correlated * l.342: wildly studied -> widely studied (?) Confidence: 4 Soundness: 2 Presentation: 3 Contribution: 3 Limitations: The authors address the limitations of their work and acknowledge that their analyses are "not precise and fully formal". Flag For Ethics Review: ['No ethics review needed.'] Rating: 7 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We would like to thank the reviewer for recognizing the relevance of our work to contemporary ML research and AI applications. Indeed, our focus is developing a theoretical understanding of in-context learning (ICL) that is insightful for practical settings. Below, we respond to the concerns raised by the reviewer point by point. W1/W2: We have clarified these in our Common Response to all reviewers. W3: We acknowledge this concern and will better explain Figure 3. Proposition 1 establishes the equivalence of linear attention (ATT) and SSM under Assumptions 1 and 2, namely, $(i)$ all ICL examples within the prompt share the same task vector, $(ii)$ they are conditioned on the task and query vector, and $(iii)$ the model is trained to predict query at a fixed time $n+1$. The point of Fig. 3 is that, when these assumptions are not valid, SSM and ATT are not necessarily equivalent and also that SSM can implement a more general predictor, namely weighted PGD. - In Fig 3(a,b,c) we train ATT and SSM using the average loss where we treat all examples within the prompt as a query. That is, rather than fitting only the last token (index $n+1$), we fit all timesteps where we predict time $i+1$ using examples from $1$ to $i$. This is in line with how LLMs are trained seq2seq in practice as well as ICL experiments of [[Garg et al. NeurIPS’22]](https://proceedings.neurips.cc/paper_files/paper/2022/file/c529dba08a146ea8d6cf715ae8930cbe-Paper-Conference.pdf). Under this setting, Fig 3(a,b) demonstrate that SSM and ATT have different behavior whereas Fig 3(c) demonstrates that SSM achieves a strictly lower risk which is in line with our statement (as weighted PGD is more general than PGD). Here, the intuition is that SSM can intelligently weight different positions in the prompt to better minimize average risk. - Finally, Fig 3(d) showcases another setting where Proposition 1 does not hold: The task vector $\beta$ is not shared and is varying along the prompt (i.e., temporally-drifting distribution). In this setting, SSM outperforms ATT for the same reason. The difference becomes visible as $n$ grows and there is more room for weighing the context window via weighted PGD. We will revise the current text so that these points come across more clearly. Q1: We clarified this in the Common Response: This approximation is valid when $\alpha=\mathcal{O}(1/\sqrt{d}), n/d=\mathcal{O}(1)$ and $d$ is sufficiently large. We will include a full theorem stating the precise risk in the revised manuscript. Q2: (A3) follows from Lemma 1 and Eq. (14). Recall that $\Sigma$ and $\Sigma^{new}$ are the effective covariances of the source and target distributions. Here, effective covariance refers to the product of the task and feature covariances. Lemma 1 studies the landscape of low-rank parameterized attention: Eq. (13) is a strict generalization of the risk formula $\cal{L}_$ of Theorem 1 where the model reduces the risk precisely along the top-$r$ eigendirections. Similarly, Eq. (14) considers the fine-tuning setting and shows that LoRA can improve the risk of an initial model by updating the attention weights along the top-$r$ eigendirections with the maximal gain. Additional discussion is under the Common Response. Q3:* The notion of $\alpha$-correlation is defined in Eq.(11) where $\mathbb{E}[\cos(x_i,x)]=\alpha$. Thus, $\alpha\in[0,1]$ is the Pearson correlation coefficient. We will revise the manuscript ensuring that the term "$\alpha$-correlated" is clearly linked to its definition, especially in the introduction section. Q4: Thanks for identifying the typo error and we have corrected it. We hope that the clarifications provided in our responses have addressed the concerns highlighted by the reviewer. We would be happy to respond to further feedback. --- Rebuttal Comment 1.1: Title: Revised ranking to Accept Comment: Thank you for your response. My concerns have been addressed and I updated my ranking to Accept. --- Reply to Comment 1.1.1: Comment: Many thanks for your encouraging assessment and positive recommendations! Your suggestions have been very valuable, and we will revise and further improve our manuscript based on them.	Summary: The authors examine the capabilities of Transformers with linear attention in performing in-context learning (ICL) by implementing a linear estimator through gradient descent. The existing studies mostly consider IID task and feature vectors and fully parameterized attention weights. This work expands on these studies by analyzing: 1. The landscape of 1-layer linear attention and 1-layer H3 (a state-space model), proving both can implement 1-step preconditioned gradient descent. 2. New risk bounds for retrieval augmented generation (RAG), showing benefits from distributional alignment. 3. The optimal risk for low-rank parameterized attention weights, illustrating how LoRA adapts to new distributions. Strengths: The authors provide a comprehensive theoretical analysis of the landscape for ICL, extending existing results to more practical settings. It offers new insights into the benefits of distributional alignment and the capabilities of low-rank parameterization in attention mechanisms. The theoretical findings are supported by experimental results, enhancing the credibility of the conclusions. Weaknesses: 1. The analysis is limited to single-layer linear attention and linear distribution data, which might not fully capture the complexities of multi-layer architectures. 2. The RAG and LoRA analyses are not precise and fully formal. Technical Quality: 3 Clarity: 3 Questions for Authors: 1. How does the distributional alignment in RAG quantitatively affect the sample complexity of ICL? Are there specific metrics or case studies that illustrate these benefits in practical applications? 2. What are the limitations of LoRA in your cases? I believe LoRA significantly underperforms compared to fully parameterized models. If there are no limitations compared with the fully parameterized models because of the linear distribution data, can you talk about the relationship between the data distribution and LoRA. 3. The paper focuses on single-layer (linear) models. How do the findings extend to multi-layer (linear) architectures? 4. The paper discusses LoRA adaptation for distribution shifts. Are there practical examples or case studies where LoRA has been successfully applied to handle real-world distribution shifts? 5. The study primarily considers linear regression tasks. How would the insights gained from this study apply to more complex tasks, such as natural language processing or computer vision tasks that involve non-linear relationships? Confidence: 4 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: As mentioned in Weakness: 1. The analysis is limited to single-layer linear attention and linear distribution data, which might not fully capture the complexities of multi-layer architectures. 2. The RAG and LoRA analyses are not precise and fully formal. Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We appreciate the reviewer's constructive feedback and the recognition of the credibility of our findings. Below, we address the questions and concerns raised by the reviewer. W1: While our study focuses on single-layer architectures similar to previous work, it nonetheless presents novel and insightful conclusions for various practical settings pertaining to in-context learning (ICL). We have briefly summarized the contributions in the Common Response. Although there exist works investigating the ICL performance using multi-layer models [Ahn et al., Von Oswald et al.], analyzing multi-layer architectures still introduces challenges, particularly when exploring the optimization landscape rather than existence or approximation results. We acknowledge this as an important area for future research. W2: Given that this concern appears to be common, we have derived the exact formula for the RAG setting and provided empirical evidence supporting its validity in the Common Response. We have also included additional discussion regarding the LoRA setting. The agreement between our theoretical analysis and empirical results in the figures (Fig. 1 for RAG in the provided pdf file and Fig. 2(c) for LoRA in the submission) further validate our findings. Q1: The benefit of RAG has been empirically studied in existing literature [Lewis et al., Nakano et al., Izacard et al.] which allows for selecting relevant demonstrations from a collection of instances. In our work, we model it within a linear data setting in Eq. (11) such that query token $x$ is correlated with the in-context features $(x_i)_{i=1}^n$. We derive results showing that $\alpha$-correlated RAG data model achieves a reduction in sample complexity by a factor of $(\alpha^2d+1)$. This implies that highly relevant demonstrations require significantly fewer in-context samples to achieve comparable performance compared to unrelated/fixed demonstrations. Q2: It is acknowledged that LoRA underperforms fully parameterized models, and our results reflect this. As demonstrated in Eq. (14), setting $r=d$ returns the optimal risk when the model is fully parameterized, highlighting a discrepancy when $r\neq d$ or when $\Sigma^{old}\neq\Sigma^{new}$. Fig. 2(c) also empirically validates this by varying the rank of LoRA weights; a rank of 20 corresponds to full parameterization, achieving the lowest risk. Reducing the rank $r$ increases the test risk, as expected. Q3: Multilayer architectures will be more challenging due to their non-convex loss landscape. The most related work is by Ahn et al. who characterizes the critical points of ICL with multi-layer linear attention. Extrapolating from our and their results, we suspect that certain critical points of the $L$-layer SSM would correspond to implementing a $L$-step weighted PGD algorithm. Ahn et al. also makes stringent assumptions on data (besides the IID data model, they also require specific covariance structures as shown in their Table 1). It would be interesting to see if our general data assumptions can be imported to the multilayer models. Q4: LoRA [Hu et al.] is a popular technique proposed to adapt large models to downstream tasks in a parameter-efficient fashion. Typically, downstream tasks have different distributions from pretrained tasks, and traditional methods such as fine-tuning the entire model over the new task are expensive and inefficient. LoRA uses fewer parameters to adapt the model to new tasks without modifying the model weights, making it both memory- and parameter-efficient. In practice, it has been widely applied and shown to realize significant improvements (e.g., Llama3). Our focus is on studying its theoretical aspects for ICL settings. Q5: As noted in our discussion of relevant literature in the submission, Mahankali et al.'s work has demonstrated that even for nonlinear in-context tasks, a linear attention model still implements one step of gradient descent on the linear regression objective. This suggests that for nonlinear tasks, a single-layer linear attention model may still achieve the same optimal loss as optimally PGD, and the challenge lies in finding the optimal preconditioning weight. While we have not yet explored nonlinear settings, we recognize this as an important direction for future research. We hope this response adequately addresses the reviewer's concerns. We are committed to enhancing our submission based on their feedback.	Summary: This paper studies how transformers can use ICL to solve linear regression problems. It is shown that state space models, transformers are both capable of performing linear regression as well as gradient descent (which implements the least squares solution). There are results about LORA and RAG. Strengths: The presentation is clear. Weaknesses: The analysis of vanilla regression is not novel, I think, and has been done for instance in Theorem 1 of Ahn et al (transformers learn to do preconditioned gradient descent) where it is even shown that a pretrained transformer (with GD) learns to do this in a similar setting (which is stronger than showing that it can learn it). Maybe for SSMs this result is new but I am not sure how strong of a contribution this is considering that it only establishes the existence of such a solution. Technical Quality: 3 Clarity: 3 Questions for Authors: What is the significance of the RAG result? It makes sense that giving the learner access to $\beta$ through $X$ directly improves the error. A comparison to some reasonable baseline might be a nice story if it came about that transformers are able to leverage this kind of side information particularly well. What i mean is to compare this with a loss like $\Vert y-X\beta\Vert^2 -\Vert X\beta\Vert^2 + \Vert \beta\Vert^2$ which is aware of this kind of side information. What is the significance of the LoRA result? A decomposition of the optimal error in terms of the singular values of a covariance is shown and then it is observed that if the covariance changes in a specific way (I think the singular vectors have to remain fixed), then LoRA can help. What role does Lemma 1 play? This is not about LoRA, but rather for an entirely low rank model. Why does linear attention perform better for shorter contexts than it was trained on (Figure 3a)? How does Figure 3b show that H3 is better at generalizing to a longer context length if the plot is discontinued at the context it was trained on? I noticed that the performance on varying $\beta$ in Figure 3d better than the iid noiseless case. Can you elaborate on that? Is this for a fixed (that is, rather than varying) $\beta$? Confidence: 3 Soundness: 3 Presentation: 3 Contribution: 1 Limitations: Yes Flag For Ethics Review: ['No ethics review needed.'] Rating: 5 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We thank the reviewer for the detailed comments and questions on our submission. W1: Novelty of contribution and prior art. Ahn et al. analyzes the loss landscape of linear transformers. Their results only apply to special IID data models (see their Table 1) but they also characterize critical points of multilayer attention. Also the more recent [[Zhang et al.]](https://arxiv.org/pdf/2306.09927) studies gradient flow under the IID data model with isotropic task covariance. Similar to Ahn et al., we characterize the loss landscape. While our Theorem 1 may appear similar to them, our work makes multiple novel contributions discussed under the Common Response. To recap, - We provide the first analysis establishing the optimization-theoretic equivalence of single-layer SSM/H3, linear attention, and optimal PGD under suitable data settings. - Comparable works assume IID data whereas our theory allows for correlated data under Assumptions 1 and 2. - Comparable works assume full parameterization ($d\times d$ weights). This is not realistic in practice. Our Lemma 1 characterizes the landscape of low-rank attention for the first time. W2: "The results only show the existence of the solution." As explained above, our study extends beyond showing the existence of a solution and thoroughly investigates the optimization landscape of ICL across various architectures and data settings. Q1: Significance of RAG. In a typical RAG setup, given a query, one retrieves relevant demonstrations to create in-context prompts which often leads to improved performance compared to utilizing query-independent demonstrations for ICL. With the objective of providing a theoretical justification to this phenomenon observed in practice, we consider the data model in Eq. (11), where the query token $x$ and the in-context features $(x_i)_{i=1}^n$ are defined to be relevant with $\alpha$ being the Pearson correlation coefficient among them (motivated by RAG practice). However, it’s important to note that the query and in-context features are all independent of the task vector $\beta$. Thus, contrary to the reviewer’s statement, we are not providing the learner access to $\beta$ through $X$. As a key contribution, under the data model in Eq. (11), we theoretically quantify the improvement in the ICL sample complexity realized via RAG. As shown in Eq. (12), the improvement factor is $(\alpha^2d+1)$. Higher $\alpha$ values lead to greater sample efficiency benefits. Fair comparison baseline: In Eq. (11), we sample $(x_i)_{i=1}^n$ via $\mathcal{N}(\alpha x,(1−\alpha^2)I)$ to ensure that the features and labels for different correlation coefficients $\alpha$ have the same norm, i.e., $E[\|\|x_i\|\|^2]=d$ and $E[y_i^2]=d+\sigma^2$. With this normalization, $\alpha=0$ serves as the baseline, and increasing $\alpha$ benefits in reducing the sample complexity of ICL as validated in Fig 1(b). Finally, under the RAG setting where $X$ and $\beta$ are independent, the loss $\|\|y-X\beta\|\|^2-\|\|X\beta\|\|^2+\|\|\beta\|\|^2$ suggested by the reviewer will yield similar results, subject to some normalization. Q2: Significance of LoRA. We study low-rank attention and LoRA adaptation because these are typically what is used in real applications. Lemma 1 shows that attention with low-rank parameterization learns to recover the $r$-dimensional data-task eigenspace corresponding to the top eigenvalues to achieve optimal loss. Based on this insight from Lemma 1, we derive the LoRA result, by considering the distance between the old and new covariance eigen-spectrums. Also see Common Response. Q3: Questions on Fig.3. > - Why linear attention performs better for shorter contexts. In Fig. 3(a)-(c), we train the model to minimize the average risk over all in-context instances, not just the last/query token as in other results. According to Theorem 1, the optimal weight $W_\star$ varies with context length $n$, so the model must learn to optimize $W$ across different lengths. Thus, a model trained on shorter in-context windows can outperform the one trained on longer windows when tested on shorter contexts. > - How Fig. 3(b) shows that H3 is better for longer context... Inspired by this question, we tested both linear attention and SSM models (from Fig. 3 (a) and (b)) over unseen context lengths up to $100$. The results are provided in Fig. 2 of the provided pdf file in the Common Response. These results clearly show that, compared to linear attention, SSM generalizes much better to unseen context lengths, supporting our claim. Additionally, these results indicate that while models trained on shorter in-context lengths perform better over shorter ranges, their performance does not compare as well over longer ranges. We will include these new experiments. > - ...performance on varying $\beta$ in Fig. 3(d) better than the IID case… Nice question! The reviewer is correct in suggesting an IID baseline. We have identified that the experimental setup led to varying expected norms of the label $E[y_i^2]$ due to the way $\beta_i$ is defined in the experiment section. To recap, given two random vectors $\beta_1$ and $\beta_2$ that follow the standard Gaussian distribution, the task vector at the $i$th position is defined by $\beta_i=\alpha_i\beta_1+(1-\alpha_i)\beta_2$. Then the expected norm of the label $E[y_i^2]=\alpha_i^2E[\|\|x^\top\beta_1\|\|]+(1-\alpha_i)^2E[\|\|x^\top\beta_2\|\|^2]=(\alpha_i^2+(1-\alpha_i)^2)d$ varies with $i$. Based on this finding, we have updated the black curve of Fig.3(d) in the paper and the revised figure is shown in the Fig.3 in the provided pdf file. The updated curve now provides a reasonable IID baseline. We appreciate the reviewer's helpful comment and will revise the paper accordingly. We thank the reviewer for the detailed comments. Although the reviewer noted a lack of novelty and recommended rejection, we hope our responses highlight the novelty and clarify any misunderstandings. --- Rebuttal Comment 1.1: Comment: I had a misunderstanding about the RAG section. Please disregard what I said. I thought $x$ was sampled as $\mathcal(\alpha \beta, (1-\alpha)I)$. Because of the clarifications on the experiments, and the contributions relating to SSMs (equivalences to PGD), I change my score from 3 to 5. However, I still feel the following about the Transformer results: 1. The LoRA discussion holds only when the new and old $\Sigma$ are jointly diagonalizable and this makes the message a little weak, but it is possible that this is truly the only benefit from LoRA on a single layer. 2. Using the present model to study RAG still seems somewhat simplistic to me. These are difficult to fix, since this is a critique of the model itself. The results themselves seem reasonable to me. --- Reply to Comment 1.1.1: Comment: We thank the reviewer for reassessing our work and for raising their rating. We would like to provide further clarifications on their comments: 1. For low-rank parameterization, our main technical result is Lemma 1. This result does not rely on the assumption of diagonalization and establishes a natural generalization of fully-parameterized results (Theorem 1) to the low-rank setting. In LoRA analysis, we assume diagonalization mostly to be able to provide closed form interpretable bounds on its impact in terms of eigenspace. We anticipate one can establish similar upper bounds on the impact of LoRA in terms of the angle between the eigenspaces of old and new covariance matrices. Finally, it worths noting that the analysis of general covariance is challenging even for vanilla ICL (fully parameterized) as both Ahn et al. and Mahankali et al. make assumptions about the covariance of $\beta$ and/or $x$ while we allow for arbitrary $x$, $\beta$ covariances in our main results. 2. Our RAG model is inspired from real RAG systems [[Karpukhin et al.](https://arxiv.org/pdf/2004.04906), [Lewis et al.](https://arxiv.org/pdf/2005.11401)], which utilize cosine similarity of the document (feature) embeddings. However, we agree with the reviewer that exploring more complex RAG models, either with general covariance models or allowing for nonlinear feature/document representations learned via a separate retriever model, would be valuable. This is certainly an exciting direction for future work and we appreciate their suggestion.	null	null	Rebuttal 1: Rebuttal: ## Common response to all the reviewers We thank the reviewers for their constructive comments and insightful questions. We are glad that Reviewer HGXo acknowledges the credibility of our conclusions and Reviewer EE6j notes the high relevance of our study to contemporary machine learning research and applications of AI. Here, we would like to restate our key contributions and respond to other shared concerns raised by the reviewers. Our main contributions are as follows: 1. We establish the theoretical and empirical equivalence among optimizing single-layer linear attention, single-layer SSM, and implementing optimally-preconditioned gradient descent (PGD). While previous works (e.g., Ahn et al.) have noted the equivalence between linear attention and PGD, to the best of our knowledge, our work is the first to elucidate the equivalence between SSM and PGD. 2. Our Proposition 1 extends the equivalence among attention, SSM, and PGD to more general and realistic settings, subsuming but going beyond the independent data scenario (as in Ahn et al. and our Theorem 1). Two key contributions are: - Our contribution on SSM is entirely novel and relies on establishing an optimization-theoretic equivalence between gating (within the SSM) and linear attention. - By considering dependent data (e.g., in RAG) and low-rank parameterizations (e.g., for LoRA adaption) — factors not assumed or analyzed in previous studies — we enhance the understanding of model behavior under more complex yet highly practical settings. 3. The alignments between theoretical predictions and empirical results demonstrate the accuracy and value of our theoretical insights. To proceed, we address the shared concerns by the reviewers: - The exact formula of RAG analysis: We recognize that the analysis of the RAG setting in our submission is not fully precise due to the complexity involved in the high-order (up to 6-order) moments of $x$, $X$, and $\beta$. To address this main concern, we have recalculated the exact formulations for the RAG data setting. In particular, the final solution takes the following exact form: $$W_\star=cI\quad\text{and}\quad {\mathcal{L}_\star}=d+\sigma^2-cnd(\alpha^2(d+1)+1)$$ where $$c=\frac{\alpha^2(d+1)+1}{\alpha^4n(d+2)(d+4)+\alpha^2(1-\alpha^2)(d+2)(d+2n+3)+(1-\alpha^2)^2(d+n+1)+\sigma^2(\alpha^2(d+1)+1)}.$$ Here $\alpha=0$ corresponds to iid setting. Note that Eq. (12) in our submission provided an approximate solution assuming that $\alpha=\mathcal{O}(1/\sqrt{d})$, $d/n=\mathcal{O}(1)$ and large enough $d$. Additionally, we have also updated the RAG figure (Fig.1(b)) based on the exact formula provided above and the results are shown in Fig.1 of the provided pdf file. The new theoretical predictions now perfectly align with empirical observations. We will incorporate these updates in the final version of the paper. - The interpretation of LoRA results: Our LoRA results in Eq. (14) show that when there is distribution shift and the joint diagonalizability assumption holds, LoRA leads to the model adapting the initial preconditioning weights to a target distribution over the principal $r$-dimensional eigenspace with the maximal gain. Though the analysis presents only an upper bound due to the complexity of arbitrary distribution shifts (e.g., arbitrary $\Sigma^{old}$ and $\Sigma^{new}$ matrices), it marks the first optimization-theoretic exploration of LoRA in ICL. Our empirical results in Figure 2(c) validate the tightness of the prediction of Eq. (14). Notably, Lemma 1 on low-rank attention provides a stronger theoretical guarantee on the landscape, namely, Eq. (13). This can be viewed as a special case of the LoRA setting where the weights of the initial model are set to zero. We will enhance the discussion of our LoRA results. We believe that extending our analysis of LoRA to broader settings is an interesting avenue for future work. Pdf: /pdf/7fc85095c7659be9f8583aa4c72c7601a2bbe3c0.pdf	NeurIPS_2024_submissions_huggingface	2,024	null	null	null	null	null	null	null	null
Full-Distance Evasion of Pedestrian Detectors in the Physical World	Accept (poster)	Summary: This paper is dedicated to improving the field of physical adversarial attacks on pedestrian detection, focusing on the robustness of attack performance over distance. To bridge the appearance gap caused by distance between the digital and physical spaces, the authors propose a distant image converter (DIC). To address the inconsistency of features required for short and long distances, the authors propose multi-frequency optimization (MFO). The authors tested the attack performance in the physical world at distances ranging from 4 to 40 meters. Experiments show that the proposed DIC and MFO can improve the attack success rate in scenarios with varying distances. Strengths: 1. The authors focused on a key factor affecting the performance of physical adversarial attacks: distance. This is an important exploration. 2. The authors conducted ablation experiments and comprehensive evaluations of the DIC and MFO modules, making the method's design convincing. 3. The DIC designed by the authors considers atmospheric conditions, camera effects, and effect filters, which is a reasonable and thorough modeling approach. 4. The method proposed in this paper has been validated through extensive experiments in the physical world, and the authors provided a quantitative evaluation by collecting data from the physical world. Weaknesses: 1. From the videos provided in the supplementary materials of this paper, the performance of the attack does not seem to be as good as claimed in the paper. For example, in 8m.mp4 and 14m.mp4, in most frames, the model can detect pedestrian instances with adversarial patches, indicating that the attack failed. In the quantitative evaluation of the paper, the ASR at a distance of 14 meters reached over 50%. This inconsistency undermines the effectiveness of the proposed method. 2. This paper lacks a comparison with other popular methods, such as AdvPatch [1], Adv-Tshirt [2], NAP [3], and T-SEA [4]. 3. The DIC proposed in this paper is physics-based modeling. Such modeling generally involves non-differentiable operations. The authors used SGD during the training of DIC, and gradient propagation is also required when updating the patch. How did the authors address this issue? 4. From the code provided in the supplementary materials, it can be seen that the method proposed in this paper appears to be related to YOLOv2. However, YOLOv2 is not mentioned in the paper. Please explain this issue. 5. Unclear description: It's not clear what "D" refers to on line 205. 6. Some writing issues: on line 32, "In DIC, We find...". Some abbreviations are not explained the first time they appear, such as "DNN" on line 18 and "FDA" on line 39. 7. Lack of important references, such as [4, 5]. [1] Thys S, Van Ranst W, Goedemé T. Fooling automated surveillance cameras: adversarial patches to attack person detection[C]//Proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops. 2019: 0-0. [2] Xu K, Zhang G, Liu S, et al. Adversarial t-shirt! evading person detectors in a physical world[C]//Computer Vision–ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part V 16. Springer International Publishing, 2020: 665-681. [3] Hu Y C T, Kung B H, Tan D S, et al. Naturalistic physical adversarial patch for object detectors[C]//Proceedings of the IEEE/CVF International Conference on Computer Vision. 2021: 7848-7857. [4] Huang H, Chen Z, Chen H, et al. T-sea: Transfer-based self-ensemble attack on object detection[C]//Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. 2023: 20514-20523. [5] Wei H, Tang H, Jia X, et al. Physical adversarial attack meets computer vision: A decade survey[J]. arXiv preprint arXiv:2209.15179, 2022. Technical Quality: 3 Clarity: 3 Questions for Authors: Please address the weakness raised in Weaknesses. Confidence: 5 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: The authors discussed limitations and societal impact. Flag For Ethics Review: ['Ethics review needed: Safety and security'] Rating: 5 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Question: From the videos provided in the supplementary materials of this paper, the performance of the attack does not seem to be as good as claimed in the paper at 8m and 14m. Answer: We have extracted all frames in the 8-meter and 14-meter demonstration videos and evaluated the ASRs by counting. At 8 meters, the attack was successful in 54 out of 132 frames, resulting in an ASR of 41\%. At 14 meters, the attack succeeded in 111 out of 171 frames, yielding an ASR of 65\%. Both results were comparable to the results reported in Figure 8(a). Please note that because of the persistence of vision, the bounding boxes would linger in our perception longer than they do in the videos. This phenomenon may explain why the subject holding the patch seemed to be detected in most frames of the 14-meter demonstration video. Question: This paper lacks a comparison with other popular methods, such as AdvPatch, Adv-Tshirt, NAP, and T-SEA. Answer: Please note that in Figure 8(a), we conducted a head-to-head comparison between FDA and Adv-Tshirt. In Figure 9(a), we compared FDA with TCA [r1], which is also a commonly used attack method for full-angle clothing attack. According to your suggestion, we obtained the AdvPatch, NAP, and T-SEA patterns targeting YOLOV5 by either utilizing the patterns provided in the respective papers or executing the codes provided. These patterns were evaluated following the physical-world patch attack setting described in Section 5.2, where the results were averaged over three different pedestrians. The AdvPatch, NAP, and T-SEA patterns achieved average ASRs of 19\%, 19\% and 42\%, respectively, all of which were lower than the 75\% average ASR achieved by the FDA pattern. We will include these results in the final paper. Question: The DIC proposed in this paper is physics-based modeling. Such modeling generally involves non-differentiable operations. How did the authors propagate gradient through the DIC? Answer: In our work, we implemented all three modules in DIC with differentiable computations. The atmospheric perspective module is a differentiable function (Equation 1, line 119). The camera simulation module is constructed using two convolutional layers (Equation 6, line 170). The style filter simulation module comprises a sequence of differentiable functions that simulate various style filters (Equation 7, line 175). If there are any details that require clarification, please feel free to discuss them with us during the discussion period. Question: Please explain the YOLOV2 related codes in the supplementary material. Answer: We built our code based on the TCA [r1] code which treated the YOLOV2 detector as the main target model. In our experiment, we forgot to remove some YOLOV2-related codes imported by the TCA authors and leveraged some general utility functions they provided. Please note that since YOLOV2 is now commonly considered outdated, we have not performed any experiments on the model. A tidier version of the code will be included in the final paper. Question: Unclear description: It's not clear what "D" refers to on line 205. Answer: D refers to the number of distances optimized. We will clarify it in the final version. Question: Some writing issues: on line 32, "In DIC, We find...". Some abbreviations are not explained the first time they appear, such as "DNN" on line 18 and "FDA" on line 39. Answer: Thank you for your careful reading. We will correct them in the final version. Question: Lack of important references, such as [4, 5]. Answer: Thank you for highlighting the two interesting pieces of research. We will discuss our work in respect to the two papers in the final version. [r1] Hu, Z., Huang, S., Zhu, X., Sun, F., Zhang, B., & Hu, X. (2022). Adversarial Texture for Fooling Person Detectors in the Physical World. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (pp. 13307-13316). --- Rebuttal Comment 1.1: Comment: Dear reviewer ouf1, have our answers in our rebuttal properly addressed your concerns? If there are any follow-up questions, please feel free to discuss them with us during this discussion period. --- Rebuttal Comment 1.2: Comment: Thank you for your response. My concerns have been addressed. I will maintain my current rating in support of accepting this paper. --- Rebuttal 2: Comment: Dear Reviewer, The authors have provided a rebuttal. Can you please provide your feedback after reading the rebuttal? The deadline is approaching fast. Thanks, AC	Summary: To achieve full-distance attacks, the authors summarize three factors that distort the performance including atmospheric, camera hardware, and effect filters. Then, the authors simulate these factors in the digital world via DIC. To overcome the conflict of different distances requiring different low-frequency patterns, the authors further propose a Multi-Frequency Optimization (MFO) technique. Strengths: 1. The summarized factors for full-distance attack are precise and important, and the DIC simulation for these factors conducted by the authors looks reasonable. 2. The victim detectors used in this paper are advanced, which better illustrates the effectiveness of the proposed method. Weaknesses: Weakness: 1. The best-performing approach in refer [1] is TC-EGA, while TCA is the suboptimal one. Why does this paper use TCA as a comparison method instead of TC-EGA? Additionally, why does this paper not adopt the TC-EGA approach to generate Expandable FDA, which is supposed to yield better results? 2. In Fig. 9, I notice that the FDA pattern looks like many abstract human heads stacked together, and the detector might also perceive it as a combination of multiple small human heads. Considering that the IoU threshold for calculating ASR in this paper is only IoU > 0.5, these small human heads do not meet this condition and are therefore directly filtered out. However, these heads can indeed impact the attack's effectiveness. Namely, if the heads are detected, the attack should also be considered a failure. Therefore, using a single IoU threshold to measure ASR may not accurately reflect the attack's effectiveness. I hope the authors can address this issue, for instance, by providing ASR results at different IoU thresholds. 3. The pattern of FDA appears very different with and without TC. Does this imply that FDA has a considerable number of different local optima during optimization? 4. Compared to ASR, mASR (mean ASR with different confidence scores) [1] can more comprehensively reflect the attack's effectiveness. 5. Providing experimental comparison results in the digital world can better illustrate the simulation effectiveness of the proposed method during training. [1] Hu Z, Huang S, Zhu X, et al. Adversarial texture for fooling person detectors in the physical world[C]//CVPR 2022. Technical Quality: 3 Clarity: 3 Questions for Authors: Please refer to the weakness. Confidence: 5 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: The authors didn't provide the limitations. But they have explained their reason in the checklist, which sounds acceptable. Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Question: Why does this paper not adopt the TC-EGA approach as a baseline and generate an Expandable FDA? Answer: Unlike the adversarial texture paper [1] which used the YOLOV2 and V3 models as the target models, we used YOLOV5 as our main target model. Table S3 of the NAPGuard Appendix [4] shows that TC-EGA performed poorly against YOLOV5, as it was only able to reduce the model’s AP from 100\% to 65\%. We confirmed this in our own experiments, making TC-EGA unsuitable as our baseline. In contrast, TCA obtained a good AP of 36\% on YOLOV5 and it also supports full-angle clothing attacks, so we used TCA as our baseline method. Question: The FDA pattern looks like many abstract human heads stacked together, which might cause small detection boxes to show up. So, using a single IoU threshold to measure ASR may not accurately reflect the attack's effectiveness. Answer: First of all, please note that it is a convention to use an IoU threshold of around 0.5 in physical attack evaluations [1,5], since in the literature of detectors, if a predicted bounding box has a small IoU relative to the ground truth, it is conventionally considered a false positive. We had to follow this convention to reproduce existing results and perform comparisons with the baselines. Second, please note that with the TC (toroidal cropping) design, depending on the target model used, it is common for abstract human-like patterns to show up on adversarial patterns. For example, in Figure S1 (d), (f), (g) and (h) of the TCA paper’s Appendix [1], the adversarial pattern targeting YOLO, FasterRCNN and MaskRCNN are formed by abstract human heads and trunks stacked together. In contrast, when we changed the target model to be Deformable DETR or RetinaNet with PVT backbone, no human-like patterns showed up (Figs. 1 and 2 of the rebuttal PDF uploaded). So, we found that when we used YOLOV5 as the target model, due to the presence of abstract human heads, as the IoU threshold decreases from 0.5 to 0.4, 0.3, 0.2, 0.1 and 0.0, more and more small detection boxes corresponding to the abstract human heads were no longer filtered out, causing the average ASRs to decrease from 73\% to 65\%, 40\%, 39\% and 38\% respectively (the average ASR of normal clothing was 0\% regardless of the IoU threshold used). Please note that even at the most extreme case of using an IoU threshold of 0, where the attack would be considered a failure if any small bounding boxes showed up on the subject due to the abstract human heads, the FDA pattern still obtained an average ASR of 38\%. Additionally, on detectors where abstract human heads did not appear in the optimized adversarial pattern, the average ASR did not drop significantly as IoU threshold decreased. That is, when treating Deformable DETR as the target model, at IoU thresholds of 0.5, 0.4, 0.3, 0.2, 0.1 and 0.0, the average ASRs were 71\%, 70\%, 68\%, 68\% and 68\% respectively. Similarly, when treating RetinaNet with PVT backbone as the target model, the average ASRs were 73\%, 73\%, 72\%,70\% and 67\% respectively. Question: The pattern of FDA appears very different with and without TC. Does this imply that FDA has a considerable number of different local optima during optimization? Answer: Please note that with and without TC, the feasible region of optimization is different (TC imposes a constraint on the feasible region), so it is quite normal to obtain different adversarial patterns in the two situations. It implies that multiple points in the feasible region of the unconstrained FDA optimization problem are effective in attacking the detectors. So, your hypothesis might be correct, but it requires additional analysis to confirm. Question: Compared to ASR, mASR can more comprehensively reflect the attack's effectiveness. Answer: Following the TCA paper [1], we calculated the mean average ASRs across confidence thresholds of 0.1, 0.2, …, 0.9. The mean average ASR of the FDA clothing and TCA clothing was 78\% and 43\%, respectively. Question: Providing experimental comparison results in the digital world can better illustrate the simulation effectiveness of the proposed method. Answer: For the result in Figure 8(a), the corresponding digital-world ASRs of the FDA pattern at different distances (4m, 8m, 14m, 20m, 26m, 34m and 40m) were 79\%, 53\%, 76\%, 88\%, 66\%, 64\%, 88\%, which resulted in an average ASR of 73\%. The digital-world ASRs of the Adv-Tshirt pattern at different distances were 83\%, 66\%, 17\%, 8\%, 3\%, 8\%, 13\%, which resulted in an average ASR of 28\%. We observed that the physical-world attack results were consistent with the digital-world attack results. We will include this result in the final paper. [1] Hu, Z., Huang, S., Zhu, X., Sun, F., Zhang, B., & Hu, X. (2022). Adversarial Texture for Fooling Person Detectors in the Physical World. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (pp. 13307-13316). [2] Wang, W., Xie, E., Li, X., Fan, D.P., Song, K., Liang, D., Lu, T., Luo, P., & Shao, L. (2021). Pyramid Vision Transformer: A Versatile Backbone for Dense Prediction Without Convolutions. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) (pp. 568-578). [3] Xizhou Zhu, Weĳie Su, Lewei Lu, Bin Li, Xiaogang Wang, & Jifeng Dai (2021). Deformable \DETR\: Deformable Transformers for End-to-End Object Detection. In International Conference on Learning Representations. [4] Wu, S., Wang, J., Zhao, J., Wang, Y., & Liu, X. (2024). NAPGuard: Towards Detecting Naturalistic Adversarial Patches. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (pp. 24367-24376). [5] Zhu, X., Hu, Z., Huang, S., Li, J., & Hu, X. (2022). Infrared Invisible Clothing: Hiding From Infrared Detectors at Multiple Angles in Real World. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) . --- Rebuttal Comment 1.1: Comment: Thanks for the response. My Weaknesses 1, 3, 4, are adequately addressed by the authors. For Weakness 5, I think AP may be a more precise and commonly-used metric in the digital world, despite the results in ASR are also acceptable. I still have questions about W2. - Firstly, I think FDA is much more similar to human heads compared with TCA, considering FDA even looks like the human head has two eyes. - Secondly, I wonder how could the ASR remain 38% when iou=0, does that mean the network sometimes may not consider them as human heads like we perceive? - Thirdly, in real-world scenarios, IoU cannot be calculated since we don't have the ground truth. The detector only cares about if bbox of the human class exists. Therefore, I think poor performance in low IoU would essentially undermine the effectiveness of FDA in real-world applications. --- Reply to Comment 1.1.1: Comment: We sincerely thank you for your effort in answering our rebuttal. We will include the corresponding digital world APs at different distances in the final version. Regarding the three new questions, our answers are as follows. Question: I think FDA is much more similar to human heads compared with TCA, considering FDA even looks like the human head has two eyes. Answer: If you zoom in on Figure S1 (g) and (h) of TCA paper’s appendix [1], similar abstract human head patterns with two eyes can also be identified. Question: I wonder how could the ASR remain 38% when iou=0, does that mean the network sometimes may not consider them as human heads like we perceive? Answer: Yes, when IoU was 0, an ASR of 38\% remained, since the network sometimes may not consider them as human heads. Please note that those human-head-like patterns are abstract. Question: Thirdly, in real-world scenarios, IoU cannot be calculated since we don't have the ground truth. The detector only cares about if bbox of the human class exists. Therefore, I think poor performance in low IoU would essentially undermine the effectiveness of FDA in real-world applications. Answer: Thank you for stressing the importance of using a low IoU threshold. Although we used an IoU threshold of around 0.5 following the existing works, but from our discussions, we agree that a low IoU threshold should be used to better demonstrate the performance of attack. In the final version, we will include results at IoU thresholds of 0 to 0.5. Regarding the question, we wish to bring two points to your attention. (1) Although the FDA pattern targeting YOLOV5 had a lowered ASR of 38\% when the IoU threshold decreased to 0, but on Deformable DETR [3] and RetinaNet with PVT backbone [2], the average ASRs of the FDA patterns were 68\% and 67\%, which we believe is strong evidence that when the “abstract human heads” are not present in the adversarial pattern, FDA attack can be very effective even on this harder metric. We will include the corresponding results in the final version. (2) When the target model was YOLOV5 and IoU threshold was 0, the average ASRs of FDA clothing, TCA clothing, random clothing and normal clothing were 38\%, 8\%, 2\% and 0\% respectively. Considering the goal of our work is to find a method that boosts full distance attack performance of existing methods under identical setting, we find our method to have fulfilled this goal. --- Rebuttal 2: Comment: Thanks for the in-depth analysis in your response. I wish the analysis from the rebuttal could be properly incorporated into the camera-ready version for better comprehension. Despite the human-heads problems in this paper are still severe from my perspective, I do agree that it is a universal problem in the community that needs to be researched in the future. I will raise my final rating. --- Rebuttal Comment 2.1: Comment: Thank you for your effort in reviewing our paper and providing us with feedback. We will try our best to properly incorporate our analysis from the rebuttal into the final version of our paper.	Summary: The presents an adversarial attack that works in real world at different distances and fools object detectors. The method utilizes several advanced techniques like atmospheric perspective, camera and filter simulations, multi-frequency optimization to bring the method closer to real-world scenarios. The method is evaluated to fool YOLO-V5, Mask RCNN, and Deformable DETR. Strengths: 1) Easy-to-follow writing with vivid illustrative figures and clear background and all text. It really helps to grasp the ideas. 2) The change of the image from different distances is indeed a problem for fooling deep detectors. This paper leverages principles of physics and camera hardware design to transform patch-based digital attacks, that do not usually transform to physical world, to real-world scenarios. 3) The method is accessed across different kinds of detectors and it seems it works at different distances most of the time. Evenmore, the crafted clothes fool across networks, making them black-box. Several ablations studies in the appendix are also interesting. Weaknesses: 1) The adversarial patches/clothes are really massive. 2) At some distances (5-10m) adv tshirt is better than the proposed FDA Technical Quality: 3 Clarity: 3 Questions for Authors: 1) Did you consider variation not across distance, but across different angle of views, or other variations? If a person wearing the proposed adversarial clothes is still going to be adversarial if they change their pose, angle towards camera, etc? 2) How are you going to propagate gradients for the detectors that use some non-differentiable operations like in MTCNN? 3) How long does it take to optimize an attack? Confidence: 3 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: Authors mention some limitations in the NeurIPS checklist and potential social impact is discussed in the Appendix L. Flag For Ethics Review: ['No ethics review needed.'] Rating: 5 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Question: The adversarial patches/clothes are really massive. Answer: We used relatively large patches and full-body clothing in our experiments for two reasons. First, it is now a common practice to employ large adversarial patterns such as full-body clothing [5][6] or fully-covered car paint [3][4], to conduct adversarial attacks. Second, considering the goal of this study is to enhance attack performance under a wide range of distances, when the subject is far away, due to the scale reduction, the blurred small patch or small piece of cloth cannot contain a sufficiently complex pattern for performing an adversarial attack. As demonstrated in Figures 8(a) and 9(a), when printed with random patterns, at the current sizes, the patches and clothing had negligible attack performance. Additionally, at the current patch and clothing sizes, the baseline methods had limited full distance attack performance. This indicates that the current patch and clothing sizes do not contribute much to our full distance attack performance. Question: At some distances (5-10m) adv tshirt is better than the proposed FDA. Answer: Yes, you are right. Please note the goal of this study is to maximize the average ASR across different distances, so we find it acceptable for the proposed method to not perform perfectly at a small number of distances. But your comment reminds us that there is space for further improvement, and we’ll explore it in the future. Question: Is the FDA clothing adversarial under different angles and poses? Answer: Yes, the FDA clothing is adversarial under different angles and poses. Regarding different angles, other than evaluating the clothing from the front and back view, we have also evaluated the performances of FDA, TCA and random clothing from the side view. The average ASRs were 61%, 46% and 0% respectively (lines 336-337 of the paper). Regarding different poses, when performing the clothing experiments, at various distances, we asked the subjects to walk with their arms swinging to collect the testing images. Therefore, the results in Figure 9 (a) were already averaged over different poses. Question: How are you going to propagate gradients for the detectors that use some non-differentiable operations like in MTCNN? Answer: To the best of our knowledge, over the last 5 years, most mainstream detectors are differentiable, which is why we performed FDA by gradient backpropagation. But we found that there are several potential ways to perform FDA if there are any non-differentiable operations in the target model. Firstly, by using the ensemble attack method [1], FDA patterns that generalize well across detectors can be generated, enabling attacks on models with non-differentiable operations. Additionally, there are works that estimate adversarial attack gradients (BPDA) [2], generate attack gradients with a surrogate network [3], perform attacks using search algorithms [4], or propose a general adaptive attack method [7] designed accounting for the presence of non-differentiable operations. These approaches can potentially be leveraged to craft FDA patterns for models with non-differentiable operations. Question: How long does it take to optimize an attack? Answer: It usually takes about 16 hours on a single NVIDIA 3090 GPU to optimize an FDA pattern. With the same set of code, the AdvTshirt and TCA patterns also required a comparable amount of time to properly optimize. We note that the optimization speed was not our main concern in this work and it could certainly be improved. [1] Yanpei Liu, Xinyun Chen, Chang Liu, & Dawn Song (2017). Delving into Transferable Adversarial Examples and Black-box Attacks. In International Conference on Learning Representations. [2] Anish Athalye, Nicholas Carlini, & David A. Wagner (2018). Obfuscated Gradients Give a False Sense of Security: Circumventing Defenses to Adversarial Examples. In International Conference on Machine Learning. [3] Yang Zhang, Hassan Foroosh, Philip David, & Boqing Gong (2019). CAMOU: Learning Physical Vehicle Camouflages to Adversarially Attack Detectors in the Wild. In International Conference on Learning Representations. [4] Tong Wu, Xuefei Ning, Wenshuo Li, Ranran Huang, Huazhong Yang, & Yu Wang (2020). Physical Adversarial Attack on Vehicle Detector in the Carla Simulator. ArXiv, abs/2007.16118. [5] Hu, Z., Chu, W., Zhu, X., Zhang, H., Zhang, B., & Hu, X. (2023). Physically Realizable Natural-Looking Clothing Textures Evade Person Detectors via 3D Modeling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (pp. 16975-16984). [6] Jing Li, Sisi Zhang, Xingbin Wang, and Rui Hou. 2023. Mimic Octopus Attack: Dynamic Camouflage Adversarial Examples Using Mimetic Feature for 3D Humans. In Information Security and Cryptology: 18th International Conference, Inscrypt 2022, Beijing, China, December 11–13, 2022, Revised Selected Papers. Springer-Verlag, Berlin, Heidelberg, 429–444. https://doi.org/10.1007/978-3-031-26553-2_23 [7] Yao, C., Bielik, P., Tsankov, P., & Vechev, M. (2021). Automated Discovery of Adaptive Attacks on Adversarial Defenses. In Advances in Neural Information Processing Systems (pp. 26858–26870). Curran Associates, Inc.. --- Rebuttal Comment 1.1: Comment: Thanks for your answers. After reading other reviews and replies, I am keeping my score 5 with leaning to accept the paper, but it can go either way. --- Reply to Comment 1.1.1: Comment: Thank you for your effort in reviewing our paper and providing us with feedback.	null	null	Rebuttal 1: Rebuttal: We sincerely appreciate the time and effort the reviewers have put into this manuscript, which will help us improve the quality of the revised paper. It is encouraging that the reviewers find our work to be “important” [vr4j, ouf1], our design to be “reasonable” and “convincing” [vr4j, ouf1], our presentation to be “vivid” and “clear” [e7ej]. We addressed the concerns in each reviewer’s individual rebuttal area and the paper will be updated accordingly. We hope you are satisfied with our response. We welcome any further comments and we would be happy to answer them in the discussion period. Please note that the PDF file uploaded is to provide supporting figures in response to a question from reviewer vr4j. Pdf: /pdf/9ffda382550329e8bee6852f923c3a656180e641.pdf	NeurIPS_2024_submissions_huggingface	2,024	null	null	null	null	null	null	null	null
From Chaos to Clarity: 3DGS in the Dark	Accept (poster)	Summary: This paper introduces a new framework for 3D reconstruction and denoising from raw images. Specifically, a noise extractor and a noise-robust reconstruction loss are proposed to deal with the overfitting issue of 3DGS to the noises heavily distributed in raw images. Experiments on the rawnerf dataset and ablation studies are provided to demonstrate the effectiveness of the proposed methods. Strengths: 1. Intuitive noise decomposition and informative noise removal (or estimation) method 2. Better reconstruction and denoising performance compared to 3DGS 3. Good writing and presentation Weaknesses: 1. The innovation may not be enough for this conference. The major contribution of this paper is the denoising part, where the noise estimator needs pre-training for good performance (extra data is required). 2. The performance improvement over baseline with rawnerf loss is marginal and even no synthesis quality improvement compared to the reported scores in the original RawNeRF paper (which is based on NeRF). Technical Quality: 3 Clarity: 4 Questions for Authors: Please refer to the weakness part. Confidence: 4 Soundness: 3 Presentation: 4 Contribution: 2 Limitations: The limitations mentioned in the paper are not easy to address, so it did not affect my rating for this paper. Flag For Ethics Review: ['No ethics review needed.'] Rating: 5 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We appreciate your valuable suggestions and comments. Here are our detailed responses to your feedback: 1. Novelty Limited Due to the Need for Noise-Clean Paired Data: We want to clarify that the pretrained model aims to provide a good initialization and does not have strict restrictions on the data used for training (e.g., we can use data from different devices or noise-only data). Specifically, our pretrained dataset, SID (captured by a DSLR camera), differs significantly from the RawNeRF dataset (captured by an iPhone X) in terms of color space and bit depth. To further verify the impact, we also experimented with a self-supervised pretraining method like Neighbor2Neighbor with noise-only images in SID. The results, shown in Table R1, indicate no significant difference, supporting that noise-clean paired images are not strictly necessary during the pretraining stage. 2. Comparison with Baselines and RawNeRF: We want to clarify that in the few-shot settings (which are more common in real-world scenarios like 3D reconstruction using multi-camera setups or for autonomous driving) on the RawNeRF dataset, we achieved a 3-4 dB gain compared to the rawnerf loss baselines as presented in Figures 5 and 6. In full-view settings, we also observed around a 0.9 dB gain compared to 3DGS with the rawnerf loss baseline. While our method's reconstruction quality is similar to RawNeRF, we significantly reduced the training time by 100x and boosted the rendering speed by 5000x, as shown in Table R2, making it more feasible for real-world applications. Once again, we sincerely thank you for your constructive feedback, which has been crucial in refining our work. * Tables: Table R1: Ablation Study of F Pretraining on RawNeRF Datasets (Full Training Views) \| Method \| F Pretrained Dataset \| F Pretrained Method \| Raw PSNR \| RGB PSNR \| RGB SSIM \| RGB LPIPS \| \|------------------\|---------------------\|-------------\|:--------:\|:--------:\|:--------:\|:---------:\| \| Ours \| SID (noise) \| self-supervised (Neighbor2Neighbor) \| 59.32 \| 23.45 \| 0.530 \| 0.505 \| \| Ours \| SID (noise-clean) \| supervised \| 59.49 \| 23.53 \| 0.535 \| 0.499 \| Table R2: Quantitative Comparison on RawNeRF Datasets (Full Training Views) \| Method \| Loss \| Training Time \| FPS \| Raw PSNR \| RGB PSNR \| RGB SSIM \| RGB LPIPS \| \|------------------\|---------------------\|:-------------:\|:----:\|:--------:\|:--------:\|:--------:\|:---------:\| \| RawNeRF \| $L_{\text{RawNeRF}}$ \| 140h \| 0.01 \| 59.07 \| 23.53 \| 0.538 \| 0.500 \| \| HDR Scaffold-GS \| $L_{\text{RawNeRF}}$ \| 1.6h \| 73 \| 58.08 \| 22.69 \| 0.521 \| 0.513 \| \| Ours \| $L_{\text{nrr}}$ \| 3.1h \| 80 \| 59.49 \| 23.53 \| 0.535 \| 0.499** \| --- Rebuttal Comment 1.1: Comment: Thanks for the author's reply. I still have some concerns regarding the performance of this paper: 1) In the setting of full training views, without using extra clean data, your model cannot outperform RawNeRF in terms of post-processed RGB metrics. 2) What is the performance of RawNeRF in the few-shot setting? Considering the major contribution of this paper is a new training paradigm without modifying the original novel view synthesis model, I think it would be better to apply this method to NeRF and compare it with RawNeRF to fully demonstrate its effectiveness in the few-shot setting. 3) The improved training and inference speeds mainly originated from the 3DGS itself, instead of the design in this paper. However, given the effectiveness of this method on 3DGS and the excellent writing of this paper, I am inclined to give borderline acceptance. --- Reply to Comment 1.1.1: Title: Thanks for your reply! Comment: Dear Reviewer h6sN, Thank you sincerely for your reply and suggestions! ### Few-shot Setting of RawNeRF Training RawNeRF takes about one week per scene, so we currently do not have enough time to evaluate its performance. However, given that our method shows more significant improvements in the few-shot setting, we anticipate it may also perform better in this context. We will include this comparison in the final version. ### Improved Speed We acknowledge that the improved speed primarily comes from 3DGS. However, when compared to both 3DGS and 3DGS+RawNeRF loss, our method achieves enhanced rendering speed (approximately 4 times faster with a limited 4-view setting). ### Intrinsic Performance Gap Between 3DGS and NeRF 3DGS and NeRF represent different approaches to 3D representation. Prior to Scaffold-GS, 3DGS methods had lower reconstruction quality compared to NeRF, despite their faster rendering speeds. While 3DGS offered faster inference than RawNeRF, it struggled with noise overfitting, leading to a significant gap in RGB reconstruction quality (22.69 vs. 23.53 in PSNR). Our key contribution addresses this gap: we analyzed why 3DGS overfits to noise and proposed a self-supervised solution that narrows this gap, achieving parity with RawNeRF in RGB reconstruction quality (23.53 vs. 23.53 in PSNR). Additionally, our method further improves the already fast rendering speed of 3DGS. ___ We hope this clarifies the unique contributions and impact of our work. Best regards, Authors	Summary: The paper tries to use raw images with high dynamic range (HDR) for training 3D Gaussian Splatting (3dgs). It first analyzes how noise from raw image affects the optimization of 3dgs especially when the number of training views is small. To address this issue, first it uses a lens distortion correction network before training to correct the distortions of raw images. Then it introduces a noise extractor to predict noise from raw image and presents a novel noise-robust reconstruction loss which consists of a term of RawNeRF loss, a NLL measuring the divergence between the estimated noise and the physical noise, and a term for decorrelating noises across pixels. The method outperforms baselines on RawNeRF dataset on rendering quality and inference speed. Strengths: 1. The paper mainly tries to use a prior noise model to eliminate the effects of noise in raw image, pushing a step forward to utilizing HDR raw images in 3dgs. 2.The paper includes extensive comparison with baselines including LDR Scaffold-GS, HDR Scaffold-GS and two-stage methods with various denoiser as first stage. And it achieves supreme results compared to these baselines. 3. The paper analyzes and demonstrates the effect of noises from raw images on 3dgs optimization and its relationship with the number of training views. It does experiments on both full views and limited views. The whole structure is well-organized and complete. Weaknesses: 1. The method assumes clean pixel values of a 3d point projected on different images remain the same. But these images are taken from different viewpoints. How does it handle the effects of viewing directions for non-Lambertian surfaces? 2. The $L_{nd}$ loss consists of two terms. I am especially curious about the effect of the $L_{cov}$. How much would it help with the method? Technical Quality: 3 Clarity: 3 Questions for Authors: Please see the weaknesses. Confidence: 3 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: 1. The calibration process for camera noise model parameters seems to be pretty time-consuming and requires capture a lot of images, e.g. in L413 "capture 100 dark frames at each ISO in a dark room". Wonder if this method can be applied to images captured in a more home-like setting or it always requires a dataset prepared in a professional setting. Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We appreciate your valuable suggestions and comments. Here are our detailed responses to your feedback: 1. Lambertian Surface Assumption: We use the Lambertian surface assumption to simplify our analysis. However, our method can also be applied to non-Lambertian surfaces, such as mirror reflections on a piano. The core of our method is the integration of a noise model to combine the noise prior with 3DGS reconstruction. Since 3DGS supports non-Lambertian surfaces, our method is also applicable in such scenarios. 2. Effect of $L_{cov}$: The $L_{cov}$ term ensures that noise remains independent across pixels and maintains the standard deviation of the noise distribution within a pixel. We want to clarify that the ablation of $L_{cov}$ ($\lambda_{cov}$) is already presented in Table 4 of our paper. We also provide a visual illustration in Figure R3 of our submitted PDF. As shown, $L_{cov}$ ensures the noise remains independent across different pixels and adjusts the extracted noise distribution within a pixel to follow a standard Gaussian distribution after normalization. Since $L_{nd}$ only ensures the mean of extracted noise is zero, it does not control the standard deviation. Therefore, constraining the diagonal of the covariance matrix in $L_{cov}$ is necessary to ensure the standard deviation is one. 3. Home-like Setting: We have found that the noise model can also be calibrated in a home-like setting as follows: a. Flat Frames Capture: Use a MacBook screen to capture flat frames. Adjust the focus of the mobile phone (iPhone X) camera to infinity, attach it tightly to the screen, and adjust the screen brightness to capture a series of flat frames at different ISO and exposure levels. This process is illustrated in Figure R4(a) of the submitted PDF. b. Dark Frames Capture: Use a drawer and a headphone to capture dark frames. Connect the mobile phone (iPhone X) with a (wired/wireless) headphone, open the camera app, place the phone in the drawer, close the drawer, and use the headphone's play button to remotely take photos. Capture 50 dark frames at different ISOs. This setup is shown in Figure R4(b) of the submitted PDF. These settings can be easily achieved in a home environment. We found that the noise parameters calibrated this way are very close to those calibrated in a lab setting (dark room) ($\sigma_{read}$ 7.38 vs. 7.06). The impact on the final 3DGS reconstruction performance on RawNeRF is less than 0.2dB. Additionally, we have developed a semi-automatic labeling app for Android mobile phones, as shown in Figure R4\(c) of the submitted PDF, which we will release upon the paper's acceptance. We hope this tool will enable more users to enjoy 3DGS reconstruction in low-light conditions using their own mobile phones. Once again, we sincerely thank you for your constructive feedback, which has been crucial in refining our work. --- Rebuttal Comment 1.1: Comment: Thank you for answering my questions. My questions are solved. So I'll keep my positive rating. --- Rebuttal 2: Title: reply to the rebuttal Comment: Dear reviewer, The authors have submitted a rebuttal to your comments. Could you please read it and respond? The deadline for your reply is within one day. Thank you. AC	Summary: This paper proposes a novel self-supervised learning framework to reconstruct HDR 3D Gaussian Splatting (3DGS) from noisy raw images. This addresses the issue of noise degrading reconstruction quality and inference speed in 3DGS, especially in scenarios with limited views. The proposed method demonstrates superior performance compared to LDR/HDR 3DGS and previous state-of-the-art models in both reconstruction quality and inference speed on the RawNeRF dataset. Strengths: 1) The paper is well written, the methodology is well motivated and technical sound. 2) The overall problem is relevant and the introduction of a noise-free loss functions for 3DGS is plausible and valid contribution. The approach of trying to model the noise for a noise-robust representation is a original idea. Weaknesses: 1) The experimental evaluation misses a comparison to NeRF-based methods on RAW data, such as RawNeRF. While the NeRF-based methods have their downsides, they are still a good baselines to judge the actual denoising capabilities of 3DGS-based methods. 2) The impact of the noise extractor F is still unclear to me. The authors mention that this needs to be pre-trained with a paired dataset of noisy and noise-free image pairs. It would be good to investigate this component more by e.g. providing ablation of this part. Technical Quality: 3 Clarity: 3 Questions for Authors: Figure 5 is hard to understand since there are these Loss function symbols. It remains unclear to me what exactly is visible there. Maybe it would be good to show only the most relevant baseline in these figures. Confidence: 3 Soundness: 3 Presentation: 3 Contribution: 2 Limitations: The authors provide a section for discussing the limitiations their methods. Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We sincerely thank you for your thorough review and valuable suggestions. Here are our detailed responses to your comments: 1. Comparison with NeRF Baselines: We acknowledge the importance of comparing our method with NeRF-based methods. We have conducted a comparison between NeRF and HDR Scaffold-GS baselines and our proposed method in Table R1. We observed that the unmodified 3DGS (HDR Scaffold-GS) exhibited a gap in rendering quality compared to the NeRF-based method (RawNeRF), as 3DGS is more prone to noise. This is discussed in our manuscript and illustrated in Figure R2 of our submitted PDF. Our method bridges this gap, achieving similar performance to RawNeRF while significantly reducing training time (by a factor of 100) and rendering speed (by a factor of 5000), making raw 3D reconstruction feasible in real-world applications. 2. Ablation of the Noise Extractor F: We want to clarify that pretraining the noise extractor F is necessary because U-Net struggles to converge and extract high-frequency noise within 30,000 iterations during 3DGS training. Pretraining F ensures it can effectively extract high-frequency noise. Our pretrained dataset, SID (captured by a DSLR camera), differs significantly from the RawNeRF dataset (captured by an iPhone X) in terms of color space and bit depth. To verify the impact, we also experimented with a self-supervised method like Neighbor2Neighbor (using one output as the clean image) with noise-only images in SID. The results, shown in Table R2, indicate no significant difference, supporting noise-clean paired images are not strictly necessary during the pretraining stage. 3. Clarification of Figure 5: We apologize for the unclear presentation in Figure 5. The solid line in Figure 5 represents the one-stage method, corresponding to the last three rows in Table 1. The dashed line denotes the two-stage method, corresponding to the rest of Table 1. The one-stage method directly trains the 3DGS on the noisy images with different loss functions. In the two-stage method, the input is first denoised using a pretrained denoiser, and then the 3DGS is trained with RawNeRF loss functions. We will clarify this in the revised version of our manuscript for better clarity. Once again, we sincerely thank you for your constructive feedback, which has been crucial in refining our work. * Tables: Table R1: Quantitative Comparison on RawNeRF Datasets (Full Training Views) \| Method \| Loss \| Training Time \| FPS \| Raw PSNR \| RGB PSNR \| RGB SSIM \| RGB LPIPS \| \|------------------\|---------------------\|:-------------:\|:----:\|:--------:\|:--------:\|:--------:\|:---------:\| \| RawNeRF \| $L_{\text{RawNeRF}}$ \| 140h \| 0.01 \| 59.07 \| 23.53 \| 0.538 \| 0.500 \| \| HDR Scaffold-GS \| $L_{\text{RawNeRF}}$ \| 1.6h \| 73 \| 58.08 \| 22.69 \| 0.521 \| 0.513 \| \| Ours \| $L_{\text{nrr}}$ \| 3.1h \| 80 \| 59.49 \| 23.53 \| 0.535 \| 0.499 \| Table R2:** Quantitative Comparison on RawNeRF Datasets (Full Training Views) \| Method \| F Pretrained Dataset \| F Pretrained Method \| Raw PSNR \| RGB PSNR \| RGB SSIM \| RGB LPIPS \| \|------------------\|---------------------\|-------------\|:--------:\|:--------:\|:--------:\|:---------:\| \| Ours \| SID (noise) \| self-supervised (Neighbor2Neighbor) \| 59.32 \| 23.45 \| 0.530 \| 0.505 \| \| Ours \| SID (noise-clean) \| supervised \| 59.49 \| 23.53 \| 0.535 \| 0.499 \| --- Rebuttal Comment 1.1: Comment: Dear authors, Thank you , I appreciate your efforts in answering my concerns. The Table R1 indeed shows that the quality is on par with RawNeRF and the explanation of the Noise extractor f makes sense. Thus I increase my rating to a weak accept. Best Reviewer JuHQ --- Reply to Comment 1.1.1: Comment: Dear Reviewer JuHQ, We would like to express our deepest gratitude for your thorough review and invaluable feedback on our paper. We thank you for recognizing the novelty and our comparative experiments with RawNeRF, which is highly encouraging and serves as a great motivation for our continued research efforts. We are also sincerely thankful for your prompt response, which has greatly contributed to the improvement of our paper. Best regards,	Summary: This paper investigates the issue of 3DGS overfitting noises in input images, and proposes a self-supervised learning framework as the solution. The paper integrates a noise model as prior to relax the constraints in the 3DGS optimization framework. Strengths: 1. The paper provides a detailed analysis of how noise impacts the optimization of 3DGS. 2. The proposed framework leverages a physical-based noise model to jointly denoise and enhance 3DGS with noisy inputs. 3. The paper is well-written. Weaknesses: 1. It is better to introduce comparisons on more sparse-view settings such as LLFF datasets. Technical Quality: 3 Clarity: 3 Questions for Authors: see weakness Confidence: 3 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: The authors have discussed the limitations. Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We sincerely thank you for your review and valuable suggestions on our paper. In response to your suggestion regarding comparisons on more sparse-view settings, we have conducted additional quantitative comparisons using the LLFF dataset with simulated noisy raw images. Below are the details of our approach and findings: To address the more sparse-view settings (3-views), we used the inverse ISP proposed in \[1] to convert the LLFF dataset RGB images into RAW images. We then added synthetic noise following the iPhone X noise model and adhered to the baseline methods' settings \[2] for a fair comparison. Specifically, we used FSGS \[2] as the 3D representation baseline and only modified the loss function to our proposed $L_{\text{nrr}}$. The results, presented in Tables R1, demonstrate that our method consistently outperforms other methods even with limited view settings. Once again, we appreciate your constructive feedback and hope that our response addresses your concerns. * Tables and References: Table R1: Quantitative Comparison on LLFF Datasets (3 Training Views) \| Method \| Loss \| Raw PSNR (1/4)\* \| RGB PSNR (1/4)\* \| FPS (1/4)\* \| Raw PSNR (1/8)\* \| RGB PSNR (1/8)\* \| FPS (1/8)\* \| \| ----------------- \| --------------------- \| :----------------: \| :----------------: \| :-----------: \| :----------------: \| :----------------: \| :-----------: \| \| BM3D \| $L_{\text{RawNeRF}}$ \| 42.902 \| 18.032 \| 163 \| 22.596 \| 17.242 \| 272 \| \| PMN \| $L_{\text{RawNeRF}}$ \| 42.585 \| 18.064 \| 201 \| 23.208 \| 18.572 \| 389 \| \| Neighbor2Neighbor \| $L_{\text{RawNeRF}}$ \| 42.774 \| 17.977 \| 152 \| 23.008 \| 18.297 \| 372 \| \| FSGS \| $L_{\text{RawNeRF}}$ \| 41.197 \| 15.496 \| 65 \| 21.183 \| 15.166 \| 132 \| \| Ours \| $L_{\text{nrr}}$ \| 43.418 \| 18.753 \| 216 \| 23.799 \| 18.902** \| 370 \| \* * denotes the downsample ratio of the resolution following the same settings as in \[2]. References: [1] Reverse Imaging Pipeline for Raw RGB Image Augmentation; Samu Koskinen, Dan Yang, and Joni-Kristian Kämäräinen, ICIP 2019 [2] FSGS: Real-Time Few-Shot View Synthesis using Gaussian Splatting; Zehao Zhu, Zhiwen Fan, Yifan Jiang, and Zhangyang Wang, ECCV 2024	Rebuttal 1: Rebuttal: # Thanks to All the Reviewers for the Insightful Comments We would like to thank the reviewers for their efforts and insightful comments. We appreciate the reviewers’ acknowledgment of the novelty, performance, and presentation of our proposed method. For example: - Reviewer JuHQ noted that "the paper is well written, the methodology is well motivated, and technically sound." - Reviewer MTeY highlighted that "the paper includes extensive comparison ... and achieves supreme results compared to these baselines" and that "the whole structure is well-organized and complete." - Reviewer h6sN pointed out the "better reconstruction and denoising performance compared to 3DGS" and praised the "good writing and presentation." - Reviewer ZDPY appreciated the "detailed analysis of how noise impacts the optimization of 3DGS" and remarked that "the paper is well-written." The questions or weaknesses mentioned by each reviewer are addressed separately in our responses. Please feel free to discuss with us if you have any further concerns or questions. Pdf: /pdf/41399df79fc7dddcf2726f8183125eb6dded11e1.pdf	NeurIPS_2024_submissions_huggingface	2,024	Summary: This paper presents a self-supervised learning framework for reconstructing HDR 3DGS from noisy raw images. The proposed method integrates a noise extractor and a noise-robust reconstruction loss to mitigate the effects of noise in raw images. By leveraging a noise distribution prior, the framework improves both the reconstruction quality and inference speed of 3DGS, particularly in scenarios with limited training views. Experimental results on the RawNeRF dataset show that the proposed approach significantly outperforms existing state-of-the-art methods in terms of rendering quality and speed, providing a robust solution for HDR 3DGS reconstruction in challenging lighting conditions. Strengths: 1. Proposed to both denoise first and denoise on-the-fly, then using 3DGS based method to reconstruct the scene. Perform better than the other denoising method. 2. Gain better performance with fewer views. Weaknesses: 1. It seems that there are only two scenes for qualitative comparison. To the best of my knowledge, there are only three scenes for testing in the RawNeRF dataset, with each scene containing only one view. This suggests a potential lack of evaluation. The author should provide more results, at least for qualitative comparison. 2. The proposed analysis seems to be not very relevant to the main text, and I did not find its connection with the proposed noise-robust reconstruction loss. Why is a self-supervised denoiser needed if N (in Eq. 9) is small? And if N is large, is a self-supervised denoiser still necessary? Additionally, the proposed analysis is very similar to the motivation of RawNeRF and burst denoising, which uses more (large N in Eq. 9) unstabilized noisy views for denoising. Therefore, I do not see the novelty in the first contribution. 3. the author seems to have not retrained the SOTA denoisers (e.g., ELD, PMN) with the calibrated noise model of the iPhone X. This presents an unfair comparison. I think the authors should provide results using the iPhone X noise model retrained, given that the proposed method also requires noise calibration. If the results are comparable to those of the proposed method, then why is the proposed method needed? Wouldn't it be sufficient to perform denoising first and then reconstruction simply? 4. the author did not provide results compared with the original 3DGS. Why is that? It is not a big deal, but I think a comparison with RawNeRF should be provided (at least in Table 1, the results with full training views). 5. There might be a typo: if the authors directly use the pretrained checkpoint (as mentioned in lines 224-225), why is the loss for ELD, PMN, etc. denoted as $L_\text{RawNeRF}$ in Table 1? Technical Quality: 2 Clarity: 2 Questions for Authors: See the weakness. Confidence: 4 Soundness: 2 Presentation: 2 Contribution: 2 Limitations: Yes Flag For Ethics Review: ['No ethics review needed.'] Rating: 3 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We deeply appreciate your thorough review and valuable feedback on our submission. Here are our detailed responses: 1. More Quantitative and Qualitative Results: Given the limitations of the RawNeRF dataset, we conducted additional quantitative comparisons using the LLFF dataset with simulated noisy raw images and provided qualitative comparisons for the training scenes in the RawNeRF dataset. Specifically, we used the inverse ISP proposed in \[1] to convert LLFF RGB images into RAW images, added synthetic noise following the iPhone X noise model, and adhered to baseline methods' settings \[2]. We used FSGS \[2] as the 3D representation baseline, modifying only the loss function to our proposed $L_{\text{nrr}}$. The results are listed in Table R1. Our method consistently outperforms others even with limited view settings and diverse training scenes. For qualitative results, please refer to Fig. R1 in our submitted PDF, which shows that our method produces smoother results and reduces artifacts. 2. Connection Between the Analysis and Main Text: Our analysis highlights an overlooked issue: 3DGS is more vulnerable to noise in few-shot settings compared to neural radiance field-based 3D reconstruction. Unlike RawNeRF, 3DGS requires explicit noise regularization. Specifically, in Sec. 4.1, we illustrate the connection between (N) and the bias of the 3DGS optimal target and real-world point color. As (N) decreases, the bias increases, necessitating a denoiser to prevent 3DGS points from overfitting to noisy training views, as shown in Fig. 6 of our manuscript. This issue is common in scenarios like multi-camera 3D reconstruction and street scene reconstruction in autonomous driving. Although the effect diminishes with larger (N), it remains significant, as evidenced by our results with 100-view settings (Fig. 7 and Table 1). 3. Novelty of First Contribution: The impact of noise is unique in 3DGS compared to NeRF-based methods. MLPs in NeRF can perform low-frequency filtering, making it challenging to overfit to high-frequency noise, as explored in \[3]. RawNeRF requires ~5 days to train a single scene and 2 minutes to render a new view in 4K resolution, whereas 3DGS can achieve training in 2 hours and 50 FPS rendering speed. However, 3DGS tends to generate thin-flat noise, degrading rendering speed and quality. Visual details comparing NeRF and 3DGS can be found in Fig. R2 of our submitted PDF. We revise our first contribution in the intro. as follows: > We explore the unique impact of noise on 3DGS and its relationship with the number of input views. We highlight that 3DGS is more vulnerable to noise than neural radiance field-based 3D reconstruction, especially in few-shot settings, and provide a detailed analysis of how noise impacts the optimization of 3DGS, modeling its relationship with the number of training views and the noise distribution. 4. Training ELD/PMN with iPhone X Noise Model: There is a gap between simulated iPhone X noise on DSLR images and real noisy iPhone X images due to differences in bit depth (14-bits vs. 12-bits). Following your suggestion, we simulated noise-clean paired data using the iPhone X noise model and performed denoising first, followed by reconstruction. The results, listed in Table R2, do not show significant performance improvement, likely due to sensor differences between the iPhone X and DSLR cameras. 5. Comparison with 3DGS and RawNeRF: We agree that a comparison with 3DGS and RawNeRF is valuable. We have added this comparison in Table R2 under full-view settings. Our method shows that the original 3DGS is more prone to noise, resulting in lower reconstruction quality. Although NeRF-based methods achieve similar results due to their low-pass characteristics, they require significantly more time (5000x for rendering a single image and 100x for training) compared to our method, making them impractical for real-world applications. 6. Pretrained Checkpoint and Loss Function: The denoising method in Table 1 is used to first denoise the input, and then the loss function is applied to 3DGS for the 3D reconstruction loss. Thus, the loss function is only used during the 3D reconstruction training period. Once again, we sincerely thank you for your constructive feedback. * Tables and References: Table R1: Comparison on LLFF Datasets (3 Training Views) \|Method\| Loss\|Raw PSNR (1/4\)\|RGB PSNR (1/4\)\|FPS (1/4\)\|Raw PSNR (1/8\)\|RGB PSNR (1/8\)\|FPS (1/8\)\| \|---\|---\|:---:\|:---:\|:---:\|:---:\|:---:\|:---:\| \|BM3D\|$L_{\text{RawNeRF}}$\|42.902\|18.032\|163\|22.596\|17.242\|272\| \|PMN\|$L_{\text{RawNeRF}}$\|42.585\|18.064\|201\|23.208\|18.572\|389\| \|Neighbor2Neighbor\|$L_{\text{RawNeRF}}$\|42.774\|17.977\|152\|23.008\|18.297\|372\| \|FSGS\|$L_{\text{RawNeRF}}$\|41.197\|15.496\|65\|21.183\|15.166\|132\| \|Ours\|$L_{\text{nrr}}$\|43.418\|18.753\|216\|23.799\|18.902*\|370\| denotes the downsample ratio of the resolution, as same as [2]. Table R2: Comparison on RawNeRF Datasets (Full Training Views) \|Method\|Loss\|TrainingTime\|FPS\|RawPSNR\|RGBPSNR\|RGBSSIM\|RGBLPIPS\| \|---\|---\|:---:\|:---:\|:---:\|:---:\|:---:\|:---:\| \|ELD\|$L_{\text{RawNeRF}}$\|1.5h\|80\|54.70\|19.82\|0.511\|0.544\| \|PMN\|$L_{\text{RawNeRF}}$\|1.4h\|94\|53.69\|19.00\|0.498\|0.584\| \|ELD(fine-tuned)\|$L_{\text{RawNeRF}}$\|17.7h\|79\|55.01\|19.96\|0.514\|0.531\| \|PMN(fine-tuned)\|$L_{\text{RawNeRF}}$\|18.6h\|92\|53.97\|19.22\|0.503\|0.575\| \|HDR3DGS\|$L_{\text{RawNeRF}}$\|1.1h\|75\|57.39\|20.83\|0.518\|0.569\| \|RawNeRF\|$L_{\text{RawNeRF}}$\|140h\|0.01\|59.07\|23.53\|0.538\|0.500\| \|HDRScaffold-GS\|$L_{\text{RawNeRF}}$\|1.6h\|73\|58.08\|22.69\|0.521\|0.513\| \|Ours\|$L_{\text{nrr}}$\|3.1h\|80\|59.49\|23.53\|0.535\|0.499\| References: [1] Reverse Imaging Pipeline for Raw RGB Image Augmentation [2] FSGS: Real-Time Few-Shot View Synthesis using Gaussian Splatting [3] DINER: Disorder-Invariant Implicit Neural Representation --- Rebuttal Comment 1.1: Comment: Dear Reviewer m8Xp, Thank you very much for your insightful review and the valuable suggestions you provided. We have carefully considered your feedback and have addressed your comments thoroughly in our rebuttal. We believe that the clarifications and improvements we’ve made should effectively address the concerns you raised. If you find that our responses have satisfactorily resolved the issues, we kindly ask you to consider adjusting your rating accordingly. We are more than happy to provide any additional information or discuss further if needed. Thank you again for your time and thoughtful consideration. Best regards, --- Rebuttal Comment 1.2: Comment: Thanks for the author's response and thanks for the highlighted lines. However, there are still some aspects of your reply that i dont fully understand. Hope the authors would provide further clasrification. 1. As for the newly provided qualitative and quantatitive results: 1. What led the author to not provide the qualitative results of RawNeRF in Fig. R1?: I did not find the RawNeRF reconstruction results from the corresponding viewpoints in Fig R1. However, from my best knowledge, the reconstruction results of RawNeRF are much better than those of the authors' proposed method (at least at the full-view settings). As the authors mentioned that "We agree that a comparison with 3DGS and RawNeRF is valuable.", i think it is important to provide the results from RawNeRF. At least dicuss the reasons that led to the visual quality being inferior to RawNeRF (primarily due to the retention of low-frequency noise, which manifests in the results as a global color cast). 2. The qualitative results provided by the authors do not effectively demonstrate the improvements over other methods. Especially in the second row, where HDR-Scaffold clearly outperforms the proposed method in terms of reconstruction quality (e.g., the grass next to the statue). 2. Training ELD/PMN with iPhone X Noise Model: i don't think bit depth is the primary cause of the performance degraded. Moreover, quantizing from 14-bit to 12-bit can be achieved in a very simple way. 3. Sth. about the analysis: As the authors mentioned that "As (N) decreases, the bias increases, necessitating a denoiser to prevent 3DGS points from overfitting to noisy training views", than why is that all the denoise+3DGS method still suffers from color shift (mainly caused by the bias), as shown in Fig R1? Besides, the analysis only mentioned that denoising is necessary. Could you clarify the connection between the analysis and the proposed self-supervised method? --- Reply to Comment 1.2.1: Comment: Dear Reviewer m8Xp, Thank you for your follow-up questions, allowing us to clarify further. ### Comparison with RawNeRF in the Training Views: Why RawNeRF Results Are Not in Fig. R1 - The extensive training time required for RawNeRF (approximately a week per scene) limited our ability to complete all desired scenes within the rebuttal period. We used all three A100s to train the testing scenes in RawNeRF dataset and report the results in Table R2. As a result, Fig. R1 could not include additional RawNeRF comparisons for those specific viewpoints. - The notable color bias observed between RawNeRF visualizations and ours in Fig. R1 might be attributed to differences in white balance settings. In RawNeRF's test sets, the output colors are finely adjusted using reference images. However, we can not adjust the color in Fig. R1 in such a way since it is not feasible for training views due to the absence of comparable reference data. Concerning Full-view Results Comparison: - Our study primarily focuses on few-shot scenarios, yet we still provide full-view results for completeness. Our RawNeRF is trained using exactly the official code. The visualization of our result is similar with that in the RawNeRF paper (the third row in the RawNeRF Fig. A1), e.g., the appearance of the artifacts. Our implementation achieves almost the same results in the rendered RGB space as follows. Therefore, we believe the correctness of our implementation given the above consideration. \|\|PSNR\|SSIM\|LPIPS\| \|---\|---\|---\|---\| \|RawNeRF in Paper (Table 1)\|23.53\|0.536\|0.501\| \|Ours (Table R2 in Rebuttal)\|23.53\|0.538\|0.500\| - As depicted in Fig. R2, our results achieve comparable reconstruction quality to RawNeRF. However, a significant enhancement in rendering speed from 0.1 FPS to 134 FPS underscores the practical improvements our method offers. Moreover, for consistent assessment across all methods, the same color correction applied in RawNeRF is used for quantitative and qualitative comparisons in our main paper. ### Trade-off Between Preserving Sharp Details and Reducing Artifacts, Enhancing Rendering Speed, and Boosting PSNR We want to clarify that our method shows significant improvements, particularly in scenarios with limited training views, as you noted. We have effectively reduced artifacts in such cases, as seen in the first row of Fig. R1. However, in full-view settings, some blurring can occur, as mentioned in our manuscript's limitations. This is a trade-off between preserving sharp details and reducing artifacts, boosting rendering speed, and improving PSNR. For instance, in the second row of Fig. R1, the rendering speed of HDR Scaffold-GS versus our method is 52 FPS versus 96 FPS, respectively. In the testing views, we increase the PSNR of HDR Scaffold-GS from 58.08dB to 59.49dB. ### ELD/PMN Domain Gap As mentioned in our rebuttal, the domain gap is not only due to bit depth differences but also other sensor characteristics, such as the color response curve. This color bias is also noted in the RawNeRF paper: > "Since each method was trained on raw data from a different source, they impart different color tints to the output ... we calculate a per-color-channel affine transform that best matches each method’s raw output to the ground truth raw image." In training scenes, we do not have the ground truth to perform a per-color-channel affine transform, which may result in the color bias observed in our response figures. ### The Analysis The reason to introduce the denoising prior when $N$ is small is to reduce the high-frequency artifacts, e.g., needle-like artifacts in HDR-Scaffold. As shown in the figures, introducing a denoiser effectively reduces these artifacts. The color bias is mainly due to the DC component, i.e., the bias you mentioned. For the reason of using a self-supervised manner, the rationale is as follows: - Self-supervised denoising is a meaningful topic that many researchers are working on. It has unique advantages in cases where the clean image domain is unavailable, e.g., 4D Gaussian modeling of dynamic scenes in the dark. - While the noise model in raw space is relatively clear, only having a noise prior cannot obtain the clean image/scene. Similar to the deep image prior (DIP) \[a\], we are the first to find that 3DGS also has a similar 3D scene prior as shown in our analysis. - While using 3DGS alone can somehow denoise the scene by controlling the number of iterations as in Fig. 3, it has the risk of overfitting to the noise. We are also the first to combine the noise prior and 3DGS prior in a self-supervised manner. In such a way, we can avoid overfitting to the noise and achieve better results, e.g., superior PSNR and FPS. We hope these address your concerns. Please feel free to reach out with further questions. \[a\] Deep image prior, CVPR 2018 --- Rebuttal 2: Title: Thanks for your reply! Comment: Dear Reviewer m8Xp, We sincerely thank you for your review and feedback. ### Color bias We believe that the ISP used in RawNeRF for training views are different from the ones we used, this shall be the cause of the color bias. The detailed reasons are as follows: - Even the visualizations of the input images are different from the ones in RawNeRF as shown in Fig. 9 of their arixv paper. - We observed the same color bias as shown in our Fig. R1 input images using `provided ISP w/o color correction` `default viewer of Windows`, `default viewer of MacOS`, and `rawpy with in-camera white balance`. - We achieved different and inaccurate white balances using `rawpy with its auto white balance`. Since the ISPs of training views are not released (the provided ISP for testing needs GT images for color correction as you mentioned), we will be grateful if you have any idea to obtain the same results in their paper and we will update the results in the revised version. Even if there are differences in the usage of ISP in training views, we use the same provided ISP for the test views in the test scenes so all the results in the paper do not have this issue. ### Loss of details - We acknowledge that some details are lost in the grass of the second row. However, as seen in other figures—including those in the manuscript and rebuttal PDF, such as the spokes of the bicycle wheel and the text on the board—we generally preserve details well. - The loss of detail in the grass in the second row of Fig. R1 might be coincidental. If it were due to the over-smoothing nature of our method, this effect would likely be visible in all other figures as well. - However, our method achieves better PSNR, SSIM, and LPIPS compared with HDR Scaffold-GS. The quantitive results demonstrate that our method can achieve overall better detail preserving, and over-smoothing is not a common behavior of our method. - By removing the needle-like artifact, we can significantly improve the rendering FPS, e.g., increases from 53 FPS to 134 FPS even for full-view training in Fig. R2 and ~4x FPS increase for 4-view training. We hope this clarifies the unique contributions and impact of our work. Best regards, Authors	null	null	null	null	null	null
Accelerated Regularized Learning in Finite N-Person Games	Accept (poster)	Summary: The paper studies the extension of Nesterov’s accelerated gradient algorithm to the solution of N-player games, named "follow the accelerated leader" (FTXL). The method is studied in both continuous and discrete time, and under various information oracles. The convergence of the algorithm is super-linear when started from an initialization close enough to the NE. Simulations are conducted which confirms the accelerated convergence numerically. Strengths: The paper is very well-written and well-motivated. The core contribution, which is the accelerated algorithm with local super-linear convergence, is clearly presented, novel, and significant to my knowledge. That the method works under different information oracles, especially bandit feedback, is also important. Despite the simplicity, the simulations do numerically verify the accelerated rate established in theory. Weaknesses: The convergence of FTXL in all settings requires sufficiently close initialization or has to do with the neighborhood $\mathcal{U}$. The exact definition of "sufficiently close" or this neighborhood is not given in the main paper. I believe this should be added in the statement of the theorems for clarity. The authors should also discuss how realistic it is that the initial point satisfies the requirement (and/or how likely the iterates of FTXL or FTRL may run into this "good neighborhood"). It seems that no assumption on the convexity of the payoff function is made. With a general non-convex payoff function, I wonder what is the problem structure and technical apparatus that allow the authors to derive the quadratic convergence (or even convergence at all). This is currently not sufficiently discussed in the main paper. While I am happy to study the detailed analysis, I would appreciate a sketch of the proof which highlights the main techniques. Technical Quality: 3 Clarity: 3 Questions for Authors: Please see Weaknesses section. Confidence: 3 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: Paper of theoretical nature. No negative social impact is foreseeable. Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Dear Reviewer sdz4, Thank you for your positive evaluation and constructive comments! We reply to your questions below: > The convergence of FTXL in all settings requires sufficiently close initialization or has to do with the neighborhood. The exact definition of "sufficiently close" or this neighborhood is not given in the main paper. I believe this should be added in the statement of the theorems for clarity. The authors should also discuss how realistic it is that the initial point satisfies the requirement (and/or how likely the iterates of FTXL or FTRL may run into this "good neighborhood"). The determining factor for the basin of attraction of a given equilibrium $x^$ is the minimum payoff difference $d$ at equilibrium, as per Eq. (B.1) of our paper. The neighborhood $\mathcal{U}$ in question is roughly $\mathcal{O}(d)$ in $L^1$ diameter, and it is essentially determined by the equation $u_i(x_i^;x_{-i}) - u_i(x) \geq d/2$. We provide the relevant details in p.13, Lemma B.1 in Appendix B, and we will be happy to transfer the corresponding expressions to the main body of the paper. As for the likelihood of (FTXL) being captured by the basin of attraction of a given equilibrium, this depends on the existence or not of non-equilibrium attracting sets – such as the heteroclinic limit cycle in Jordan's matching pennies, which is universally attracting under the replicator dynamics, that is, the first-order version of (FTXL) with entropic regularization. In typical coordination / anti-coordination scenarios, the state space of the game is partitioned into basins of attraction of different strict equilibria, so convergence is almost surely guaranteed. Otherwise, if there are non-equilibrium attracting (or chain-recurrent) sets, we conjecture that (FTXL) could be captured by a spurious attractor on the boundary and fail to converge (as in Jordan's matching pennies). To the best of our knowledge, there is no theory providing a coherent characterization of when such spurious attractors may arise in finite games; this is a very difficult problem on which very little progress has been done since the 1950's, so we cannot provide more insights here. > It seems that no assumption on the convexity of the payoff function is made. With a general non-convex payoff function, I wonder what is the problem structure and technical apparatus that allow the authors to derive the quadratic convergence (or even convergence at all). This is currently not sufficiently discussed in the main paper. While I am happy to study the detailed analysis, I would appreciate a sketch of the proof which highlights the main techniques. Please note that our paper focuses throughout on finite games in normal form, so the players' payoff functions are, de facto, multilinear in the players' mixed strategies – and, in particular, linear in each individual player's strategy. In the case of continuous games – that is, games with a finite number of players and a continuum of actions per player – things are dramatically different, especially if (as you suggest) there are no individual convexity assumptions made on the players' payoff / cost functions. The extension of FTXL to continuous games (possibly with a monotone structure) is a very fruitful research direction, but it would require a drastic departure from the finite game setting of the current paper, so it lies well beyond the scope of our work. Now, going back to our particular setting, the basic insight is that strict Nash equilibria have a "sharp" variational characterization, in the sense that the players' payoff vector lies in the interior of the normal cone to the point in hand (akin to Polyak's notion of sharpness for minimization problems). This allows the buildup of significant momentum accelerating the algorithm, and since strict equilibria are extreme points (all positivity constraints are saturated except one, owing to the simplex equality constraint), there is no danger of overshooting the point in question. This is the reason that the lack of friction does not hinder convergence – and, of course, it is inextricably tied to the (multi)linear structure of finite games and their strategy domains. We will of course be happy to take advantage of the extra page of the first revision opportunity to describe all this in more detail and provide a technical roadmap of the proof. --- Please let us know if you have any follow-up questions, and thank you again for your time and positive evaluation! Kind regards, The authors --- Rebuttal Comment 1.1: Comment: I thank the authors for the clarification and would like to maintain my score. I support the acceptance of the paper. --- Reply to Comment 1.1.1: Title: Thank you Comment: Thank you for your time and your support! Kind regards, The authors	Summary: The submission studies the convergence of finite N-person games. Combining the idea of (NAG) and (FTRL), the submission proposes a continuous-time dynamic (FTXL-D) and its discrete-time scheme (FTXL). The novelty of integrating momentum and regularization into an algorithm allows the construction of quadratic rate methods. Strengths: - A novel viewpoint that uses (HBVF)’s interpretation to link (NAG) and (FTRL-D) to form (FTXL-D). - Improved converge rates from linear to quadratic are shown in Theorems 1 – 4. Theorems 2 – 4 cover common practical scenarios. - Clear comparison between the proposed method and the previous methods, especially at the intuition level. - Pinpoint the critical term to focus on the technical details (U in Theorem 3). This greatly increases the time to understand the analysis. - Rigorous analysis is provided in the Appendix. Weaknesses: - A smoother transition toward (FTXL-D) in lines 182–186 is desired. In a game, $x$ is the response and $y$ is an aggregated payoff. For (HBVF), the interpretation is for a moving object. - What is the correspondence of $x$ and $y$ of a game in (HBVF)? Do we correspond $x$ (position) in (HBVF) to $y$ (payoff) in a game? If so, what corresponds to the response ($x$ in a game) in (HBVF)? Is it the control $w$ (my naive guess)? - These analogs, with the aid of a clearer transition, would help us to better understand the first paragraph of Section 3.2. Maybe the response x in a game and the position x in (HBVF) introduces a bit of confusion, too? - There are two sets of notations, one for the continuous-time regime and the other for the discrete-time regime. A small table to distinguish them might help the readability. - Although in general, the submission is easy to read, a minor aspect of improving readability could be explaining the meaning $\dot{y}$ (lines 192 and 196) and $\ddot{z}$ (line 486), which could ease the reading barrier for people who are more comfortable with discrete-time analysis. Technical Quality: 3 Clarity: 3 Questions for Authors: - In line 245, the friction parameter is set to zero ($r=0$) according to the discussion of Example 3. This setting, by direct substitution, would cancel out the desired effect of the momentum in (FTXL-D) in line 186 or (4) in line 197. However, in (FTXL) (line 258), the momentum appears to play an important role in the leftmost equality $y_{n+1} = y_n + \gamma p_n$. Did I misunderstand something? - In Figure 1, why does FTXL perform poorly in the early rounds? Is it due to the constants in the bound or is there another reason? Confidence: 4 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: Yes. Flag For Ethics Review: ['No ethics review needed.'] Rating: 7 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Dear Reviewer 3QUL, Thank you for your strong positive evaluation and constructive comments! We reply to your questions below: > A smoother transition toward (FTXL-D) in lines 182–186 is desired. [...] What is the correspondence of $x$ and $y$ of a game in (HBVF)? [...] These analogs, with the aid of a clearer transition, would help us to better understand the first paragraph of Section 3.2. The key insight here is that, in the case of regularized learning, the algorithm's latent, state variable – that is, the variable which determines the evolution of the system in an autonomous way – is the "score variable" $y$, not the "strategy variable" $x$ (which is an ancillary variable obtained from $y$ via the regularized choice map $Q$). In particular, this means that, even though results are stated in terms of the strategy variables $x$ (which are the primary objects of interest), the true state variable of the algorithm's iterative structure is $y$, not $x$. Put differently, the $x$ in (NAG) and (HBVF) should be compared and contrasted to $y$ in (FTXL), not $x$. We made this choice of notation to be as close as possible to the work of Su, Boyd and Candès, but we understand that it may have occluded intuition. We will make sure to update the presentation here and provide more details along the above lines at the first revision opportunity. > There are two sets of notations, one for the continuous-time regime and the other for the discrete-time regime. A small table to distinguish them might help the readability. This is a great idea, thanks! We were (very sharply) constrained for space in the original submission, but we will be happy to take advantage of the extra page to include such a table. Thanks for the suggestion! > Although in general, the submission is easy to read, a minor aspect of improving readability could be explaining the meaning of $\dot y$ (lines 192 and 196) and $\ddot z$ (line 486), which could ease the reading barrier for people who are more comfortable with discrete-time analysis. Point well taken. We will include a note about this, and also make a reference to it in the proposed table of notation above. > In line 245, the friction parameter is set to zero ($r=0$) according to the discussion of Example 3. This setting, by direct substitution, would cancel out the desired effect of the momentum in (FTXL-D) in line 186 or (4) in line 197. However, in (FTXL) (line 258), the momentum appears to play an important role in the leftmost equality $y_{n+1} = y_n + \gamma p_n$. Did I misunderstand something? The desired effect of the momentum in (FTXL-D) – but also (NAG) and (HBVF) – is not the friction term, but the fact that the system is second-order in time, so the effects of the momentum are "baked in" the algorithm's update structure. The friction term $r\dot x$ (or $r\dot y$) has a "dampening effect" intended to mitigate the algorithm overshooting a desired state. We hope this is clearer now, please let us know if we missed or misunderstood something in your question! > In Figure 1, why does FTXL perform poorly in the early rounds? Is it due to the constants in the bound or is there another reason? The reason is that the algorithm is building up momentum in the initial iterations, much like uniformly accelerated motion (like pushing a crate with constant force): the algorithm starts at zero speed and is initially slow, but as it gains more and more momentum, it rapidly accelerates and gains more speed and momentum, which is ultimately responsible for the method's quadratic convergence rate. In this analogy, FTRL would correspond to uniform motion – always moving at "constant" speed, so it is faster in the initial iterations, but much slower overall since there is no momentum build-up. --- Thank you again for your strong positive evaluation and encouraging remarks – and please let us know if you have any follow-up questions! Kind regards, The authors --- Rebuttal Comment 1.1: Comment: Thank you for your response. I have no further questions. --- Reply to Comment 1.1.1: Comment: Thank you again for your time, input, and positive evaluation! Kind regards, The authors	Summary: The paper proposes a momentum-based follow-the-regularized-leader (FTRL) type algorithm for finding the Nash equilibrium in finite games. The paper first investigates the continuous second order ODE and a concrete game, and then devises the appropriate FTRL scheme. The convergence rate for the proposed algorithm in all three information settings of reward is quadratic (O( exp(-T^2))). Strengths: 1. A rich collection of results on momentum methods for finite-games. The paper has three sets of theoretical results. First, convergence results for the second-order ODE dynamics with vanishing friction (Theorem 1) and non-vanishing friction (Theorem B.1), both under the logit best response. Second, an analysis of a momentum FTRL under full information on a concrete one-player game with vanishing (Prop C.1) and non-vanishing friction (Prop C.2). And finally, analysis momentum FTRL with three kinds of info feedback. 2. The paper discovered an important phenomenon that, in the continuous second order ODE, "the friction term hinders convergence". This claim is backed by a concrete analysis of momentum FTRL on a concrete game. In my opinion this phenomenon is interesting to the game learning community. 3. The writing is flowing and smooth. Weaknesses: See question. Technical Quality: 4 Clarity: 4 Questions for Authors: Minor comments 1. Are lower bounds of learning a finite game with the three kinds feedbacks known? If so, it would be great to provide a survey. 2. In line 305 the phrasing "second-order in space" is a bit confusing. In my opinion, it would be better to say "using Hessian information of relevant functions". Typo 1. At Eq FTRL, near line 131, the second equation should be x_i,n = Q_i (y_i, n). Confidence: 3 Soundness: 4 Presentation: 4 Contribution: 4 Limitations: The paper provides a extensive discussion of momentum method for finite game solving, and does not have major limitations in my opnion. Flag For Ethics Review: ['No ethics review needed.'] Rating: 8 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Dear Reviewer qtGD, Thank you for your overwhelmingly positive evaluation and constructive comments. We reply to your questions below: > Are lower bounds of learning a finite game with the three kinds feedbacks known? If so, it would be great to provide a survey. The closest lower bounds that we're aware of are algorithm-specific, e.g., as in the paper of Giannou et al. [15] where the authors argue (but do not prove) a geometric lower bound for FTRL with full information feedback, which is also achieved by FTRL with bandit feedback (so it is tight for both models). We fully share the reviewer's point of view but, regrettably, we are not aware of a more general lower-bound analysis as in the case of convex minimization. > In line 305 the phrasing "second-order in space" is a bit confusing. In my opinion, it would be better to say "using Hessian information of relevant functions". Point well taken, we will adjust the phrasing accordingly to make sure there is no confusion. > At Eq FTRL, near line 131, the second equation should be x_i,n = Q_i (y_i, n). Oh, great catch, many thanks - will fix! --- Please let us know if you have any follow-up questions, and thank you again for your time and overwhelmingly positive evaluation! Kind regards, The authors	Summary: This paper primarily focuses on introducing a Nesterov's accelerated gradient (NAG) algorithm for online learning in games. Initially, the author shows that a continuous-time version of NAG converges to a strict Nash equilibrium at a rate that is quadratic. This rate of convergence is notably faster than that of standard FTRL algorithms. Following this, the paper demonstrates that the convergence rate of NAG is preserved even in a bandit feedback setting. The experimental results confirm that the proposed algorithm converges to an equilibrium more quickly than a FTRL algorithm. Strengths: * Despite the significant importance of introducing Nesterov's accelerated gradient, a method highly esteemed in convex optimization problems, into the context of learning in games, such an approach has not been pursued until now. * The intuition behind the proposed method is thoroughly explained. Furthermore, the proposed algorithm appears to be straightforward to implement. * The derived convergence rates are confirmed to be tight by deriving the exact convergence rate in a simple instance. Weaknesses: My primary concern is that the theoretical results are only applicable to games with a strict Nash equilibrium. This is somewhat problematic as even in a simple game like rock-paper-scissors, a strict Nash equilibrium does not exist. Moreover, the convergence in games that do not have a strict equilibrium has been the subject of extensive research. Therefore, I'm wondering how efficient NAG algorithms are in two-player zero-sum games and monotone games. Furthermore, Theorems 1, 2, 3, and 4 hinge on the presumption that the initial strategy is in close proximity to an equilibrium. Could you provide some insight into how close the initial point needs to be to the equilibrium? Technical Quality: 3 Clarity: 3 Questions for Authors: My major concerns and questions are stated in Weaknesses. I also have the following additional questions: * Why is the update rule of $y_{i,n+1}=y_{i,n}+\gamma p_{i,n+1}$ in (FTXL) derived? Looking at the original NAG's update rule in (NAG), it seems more natural that $y_{i,n+1}=y_{i,n}+\gamma (p_{i,n+1}-p_{i,n})$. * Is it possible to recover FTRL algorithms by setting $r$ appropriately in FTXL? Confidence: 3 Soundness: 3 Presentation: 3 Contribution: 2 Limitations: I see no negative societal impacts that need to be addressed. Flag For Ethics Review: ['No ethics review needed.'] Rating: 5 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Dear Reviewer yttq, Thank you for your input. For your convenience (and that of the committee), we reproduce and reply to your comments below one-by-one. [Note: all bibliography reference numbers are as in our paper] > Results are only applicable to games with a strict Nash equilibrium. [...] Therefore, I'm wondering how efficient NAG algorithms are in two-player zero-sum games and monotone games. There are several points here – related, but distinct: 1. On the lack of strict equilibria. We should begin here by stating that, by the impossibility theorem of Hart & Mas-Colell [17], there are no uncoupled learning dynamics that converge to Nash equilibrium in all games. In this regard, at least some assumption needs to be made in order to have a hope of converging to a Nash equilibrium. Our focus on games with strict Nash equilibria is motivated by the results of [14] which state that, in the presence of uncertainty – i.e., with anything other than full, perfect information – *only* strict Nash equilibria can be stable and attracting with high probability. In particular, if a game does not possess a strict equilibrium, the sequence of play generated by regularized learning schemes provably fails to converge with positive probability. Thus, given that we are interested in robust convergence results that continue to hold in the presence of uncertainty, the focus on strict Nash equilibria is, in this sense, unavoidable. 2. On zero-sum games. A first point here is that zero-sum games may well admit strict equilibria, in which case our results apply verbatim -- for example, consider the min-max payoff matrix $[0 \quad 1; -1 \quad 0]$. In general zero-sum games, it can further be shown that FTXL identifies the support of the set of Nash equilibria of a zero-sum game at a quadratic, superlinear rate – though, in view of the discussion above, high probability convergence to equilibrium under uncertainty should not be expected if the support is not a singleton (e.g., as in RPS). [We did not include this more specialized result because it was beyond the scope of our submission, but we will be happy to include the above discussion in the first revision opportunity.] Finally, even though there is an extensive literature on (usually optimistic) regularized learning methods that converge to equilibrium in zero-sum games with fully mixed Nash equilibria, it should be noted that these results typically concern learning with full, perfect information, as per model (10a) in our paper. These results are inherently deterministic and collapse in the presence of persistent randomness and uncertainty, see e.g., https://arxiv.org/abs/2206.06015 or https://arxiv.org/abs/2110.02134, so it is in general quite difficult to achieve convergence to Nash equilibrium without full information in zero-sum games. 3. On monotone games. Monotonicity is a condition that primarily concerns continuous games, that is, games with a finite number of players and continuous action spaces. By contrast, our paper focuses throughout on finite games: the mixed extension of a finite game may indeed be seen as a continuous game, but the players' (mixed) payoff functions are necessarily multilinear in their strategies. As a result, except for trivial cases (like the zero game), only two-player games can be monotone, and this only if the players' payoff matrices satisfy very specific conditions that ultimately make the game strategically equivalent to a zero-sum game. The extension of FTXL to (monotone) games with continuous action spaces is a very interesting one, but this is otherwise a completely different framework which lies well beyond the scope of our work. > Could you provide some insight into how close the initial point needs to be to the equilibrium? The determining factor is the minimum payoff difference $d$ at an equilibrium $x^$ as per Eq. (B.1) of our paper. The neighborhood in question is roughly $\mathcal{O}(d)$ in $L^1$ diameter, and it is essentially determined by the equation $u_i(x_i^;x_{-i}) - u_i(x) \geq d/2$. We provide the relevant details in p.13, Lemma B.1 in Appendix B, and we will be happy to transfer the corresponding expressions to the main body of the paper. > Why is the update rule of $y_{n+1} = y_n + \gamma p_{n+1}$ in (FTXL) derived? [...] It seems more natural that $y_{n+1} = y_n + \gamma (p_{n+1} - p_n)$ The update rule in (FTXL) is derived by setting $p_n = (y_n - y_{n-1})/\gamma$ in (9), which is the direct discretization of (FTXL-D), the game-theoretic analogue of (HBVF). [Put differently, mutatis mutandis, (FTXL) is related to (FTXL-D) in the same way that (NAG) is related to (HBVF).] In our original submission, we had devoted Section 3 to provide a detailed discussion on the connection of (FTXL-D) with (NAG) and (HBVF), along with the intuition behind it, referring to the paper of Su, Boyd and Candès for the detailed relation between (HBVF) and (NAG). We will be happy to take advantage of the extra page in the revision to include the details of the link between (HBVF) and (NAG) for completeness. > Is it possible to recover FTRL algorithms by setting $r$ appropriately in FTXL? That's an interesting question. We tried to find a way to see if it's possible to express FTRL as a limit case of FTXL, but we didn't find one. --- Please let us know if you have any follow-up questions, and thank you again for your time! Kind regards, The authors --- Rebuttal Comment 1.1: Comment: Thank you for your response. I have no further questions, and I will keep my score. --- Reply to Comment 1.1.1: Comment: Thank you again for your time and input. Kind regards, The authors	Rebuttal 1: Rebuttal: Dear reviewers, dear AC, We are sincerely grateful for your time, comments, and positive evaluation! To streamline the discussion phase, we replied to each of your questions and comments in a separate rebuttal below, and we will of course integrate all applicable points in the next revision opportunity. Thank you again for your input and encouraging remarks, and we are looking forward to a continued constructive exchange during the discussion phase. Kind regards, The authors	NeurIPS_2024_submissions_huggingface	2,024	null	null	null	null	null	null	null	null
Model Fusion through Bayesian Optimization in Language Model Fine-Tuning	Accept (spotlight)	Summary: The main contributions - Using Bayesian Optimization for hyperparameter search on LoRA’s for full-model fine-tuning - Using EVHI to learn coefficients to fuse models from within a checkpoint Other contributions - Finding discrepancy between metric and loss Strengths: - Paper is well-written and easy to follow - Bayesian method for model fusion is interesting Weaknesses: - Empirical observations of discrepancy between metric and loss landscape are on one specific outdated model and not sure how accuracy is computed (but could be an outdated way). Same for hyperparameter alignment claim. See questions below - Extra compute required for BOMF - Performance improvements of BOMF are very small Technical Quality: 2 Clarity: 3 Questions for Authors: Major - How much extra time is required due to the sequential nature of HBPO compared to grid search? - Some discussion of the extra compute required for BOMF due to the extra evaluations. Does 75 iterations mean 75 evaluations are needed? - How is accuracy computed on RoBERTa? With a classifier head on top of the representations? Nowadays, for NLP, all tasks are cast as text-to-text and accuracy is done by comparing the log probs of the tokens corresponding to the different labels. - The claim that the best hyperparameters for fine-tuning LoRA can transfer to the full model are not that interesting for RoBERTa. Since RoBERTa is a pretty small model overall, LoRA is generally not needed and full-model fine-tuning is needed. LoRA is usually used for larger models, so showing this for the LLAMA-2 experiments would be interesting. Minor - Can you an algorithms description of MOBO in the main paper? Specifically, how is qNEHVI done? - Where is freeze in Table 1 and rank 4 in Table 2? (Mentioned in line 298/315) - What are the two variants of BOMF in Table 1 and 2? - Why are there are no grid fine-tune results in Table 2? Confidence: 3 Soundness: 2 Presentation: 3 Contribution: 3 Limitations: NA Flag For Ethics Review: ['No ethics review needed.'] Rating: 7 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We appreciate your constructive feedback. We have answered your questions and concerns in this response. Please let us know if you have any follow-up questions. [W1, Q3, Q4] Empirical observations of discrepancy between metric and loss landscape, and hyperparameter alignment are performed on the RoBERTa model. Thus it requires results on Llama models. Also how does the accuracy computed on RoBERTa? First, we want to clarify that we use a classifier head for the classification task because our task involves multiple-choice questions, allowing us to directly output an answer and measure accuracy. However, we want to emphasize that BOMF is not limited to this specific scenario. It can be applied to any model as long as the loss and metrics can be calculated. Since fusion does not require backpropagation, BOMF is equally applicable in situations where there is no classifier head. And in our paper, we demonstrated that both full model fine-tuning and LoRA fine-tuning exhibit 1) metric and loss misalignment and 2) optimal hyperparameters alignment even when freeze layers or ranks change. However, we agree that extending these experiments to larger-scale models would strengthen our findings. Therefore, we conducted additional experiments on the Llama-2 model to verify if 1) and 2) still hold. Firstly, regarding point 2), the alignment of hyperparameters can be easily observed in Figures A.1. For point 1), the misalignment between loss and metrics is evident in Figure A.2. To further show the misalignment between loss and metrics, as well as among different metrics, we measured Spearman’s correlation using the loss and metrics of 20 sampled models. Also, we empirically prove that BOMF significantly improves this correlation by considering both loss and metrics simultaneously. Refer to Tables 9, 10, and 11 in Appendix C for full experimental results. We have included a portion of these results here. According to Table R.2 (c), (d), and (e), language tasks involving the Llama model show significantly lower correlation compared to vision tasks (a) and (b). Then, Table R.3 shows that BOMF successfully increases the correlation between loss and metrics compared to other baseline methods. [W2, Q1, Q2] How much extra time is required for BOMF. Please refer to our global response for answers. [W3] Performance improvements of BOMF are very small The improvement of BOMF is not negligible compared to other baselines. First, we would like to emphasize that our method consistently outperforms various baselines across a wide range of task types (i.e., classification, question answering, and the medical domain) and models (i.e., T5, LLaMA3). While the degree of improvement might appear marginal in some cases, it is essential to recognize that the extent of improvement, whether small or large, is relative. The fact that our method consistently achieves superior performance in nearly every case serves as an absolute metric that is universally recognizable. Our method consistently outperforms the baselines in these diverse scenarios. Specifically, BOMF achieved a 13%, 14%, 10% error improvement in the GLUE tasks, SQuAD task and KorMedQA tasks, respectively. Furthermore, the recent fusion research [1,2,3] in various scenarios has shown similar tendency to our results, showing that even if the improvement in specific tasks seems marginal, they exhibit overall superior performance compared to the baseline, proving the necessity of such methods. In particular, even when compared to these methods, our method demonstrates improved performance in almost all experimental results, indicating its effectiveness. [Q5] Can you provide an algorithm description of MOBO in the main paper? Specifically, how is qNEHVI done? We have already described the details of our algorithm in Sec A.1.3, B, and Algorithm 1 of the appendices. Due to the limited space of the NeurIPS submission, they should be located in the appendices. We will reorganize our contents in the final version. As described in our manuscript, qNEHVI sequentially samples optimum candidates by maximizing the expected hypervolume improvement, which can be considered as an acquisition function in Bayesian optimization. At every iteration of MOBO, it chooses the maximizer of the alternative objective rather than one of multiple unknown target objectives. We will clarify it in the final version. [Q6, Q7] Where is freeze in Table 1 and rank 4 in Table 2? Thank you for pointing out the confusion. The dagger symbol ($\dagger$) in Tables 1 and 2 indicates whether low rank or freeze layers were used. Specifically, "BOMF with dagger symbol ($\dagger$)" refers to cases where the best hyperparameters were determined using full rank or full layer settings, and these hyperparameters were then used for training before performing fusion. On the other hand, "BOMF without a dagger symbol" refers to cases where low rank or freeze layers were used to find the best hyperparameters. This information is clarified in the footnote on Page 8 of the paper. [Q8] Why are there no grid fine-tune results in Table 2? The models used in Table 2 are LLMs, so even when tuning with LoRA, the forward path takes a considerable amount of time. This made it time-prohibitive to perform grid search, which requires a relatively large number of iterations to find the best-performing hyperparameters. However, to strengthen our argument in line with your suggestion, we are running the grid search in the remaining time and will present the results as soon as they are available. References [1] Wortsman, et al. Model soups: averaging weights of multiple fine-tuned models improves accuracy without increasing inference time. ICML, 2022. [2] Weng, R., et al. G-tuning: Improving generalization of pre-trained language models with generative adversarial network. ACL, 2023. [3] Malladi, S., et al. Fine-tuning language models with just forward passes. NeurIPS 2023. --- Rebuttal Comment 1.1: Comment: Thanks for the response and for running the additional experiments on llama-3 for the empirical observations. I have increased my score from a 6 to a 7. --- Reply to Comment 1.1.1: Comment: Thank you for the positive feedback on our paper. We will incorporate all the discussions into the final manuscript. Also, in your response, you mentioned that you would raise the score to 7, but it seems that it wasn’t updated. Could you please adjust the score accordingly?	Summary: The paper presents a novel approach to model fusion through Bayesian Optimization for fine-tuning pre-trained language models on downstream tasks. The authors address the challenges associated with hyperparameter selection and the discrepancy between loss and metric landscapes during the fine-tuning process. They propose a two-stage Bayesian optimization framework that first identifies optimal hyperparameters for fine-tuning and then uses multi-objective Bayesian optimization to find the best combination of models in the parameter space. The paper provides a comprehensive experimental evaluation to demonstrate the properties and benefits of the proposed approach. Strengths: 1. The paper introduces a novel approach to model fusion through Bayesian Optimization (BOMF), which is a creative application of Bayesian optimization to the problem of fine-tuning pre-trained language models. The originality lies in the development of a two-stage Bayesian optimization process that addresses the specific challenges of hyperparameter tuning and model combination in NLP tasks. The method is innovative as it optimizes both the loss and the desired metrics simultaneously, which is a new perspective in the context of model fusion. The empirical finding that a large discrepancy exists between loss and metric in fine-tuning LLM is valuable to the community. 2. The quality of the paper is evident in its rigorous experimental design and comprehensive evaluation. The authors have tested BOMF across various NLP tasks and models, demonstrating its robustness and effectiveness. 3. The paper is well-structured, with a clear presentation of the problem, the proposed solution, and the experimental setup. Overall, it is well-written and easy to follow. Weaknesses: 1. There are no discussions on computational efficiency: Although the paper mentions a two-stage Bayesian optimization process, it could further elaborate on the computational efficiency of BOMF, especially when compared to traditional fine-tuning methods and the baselines. More detailed complexity analysis or comparisons could help readers understand the practicality of applying BOMF at scale. 2. Hyperparameter sensitivity analysis: The paper could provide a more in-depth analysis of how sensitive BOMF is to the choice of hyperparameters for the Bayesian optimization process itself, e.g., the choice of kernel of Gaussian process regression. Understanding the stability of the method under different settings can be crucial for practitioners. Technical Quality: 3 Clarity: 3 Questions for Authors: See weakness. Confidence: 4 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: The authors have adequately discussed the limitations of their work. Flag For Ethics Review: ['No ethics review needed.'] Rating: 7 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We appreciate your constructive feedback. We have answered your questions and concerns in this response. Please let us know if you have any follow-up questions. [W1] There are no discussions on computational efficiency: More detailed complexity analysis or comparisons could help readers understand the practicality of applying BOMF at scale Please refer to the general response for the analysis and additional experiments regarding the computational cost. [W2] Hyperparameter sensitivity analysis: The paper could provide a more in-depth analysis of how sensitive BOMF is to the choice of hyperparameters for the Bayesian optimization process itself Since the test suites used in this work require huge compute resources, it is difficult to provide hyperparameter sensitive analysis on Bayesian optimization processes. It is noteworthy that the hyperparameters of Bayesian optimization can be thought of as meta-hyperparameters, which should be tuned with multiple tasks and multiple domains. It implies that it necessitates utilizing more compute resources. In addition, from the standpoint of black-box optimization, the sensitive analysis assumes that the specifics of objective functions are known, which is generally infeasible in black-box optimization. Therefore, instead of analyzing the sensitivity of the hyperparameters of Bayesian optimization, we followed their conventional settings. All hyperparameters related to Gaussian processes are optimized via model selection. Matern 5/2 kernel for Gaussian processes and the logarithmic form of expected improvement are used. --- Rebuttal Comment 1.1: Comment: I thank the authors for providing responses to the questions. This rebuttal essentially clarifies my concerns, and I will raise my score to 7 accordingly. --- Reply to Comment 1.1.1: Comment: Thank you for the positive feedback on our paper. We will incorporate all the discussions into the final manuscript.	Summary: This paper proposes "Bayesian Optimization Model Fusion" (BOMF), which is a method to fuse model weights using Bayesian Optimization over a set of metrics for a given target task. The authors motivate the necessity of BOMF for model fusion in large language model fine-tuning by providing evidence that there exists a large misalignment between the loss and metric surface in NLP tasks for pre-trained models. Additionally, the authors demonstrate initially that, for a given target task, the optimal set of hyperparameters for fine-tuning appears to remain consistent across different levels of fine-tuning capacity (# of frozen layers or LoRA rank). Drawing on these insights, the total proposed method operates in 3 steps: - First, Bayesian Optimization is used to select the optimal set of hyperparameters for a given task and setting, optimizing the hyperparameters over the target-task metric, rather than validation loss. Additionally, drawing on their earlier observation, the authors perform this BO procedure on a lightweight model (freezing lower layers of the pre-trained model), and then applying the found hyperparameters to the larger model. - Then, the parameters are fine-tuned on the target-task, and various checkpoints during this fine-tuning procedure are uniformly sampled (towards the end of training) for model fusion. - Finally, to select the best set of weights to use when fusing the models together (i.e. taking a weighted average of the sampled checkpoint parameters), multi-objective Bayesian Optimization is used to optimize a set of weights that maximizes the Pareto frontier hyper-volume of a set of metrics (as well as the loss) of the target-task validation data. The authors demonstrate that this procedure results in a fine-tuned model that outperforms standard fine-tuning and other comparable model fusion techniques. Additionally, the authors provide an ablation study, showing the benefit of optimizing for multiple objectives in the model fusion weights, and Strengths: - While the individual components of the proposed method are not novel in and of themselves, the proposed method results in a new technique for LLM fine-tuning that appears to convincingly improve upon previous relevant methods. - The authors provide ablation studies for their method, motivating each step of their method. - The paper provides some novel and interesting insights regarding the alignment between loss and metric surfaces for NLP tasks, and highlights a flaw in loss-based aggregation and fusion methods in LLM fine-tuning, before proposing a solution. Weaknesses: - The 3 individual components of the proposed method are fairly distinct and disconnected from one another. As a result, it feels as though there are two distinct contributions that are not necessarily related (hyperparameter BO using lower capacity models, and BO for model fusion). - Because of this disconnect, one of these contributions (hyperparameter transfer using BO) feels less important than the other (Multi-Objective BO for model fusion), resulting in the MOBO for MF feeling under-explored as the "main contribution" (for instance, it would be interesting to see how BOMF works for "model soups", models obtained from different training runs). Instead, both contributions get comparatively equal attention in this work. - The presentation of the insights used to motivate the proposed method (both the misalignment between metric and loss surface, and the alignment of optimal hyperparameters over capacity) are only showed qualitatively, i.e. over a single dataset and model setting. It is not clear whether these findings generalize well across other NLP tasks, although it could be argued that the success of the proposed method indicates that they generalize well. - While there are limitations due to the size of the model, unless I am mistaken, all results are taken from a single run, e.g. results are not averaged over different random seeds and it is therefore less clear how meaningful some of the improvements that BOMF makes over it's baselines are. Technical Quality: 3 Clarity: 3 Questions for Authors: n/a Confidence: 4 Soundness: 3 Presentation: 3 Contribution: 4 Limitations: Yes Flag For Ethics Review: ['No ethics review needed.'] Rating: 8 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We appreciate your constructive feedback. We have answered your questions and concerns in this response. Please let us know if you have any follow-up questions. [W1, W2] It feels as though there are two distinct contributions that are not necessarily related. Because of this disconnect, one of these contributions feels less important than the other, resulting in the MOBO for MF feeling under-explored as the “main contribution”. To simplify the notations, we will refer to BO-based hyperparameter selection as `outer BO` and BO-based coefficient search as `inner BO`. The bottom line is that *the outer BO process and the inner BO process are coupled with each other, so they were not conducted for different purposes. As outlined in Section 1, BOMF aims to construct the best-performing single model through model fusion within a parameter space. As detailed in Section 4.2, different combinations of fine-tuning hyperparameters (such as learning rate and batch size) yield varied generalization performances after the fine-tuning process. Figure 3 demonstrates that the generalization performances of models before and after fusion align well. Therefore, to identify the best-performing single model after the model fusion, we must first determine the optimal hyperparameters that yield the best-performing single model before the fusion. For a simpler explanation, let us consider two fine-tuning hyperparameter configurations, H1 and H2. Figure 3 illustrates that if the performance of the fine-tuned solution from the training trajectory with H1 surpasses that of H2, then the performance after the model fusion with fusion members S1, sampled from the trajectory with H1, remains higher than that of the fusion members S2, sampled from the trajectory with H2. This finding, coupled with the non-differentiable nature of metric functions, underscores the necessity of identifying the best hyperparameter configurations through outer BO to achieve BOMF's ultimate objective with inner BO. [W3] Both the misalignment between metric and loss surface, and the alignment of optimal hyperparameters over capacity are only shown over a single dataset and model setting.* In our paper, we demonstrated that both full model fine-tuning and LoRA fine-tuning exhibit 1) metric and loss misalignment and 2) optimal hyperparameters alignment even when freeze layers or ranks change. However, we agree that extending these experiments to larger-scale models would strengthen our findings. Therefore, we conducted additional experiments on the Llama-2 model to verify if 1) and 2) still hold. Firstly, regarding point 2), the alignment of hyperparameters can be easily observed in Figures A.1. For point 1), the misalignment between loss and metrics is evident in Figure A.2. To further demonstrate the misalignment between loss and metrics, as well as among different metrics, we measured Spearman’s correlation using the loss and metrics of 20 sampled models. Also, we empirically demonstrate that BOMF significantly improves this correlation by considering both loss and metrics simultaneously. Please refer to Tables 9, 10, and 11 in Appendix C for full experimental results. We have included a portion of these results here. According to Table R.2 (c), (d), and (e), language tasks involving the Llama model show significantly lower correlation compared to vision tasks (a) and (b). Then, Table R.3 shows that BOMF successfully increases the correlation between loss and metrics compared to other baseline methods. [W4] All results are taken from a single run. Thank you for your insightful comment regarding the use of random seeds. We agree that running multiple experiments with different seeds is crucial for ensuring the reliability and completeness of the results. Accordingly, we will provide experimental results obtained using multiple seeds for each experiment. Due to computational limitations during the rebuttal period, we were only able to add results for the medium-sized language models in Table 1 in the main paper, using five seeds per method and dataset. This result can be found in Table R.6 and Table R.7. Additionally, we plan to conduct further experiments on large language models and update all tables with results from multiple seeds in the camera-ready version of this paper. --- Rebuttal Comment 1.1: Title: Rebuttal Response Comment: Thanks for the response and clarification. > the outer BO process and the inner BO process are coupled with each other I do not mean to suggest that the impact of one does not have an impact on the other, and I agree that the paper clearly demonstrates that if the performance of H1 is greater than H2, then fusion over S1 will be greater than fusion over S2. However, I disagree that they are "coupled" with each other in the sense that we could drop in some other HPO technique in place of the proposed BO procedure and, as long as it results in an effective hyper-parameter search, model fusion would still perform well. In other words, we want to find H2 to do BOMF, but we can find H2 without using BO for hyper-parameter observation. This is more of a stylistic concern however, as the BO for hyper-parameter is still a useful techniques that is backed by empirical observations, and I am not implying that this contribution is not valuable. Only that the paper makes them feel more entangled than they truly are. I think my other listed weaknesses have been addressed, and I appreciate the amount of effort put into the overall rebuttal, which contains results from many new experiments, particularly including results over many random seeds, which I greatly appreciate; I will raise my score to an 8. --- Reply to Comment 1.1.1: Comment: Thank you for the positive feedback on our paper. We will incorporate these discussions into the final manuscript and make the necessary revisions.	Summary: This paper introduces Bayesian Optimization Model Fusion (BOMF), a method for improving fine-tuning of pre-trained language models. BOMF addresses the challenge of selecting optimal models and hyperparameters by utilizing multi-objective Bayesian optimization to consider both loss and desired metrics during model fusion. It employs a two-stage approach, first optimizing hyperparameters for fine-tuning, then focusing on model fusion. Experiments across various NLP tasks demonstrate significant performance gains using BOMF, showcasing its effectiveness in both Natural Language Understanding and Generation. Strengths: Using Multi-Objective Bayesian Optimization (MOBO) that considers both metrics and loss functions for model fusion seems like an interesting idea. Weaknesses: Limited Performance Gain: The observed improvements from the proposed BOMF method are relatively small. Outdated Models and Datasets: The study relies on older base models (RoBERTa and T5) and simple datasets that may not reflect the current state-of-the-art in terms of model architectures, size, and task complexity. Incomplete Evaluation for Generative Tasks: The evaluation of generative tasks lacks the use of LLM-based evaluation or human evaluation, which are crucial for assessing the quality and diversity of generated text. Relying solely on computational metrics might not capture the nuances and potential biases of generated outputs. Questionable Approach: The use of model fusion to address the discrepancy between loss and desired metrics is not necessarily the most effective solution. Alternative approaches like Reinforcement Learning from Human Feedback (RLHF) and Direct Preference Optimization (DPO) can directly optimize towards desired metrics, potentially bypassing the need for fusion altogether. Technical Quality: 2 Clarity: 3 Questions for Authors: 1) what if the metrics to be optimized can not be directly computed? How to deal with complex evaluation criteria (e.g. grounding, coherence) commonly seen in LLM evaluation? 2) How the method perform in recent models such as Llama family? Confidence: 4 Soundness: 2 Presentation: 3 Contribution: 2 Limitations: Limited Performance Gain: The observed improvements from the proposed BOMF method are relatively small. Outdated Models and Datasets: The study relies on older base models (RoBERTa and T5) and simple datasets that may not reflect the current state-of-the-art in terms of model architectures, size, and task complexity. Incomplete Evaluation for Generative Tasks: The evaluation of generative tasks lacks the use of LLM-based evaluation or human evaluation, which are crucial for assessing the quality and diversity of generated text. Relying solely on computational metrics might not capture the nuances and potential biases of generated outputs. Questionable Approach: The use of model fusion to address the discrepancy between loss and desired metrics is not necessarily the most effective solution. Alternative approaches like Reinforcement Learning from Human Feedback (RLHF) and Direct Preference Optimization (DPO) can directly optimize towards desired metrics, potentially bypassing the need for fusion altogether. Flag For Ethics Review: ['No ethics review needed.'] Rating: 2 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We appreciate your constructive feedback. We have answered your questions and concerns in this response. Please let us know if you have any follow-up questions. [W1] The observed improvements from the proposed BOMF method are relatively small The improvement of BOMF is not negligible compared to other baselines. First, we would like to emphasize that our method consistently outperforms various baselines across a wide range of task types (i.e., classification, question answering, and the medical domain) and models (i.e., T5, LLaMA3). While the degree of improvement might appear marginal in some cases, it is essential to recognize that the extent of improvement, whether small or large, is relative. The fact that our method consistently achieves superior performance in nearly every case serves as an absolute metric that is universally recognizable. Our method consistently outperforms the baselines in these diverse scenarios. Specifically, BOMF achieved a 13%, 14%, 10% error improvement in the GLUE tasks, SQuAD task and KorMedQA tasks, respectively. Furthermore, the recent fusion research [1,2,3] in various scenarios has shown similar tendency to our results, showing that even if the improvement in specific tasks seems marginal, they exhibit overall superior performance compared to the baseline, proving the necessity of such methods. In particular, even when compared to these methods, our method demonstrates improved performance in almost all experimental results, indicating its effectiveness. [W2, Q2] The study relies on older base models (RoBERTa and T5) and simple datasets. How did the method perform in recent models such as the Llama family? We have already conducted experiments using Llama-2 and Llama-3 in Sec 6.2 of the main paper, and included the results for summarization, Korean multi-choice medical question answering, and dialogue generation. In most cases of these three experimental settings, our method outperforms other baseline methods. For detailed experimental setting and results, refer to Table 2 in the main text and Tables 15 and 16 in the appendices. [W3, Q1] The evaluation of generative tasks lacks the use of LLM-based evaluation or human evaluation. How to deal with complex evaluation criteria commonly seen in LLM evaluation? Thank you for suggesting additional evaluations on generative tasks. We would like to address this question from two major perspectives: 1. When a large language model (LLM) is used to evaluate the outputs of various models, BOMF demonstrates superior performance compared to the existing baselines. 2. BOMF can utilize more complex evaluation criteria such as human evaluation or LLM assessment in the process of model fusion. Regarding the first one, following your suggestion, we conducted an evaluation with a ChatGPT-turbo-3.5-based approach. This evaluation method involves generating scores by asking the LLM to assess the similarity between the generated responses and the ground-truth answers. Using this, we compared BOMF's performance with other baselines. Here, BOMF refers to the model optimized with the R1, R2, and RL metrics. As shown in Table R.1, even when optimized with this specific metric, BOMF outperformed other baselines. This indicates that our approach not only excels with traditional metrics but also adapts well to the latest evaluation methods. Additionally, the consideration of both loss and multiple metrics helps the model remain robust across various unseen metrics. Furthermore, since BOMF uses BO to optimize combination coefficients, it requires only the evaluation metric values corresponding to each set of combination coefficients to update them. This means that the optimization process does not rely on a backward process through the metric; as long as evaluation values are available, BOMF can optimize regardless of the complexity of the evaluation procedure. To illustrate this, we conducted optimization using metrics evaluated by LLMs, denoted as the column of ChatGPT BOMF in Table R.1. The result shows that BOMF can optimize the coefficients and achieve performance improvements even with these complex metrics. [W4] RLHF and DPO can directly optimize towards desired metrics, potentially bypassing the need for fusion altogether. We agree that methods like RLHF can be used to optimize desired metrics. However, we want to emphasize that our method complements rather than contradicts methods like RLHF or DPO. For instance, the hyperparameters used in RLHF or DPO can be optimized more efficiently using BO with a lower rank or fewer layers. Also, when performing RLHF or DPO, multiple checkpoints can be generated during training procedure. Typically, the best validation performance checkpoint is selected from these. However, with BOMF, these checkpoints can be fused to produce better results. By combining multiple checkpoints through BOMF, better generalization performance can be achieved compared to selecting a single checkpoint. This capability is well illustrated in our experiments, as shown in Tables 2 and 4. Another important point is that BOMF can consider multiple metrics simultaneously. While methods like RLHF or DPO require detailed preference settings for each metric in different scenarios, BOMF automatically balances various metrics and loss functions to find the optimal combination. To demonstrate that RLHF and BOMF can indeed be combined, we conducted additional experiments. According to Table R.2, performing various fusion methods on RLHF training checkpoints showed that BOMF outperformed other methods. References [1] Wortsman, et al. Model soups: averaging weights of multiple fine-tuned models improves accuracy without increasing inference time. ICML, 2022. [2] Weng, R., et al. G-tuning: Improving generalization of pre-trained language models with generative adversarial network. ACL, 2023. [3] Malladi, S., et al. Fine-tuning language models with just forward passes. NeurIPS 2023. --- Rebuttal Comment 1.1: Comment: Dear Reviewer, You have not responded to the authors yet. Your review is in very strong disagreement with the other 3 reviews. The points mentioned in both Weaknesses and Limitations are identical and quite generic. To ensure that this is not an LLM-generated review, you need to substantiate your decision and better explain it to the authors. Thank you, Your AC	Rebuttal 1: Rebuttal: We thank all reviewers for their valuable and constructive comments. The reviewers acknowledge that the paper is well-written, easy to follow, and recognize the originality of using Bayesian optimization for model fusion in BOMF (R-ut18, R-csBg, R-WU7w, R-JjKU). Reviewers also recognize the paper for providing a thorough study of this method, motivating each step of the approach, and offering new and interesting insights into the alignment between loss and metric surfaces for NLP tasks (R-csBg, R-WU7w, R-JjKU). Additionally, reviewers agree that the authors clearly demonstrate the quality of the paper through rigorous experimental design and comprehensive evaluation. They acknowledge that the authors test BOMF across various NLP tasks and models, proving its robustness and effectiveness (R-csBg, R-WU7w). All supplementary experimental outcomes and discussions will be incorporated into the final manuscript. Regarding the computational efficiency of BOMF In response to the reviewers' concerns regarding the computational efficiency of BOMF, we analyze these points as follows: The additional computational cost incurred by BOMF can be analyzed in two ways: 1) in comparison to basic fine-tuning and 2) relative to other fusion baseline methods. First, compared to basic fine-tuning, the computational cost during hyperparameter tuning in BOMF is similar to that of grid search in fine-tuning. Given that we used the same number of iterations for the BO step and the same number of grid points for the grid search method. While grid search can theoretically reduce search time through parallelization with sufficient GPUs, the increasing memory and computational demands of large language models make it challenging to perform parallel grid searches without a significant number of GPUs. Consequently, the time difference between BOMF and basic fine-tuning may not be substantial. Thus, the additional computational cost of BOMF mainly consists of the cost associated with calculating the combination coefficient. Specifically, during the MOBO step, if performed K times, the additional computational cost includes K evaluations of the validation set metrics and losses, as well as the computations required for Bayesian Optimization. When compared to other fusion methods, BOMF, like other baselines (excluding SWA), requires multiple forward passes to find the best combination coefficient. Methods like learned SWA, which optimize combination coefficients via backward passes, incur additional backward computation costs. Table R.5 shows the empirical results comparing the time required for combination coefficient optimization between BOMF and other fusion baselines. The experiments were conducted using the Llama-3 model on a Korean multi-choice medical question answering task to demonstrate the scenario with a large language model. Our additional experiments are included in the below PDF. Pdf: /pdf/29bc06ee1c12cb8415ef20792a9c106df1edb9f9.pdf	NeurIPS_2024_submissions_huggingface	2,024	null	null	null	null	null	null	null	null
Listenable Maps for Zero-Shot Audio Classifiers	Accept (poster)	Summary: This paper describes an extension of the LMAC method for explaining decisions made by audio classifiers. The novelty in the present article is to extend from fixed-vocabulary settings to zero-shot / open text description settings. To accomplish this, the authors propose a training objective that aims to preserve proximity of (learned) masked audio embeddings to embeddings of corresponding text descriptions. The method is evaluated on the CLAP model over two standard sound event detection datasets, and compares quite favorably to prior work along several evaluation criteria. Strengths: Overall, I found this paper to be well written and easy to follow. The topic of the paper is timely and relevant, as much audio analysis research does seem to be trending toward open vocabulary settings. The empirical evaluations are appropriate and generally convincing. The proposed method makes a lot of intuitive sense, and appears to be a natural extension of the fixed vocabulary setting (section 2.2). Weaknesses: The main weakness I see in this work is a somewhat shallow investigation of the proposed method itself, as opposed to high-level comparisons to prior work. The core technical contribution is described in equations 5-7, and consists of 3 terms: one to approximately preserve the embedding of audio (relative to text embedding) after masking, one to promote sparsity on the masks, and one to promote diversity of masks when conditioning on different text embeddings. As in CLAP, contrast is obtained by comparing amongst other samples within the training batch. This raises a handful of questions about the various components of the training objective which I will detail below; however, the critical issue here is that no ablation study is conducted to measure the importance of the various terms. This leads the reader at a bit of a loss to understand how these components interact with each other, whether they're all necessary, and so on. Technical Quality: 3 Clarity: 3 Questions for Authors: My questions primarily center around equation 6: - Notational quirk: $C_{i,j}$ is a scalar value (equation 5), right? Why are norm bars ($\\|$) used in eq 6 when computing the loss on approximating $C_{i,j}$? - Reasoning through the first term of eq6, a few things jump out at me. First, expanding out the definition in eq5, the summand should be equivalent to $\left\|t_i^\mathsf{T} \left(f_\text{audio}(X_{\text{audio},j}) - f_\text{audio}(M_\theta(t_i, h_j) \odot X_{\text{audio},j}) \right)\right\|$. That is, the gap between the original and masked audio embedding should be small (ie the mask doesn't do much) or approximately orthogonal to $t_i$, and this should hold for all $t_i$ in batch when the outer sum is computed. - What is the effect of the batch size / batch diversity on this loss? Presumably, the more (and more diverse) text embeddings we compute against, the harder it will be to achieve this orthogonality while having nontrivial masks. Similarly, a too-small batch could make it too easy to achieve this orthogonality in a high-dimensional embedding space while not producing useful masks. Is there a sweet spot or range sizes where this seems to work (on CLAP)? - What happens if the third term is left out of the loss? Similarly, how sensitive is this term to size and diversity of the batch? A couple of minor questions: - Line 155 describes the audio reconstruction by ISTFT. It's not clear how phase is treated here, since $X$ appears to be magnitude spectra. Presumably phase is retained from the STFT and propagated through, but if some other method is used here (eg griffin-lim or the like) it should be made explicit. - In general, the spectrogram visualizations (eg figure 4) are quite difficult to read in this paper. It would be helpful to A) label the axes with appropriate units, B) use a log-scale frequency axis for display (even if the underlying method uses linear spectra, the results will be more visible), and C) use a log amplitude scaling on the magnitudes to enhance visual contrast. Confidence: 4 Soundness: 3 Presentation: 3 Contribution: 2 Limitations: The limitations and impacts are well and appropriately described by the authors. Flag For Ethics Review: ['No ethics review needed.'] Rating: 5 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We thank you for your comments. Our replies are below: - Regarding your comment on the lack of ablation study regarding different terms of the loss function: We have conducted an ablation study (both quantitative and qualitative) to showcase the relevance of the unimodal diversity loss introduced in Eq. 7. - Regarding the notational quirk on $C_{ij}$ you make reference to: You are correct. We will remove the norm bars in Equation 6. - Regarding your comment on the first term in equation 6, and whether a trivial, all-ones mask would minimize this loss function: It definitely would. Because of this, the second term in Equation 6 is very important, and as we indicate in line 142 of the submission, this term is needed to avoid trivial solutions (all-ones mask) : `The second term in Equation 6 promotes sparsity in the generated mask to avoid trivial solutions`. - Regarding your comment, > What happens if the third term is left out of the loss? The third term in the loss function is responsible for making the decoder more sensitive to text prompts. To validate its impact on the generated explanations, we have carried out a qualitative and quantitative ablation study for this. Please see the qualitative results in the .pdf document attached for the rebuttal. The quantitative comparison indicates that by introducing the third loss term we marginally decrease the explanation’s faithfulness, while increasing the sensitivity to text prompts significantly (Fig 3 in the attached pdf). - > Similarly, how sensitive is this term (third term) to size and diversity of the batch? First of all, note that the proposed loss function is not contrastive. It is rather a loss term that matches two matrices. So, the behavior is expected to be more similar to regular loss functions. To quantify the effect of batch size on the end performance of the model, we have also trained a model with BS=4 on the AudioCaps dataset and reported the results in Table 5. We observe that with BS=4 in fact the faithfulness numbers are slightly worse compared to BS=2 (potentially due to suboptimal hyperparameters). - Regarding your comment, > Line 155 describes the audio reconstruction by ISTFT. It's not clear how phase is treated here, since appears to be magnitude spectra. Presumably phase is retained from the STFT and propagated through, but if some other method is used here (eg griffin-lim or the like) it should be made explicit. The phase of the original input audio is used to reconstruct the listenable audio for the explanation (same way it is done in the original L-MAC paper, equation 3.) We will clarify this in the final version of the paper. - Regarding your comments, > In general, the spectrogram visualizations (eg figure 4) are quite difficult to read in this paper. It would be helpful to A) label the axes with appropriate units, B) use a log-scale frequency axis for display (even if the underlying method uses linear spectra, the results will be more visible), and C) use a log amplitude scaling on the magnitudes to enhance visual contrast. Thank you for your comment. We will use a log-amplitude scaling on the magnitudes / log-scale frequency axis for display, and label the axes. --- Rebuttal Comment 1.1: Comment: Thanks for your responses - these address all of my questions above. --- Rebuttal 2: Comment: Thanks a lot for your positive feedback on our rebuttal! Please consider raising your score as it might help with the final decision. Thank you very much consideration.	Summary: The paper introduces a post-hoc interpretation method for zero-shot audio classifiers, named LMAC-ZS (Listenable Maps for Audio Classifiers in the Zero-Shot context). It addresses the challenge of interpreting predictions from zero-shot audio classifiers that define audio classes based on textual prompts, where labels are not predefined but generated dynamically. LMAC-ZS outputs saliency maps that highlight important regions within input audio, correlating these with corresponding text prompts. The method involves a novel loss function that maintains the original audio-text similarity, enhancing interpretability. Experiments were conducted using the CLAP model on standard datasets like ESC50 and UrbanSound8K. Strengths: 1. Motivation is Clearly articulated. It stresses the importance of interpretability in AI, particularly for models used in critical decision-making areas, focusing on the less-explored zero-shot audio classification. 2. The problem of providing interpretable explanations for zero-shot audio classifiers is clearly formulated and identified as a novel challenge in the field. 3. The paper compares LMAC-ZS with several baseline methods like GradCAM++, SmoothGrad, and Integrated Gradients, providing a thorough comparative analysis and justifying the selection based on their relevance and common use in related tasks. Weaknesses: 1. It is observed that in some scenarios LMAC-ZS (CT) performs better than LMAC-ZS (FULL) in metrics such as AI, AD, and AG. This is counterintuitive, as one would expect the model trained on the full CLAP dataset to perform better. An explanation for this discrepancy is needed. 2. The paper mentions exploring the training of LMAC-ZS only on the Clotho dataset to simulate a limited computational budget scenario. It raises the question of whether training on other individual datasets or comparing the performance across different single datasets versus the full dataset would yield different insights. This aspect needs further exploration and clarification. 3. In Table 2, the use of ESC50 contamination for ZS classification on the ESC50 dataset seems confusing. The rationale behind using the same dataset for contamination needs to be explained in detail. Similarly, in Table 3, the reason for using US8K contamination for Mel-Masking and ESC50 contamination for STFT-Masking in ZS classification on US8K requires clarification. Technical Quality: 3 Clarity: 3 Questions for Authors: 1. In Figure 7, the visual explanations for 'Toilet flushing' and 'Water drops' look very similar despite different similarity scores (Sim=0.69 and Sim=0.14). How does the model ensure distinguishable and meaningful explanations, and can more distinct examples be provided to demonstrate the method's effectiveness? 2. I have a question about a set of textual prompts. What happens if there are sounds that are not in the text? Confidence: 4 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: The author has addressed the limitations. Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We thank you for your constructive and positive comments. Our replies are below: - Regarding your comment as to the discrepancy between the L-MAC-ZS Full and L-MAC-ZS (CT) in some cases: This is potentially due to optimization, and choice of hyperparameters (note also that this is mainly observed for mel-masking and not STFT masking). From preliminary results on training L-MAC-ZS (Full) for few more epochs for mel-masking, and we see that the results seem to be on par. - Regarding your comment on training L-MAC-ZS with other subsets (subsets other than Clotho): In Tables 3 and 4 (the general reply above) we report the results obtained by training the decoder on different subsets. We experimented both with randomly sampled subsets of the whole CLAP dataset and with the individual datasets comprising the CLAP training data (namely AudioCaps, FSD50K, MACs, and Clotho). As mentioned above, we note that the explanation’s faithfulness is comparable when training the decoder on the full training data or a subset. We note that the training with only MACs dataset results in the lowest performance on ESC50. We note that this is likely related to the differences in data distributions. In fact, the MACs / ESC-50 Frechet Audio Distance (computed using CLAP embeddings) is the highest among all the subsets (Table 6 in the general rebuttal above). We conclude that training on a subset of the training data, the decoder can generate faithful explanations for the ZS classifier. - Regarding your comment: > In Table 2, the use of ESC50 contamination for ZS classification on the ESC50 dataset seems confusing. This setting refers to the case where we create sound mixtures using two recordings with different classes from ESC50. Same is true for the US8k dataset as well. The goal is to investigate whether L-MAC-ZS is able to point out the important parts of the recording in a more complicated audio recording. - Regarding your comment: > In Figure 7, the visual explanations for 'Toilet flushing' and 'Water drops' look very similar despite different similarity scores (Sim=0.69 and Sim=0.14). How does the model ensure distinguishable and meaningful explanations, and can more distinct examples be provided to demonstrate the method's effectiveness? We have provided additional qualitative examples in the .pdf document provided for the general rebuttal. - Regarding your question > I have a question about a set of textual prompts. What happens if there are sounds that are not in the text? In this case, as demonstrated in the example in Figure 4 and more generally Figure 3, L-MAC-ZS is able to return an empty mask when the similarity between the text prompt and the input audio is low. We have provided additional evidence to this in the pdf attached to the general rebuttal. --- Rebuttal Comment 1.1: Comment: Thank you for addressing the questions! I am satisfied with the responses and have no further questions at this time.	Summary: In this paper, the authors focus on interpreting the decisions of zero-shot audio classifiers, particularly ones based on the Contrastive Language-Audio Pretraining (CLAP) model. To achieve this, the authors propose to learn a decoder that predicts a "listenable" audio saliency map (a mask on the input spectrogram) from an audio-text pair. The decoder is trained with a novel loss function that remains faithful to the characteristics of CLAP features. The authors demonstrate the effectiveness of their approach with extensive experiments across various datasets. Strengths: ### Strengths 1. Although there is a significant body of research focusing on interpretability of audio classifiers, interpretability of zero-shot audio classifiers (and zero-shot classifiers in general) is an under-explored area, which certainly merits more investigation. 2. From a technical perspective, the proposed approach is sound and has no glaring weaknesses. 3. The qualitative and quantitative results show that the method indeed works as claimed. - Particularly the anonymized qualitative samples provided are highly appreciated. Weaknesses: ### Weaknesses 1. Even after reading two out of the three subsections in Section 2: Methodology, a reader learns nothing new from the paper. To be more specific: - Section 2.1 provides a high-level overview of "Contrastive Language Audio Pretraining" (CLAP) [10], while Section 2.2 gives a brief summary of the approach of the paper "Listenable Maps for Audio Classification" (L-MAC) [8]. Both these sections discuss the previous papers as is, without any new perspective. - It is solely Section 2.3 that pertains to the method proposed by the paper. Even in Section 2.3, the sub-subsection @line-152, "Producing Listenable Interpretations" is not something new achieved by the proposed method, but rather a feature of the previous L-MAC paper [8]. - Notably, [8] also has an explicit sub-section with the exact same title; Section 2.2: Producing Listenable Interpretations. The authors make no effort in lines 152-156 to clarify that the "listenable" feature comes from [8]; thus this is not only redundant, but also very close to plagiarism. - Suggestion: I strongly suggest that Sections 2.1, 2.2, and the sub-subsection (@lines 152-156) be discussed separately as a "Background" or "Preliminaries" section, instead of Methodology. Specifically, since the proposed approach builds significantly upon L-MAC [8], it is necessary to disentangle your contributions from that of the authors of [8]. 2. The technical novelty is limited by the L-MAC paper. - From my understanding, the only novelty in the method is the loss function in Equation 6. And even that is mostly changing the cross entropy-based objectives in L-MAC (Equation (2) in [8]) to contrastive losses characteristic of CLAP. Overall, the same min-max objective is retained, and the same regularization term is added for mask sparsity. The same general framework from L-MAC is followed: a decoder is trained on the same dataset as the encoder to predict a mask on the input audio spectrogram. 3. (minor weakness) Comparison against baselines. - The authors are comparing against the baselines of GradCAM, GradCAM++, SmoothGrad, Integrated Gradients. While all these methods were also used for comparison with the original L-MAC in [8], that was in a regular classification setting; these methods are, by nature, poorly suited to zero-shot settings. - For instance, for both "cat" and "glass-breaking", the CLAP model needs to "look" at the important regions of the audio to give an encoded audio feature---which can be observed in Figure 4 (for GradGAM) in the paper. - Suggestion: A better baseline would perhaps be from the paper "gScoreCAM: What objects is CLIP looking at?" [a]. The paper empirically establishes that for CLIP, in zero-shot settings, ScoreCAM performs better than the other forms of CAM. - Note: I understand that in general, there is a lack of baselines to compare to (for instance, listen-to-interpret and L-MAC are not applicable). I do not expect the authors to conduct any additional experiments. --- In general, the paper has a significant amount of similarity/redundancy with the L-MAC paper [8], from the naming of sections and content (noted in Weakness 1), figures (Fig 1 in [8] and fig 2 in current paper), and even the exact set of metrics posed in the identical order in Section 3.1, which makes it difficult to discern the novelty and contributions. --- [a] Chen, Peijie, et al. "gscorecam: What objects is clip looking at?." Proceedings of the Asian Conference on Computer Vision. 2022. Technical Quality: 3 Clarity: 1 Questions for Authors: ### Questions 1. line 147 "The intuition is that the similarity between two text prompts should be reflected in the similarity of the audio embeddings from the corresponding masked spectrograms". Equation (7) is the only part of the loss function that is distinctly novel, as it adds a third term to the original loss in [8]. Have you tried training the overall decoder without the third term? It would be helpful to the paper if you can show that adding the third term yields a noticeable improvement over just using the first two terms; otherwise, the hypothesis remains unsupported by evidence. 2. It is common to perform zero-shot classification with foundation models trained on a large dataset. Suppose we have one such hypothetical model, trained on a dataset with millions of samples. - Is it necessary to train the decoder on the same large dataset? If not, can an estimation be made as to what percentage of the training data the decoder needs to see to achieve reliable performance? (For instance, it may be the case that after seeing 20% of the data, the model achieves 80% of its full performance). - If it is necessary to train the decoder on the pre-training dataset, using LMAC-ZS becomes expensive in many cases. Can LMAC-ZS be jointly learned during the pre-training process of the base model (e.g. CLAP)? - Note that this may not be so straightforward, as the first term in the loss objective will try to match the decoder predictions to faulty entries of $C$ at the start of the training. ### General Suggestions - line-114; line-122: The authors note that they omitted a part of Equation (4) for brevity. I strongly suggest against doing this; clarity is more important than brevity when posing a loss function or optimization objective. Equation (4) is supposed to be a min-max objective; the goal is to maximize the classification confidence of the masked-in (salient) part of the audio, while minimizing the confidence of the masked-out part. Thus the whole equation should be written together, and parts of it should not be omitted. If brevity is desired, it may be achieved by abbreviating $\text{CrossEntropy}$ to $CE$ or $\mathcal{L}_{CE}$. - line 120: The abbreviation L-MAC is used, but it has not been defined previously. The authors should define the abbreviation on line 70. - line 62-63: The citation style is inconsistent. For example, in lines 59-60, it has been written "Key approaches in this category include [19, 20, 21, 22],". The same citation style can be followed in line 64: "Notable attempts in this vein include [23, 24, 25]." --- [b] Shimada, Kazuki, et al. "Zero-and Few-Shot Sound Event Localization and Detection." ICASSP 2024-2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2024. Confidence: 4 Soundness: 3 Presentation: 1 Contribution: 2 Limitations: ### Limitations - One limitation that I feel is not addressed is that the approach needs to be specifically trained on CLAP's data; it is not plug-and-play like Grad-CAM or similar approaches. - In line 216 It is mentioned that the decoder uses CNN14 layers, presumably because the audio part of CLAP is based on CNN14. - Now if we have another zero-shot audio classifier, LAION-CLAP, that (suppose) has the same dataset but a different transformer-like architecture, then the decoder may not be transferable. Another decoder architecture needs to be designed to suit the alternative zero-shot classifier. So for different zero-shot foundation models, it may become necessary to have different architectures, which can be a hurdle. - As noted earlier, for same architecture but different datasets, it is still needed to train the decoder again. Unless of course, the decoder from one pre-training dataset transfers to another dataset, as a general purpose audio method. Apart from these scalability concerns, I believe the remaining limitations are adequately addressed. --- Update: The rebuttal addresses most of my concerns, so I will raise my score to 6. Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We thank you for your comments. Our replies are below: - Regarding your comments on the organization of the paper, and the fact that the paper contains sections pertaining to L-MAC and CLAP: Our aim was to make this submission as self-contained as possible, and to provide the reader with the necessary preliminaries. We agree with your suggestion, and we will put Sections 2.1 and 2.2 under a new section called ‘Preliminaries’. We have also noted your comment on the subsection called ‘Producing Listenable Interpretations’, and we will clarify the overlaps with the original L-MAC paper. - Regarding your comment on limited novelty on top of L-MAC: This submission is an extension of the L-MAC framework to zero-shot classifiers, and hence it is called L-MAC-ZS. There are obvious similarities, like L-MAC, L-MAC-ZS is also a masking method that uses a decoder. However, beyond using a decoder network, using masking, and the fact that the masking is done in the STFT domain (to be able to produce listenable explanations) this paper addresses another, and more challenging problem compared to the original L-MAC paper. We would like to also state that, to the best of our knowledge, this is the first method that explores using a decoder based method to explain an audio foundation model like CLAP. Also, note that the loss function is significantly different from L-MAC. In L-MAC, the loss function tries to maximally retain logit similarity for mask-in, minimize the logit similarity for mask-out. In L-MAC-ZS however the interpreter is not trained with data with a fixed set of labels, and the loss function aims to match the original similarities between the text prompts and the audios in the batch, with the similarities obtained after masking. We also introduced a unimodal similarity term in Eq. 7, which we show is important for our results in this rebuttal (Tables 1 and 2 in the general rebuttal) . - Comparisons with baselines: We thank you for your suggestion to compare with gScoreCAM. We have added a comparison with gScoreCAM and ScoreCAM in Tables 1 and 2 of the general rebuttal. We observe that both gScoreCAM and ScoreCAM faithfulness scores are lower than L-MAC-ZS. - Regarding your question whether using the third term (given in equation 7) in the loss function is useful: We have conducted a qualitative (the rebuttal .pdf file) and quantitative ablation study (Tables 1 and 2 in the general rebuttal). We observe that by introducing the unimodal alignment loss term explanations become sensitive to the text prompt (Figure 3 in the pdf) while maintaining comparable faithfulness metrics (Tables 1 and 2). This is confirmed qualitatively as we note that explanations generated without the unimodal diversity term remain unchanged regardless of the provided text prompt. - Regarding your question on if we need to train the decoder on the same dataset as the foundation model to be interpreted (e.g. CLAP): Please note that we had partially investigated this in the submission. In Tables 1, 2, 3 of the original manuscript we have provided results with a decoder that had only been trained with Clotho (approximately ¼ of the whole training dataset). We see that results do not significantly drop. To further investigate this, we conducted an ablation study on the amount of data needed to train the decoder (Tables 3 and 4 in the general rebuttal). We experimented with randomly sampled subsets of the CLAP dataset and with the individual datasets comprising the CLAP training data (namely AudioCaps, FSD50K, MACs, and Clotho). As observed in the submission, we note that the explanation’s faithfulness is comparable when training the decoder on the full training data or a subset. The MACs-only training results in the lowest performance on ESC50. We note that this is likely related to the differences in data distributions. In fact, the MACs / ESC-50 Frechet Audio Distance (computed using CLAP embeddings) is the highest among all the subsets (Table 6). We conclude that training on a subset of the training data, the decoder can generate faithful explanations for the ZS classifier. - > If it is necessary to train the decoder on the pre-training dataset, using LMAC-ZS becomes expensive in many cases. Can LMAC-ZS be jointly learned during the pre-training process of the base model (e.g. CLAP)? In this paper we have not experimented with jointly training CLAP and L-MAC-ZS. However, this is an interesting future direction. Also, please note that in response to your earlier comment, it is also possible to train LMAC-ZS on a subset of the pre-training dataset. - Regarding your comment, > line-114; line-122: The authors note that they omitted a part of Equation (4) for brevity. I strongly suggest against doing this; clarity is more important than brevity when posing a loss function or optimization objective. We will include the full loss function in Equation 4. - Regarding your comments on citation style consistency, and undefined abbreviation in line 120: We will fix these issues in the final version of the paper. - Regarding your comment: > One limitation that I feel is not addressed is that the approach needs to be specifically trained on CLAP's data; it is not plug-and-play like Grad-CAM or similar approaches. and your related comments regarding generalizability to different foundation model architectures, and different datasets: Generalizability to different audio encoder architectures: There’s no theoretical reason as to why the convolutional decoder would not provide faithful and understandable explanations for different foundation model architectures. However, it also remains feasible to design new decoder architectures based for a different foundation model. Regarding different training datasets: As we mentioned earlier above, we have experimented with training the decoder with subsets of the training data, and we see that the results remain comparable. --- Rebuttal Comment 1.1: Comment: Thank you, my primary concerns have been addressed. I will raise my score. --- Rebuttal 2: Comment: Thank you for your detailed review, and your prompt response!	null	null	Rebuttal 1: Rebuttal: We thank all the reviewers for their comments. We provide a reply to each reviewer in the corresponding rebuttal. Here, we report the additional quantitative results obtained during the rebuttal period to address the reviewer's concerns. The reviewer replies refer to the table numbers listed below. Table 1. In-Domain Results on ESC-50, LMAC-ZS (CT) refers to LMAC-ZS trained on Clotho (with the diversity loss term), LMAC-ZS (CT) NoDiv refers to LMAC-ZS trained on Clotho without the diversity term (Eq. 7 in the paper). We only show the baselines that perform close to LMAC-ZS. \| Metric \| AI (↑) \| AD (↓) \| AG (↑) \| FF (↑) \| Fid-In (↑) \| SPS (↑) \| COMP (↓) \| MM \| \|-----------------------------\|--------\|--------\|--------\|--------\|------------\|---------\|----------\|------\| \| GradCam \| 20.30 \| 23.75 \| 7.77 \| 0.78 \| 0.58 \| 0.72 \| 11.54 \| 0.14 \| \| GradCam++ \| 32.50 \| 8.97 \| 7.95 \| 0.79 \| 0.84 \| 0.41 \| 12.41 \| 0.35 \| \| ScoreCAM \| 29.97 \| 12.14 \| 8.82 \| 0.70 \| 0.75 \| 0.32 \| 12.59 \| 0.41 \| \| GScoreCAM \| 29.64 \| 8.56 \| 6.62 \| 0.79 \| 0.84 \| 0.36 \| 12.52 \| 0.39 \| \| LMAC-ZS (CT) \| 37.40 \| 7.43 \| 11.26 \| 0.78 \| 0.86 \| 0.50 \| 12.29 \| 0.11 \| \| LMAC-ZS (CT) NoDiv \| 37.54 \| 6.38 \| 11.70\| 0.77 \| 0.88 \| 0.72 \| 11.59 \| 0.02 \| Table 2. Out-of-Domain Results on ESC50 Mixtures \| Metric \| AI (↑) \| AD (↓) \| AG (↑) \| FF (↑) \| Fid-In (↑) \| SPS (↑) \| COMP (↓) \| MM \| \|-----------------------------\|--------\|--------\|--------\|--------\|------------\|---------\|----------\|------\| \| GradCam \| 23.77 \| 25.25 \| 12.24 \| 0.69 \| 0.49 \| 0.69 \| 11.73 \| 0.17 \| \| GradCam++ \| 29.52 \| 14.84 \| 10.17 \| 0.70 \| 0.70 \| 0.39 \| 12.48 \| 0.35 \| \| ScoreCAM \| 31.39 \| 7.03 \| 7.05 \| 0.79 \| 0.87 \| 0.36 \| 12.52 \| 0.39 \| \| GScoreCAM \| 28.07 \| 13.74 \| 8.42 \| 0.70 \| 0.73 \| 0.32 \| 12.59 \| 0.41 \| \| LMAC-ZS (CT) \| 35.65 \| 12.23 \| 13.04 \| 0.69 \| 0.74 \| 0.53 \| 12.18 \| 0.09 \| \| LMAC-ZS (CT) NoDiv \| 37.35 \| 9.55 \| 13.53 \| 0.67 \| 0.79 \| 0.75 \| 11.46 \| 0.02 \| Table 3. Ablation of the amount of training data (the model is trained only on the indicated dataset) - In-Domain Setting (ESC50) \| Metric \| AI (↑) \| AD (↓) \| AG (↑) \| FF (↑) \| Fid-In (↑) \| SPS (↑) \| COMP (↓) \| MM \| \|-----------------------------\|--------\|--------\|--------\|--------\|------------\|---------\|----------\|------\| \| LMAC-ZS Clotho \| 37.40 \| 7.43 \| 11.26 \| 0.78 \| 0.86 \| 0.50 \| 12.29 \| 0.11 \| \| LMAC-ZS FSD50K \| 34.00 \| 8.33 \| 10.12 \| 0.77 \| 0.83 \| 0.61 \| 11.83 \| 0.04 \| \| LMAC-ZS AudioCaps \| 39.00 \| 5.93 \| 10.43 \| 0.78 \| 0.88 \| 0.68 \| 11.67 \| 0.07 \| \| LMAC-ZS MACs \| 15.61 \| 22.86 \| 5.32 \| 0.78 \| 0.61 \| 0.42 \| 12.42 \| 0.04 \| \| LMAC-ZS Subset (25%) \| 41.50 \| 3.48 \| 7.99 \| 0.79 \| 0.92 \| 0.65 \| 11.91 \| 0.22 \| \| LMAC-ZS Subset (50%) \| 43.70 \| 3.54 \| 7.86 \| 0.79 \| 0.91 \| 0.63 \| 11.97 \| 0.19 \| \| LMAC-ZS Subset (75%) \| 40.60 \| 4.74 \| 7.73 \| 0.79 \| 0.89 \| 0.66 \| 11.84 \| 0.17 \| \| LMAC-ZS CLAP Data \| 43.35 \| 4.29 \| 10.57 \| 0.78 \| 0.90 \| 0.65 \| 11.86 \| 0.10 \| Table 4. Ablation of the amount of training data (the model is trained only on the indicated dataset) - Out-of-Domain Setting (ESC50 Mixtures) \| Metric \| AI (↑) \| AD (↓) \| AG (↑) \| FF (↑) \| Fid-In (↑) \| SPS (↑) \| COMP (↓) \| MM \| \|-----------------------------\|--------\|--------\|--------\|--------\|------------\|---------\|----------\|------\| \| LMAC-ZS Clotho \| 35.65 \| 12.23 \| 13.04 \| 0.69 \| 0.74 \| 0.53 \| 12.18 \| 0.09 \| \| LMAC-ZS AudioCaps \| 35.97 \| 10.35 \| 11.42 \| 0.68 \| 0.76 \| 0.71 \| 11.63 \| 0.07 \| \| LMAC-ZS FSD50K \| 26.95 \| 16.26 \| 9.97 \| 0.67 \| 0.65 \| 0.66 \| 11.59 \| 0.03 \| \| LMAC-ZS MACS \| 11.38 \| 31.54 \| 4.42 \| 0.68 \| 0.38 \| 0.44 \| 12.41 \| 0.05 \| \| LMAC-ZS Subset (25%) \| 42.65 \| 5.99 \| 9.81 \| 0.70 \| 0.84 \| 0.66 \| 11.90 \| 0.20 \| \| LMAC-ZS Subset (50%) \| 39.47 \| 7.52 \| 9.05 \| 0.71 \| 0.81 \| 0.66 \| 11.88 \| 0.16 \| \| LMAC-ZS Subset (75%) \| 40.42 \| 7.07 \| 8.84 \| 0.70 \| 0.83 \| 0.68 \| 11.80 \| 0.16 \| \| LMAC-ZS CLAP Data \| 39.47 \| 8.28 \| 11.81 \| 0.69 \| 0.80 \| 0.67 \| 11.79 \| 0.09 \| Table 5. Ablation on the batch size \| Metric \| AI (↑) \| AD (↓) \| AG (↑) \| FF (↑) \| Fid-In (↑) \| SPS (↑) \| COMP (↓) \| MM \| \|-----------------------------\|--------\|--------\|--------\|--------\|------------\|---------\|----------\|------\| \| AudioCaps BS=4 OOD \| 27.15 \| 18.00 \| 10.10 \| 0.66 \| 0.63 \| 0.70 \| 11.75 \| 0.03 \| \| AudioCaps BS=2 OOD \| 35.97 \| 10.35 \| 11.42 \| 0.68 \| 0.76 \| 0.71 \| 11.63 \| 0.07 \| \| AudioCaps BS=4 ID \| 30.00 \| 12.03 \| 8.92 \| 0.76 \| 0.78 \| 0.64 \| 11.75 \| 0.06 \| \| AudioCaps BS=2 ID \| 39.00 \| 5.93 \| 10.43 \| 0.78 \| 0.88 \| 0.68 \| 11.67 \| 0.07 \| Table 6. Frechet Audio Distance between ESC-50 and the datasets, computed using CLAP representations \| Comparison \| FAD \| \|----------------------------\|-------\| \| MACs \| 3.33 \| \| FSD50K \| 3.04 \| \| Clotho \| 3.09 \| \| AudioCaps \| 3.11 \| \| CLAP Subset (25%) \| 3.18 \| Pdf: /pdf/c59ad9f06e6b11eb2bd764d57f018f861d52aa81.pdf	NeurIPS_2024_submissions_huggingface	2,024	null	null	null	null	null	null	null	null
From Alexnet to Transformers: Measuring the Non-linearity of Deep Neural Networks with Affine Optimal Transport	Reject	Summary: This paper proposes the affinity score, which measures the non-linearity of an activation function $\sigma(X)$ given the distribution of $X$. The affinity score is defined based on how well the 2-Wasserstein distance $W_2(X, Y)$, where $Y=\sigma(X)$, is approximated by $W_2(N_X, N_Y)$, where $N_X$ and $N_Y$ are Gaussian approximations of the distributions of $X$ and $Y$, respectively. Note that $W_2(N_X, N_Y)$ has a closed-form solution, and it holds that $W_2(X, Y) = W_2(N_X, N_Y)$ if the relation between $X$ and $Y$ is locally affine on the support of the given $X$. The authors then propose to characterize a DNN model by the set of affinity scores of activation functions in the model under a given input distribution. Experimental results suggest that the affinity scores are relatively low in transformer-based vision models, meaning that the activation functions are used in a more non-linear region compared to CNN models. Strengths: * The proposed score presents an interesting insight in comparing the series of CNN models and transformer-based models. Experiments suggest that transformer-based models utilize the non-linearity of activation functions more efficiently, leading to the higher prediction performance. Weaknesses: * It is empirically shown that the proposed score has a low correlation with existing non-linearity metrics such as R^2, but it is unclear whether the existing metrics are insufficient to analyze the DNN models in the way proposed in this paper. I would like to see how the distribution in Fig. 3(C) changes when other metrics such as R^2 are used instead of the proposed $\rho_{aff}$. * In my opinion, one would expect the nonlinearity score to behave symmetrically at $x=0$ for activation functions like ReLU, but the proposed affinity score seems to have a lower score at negative $x$, as shown in Fig.2 or Fig.6. Is there any reasonable explanation for such a behavior of the proposed score? Technical Quality: 3 Clarity: 4 Questions for Authors: * I could not fully understand the definition of the non-linearity signature in Def. 3.1. It is defined based on $F_i \cap \mathcal {A}$, but $F_i$ is a layer in N, so it seems to be empty. * This is just a comment out of interest, I would like to see if the affinity score of a neuron in the same model changes depending on the input class. Confidence: 2 Soundness: 3 Presentation: 4 Contribution: 3 Limitations: See weakness above. Flag For Ethics Review: ['No ethics review needed.'] Rating: 7 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We thank the reviewer for taking the time to review our work and for their comments. > I would like to see how the distribution in Fig. 3(C) changes when other metrics such as R^2 are used instead of the proposed $\rho_{aff}$. We thank the reviewer for this suggestion. We've added histograms for $R^2$ score similar to those presented in Figure 3 for the affinity score in the Fig.1 of the PDF included in the global rebuttal. We can see vividly that $R^2$ score is less informative, with several models having modes around negative values. This is in line with our Remark 3.6 explaining why such a metric may be inappropriate for the considered task. > one would expect the nonlinearity score to behave symmetrically at $x=0$ for activation functions like ReLU This is an interesting question, as it is a counterintuitive behavior. We can study this behaviour analytically by restricting to $X \sim \mathcal{U}[b,a]$, $b < 0$ and $a > 0$, $f: x \mapsto \mathrm{ReLU}(x)$, $Y = f(X)$ and verifying whether $\rho_{aff}(X, Y)$ is symmetrical in $a$ and $b$, that is, if it's invariant with respect to the assignment $(a,b)\mapsto (-b, -a)$. A straightforward computation yields - $\mu(X) = \frac{a+b}{2}$, $\Sigma(X) = \frac{(a-b)^2}{12}$, - $\mu(Y) =\frac{a^2}{2(a-b)}$ and $\Sigma(Y) = \frac{a^3(a-4b)}{12(a-b)^2}$. Plugging these in the affine transport formula yields - $A_{aff}=\frac{\sqrt{a^3(a-4b)}}{(a-b)^2}$ - $b_{aff} = \frac{a}{2(a-b)}\left(a - \sqrt{a(a-4b)}\left(\frac{a+b}{a-b}\right)\right)$. Finally, we can compute $W_2^2(T_{aff}(X), Y) = \frac{a^3}{6(a-b)^2}\left((a-4b) + \sqrt{a(a-4b)}\left(\frac{-a+3b}{a-b}\right)\right)$. Here we can see that this term is not strictly symmetric, but empirically we can assess that this term is not too far from being symmetric (see leftmost part of Figure 3 in the attached PDF). A greater source of asymmetry in the nonlinearity score $\rho_{aff}(X,Y) = 1 - \frac{W_2(T_{aff}(X), Y)}{\sqrt{2Tr(\Sigma(Y))}}$ is the trace term, as can be seen from the expression for $\Sigma(Y)$ and the middle plot of Figure 3. We illustrate this in the attached PDF in the global rebuttal. > I could not fully understand the definition of the non-linearity signature in Def. 3.1. It is defined based on $F_i \cap \mathcal{A}$, but $F_i$ is a layer in N, so it seems to be empty. We consider a layer $i$ as a function $F_i$ acting as a block used in a repeated way in a given architecture. In that sense, $F_i \cap \mathcal{A}$ would just be a set of activation functions in such block without all other eventual operations (such as for instance, convolutions). > Affinity score of a neuron in the same model depending on the input class. We thank the reviewer for this suggestion. We provide a visualization (Fig.4 of the attached PDF in the general reply) of the variance of the affinity scores when passing batches consisting of samples belonging to a single class of CIFAR10 dataset (so we have 10 batches, each batch contains samples from 1 class). For the sake of clarity, we use models with less than 40 layers. One can observe that early layers seem to be the most sensitive to the class distribution for all models. For ViT, some MLP block of intermediate layers also exhibit a high variability potentially hinting at their mono- and poly-semanticity. --- Rebuttal Comment 1.1: Comment: Thank you for your responses. From Figure 1 in the supplemental pdf, it appears that R^2 has a similar tendency to $\rho_{aff}$, e.g., the bimodality of ResNet152 and the less linear behavior of Vit Huge 14. As the author notes, it is curious that R^2 becomes negative. Figure 4 in the supplemental pdf is interesting, showing another difference in Conv nets and ViTs. I would like the authors to include these results in the main manuscript. I would like to keep my score at 7. --- Reply to Comment 1.1.1: Title: Thank you Comment: We thank the reviewer for their reply. We are happy that the reviewers found the provided illustrations insightful and we will include them in the new revision as suggested.	Summary: This study proposes empirical statistics about different DNN architectures in the hope to shed some light into why some architectures are better than others for some computer vision tasks. To do so, the study leverages common optimal transport results on DNN's internal representations, under some strong assumption about the distribution of those representations. Strengths: The paper proposes to consider an interesting and useful question of going to the bottom of why some architectures are better than others as measured by some restricted downstream task. Weaknesses: - I do not agree with the following statement `Without non-linear activation functions, 84 most of DNNs, no matter how deep, reduce to a linear function unable to learn complex patterns.` as to me, models such as transformers with linear attention and linear MLP blocks have no actual nonlinearity but are higher order polynomial of the input, i.e., are not linear. Could the authors provide clarifications on that statement or did I misunderstand something? - I also disagree with the following `Activation functions were also early identified [29, 30, 31, 32] as a key to making even a shallow 86 network capable of approximating any function, however complex it may be, to arbitrary precision.` since again, Fourier series for example can approximate any function as well. Hence DNN nonlinearities are certainly not the key ingredient to function approximation in general - Many formal results such as Theorem 3.3 are well known and have been established for years (even decades) but no reference is provided which is misleading to the reader. - Fig 2. is also misleading since the "nonlinearity" of any activation function depends on the range of the inputs. The only case that wouldn't be true is e.g. for ones with constant second derivatives, i.e., a linear activation function.... hence again that statement is highly misleading in presenting ReLU as inherently benefiting form that property compared to others - the statement `No other metric extracted from the activation functions of the 260 considered networks exhibits a strong consistent correlation with the non-linearity signature.` is again an overstatement as the authors only compare with a few alternatives and theorem is provided to support such a statement - the statement `We proposed the first sound approach to measure non-linearity of activation functions in neural 270 networks` is also incorrect, see e.g. - https://jmlr.org/papers/v20/18-418.html - https://arxiv.org/pdf/2301.09554 - https://arxiv.org/abs/1810.09274 all the above works have been published in peer reviewed journals/conferences Technical Quality: 2 Clarity: 2 Questions for Authors: Please see the Weaknesses section Confidence: 4 Soundness: 2 Presentation: 2 Contribution: 1 Limitations: In addition to my concerns expressed above, the study does not provide any actionable insights or understanding on the "why" of different architectures performing differently beyond the proposed statistical numbers. How could one use the provided analysis to better design model architectures or for model selection? Also, the paper does not provide any novel theoretical results. All the major theorems and results are already widely known within the community, yet they are presented as part of the contributions. With that in mind, the paper solely leverages existing OT tools, with some underlying simplifications on the DNN's data distribution, and report computed metrics. Hence the study falls below acceptance level in my opinion and would need a major rewriting + additional novel contributions to be worth acceptance. The writing style is also filled with unsupported claims and highly misleading statements (see the Weaknesses examples). Flag For Ethics Review: ['No ethics review needed.'] Rating: 4 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We thank the reviewer for taking the time to review our work and for their comments. > I do not agree with the following statement `Without non-linear activation functions, most of DNNs, no matter how deep, reduce to a linear function unable to learn complex patterns.` We believe that there is a misunderstanding regarding this claim. First, we believe that the presence of activation functions in all models available on Torchvision suggests that they are necessary to learn complex patterns. We are not aware of any DNN architecture that is widely acclaimed and that removes the activation functions altogether. Yet, we agree that for some models, such as transformers, even removing the activation function from attention and MLP blocks, doesn't make them linear due to the quadratic term in the attention equation. We propose to reformulate this phrase by saying: "For instance, without non-linear activation functions, popular multi-layer perceptrons, no matter how deep, reduce to a linear function unable to learn complex patterns." > I also disagree with the following `Activation functions were also early identified [29, 30, 31, 32] as a key to making even a shallow network capable of approximating any function, however complex it may be, to arbitrary precision.` We do not claim that "nonlinearity is the key ingredient to function approximation in general" as the reviewer says. Here, we refer to universal approximation theorems [29, 30, 31, 32] of feedforward networks that require at least one hidden layer and a non-linearity to be proved. If there are other universal approximation theorems that the reviewer has in mind and that can be proved for models without an activation function, we will gladly discuss them in our work and adjust the phrasing accordingly. > Many formal results such as Theorem 3.3 are well known We respectfully disagree with this comment as they seem to confuse the common results for OT between Gaussian distributions, duly cited in our work, and Theorem 3.3. Our statement holds for arbitrary (non-gaussian) random variables $X$ and $Y$ that satisfy $Y=T(X)$ for a symmetric positive-definite linear transformation $T$. The classical result about normal distributions is then a direct corollary: if $X$ and $Y$ are normally distributed, then such a $T$ can always be found, and the classical result follows. We also note that "Computational Optimal Transport" by Peyré and Cuturi (2019), which arguably represents a modern reference for applied OT doesn't contain any results having the generality of Theorem 3.3. Yet, it does refer to the classical result for OT between Gaussians in Remark 2.31 as well as its generalization to elliptical distributions in Remark 2.32. > Fig 2. is also misleading since the "nonlinearity" of any activation function depends on the range of the inputs. Kindly note that we illustrate exactly what the reviewer claims in Fig. 2 (A) for ReLU and in Appendix D Fig.6 for 8 more activation functions. In Appendix D, we explicitly say, "We observe that each activation function can be characterized by 1) the lowest values of its non-linearity obtained for some subdomain of the considered interval and 2) the width of the interval in which it maintains its non-linearity." (lines 514-516). We propose to clarify this in the caption of Fig. 2 (A) by saying "Non-linearity of activation functions depends on the range of input values (red), illustrated here on ReLU". >the statement No other metric extracted from the activation functions of the considered networks exhibits a strong consistent correlation with the non-linearity signature. is again an overstatement As explained in our work, we are not aware of any other quantifiable way to measure the non-linearity of an activation function. Yet, we considered reasonable alternatives to it: such as $R^2$ score, CKA commonly used to measure the similarity across the layers of NNs, and sparsity of activations recently used to better understand the semantic properties of neurons in MLPs and transformers. Our results show that the non-linearity signatures capture something different from these metrics, just as we claim it. We will gladly discuss this with the reviewer. > the statement We proposed the first sound approach to measure non-linearity of activation functions in neural networks is also incorrect None of the works mentioned by the reviewers provides a measure of non-linearity for activation functions. If we missed the introduction of such a metric upon reading the above-mentioned works, we would like to kindly ask the reviewer to point out the exact equations where it is defined in each of these works. > the study does not provide any actionable insights or understanding on the "why" Our experiments put forward several novel contributions factually explaining "why" different neural architectures differ from the lens of their activation functions' non-linearity. We show that DNNs of the last decade have several trends for which we do not find any mention in prior work: 1) Till 2020, state-of-the-art CNNs included more (on average) linear activation functions, 2) ViTs behave differently from CNNs with a much more non-linear behavior of the activation function in their MLP blocs, 3) skip connections help the model learn the (linear) identity function, as suggested by He et al. (2016). As highlighted in the Discussion section, the affinity score does provide actionable insights: it allows comparing neural architectures and identifying those that have a potential to disrupt the ML field, just as the transformers did. > the study leverages common optimal transport results on DNN's internal representations, under some strong assumption about the distribution of those representations. We stand by our claim expressed before and hope that our reply regarding Theorem 3.3 which holds for arbitrary random variables will clear out the misunderstanding that the reviewer insists on. --- Rebuttal Comment 1.1: Title: Thank you Comment: I want to thank the authors for providing clear answers to each of my concerns. In light of those findings and while considering the other reviews I have updated my score. I still believe that some improvements could be done in terms of presentation (which would have helped my initial assessment) and I am still unconvinced that the method allows for "actionable results". Also, I find that many insights are echoing previous studies and are not surprising (which is why I raised my score to 4 and not 5). --- Reply to Comment 1.1.1: Title: Thank you for your reply and reconsidering the score Comment: We would like to thank the reviewer for reconsidering their score. > I find that many insights are echoing previous studies If our study echoes other known results on this topic, we would gladly discuss and contextualize them in our work if the reviewer kindly provides us with more specific pointers to such studies. There seem to be no other metrics for measuring non-linearity as we propose to do it but we remain open to discovering other works deriving similar insights with other tools. We genuinely tried to find reasonably related works for this contribution but didn't manage to find them. > I am still unconvinced that the method allows for "actionable results" As for the providing actionable results, can the reviewer elaborate a bit more on what is considered actionable? We believe that our study provides comparisons between models in a theoretically-grounded way. It is thus actionable as it allows to quantify a quantity of interest and use this quantity in a variety of ways (to compare models, to understand the role of different parts of the network etc). We believe that such a comparison can be done in future works as well, acting as a tool in them. We remain open to discussing it further until the end of the rebuttal period if the reviewer shares with us they understand by actionable.	Summary: This paper introduces a novel method for quantifying the non-linearity of activation functions in neural networks, termed the "non-linearity signature." Using an affinity score derived from optimal transport theory, it measures the non-linearity of individual activation functions. It defines the non-linearity signature as a comprehensive set of these scores across all functions in a deep neural network (DNN). The study compares these signatures across a range of popular DNN architectures in computer vision, revealing clear patterns in their evolution over the past decade, notably showing a trend towards decreasing non-linearity until the disruptive impact of vision transformers. It emphasizes the uniqueness of their measure, as it does not strongly correlate with other metrics across different architectures. The approach could potentially be applied to analyze the non-linearity of newer large language models (LLMs) and identify innovative neural architectures that optimize internal non-linear characteristics for enhanced performance, crucial in the era of costly experiments with large-scale model optimizations. Strengths: 1. Novelty and importance. The paper introduces a theoretically grounded measure, the affinity score, for quantifying the non-linearity of activation functions using optimal transport theory, providing a robust framework for analysis. This is the first approach to approximately measure the non-linearity of DNNs, which is crucial for understanding their inner workings. 2. Solid theoretical and experimental validation. The method is grounded in optimal transport theory, providing a rigorous theoretical foundation for the proposed non-linearity signature. This enhances the credibility and robustness of the findings. The experimental results demonstrate the practical utility of the non-linearity signature. It can predict DNN performance and meaningfully identify the family of approaches to which a given DNN belongs, making it a valuable tool for researchers and practitioners. 3. Clear Writing. The structure of the paper is well-organized, with a clear presentation of background knowledge, theoretical properties, experimental evaluations, and conclusions. This clarity aids in understanding the contributions and implications of the research. The paper's figures and tables are comprehensive, providing clear and precise information, and the writing maintains a coherent logical sequence. Weaknesses: 1. The authors should discuss more activation functions. Currently, only ReLU, Tanh, and Sigmoid are included. While these are among the most commonly used activation functions in neural networks, many other activation functions have been introduced and shown to be effective in various contexts, like GELU. Including a more comprehensive analysis of a diverse set of activation functions would enhance the robustness and applicability of their proposed method. 2. There is currently some research on the nonlinearity of deep neural networks that should be compared and discussed. 3. It would be beneficial to showcase examples from domains beyond computer vision. While the paper focuses on computer vision tasks, it may not address the non-linearity signature's applicability to other domains such as NLP, speech recognition, or reinforcement learning. The findings might be less generalizable if the proposed measure does not perform equally well across diverse types of tasks and data. Technical Quality: 4 Clarity: 3 Questions for Authors: How would the model’s nonlinearity affect different downstream tasks? Does the impact vary across datasets? Confidence: 4 Soundness: 4 Presentation: 3 Contribution: 4 Limitations: Yes, they have discussed the assumption of Theorem 3.3. Flag For Ethics Review: ['No ethics review needed.'] Rating: 7 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We thank the reviewer for their positive comments about our work. > The authors should discuss more activation functions We thank the reviewer for this suggestion. We kindly note that the individual behavior of 9 different activation functions (Sigmoid, ReLU, GeLU, ReLU6, LeakyReLU, tanh, hardtanh, SiLU and Hardswish) is shown in Fig. 6 in Appendix D. Additionally, we also consider the original activation functions used for each neural architecture studied further in our work (GeLU for ViTs, Swin, ConvNext, and SiLU in EfficientNets). > There is currently some research on the nonlinearity of deep neural networks that should be compared and discussed. There was a recent paper by Teney et al. ("Neural Redshift: Random Networks are not Random Functions", CVPR, June 2024) that discusses the biases of the different activation functions for randomly initialized DNNs. Although it doesn't introduce a non-linearity measure for activation functions, it does illustrate the inductive bias of varying activation functions and their sensitivity to a change of weight magnitudes in Glorot initialization for randomly initialized MLPs. The authors of the latter work conclude that different activation functions visually carry different inductive biases (low complexity for ReLU, higher complexity for Tanh) that are more or less sensitive to weight magnitudes. Our work allows quantifying their findings using the affinity score (panel (a) of Fig.2 in the attached PDF). We also extend their work by showing that these biases can be changed by shifting the domain of the weight initialization (toward negative values) following the intuition provided in our work (panel (b) of Fig.2). This also affects the affinity score that becomes lower confirming our findings from the main manuscript. We are open to discussing any other relevant work we may have missed further. > It would be beneficial to showcase examples from domains beyond computer vision. We thank the reviewer for this suggestion. We considered the computer vision task, as the torchvision repository allows for a reproducible comparison of state-of-the-art models trained on Imagenet dataset. Additionally, the available models are very varied and span a decade of research. We agree that extending our analysis to other tasks mentioned by the reviewer is also very important, and we believe that our work can spur such contributions in the future. We will mention this in the Discussion. > How would the model’s nonlinearity affect different downstream tasks? Does the impact vary across datasets? We show in Appendix H and Fig. 15 the deviation of non-linearity signatures for the same model between ImageNet and different datasets (CIFAR10/CIFAR100/Random data). We can see that deviations between ImageNet and CIFAR10/CIFAR100 are much lower than for random data, implying that these datasets are processed similarly to Imagenet. Measuring the distance between the non-linearity signature of the same DNN for different datasets thus could be a promising way to measure semantic similarity between datasets with potential applications in meta- and transfer learning where this notion is of primary importance.	null	null	Rebuttal 1: Rebuttal: We thank the reviewers for their valuable feedback. We are glad to know that they found our work novel (Reviewer 2p6h), insightful (Reviewer 8vYn), our experiments solid (Reviewer 2p6h), and the writing clear (Reviewer 2p6h). Below, we summarize the additions that we present in the attached PDF following reviewers' questions. 1. We provide a comparison to a very recently published work of Teney et al. (CVPR'24): "Neural Redshift: Random Networks are not Random Functions" showing how the neural redshift phenomenon can be understood through the lens of the affinity score, although it doesn't introduce a non-linearity measure contrary to our work. The latter paper illustrates the inductive bias of different activation functions by passing an image formed by a 2D grid with values distributed uniformly in $[-1,1]^2$ interval and their sensitivity to a change of weight magnitudes in Glorot initialization for randomly initialized MLPs. The authors of the latter work then conclude that different activation functions visually carry different inductive biases (low complexity for ReLU, higher complexity for Tanh) that are more or less sensitive to weigh magnitudes (panel (a) of Fig.2). We quantify them with our affinity score and confirm quantitatively their findings. We also extend their work by showing that these biases can be changed by shifting the domain of the weight initialization (toward negative values) following the intuition provided in our work (panel (b) of Fig.2). This also affects the affinity score that becomes lower confirming our findings from the main manuscript. 2. Following the request of Reviewer 8vYn, we've added histograms for $R^2$ score in Fig.1 similar to those presented in Figure 3 for the affinity score. We can see vividly that $R^2$ score is less informative with several models having modes around negative values. This is in line with our Remark 3.6. 3. We added a plot (Fig.3) showing the evolution of the two terms defining the affinity score to explain its asymmetry. The idea behind it is to illustrate how the two evolve in a simple 1d case and when the affinity score approaches 0. A more detailed analytical analysis for this is given in the response to Reviewer 8vYn. 4. Following the request of Reviewer 8vYn, we added a plot (Fig.4 in the attached PDF) showing how the affinity scores vary in the different layers of several DNNs when samples from different classes of CIFAR10 dataset are passed through it. One can observe that early layers seem to be the most sensitive to the class distribution for all models. For ViT some MLP block of intermediate layers also exhibit a high variability. We hope that these additional experiments and the explanations provided to reviewers in their individual responses address their questions. We respectfully encourage the reviewers to endorse our paper if our replies are satisfactory to them. Pdf: /pdf/3cefcc82cb3dc0b1d17a1869c8bbdcda7ff37656.pdf	NeurIPS_2024_submissions_huggingface	2,024	null	null	null	null	null	null	null	null
Kangaroo: Lossless Self-Speculative Decoding for Accelerating LLMs via Double Early Exiting	Accept (poster)	Summary: This paper proposes a novel self-speculative decoding framework Kangaroo with a double early exiting strategy for accelerating LLM inference. It addresses the challenge of inference latency and shows effectiveness through extensive experiments, achieving significant speedups and outperforming the competitors with fewer parameters. Strengths: - The paper introduces a novel double early exiting strategy Kangaroo, which is a unique approach to improving the efficiency of LLMs without compromising performance. - The framework significantly accelerates inference by reducing the computational overhead associated with autoregressive decoding in LLMs, and achieves remarkable speedups compared to existing methods, with empirical results showing up to 2.04× wall-time speedups. - The proposed method can achieve lossless acceleration. The method maintains the same sampling distribution as the target LLM, ensuring that the quality of the generated text is not sacrificed for speed. Weaknesses: - This paper introduces extension to tree decoding for Kangaroo, but the implementation details are not fully provided. More explanation of that would be helpful for readers to understand the mechanism. - The optimal settings for the early exit layer and the dynamic threshold η may require careful tuning, which could be resource-intensive and may not generalize across different tasks or datasets. - This paper focuses on Vicuna-7B and Vicuna-13B models. More experiments on a wider range of model sizes could strengthen the evidence of Kangaroo's scalability and generalizability. Technical Quality: 3 Clarity: 3 Questions for Authors: See questions in the weaknesses. Confidence: 4 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: See questions in the weaknesses. Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We thank the reviewer for the detailed comments and valuable suggestions. We answer each of these comments below. ### More explanation of the extension to tree decoding for Kangaroo Thank you for your valuable suggestions. We will provide a more detailed formal expression and description of this extension in the final version. Kangaroo uses a static threshold to determine the timing of the second early stopping, based on the observation that the confidence of draft tokens in the small model is strongly correlated with their acceptance by the large model. Therefore, we approximate the probability of a token being accepted by the large model using the top-1 confidence of each token in the token tree. Considering the contextual dependency of tokens in speculative decoding, we model the probability of a token being accepted by the large model as the product of the top-1 confidences from the root node to that token. This approach is inspired by similar ideas in [1] > Sequoia: Scalable, Robust, and Hardware-aware Speculative Decoding. ### The optimal settings for the early exit layer and the dynamic threshold $\eta$ may require careful tuning Please refer to section 1.1 in the global response. ### More experiments on a wider range of model sizes. To demonstrate the generalizability of Kangaroo’s approach (reusing shallow parameters + learning a lightweight adapter), we conducted experiments on Llama2 and Llama3 as suggested by the reviewer. The results are as follows: \| $\qquad $ Method \| Translation \| $\qquad $ QA \| Summarization \| $\quad $ Math \| $\quad $ RAG \| MT Bench \| Avg. \| \| :-------------------: \| :------------------: \| :------------------: \| :------------------: \| :------------------: \| :------------------: \| :------------------: \| :------------: \| \| Draft & Verify \| $1.22\times (2.61)$ \| $1.02 \times (2.36)$ \| $1.13 \times (2.84)$ \| $1.08 \times (2.47)$ \| $1.15 \times (2.44)$ \| $1.12 \times (2.46)$ \| $1.12 \times $ \| \| SpS \| $1.26\times (1.63)$ \| $1.34\times (1.69)$ \| $1.14\times (1.47)$ \| $1.34\times (1.77)$ \| $1.32\times (1.81)$ \| $1.28\times (1.67)$ \| $1.28 \times $ \| \| Medusa$^$ w/o Tree* \| $1.53\times (1.85)$ \| $1.27\times (1.55)$ \| $1.19\times (1.48)$ \| $1.36\times (1.76)$ \| $1.25\times (1.54)$ \| $1.43\times (1.72)$ \| $1.34 \times $ \| \| Kangaroo w/o Tree \| $1.46 \times (1.97)$ \| $1.40 \times (1.87)$ \| $1.35 \times (1.97)$ \| $1.52 \times (2.22)$ \| $1.36 \times (2.05)$ \| $1.58 \times (2.28)$ \| $1.45 \times $ \| > Speedup comparison of various speculative decoding methods on Spec-Bench [22] for Llama2-13B-Chat. Values outside the parentheses indicate speedup, while those inside parentheses indicate the compression rate (CR). $^$ denotes reproduction result. SpS takes Llama-68M as the draft model. \| $\quad $ Model \| $\qquad$ Method \| Translation \| $\qquad $ QA \| Summarization \| $\quad $ Math \| $\quad $ RAG \| MT Bench \| Avg. \| \| :-----------------: \| :-----------------: \| :------------------: \| :------------------: \| :------------------: \| :------------------: \| :------------------: \| :------------------: \| :------------: \| \| Llama-3-8B-Instruct \| Kangaroo w/o Tree* \| $1.46\times (2.10)$ \| $1.49 \times (2.21)$ \| $1.47 \times (2.33)$ \| $1.61 \times (2.51)$ \| $1.38 \times (2.44)$ \| $1.64\times (2.44)$ \| $1.51 \times $ \| \| Llama-3-8B-Instruct \| Kangaroo \| $1.57 \times (2.32)$ \| $1.62 \times (2.43)$ \| $1.61 \times (2.61)$ \| $1.92 \times (2.95)$ \| $1.87 \times (2.85)$ \| $1.93 \times (2.87)$ \| $1.75 \times $ \| --- Rebuttal Comment 1.1: Title: After the rebuttals Comment: Thank the authors for your rebuttals carefully. My concerns about unclear details and experiments are well addressed. Thus, I keep my original rate to accept this paper. --- Reply to Comment 1.1.1: Title: Thanks Comment: We sincerely appreciate the time you took to provide valuable comments on our paper.	Summary: The authors introduce "Kangaroo" a novel self-speculative decoding framework designed to accelerate LLMs using a double early exiting strategy. This approach leverages the shallow sub-network and LM Head of the target LLM to construct a self-drafting model and employs a dynamic early exiting mechanism to enhance token acceptance rates and overall efficiency. The method shows promising results, achieving significant speedups and outperforming existing techniques with fewer additional parameters Strengths: 1. New Self-Speculative Decoding: The Kangaroo framework introduces a novel double early exiting strategy that effectively combines self-drafting and verification stages within the same model, significantly enhancing decoding efficiency. 2. Efficiency Gains: Experimental results demonstrate substantial walltime speedups over prev methods, with Kangaroo outperforming Medusa by using 88.7% fewer additional parameters. 3. Dynamic Drafting: The introduction of a dynamic early exiting mechanism tailored to both single-sequence and tree decoding scenarios ensures optimal balance between token acceptance rate and drafting efficiency. 4. Comprehensive Empirical Validation: Extensive experiments on Spec-Bench provide robust evidence of Kangaroo's superior performance across multiple tasks, including mathematical reasoning and retrieval-augmented generation. Weaknesses: 1. Adapter-Network Design: The design of the adapter network is heuristic and seems verified only on LLama2 and Vicuna models. More justification and exploration across different model architectures would enhance the generalizability of the approach. (new results for Gemma1~2, Phi3, LaMMa3, etc..) 2. Size of Shallow Network: The paper does not provide a rational basis for selecting the size of the shallow network. Clear criteria or experimental validation for this choice are needed. 3. Task-Specific Performance: While Medusa performs better on translation tasks, Kangaroo excels in mathematical reasoning. A deeper discussion on the strengths and weaknesses of each method across different tasks would provide more insights. 4. Training Time Justification: The justification for the chosen number of training epochs (10 epochs) compared to previous methods is unclear. A comparative analysis of training times and their impact on performance would be beneficial. 5. Speedup Verification: Speedup claims are tricky and should be verified over multiple GPUs to ensure robustness. More comprehensive benchmarking in diverse hardware settings is required. 6. Experimental Results with Temperature > 0: The paper lacks discussion on the impact of different temperature settings during inference. Additional experimental results with temperature > 0 would provide a more complete evaluation of the method. Technical Quality: 3 Clarity: 2 Questions for Authors: 1. Adapter-Network Design: - Question: Can you provide more detailed justification and exploration for the design of the adapter network? Specifically, how did you determine the architecture (one multi-head attention layer and two normalization layers) and its suitability across different model architectures? - Suggestion: Include additional experiments verifying the effectiveness of the adapter network on a broader range of LLMs beyond LLama2 and Vicuna. This could enhance the generalizability of your approach. 2. Size of Shallow Network: - Question: What criteria or experimental validations did you use to select the size of the shallow sub-network? Is there an optimal depth, and how does it vary with different models or tasks? - Suggestion: Provide a more rational basis or empirical analysis for the chosen size of the shallow network. Include ablation studies or sensitivity analyses showing the impact of varying the depth on performance and efficiency. 3. Task-Specific Performance: - Question: Why does Medusa perform better on translation tasks compared to Kangaroo, and how does Kangaroo's performance in mathematical reasoning differ? Can you elaborate on the strengths and weaknesses of each method across different tasks? - Suggestion: Include a more detailed discussion and comparative analysis of the performance of Kangaroo and other methods across various tasks. Highlight the specific attributes that make Kangaroo excel in mathematical reasoning and discuss potential improvements for translation tasks. 4. Training Time Justification: - Question: How did you justify the chosen number of training epochs (10 epochs) compared to previous methods? How does this training regimen impact efficiency and performance? - Suggestion: Provide a comparative analysis of training times and their impact on performance. Include benchmarks or references to previous methods to justify the training duration and its effectiveness. 5. Speedup Verification: - Question: Have you verified the speedup claims over multiple GPUs? How does the performance scale with different hardware configurations? - Suggestion: Conduct and report additional experiments verifying the speedup over multiple GPUs and diverse hardware setups. This would strengthen the robustness and reliability of your speedup claims. 6. Experimental Results with Temperature > 0: - Question: How does Kangaroo perform with different temperature settings during inference? Have you evaluated the impact of varying the temperature on token acceptance rates and overall efficiency? - Suggestion: Include experimental results and analysis with different temperature settings (e.g., temperature > 0) to provide a more comprehensive evaluation of the method. Discuss how temperature variations influence the performance and stability of Kangaroo. 7. Generalizability and Scalability: - Question: How does the proposed method generalize to even larger models and more diverse application scenarios? Have you explored its scalability and potential limitations in real-world deployments? - Suggestion: Extend your experimental evaluations to include larger models and a wider range of application scenarios. Discuss any potential scalability issues and provide insights into how the method can be adapted or improved for broader applicability. Confidence: 5 Soundness: 3 Presentation: 2 Contribution: 2 Limitations: See Weakness. Flag For Ethics Review: ['No ethics review needed.'] Rating: 4 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We thank the reviewer for the detailed comments and valuable suggestions. We answer these comments below. ### How did you determine the architecture and its suitability across different model architectures. The iterative process of designing the adapter network in Kangaroo is reflected in Table 2 of our paper, as follows: 1. Initial Approach with MLP: We initially considered using a simple MLP to map shallow features to the LM head. However, this approach did not yield satisfactory results as it failed to leverage token context effectively. 2. Transformer Layer: Next, we replaced the MLP with a single transformer layer to enhance expressive power. This improved performance but still fell short of our expectations. 3. Removing FFN Module: Observing that the FFN module in the LLM Decoder significantly contributes to latency without leveraging token context, we removed the FFN module from the transformer and introduced a new LM Head to maintain parameter count. This strategy increased the speedup from 1.37 to 1.44. 4. Reusing LLM’s Language Head: Finally, we found that reusing the LLM’s own language head was highly effective, significantly boosting acceleration performance. To demonstrate the generalizability of Kangaroo’s approach (reusing shallow parameters + learning a lightweight adapter), we conducted experiments on Llama2 and Llama3 as suggested by the reviewer. The results are as follows: \| $\qquad $ Method \| Translation \| $\qquad $ QA \| Summarization \| $\quad $ Math \| $\quad $ RAG \| MT Bench \| Avg. \| \| :-------------------: \| :------------------: \| :------------------: \| :------------------: \| :------------------: \| :------------------: \| :------------------: \| :------------: \| \| Draft & Verify \| $1.22\times (2.61)$ \| $1.02 \times (2.36)$ \| $1.13 \times (2.84)$ \| $1.08 \times (2.47)$ \| $1.15 \times (2.44)$ \| $1.12 \times (2.46)$ \| $1.12 \times $ \| \| SpS \| $1.26\times (1.63)$ \| $1.34\times (1.69)$ \| $1.14\times (1.47)$ \| $1.34\times (1.77)$ \| $1.32\times (1.81)$ \| $1.28\times (1.67)$ \| $1.28 \times $ \| \| Medusa$^$ w/o Tree* \| $1.53\times (1.85)$ \| $1.27\times (1.55)$ \| $1.19\times (1.48)$ \| $1.36\times (1.76)$ \| $1.25\times (1.54)$ \| $1.43\times (1.72)$ \| $1.34 \times $ \| \| Kangaroo w/o Tree \| $1.46 \times (1.97)$ \| $1.40 \times (1.87)$ \| $1.35 \times (1.97)$ \| $1.52 \times (2.22)$ \| $1.36 \times (2.05)$ \| $1.58 \times (2.28)$ \| $1.45 \times $ \| > Speedup comparison of various speculative decoding methods on Spec-Bench [22] for Llama2-13B-Chat. Values outside the parentheses indicate speedup, while those inside parentheses indicate the compression rate (CR). $^$ denotes reproduction result. SpS takes Llama-68M as the draft model. \| $\quad $ Model \| $\qquad$ Method \| Translation \| $\qquad $ QA \| Summarization \| $\quad $ Math \| $\quad $ RAG \| MT Bench \| Avg. \| \| :-----------------: \| :-----------------: \| :------------------: \| :------------------: \| :------------------: \| :------------------: \| :------------------: \| :------------------: \| :------------: \| \| Llama-3-8B-Instruct \| Kangaroo w/o Tree* \| $1.46\times (2.10)$ \| $1.49 \times (2.21)$ \| $1.47 \times (2.33)$ \| $1.61 \times (2.51)$ \| $1.38 \times (2.44)$ \| $1.64\times (2.44)$ \| $1.51 \times $ \| \| Llama-3-8B-Instruct \| Kangaroo \| $1.57 \times (2.32)$ \| $1.62 \times (2.43)$ \| $1.61 \times (2.61)$ \| $1.92 \times (2.95)$ \| $1.87 \times (2.85)$ \| $1.93 \times (2.87)$ \| $1.75 \times $ \| ### Provide a more rational basis or empirical analysis for the chosen size of the shallow network. Please refer to section 1.1 in the global response, where we conducted multiple comparative experiments across models with different architectures and sizes. It can be seen that deeper models (with depth $N$) have larger optimal early exit layers $\ell$, but the ratio $\ell / N$ is relatively consistent. This implies that the latency cost for the self-speculative small model compared to the full model is fairly stable. Therefore, we recommend setting $\ell / N$ between 1/16 and 1/10 in practical applications. Regarding whether different tasks should have different optimal early exit layers $\ell$, we summarized the changes in speedup across different datasets with varying early exit layers for Kangaroo@Vicuna-7B, as shown in the table below: \| Exit Layer $\ell$ \| Translation \| QA \| Summarization \| Math \| RAG \| MT Bench \| Avg. \| \| :---------------: \| :---------: \| :--: \| :-----------: \| :--: \| :--: \| :------: \| :--: \| \| 1 \| 1.10x \| 1.35x \| 1.23x \| 1.43x \| 1.31x \| 1.49x \| 1.32x \| \| 2 \| 1.24x \| 1.43x \| 1.50x \| 1.61x \| 1.52x \| 1.68x \| 1.50x \| \| 3 \| 1.19x \| 1.41x \| 1.53x \| 1.62x \| 1.49x \| 1.63x \| 1.47x \| \| 4 \| 1.12x \| 1.34x \| 1.47x \| 1.56x \| 1.44x \| 1.60x \| 1.43x \| \| 5 \| 1.11x \| 1.29x \| 1.39x \| 1.46x \| 1.37x \| 1.49x \| 1.35x \| It can be seen that the optimal early exit layers for different subtasks are not exactly the same but are generally close. ### Verifying the speedup over multiple GPUs and diverse hardware setups. We conducted inference on Vicuna-33B using 4 NVIDIA V100 GPUs, achieving a speed of approximately 7.6 tokens per second. With Kangaroo, using early exit at the 5th layer and training the adapter, the inference speed increased to 11.4 tokens per second, resulting in a speedup of about 1.5x. Sincerely, Paper 2124 Authors --- Rebuttal 2: Comment: Thank you for the detailed rebuttal, and I appreciate the effort in addressing the comments. However, I still have concerns and a few follow-up questions and suggestions: 1. Tree-Attention Experiments: - I noticed that your added results primarily focus on the non-tree attention variant of Kangaroo. Could you clarify why the tree attention variant was not included in the comparisons? My hypothesis is that adding tree attention might shift the computation from being `memory-bound` to more `compute-bound`, potentially diminishing some of Kangaroo's advantages. Could you provide insights or results regarding this? 2. Temperature Sampling Experiments: - The lack of experiments involving `temperature > 0` during inference remains a concern. In real-world applications, varying temperature settings are commonly used, and understanding how Kangaroo performs under these conditions is crucial. Could you elaborate on why these experiments were not conducted or share any preliminary results if available? 3. Interpretation of Ratio $ l/N $: - You mentioned that the optimal early exit layers for different subtasks are "generally close." However, looking at the results, it seems that other methodologies might actually perform better in certain scenarios. Doesn't this indicate that the differences are more significant than "generally close"? Could you clarify this point? 4. Task-Specific Performance: - The rebuttal tables highlight Kangaroo's strengths effectively, but I'm interested in understanding if there are scenarios or tasks where Kangaroo does not perform as well. Acknowledging these limitations is important, as no single method can excel in every task. It would be valuable to have a deeper discussion on where Kangaroo excels and where it might need further improvement. This would provide a more balanced and comprehensive perspective on the method's contributions. I will definitely raise your score once I receive answers to these questions. However, for now, I will lower your score because I still have additional doubts and concerns. --- Rebuttal Comment 2.1: Title: Discussion Inquiry Comment: Dear Reviewer J1nq, Thank you for your ongoing efforts in helping us improve the quality of this manuscript. We greatly appreciate the time and attention you have dedicated. We have responded to your latest comments. As the discussion period draws to a close, we would like to reach out to see if you have any remaining questions or unresolved issues. If everything is now clear, we would be grateful if you could consider updating your evaluation to reflect this. Once again, thank you for your constructive feedback and for your invaluable contribution to the development of this manuscript. We look forward to hearing from you soon. --- Rebuttal 3: Title: Rebuttal by Authors (Part 1/2) Comment: We have conducted experiments for the case where temperature > 0 and explored a comparative analysis of the performance of Kangaroo and other methods across various tasks. We hope that the following responses can address your concerns. ### Tree-Attention Experiments In most of the rebuttal tables, we only reported the non-tree attention variant of Kangaroo. This is because most of the reviewers’ concerns in the rebuttal focused on the robustness of Kangaroo's two hyperparameters (i.e., $\eta$ and $\ell$), which are not closely related to the use of tree attention. Therefore, in the comparison table for Llama2-13B-Chat, we compared Kangaroo and three other algorithms (Draft & Verify[1], SpS, Medusa-1) that also do not use tree attention. It's indeed true that adding tree attention might shift the computation from being memory-bound to more compute-bound. However, there is no evidence that tree attention might potentially diminish some of Kangaroo's advantages. To alleviate your concerns, we have summarized the performance of Kangaroo with and without tree attention across different model architectures: \| $\quad$ Model \| Tree \| Translation \| $\qquad $ QA \| Summarization \| $\quad $ Math \| $\quad$ RAG \| MT Bench \| Avg. \| \| :-------: \| :------------: \| :---------: \| :--: \| :-----------: \| :--: \| :--: \| :------: \| :--: \| \| Vicuna-7B \| False \| $1.24 \times (1.41)$ \| $1.43 \times (1.87)$ \| $1.50 \times (1.87)$ \| $1.61 \times (2.14)$ \| $1.52 \times (2.05)$ \| $1.68 \times (2.22)$ \| $1.50 \times $ \| \| Vicuna-7B \| True \| $1.43 \times (1.76)$ \| $1.71 \times (2.32)$ \| $1.68\times (2.31)$ \| $2.04 \times (2.76)$ \| $1.75 \times (2.37)$ \| $1.93 \times (2.67)$ \| $1.72 \times $ \| \| Llama2-13B-Chat \| False \| $1.46 \times (1.97)$ \| $1.40 \times (1.87)$ \| $1.35 \times (1.97)$ \| $1.52 \times (2.22)$ \| $1.36 \times (2.05)$ \| $1.58 \times (2.28)$ \| $1.45 \times $ \| \| Llama2-13B-Chat \| True \| $1.62 \times (2.23)$ \| $1.67 \times (2.07)$ \| $1.64 \times (2.15)$ \| $1.76 \times (2.53)$ \| $1.63 \times (2.37)$ \| $1.81 \times (2.59)$ \| $1.69 \times $ \| \| Llama-3-8B-Instruct \| False \| $1.46\times (2.10)$ \| $1.49 \times (2.21)$ \| $1.47 \times (2.33)$ \| $1.61 \times (2.51)$ \| $1.38 \times (2.44)$ \| $1.64\times (2.44)$ \| $1.51 \times $ \| \| Llama-3-8B-Instruct \| True \| $1.57 \times (2.32)$ \| $1.62 \times (2.43)$ \| $1.61 \times (2.61)$ \| $1.92 \times (2.95)$ \| $1.87 \times (2.85)$ \| $1.93 \times (2.87)$ \| $1.75 \times $ \| > The second column ``Tree`` denotes whether tree attention is used. Values outside the parentheses indicate speedup, while those inside parentheses indicate the compression rate (CR). ### Temperature Sampling Experiments In line 118 of the main paper, we claimed that in this work we focus on greedy decoding while our methodology can be easily extended to the rejection sampling case. To further address your concern, we followed the setting in Draft & Verify and set temperature = 0.2 to conduct two sets of comparative experiments on Kangaroo (without tree attention). One set used the original top-1 confidence as the variable to determine early stopping, while the other used adjusted top-1 confidence (softmax of the adjusted logits divided by the temperature). The results are shown in the table below: \| Temperature \| Confidence \| Translation \| $\qquad $ QA \| Summarization \| $\quad $ Math \| $\quad$ RAG \| MT Bench \| Avg. \| \| :-------: \| :------------: \| :---------: \| :--: \| :-----------: \| :--: \| :--: \| :------: \| :--: \| \| 0.0 \| Original \| $1.24 \times (1.41)$ \| $1.43 \times (1.87)$ \| $1.50 \times (1.87)$ \| $1.61 \times (2.14)$ \| $1.52 \times (2.05)$ \| $1.68 \times (2.22)$ \| $1.50 \times $ \| \| 0.2 \| Original \| $1.23 \times (1.41)$ \| $1.41 \times (1.88)$ \| $1.48\times (1.88)$ \| $1.58 \times (2.19)$ \| $1.50 \times (2.01)$ \| $1.67 \times (2.21)$ \| $1.48 \times $ \| \| 0.2 \| Adjusted \| $1.04 \times (1.45)$ \| $1.25 \times (1.90)$ \| $1.31 \times (2.00)$ \| $1.48 \times (2.32)$ \| $1.40 \times (2.18)$ \| $1.53 \times (2.41)$ \| $1.34 \times $ \| \| 0.5 \| Original \| $1.18 \times (1.41)$ \| $1.38\times (1.86)$ \| $1.43 \times (1.88)$ \| $1.55 \times (2.21)$ \| $1.46 \times (2.01)$ \| $1.62 \times (2.23)$ \| $1.43 \times $ \| From the analysis of the table, it can be seen that Kangaroo still achieves a speedup effect very similar to that of greedy decoding when $T = 0.2$. Moreover, the top-1 confidence used for the second early stopping mechanism should remain unadjusted because when using temperatures less than 1, the adjusted top-1 confidence tends to be overestimated, making it more difficult to trigger the early stopping mechanism (even if we use a large threshold of $\eta = 0.96$ for the adjusted confidence). Besides, the acceleration effect of Kangaroo decreases with an increase in sampling temperature. This is attributed to the increased computational complexity of the speculative sampling criterion at higher temperatures, as revealed in prior research[2]. --- Rebuttal 4: Title: Rebuttal by Authors (Part 2/2) Comment: ### Interpretation of Ratio $\frac{\ell}{N}$ By ``generally close``, we refer to cases where the early exit layer falls within our recommended robust range of $[1/16, 1/10]$. For the 32-layer Vicuna-7B, only the second and third layers meet this criterion. On Vicuna-7B, the best early exit layers for Summarization and Math tasks fall on the third layer, but their optimal speedup performance (1.53x and 1.62x) is generally close to that of the speedup performance using the second layer for early exit (1.50x and 1.61x). It's unfair to compare the performance of other methodologies against Kangaroo when it exits early at the fifth layer on Vicuna-7B. ### Task-Specific Performance After a thorough analysis, we could explain why Medusa performs better on translation tasks, while Kangaroo excels in mathematical reasoning, summarization, and retrieval-augmented generation from two perspectives: 1. The usage of an auto-regressive adapter. It should be noted that the draft tokens in Kangaroo are generated in an auto-regressive manner, while Medusa uses time-independent FFN heads. Since Kangaroo can better leverage long-distance contextual relationships, it excels in subtasks with high similarities between input and output, such as summarization, retrieval-augmented generation, and mathematical reasoning. For the translation subtask, the input and output are the least similar. 2. The depth of the feature used for drafting. Medusa uses second-to-top-layer features, while Kangaroo performs early exits at shallow layers and trains a lightweight adapter to generate draft tokens. As Reviewer Tuhh also pointed out [3,4], the shallow layers of LLM may learn relatively simple tasks, while deep-layer features learn more abstract capabilities. Therefore, for the more abstract translation task, Medusa shows higher accuracy, reflected in its compression rate (CR) being higher than Kangaroo's. On the other hand, Medusa's compression rate is lower than Kangaroo's in all other subtasks. > [1] Draft & Verify: Lossless Large Language Model Acceleration via Self-Speculative Decoding > > [2] Benjamin Spector and Chris Re. 2023. Accelerating LLM inference with staged speculative decoding > > [3] The Unreasonable Ineffectiveness of the Deeper Layers > > [4] Exploring Concept Depth: How Large Language Models Acquire Knowledge at Different Layers? We would like to express our sincere appreciation again for the valuable suggestions from the reviewer, which have greatly enriched the quality of this manuscript. Sincerely, Paper 2124 Authors --- Rebuttal Comment 4.1: Title: Anticipating your response Comment: Dear Reviewer J1nq, Sorry to bother you again. We appreciate the time and attention you have dedicated to this manuscript. As mentioned in your last comment, ``I will definitely raise your score once I receive answers to these questions. `` With only one day left in the discussion period, we are eager to hear your feedback on whether our recent response has addressed your concerns. If you have any remaining concerns, please do not hesitate to let us know; we are more than happy to clarify and respond. Engaging in this discussion with you has been a rewarding experience, and your feedback has significantly improved the quality of this manuscript. We look forward to your feedback. Best regards, Authors --- Rebuttal 5: Title: Author Response Comment: Dear Reviewer J1nq, We greatly appreciate your ongoing efforts in helping us improve the quality of this manuscript. We will add more baseline methods and conduct a comprehensive comparison for temperature sampling in the final version as suggested. We would like to point out that Medusa uses `typical sampling` when the temperature > 0, which is a lossy method that does not guarantee consistency with the original model’s sampling distribution. In contrast, Kangaroo is a lossless speculative decoding approach that exhibits a larger speedup ratio than Medusa. As claimed in Section 2.3.1 of Medusa[1], ``We ascertain that it is typically unnecessary to match the distribution of the original model. Thus, we propose employing a typical acceptance scheme to select plausible candidates rather than using rejection sampling.` > [1] MEDUSA: Simple LLM Inference Acceleration Framework with Multiple Decoding Heads By the way, as you mentioned, could you please adjust the score to 5 in your review? Sincerely, Paper 2124 Authors --- Rebuttal Comment 5.1: Comment: I updated the score. To all reviewers: Please respect the authors' efforts by providing feedback. Regardless of score changes, they put significant work into their rebuttals. If you're also an author, remember that we should all support each other by offering constructive comments. This applies to all submissions.	Summary: Authors proposed a new method to speed up large language model inference called self-speculative decoding. Instead of training a separate, costly draft model to maintain token acceptance rates, Kangaroo uses a shallow sub-network of the large model itself as the draft model. A lightweight adapter module is trained to bridge the representation gap between the sub-network and the full model. To improve efficiency, the method introduces an early exiting mechanism, halting predictions if the confidence level for the current token is too low. Experiments show that Kangaroo achieves up to 1.68× speedup with significantly fewer additional parameters compared to another method, Medusa-1. Strengths: * Compared to its prior work Medusa-1, the model is extremely lightweight (88.7% less parameter) but the performance is still outperforming. * The proposed auto-regressive self-drafting mode design is novel, and low cost. Their simple adapter network design sounds like a kick. Verification on a large language model can start where the self-drafting model has stopped and this can be beneficial as well. * Self-drafting models can share KV cache thus can be memory efficient. * Well written paper. Weaknesses: * It will be nice to evaluate more LLM families other than Vicuna. * Evaluation is not comprehensive. It misses evaluations on double-early exiting effect – see question Technical Quality: 4 Clarity: 3 Questions for Authors: * How frequently does the top-1 probability of small model goes below the predefine threshold? Do you have some observations on the efficacy of double early exiting? Confidence: 3 Soundness: 4 Presentation: 3 Contribution: 3 Limitations: * Requires some hyperparameters search (e.g., exit-layer l) Flag For Ethics Review: ['No ethics review needed.'] Rating: 5 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We thank the reviewer for the detailed comments and answer each of these comments below. ### Evaluate more LLM families other than Vicuna To validate the robustness of Kangaroo across different architectures, we additionally trained adapter networks on Llama2-13B-Chat and Llama-3-8B-Instruct and evaluated the performance on Spec-Bench. The results are shown in the tables below: \| $\qquad $ Method \| Translation \| $\qquad $ QA \| Summarization \| $\quad $ Math \| $\quad $ RAG \| MT Bench \| Avg. \| \| :-------------------: \| :------------------: \| :------------------: \| :------------------: \| :------------------: \| :------------------: \| :------------------: \| :------------: \| \| Draft & Verify \| $1.22\times (2.61)$ \| $1.02 \times (2.36)$ \| $1.13 \times (2.84)$ \| $1.08 \times (2.47)$ \| $1.15 \times (2.44)$ \| $1.12 \times (2.46)$ \| $1.12 \times $ \| \| SpS \| $1.26\times (1.63)$ \| $1.34\times (1.69)$ \| $1.14\times (1.47)$ \| $1.34\times (1.77)$ \| $1.32\times (1.81)$ \| $1.28\times (1.67)$ \| $1.28 \times $ \| \| Medusa$^$ w/o Tree* \| $1.53\times (1.85)$ \| $1.27\times (1.55)$ \| $1.19\times (1.48)$ \| $1.36\times (1.76)$ \| $1.25\times (1.54)$ \| $1.43\times (1.72)$ \| $1.34 \times $ \| \| Kangaroo w/o Tree \| $1.46 \times (1.97)$ \| $1.40 \times (1.87)$ \| $1.35 \times (1.97)$ \| $1.52 \times (2.22)$ \| $1.36 \times (2.05)$ \| $1.58 \times (2.28)$ \| $1.45 \times $ \| > Speedup comparison of various speculative decoding methods on Spec-Bench [22] for Llama2-13B-Chat. Values outside the parentheses indicate speedup, while those inside parentheses indicate the compression rate (CR). $^$ denotes reproduction result. SpS takes Llama-68M as the draft model. \| $\quad $ Model \| $\qquad$ Method \| Translation \| $\qquad $ QA \| Summarization \| $\quad $ Math \| $\quad $ RAG \| MT Bench \| Avg. \| \| :-----------------: \| :-----------------: \| :------------------: \| :------------------: \| :------------------: \| :------------------: \| :------------------: \| :------------------: \| :------------: \| \| Llama-3-8B-Instruct \| Kangaroo w/o Tree* \| $1.46\times (2.10)$ \| $1.49 \times (2.21)$ \| $1.47 \times (2.33)$ \| $1.61 \times (2.51)$ \| $1.38 \times (2.44)$ \| $1.64\times (2.44)$ \| $1.51 \times $ \| \| Llama-3-8B-Instruct \| Kangaroo \| $1.57 \times (2.32)$ \| $1.62 \times (2.43)$ \| $1.61 \times (2.61)$ \| $1.92 \times (2.95)$ \| $1.87 \times (2.85)$ \| $1.93 \times (2.87)$ \| $1.75 \times $ \| ### How frequently does the top-1 probability of small model goes below the predefine threshold? Do you have some observations on the efficacy of double early exiting? The frequency with which the small model's top-1 confidence falls below the predefined threshold varies depending on the model and the evaluation dataset. For instance, in Figure 4 of the main paper, about 40% of tokens have a confidence lower than the threshold $\eta$, but most of these tokens are "hard tokens" that would not be accepted by the large model. More confidence distribution graphs are available in the Figures in the global response PDF. The purpose of the second early exit mechanism is to reduce the time spent on difficult tokens, allowing the model to speculate further on simpler token sequences. As shown in Figure 6(b), when the stride $\gamma = 6$ but the dynamic early exit mechanism is not used ($\eta = 0$), the speedup is 1.2. When $\eta = 0.6$, the speedup increases to 1.5, indicating that the second early exit mechanism can provide approximately a 25% performance improvement. ### Requires some hyperparameters search (e.g., exit-layer l) We argue that in practical applications, the early exit layer $\ell$ can be set based on an empirical ratio derived from the target model's depth. To determine the optimal depth for shallow sub-networks relative to the full model depth, we conducted multiple comparative experiments across models with different architectures and sizes. The table below records the average speedup achieved on Spec-Bench with early exits at various depths, and the final column shows the optimal ratio of the early exit layer to the total model depth. \| $\qquad $ Model \| $\quad $ Group 1 \| $\quad $ Group 2 \| $\quad $ Group 3 \| $\quad $ Group 4 \| $\quad $ Group 5 \| Optimal Ratio \| \| :-----------------------: \| :------------: \| :------------: \| :------------: \| :------------: \| :-------------: \| :-------------: \| \| Vicuna-7B @32 layer \| 1.32x @1 layer \| 1.50x @2 layer \| 1.47x @3 layer \| 1.43x @4 layer \| 1.35x @5 layer \| 2 / 32 = 0.0625 \| \| Vicuna-13B @40 layer \| 1.26x @1 layer \| 1.36x @2 layer \| 1.44x @3 layer \| 1.34x @4 layer \| 1.31x @5 layer \| 3 / 40 = 0.075 \| \| Vicuna-33B @60 layer \| - \| 1.44x @4 layer \| 1.49x @5 layer \| 1.38x @6 layer \| - \| 5 / 60 = 0.083 \| \| Llama2-13B-Chat @40 layer \| 1.18x @1 layer \| 1.42x @2 layer \| 1.45x @3 layer \| 1.39x @4 layer \| 1.27x @5 layer- \| 3 / 40 = 0.075 \| It can be seen that deeper models (with depth $N$) have larger optimal early exit layers $\ell$, but the ratio $\ell / N$ is relatively consistent. This implies that the latency cost for the self-speculative small model compared to the full model is fairly stable. Therefore, we recommend setting $\ell / N$ between 1/16 and 1/10 in practical applications. --- Rebuttal 2: Title: Looking forward to your feedback Comment: Dear Reviewer oUCN, With the discussion phase nearing the end, we would like to know whether the responses have addressed your concerns. Should this be the case, we are encouraged that you raise the final rating to reflect this. If there are any remaining concerns, please let us know. We are more than willing to engage in further discussion and address any remaining concerns to the best of our abilities. We are looking forward to your reply. Thank you for your efforts in this paper. Best regards, Authors --- Rebuttal 3: Comment: I appreciate the authors' feedback and acknowledge the generalizability of the method as well as the impact of its various hyperparameters. I will maintain my original rating for now but am open to revising the score if necessary during the reviewer-AC discussion phase. --- Rebuttal Comment 3.1: Title: Thanks Comment: Dear Reviewer oUCN, Thank you for your continued efforts in helping us improve the quality of this manuscript. We greatly appreciate the time and attention you have dedicated to this process. Sincerely, Paper 2124 Authors	Summary: The paper presents Kangaroo, a novel self-speculative decoding framework designed to accelerate the inference of LLMs while maintaining an identical sampling distribution: a new speculative decoding method: self-speculative decoding method using the model's own subnetwork as the speculative small model. The drafting stage is dynamic based on the confidence level for the current step. To fully utilize GPU, it proposes two decoding method: single-sequence and tree decoding for the verification phase. Extensive experiments are done to demonstrate the effectiveness of the method. Strengths: 1. This is a pretty novel speculative decoding method, smartly utilizing self-drafting by the model's own layers. It resolves both the issue of having to accommodate another drafting model by reducing extra parameters needed and the accuracy of the self-drafted model. 2. Extensive experiments on multiple tasks and baselines 3. The proposed tree decoding method to enhance parallelism and improve GPU utilization is also pretty interesting 4. Clear presentation and explanation Weaknesses: 1. This paper seems largely similar to: Draft & Verify: Lossless Large Language Model Acceleration via Self-Speculative Decoding. What is the main difference? Technical Quality: 3 Clarity: 4 Questions for Authors: Is the optimal layer dataset specific or domain specific? The early exit strategy adopted in self-drafting here reminds me of depth of learnt knowledge/concept in LLM [1, 2]. Thus in ablation study on the depth of shallow sub-network, I wonder how general the results can be seen? Also, is it also possible that the early exit layer also becomes dynamic? [1] The Unreasonable Ineffectiveness of the Deeper Layers [2] Exploring Concept Depth: How Large Language Models Acquire Knowledge at Different Layers? Confidence: 4 Soundness: 3 Presentation: 4 Contribution: 3 Limitations: N/A Flag For Ethics Review: ['No ethics review needed.'] Rating: 7 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We thank the reviewer for the detailed comments and answer each of these comments below. ### The main difference between Kangaroo and Draft & Verify Both Draft & Verify and Kangaroo are self-speculative decoding algorithms that construct a small model by reusing the parameters of the original model. However, the starting points differ: Draft & Verify aims to approximate the full model by removing redundant middle layers, leaving a small model with about half the layers of the original model. This creates a trade-off between the small model's accuracy and inference latency. In contrast, Kangaroo is designed to minimize the layers of shared parameters while maintaining small model accuracy. Therefore, Kangaroo reuses only the initial shallow layers of the LLM, which are crucial for feature extraction [3]. Additionally, the lightweight adapter in Kangaroo is designed to enhance the small model's accuracy without significantly impacting latency. > [3] Tang Y, Liu F, Ni Y, et al. Rethinking optimization and architecture for tiny language models[J]. ICML, 2024. > > For the benefits of reusing the initial layers, please also refer to the question "Why starting from shallow layers works better?" raised by Reviewer NQcw. ### Is the optimal layer dataset specific or domain specific? Thank you for pointing out this insightful question and for the valuable references. According to [2], LLMs learn different difficulty levels of tasks at different layers. This raises the question of whether the optimal early exit layer varies across datasets (domains). To address this, we summarized the changes in speedup across different datasets with varying early exit layers for Kangaroo@Vicuna-7B, as shown in the table below: \| Exit Layer $\ell$ \| Translation \| QA \| Summarization \| Math \| RAG \| MT Bench \| Avg. \| \| :---------------: \| :---------: \| :--: \| :-----------: \| :--: \| :--: \| :------: \| :--: \| \| 1 \| 1.10x \| 1.35x \| 1.23x \| 1.43x \| 1.31x \| 1.49x \| 1.32x \| \| 2 \| 1.24x \| 1.43x \| 1.50x \| 1.61x \| 1.52x \| 1.68x \| 1.50x \| \| 3 \| 1.19x \| 1.41x \| 1.53x \| 1.62x \| 1.49x \| 1.63x \| 1.47x \| \| 4 \| 1.12x \| 1.34x \| 1.47x \| 1.56x \| 1.44x \| 1.60x \| 1.43x \| \| 5 \| 1.11x \| 1.29x \| 1.39x \| 1.46x \| 1.37x \| 1.49x \| 1.35x \| It can be seen that the optimal early exit layers for different subtasks are not exactly the same but are generally close. ### Is it also possible that the early exit layer also becomes dynamic? Early exiting methods often exit at different layers based on the difficulty of the samples [4]. These methods save inference costs by exiting earlier for easier samples. However, in self-speculative inference, the small model only needs to align with the large model on simple tokens. Pursuing higher accuracy by exiting at deeper layers increases the small model's inference cost. Therefore, we chose to exit at a fixed shallow layer to balance inference cost and framework simplicity. Nevertheless, considering different exit layers within the shallow sub-network for different tokens is a valuable direction for future research. > [4] Teerapittayanon S, McDanel B, Kung H T. Branchynet: Fast inference via early exiting from deep neural networks[C]//2016 23rd international conference on pattern recognition (ICPR). IEEE, 2016: 2464-2469. Sincerely, Paper 2124 Authors --- Rebuttal 2: Title: Looking forward to your feedback Comment: Dear Reviewer Tuhh, With the discussion phase nearing the end, we would like to know whether the responses have addressed your concerns. Should this be the case, we are encouraged that you raise the final rating to reflect this. If there are any remaining concerns, please let us know. We are more than willing to engage in further discussion and address any remaining concerns to the best of our abilities. We are looking forward to your reply. Thank you for your efforts in this paper. Best regards, Authors	Rebuttal 1: Rebuttal: We thank all the reviewers for their insightful comments and valuable feedback. We are encouraged by the positive reception of Kangaroo, as reflected in comments like ``The proposed method is well-motivated and the proposed token-level early exiting mechanism is interesting`` from Reviewer NQcw, ``This is a pretty novel speculative decoding method, smartly utilizing self-drafting by the model's own layers.`` from Reviewer Tuhh and ``Their simple adapter network design sounds like a kick. `` from Reviewer oUCN. Below, we briefly summarize our response to the common concerns raised by reviewers. ### 1. Robustness and sensitivity of Kangaroo's hyperparameters across different models Both Reviewers xb3n and oUCN raise concerns about the optimal early exit layer $\ell$ and the early-stopping threshold $\eta$ potentially requiring careful tuning. We argue that in practical applications, the early exit layer $\ell$ can be set based on an empirical ratio derived from the target model's depth, while the optimal threshold $\eta$ is robust across different model architectures and datasets. ### 1.1 The choice of the exit layer $\ell$ To determine the optimal depth for shallow sub-networks relative to the full model depth, we conducted multiple comparative experiments across models with different architectures and sizes. The table below records the average speedup achieved on Spec-Bench with early exits at various depths, and the final column shows the optimal ratio of the early exit layer to the total model depth. \| $\qquad $ Model \| $\quad $ Group 1 \| $\quad $ Group 2 \| $\quad $ Group 3 \| $\quad $ Group 4 \| $\quad $ Group 5 \| Optimal Ratio \| \| :-----------------------: \| :------------: \| :------------: \| :------------: \| :------------: \| :-------------: \| :-------------: \| \| Vicuna-7B @32 layer \| 1.32x @1 layer \| 1.50x @2 layer \| 1.47x @3 layer \| 1.43x @4 layer \| 1.35x @5 layer \| 2 / 32 = 0.0625 \| \| Vicuna-13B @40 layer \| 1.26x @1 layer \| 1.36x @2 layer \| 1.44x @3 layer \| 1.34x @4 layer \| 1.31x @5 layer \| 3 / 40 = 0.075 \| \| Vicuna-33B @60 layer \| - \| 1.44x @4 layer \| 1.49x @5 layer \| 1.38x @6 layer \| - \| 5 / 60 = 0.083 \| \| Llama2-13B-Chat @40 layer \| 1.18x @1 layer \| 1.42x @2 layer \| 1.45x @3 layer \| 1.39x @4 layer \| 1.27x @5 layer- \| 3 / 40 = 0.075 \| It can be seen that deeper models (with depth $N$) have larger optimal early exit layers $\ell$, but the ratio $\ell / N$ is relatively consistent. This implies that the latency cost for the self-speculative small model compared to the full model is fairly stable. Therefore, we recommend setting $\ell / N$ between 1/16 and 1/10 in practical applications. ### 1.2 The early-stopping threshold $\eta$ Kangaroo uses a static threshold to determine the timing of the second early stopping, based on the observation that the confidence of draft tokens in the small model is strongly correlated with their acceptance by the large model. For example, Figure 4 in the main paper shows that the Top-1 confidence distribution of the speculative small model on the mathematical reasoning subtask exhibits a clear bimodal distribution. Selecting the intersection of these distributions as the threshold allows us to retain a high proportion of tokens likely to be accepted while minimizing the proportion of tokens likely to be rejected. To explore the sensitivity of this threshold, we visualized the conditional distributions of the small model's Top-1 confidence across different model architectures and subtasks (see Figures in the global response PDF). Interestingly, we found this threshold to be stable (ranging from 0.6 to 0.8) across various models and subtasks. ### 2. More comprehensive comparison Following the suggestions from Reviewer xb3n and Reviewer oUCN, we conducted experiments on a wider range of model architectures and larger models on Spec-Bench by including Llama2-13B-Chat. \| $\qquad $ Method \| Translation \| $\qquad $ QA \| Summarization \| $\quad $ Math \| $\quad $ RAG \| MT Bench \| Avg. \| \| :-------------------: \| :------------------: \| :------------------: \| :------------------: \| :------------------: \| :------------------: \| :------------------: \| :------------: \| \| Draft & Verify \| $1.22\times (2.61)$ \| $1.02 \times (2.36)$ \| $1.13 \times (2.84)$ \| $1.08 \times (2.47)$ \| $1.15 \times (2.44)$ \| $1.12 \times (2.46)$ \| $1.12 \times $ \| \| SpS \| $1.26\times (1.63)$ \| $1.34\times (1.69)$ \| $1.14\times (1.47)$ \| $1.34\times (1.77)$ \| $1.32\times (1.81)$ \| $1.28\times (1.67)$ \| $1.28 \times $ \| \| Medusa$^$ w/o Tree* \| $1.53\times (1.85)$ \| $1.27\times (1.55)$ \| $1.19\times (1.48)$ \| $1.36\times (1.76)$ \| $1.25\times (1.54)$ \| $1.43\times (1.72)$ \| $1.34 \times $ \| \| Kangaroo w/o Tree \| $1.46 \times (1.97)$ \| $1.40 \times (1.87)$ \| $1.35 \times (1.97)$ \| $1.52 \times (2.22)$ \| $1.36 \times (2.05)$ \| $1.58 \times (2.28)$ \| $1.45 \times $ \| > Speedup comparison of various speculative decoding methods on Spec-Bench [22] for Llama2-13B-Chat. Values outside the parentheses indicate speedup, while those inside parentheses indicate the compression rate (CR). $^*$ denotes reproduction result. SpS takes Llama-68M as the draft model. From the experimental results in the table, we further validate the effectiveness and generalization of Kangaroo. Compared to Table 1 in the main paper, an interesting phenomenon emerges: most algorithms applied to Vicuna-13B show lower speedup in translation tasks relative to other datasets, but when the target model is switched to Llama2-13B-Chat, the speedup significantly improves. This may be related to the different SFT data used by these models. We will incorporate the valuable suggestions provided by the reviewers in our final version and further refine the ablation experiments and model evaluations. We sincerely hope our response addresses the concerns raised by the reviewers. Sincerely, Paper 2124 Authors Pdf: /pdf/0e048c22e93974476117e2caece8501626366137.pdf	NeurIPS_2024_submissions_huggingface	2,024	Summary: The authors propose a new self-speculative decoding framework, Kangaroo, for accelerating large language models (LLMs) by leveraging a double early exiting strategy. Kangaroo is able to enhance inference efficiency without the need for a separate draft model. It utilizes the shallow sub-network and the LM Head of the target LLM to construct a self-drafting model and introduces a dynamic token-level early exiting mechanism to minimize latency. The proposed approach achieves significant speedups, up to 2.04x, compared to existing methods, with substantially fewer additional parameters. The paper provides empirical validation mainly on Spec-Bench for Vicuna, showing Kangaroo achieves better wall-clock speedup. Strengths: 1. The paper is well-written and clearly presents the proposed framework. The related works are well-discussed, and the proposed method is well-motivated. 2. Novelty: the proposed token-level early exiting mechanism is interesting. It dynamically adjusts the decoding process based on the confidence levels of the predictions, which leads to better inference efficiency. 3. The authors conduct extensive experiments on multiple benchmarks, providing a thorough comparison with state-of-the-art methods. Weaknesses: 1. Insufficient explanation and comparison to existing early-exiting methods. The authors suggest that their work differs from the previous early exiting methods [1] [2] by using early exiting from shallow layers. However, the paper lacks a detailed explanation of the differences and a comparison with these methods empirically. In Table 1, the main evaluation doesn't include the comparison with these early exiting methods. It would be great to have a detailed comparison. 2. Double early exiting strategy involves additional hyperparameters, such as the exiting layer and the threshold for early exiting. Although the authors provide ablation studies on the selection of hyperparameters, it would be helpful to discuss their sensitivity across different models and tasks. [1] Jun Zhang, Jue Wang, Huan Li, Lidan Shou, Ke Chen, Gang Chen, and Sharad Mehrotra. Draft & verify: Lossless large language model acceleration via self-speculative decoding. arXiv preprint arXiv:2309.08168, 2023. [2] Coleman Hooper, Sehoon Kim, Hiva Mohammadzadeh, Hasan Genc, Kurt Keutzer, Amir Gholami, and Sophia Shao. Speed: Speculative pipelined execution for efficient decoding. arXiv preprint arXiv:2310.12072, 2023. Technical Quality: 3 Clarity: 3 Questions for Authors: 1. According to Figure 4, it seems there is no clear separation between "Accept" and "Reject" draft tokens. This makes choosing an appropriate threshold difficult. Do different down-streaming tasks have different favorable threshold and distribution? 2. How does the proposed double early exiting mechanism differ fundamentally from existing early exiting strategies in speculative decoding? Why starting from shallow layers works better? 3. How robust is Kangaroo in terms of hyperparameters under different models and tasks? Confidence: 3 Soundness: 3 Presentation: 3 Contribution: 2 Limitations: 1. Reproducibility: the paper doesn't provide source code, which makes it difficult to reproduce the results. 2. Evaluation: the paper mainly evaluates the proposed method on Spec-Bench for Vicuna. It would be great to see more evaluation on other benchmarks and models to provide a more comprehensive comparison. Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We thank the reviewer for the detailed comments and answer each of these comments below. ### 1. Insufficient explanation and comparison to existing early-exiting methods. We highlight the fundamental differences between Kangaroo and existing early-exiting methods in the following table: \| $\qquad$ Method \| $\qquad$ $\quad$ Early-Exiting \| $\qquad$ Architecture \| Plug and Play \| \| :----------------: \| :-----------------------------: \| :-------------------: \| :-----------: \| \| Kangaroo \| shallow layers \| General LLM \| True \| \| Draft & Verify [1] \| skipping redundant layers \| General LLM \| True \| \| SPEED [2] \| Speculative Pipelined Execution \| Parameter sharing LLM \| False \| SPEED specifically targets parameter-sharing (block-reuse) architectures, combining speculative decoding with pipelined execution for inference acceleration. Parameter sharing allows deeper decoders to enhance model performance. The combination of speculative decoding and pipelined execution offers SPEED a speed advantage over larger models without parameter sharing. However, Draft & Verify and Kangaroo are architecture-agnostic and can be plugged into mainstream large language models (LLMs). Draft & Verify employs Bayesian optimization to pre-select less critical layers in the target LLM for skipping, while Kangaroo implements early exits in the continuous shallow layers of the target LLM and uses a lightweight adapter to bridge the expressiveness gap between the self-speculative small model and the large model. Due to SPEED’s limited applicability and there is no released code, we focus on a detailed performance comparison between Kangaroo and Draft & Verify. \| Model \| $\quad$ Method \| Translation \| $\qquad$ QA \| Summarization \| $\quad$ Math \| $\quad$ RAG \| MT Bench \| Avg. \| \| :-------: \| :------------: \| :------------------: \| :------------------: \| :------------------: \| :------------------: \| :------------------: \| :------------------: \| :------------: \| \| Vicuna-7B \| Draft & Verify \| $1.22\times (2.61)$ \| $1.02 \times (2.36)$ \| $1.13 \times (2.84)$ \| $1.08 \times (2.47)$ \| $1.15 \times (2.44)$ \| $1.12 \times (2.46)$ \| $1.12 \times $ \| \| Vicuna-7B \| Kangaroo \| $1.24 \times (1.41)$ \| $1.43 \times (1.87)$ \| $1.50 \times (1.87)$ \| $1.61 \times (2.14)$ \| $1.52 \times (2.05)$ \| $1.68 \times (2.22)$ \| $1.50 \times $ \| \| Model \| Method \| Translation \| $\qquad$QA \| Summarization \| $\quad$ Math \| $\quad$ RAG \| MT Bench \| Avg. \| \| :-------: \| :------------: \| :---------: \| :--: \| :-----------: \| :--: \| :--: \| :------: \| :--: \| \| Llama2-13B-Chat \| Draft & Verify \| $1.11 \times (2.34)$ \| $1.06 \times (2.07)$ \| $1.02 \times (2.25)$ \| $1.08 \times (2.39)$ \| $1.05 \times (2.32)$ \| $1.07 \times (2.34)$ \| $1.07 \times $ \| \| Llama2-13B-Chat \| Kangaroo \| $1.46 \times (1.97)$ \| $1.40 \times (1.87)$ \| $1.35 \times (1.97)$ \| $1.52 \times (2.22)$ \| $1.36 \times (2.05)$ \| $1.58 \times (2.28)$ \| $1.45 \times $ \| > We used the official code from the Draft & Verify repository to optimize the redundant layers for Vicuna-7B and used the official optimized layers for Llama2-13B-Chat. Values outside the parentheses indicate speedup, while those inside parentheses indicate the compression rate (CR). While Draft & Verify achieves a higher compression rate, indicating the retained layers in the draft model better approximate the full model, the high latency of the small model limits its end-to-end speedup. ### 2. Why starting from shallow layers works better? Theoretical analysis of speculative inference [3] highlights two key factors: the accuracy of the small model and its inference latency. Although Draft & Verify benefits from higher accuracy by leveraging deeper features, its small model incurs a high cost, almost half that of the large model (compared to about 1/10 in Kangaroo), limiting its overall speedup. The reasons for the effectiveness of shallow layers in Kangaroo may include: * The initial layers directly connect token embeddings to subsequent decoder layers, playing a critical role in feature representation. References provided by Reviewer Tuhh [4, 5] suggest that the shallow layers of LLMs are sufficient for understanding simple tasks, aligning with the requirement for speculative decoding where the draft model needs only to align with the large model on simple tokens. * The autoregressive adapter used in Kangaroo enhances the expressiveness of shallow layers, and its lightweight design reduces inference latency. > [3] Leviathan Y, Kalman M, Matias Y. Fast inference from transformers via speculative decoding[C]. ICML 2023. > > [4] The Unreasonable Ineffectiveness of the Deeper Layers > > [5] Exploring Concept Depth: How Large Language Models Acquire Knowledge at Different Layers? ### 3. Do different down-streaming tasks have different favorable thresholds and distribution? As shown in Figure 4 in the main paper, "Accept" and "Reject" draft tokens cannot be perfectly separated by a fixed threshold. However, selecting the intersection of these distributions as the threshold allows us to retain a high proportion of tokens likely to be accepted while minimizing the proportion of tokens likely to be rejected. To explore the sensitivity of this threshold, we visualized the conditional distributions of the small model's Top-1 confidence across different model architectures and subtasks (see Figures in the global response PDF). The confidence distributions across different subtasks suggest that selecting $\eta \in [0.6, 0.8]$ is a robust choice. ### 4. Robustness and sensitivity of Kangaroo's hyperparameters Please refer to the common concerns in the global response. ### Reproducibility We will make our training and evaluation code publicly available in the camera-ready version. --- Rebuttal 2: Title: Looking forward to your feedback Comment: Dear Reviewer NQcw, With the discussion phase nearing the end, we would like to know whether the responses have addressed your concerns. Should this be the case, we are encouraged that you raise the final rating to reflect this. If there are any remaining concerns, please let us know. We are more than willing to engage in further discussion and address any remaining concerns to the best of our abilities. We are looking forward to your reply. Thank you for your efforts in this paper. Best regards, Authors --- Rebuttal Comment 2.1: Comment: Thanks for your detailed response, and this addresses most of my concerns. I am happy to raise my score based on the rebuttal and discussion with ACs and other reviewers (especially the upcoming response from reviewer J1nq). --- Reply to Comment 2.1.1: Title: Official Comment by Authors Comment: Dear Reviewer NQcw, Thank you for your continued efforts in helping us improve the quality of this manuscript. We greatly appreciate the time and attention you have dedicated to this process. You mentioned that you would consider raising your score based on the upcoming response from Reviewer J1nq. We are pleased to report that Reviewer J1nq has expressed satisfaction with our response, indicating that the concerns have been also adequately addressed. As the discussion period draws to a close, we would like to reach out to see if you have any remaining questions or unresolved issues. If everything is now clear, we would be grateful if you could consider updating your evaluation to reflect this. Once again, thank you for your constructive feedback and for your invaluable contribution to the development of this manuscript. We look forward to hearing from you soon. Sincerely, Paper 2124 Authors	null	null	null	null	null	null
D2R2: Diffusion-based Representation with Random Distance Matching for Tabular Few-shot Learning	Accept (poster)	Summary: This paper proposes a novel approach named Diffusion-based Representation with Random Distance matching (D2R2) for tabular few-shot learning. It leverages the powerful expression ability of diffusion models to extract essential semantic knowledge crucial for the denoising process. During the training process of the designed diffusion model, it introduces a random distance matching to preserve distance information in the embeddings, thereby improving the effectiveness of classification. During the classification stage, it introduces an instance-wise iterative prototype scheme to improve performance by accommodating the multimodality of embeddings and increasing clustering robustness. Experiments demonstrate the effectiveness of the proposed method. The main contributions of this paper are: - It may be the first to propose a specifically designed diffusion method to learn semantic knowledge for tabular data. - It proposes an innovative framework, D2R2, to extract representations in tabular few-shot learning. - It introduces a novel classifier with instance-wise iteration prototypes to further improve few-shot classification performance. Strengths: - The problem studied in this paper is interesting and valuable. - This paper is well written and in good sharp, which is easy to follow. - Experimental results are promising and can validate the effectiveness of the proposed method. Weaknesses: - In my opinion, the description of the shortcomings of existing tabular few-shot learning methods in the introduction is not sufficient. Specifically, the authors should further emphasize the problems the heterologous feature types bring that SOTA tabular few-shot learning methods ignore, to demonstrate the necessity and innovation of the proposed method. - On the basis of the first issue, the introduction of diffusion models should be based on the drawbacks of existing problems, rather than solely on their strong expressiveness. The authors should point out the motivation for the leverage of diffusion models. - The emergences of semi-supervised learning and self-supervised learning mentioned in this paper are somewhat abrupt. The authors should explain the relationship between them and tabular few-shot learning, as well as their relevance to the D2R2 proposed in this paper. Technical Quality: 2 Clarity: 3 Questions for Authors: 1. The authors should further emphasize the problems the heterologous feature types bring that SOTA tabular few-shot learning methods ignore, to demonstrate the necessity and innovation of the proposed method. 2. The authors should point out the motivation for the leverage of diffusion models. 3. The authors should explain the relationship between semi-supervised learning, self-supervised learning, and tabular few-shot learning, as well as their relevance to the D2R2 proposed in this paper. 4. Tabular few-shot learning seems not to emphasize the variation of categories in the training (base ) and test (novel) datasets, does this indicate that it ignores the requirement for the generalization of the algorithm? Confidence: 3 Soundness: 2 Presentation: 3 Contribution: 3 Limitations: The authors point out the limitations of this work. No obvious potential negative societal impact remains. Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thanks for your comments! Q1. We plan to add more explanations about shortcomings of existing methods as follows. Firstly, tabular data comprises heterologous features, which underscores the importance of simultaneously modeling continuous and categorical features. The current SOTA method STUNT[28] in our paper, and other existing few-shot learning methods designed for images or texts, treat all features as the same type, neglecting the unique information in different features. Failing to address heterogeneous features hinders the model's ability to capture complex data patterns and relationships. Moreover, the heterogeneous features make tabular data lacking in straightforward augmentation techniques, which are easily implemented in images, as demonstrated by UMTRA[21]. Secondly, unlike images and texts, tabular data lacks strong spatial and sequential relationships between features. While STUNT[28] generates pseudo-labels based on the assumption that substitutable features, such as "occupation" acting as a proxy for "income", may not universally apply to tabular data. Besides, the arbitrary permutation of columns in tabular data highlights the model's robustness, which is ignored by existing methods. Thirdly, the influence of column permutation and multimodal behavior within same class are all ignored by existing methods. Q2. We will add more explanations about motivations as follows. Existing methods failing to address issues in Q1. thus we propose a novel approach, D2R2, to tackle above challenges in tabular data. Considering the first and second challenges above, we avoid to rely on proxy methods or augmentation methods, which are limited by the lack of relationships or augmentation techniques in tabular data. Instead, we create an information bottleneck for extracting semantic knowledge, named D2R2, which leverages the strong expressiveness of the diffusion model and distance information from pairwise comparison. Notably, we handle numerical and categorical data types separately. For example, we introduce two distinct noises and generate two distinct random projections for different feature types. Adding small noise to the input also enhances the model's robustness. For the third challenge in Q1, we propose an instance-wise iteration prototypes classifier, which can construct stable prototypes while revealing the multimodal behavior within same class. Q3. Based on the unique characteristics of tabular data, as detailed in Q4, we address scenarios where there is an unlabeled training set and a very limited labeled support set (K-shot) to help predict class labels for a testing query set. This follows the problem definition in [28] and is stated in Section 3 of our paper. Such scenarios are common in critical applications like credit risk assessment[1] and diagnosing patients with rare or new diseases[2]. Based on our research, three types of methods can address these scenarios: few-shot learning, semi-supervised learning, and self-supervised learning. Meta-learning is one of the most common techniques in few-shot learning. The SOTA method [28] for few-shot tabular learning in our paper follows this scheme. On the other hand, our method innovates within the prototype scheme by enhancing embedding generation and prototype classification. In semi-supervised learning, a model is trained on labeled data then to predict the unlabeled data, where predicted labels are treated as true labels in further training rounds. This method aligns with our problem definition; however, it requires more labeled data compared to few-shot learning. Self-supervised learning excels in generating robust representations. Since our approach involves generating representations, we compare our method with the SOTA self-supervised learning methods[54,4,47] in the tabular domain. The mentioned methods serve to position our work within the broader context of existing methodologies that address similar problems. We have already outlined the relevant methods in Section 2. We will include the above more detailed discussion in the updated version. [1]Chen N, Ribeiro B, Chen A. Financial credit risk assessment: a recent review[J]. Artificial Intelligence Review, 2016, 45: 1-23. [2]Cereda D, Tirani M, Rovida F, et al. The early phase of the COVID-19 outbreak in Lombardy, Italy[J]. 2003.09320, 2020. Q4. Our problem setting follows the definition provided in [28], where there is no variation of categories between the training (base) and test (novel) datasets, but the training set is unlabeled, which is a more common scenario in tabular data. Tabular datasets consist of numerical or categorical inputs that have specific and explicit meanings, where classes are well-defined and consistent across the dataset. it is rare to encounter new categories based solely on existing features. New categories usually emerge only with the introduction of new features, which is different from domains like image or text data, which have spatial and sequential relationships. For example, new image categories can be effectively classified based on the learned spatial structure patterns between pixels. Our study focus on the most common scenario in tabular few-shot learning, as mentioned in Section 3 of our paper and supported by the problem setting in [28]. However, we also conducted experiments to test the generalization of our method regarding the variation of categories in the training (base) and test (novel) datasets. For the dataset for 10-class classification 'optdigit', we randomly removed one classes from the train set and validated the effectiveness of the test results. The experimental results are as follows, demonstrating the robustness and effectiveness of our method. \| Scenario \| shot=1 \|shot=5 \| \|---------------\|--------\|--------\| \| Remove Class #1\| 75.21\| 87.37\| \| Remove Class #2\| 72.15 \| 84.63\| \| Remove Class #6\| 77.16\| 88.62\| \| Non-Remove \| 81.13\| 90.73\| --- Rebuttal Comment 1.1: Comment: Thanks for the response. The response addresses the majority of my concerns. Therefore, I update my score to "Weak Accept". --- Reply to Comment 1.1.1: Comment: Thank you for your positive feedback and for recognizing the value of our work! We will certainly enhance the clarity of our manuscript based on your suggestions.	Summary: The paper proposed a new diffusion-based representation learning method for tabular data, namely D2R2, specifically for few-shot learning. It is the first paper to use the diffusion model for tabular data representation learning. The method trains a conditional diffusion model with a combined loss of vanilla diffusion reconstruction loss and a random distance match loss to learn an embedding function $z_{\theta}$. $z_{\theta}$ encodes a representation of the data with different noise levels. The resulting embedding is used as the conditional information to the diffusion model. Further, random linear projections are performed on the original data, and the random distance match loss is to preserve distances in the embedding space and the random projection space. Two distinct projections are used for numerical and categorical data. After training, instance-wise iteration prototypes are generated for more accurate and stable few-shot learning. The paper conducts experiments on 7 datasets from OpenML-CC18, showing superior performances than a variety of baselines. Strengths: 1. The paper introduced the first diffusion-based method for tabular few-shot representation learning. 2. The proposed method has superior performance compared to SOTA methods. 3. The paper is well-written and easy to follow. Weaknesses: 1. Even though D2R2 is the first method to utilize the diffusion model for representation learning, it is not the first paper to analyze/utilize the diffusion model for representation learning. The paper lacks discussion on related works in related areas such as [1] [2] [3]. 2. After training, the paper only extracts features of clean tabular data during few-shot learning, while the embedding function is learned for different noise levels. There is no analysis of the effect of different noise levels as input to the embedding function for few-shot learning. 3. The method is limited to only tabular data, such as the specific projection head design for numerical and categorical data. [1] Xinlei Chen, Zhuang Liu, Saining Xie, Kaiming He. Deconstructing Denoising Diffusion Models for Self-Supervised Learning. 2024. [2] Sarthak Mittal, Korbinian Abstreiter, Stefan Bauer, Bernhard Schölkopf, Arash Mehrjou. Diffusion Based Representation Learning. 2023. [3] Xingyi Yang, Xinchao Wang. Diffusion Model as Representation Learner. 2023. Technical Quality: 3 Clarity: 3 Questions for Authors: 1. See weakness 1. What's the relationship and key novelty in the method compared to existing methods? 2. Will other noise levels be more useful for few-shot learning? Or will a combination of different noise levels be better? Confidence: 3 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: Limitation is discussed and the paper states to have discussed potential negative societal impact. Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thanks for your comments! Q1. What's the relationship and key novelty in the method compared to existing methods? Existing papers about representation learnings are all designed for image data. Paper [1] introduces a "latent Denoising Autoencoder" (LDAE) architecture where the learned representations are used for Denoising Autoencoder. Unlike LDAE, we learn a latent z that supports image diffusion process, rather than Autoencoder reconstruction process, which is less suitable for tabular data as discussed in [4]. The success of the representations in paper [2] depends on the quality of the generated images, meaning accurate image generation is crucial for performance. Generating high-quality data is especially difficult for tabular data due to its varied features and the absence of strong spatial or sequential relationships [5]. Our approach, however, does not rely on data quality for results, making our model more effective with tabular data. Paper [3] extracts representations from noisy data at a specific time step t in the diffusion process, using a student model trained with task labels. Our method differs by deriving representations directly from the original feature space without requiring a student model, eliminating the need for labels during training. This is especially critical for tabular few-shot learning where labeled data is scarce. Moreover, different from existing papers, our work effectively modify diffusion models according to tabular data characteristics, including heterogeneous features, lack of strong spatial and sequential relationships between features, the influence of column permutation, and multimodality within the same class. We design a conditional diffusion process with distance matching to extract representations from original space, which is not influenced by the new generated samples' quality. Besides, tabular columns can be arbitrarily permuted without altering the underlying information, thus models for tabular data need to be robust to such permutations. We perturbed original data with two distinct types of relatively small noises for numerical and categorical input respectively (Section 4.1), which enhanced the robustness of the performance. Moreover, deriving representations solely from the "diffusion-only" model is not effective enough, as shown in our manuscript's Ablation study (Table 2). Therefore, we align the comparison of pairwise distances between the embedding space and a randomly projected space with the diffusion training process, using different projection matrices for various feature types. Additionally, we propose instance-wise iterative prototype to address multimodality for tabular data. [4]Nam J, Tack J, Lee K, et al. Stunt: Few-shot tabular learning with self-generated tasks from unlabeled tables[J]. arXiv preprint arXiv:2303.00918, 2023. [5]Xu L, Skoularidou M, Cuesta-Infante A, et al. Modeling tabular data using conditional gan[J]. Advances in neural information processing systems, 2019, 32. Q2. Will other noise levels be more useful for few-shot learning? Or will a combination of different noise levels be better? The timestep t = 1, ..., T indicates noise levels in diffusion models. In our experimental analysis, we have noted a trend where performance sees an improvement with an increase in t during the initial stages. However, this trend of performance enhancement plateaus after the first few timesteps, which tends to stabilize with minimal fluctuations, as shown in the following table. We speculate that this is because, as t approaches 0, the noise levels decrease, potentially causing only the fine-grained details to be lost. Hence, the representation learns to keep the information needed to recover these fine-grained details. On the contrary, as t approaches T, noise levels rise, and the mutual information between xt and x0 begins to decrease. In these situations, effective denoising requires that all information about x0 be thoroughly encoded, which contains the most information about the data class. Thus, as stated in line 200 of our manuscript, "We focus on larger timesteps thus extract the low-frequency semantic information rather than details." In our experiments, we selected the last step (T) for all datasets to ensure stable performance. We will include a discussion on this observation in the updated version of our manuscript. \| Dataset \| Shot \| First Step \| Middle Step \| 80th Percentile Step \| Last Step \| Average of All Steps \| \|-----------\|------\|------------\|-------------\|----------------------\|-----------\|----------------------\| \| optgidits \| 1 \| 72.39 \| 79.21 \| 80.93 \| 81.13 \| 80.66 \| \| \| 5 \| 84.57 \| 88.67 \| 91.73 \| 90.73 \| 89.66 \| \| cmc \| 1 \| 37.19 \| 42.26 \| 42.86 \| 42.88 \| 42.74 \| \| \| 5 \| 36.28 \| 42.73 \| 43.04 \| 43.39 \| 42.37 \| In the updated version, we will include the above discussion following your valuable comments. --- Rebuttal Comment 1.1: Comment: Thanks for the response. I've read through other reviewers' feedback and responses as well. I agree with some reviewers' opinions that there lack of method motivation, experiments on scalability, and some technical details, but I am overall satisfied with the authors' responses to address the existing problems and will keep the score as it is. --- Reply to Comment 1.1.1: Comment: Thank you for your positive feedback and for recognizing the value of our work! We appreciate your acknowledgment of our responses in addressing the existing problems.	Summary: The paper introduces a novel method, Diffusion-based Representation and Random Distance Matching (D2R2), to address the challenge of few-shot learning on tabular data. This approach leverages the robust representational capacity of diffusion models to extract essential semantic knowledge from tabular data, thereby enhancing the performance of downstream classification tasks. Additionally, the paper presents the Random Distance Matching (RDM) loss function, which preserves distance information within embedding vectors during diffusion model training, further boosting classification performance. During the classification phase, the authors propose an instance-based iterative prototype scheme to accommodate the multimodal characteristics of the embedding vectors, thus improving clustering robustness. Extensive experiments conducted on various few-shot learning benchmarks for tabular data demonstrate that D2R2 surpasses current benchmark methods. In conclusion, this paper introduces an innovative few-shot learning method for tabular data, achieving significant advancements in extracting effective semantic representations and improving classification performance. Strengths: - The use of diffusion models to extract effective semantic representations from tabular data is a novel approach that overcomes the limitations of traditional supervised learning methods. - The introduction of the Random Distance Matching (RDM) loss function, which preserves distance information within the embedding vectors during diffusion model training, further enhances classification performance. - Extensive experiments conducted on multiple few-shot learning benchmarks demonstrate that D2R2 outperforms the latest benchmark methods, validating the effectiveness of the proposed method. - The paper provides comprehensive explanations and descriptions of the various components of the method, including the expressive power of the diffusion model, the role of the RDM loss function, and the principles of the iterative prototype scheme, making the methodology and process clear and understandable. Weaknesses: - The paper lacks a clear description of the input format for tabular data. Specifically, it is unclear whether the tabular data is input in text form or as images. Clarification on this point is essential, as the input format may significantly impact performance. - The paper utilizes a diffusion model to enhance the semantic representational capacity of tabular data. However, it remains uncertain whether the diffusion model is effective across all types of tabular data. Detailed demonstrations and explanations of how the model captures and enhances these semantic features are needed to substantiate its broad applicability. - The authors introduce a diffusion model that differs from existing methods. It is crucial to provide additional information regarding the model's parameter count and FLOPs (floating point operations per second) to allow for a thorough comparison of its computational efficiency with that of other models. This would help in understanding the trade-offs between performance gains and computational costs. Technical Quality: 2 Clarity: 3 Questions for Authors: Please see the Weakness section. Confidence: 4 Soundness: 2 Presentation: 3 Contribution: 2 Limitations: The author has already provided a description of the limitations, but it needs to be further elaborated. Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thanks for your comments! Q1. The paper lacks a clear description of the input format for tabular data. In our study, "tabular data" refers to the dataset organized in tables, which is a structured format that presents information in rows and columns. It is defined as $D = \{\boldsymbol{x} _ i\} _ {i=1}^n\subset \mathcal{R}^{d}$ consisting of $n$ instances (rows) and $d$ dimensional features (columns). Each data instance $\boldsymbol{x} _ i = (x_i^1, x_i^2, ... , x_i^d)$ may or may not hold strong relationship among features. The feature types of tabular data can be a mixtures of numerical numbers or categorical indicators. For example, the "income" dataset classifies individuals' incomes based on features like "age" (numerical)，"education" (categorical) and so on. Table 3 in Appendix provides a detailed information on each dataset. Tabular data is widely used in practical applications, and few-shot learning is particularly important in the context of tabular data, as limited labeled tabular data is inherently common in many real-world applications, such as financial fraud detection [1], disease diagnosis [2], and social science [3]. Different from CV and NLP, tabular data contains a mixture of numerical and categorical features; tabular data lacks spatial and sequential relationships between columns, making it more challenging to extract semantic knowledge; tabular column order can be arbitrary permuted without affecting the tabular information; tabular embeddings within the same class may exhibit multiple modes, as demonstrated in Figure 3 of our manuscript. The significance and challenges of tabular few-shot learning motivate us to conduct this work. Q2. Whether the diffusion model is effective across all types of tabular data? From the experimental analysis, our datasets (seven in the paper and two additional datasets in the response to Review wij1's Q1) involve various types of tabular data. These include varying sample sizes (from 768 to 48842), feature dimensions (from 8 to 24482), feature types (numerical, categorical or the mixture of both), and number of classes (from 2 to 10). Our experimental results demonstrate the effectiveness of our proposed method and its scalability across various types of those tabular data. From a theoretical analysis perspective, reasons that the designed diffusion model can extract semantic knowledge come from twofold. Firstly, the diffusion model with powerful expressiveness encodes the information needed for denoising. Specifically, in conditional diffusion models, the noise reconstruction loss $\mathbf{E} _ {t, \mathbf{x} _ 0, \mathbf{x} _ t} \|\|\epsilon_\phi(\mathbf{x} _ t, t,c) - \epsilon\|\|^2_2$ trains the noise prediction function $\epsilon_\phi(\mathbf{x} _ t, t,c)$ to predict the true noise $\epsilon(\mathbf{x} _ t, t, \mathbf{x} _ 0)$ given the noisy sample $\mathbf{x} _ t$, the knowing $t$ and the condition information $c$. If $c = \mathbf{x} _ 0$, we could expect $\epsilon_\phi$ can almost perfectly recover $\epsilon$. By replacing the conditional information $c$ by a function $z_\theta(\mathbf{x} _ 0)$ that maps to a embedding space with lower dimension than $\mathbf{x} _ 0$, we introduce an information bottleneck to the noise reconstruction process. This forces $z_\theta$ to extract effective information for denoising from $\mathbf{x} _ 0$, leading to representation learning through the noise reconstruction loss. Q3. what is model's computational costs？ The model complexity is manageable through our framework's utilization of two neural networks: the embedding network and the noise prediction network, both of which are demonstrated as 3-layer MLPs in our paper. These settings (hidden dimensions, embedding dimensions and model structure) can be adjusted to meet different needs. In terms of computational efficiency, The main time-consuming factor after the embedding model is the calculation of the instance-wise iterative prototype. For an N-way K-shot problem with L iterations and a query set of size Q, the computation complexity is O(NKLQ). In few-shot learning, where N, K, L are small, the computation complexity is linear in the size of the query set. In the updated version, we will include the above discussion following your valuable comments. [1]West J, Bhattacharya M. Intelligent financial fraud detection: a comprehensive review[J]. Computers & security, 2016, 57: 47-66. [2]Kim S J, Choi S J, Jang J S, et al. Innovative nanosensor for disease diagnosis[J]. Accounts of Chemical Research, 2017, 50(7): 1587-1596. [3]Hicks D. The four literatures of social science[J]. Handbook of quantitative science and technology research: The use of publication and patent statistics in studies of S&T systems, 2004: 473-496. --- Rebuttal Comment 1.1: Title: Reply to the authors' rebuttal Comment: Dear authors, I greatly appreciate your responses and the additional results presented in the PDF. I think the authors addressed all my comments and I think this work should be accepted. --- Reply to Comment 1.1.1: Comment: We sincerely appreciate your positive feedback and are pleased that our responses addressed your comments. Thank you for recommending our work for acceptance!	Summary: The paper proposed a few-shot learning framework for tabular data by designing a diffusion based semantic knowledge encoder and introducing a random distance matching mechanism to preserve distance information in the embeddings. During classification, an instance-wise iterative prototype scheme is utilized to improve performance by accommodating the multimodality of embeddings and increasing clustering robustness. Extensive experiments on multiple datasets show that D2R2 achieves state-of-the-art performance compared to other schemes. Strengths: 1. The proposed approach that combines the diffusion model with random distance matching for tabular few-shot learning is relatively unexplored compared to classical few-shot learning tasks. 2. The paper is well-written and organized, with detailed explanations of the challenges, methodology, and experimental results. Weaknesses: 1. The feature dimensions of the benchmark datasets are relatively low, which leaves the the scalability to large datasets or those with extremely high dimensionality not fully explored. 2. The pseudo-label validation scheme for hyperparameter selection relies on the assumption that the clustering of raw features can provide a reliable proxy for true labels, which might not always hold in practice. Technical Quality: 3 Clarity: 3 Questions for Authors: 1. How does the proposed method scale with larger datasets or those with higher feature dimensions? Are there any practical limitations or considerations for scaling? 2. Could the authors provide more empirical evidence or analysis on the pseudo-label validation scheme, such as when the clustering might not accurately reflect the true class structure? Confidence: 2 Soundness: 3 Presentation: 3 Contribution: 2 Limitations: No potential negative societal impact is observed. Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thanks for your comments! Q1. Does the proposed method scale with larger datasets or those with higher feature dimensions? In our experiments, we utilized seven datasets that are widely used in tabular data research, as referenced in popular papers [28], [2], and [54] in our paper. Among these, the "income" dataset has the most samples (48842), and the "pixel" dataset has the highest feature dimension (240). To further investigate the scalability of our method, we added two new datasets to our experiments. The "nomao" dataset has a larger size of 34465 samples, and the "breast" dataset has a significantly higher feature dimension of 24482. The following table shows their results. Our method consistently outperforms other baselines. These results will be included in the updated version of our paper, demonstrating that our method scales well to larger and high dimensional datasets. \| Dataset \| Shot \| CatBoost \| KNN \| SubTab \| VIME \| SCARF \| RTDL \| STUNT \| D2R2 \| \|---------\|--------\|----------\|------\|--------\|------\|-------\|------\|-------\|------\| \| nomao \| 1-shot \| 0.636 \| 0.635 \| 0.676 \| 0.647 \| 0.689 \| 0.683 \| 0.715 \| 0.794 \| \| \| 5-shot \| 0.753 \| 0.737 \| 0.761 \| 0.749 \| 0.776 \| 0.736 \| 0.814 \| 0.826 \| \| brest \| 1-shot \| 0.697 \| 0.718 \| 0.729 \| 0.701 \| 0.753 \| 0.763 \| 0.769 \| 0.776 \| \| \| 5-shot \| 0.770 \| 0.794 \| 0.827 \| 0.858 \| 0.844 \| 0.796 \| 0.868 \| 0.882 \| Q2. Could the authors provide more empirical evidence or analysis on the pseudo-label validation scheme？ In the few-shot learning scenario, particularly in 1-shot classification, there are no labeled samples available for validation, which leads most researchers to rely on fixed parameters. However, we argue that a better set of hyper-parameters can always be identified for a specific dataset. To this end, we generate pseudo-labels for a validation set using the fundamental principles of the soft k-means algorithm. Firstly, achieving complete consistency between the validation labels and the true labels is indeed challenging. Nonetheless, existing studies have demonstrated satisfactory performance despite minor discrepancies between the validation and test sets [1]. It is important to note that we have consciously avoided over-tuning the hyper-parameters on the validation set, limiting our adjustments to only three parameters, as detailed in the appendix. Secondly, we posit that raw features possess inherent clustering properties, and the soft k-means method has been proven effective and widely adopted in various contexts, as supported by references [3][4]. Our empirical evidence further substantiates this claim. In our experiments, the accuracy of pseudo-labels was evaluated on several datasets, all showing relative high accuracy. This leads us to believe that pseudo-labels serve as a relative reliable proxy for true labels during validation. For the dataset "income" , where accuracy was comparatively lower compared to the classes number, we conducted experiments without using a validation set but with fixed hyper-parameters. The results indicated that selecting parameters using a pseudo-label validation set achieved better performance than using fixed hyper-parameters. The robustness of our method, demonstrated across multiple datasets, enables our approach to be applicable to a wide variety of tabular datasets. Table 1: Accuracy of pseudo-labels \| Dataset \|DNA \| Income \| Karkunen \| Optdigits \| Nomao \| Breast \| \|-----------\|-----\|--------\|----------\|-----------\|-------\|--------\| \| Accuracy \|0.37\| 0.52 \| 0.53 \| 0.47 \| 0.59 \| 0.63 \| Table 2: Fixed hyper-parameters VS Tuned hyper-parameters \| \| $\alpha$ \| $d_z$ \| $\tau$ \| Accuracy \| \|-------\|----------\|-------\|--------\|-------\| \| Fix \| 0.1 \| 10 \| 0.1 \| 0.742 \| \| Fix \| 0.1 \| 80 \| 0.1 \| 0.755 \| \| Fix \| 0.5 \| 10 \| 0.1 \| 0.739 \| \| Fix \| 0.5 \| 80 \| 0.5 \| 0.721 \| \| Fix \| 0.9 \| 10 \| 0.1 \| 0.732 \| \| Fix \| 0.9 \| 80 \| 0.9 \| 0.743 \| \| Tune \| 0.775 \| 10 \| 0.325 \| 0.758 \| [1]Peng X, Usman B, Kaushik N, et al. Visda: The visual domain adaptation challenge[J]. arXiv preprint arXiv:1710.06924, 2017. [3]Ferraro M B, Giordani P. Soft clustering[J]. Wiley Interdisciplinary Reviews: Computational Statistics, 2020, 12(1): e1480. [4]Singh V K, Tiwari N, Garg S. Document clustering using k-means, heuristic k-means and fuzzy c-means[C]//2011 International conference on computational intelligence and communication networks. IEEE, 2011: 297-301. --- Rebuttal Comment 1.1: Comment: Thanks for the authors' response. The authors did additional experiments to answer my questions and they partially address my question related to the dimension of studied data. I have updated my scores. --- Reply to Comment 1.1.1: Comment: Thank you for your positive feedback and for recognizing the value of our work! We will incorporate the new experimental results into our revised version.	null	NeurIPS_2024_submissions_huggingface	2,024	null	null	null	null	null	null	null	null
Order-Independence Without Fine Tuning	Accept (poster)	Summary: State-of-the-art language models (LMs) are now used to perform tasks (such as, e.g., question answering) with no fine-tuning; these models are fed the question and target options in their context, and their next-token distribution is then used to select an answer from this set of options. LM outputs, however, are known to vary significantly based on the order in which these options are fed to it. This paper proposes set-based prompting, a method to make LMs invariant to such option reorderings. In short, the method feeds all options to a transformer model “in parallel” by modifying their position indexes to start from the same value, and by masking the attention so that options cannot attend to each other. In practice, this makes the model run slightly out of distribution, as tokens presented to the model after the QA options will be able to attend to multiple tokens with the exact same position (as the multiple options are assigned this equivalent position indexes). The paper has experiments on two tasks, question answering and language understanding. Strengths: The paper presents a simple solution to a well known problem in NLP. The paper runs experiments with a reasonable number of language model families: gpt-2, llama2, llama3, and mistral. Weaknesses: In my opinion, this paper presents what is a relatively simple solution as being more complex than it is. (To be clear, I think this simplicity is one of the solution’s positive aspects, and not a negative aspect. Its presentation, however, could be considerably simpler.) Further, the proof of order invariance is sold as an important contribution, but it is also relatively straightforward. Since the methodological/theoretical contributions of this paper are relatively minor, though, I would expect it to have more experiments to confirm its efficacy. The paper, however, only presents two experiments, which are not very comprehensive in assessing the method's performance. More details below. * The paper’s contribution is a simple (yet potentially efficient) method to make LMs invariant to prompt ordering. This solution amounts to reindexing token positions and slightly changing the model’s attention mask. However, the paper takes almost 6 pages to present this relatively simple solution. * The paper only runs experiments on two tasks, and with one dataset each. This is not a comprehensive evaluation. The paper could run experiments with other tasks. Specifically, analysing the performance of this method on long-context information retrieval, as in Liu et al. (2024) could be interesting. * In-context learning is another important setting in which models are susceptible to input order variation. However, the paper only focuses its experiments on zero-shot experiments. Liu et al. (2024). Lost in the middle: How language models use long contexts ------------ Edit after author responses: I thank the authors for their response, and apologise for only replying to it after the discussion deadline. While I still think that some aspects of the paper could be significantly improved, e.g., simplifying the presentation, the newly added experiments make the paper's evaluation much more comprehensible, so I am raising my scores (from 3 to 5). The experiments showing the performance degradation on the long-context experiment are quite interesting (even if they present negative results), and experiments with ICL show another important setting for which this method can be useful. Technical Quality: 3 Clarity: 1 Questions for Authors: I don't have any specific questions. But as a suggestion, I think the stacked box plots in Figs 3 and 5 are quite hard to read. It would likely be easier to read them if the different conditions were presented as separate bars. Confidence: 4 Soundness: 3 Presentation: 1 Contribution: 1 Limitations: The authors mention that their solution will make models run slightly out of distribution. Expanding on this point could be interesting, maybe discussing possible ways to mitigate this issue, such as fine-tuning models. Flag For Ethics Review: ['No ethics review needed.'] Rating: 5 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We thank the reviewer for their time in reading our paper and providing some useful questions. # Relevancy Order dependency is an important open problem in the LLM NLP literature [1,2,3,4]. Known motivations include increasing model robustness and reducing ordering biases. To these we add applications from algorithmic fairness. To our knowledge, we are the first to obtain a theoretically validated solution. Moreover, we demonstrate that our solution incurs essentially no cost in accuracy. The robustness problem is a major concern for LLMs and that they may only know the answer in one ordering [5] suggests that the models are not learning useful representations. Having a method that forces the LLM to answer without using ordering means we can better evaluate LLMs which will affect all LLM evaluations. # Methodological Elegance That the method seems obvious in hindsight is a strength, not a weakness! This elegant solution allows for implementation by adding only three new “tokens” to the tokenizer’s dictionary. We will endeavor to increase clarity, and any suggestions for where to focus our attention are enthusiastically welcomed. # Experimental Scope Inspired by [1,2,3,4], we tested our method across multiple LLM implementations with the 57 subject MMLU as our main target. We also spent considerable time and effort examining which components of our intervention were required in practice: our original hypothesis was that we would also need to modify the padding so that the different sub-sequences were aligned, but we found that this was not needed. We also showed that our method allows for expanding the context window of the model (Sec 4.3), but as that follows directly from the positional encoding formulation we did not emphasize this result. The reviewer suggests two non-multiple choice question tasks. We have implemented both of them and plots showing our results are available in the main response PDF. ## Long-Context Information Retrieval We implemented the long-context information retrieval task described in the paper by Liu et al. [6]. To do this, we generated 10 document sequences with the answer hidden in one of them. Then we moved where the answer was located from the first to the last document and observed the retrieval accuracy. We used the same templates and dataset as Liu et al. for this. To test the effects of our method we ran the documents in parallel, either all 10, 5 groups (2,2,2,2), four groups (3,3,3,1) or two groups (5,5). When running the sets of documents in parallel there are two opposing forces affecting the accuracy: (1) parallelism, naturally, reduces order-dependence, which helps accuracy; (2) at the same time, the intervention moves the inputs farther out of distribution, reducing accuracy. Our plot suggests that limited intervention is a sort of “sweet spot.” More interestingly, this suggests that our method can be used to evaluate the robustness of the model’s learned manifold. This is an exciting new direction. Thanks again! We will include more details along with the data in the code release in the final paper. ## In-Context Learning In response to the reviewer’s suggestions, we implemented a sentiment classification task with in context learning. The model was provided with 4 samples with labels, then asked to classify the fifth. The dataset is Financial Phrase Bank [7] so all statements are finance related and the model is attempting to classify them as positive or negative. To look at the effects of ordering on the model performance we always had three samples with the same label and 1 with the other label, with the different label always being first or last. When we ran this test, the in-context learning improved over the 1 shot test, showing that the models are using the examples. We found that putting the examples in parallel often improves the accuracy over the non-parallel case, or is as good. This was a very strong showing by our method and we thank the reviewer for suggesting it! The plot can be seen in the main response PDF. We will include the plot and details of the test in the paper along with releasing the code. Bar plots We have redone them as separate bars, see main response PDF, and will update the paper with the new style for all plots. # Discussion of fine-tuning We briefly discussed fine-tuning in section 5.1, but mostly to say we did not examine it. We agree that fine-tuning is the likely best way to mitigate the accuracy degradation. We will add more to the discussion on this. If there are other areas we can improve the discussion on please let us know and we can discuss them during the decision period. # Citations [1] Alzahrani, N., Alyahya, H.A., Alnumay, Y., Alrashed, S., Alsubaie, S., Almushaykeh, Y., Mirza, F., Alotaibi, N., Altwairesh, N., Alowisheq, A., Bari, M.S., Khan, H., 2024. When Benchmarks are Targets: Revealing the Sensitivity of Large Language Model Leaderboards. [2] Chen, X., Chi, R.A., Wang, X., Zhou, D., 2024. Premise Order Matters in Reasoning with Large Language Models. https://doi.org/10.48550/arXiv.2402.08939 [3] Pezeshkpour, P., Hruschka, E., 2023. Large Language Models Sensitivity to The Order of Options in Multiple-Choice Questions. https://doi.org/10.48550/arXiv.2308.11483 [4] Zheng, C., Zhou, H., Meng, F., Zhou, J., Huang, M., 2023. Large Language Models Are Not Robust Multiple Choice Selectors. http://arxiv.org/abs/2309.03882 [5] Oren, Y., Meister, N., Chatterji, N.S., Ladhak, F. and Hashimoto, T., Proving Test Set Contamination in Black-Box Language Models. In The Twelfth International Conference on Learning Representations. [6] Liu et al. (2024). Lost in the middle: How language models use long contexts [7] Malo, P., Sinha, A., Takala, P., Korhonen, P.J. and Wallenius, J., 2013. Good debt or bad debt: Detecting semantic orientations in economic texts. CoRR abs/1307.5336. URL: http://arxiv. org/abs/1307.5336.	Summary: The paper addresses the problem of order dependency in Large Language Models (LLMs), which causes inconsistency in outputs when the order of semantically identical sub-sequences is changed. The authors propose a technique that eliminates order dependency in transformer-based LLMs. The method is both theoretically sound and experimentally validated, showing minimal impact on accuracy despite the inputs being out of distribution. The technique can be seamlessly integrated into fully trained models. The paper also explores the potential of enhancing LLMs by incorporating metadata information similar to positional encoding. Strengths: - The proposed method elegantly and effectively addresses the order dependency issue in LLMs with only a minor performance drop. - The method is supported by thorough experimental evaluation and theoretical guarantees, adding to its credibility. - The paper is well-written, clearly explaining the methodology and its implications. - Section 5.2 introduces an innovative idea of improving LLMs by using metadata information in a manner akin to positional encoding, which opens up exciting new avenues for research. Weaknesses: The review does not identify any significant weaknesses in the paper. However, there are a few points of clarification needed, as outlined in the Questions section. Technical Quality: 4 Clarity: 4 Questions for Authors: - What is the impact of not enumerating the candidate answers (e.g., using A, B, C, D) versus simply using “”? This aspect does not appear to be addressed in the main text. - Regarding Figure 3, the proposed method seems to perform better than the worst-case scenario. However, if the goal is to achieve the highest score, it appears that the order-dependent method might still be preferable. Can the authors elaborate on this? - Why do the results differ in llama3? A more detailed explanation would help clarify this variation. This paper presents a significant advancement in addressing order dependency in LLMs. The method is both theoretically and experimentally robust, providing a practical solution with minimal performance trade-offs. The clear writing and innovative approach, particularly in Section 5.2, add to the paper’s strengths. Addressing the questions and clarifying the identified points would further enhance the paper. Overall, this work is a valuable contribution to the field and is strongly recommended for acceptance. [edit: I increase my rating to Very Strong Accept, see my comment on author's rebuttal for justification] Confidence: 4 Soundness: 4 Presentation: 4 Contribution: 4 Limitations: Limitations are adressed in the paper Flag For Ethics Review: ['No ethics review needed.'] Rating: 9 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thank you for the kind review and we are very glad to see other people are excited by the directions this work suggests. As the other reviewers posed many questions, we have some more results to show in the general rebuttal document. # Question Responses ## Impact of Enumeration We didn’t test with the questions numbered in the paper since that leaks positional information. We did some testing to see what quoting methods worked best, notably finding that no quoting has significant degradation across all models and conditions. We re-ran the first 20 MMLU questions (our test set for Figure 4) with the numbers added. This does improve performance across the board. We believe this is due to the models being trained with numbers on multiple choice questions. The order-independent results were within the error bars as in the other tests, but it’s harder to interpret those results as there is positional information leakage. To see the figure please look at the PDF for the main response. ## Figure 3 Yes, without fine-tuning our method is likely to slightly underperform non-parallel inputs, but as we and others have shown the order dependent method leads to a significant increase in variance under reordering changes. Our technique reduces variance. To make the bias reduction concerns more concrete, consider a system that uses an LLM to compare essays. This system will likely be biased towards preferring the earlier essays (thanks to the Lost Middle Phenomenon). So if the essays are provided in alphabetical order the system will likely be biased against Chinese writers due to their names more often being later in the (English) alphabet. Making systems that don’t have these subtle biases is an important part making deployment of LLMs possible in the real world. ## Llama3 We found the lower performance on Llama3 very interesting too. We hypothesize that this is due to Llama3 being overtrained relative to Llama2, making it more sensitive to out of distribution inputs. Interestingly, qualitatively we observed that our method also tends to make the chat fine-tuned models behave more like their base models (less terse responses was the main quantitative observation), suggesting that the fine-tuning was also partially interfered with by our method, resulting in the model responding like its base under our method. We think this suggests some methods for evaluating the robustness of trained models, and will add some discussion of these results to the main paper. # Final Thoughts Thank you for your comments and new experiment suggestion. If there are any more questions we will try to answer them during the discussion period. --- Rebuttal Comment 1.1: Comment: Thank you for the rebuttal and the clarifications to my questions. I must admit that I am surprised by the low ratings given in the other reviews. This paper proposes an elegant and creative method for addressing an important issue observed in LLMs (the authors convincingly explained why order-dependency in language models is problematic). It is supported by thorough experimental evaluation and theoretical guarantees. Moreover, it paves the way for promising applications that extend beyond the specific case of multiple-choice questions, as discussed in Section 5.2 and supported by the additional results for in-context learning provided during this rebuttal. Given these strengths, I believe this paper deserves to be accepted at the conference. Based on the rebuttal, I have decided to increase my initial rating to 9, hoping that this might encourage other reviewers to reconsider their ratings. --- Reply to Comment 1.1.1: Comment: Thank you for the kind response. We agree with your summary of this work.	Summary: The paper aims to address an issue within large language models (LLMs): their sensitivity to the order of input sequences, known as order dependency. This problem causes LLMs to produce inconsistent outputs when the order of semantically identical inputs is changed. ### Key Contributions: 1. Set-Based Prompting Technique: The authors propose a novel technique called Set-Based Prompting, which ensures that the output of an LLM is invariant to the order of a specified set of sub-sequences. This method modifies the input representation by removing the order information from the inputs. 2. Theoretical Guarantees: The paper provides theoretical proofs demonstrating that Set-Based Prompting eliminates order dependency for any transformer-based LLM. 3. Empirical Evaluation: The authors test their method on multiple LLMs, including GPT-2, Llama 2, Llama 3, and Mistral, using tasks like multiple-choice questions from the CommonSenseQA and MMLU datasets. They show that while Set-Based Prompting can slightly impact performance, it generally maintains accuracy within acceptable bounds and eliminates the variation caused by different orderings. ------------------------------------------------------------------------ Thank you for your replies and I raised my score accordingly. Strengths: ### Strengths - Theoretical Rigor: The paper includes rigorous theoretical proofs that validate the effectiveness of Set-Based Prompting in eliminating order dependency. - Empirical Validation: The extensive empirical evaluation on multiple models (GPT-2, Llama 2, Llama 3, and Mistral) on two datasets (CommonSenseQA and MMLU) demonstrates the robustness and applicability of the method. Weaknesses: ### Weaknesses #### Implementation Complexity - Engineering Trick Perception: The method may be perceived as an engineering workaround rather than a fundamental advancement. It requires users to input sub-sequence information, which can be seen as an additional burden and may limit the method's practical applicability. Providing a more automated or integrated approach could make the technique more user-friendly. #### Limited Improvement in Performance - Marginal Accuracy Gains: While the method ensures order independence, the actual improvement in model performance (accuracy) is limited. The results show that Set-Based Prompting maintains accuracy within the variation caused by different orderings but does not significantly enhance it. Highlighting specific scenarios where this method offers substantial performance gains could strengthen the paper's impact. #### Dependency on User Input - Manual Sub-Sequence Specification: The requirement for users to specify sub-sequences manually can be a significant limitation. This dependency on user input reduces the method's usability and scalability, particularly for large datasets or applications where manual specification is impractical. Exploring ways to automate the identification of sub-sequences would be a valuable enhancement. #### Evaluation Scope - Limited Dataset Variety: The evaluation is primarily conducted on multiple-choice question datasets (CommonSenseQA and MMLU). While these are standard benchmarks, the scope of evaluation could be broadened. #### Theoretical vs. Practical Benefits - Practical Utility of Theoretical Guarantees: While the theoretical guarantees are robust, the practical benefits might not be as compelling if the accuracy remains within the range of existing variations. Providing more concrete examples or case studies where order independence significantly improves real-world applications could make the practical utility of the method more apparent. #### Misleading Paper Title - Title Naming Concerns: I do not think that the proposed method is kind of "prompting" paper. It does change the internal implementation of LLMs instead of simply "prompting" them. Technical Quality: 3 Clarity: 2 Questions for Authors: What will be the benchmark performance under chain of thought prompting of this method compared to default chain-of-thought prompting baseline? Confidence: 4 Soundness: 3 Presentation: 2 Contribution: 2 Limitations: See more details in Weaknesses section. Flag For Ethics Review: ['No ethics review needed.'] Rating: 5 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thank you for taking the time to go over the paper and provide some useful feedback. # Relevancy Order dependency is an important open problem in the LLM NLP literature [1,2,3,4]. Known motivations include increasing model robustness [4] and reducing ordering biases [2]. To these we add applications from algorithmic fairness. To our knowledge, we are the first to obtain a theoretically rigorous solution. Moreover, we demonstrate that our solution incurs essentially no cost in accuracy. The robustness problem is a major concern for LLMs, and that they may only know the answer in one ordering [5] suggests that the models are not learning semantically useful representations. Having a method that forces the LLM to answer without using ordering means we can better evaluate LLMs. To make the fairness concerns more concrete, consider a system that uses an LLM to compare essays. This system will likely be biased towards preferring the earlier essays (thanks to the Lost Middle Phenomenon). So if the essays are provided in alphabetical order the system will likely be biased against Chinese writers due to their names more often being later in the (English) alphabet. Making systems that don’t have these subtle biases is an important part making deployment of LLMs possible in the real world. # Additional Comments The reviewer raises many other points in their review and we address them here. ## Engineering Trick Perception Most methods that interact with LLM internals are simple engineering tricks, since the systems are complex, and simple linear methods mean the interventions are well understood. For example, the recent Golden Gate Claude [6] is implemented as a simple linear addition to the model’s activations, and the added vector is learned via a single hidden layer autoencoder. More generally, linear probes are very common in the explainable AI literature and many works have been published about them. Thus, there is a strong tradition in deep learning using simple methods to enhance precision. Our technique, too, is a simple intervention; moreover, it enjoys a rigorous correctness proof. See Sub-Sequence Specification below for discussion of how our method can be easily integrated into existing workflows. ## Accuracy Gains We do not expect this method to improve accuracy. Our goal is reducing variance and information leakage. We are not providing additional information or additional “thinking time”, we are removing a cognitive defect from the LLM. The results shown in this paper greatly exceeded our hopes, it is very rare to find an intervention that doesn’t completely destroy the model’s performance. So we were hoping to get near the bottom of the error bars, not even fully within the range for all models. We think that this method having such a minor impact is in and of itself something that the community should be aware of, as we discuss a little bit in section 5, reviewer gPKU commented on this also. ## Sub-Sequence Specification We have implemented a method to input parallel sub-sequences that works like the special tokens already commonly used by most LLMs. Our system is implemented with three special “tokens”, a start parallel (`<\|start_2d\|>`), new sub-sequence (`<\|split_2d\|>`), and an end of parallel (`<\|end_2d\|>`). During inference our tokenizer does the splitting itself. This should be a simple addition to any tokenizer. The main difficulty for us was verifying there was no accidental information leakage. Making a nicer interface made the analysis much easier, as we simply ignore the parallel tokens during “normal” runs! The main results in the paper are all done with automated extraction of the parallel sequences, we expect automating the insertion of parallel sub-sequences to be straightforward as many LLM inputs are already constructed via concatenation of strings. ## Limited Dataset Variety We also have conducted two non-multiple choice question tasks at the request of the other reviewers. We have implemented a few-shot sentiment analysis task on which our method often outperforms the baseline, and a needle-in-a-haystack test where we match the performance of the baseline. Please see the general response for more details. These two new results will also be added to the paper. _Please see the global response for more details._ ## Theoretical Guarantees Theoretical guarantees are few and far between in this exciting literature. In our case, they lead to elegant solutions to a real-life problem, if the model is not robust to our method then it is likely not robust enough in the real world. As we discuss in the Relevancy section there are many areas where raw performance is traded for lower variance. Note, also that we only look at un-fine tuned models, we expect fine-tuning to significantly reduce the performance gap, if the models are robust. ## Title Naming Concerns We agree that the name can be improved! Thank you for pointing this out. Our current revised candidate name is __Order Independent Input Representations__ . We are happy to discuss this more during the discussion period. ## Chain of Thought Prompting In response to the reviewer’s suggestion we implemented a simple chain of thought prompt (“A: Let's think step by step”) for the first 20 MMLU question sets. We were happy to see that the results match the other MMLU results from the paper, with chain-of-thought providing a moderate uplift on performance for both order dependent and order independent cases! This was our first run with chain of thought, with no tuning. We have added the plot to the general response PDF and will include it in the paper, along with the code and data. # Final Thoughts Thank you and if there are any more questions we will try to answer them during the discussion period. # Citations _Please see global response_	null	null	Rebuttal 1: Rebuttal: We thank the reviewers for their thoughtful comments and are glad that everyone appreciates that we have obtained the first rigorously proven solution to an open problem – achieving order-independence – in NLP. Our solution is conceptually intuitive, and while the proof requires careful attention, it aligns well with this intuition. Attached hereto are the plots associated with the 4 additional experiments the reviewers mentioned. All four show that our method is well within the range suggested by the original experiments, indicating that our method is robust to variations in task as well as in model family. Notably, our method outperforms the best-of-two results for the non-parallel inputs on the few shot learning test. We provide additional, more tailored, responses to each interviewer individually. # Relationship to the Literature The order-dependency problem is well studied in the literature [1,2,3,4]. Prior work focuses on the robustness implications, as a model’s high sensitivity to input orderings mean that tests cannot be trusted to fully represent the model’s underlying understanding of the world. In this paper we add a new motivation for solving this problem: the fairness concerns raised by order dependency (for example, alphabetical ordering of candidates may lead to discrimination against individuals with Chinese names). Both these concerns are significant to the field; our method is the first to completely solve them, and our solution provably will always work. There has also been further work on the order dependency problem since this paper was submitted, one of the top cited papers out of ICLR this year looked at how order dependency reveals a lack of robustness in models [5], specifically they were interested in looking for training set contamination as the cause. # Solution Elegance Our method requires two interventions to the input representation, proving that these have the required effects takes a few pages since we prove our results for the general case of all transformer based LLMs. We took significant effort to make the methods clear and easy to understand and we are happy to see that our results have the property of seeming obvious in hindsight. Our implementation seeks to make our methodology easy to use, and our code release implements the system as a modification to the tokenizer, meaning that we simply add three special tokens (start, break, end) to our texts to make sub-sequences order independent. # Limited Dataset Variety Our paper is addressing works highlighting the order dependency problem in multiple choice questions as those are the most common benchmarks for LLMs, so we focused our efforts on providing a broad set of models. Additionally MMLU tests a broad range of tasks (math, law, general knowledge), we reported only the top level numbers due to space constraints, but it is often broken down into the 57 different tests. We also showed that our method allows for expanding the context window without any additional training (see Sec 4.3), as this paper was primarily focused on addressing the problems on the same terms as the other works in the literature; we only mention this briefly. We will add more details in the final paper on the expanded context window results. # Figures The attached PDF has 4 figures. Note that for all figures we will provide code and data release. We focused on the Llama and gpt-2 models for these analyses. If the reviewers have any questions we will seek to address them in the discussion period. ## Fig 1, Chain of Thought Prompting We implemented a simple chain of thought prompt (“A: Let's think step by step”) for the first 20 MMLU question sets. We were happy to see that the results match the other MMLU results from the paper, with chain-of-thought providing a moderate uplift on performance for both order dependent and order independent cases! This was our first run with chain of thought, with no tuning. ## Fig 2, Impact of Enumeration We re-ran the first 20 MMLU questions (our test set for Figure 4) with the numbers added. This shows improved performance across the board. We believe this is due to the models being trained with numbers on multiple choice questions. ## Fig 3, Long-Context Information Retrieval Accuracy on extracting the correct pieces of information from a set of 10 documents. The 0 parallel batches case is the fully order dependent model, with others being different splitting locations for the parallel batches. Note that order dependent model shows the characteristic reduction in performance for documents further from the start. While the less order dependent inputs perform as well or better than the worst case for the order dependent inputs. The fully order independent inputs lead to significant degradation, as that is a full 10 parallel inputs. We hypothesize that this is due to the inputs being too far out of distribution. For this analysis we looked at Llama 3 8B Instruct as the non-instruct models struggled with this task. ## Fig 4, In-Context Learning We implemented a sentiment classification task with in-context learning. The model was provided with 4 samples with labels, then asked to classify the fifth. The dataset is the Financial Phrase Bank [7] and with model attempting to classify them as positive or negative. To look at the effects of ordering on the model performance we always had three samples with the same label and 1 with the other label, with the different label always being first or last. When we ran this test, the in-context learning improved over the 1 shot test, showing that the models are using the examples. We did 4 batches of 100 questions, varying the valence of the 3 samples and the valence of the sample. We found that putting the examples in parallel often improves the accuracy over the non-parallel case, or is as good. This was a very strong showing by our method and we thank the reviewer for suggesting it! # Citations See CaKN response for citations. Pdf: /pdf/eba175d8d90c1e5561c0b895bb15c31d8001e0c2.pdf	NeurIPS_2024_submissions_huggingface	2,024	null	null	null	null	null	null	null	null
The Selective $G$-Bispectrum and its Inversion: Applications to $G$-Invariant Networks	Accept (poster)	Summary: The paper proposes a novel invariant layer for group-convolution networks (GCNNs) based on the "selective bi-spectrum". With respect to the expensive bispectrum which has $O(\|G\|^2)$ coefficients, the selective bispectrum only contains a subset of $O(\|G\|)$ coefficients which still ensure the completeness of the invariant features and can be computed in $O(\|G\| log \|G\|)$ when a FFT algorithm is available. Some preliminary experiments show the proposed method has comparable runtime and performance to the typically used max-pooling operator. Strengths: The manuscript is clearly written and the proposed method is original and interesting. The paper also includes some novel theoretical results, which could be useful for a wider audience. The empirical validation is in its very early stage, but the preliminary results are encouraging. Weaknesses: I feel like the empirical evaluation of the method is still too weak for acceptance. First, the authors only experiment with variations of the MNIST dataset: while this is a good dataset for some preliminary experiments to validate an idea is a simple and easily reproducible environment, I don't think it is sufficient to draw significant conclusions, especially about properties such as expressiveness of an operator or completeness of a representation. The authors should consider experimenting with more realistic datasets: for example, previous works typically used medical datasets, which still show rotational symmetries but provide higher resolution images and more challenging classification tasks. Because the contributions of this paper include the theoretical description of the selective bispectrum for some 3D groups, the authors should include some experiments on 3D datasets to validate that too; these datasets could also provide more convincing evidence of the value of the proposed method. Second, even if larger datasets were used, the current experimental setting doesn't provide conclusive answers on when one should use max-pooling or bi-spectrum. Indeed, the comparison in Fig. 5 is a bit misleading as the accuracy is plotted as a function of the number of filters K, but the computational cost of the two methods is different when varying K. Instead, a plot showing the accuracy vs (some proxy for) the computational cost seems a more suitable choice, drawing a more complete picture and quantifying the trade-offs discussed between lines 259-263. Overall, while I am very positive about the proposed idea, I think that deserves a much more thorough empirical validation. Technical Quality: 2 Clarity: 3 Questions for Authors: Fig. 6: this seems just a qualitative results useful to demonstrate the robustness of the selective bispectrum, but a proper quantitative validation of the robustness is still missing. A more comprehensive and thorough comparison between selective bispectrum, normal bispectrum and, maybe, max-pooling seems necessary. Why not reporting some statistics aggregated over a much larger dataset of random signals rather than visualising just 6 examples? It would be even more convincing to do so using the deep features extracted by a GCNN from a dataset of natural images, to prove the completeness is relevant beyond simple synthetic settings. Theorem 4.5: Is it correct that the (4) G-Bispectral coefficients are matrices? I find the choice of counting the matrices instead of individual scalars to be a bit misleading, as these matrices have very different shapes, making this number not very informative. Why not reporting only the number of scalar coefficients? This is a quantity which is meaningful across different groups and that can be directly compared with the group size \|G\| (the Fourier transform over the group G always has \|G\| coefficients, regardless of which finite group we consider). Theorem 2.4: alpha is used but never defined Table 1: why is K not included inside the big-O notation? Line 242: "... so that the G-Bispectrum scales worth". This sentence is a bit unclear. Confidence: 5 Soundness: 2 Presentation: 3 Contribution: 3 Limitations: The authors could discuss practical challenges in an efficient or parallel implementation of the selective bispectrum operator due to the sparsity and irregular shapes of the matrices involved. Flag For Ethics Review: ['No ethics review needed.'] Rating: 4 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Dear Reviewer Wppf, We thank you very much for your review and for your time. We address your comments and questions below. ### Experiments on larger datasets We address this concern in our global response to all reviewers. ### Accuracy versus computational cost We agree that showing accuracy versus computational cost is necessary to make our main point much stronger. We have added a scatter plot quantifying this trade-off for the different pooling methods. We describe it in our global reponse to all reviewers. We thank you for the great suggestion. ### Question on evaluation of robustness/completeness Thank you for raising this important point. We appreciate your question, as it helped us realize that the purpose of Figure 6 was not clearly presented. We clarify it here, and also refer you to the additional experiments on robustness/completeness that we discuss in our global response to all reviewers. Figure 6 is _not_ meant to be a _validation_ of the completeness of the G-Bispectrum, but rather an _illustration_. This is why, as you correctly noted, Figure 6 is qualitative and not quantitative. Our point here is that Theorem 4.1 (completeness of the selective G-bispectrum on $C_n$) is known to be true since it is proved: all signals that have the same selective $G$-bispectrum are known to be identical up to group action. Hence, numerical experiments can only validate it. In consequence, Fig. 6 is not a statistical validation but an illustrative example to convince the reader. This is why we visualize examples and do not report statistics aggregated over a much larger dataset of random signals. Besides, to provide a more comprehensive picture of the robustness of the selective $G$-bispectrum, we have added in the "Answer to all reviewer" and the PDF attached an experiment where, this time, input image are recovered. We hope this re-framing of Figure 6 clarifies its purpose. We will integrate this wording into our revised manuscript. Additionally, we provide novel experiments illustrating robustness and completeness in our global response to all reviewers. ### Question on memory output: report the number of scalar coefficients Yes, the G-bispectral (Bsp) coefficients are matrices in general, although they are scalars for commutative groups. We originally chose to report the number of G-bispectral coefficients to report a single value across groups, and also because the size of these matrices does not change the O complexity. Yet, you are absolutely right that it would be very useful (and potentially less misleading) to report the number of scalar coefficients as well. We do so below. The number of scalar coefficients in the G-TC is $K * \|G\| * \frac{\|G\| + 1}{2}$. This formula applies to all the groups discussed on our paper: $C_n$, $D_n$ the octahedral group $O_h$ and the full octahdral group $FO_h$. The number of scalar coefficients in the Max G-pool is $K$, also the same across the groups $C_n$, $D_n$, $O_h$ and $FO_h$. The number of scalar coefficients in the selective G-Bsp varies depending on the group. - $C_n$: $K * \|G\|$ scalar coefficients, - $D_n$: $\approx K * 4\|G\|$ scalar coefficients, - $O_h$: $K * 172$ scalar coefficients ($\approx K* 7 \|G\|$), - $FO_h$: $K334$ scalar coefficients ($\approx K 7 \|G\|$). These numbers are computed by considering which the G-Bsp coefficients are used in the selective G-Bsp (which is given in our theorems and proofs), and by summing the scalar coefficients of their respective matrices. ### Discussion on practical challenges in parallel implementation Thank you for raising this point. Our main contribution is about computational gains, made possible thanks to new mathematical theorems on group theory. Thus, we agree that it is important to discuss other ways to obtain computational gains. Parallel computations is one of them. Following your suggestion, we discuss this point below. (Notation: Bsp = Bispectrum) To further gain in computational efficiency, we could compute the G-Bsp coefficients in parallel. This can be done both for the full G-Bsp and for the selective G-Bsp. Indeed, for both, we know _which_ G-Bsp coefficients are needed: we need all of them for the full G-Bsp (or, half of them, given the symmetry rho_i, rho_j <> rho_j, rho_i), and a subset of them for the selective G-Bsp where the subset is given by our theorems. Consequently, each G-Bsp coefficient can be computed with a separate process. One would need to account for the fact that each G-Bsp coefficient has a different memory footprint. Indeed, each G-Bsp coefficient is a matrix with a size depending on the nature of the coefficient. Since this parallelization would speed up both the full G-Bsp and the selective G-Bsp, the selective G-Bsp would still be faster. Lastly, since the matrices defining the G-Bsp coefficients are generally non-sparse, no further computational gain that can be obtained there. ### Notations Thank you for reporting the typos. We have corrected these. ### Conclusion Again, we thank you for the thorough review and the very interesting questions. Your feedback has been beneficial for us, and we believe that the additional experiments and discussions significantly strenghten our paper. If you have any remaining questions or comments, we would be happy to discuss further! We hope that we have addressed your concerns and we look forward to discussing these topics further. --- Rebuttal Comment 1.1: Comment: I thank the authors for the detailed answer. I understand the authors' argument about the simple and controlled experiments, but I do not see that as a good reason why not also including a larger scale experiment to validate the method. Indeed, 1) it is common to test equivariant networks on much larger and more complex datasets than MNIST by augmenting them with group transformations and still ensuring that we know exactly which groups structure the data follows (see again the recommendations in my original review). 2) I still believe that MNIST is not sufficient to draw strong conclusions about properties such as expressiveness of an operator or completeness of a representation, due to its simplicity. Unfortunately, while I am very positive about the theoretical value of this work, I think a stronger empirical evaluation - validating the important claims about the proposed method - is necessary for the acceptance of the paper. --- Reply to Comment 1.1.1: Comment: Dear reviewer WWpf, Thank you for reading and engaging with our rebuttal. We appreciate your time. While we respectfully maintain that large-scale dataset experiments may not be strictly necessary for theory-heavy research contributions, we understand and acknowledge that additional empirical validation could be valuable and convincing to you and potentially a larger class of our future potential readers. Consequently, to enhance the impact of this work, we will explore the possibility of conducting experiments on a larger dataset and will provide an update as soon as possible.	Summary: The paper addresses the problem of achieving invariance to nuisance factors in signal processing and deep learning, particularly those describable by group actions (e.g., rotations, translations). The authors propose the selective G-Bispectrum, a computationally efficient variant of the G-Bispectrum, which reduces complexity from $O(\|G\|^2)$ to $O(\|G\|)$ in space and $O(\|G\| \log \|G\|)$ in time when an FFT is available on the group $G$. Strengths: The paper presents a novel approach to reducing the computational complexity of the G-Bispectrum, a key operation in achieving G-invariance in signal processing and deep learning. This reduction is significant, lowering the complexity from $O(\|G\|^2)$ to $O(\|G\|)$ in space and from $O(\|G\|^2)$ to $O(\|G\|\log\|G\|)$ in time, assuming the availability of an FFT on $G$. The selective G-Bispectrum is a creative combination of existing ideas in group theory and signal processing, applied in a new way to enhance deep learning architectures. This work stands out by addressing computational limitations that have hindered the widespread adoption of the G-Bispectrum. The paper is well-structured and thorough in its theoretical contributions. The authors provide rigorous proofs of the mathematical properties of the selective G-Bispectrum. The significance of this work lies in its potential to advance the field of geometric deep learning and signal processing on groups. By reducing the computational complexity of the G-Bispectrum, the authors make this tool more practical for real-world applications. Weaknesses: The paper primarily evaluates the proposed selective G-Bispectrum on the MNIST and EMNIST datasets. While these datasets are standard benchmarks in machine learning, they are relatively simple and may not fully stress-test the proposed method’s effectiveness in more complex and varied scenarios, especially larger group such as 3D group. Although the paper claims that the selective G-Bispectrum has higher computational efficiency, it does not explicitly compare the computational costs and performance together in the experiments. This omission undermines the argument for computational efficiency, particularly given that the performance of the selective G-Bispectrum is not as strong as the G-TC method. Technical Quality: 2 Clarity: 3 Questions for Authors: Question: It is mentioned in the paper that the selective G-Bispectrum offers greater selectivity and robustness. Could you provide experimental evidence to validate these claims? Suggestion: Include experiments and analyses specifically designed to demonstrate the greater selectivity and robustness of the selective G-Bispectrum. This could involve: - Selectivity: Experiments that show how well the method distinguishes between different classes or handles variations within the same class. - Robustness: Tests under various noise conditions, transformations, or adversarial attacks to demonstrate the method's stability and reliability. Question: Why were only MNIST and EMNIST datasets used for the evaluation? Would the selective G-Bispectrum perform similarly on more complex datasets? Suggestion: Expand the experimental section to include more diverse and challenging datasets, such as CIFAR-10, ModelNet10. This will provide a more comprehensive assessment of the method's effectiveness. Confidence: 4 Soundness: 2 Presentation: 3 Contribution: 2 Limitations: yes Flag For Ethics Review: ['No ethics review needed.'] Rating: 5 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Dear Reviewer ZVp5, We thank you for your time and thoughtful review. We address your comments and suggestions below. ### Experiments on larger datasets Thank you for raising this point. We comment on it in our global response to all reviewers. ### Showing accuracy versus computational cost We agree that showing accuracy versus computational cost would make our points much stronger. We have added a scatter plot quantifying this trade-off for the different pooling methods. See global response to all reviewers. ### Showing "selectivity" Indeed, we mention that the G-Bispectrum offers greater selectivity in our sentence: "the $G$-Bispectrum has been incorporated into deep neural network architectures as a computational primitive for $G$-invariance - akin to a pooling mechanism, but with greater selectivity and robustness." In this sentence, we use the word "selectivity" to refer to the completeness property of the G-Bispectrum. The G-Bispectrum of a signal is complete in the sense that it removes only the variation due to the action of a group G on the signal, while preserving all information about the signal's structure: it _selectively_ removes the group action without destroying the representation. We have provided empirical evidence to validate this claim in our results presented in Figure 6 and the new experiment from the PDF attached. In contrast, we believe that you understood the word "selectivity" as it is commonly used in machine learning, where it refers to "specificity", the true negative rate (TNR) of a classification task. We apologize for our poor wording that led to this confusion. To avoid misunderstanding, we propose to rephrase our original sentence as follows: "the $G$-Bispectrum has been incorporated into deep neural network architectures as a computational primitive for $G$-invariance - akin to a pooling mechanism, but with greater selectivity and robustness." Since we do not claim greater "specificity" in the classification sense, we have not provided results on this front. However, we can show the confusion matrices of our classification results in supplementary materials. We hope that we have addressed this concern. Let us know if we should provide additional details here. ### Showing robustness Thank you very much for this suggestion. We have added an experiment in our global answer to all reviewers accordingly. ### Conclusion Again, we thank you for your time and attention. We hope that we have addressed your concerns and we look forward to discussing these topics further. --- Rebuttal Comment 1.1: Title: Additional experiments addressing your concern Comment: Dear Reviewer ZVp5, Following our post-rebuttal discussion with reviewer Wppf, we have conducted additional experiments using the more complex CIFAR-10 dataset. These results are detailed in our global response to all reviewers and address one of your initial concerns. These experiments also complement our other new experiments on robustness and our addition of a cost-accuracy analysis plot – both of which were incorporated in response to your insightful feedback. Addressing your comments has been very beneficial to us, as these additions have strengthened our paper. We thank you for engaging with our work! If you agree that we have adequately addressed several of your original concerns, may we kindly ask if you would consider raising your score to support the acceptance of our paper?	Summary: The work focuses on the problem of achieving invariance to extraneous nuisance variables in signal processing and deep learning by utilizing group actions such as rotations and translations. The focus is on the G-Bispectrum, which is a tool used to extract signal characteristics that are invariant to such actions. This tool provides advantages similar to pooling techniques, but with more selectivity and robustness. The G-Bispectrum has been limited in its application because to its high computational cost, which is O(\|G\|2). However, this study presents a strategy to lower this complexity to O(\|G\|) by proposing selective G-Bispectrum. The novel methodology preserves the mathematical characteristics of the G-Bispectrum and enhances the precision and robustness of neural networks, while simultaneously delivering substantial speed ups. Strengths: The research presents a clear motivation, which is to decrease the computational complexity of the G-Bispectrum from O(\|G\|2) to O(\|G\|) without compromising its effectiveness as a computational primitive for G-invariance. The emphasis on enhancing computing efficiency while maintaining performance is clearly expressed and pertinent. The proposed method is technically valid. The assumptions employed are rational, and the research establishes a robust theoretical basis for the selected G-Bispectrum, effectively showcasing its mathematical qualities such as G-invariance, completeness, and uniqueness. The paper is very well-written and easy to follow. Figure 1 enhances the clarity of the presentation by providing a comprehensive overview of the motivation for the proposed strategy and successfully conveying the essential themes. The research presents extensive empirical evidence showing that the selective G-Bispectrum achieves significant improvements in speed compared to the whole G-Bispectrum in real-world scenarios. The enhancement in efficiency renders the selected G-Bispectrum a feasible and efficient substitute for attaining G-invariance in deep learning models. Weaknesses: The proposed approach is currently limited to discrete groups. Several contemporary equivariant networks have been applied in domains that involve continuous groups like $SO(3)$ and $SE(3)$. However, the current technique does not readily adapt to these situations, and the paper does not investigate or experiment with such extensions in these domains. Could the authors comment on this limitation and discuss whether it could be a potential direction for future research? Furthermore, the tests conducted in this research are restricted to small-scale datasets such as MNIST. There is the possibility of analyzing larger and more intricate datasets in order to more effectively showcase the advantages of the suggested approach. I strongly urge the authors to contemplate using their methodology on larger-scale image datasets in order to demonstrate its efficacy in a more thorough manner. Minor In line 183, novel theorems, Although the coefficients may not be precisely zero, they could be in close proximity to zero. When performing numerical computations, this can result in problems related to stability and precision. The formula presented in Theorem 4.4, which states that the number of bispectral matrix coefficients is equal to ⌊n-1/2⌋ + 2, may benefit from a more comprehensive explanation for readers who are less acquainted with dihedral groups. The sentence in Theorem 4.4, "This corresponds to 1 + 4 + 16 · ⌊n-1/2⌋ ≈ 4\|Dn\| scalar values," may be clarified by providing a concise explanation of the derivation of this approximation. Technical Quality: 3 Clarity: 3 Questions for Authors: The proposed approach is currently limited to discrete groupings. Various contemporary equivariant networks have been utilized for continuous groups like $SO(3)$ and $SE(3)$. Could you provide an analysis of the constraints associated with applying your approach to these continuous categories? Do you anticipate any theoretical or practical obstacles, and do you view this as a promising avenue for future investigation? Recommendation: It is advisable to evaluate the suggested technique on image datasets of greater magnitude in order to demonstrate its efficacy. To better showcase the benefits of your technique, it is advisable to conduct trials using established benchmarks or more intricate datasets. Question: What is the difference between the suggested selective G-Bispectrum and contemporary equivariant networks tailored for continuous groups? Are there any particular benefits or drawbacks that might be emphasized in the context of ongoing group actions? Confidence: 3 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: NA Flag For Ethics Review: ['No ethics review needed.'] Rating: 5 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Dear Reviewer UtSC, We thank you for your review and insightful comments and suggestions. We address your points below. ## Continuous Groups We thank you for asking the very interesting question of how to generalize to continuous groups. The formulas defining the full and selective G-Bispectra (G-Bsp) only depend on the G-Fourier transform. Since the G-Fourier transform can be defined for continuous groups, so can the G-Bsp. However, practical difficulties arise upon implementation. Indeed, the full G-Bsp has one G-bispectral coefficient per irreps of the group G. While discrete groups have a discrete number of irreps, continuous groups can have an infinite number of them. For example, SO(3) has an infinite number of irreps. Accordingly, because of memory constraints, we cannot compute the _full_ G-Bsp for SO(3), nor for other continuous groups with an infinite number of irreps. Interestingly, the _selective_ G-Bsp that we propose does not require all of the G-bispectral coefficients. Thus, it does not require to use all the irreps of the group G. As a result, it is possible that we could compute the _selective_ G-Bsp for continuous groups. Let us, therefore, assume that the selective G-Bsponly requires a finite number of irreps for a given continuous group G. In this case, we face another obstacle. This obstacle appears when we want to use this G-Bsp as the pooling mechanism of a G-convolutional neural net - as done in our classification experiments. Indeed, the G-Convolution that preceeds the (selective) G-Bsp is a so-called regular G-convolution (Group Equivariant Convolutional Networks. Cohen & Welling, 2016), which uses a discretization of the group G. Thus, the output $\Theta$ of the G-convolution would be a signal over a discrete/discretized group. Consequently, we would not be able to use the continuous selective G-Bsp there. To overcome this obstacle, one could think of using a steerable G-Convolution, instead of a regular one, since steerable G-convolutions apply to continuous groups (Steerable CNNs. Cohen & Welling, 2017). In our paper, we chose to use regular G-Convolutions because: (i) we are in fact interested in discrete groups and, (ii) a regular G-convolution outputs a signal $\Theta$ defined over a group - much needed to apply the definition of the G-Bispectrum - whereas a steerable G-Convolution outputs a more complicated mathematical object for which the G-Bispectrum needs to be generalized. We discuss this aspect below. Consider a group $G$ that is the semi-direct product $G=\mathbb{Z}_2 \ltimes H$ of the group of translations $\mathbb{Z}_2$ and a group $H$ of transformations that fixes the origin $0 \in \mathbb{Z}_2$. Consider the output of the steerable G-Convolution, which is a map $\Theta: \mathbb{Z}_2 \rightarrow \mathbb{R}^K$, where $K$ is the number of filters in the convolution. The signal $\Theta$ has the following mathematical property: $\Theta$ is a $K$-dimensional field over the grid $\mathbb{Z}_2$ that transforms according to a specific representation $\pi$ of the group $G$. This representation is induced by a representation $\rho$ of its subgroup $H$ on $R^K$ (Steerable CNNs. Cohen \& Welling, 2017), such that it is common to write $\pi=\mathrm{Ind}\_H^G \rho$. Consequently, the interaction between the signal $\Theta$ and the group $G$ is very different for the steerable case, compared to the regular case. - For a signal $\Theta$, the action of an element $(t, r) \in G = \mathbb{Z}^2 \ltimes H$ is: $[\pi\_{t, r} \Theta] (x) = \rho(r)\Theta((t, r)^{-1} x) \in \mathbb{R}^K.$ - By contrast, when we use regular G-convolution, the action of an element $g \in G$ on each coordinate of the signal $\Theta$ is: $[g \ast \Theta] (g') = \Theta((g^{-1} g) \in \mathbb{R}.$ The interaction between the signal $\Theta$ and the group $G$ is central to the definition of a G-Bsp. In fact, there is currently no definition of a G-bispectrum for the general steerable case. However, there is a definition of G-Bsp for signals defined over the homogeneous space of a group $G$ by Kakarala, which corresponds to a special steerable case for which $\rho$ is the trivial representation: $\rho(r) = r$ for all $r \in H$. Therefore, generalizing to continuous groups involves i) determining whether the selective G-Bsp only requires a finite number of coefficients, and for the classification experiments: ii) using a steerable G-convolution with trivial $\rho$ together with the G-Bsp for homogeneous spaces. These ideas could provide fruitful paths for exploration in future work. We believe, nonetheless, that presenting the theory for discrete groups is an important first step and should remain the focus on our work. Relatedly, in geometric deep learning, the paper Group-Equivariant Convolution Networks (Cohen & Welling, 2016) introduced the G-CNN for discrete groups. Only later did the paper Steerable CNNs (Cohen & Welling 2017) present a theory that applies to continuous groups. We will gladly add this discussion to the final version of the paper, thank you for this suggestion. ## Larger datasets Thank you for raising this point. We discuss it in the global answer to all reviewers. # Minor - You are right to mention that numerical zeros are a recurrent issue in the field of numerical computations, especially when used for division. However, the G-Bsp relies on additions and multiplications. Hence, we did not encounter such problems in our experiments. - Thank you for raising this point about the number of scalar coefficients. In the revised paper, we add a paragraph clarifying the number of bispectral matrices and the related number of scalar coefficients: see our answer to reviewer Wppf. Specifically, the approximation you mention comes from the fact that $\|D_n \|=2n$. Hence, $16\frac{n-1}{2} \approx 4\|D_n\|$. # Conclusion Again, we thank you for your time and attention. We hope that we have addressed your concerns. We look forward to further discussing these topics. --- Rebuttal Comment 1.1: Title: Correcting a typo + Additional experiments Comment: Dear Reviewer UtSC, We would like to correct a typo in our discussion explaining how we can extend our approach to continuous groups, answering a very interesting question from your review. The trivial representation rho is rho(r) = 1 and not rho(r) = r, such that the corrected sentence is: “which corresponds to a special steerable case for which ρ is the trivial representation: ρ(r)=1 for all r∈H”. We apologize for this typo. We hope that our discussion of the extension to continuous groups picked your interest. We would also like to point you to our global answer to all reviewers which contains additional experiments on larger datasets (CIFAR-10). We thank you for your constructive feedback, which has been very beneficial to us as it has stimulated very interesting discussions. If you believe that we have adequately addressed your concerns, would you consider increasing your score to support the acceptance of our work?	Summary: The paper proposes a new invariant layer for group convolutional neural networks (G-CNN). The proposed layer computes the spectral coefficients of the input data, that is based on higher-order spectral analysis: the bispectrum, which is the Fourier transform of the triple correlation (G-TC). While there can be up to \|Irrep\|^2 different coefficients in the bispectrum, the main contribution of the work is an algorithm that finds a subset of at most \|G\| of these coefficients. For specific groups (general abelian, dihedral, (full) octahedral), the authors proved that there exists a subset of size at most \|G\| that can reconstruct the input data (signal) perfectly, and that their algorithm, with the right choice of \rho_1, can find this set. Experiments are conducted on MNIST and EMNIST, comparing the newly proposed layer with max/sum pooling and G-TC layer. Strengths: - Novelty. The use of higher-order spectral analysis is relatively new in designing deep learning architecture over group and can be powerful since they are originally used to capture nonlinear/autocorrelational structure of the input signal. Capturing nonlinearity is nontrivial in designing G-CNNs as pointwise nonlinearity is known to not be the most appropriate choice of nonlinearity. - Theoretical contributions, especially to nonabelian groups such as Dn and O are interesting and has substantial savings in the number of necessary coefficients to be efficient in practice. I did not go through the proofs line by line but believe that the proof idea is sound. - Self-contained. The paper, together with the appendix is relatively self-contained in introducing group-theoretic and representation-theoretic concepts. The overall exposition/English is also easy to follow. - Clear experiments and diagrams. The paper demonstrate a regime of small number of filters where their methods work much better than simple group pooling, which is a clear and instructive takeaway. Diagrams to demonstrate algorithms are also clear and make for a quicker read. Weaknesses: The main issues I have with the current manuscript is its motivation and some ill-defined notion. Issues with ill-defined notions: - The term "selective G-bispectrum" is not well-defined, despite being the central notion of the whole paper. The authors defined this notion by giving an algorithm that depends crucially on a hyperparameter \rho_1. The authors themselves have noted that the choice of \rho_1 will make or break the resulting set of coefficient (being a complete system). Thus, completeness is a property of the algorithm + choice of \rho_1 and it does not make sense to claim, for e.g., that the selective G-bispectrum is complete without specifying this choice of irrep. - Algorithm 1 depends crucially on computing the Clebsch–Gordan matrices, which may be very costly in general for nonabelian groups. The authors have also acknowledged this point. Issues with motivations: - The authors motivate the paper by arguing that a complete (Definition in Theorem 2.4) invariance layer is more desirable than simple G-pooling layer (such as max or sum), which is an intuitive point. However, in the experiments, average-pooling, which is arguably more complete than max-pooling (or, to quote, less "excessively invariant") performs consistently worse than anything else, and to a large extent. Is there a good explanation for this? - The proofs of all completeness theorems essentially boils down to being able to reconstruct the group FT coefficients of the input signal from the subset of bispectrum coefficient, which begs the questions of what happens if you simply input these \|Irrep\| numbers of group FT coefficients as an invariant layer, instead of going to the bispectrum and hoping that it can reconstruct group FT coefficients? The answer to this probably lies in the reference "The work of Kakarala (1992) illuminated the relevance of the G-Bispectrum for invariant theory, as it is the lowest-degree spectral invariant that is complete (Sturmfels, 2008)" but scheming through Sturmfels' book did not lead me to any such conclusion. Can the authors provide a more specific reference, and/or discuss this point (why bispectrum but not just FT coefficients, in light of completeness) a bit more? This would be an interesting experiments to look at too. - Lastly, why spectral? Usually, signal processing literature makes use of spectral domain because of sparsity in that domain. The authors have also shown that it is faster to compute G-bispectrum coefficients than G-TC. These motivations (and perhaps others, if presence) to using spectral domain should be written up more clearly in the beginning of the paper. There are also various typos (e.g. line 26, R not defined in line 493, ...) that I did not carefully keep track of in the interest of time. Technical Quality: 3 Clarity: 2 Questions for Authors: See weakness section. Confidence: 3 Soundness: 3 Presentation: 2 Contribution: 3 Limitations: The authors have acknowledged some of the limitations in their methods and definitions (detailed in Weakness section). There are no potential negative societal impact that are specific to this work and are worth highlighting. Flag For Ethics Review: ['No ethics review needed.'] Rating: 7 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Dear reviewer P7yz, we wish to thank you again for your time. In what follows, we address the points you made in your review. # Issues with ill-defined notions: ## The term "Selective G-Bispectrum" You are right to mention that the "Selective G-bispectrum" admits degrees of freedom, that is, it is not uniquely defined and there are several "Selective G-bispectra". As you correctly pointed out, each choice of $\rho_1$ will yield a different "Selective G-bispectrum". In the revised version of the paper, we propose to clarify this point as follows. For the most common groups appearing in our theorems, we call "Selective G-bispectrum" _the_ bispectrum proposed in their respective proofs. Indeed, each theorem provides a precise set of G-bispectral coefficients. The Selective G-bispectrum will refer to that set. For other groups, we call Selective G-bispectra the G-bispectra obtained via our algorithmic procedure that yield the maximum number of G-Fourier coefficients. We thank you for this important remark on the main term of our paper. We will make sure to clarify this in the final version of our work. ## Computing the Clebsch-Gordan matrices You are correct mentioning that obtaining the Clebsch-Gordan (CG) matrices is not an easy task. However, this work needs to be done only once before the training. This is because these matrices are not data-dependent and can be stored. In consequence, obtaining the CG matrices is not a bottleneck. Additionally, we point out that - for the octahedral and the full octahedral groups - we only use the CG matrices in the proofs of our theorems. Specifically, we only use them to find out the minimal set of G-bispectral coefficients that is needed. In practice, when we implement the selective G-bispectra for these two groups, we do not even use the CG matrices. This is because the G-bispectral coefficients can be written in the following alternative form: $$\beta(\Theta)\_{\rho_1,\rho_2} = \sum\_{g_1 \in G} \sum_\{g_2 \in G} \sum\_{g \in G} \Theta^*(g) \Theta\left(g g_1\right) \Theta(g g_2)[ \rho_1(g_1)^{\dagger} \otimes \rho_2(g_2)^{\dagger}].$$ This expression comes from the definition of the G-bispectrum as the G-Fourier transform of the G-triple correlation. Since computational gain is an important part of our contribution, we feel that it is important that we discuss and clarify this important point in the final version of the paper. We thank you for this feedback. # Issues with motivations: ## Completeness of avg pooling versus max pooling We respectfully think there might me a confusion with the notion of completeness. You are right to notice that average pooling uses all data to be computed, however so does the max pooling since all data must be compared to retrieve the maximum. In consequence, the average of a list is not more complete than the maximum, one does not give more information about the distribution of a list than the other. ## Distinction between the G-bispectrum and the G-Fourier transform We are glad to precise the interest of the G-bispectrum w.r.t the Fourier transform (FT). The advantage of the G-bispectrum is its invariance to group action while the FT is not invariant: the FT is only equivariant. To achieve group-invariance of a neural network, the FT is thus not a interesting quantity, at least it does not provide specific improvement towards this goal. Kakarala (1992) proved the invariance of the $G$-Triple Correlation (G-TC) for any compact group $G$, and by extension, the invariance of the $G$-bispectrum. This line of research also proves that the G-TC is the unique, lowest-degree polynomial invariant map that is also complete. Consequently, to achieve completeness, the simplest approach is the use of the G-TC or of its Fourier transform: the G-bispectrum. This comment raises the follow-up question: why using the G-Bispectrum over the G-TC? This question is linked to your question "why a spectral approach?", which we answer below. ## Why a spectral approach? This is a very interesting question. Our motivation for proposing, and using, the "Selective G-bispectrum" is motivated by the $O(n\log(n))$ complexity that it allows to obtain. By comparison, the complexity of the G-TC is $O(n^3)$ and we were not able to design a "selective G-TC" to reduce it. Consequently, the selective G-bispectrum is the unique, lowest degree polynomial invariant map that is also complete and that is the most computationally efficient. We will insist more on this point in the final version of the paper. # Conclusion Again, we thank you for the attention given to our work and for your thorough review. We have seeked to address your concern and comments with this rebuttal. If the elements provided above enhance your appreciation for our work, may we kindly ask whether you would consider defending our paper and increasing your score? --- Rebuttal Comment 1.1: Comment: Thanks the author for the detailed reply and addressing my points. I personally think the paper is at acceptable quality for publication at the conference based on the theoretical contribution alone. I agree with other reviewers that the experiments (though clear and succinct) are still at 'proof-of-concept' stage (which explains why I'm not raising the score to award level). To make meaningful experimental contributions beyond verifying the theory, a lot more work would need to be done, such as ablation and testing on various architectures as well as larger datasets and on more varied tasks. This would make a nice extension perhaps for a journal publication but at the current stage of the manuscript, I believe keeping the current score is reasonable. --- Reply to Comment 1.1.1: Comment: Dear Reviewer P7yz, Thank you very much for your kind words and continued support. We appreciate that you value our theoretical contributions and, overall, our work. We are grateful for the attention you have dedicated to the review process!	Rebuttal 1: Rebuttal: We thank the four reviewers for their time and attention. All reviewers found our manuscript very well-written and easy to follow. We find this very encouraging especially given its substantial theoretical components in group theory. The reviewers also appreciated the novelty and depth of our theoretical results, the rigor of our proofs, and their importance for signal processing as a whole. We thank all reviewers for recognizing this contribution. We have provided responses to each reviewer for unique points made in the review. Here, we highlight a few points that appear across reviews, to avoid redundancy. ### Use of Simple, Controlled Datasets in our experiments Reviewers ZVp5 and Wppf asked for the rationale behind using simple datasets for our experiments, and asked for experiments on more complex datasets. We agree that the datasets we use here are relatively simple. We chose to use simple datasets in our analyses because they allow us to precisely control their mathematical structure. We generate our synthetic datasets by applying specific groups' actions to these data exemplars. This approach ensures that we know exactly which groups structure the data, enabling us to formulate theoretical expectations for the application of the selective G-bispectrum layer. In contrast, the CIFAR image dataset (mentioned by the reviewers) contains many variations not attributable to group actions, such as lighting changes and general differences in appearance across exemplars within the same category. Therefore, we theoretically expect less benefit from using group-invariant and group-equivariant structures with this dataset. Our goal in this work was to empirically demonstrate the theoretically expected properties of our proposed layer, which we believe we have achieved using (i) strong theoretical contributions with several new theorems, and (ii) simple, controlled datasets. ### Robustness to adversarial attacks Reviewers ZVp5 and Wppf asked for additional experiments to illustrate the robustness of the bispectrum. We first explain the link between completeness and robustness. Then, we provide new experiments and present their results here. First, we emphasize that the completeness of the the triple correlation (TC), and its Fourier transform, the bispectrum, is well-documented in signal processing literature (see Yellott & Iverson (1992) for translation-invariant case, and Kakarala (2009) for compact groups). The selective G-Bispectrum contains the same information as the G-bispectrum and is therefore also complete for the groups we consider. This completeness implies that it is inherently robust against invariance-based adversarial attacks. In such attacks, one finds non-identical inputs (up to group action) that produce identical output representations. Completeness prevents from this event, thereby ensuring robustness to these attacks. We provide empirical illustration of this property with a new set of experiments. Specifically, we perform invariance-based adversarial attacks on the models trained in our experiments. Our results are presented in Figure 1 of the PDF attached. They show that the model with selective G-bispectrum is complete and robust to these attacks, whereas the Max G-Pool model is not. Specifically, when we optimize input images to match a target selective G-bispectrum, obtained form a target image, every optimized input image is identical to the target image up to group action. By contrast, an adversarial agent can find several input images that would yield a target output in the context of the Max G-pool model because it is not complete. This analysis, initially established by Sanborn et al. (ICLR 2023), and later used in Sanborn & Miolane (2023) for the TC, empirically demonstrates the theoretically-expected completeness and robustness property in the trained model. ### Cost-Accuracy Trade-offs Reviewers ZVp5 and Wppf raised the concern that the trade-off between computational cost and accuracy was not explicit enough, since computational cost and accuracy were on two separate figures in the manuscript. We agree and address this point here. We now plot computational cost and accuracy on the same figure to draw a complete, quantified picture of their trade-offs. In the PDF attached, we provide a scatter-plot that shows the relation between computational cost (x-axis) and accuracy error (y-axis, log scale). Each dot is colored according to the pooling method, and we additionally vary the number K of filters, with K going from 2 to 20. Since the accuracy is monotone with K, one can additionally read the figure in K by going from top to bottom. This scatter-plot provides additional evidence of our claim that the selective G-Bispectrum is faster than the G-TC while being more accurate than the Max G-Pooling and the Avg G-pooling. In terms of speed, we see that the selective G-Bispectrum provides an average gain of around 15 seconds over the G-TC, which corresponds to around 10% improvement. In terms of accuracy, we see that the selective G-Bispectrum provides an average gain of 0.02 accuracy over the Max G-pooling, and 0.26 over the Avg G-pooling. We will include this plot in the main paper, and add corresponding plots for O2 and SO2/O2-EMNIST in the supplementary materials. We believe that they do, indeed, better quantify our claims. We thank the reviewers for the suggestion. ## Conclusion on experiments We believe that the additional plots and proposed additional experiments on robutness strengthen our manuscript, and we thank the reviewers for their important feedback. Yet, we would like to emphasize that our work also makes several _theoretical_ contributions, and thus should not be reduced to its experiments. We believe that the impact of our reduction of the group-bispectrum is profound, with consequences for applied mathematics, signal processing and machine learning that go beyond classification results on benchmarks. Pdf: /pdf/b8f5a27217220a3291ec99e20c4fe33d40f3874c.pdf	NeurIPS_2024_submissions_huggingface	2,024	null	null	null	null	null	null	null	null
SPEAR: Exact Gradient Inversion of Batches in Federated Learning	Accept (poster)	Summary: This paper presents a novel method to recover batched input data from gradients in fully-connected networks. Based on a detailed analysis of the rank of gradients, it utilizes matrix decomposition to first figure out the batch size and then recover the input tensor through sparsity. Good comparisons with previous methods have demonstrated the performance of such a method. Strengths: - Good writing. The structure is well-organized. - It is of great value to discover the mathematical relationships within the gradient matrixes, and then use matrix decomposition to solve this problem. - Comparisons with previous methods, both in the image and tabular data domains, make this work convincing. Weaknesses: Please refer to the Questions part. Minor problem: It is recommended that the work should be double-checked in case of typos or compiling errors. - Line 79 typo - Line 125 typo - Line 149 abnormal block Technical Quality: 4 Clarity: 3 Questions for Authors: - It seems this work relies highly on the sparsity. What if the activation function is sigmoid, leakyReLU, tanh, etc? These activations do not directly filter out negative inputs and return 0s, which may affect the sparsity property. - For experiments in 6.2, are the image labels known for the reconstruction ahead of time? - Why is it that this method may fail when batchsize is larger than 25? In my personal view, more insights could be covered about this direction in the failure probability part. Confidence: 3 Soundness: 4 Presentation: 3 Contribution: 3 Limitations: The proposed method only works for FCNs. Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We thank the reviewer for their very positive review. We are particularly happy to read that the reviewer assesses our contribution to be of great value and finds the comparison to related work convincing. We will include all proposed writing suggestions. We now address all questions of the reviewer. Q1: Can SPEAR handle activation functions different from ReLU? We direct the reviewer to our previous answer to question Q2 in the general response. In this work, we focus on the most common ReLU activation as it is the most relevant one in practice. Further, we believe that transferring our insights to other activations is a valuable item for future work. Q2: For experiments in 6.2, are the image labels known for the reconstruction ahead of time? No, we do not require the knowledge of labels at all. We do want to emphasize that the only prior we utilize is the ReLU-induced sparsity. Q3: Why can SPEAR fail when the batch size is larger than 25? Is this handled by the probability of failure analysis? We refer the review to our answer to Q1 in the main response. As outlined there, the failure probability for reasonable network widths is not a significant bottleneck for the batch sizes we experimented with. Instead, the bottleneck is the number of submatrices $L_A$ one needs to sample to recover all the correct directions $\overline{q}$, which is analyzed in Lemma 5.2. In practice, we limit this number to fit in a reasonable time limit and we report a failure otherwise. We emphasize that SPEAR, in theory, is capable of recovering inputs for batch sizes $> 25$ provided enough time. As outlined in the general response to Q1, one can use approximate reconstruction methods to guide the search for those submatrices, lifting the restriction, and allowing exact reconstructions for batch sizes as large as 100. Finally, we provide experiments in Appendix E5, showing that the predicted failure probabilities agree well with empirical results for small to moderate layer widths. We are happy to discuss any remaining or follow-up questions the reviewer might have.	Summary: This paper studied the gradient inversion problem in federated learning. In particular, an honest-but-curious server follows the federated training protocol but aims to infer private information from gradients collected from clients. The paper proposed a novel approach, SPEAR, that exploits the low-rank and sparsity properties of gradients corresponding to linear layers with ReLU activations, and recovers input data of batch size greater than 1 almost perfectly with high probability. The algorithm is a combination of matrix decomposition and a optimisation-based filtering procedure. The performance of the algorithm is explained by theoretical analysis and empirical evaluations, which is significantly better than previous methods that struggle to deal with large batch sizes. Strengths: - Recovering batched images from gradients has been a challenging task. The algorithm introduced in this paper is built on insightful observations on the gradients and solid theoretical justifications, which can inspire further work in this area. - The proposed algorithm works well empirically, under necessary requirements such as fully connected networks and moderate batch size. Sufficient ablations are provided to evaluate different aspects of the algorithm. - The exposition of the paper is nicely structured. Weaknesses: - The effectiveness of the algorithm relies on the assumption that the first layer of the network is a linear layer (while for linear layers in the middle or at the end of the network the algorithm can only recover the intermediate input to those layers, not the original images). However, this is not the case for many popular convolutional networks for image tasks. The application of the algorithm is restricted and it can be easily defended in practice. - The running time of the algorithm is exponential in the batch size, and the success probability may drop significantly for large batch size. Moreover, it is not clear if the algorithm is effective for high-resolution datasets like ImageNet. Technical Quality: 3 Clarity: 3 Questions for Authors: - When the batch size increases, is there a trade-off scheme that sacrifices the fidelity of recovered images but reduces the running time? - Can the algorithm deal with normalization layers? - Can the algorithm be adapted to convolutional networks and (vision) transformers? Confidence: 4 Soundness: 3 Presentation: 3 Contribution: 2 Limitations: The paper discussed limitations. Flag For Ethics Review: ['No ethics review needed.'] Rating: 5 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We thank the reviewer for their positive review. We are happy to read that they find our contribution insightful and inspiring future work and credit our significantly better performance and theoretical analysis. Further, we are glad the reviewer deems the paper nicely structured. Q1: Can the exponential sample complexity of SPEAR be improved? Is the probability of failure a bottleneck in practice? Yes, the exponential sample complexity can be addressed, as we discuss in the general response (Q1). Specifically, there we show that SPEAR can leverage optimization-based methods to circumvent the exponential sampling-based information retrieval to scale to much larger batch sizes while still providing (almost) perfect reconstruction, and we clarify that the failure probability is not currently a bottleneck (see also Figure 3 in the main paper). Q2: Is the algorithm effective for high-resolution datasets like ImageNet? As highlighted in the abstract, SPEAR reconstructs batches of sizes up to 25 on ImageNet exactly. We provide more corresponding results in Section 6.2 (Main Results). Specifically, we run experiments on the commonly used resolution of 224x224 for ImageNet and also on a 720x720 high-resolution setting to showcase our scalability even further. We note that this is, as shown in Table 3, at the very moderate cost of a runtime increase from 2.1 minutes to 2.6 minutes per batch. This is significantly faster than prior work as shown in Table 1, where we compare various works on ImageNet using a resolution of 256x256. Could the reviewer clarify the question in case we misunderstood it? Q3: Does the effectiveness of the algorithm rely on the assumption that the first layer of the network is a linear layer? No. As discussed in Q3 of the main response, SPEAR can be efficiently applied to any linear layer that precedes or follows ReLU activations to obtain that layer’s inputs. Even when the attacked layer is in the middle of the network the intermediate inputs recovered via SPEAR pose an increased risk to the privacy of the client data. We demonstrate the practicality of using SPEAR on linear layers in the middle of the network by attacking VGG16, as shown in Table 1 in the rebuttal PDF. See Q3 of the main response for further details. Further, we emphasize that our method is a significant theoretical advancement over prior work, providing novel insights into the gradient leakage problem as acknowledged by the reviewers. We focus on ReLU with linear layers to open the field for further research. There is already very promising future research building on the insights gained in this work. See our answer to Q6. Finally, for tabular datasets, networks starting with a linear layer are common. In this setting SPEAR outperforms all prior work by a wide margin in terms of input reconstruction quality and speed, particularly when no further priors are given. Q4: When the batch size increases, is there a trade-off scheme that sacrifices the fidelity of recovered images but reduces the running time? Yes, the threshold we use to test if an entry in $L \overline q_i$ is 0 can be tuned. As we use this test to filter out wrong proposal directions (see Section 4.1), high thresholds can permit directions computed in a numerically unstable way (e.g. due to applying Theorem 3.4 on matrices $L_A$ that are not so well-conditioned), leading to lower reconstruction quality but faster runtime. While tuning this threshold can yield a noticeable speedup, it does not address the exponential nature of our algorithm However, if one would like to resolve the exponential runtime of SPEAR, we propose a combination with an optimization-based method to get priors on the signs of $Z$. Please see our reply to question Q1 in the main response, where we discuss in detail how this allows SPEAR to scale to significantly larger batch sizes while retaining exact reconstruction. Q5: Can the algorithm deal with normalization layers? As Batch-Norm layers compute batch statistics at training time and then use these statistics to normalize the activations, they entangle the gradients of all batch elements. This makes the analysis of the gradient structure significantly more complex and prevents us from directly applying SPEAR as now the output of the BN rather than the linear layer has sparse gradients. On the other hand, this normalization and rescaling allows us to compute a very precise estimate of the expected sparsity for the ReLU inputs $Z$ and thus improve our filtering described in Section 4. While we have promising initial results addressing this issue, SPEAR requires significant non-trivial adaptations for BN-layers, which we believe is an interesting item for future work. Q6: Can the algorithm be adapted to convolutional networks and (vision) transformers? Yes, there is recently published work on Transformers relying on insights from this work, that also significantly outperforms all prior work. We contacted the chair to share this work with you. For applications to convolutional networks, please see our reply to Q3 in the main response.	Summary: This paper introduces a novel approach to input reconstruction for neural networks, focusing specifically on linear layers followed by ReLU activations. The key aspects of this method are: Low-rank decomposition: The authors leverage low-rank matrix decomposition techniques to simplify the representation of the linear layer's weight matrix. ReLU sparsity exploitation: The method takes advantage of the sparsity induced by the ReLU activation function following the linear layer. This sparsity provides additional constraints that aid in the reconstruction process. Efficient disaggregation algorithm: A core contribution of the paper is an efficient algorithm for searching the columns of the disaggregation matrix Q. This matrix plays a crucial role in the reconstruction process. Input reconstruction: By combining these elements, the authors demonstrate a method to reconstruct input examples from the output of a linear layer followed by a ReLU activation. Strengths: Its a really interesting idea to reconstruct from gradients and provide a mathematical explanation of previous work [1] (from the paper’s citation) and this work. Proof is clean and precise. Also discuss many settings, like layer width and batch size. It also provides a theory bound for failure probability in reconstruction attacks. The reconstruction result and reuse quality for reconstruction images are really good. [1] Kariyappa, Sanjay, et al. "Cocktail party attack: Breaking aggregation-based privacy in federated learning using independent component analysis." International Conference on Machine Learning. PMLR, 2023. Weaknesses: It performs better than [1] in the reconstruction result, but the batch size would be a bottleneck since the method is bound by batch size to resolve, and [1] can reach out to batch size > 64 easily. Interesting idea but miss some experiments and discussion: 1. It only provides experiments run on 1st linear layer setting. However, in many NN designs, linear layers are usually after the conv layer or before softmax. Could you provide a more detailed discussion about using your method for linear layer after conv layers in more practical setting? like [9]’s section 4 2. In the supplement code section, it only provides experiments based on relu1 and fc1, Could you run quick experiments and report performance on different/later fc layers? If the result is limited to the first fc layer, then it has some interesting innovation in method design but is not really practical. Also, if authors can provide more no cherry-picking visualization results on different fc layers or fc after conv for different datasets in the appendix,? For E6 DPSGD part, usually sigma = 1*10-4 is considered to be really large eps and too little noise, which is not considered practical privacy protection. For DPSGD on linear layer NN, sigma = 0.01 with some reasonable setting can be considered as eps = 80–100 ( I might calculate wrong but its around that number), which is still not useful for DPSGD. [1] Kariyappa, Sanjay, et al. "Cocktail party attack: Breaking aggregation-based privacy in federated learning using independent component analysis." International Conference on Machine Learning. PMLR, 2023. Technical Quality: 3 Clarity: 3 Questions for Authors: See above. Confidence: 3 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: See above. Flag For Ethics Review: ['No ethics review needed.'] Rating: 7 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We thank the reviewer for their positive review and are glad to read they found our contribution really interesting, the proofs clean and precise and the reconstruction quality really good. Q1: Can SPEAR scale to batch sizes greater than 64? Yes, we address this in the general response (Q1). Specifically, we show that SPEAR can leverage optimization-based methods to scale to significantly larger batch sizes while still providing (almost) perfect reconstruction. Q2: Can SPEAR handle a linear layer that is preceded by a conv layer? Yes. We added an experiment showing perfect reconstruction results for VGG16 and $b=20$ in Table 1 in the rebuttal PDF. Please refer to Q3 in the general response for more information about this experiment. Q3: Can you provide experiments, where SPEAR reconstructs inputs to linear layers that are not the first one? Yes. We consider a 6-layer network with 400 neurons in each layer and a batch size of 20 and report the mean absolute error (MAE) to the original activations in Table 2 in the rebuttal PDF. Our experiments indicate that we can recover almost all inputs to all layers (almost) perfectly. Please refer to Q3 in the main rebuttal for more details about this experiment. Q4: Are the provided visualizations representative? Can the authors provide more visualizations of the SPEAR’s results? We want to clarify that the visualizations provided in Figure 1 of the main paper are a subsample of the visualizations provided in Figure 9 in Appendix G, where we visualize the corresponding full batch. The batch in Figure 9 is representative as it was randomly chosen and has a PSNR similar to the overall PSNR mean. To reaffirm this point, we provide further batch visualizations in Figures 1-3 in the rebuttal PDF. These visualizations are in the same setting as Figure 1 of the main paper, but their corresponding batches are chosen to represent the $10^{\text{th}}$,$50^{\text{th}}$, and $90^{\text{th}}$ percentiles of the obtained PSNRs in this experiment. We observe that only 1 sample is not perfectly reconstructed for the $10^{\text{th}}$ percentile batch (which we show in its corresponding figure) and that the $50^{\text{th}}$ and $90^{\text{th}}$ percentile batches contain only perfect reconstructions. Further, we note that Figure 10 in Appendix G visualizes the batch on which SPEAR performed the worst on TinyImageNet. Manual inspection showed that its poor performance is caused by SPEAR failing to recover one of the directions $\overline{q}_i$. We will clarify these points and provide further visualizations in the next version of the paper. Q5: What is SPEAR’s performance under rigorous DPSGD guarantees? We want to clarify that the results in Appendix E.6 were obtained by adding Gaussian noise to the gradient, as a viability test of SPEAR’s robustness against noisy gradients. Specifically, we did not do the gradient clipping necessary for DPSGD. We will clarify this in the paper and also provide new experiments, showing SPEAR’s results in the DPSGD setting alongside the corresponding $(\epsilon,\delta)$ guarantee. --- Rebuttal 2: Comment: I did run the code provided by the author again, and the reconstruction results are fantastic. I am willing to raise my score since the authors addressed all the questions I asked. I believe this method is ground-breaking for gradient inversion attacks in general. Other recent works tend to use diffusion/GAN to enhance their reconstruction results, but this work perfectly reconstructs the linear layer's result and demonstrates the possible vulnerability of LL in many NN architectures. I just had a few detailed questions regarding implementation: 1. For VGG-16, are you running at linear shape = 1 1 4096 layer? 2. for implementation part, could you put Q_opt sperate only for extra evaluation function? I was confused and first thought you directly used GT in the reconstruction process. --- Rebuttal Comment 2.1: Comment: We deeply appreciate the reviewer’s effort in reviewing our paper and code. We are thankful for their insightful questions, the answers of which we will incorporate into the next revision of our paper. Below we respond to their additional questions above: Q1: For VGG-16, what is the shape of the attacked linear layer? We use the standard ImageNet VGG-16 from torchvision, which uses AdaptiveAvgPool2d(7, 7), resulting in 25088 input features for the attacked linear layer. The attacked linear layer has 4096 output neurons. We will clarify this. Q2: Computing $Q_\text{opt}$ in the main file of the attack is confusing We agree with the reviewer that computing $Q_\text{opt}$, as one of the first things we do in our main can give the wrong impression we use the ground truth $Q$ in our reconstruction. We emphasize this is not the case as the reviewer correctly noticed, and we apologize for the confusion this has caused. We compute $Q_\text{opt}$ so that our sampling function getQBarUniqueCol, based on Theorem 3.3, in sparse\_gradient\_reconstruction.py can display how many directions were recovered correctly in real-time. This is incredibly valuable for debugging and monitoring the progress of our method. $Q_\text{opt}$ is also used to report the success of the different filtering stages of SPEAR described in Section 4 of the paper. This is done in the function summarizeQ in sparse\_gradient\_reconstruction.py, and it is useful for both debugging and for creating Figure 6 in Appendix E.3. We reiterate that $Q_\text{opt}$ is not used for any other purpose in our code. Before making our code public, we will do our best to restructure the code such that $Q_\text{opt}$ is only used at the end of the reconstruction for evaluation to avoid confusion.	Summary: The paper presents SPEAR, a novel algorithm for exact gradient inversion of batches in federated learning. Unlike previous methods, SPEAR achieves exact reconstruction for larger batches by leveraging low-rank structure and ReLU-induced sparsity in gradients. The authors provide a theoretical foundation and an efficient GPU implementation, demonstrating the effectiveness of their approach on high-dimensional datasets and large networks. Strengths: 1. The paper introduces a significant advancement in gradient inversion attacks by enabling exact reconstruction for batch sizes greater than one. 2. The authors provide a strong theoretical basis for their method, including proofs of the low-rank nature of gradients and the sparsity induced by ReLU activations. Weaknesses: 1. The theoretical analysis only considers the ReLu and fully connected layer. 2. The method only seems to be effective for a small batch size, even though it has already been a big progress. Technical Quality: 4 Clarity: 4 Questions for Authors: SPEAR is effective for a moderate batch size of b < 25. What are the possible reasons to prevent the batch size to become larger? Confidence: 3 Soundness: 4 Presentation: 4 Contribution: 4 Limitations: See weakness. Flag For Ethics Review: ['No ethics review needed.'] Rating: 7 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We thank the reviewer for the very positive review and are excited that the reviewer finds that SPEAR marks a significant advancement for gradient inversion. We are also happy to read that the reviewer attests that our method builds on a strong theoretical basis and credits our soundness, presentation and contribution. We answer all of the reviewer's questions in the general response: Q1: SPEAR is effective for a moderate batch size of b < 25. What is preventing SPEAR from scaling beyond this? Please see our detailed reply in the main response (Q1). Q2: Can SPEAR handle activations beyond ReLU? Please see our detailed reply in the main response (Q2). Q3: Can SPEAR handle layers other than fully connected layers? Please see our detailed reply in the main response (Q3). We are happy to discuss any remaining questions or follow-up questions the reviewer might have.	Rebuttal 1: Rebuttal: We thank the reviewers for their positive and helpful feedback. We are encouraged they consider our work to mark a significant advancement (Brdw), provide valuable insight (LPF2, Xnvh), our theory to be well founded (Brdw, 38cL), our presentation to be good (Brdw, LPF2, Xnvh) and our reconstruction results to look really good (38cL, Xnvh). Based on the reviewers' suggestions, we conducted a range of additional experiments, reporting results in the attached PDF. Below, we address the topics we found to be shared amongst reviewers or ones we believe are of major importance. Q1: What prevents SPEAR from scaling to larger batch sizes, and how can this limitation be overcome? (Reviewers Brdw, 38cL, LPF2, Xnvh) First, we would like to highlight that in the common regime where the batch size is small in comparison to the dimensions of the linear layer, we prove that w.h.p. the input information is losslessly represented in the gradient (see Figure 3). SPEAR is not failing for large batch sizes because the relevant information can not be retrieved but solely because we set a timeout on sampling as exponentially many samples would be required. A theoretical analysis of this is provided in Lemma 5.2. This exponential sampling complexity can be addressed by combining SPEAR with an approximate reconstruction method to get a prior on which submatrices $L_A$ satisfy the conditions of Theorem 3.3, i.e., have corresponding matrices $A$ containing a 0-column. Using approximate, optimization-based input reconstruction techniques we can estimate the pre-activation values $Z$ and thus positions of $0$s in $\tfrac{\partial \mathcal{L}}{\partial Z}$. We can now use the estimated locations of the 0 entries in $\tfrac{\partial \mathcal{L}}{\partial Z}$ to directly sample $L_A$ matrices for which Theorem 3.3’s prerequisites are likely satisfied, drastically speeding up SPEAR’s information retrieval. We confirm the effectiveness of this approach in a preliminary study, shown in Table 3 in the rebuttal PDF, allowing SPEAR to scale to batch sizes beyond 100. Specifically, we use the modern version of Geiping et al. [1] to get priors on $Z$ and considered the most negative entries of the approximated $Z$ to be the ones with the highest likelihood of actually being negative. We consider a network with $m=2000$ and 6 layers, and batches of 100 TinyImageNet images. There, Geiping has a PSNR of 32.8 while SPEAR reconstructs 6 out of 10 random batches exactly (PSNR of 120) and reconstructs 98 or 99 directions correctly on the remaining four batches, leading to an overall PSNR of 81.5. These results confirm that SPEAR can be effectively scaled to larger batches, with further optimizations being left to future work. Q2: The work is restricted to ReLU activations. Can you handle other activations, including sigmoid, leakyReLU, tanh, etc? (Reviewers Brdw, Xnvh) In this work, we focus on ReLU activations as they are the most relevant activation function in practice. However, more broadly, the core of our method is to utilize low rankness and priors on the activation function in order to not need priors on the client input data. While the low-rankness can directly be transferred to other activations, we believe extending the priors to other activations makes for an interesting future work item. We emphasize that SPEAR is the first exact method to reconstruct network inputs from gradients originating from batches with a batch size greater than 1. Q3: Is SPEAR applicable to architectures that do not have a fully connected layer immediately following their input? (Reviewers Brdw, 38cL, LPF2) Yes, it is. While the evaluation primarily focuses on recovering inputs to the first FC layer of a fully connected network with ReLU activations, our method allows the recovery of inputs of any FC layer preceded or followed by a ReLU activation regardless of the architecture used, as emphasized in Section 3.2. This is demonstrated in Table 2 in the rebuttal PDF, where we successfully recover the inputs to all layers followed by a ReLU in a $L=6$ layer FC network with width $m=400$. While these intermediate layer inputs may not violate the client privacy as severely as directly obtaining the user data exactly, the authors of [9] demonstrate that recovering individual batch inputs to intermediate layers can further boost the performance of optimization-based gradient leakage attacks and poses additional privacy issues for the clients. Crucially, [9] makes this observation in the context of convolutional neural networks (CNNs). We also experimented with recovering the inputs to the FC layers in CNNs. Specifically, we attacked the first FC layer of a VGG16 network for ImageNet batches of size $b=20$. We present the results in Table 1 in the rebuttal PDF. Our results demonstrate that SPEAR recovers 100% of the individual linear layer inputs up to numerical precision. Thus, we are confident that SPEAR can improve optimization-based gradient leakage attacks for CNNs. We will add this to the paper. Apart from CNNs, FC layers are the predominant layers for other domains, such as tabular data, where we significantly outperform the current state-of-the-art (Table 2, main paper). Further, for transformers, there is very promising follow-up work leveraging SPEAR’s insights to achieve state-of-the-art reconstruction of text by a wide margin. We have contacted the chair to share an anonymized copy of the follow-up work with you. All of this confirms the generality and applicability of SPEAR, as well as its significant impact on the field of gradient leakage. Conclusion We hope to have been able to address the reviewers’ questions and look forward to the discussion period. Pdf: /pdf/ade8ee6fc3258e440b62be42b6298a256c3ab172.pdf	NeurIPS_2024_submissions_huggingface	2,024	null	null	null	null	null	null	null	null
Contextual Linear Optimization with Bandit Feedback	Accept (poster)	Summary: This paper studies Contextual Linear Optimization (CLO) with bandit feedback and presents a class of algorithms termed Induced Empirical Risk Minimization (IERM). The authors also derive the regret upper bounds. The regret analysis accounts for the misspecification of the policy class and incorporates a margin condition that potentially enables a faster regret rate. Strengths: 1. The extension of the CLO problem to bandit feedback is an important and novel research direction. 2. The mathematical formulations and proofs are rigorous and well-presented. 3. The analysis allows for model misspecification in the induced policy class. Weaknesses: Although I believe this paper is technically sound, the readability could be improved. The presentation of the problem and model could be clearer, especially for readers unfamiliar with the CLO problem. It would be better to provide a more detailed explanation of the problem setting before delving into mathematical formulas. For example, in the introduction, the authors could explain the interaction protocol between the learner and the environment, and specify which quantities are known or unknown to the learner/environment. Such elaborations would make the paper more accessible. Technical Quality: 3 Clarity: 3 Questions for Authors: N/A Confidence: 3 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: Limitations are adequately discussed in the paper. Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thank you for your comments. We agree that providing more background information on the CLO problem would be helpful. Here, we clarify the unknown quantities and the interaction between the learner and the environment. In classic CLO problems, the learner aims to solve the optimization problem $\min_{z\in\mathcal{Z}}\mathbb{E}[Y^\top z \mid X = x] = f_0(x)^\top z$ for all observed context values $x$, where $f_0(x)$ is the conditional expectation function $\mathbb{E}[Y \mid X = x]$. This means the learner seeks to minimize the expected total decision cost $Y^\top z$ given the observed context value $x$. The true distribution of $(Y, X)$ is unknown, making the unknown $f_0$ function the main challenge for the learner. Before making decisions, the learner needs to estimate the $f_0$ function from samples $(X_i, Y_i)_{i = 1}^n$, typically assumed to be i.i.d draws from the population distribution of $(X, Y)$. Note that this is an offline setting: the learner is provided with a batch of existing data $(X_i, Y_i)$ for $i=1, ..., n$ and uses this dataset for learning, without further interaction with the environment to generate more data while learning and making decisions. The existing literature has considered two major types of approaches to learning $f_0$ and the corresponding decision policy: estimate-then-optimize (ETO) approaches and integrated approaches. The latter directly targets the decision objective and is thus favored by many recent papers. Our paper studies a CLO problem where the cost coefficient $Y$ is not directly observed in the data. Instead, we observe only the total costs of some historical decisions implemented in the past. Specifically, we have access to some historical decisions $Z_1, \dots, Z_n$, made by some previous policies that might depend on the corresponding contexts $X_1, \dots, X_n$ and potentially other unobserved factors. The total costs $C_i = Y_i^\top Z_i$ for these decisions were recorded, but the underlying cost coefficient $Y_1, \dots, Y_n$ were not. This means the learner only observes the costs of the decisions that were actually implemented, without information on the costs of other unimplemented counterfactual decisions. These constitute the observed samples $(X_i, Z_i, C_i)_{i = 1}^n$. Again, we consider an offline learning setting where the learner is given this dataset collected under previous policies, without further interacting with the environment to generate new data. The learner does not know the population distribution of $(X, Z, Y, C)$ and aims to use the given dataset $(X_i, Z_i, C_i)$ for $i=1, ..., n$ to learn a new decision policy for future decision-making. This is akin to the offline contextual bandits widely studied in the literature [1, 2, 3]. However, the offline contextual bandit literature mostly considers the optimization over finitely many arms, while we consider a more general constrained linear programming problem. We in particular study how to extend the integrated IERM approach from regular CLO problems to CLO with partial/bandit feedback. In summary, the main unknown quantity to the learner in the CLO problems is the conditional expectation function $f_0(x) = \mathbb{E}[Y \mid X = x]$ of the cost coefficient $Y$ given contextual features $X$. Once this function is known, the CLO problems can be solved easily. Classic CLO literature considers learning this unknown function from samples of the cost coefficients and the contexts $(X_i, Y_i)$. Our paper considers the setting where only total costs $C_i$ associated with historical decisions $Z_i$ (along with the contexts $X_i$) can be observed, but the underlying cost coefficients $Y_i$'s are unobserved. Both these CLO literature and our paper consider an offline setting where the learner is given a fixed set of data to decide a policy, without further interacting with the environment to collect more data. It would be interesting to consider online settings where the learner can directly interact with the environment, generating data on the fly while learning and making decisions. But the online learning is out of the scope of our current paper. Thanks for your suggestions. We will incorporate these clarifications into our Section 1. In particular, we will highlight that our paper considers an offline learning setting in both the abstract and introduction. [1] Dudík, Miroslav, John Langford, and Lihong Li. "Doubly robust policy evaluation and learning." Proceedings of the 28th International Conference on International Conference on Machine Learning. 2011. [2] Dudík, Miroslav, et al. "Doubly Robust Policy Evaluation and Optimization." Statistical Science 29.4 (2014): 485-511. [3] Swaminathan, Adith, and Thorsten Joachims. "Batch learning from logged bandit feedback through counterfactual risk minimization." The Journal of Machine Learning Research 16.1 (2015): 1731-1755. --- Rebuttal Comment 1.1: Comment: Thanks for your responses and my concerns have been addressed. --- Reply to Comment 1.1.1: Comment: Dear Reviewer h4hQ. Thank you for reading our rebuttal. As you write that it has addressed all your concerns, we would greatly appreciate if you would raise your rating and confidence scores accordingly. Thank you. And do let us know if you have further questions -- we would do our best to answer them promptly.	Summary: This work addresses the partial bandit feedback setup for contextual linear optimization. The authors propose an algorithm based on Induced Empirical Risk Minimization (IERM), which incorporates doubly-robust estimation to handle the reward model misspecification. They provide a regret bound for the partial bandit feedback setup and, as a byproduct, offer a regret bound for the full bandit feedback setup under model misspecification. This contribution is significant due to the robustness of the theoretical results and the broad applicability of the findings. The paper is well-written and merits acceptance with only minor revisions. Strengths: This work accommodates the practicality such as partial feedback and model misspecification. In addition, the theoretical results are solid enough. The numerical experiments demonstrate the theoretical results well. Weaknesses: This setup is similar to partial monitoring. It would be great if the suggested model is compared with partial monitoring. In addition, it is hard to interpret the result to the reviewer, as the complexities of terms are not stated. The reviewer would like to see a summary of what the order of regret with respect to n is. Further, the model misspecification is not fully explained in the first two sections. The authors should explain the model misspecification before starting a theoretical analysis. Lastly, there are some minor errors in the presentation: - In line 108, "is" should be corrected. - There is no definition of $\chi^2_{n,\delta}$. - In line 74, "in that that" should be corrected. Technical Quality: 4 Clarity: 3 Questions for Authors: Based on my understanding, ETO is a family of algorithms that use estimates directly, such as plug-in policies with estimates. In contrast, IERM-type algorithms include an additional optimization step based on these estimates. This step appears to make IERM-type algorithms function similarly to optimism in the face of uncertainty (OFU)-type algorithms. What aspects of IERM-type algorithms make them more effective than ETO-type algorithms? Is this improvement related to their exploration schemes? Confidence: 4 Soundness: 4 Presentation: 3 Contribution: 4 Limitations: It is not well explained why the suggested algorithm has improved performance as compared to the existing algorithms. Flag For Ethics Review: ['No ethics review needed.'] Rating: 8 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thanks for your comments. We now address each point individually. Comparison to partial monitoring. It is interesting to compare our problem with partial monitoring. There are at least two major differences. One is that partial monitoring usually considers an online setting where the learner interacts with the environment and collects data on the fly while making decisions. In contrast, we consider an offline setting where the learner is provided with a fixed dataset without further interaction. See our response to Reviewer h4hQ for more details. The other is that partial monitoring is designed to allow the observed feedback to be different from the reward associated with each decision (a.k.a, arm) [1]. However, our setting is closer to the typical bandit setting, since our observed feedback $C = Y^\top Z$ is exactly the decision cost (i.e., negative reward) of each decision $Z$. Complexity terms in theoretical guarantees. Thank you for raising this issue. Our regret bounds are formulated in terms of the critical radius, a widely used complexity measure in statistical learning theory. This measure is very general and recovers the information-theoretically optimal rate for many function classes [2]. However, we agree that this may be abstract and could impede the readers' understanding. As we explain below Theorem 1, a concrete example is when the function classes $\mathcal{G}_1, \mathcal{G}_2$ are VC subgraph classes with VC dimension $\eta$. The corresponding critical radius is $\tilde{r} = O(\sqrt{\eta/n})$. Then the regret bound in Theorem 1 reduces to \begin{align} \text{Reg}(\hat{\pi}) \lesssim \left(\frac{\eta}{n}\right)^{\frac{\alpha + 1}{\alpha + 2}} + \text{Reg}(\tilde{\pi}^) \left(\frac{\eta}{n}\right)^{\frac{1}{2}} + \frac{\eta}{n} + \text{Rate}^N(n/2, \delta/4) + \text{Reg}(\tilde{\pi}^). \end{align} The regret rate depends on the degree of misspecification characterized by $\text{Reg}(\tilde{\pi}^)$ (i.e, the regret of the best-in-policy-class policy), the nuisance estimation error rate $\text{Rate}^N(n/2, \delta/4)$, and the margin parameter $\alpha$. Hu et al. [3] study the full-feedback setting without misspecification, so there is no nuisance estimation error or misspecification error (i.e., $\text{Rate}^N(n/2, \delta/4) = 0$ and $\text{Reg}(\tilde{\pi}^) = 0$). Our bound recovers their fast rate $O((\eta/n)^{(\alpha+1)/(\alpha+2)} + \eta/n)$ that interpolates between $O(n^{-1/2})$ and $O(n^{-1})$ according to the margin parameter $\alpha$. Our new bound shows that this fast rate could be dominated by the slow rate $O(\text{Reg}(\tilde{\pi}^) ({\eta}/{n})^{{1}/{2}})$ when there is significant misspecification. Moreover, our bound shows that in the bandit-feedback setting, we also need to estimate nuisance functions, and the corresponding error $\text{Rate}^N(n/2, \delta/4)$ can also affect the final regret bound. The order of this nuisance error rate depends on the form of the score function (see Proposition 2) and the complexity of estimating the nuisances. For example, if we use the doubly robust score and both nuisances can be estimated at rates $ o(n^{-1/4})$, then $\text{Rate}^N(n/2, \delta/4) = o(n^{-1/2})$ is negligible [4]. We plan to explain these clearly in a dedicated remark in Section 3. Model misspecification.* Thanks for your suggestion of explaining the model misspecification before our theoretical analysis. We totally agree with this suggestion, because accommodating model misspecification is one of our major technical contributions. We plan to explain model misspecification in Section 1.1 when we first introduce the induced policy class. What makes IERM-type algorithms more effective? IERM can be more effective than ETO because it more directly targets the decision-making problem. ETO first constructs an estimator $\hat f$ for the unknown function $f_0(x) = \mathbb{E}[Y \mid X = x]$ by fitting standard regressions (e.g., linear regressions), and then solve optimization $\min_{z\in\mathcal{Z}}\hat f(x)^\top z$ for each context value $x$ to get the final policy. However, the estimation step typically optimizes some regression error (e.g., least squares objective), without considering the downstream decision-making problem. In contrast, IERM considers the policies $\pi_f \in \text{argmin}_{z\in\mathcal{Z}}f(x)^\top z$ induced by candidate regression functions $f$ (e.g., linear functions) and then selects the policy achieving the smallest in-sample average cost. IERM directly optimizes the decision cost objective, thereby achieving better decision performance in many scenarios. This is quite different from the optimism in face of uncertainty type algorithms for balancing the exploration and exploitation in online decision-making. Minor errors. Thanks for pointing out the typos, which we will correct in the future. Regarding the definition of $\chi^2_{n, \delta}$, it is a vanishing sequence used to characterize the error in estimating the two nuisance functions. We will change the first two lines of Proposition 2 to the following: ''For any given $\delta \in (0, 1)$, let $\chi_{n, \delta}$ be a positive sequence converging to $0$ as $n \to \infty$, such that the mean square errors of the nuisance estimates satisfy the following with probability at least $1-\delta$''. [1] Kirschner, Johannes, Tor Lattimore, and Andreas Krause. "Linear partial monitoring for sequential decision making: Algorithms, regret bounds and applications." The Journal of Machine Learning Research (2023). [2] Wainwright, Martin J. High-dimensional statistics: A non-asymptotic viewpoint. Cambridge university press, 2019. [3] Hu, Yichun, Nathan Kallus, and Xiaojie Mao. "Fast rates for contextual linear optimization." Management Science (2022). [4] Victor Chernozhukov, Mert Demirer, Greg Lewis, and Vasilis Syrgkanis. Semi-parametric efficient policy learning with continuous actions. NeurIPS 2019. --- Rebuttal Comment 1.1: Comment: I thank the authors for their detailed response.	Summary: The paper aims at the following problem. The learner observes side information $x$ and needs to output $z$. The Nature will then generate $Y$ (which depends on $x$, but may be randomised) and the learner will suffer loss $Y'z$. The key feature of the setup is that $Y$ is not directly observable. The paper describes convincing real-life applications fitting this scenario. One is after a policy $z=\pi(x)$ minimising the loss expectation on the basis of the past data. One can obtain it by estimating the conditional expectations $f_0(x) = E(Y\mid x)$ first. However, this is not necessarily the best approach. In the style of reinforcement learning, the paper aims to estimate the policy value $V(\pi)$ directly. The paper puts forward an algorithm, or, rather an approach for this called IERM (I understand some steps in it involve optimisation, which is very hard to perform). The performance of the approach is studied in Theorem 1. It bounds the regret of the policy found by IERM w.r.t. the optimal policy in a class. The bound involves Rademacher complexity of the class. There is substantial empirical evaluation. Strengths: I believe the approach of the paper is useful and very practical. Weaknesses: The paper is very hard to read for an uninitiated person like me. I found it very hard to distinguish between given data, estimates, datapoints, random variables etc. Perhaps it would help if the paper were structured in a more traditional way with definitions and algorithms rather than the free-flowing narrative. Page 3, line 125: I am not sure I parse which we call as nuisance functions as they will be used to correctly. Did you mean which we call nuisance functions because they will be used to ? (I should say I do not understand the choice of the word "nuisance". The pun is lost on me.) Technical Quality: 3 Clarity: 2 Questions for Authors: None. Confidence: 1 Soundness: 3 Presentation: 2 Contribution: 3 Limitations: Yes Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thank you for your feedback. We now further clarify our problem and explain the data, estimates, random variables, and nuisances involved. First, let's consider a linear programming problem $\min_{z\in \mathcal{Z}} y^\top z$ where $z$ is the decision variable, $\mathcal{Z}$ is the constraint set, and $y$ is the coefficient vector in the objective. This can model many problems, such as the shortest-path problem in our paper. Specifically, consider finding the shortest path from a starting point to a destination along edges on a graph (with $d$ edges in total). Any feasible path can be represented by a vector $z \in \\{0, 1\\}^d$, with each entry corresponding to one edge and a value of 1 indicating that the edge is picked by the path. A vector $y$ gives the traveling time along each edge respectively. Thus, the total traveling time for a path $z$ is $y^\top z$. The set $\mathcal{Z}$ constrains $z$ to represent a path (or, a convex combination of paths) such that $\text{argmin}_{z\in\mathcal{Z}}y^\top z$ gives a path that achieves the shortest total traveling time. Our paper considers an uncertain environment where the objective coefficient is a random vector $Y$. For example, the traveling time could be stochastic and uncertain in practice. While the realized value of $Y$ is unknown at the time of decision-making, we may have access to some predictive contextual features $X$ (also considered a random vector). For example, although we do not know the actual traveling time in each shortest-path decision-making instance, we may observe features such as weather, time of day, and traffic conditions that are predictive of the traveling time. Contextual linear optimization (CLO) provides a formal way to incorporate features in decision-making, solving for the optimal decision that minimizes the expected cost upon observing each context value $x$: \begin{align} \min_{z\in\mathcal{Z}}~ \mathbb{E}\left[Y \mid X = x\right]^\top z. \end{align} Since the distribution of $(X, Y)$ is unknown in practice, existing CLO literature uses samples $(X_i, Y_i)_{i=1}^n$ to estimate the unknown function $f_0(x) = \mathbb{E}[Y \mid X = x]$. There are primarily two different approaches: the estimate-then-optimize approach and the integrated approach. The estimate-then-optimize approach directly fits a standard regression of $Y$ against $X$ to get an estimator $\hat f$, usually by minimizing some statistical error measures like mean squared errors. The integrated approach considers the decision policies $\pi_f$ induced by each candidate estimator $f$ (e.g., linear functions): \begin{align} \pi_f(x) \in \text{argmin}_{z\in\mathcal{Z}}~ f(x)^\top z. \end{align} Then it picks an optimal policy achieving the smallest sample average cost $\frac{1}{n}\sum_{i=1}^n Y_i^\top \pi_f(X_i)$. The integrated approach directly targets the decision cost objective and is favored in recent literature. Our paper also focuses on the integrated approach. The existing literature only considers the full-feedback setting where $Y$ is fully observed so its observations can be directly used to evaluate any decision policy. In contrast, we consider a partial or bandit feedback setting where we can only observe the total cost $C = Y^\top Z$ for some historical decision $Z$ prescribed by a previous policy. Here the historical decision $Z$ is also considered as a random vector because it may have been prescribed according to $X$ or some other uncertain factors. Therefore, we have access to samples $(X_i, Z_i, C_i)_{i=1}^n$ but without observing the underlying cost coefficients $Y_i$'s. In this case, we cannot directly evaluate the sample average cost of any induced policy $\pi_f$. Our Proposition 1 shows that under Assumption 1, we can evaluate the mean cost of a policy $\pi_f$ by using a score function $\theta$ in place of the unobserved $Y$. This score function may depend on not only the observed variables but also two unknown functions $f_0$ and $\Sigma_0$. We thus propose to first construct some estimators $\hat f$ and $\hat \Sigma$ for the unknown functions and then plug these estimators into the score function for the downstream evaluation and learning of decision policies. This process involves constructing estimators $\hat f$ and $\hat \Sigma$ in the score function and the estimator for the policy value $\frac{1}{n}\sum_{i=1}^n \theta(X_i, Z_i, C_i; \hat f, \hat \Sigma)^\top \pi_f(X_i)$. Our Section 2 considers a slightly different procedure that incorporates additional sample splitting, but the overall idea is the same. We refer to the unknown functions $f_0$ and $\Sigma_0$ in the score function $\theta$ as "nuisance" functions, because these functions are not directly used for decision-making but merely serve as intermediaries for the evaluation of the policy value. This is standard terminology in statistics and is also frequently used in off-policy evaluation and learning. For example, the off-policy evaluation of a contextual bandit policy or reinforcement learning policy often involves estimating a behavior policy function (propensity score) or a conditional reward Q function. These are often referred to as nuisance functions as well [1]. Thank you for your comments again. We will further clarify the data, estimators and random variables and explain the meaning of ''nuisance'' in the camera-ready version. [1] Uehara, Masatoshi, Chengchun Shi, and Nathan Kallus. "A review of off-policy evaluation in reinforcement learning." arXiv preprint arXiv:2212.06355 (2022). --- Rebuttal Comment 1.1: Comment: Thank you for the detailed response. It helps my understanding. Special thanks for explaining "nuisance". This needs to be incorporated in the paper as the meaning seems to be remote from that in everyday English.	null	null	Rebuttal 1: Rebuttal: Dear Reviewers, Thank you for your thoughtful comments and constructive suggestions. We are encouraged to hear that you find our theoretical results solid, our numerical experiments demonstrate the theoretical results well, and our approach useful and practical. We appreciate your suggestions for improving the readability and presentation of our paper. We have addressed each of your points individually, and we will incorporate these clarifications in the camera-ready version if our paper is accepted. We are committed to enhancing the presentation and writing of our paper and are grateful for your valuable feedback. Please let us know if you have any further comments. We are happy to discuss them. Best regards, The Authors.	NeurIPS_2024_submissions_huggingface	2,024	null	null	null	null	null	null	null	null
Optimizing over Multiple Distributions under Generalized Quasar-Convexity Condition	Accept (poster)	Summary: The paper considers a minimization problem, where the optimizing variable is composed of $d$ probability distributions. The notion of quasar convexity is generalized by allowing for different quasar-convexity parameters $\gamma_i, i=1,...,d$ for the $d$ distributions. Instead of using the gradient function for the oracle term, more generic functions $F_i, i=1,...,d$ ("internal functions") are allowed. Convergence guarantees are established for the situations where the interval functions are either Lipschitz continuous of "polynomial-like". Infinite horizon reinforcement learning is shown two be a special case of this setting. Similar results are derived for the minimax setting. Strengths: * thorough theoretical analysis * method requires strictly less iterations than mirror descent * algorithm also works for unknown $\gamma_i$ * reinforcement learning application is an important one * extensive overview of related work Weaknesses: * Algorithm 1 and Algorithm 2 are never actually described in the text, they are only provided as pseudocode. It think it would be nice to add 2-3 sentences to point out how exactly it differs from basic mirror descent. Example given, I don't understand why one randomly picks t following probability $\mathbb{P}[t] =1/T$. * There is "only" one example (the reinforcement application) for the new structure the paper introduces. I am not sure if one example is enough to justify speaking of a new structure. (But I also understand that such examples are complex and it might go beyond the scope of the paper to come up with several such examples) * the paper does not mention the computational costs of the algorithm. Is it $d$ times as much as basic mirror descent? Is this still feasible in praxis? Typos: * Headline Algorithm 1: Mirior -> Mirror * line 206: depends -> depend Technical Quality: 3 Clarity: 3 Questions for Authors: * I am confused about the following part in the introduction (line 62-65): "We then study designing efficient algorithms to solve (1). One simple case is when $\gamma_i, i=1...m$ is pre-known by the algorithms. The possible direction is to impose a $\gamma_i$-dependent update rule, such as by non-uniform sampling. However, in general cases, $\gamma_i, i=1...m$ is not known and determining $\gamma_i, i=1...m$ require non-negligible costs." Can you elaborate on this? Confidence: 2 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: I don't see that the authors explicitly mention limitations of their work, but it is also a theoretical paper, so everyone can just read the assumptions at the beginning of the theorems to see where the results apply and where not. The assumptions made for the theoretical analysis are stated clearly. Flag For Ethics Review: ['No ethics review needed.'] Rating: 7 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thank you for your comments and suggestions. We will respond to each of your comments individually in the following. ## Algorithms Descriptions: Thanks for your suggestion. We will use equations and sentences to introduce the algorithm. One randomly picks $t$ following probability means we output $x^t$ by uniformly sample from $(x^t)_{t=1}^T$. ## Examples: For minimization problems, we provide two examples of functions that simultaneously satisfy both the GQC condition and Assumption 3.3: a toy example (Example B.8 in Appendix B) and a simple neuron network over a simplex (Example 3.4 in Section 3.2). For minimax problems, In addition to infinite horizon two-player zero-sum Markov games, we have also demonstrated that general smooth convex-concave problems satisfy the GQCC condition and Assumption 4.2 (see Appendix C.3.2). Further examples await our future exploration. ## Computational Costs: Since the MD method requires computing a $\sum_{i=1}^d n_i$-dimensional gradient vector at each iteration, and Algorithm 1 also requires computing a $\sum_{i=1}^d n_i$-dimensional internal function at each iteration, so Algorithm 1 does not need more additional computational costs. ## Response to confuse about line 62-65: Sorry for the confusion. We mean that If $\{\gamma_i\}$ is known, then the update of our algorithm on each variable block will depend on $\gamma_iF_i$ rather than solely on $F_i$. However, in many applications, $\{\gamma_i\}$ is often related to the information of the optimal point (see Appendix B.3 for reinforcement learning problem, $\gamma_i $ depends on the optimal solutions). Therefore, the computational costs of estimating $\{\gamma_i\}$ may even exceed that to solve the optimization problems. This is our motivation to design adaptive algorithms without pre-known $\{\gamma_i\}$. --- Rebuttal Comment 1.1: Comment: Thank you for the clarifications. I keep my positive score.	Summary: The authors study a typical optimization model where the optimization variable is composed of multiple probability distributions. For this optimization problem, they propose a new structural condition/landscape description named generalized quasar-convexity (GQC) beyond the realms of convexity. Strengths: It looks like the authors propose a new theoretical analysis for the quasar-convex optimization. Weaknesses: It would be better if the authors added more motivation for why they focus on quasar optimization. Another point: the theoretical analysis demonstrated is not very clear. If they can show it clearer, it will be better, Technical Quality: 2 Clarity: 2 Questions for Authors: Does the assumption 4.2 can be relaxed? Can the Section 4.3 be written more? Confidence: 4 Soundness: 2 Presentation: 2 Contribution: 2 Limitations: No Flag For Ethics Review: ['No ethics review needed.'] Rating: 4 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thank you for your comments and suggestions. We will respond to each of your comments individually in the following. ## Motivation: In many non-convex optimization problems, the objective function often admits ''convexity-like'' properties, such as matrix completion, phase retrieval, and neural network under some settings (Bartlett et al., 2019; Ge et al., 2016). Therefore, exploring the structured sets of functions for which the problem can be efficiently solved is meaningful. This aligns with the ideas proposed by Hinder et al., (2020). Quasar convexity is a widely known condition in the field of optimization with two typical examples in machine learning: 1) the loss functions of neural networks satisfy the quasar convexity condition in large neighborhoods of the minimizers (Kleinberg et al., 2018; Zhou et al., 2019); 2) Under mild assumptions, the objective function for learning linear dynamical systems also admits to the quasar convexity condition (Hardt et al., 2018). Furthermore, generalized quasar convexity (GQC), as an extension of quasar convexity, encompasses not only quasar convexity but also the function structures of reinforcement learning and certain neural networks. This implies that the objective functions of more machine learning problems may satisfy the GQC condition. Our analysis provides a theoretical guarantee for the solvability of approximate global optima of these problems. As a supplement to the GQC condition, generalized quasar-convexity-concavity (GQCC) extends the concept of convexity-concavity. Based on this landscape characterization and appropriate assumptions, we propose an algorithm for solving minimax problems that have a better convergence rate in specific machine learning problems. ## Theoretical analysis is not clear: Thanks for your suggestions. We will add more intuitive discussions for our analysis. For Thm 3.5, the key idea is to reduce the impact of each variable block on the function error, which plays a crucial role in establishing global convergence. Concretely, we separate the upper bound of $\frac{1}{T}\sum_{t=1}^T(f(x^t)-f(x^{*}))$ into: a) an invariant lower order term and b) the weighted sum of variance of finite difference term $\frac{\mathcal{O}(1)}{T} \sum_{t=1}^TVar_{x_i^t}(F_i(x^t)-F_i(x^{t-1}))-\frac{\mathcal{O}(1)}{T} \sum_{t=1}^TVar_{x_i^t}(F_i(x^{t-1}))$. Furthermore, applying the property of sequence $(F_i(x^t))_{t=1}^T$ under ''polynomial-like'' structure condition, we bound (b) by a quantity that grows poly-logarithmically in $T$. For a more detailed proof sketch, please refer to the beginning of Appendix B.1. We will add more explanations for both Thm 3.5 and 4.4 in the revised version. ## Relax Assumption 4.2: Thanks for suggestion. Currently, to achieve the faster $\tilde{\mathcal{O}}(1/T)$ convergence rate, we do not find a clear way to have a good relaxation. The main challenges of multi-variable block optimization problems lie in its non-convex (non-convex-non-concave) structure and the coupling of multiple variable blocks. Therefore, a simple Lipschitz continuity assumption on internal function may not be helpful in obtaining a faster rate. However, for a slower rate, such as $\tilde{\mathcal{O}}(1/\sqrt{T})$, it is possible and we leave it as a future work. We will discuss it in the revised version. ## Written more Section 4.3: Thanks for suggestion. We will introduce more background about Markov games, including its problem formulation, relation to convex and concave optimization, and current results. --- Rebuttal 2: Comment: Dear reviewer, Thank you for your efforts in the review process. We have tried our best to address your concerns and answer the insightful questions. Please let us know if you have any other questions regarding our paper or response. If we have successfully addressed your questions, we would greatly appreciate if you can reevaluate the score. Best, Authors	Summary: This paper introduces a novel optimization model for addressing problems involving multiple probability distributions, a common scenario in fields such as policy optimization and reinforcement learning. The authors present a new structural condition called Generalized Quasar-Convexity (GQC), which extends the original quasar-convexity concept by allowing individual quasar-convex parameters for each variable block. This flexibility accommodates varying degrees of block convexity. The paper proposes an Optimistic Mirror Descent (OMD) algorithm tailored to this framework and demonstrates its efficiency in achieving an ε-suboptimal global solution. Furthermore, the paper extends the GQC concept to minimax optimization problems by introducing the Generalized Quasar-Convexity-Concavity (GQCC) condition. The theoretical findings are supported by applications in discounted Markov Decision Processes (MDPs) and Markov games, showing improved iteration complexity bounds over existing methods. Strengths: - The introduction of GQC and GQCC conditions provides a fresh perspective on optimization problems involving multiple distributions, extending beyond traditional convexity assumptions. - The paper offers rigorous theoretical analysis, including complexity bounds for the proposed OMD algorithm, and demonstrates that these bounds are tighter than those for existing methods. - The application of the proposed framework to reinforcement learning and Markov games showcases its practical relevance and potential impact on real-world problems. Weaknesses: - The assumptions made for the GQC and GQCC conditions, such as polynomial-like structures and specific Lipschitz continuity requirements, may limit the generality of the results. It would be beneficial to discuss the applicability of these assumptions in broader contexts. - While the theoretical contributions are robust, the paper lacks empirical validation through extensive experiments. Including experimental results on synthetic and real-world datasets would strengthen the paper by demonstrating the practical performance of the proposed algorithms. Technical Quality: 3 Clarity: 3 Questions for Authors: Yes Confidence: 2 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: - How do the assumptions made for the GQC and GQCC conditions compare to those in other optimization frameworks? Are there potential relaxations or alternative conditions that could be explored to broaden the applicability of the proposed methods? - Can you provide more detailed insights or examples of real-world problems where the proposed GQC and GQCC conditions are particularly beneficial? How do these conditions improve the solution quality or computational efficiency in such scenarios? Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thanks for your comments and suggestions. We will respond to each of your comments individually in the following. ## Discuss the applicability of these assumptions in broader contexts: Thanks for your suggestion. We will add more discussions on the applicability of the assumption and our framework. For minimization problems under the GQC condition, we begin with a convergence result relying on the general Lipschitz continuous condition of internal function $F$. However, under this condition, we need to additionally assume that $\max_{i\in[1:d]}\gamma_i<+\infty$. Based on the aforementioned assumption merely, according to our knowledge, it appears impossible to eliminate the coupling effects between different variable blocks without a consistent upper-bounded assumption of $\{\gamma_i\}$. In certain applications, $\max_{i\in[1:d]}\gamma_i$ can indeed approach $+\infty$. For instance, in the infinite horizon reinforcement learning problems, the discounted state visitation distribution $d_{\rho_0}^{\pi}(s)$ for a particular state $s$ may approach zero, implying that $1/\gamma_s=0$. Therefore, to address this uncertainty of unknown $\{\gamma_i\}$, we introduce the Assumption 3.3. Assumption 3.3 is not hard to achieve. In Proposition B.2 of Appendix B, we demonstrate that if the absolute value of the higher-order derivatives of a vector-valued function $F$ grows exponentially with its order, then $F$ admits Assumption 3.3. Furthermore, since the feasible domain of the functions considered in this paper is a bounded closed region, the conditions of Proposition B.2 are easily satisfied by smooth $F$, making Assumption 3.3 hold easily. Additionally, we provide two examples of functions that simultaneously satisfy both the GQC condition and Assumption 3.3: a toy example (Example B.8 in Appendix B) and a simple neuron network over a simplex (Example 3.4 in Section 3.2). For minimax problems, most algorithms that guarantee global convergence are based on a convex-concave structure. Our assumptions not only encompass the assumption of Lipschitz continuous gradients for the objective function $f$ in the convex-concave case (refer to Section C.3.2 in Appendix C) but also provide a unified landscape description for infinite horizon two-player zero-sum Markov games. It is worth noting that Algorithm 2 in this paper achieves a better convergence rate in infinite horizon two-player zero-sum Markov games compared to the results in Wei et al., (2021). ## Experiments: Please refer to response Author Rebuttal:$\textbf{Experiments}$. ## Potential relaxations or alternative conditions: Thanks for your concerns about the generality of our conditions and the request to relax the assumptions. The GQC condition include quasar convexity (QC) condition. Actually, our algorithm can achieve $\tilde{\mathcal{O}}(1/T)$ convergence rate under QC condition with Lipschitz continuous internal function $\nabla f$. Currently, to achieve the faster $\tilde{\mathcal{O}}(1/T)$ convergence rate, we do not find a clear way to have a good relaxation. The main challenges of multi-variable block optimization problems lie in its non-convex (non-convex-non-concave) structure and the coupling of multiple variable blocks. Therefore, a simple Lipschitz continuity assumption on internal function may not be helpful in obtaining a faster rate. However, for a slower rate, such as $\tilde{\mathcal{O}}(1/\sqrt{T})$, it is possible and we leave it as a future work. ## Detailed insights or examples of real-world problems where the proposed GQC and GQCC conditions are particularly beneficial? We prove that for reinforcement learning (RL) problem, GQC can be achieved. For the infinite horizon reinforcement learning problems, the discounted state visitation distribution $d_{\rho_0}^{\pi}(s)$ satisfies $\sum_{s\in S}d_{\rho_0}^{\pi}(s)=1$, implying that $\sum_{s\in S}1/\gamma_s=1$, therefore our algorithmic complexity does not rely on $\|S\|$. However, if we describe the properties of the function using the worst $1/\gamma_s$ (similar to the QC condition), then the complexity might be $\|S\|\max_{s\in S}1/\gamma_s$. Similarly, we prove that infinite horizon two-player zero-sum Markov games satisfies GQCC condition. For infinite horizon two-player zero-sum Markov games, $\sum_{s\in S}\psi_s(z)\leq 2$ for any $z\in \mathcal{Z}$. Therefore, the convergence rate of Algorithm 2 does not depend on $\|S\|$.	Summary: The paper studies the optimization of generalized quasar convex (GQC) functions, a new global structure introduced in this paper. This definition relaxes the quasar convexity (QC) condition, in a way that different components of the optimization variable satisfy the QC like condition with different parameters. The key idea is that if we simply treat a GQC function as a QC function, one needs to take d times the maximum value of these paramters which can be large. Applying GQC allows one to deal with the sum instead which could be significantly smaller than d. Leveraging this insight, the authors adapt the well-known optimistic mirror descent algorithm. The main use case appears to be in policy optimization in MDPs, where the authors show that the GQC condition is shown to be satisfied by the expected value function. The authors also define a similar condition for minimax optimization problems, and apply their framework to find the Nash equilibrium of a two-player zero sum Markov game. Overall my evaluation of the paper is quite positive. Even as a non-expert, I was able to appreciate the ideas and the contributions of the authors. The paper is also well-written and easy to follow. I have given a positive score to reflect my current assessment. However, as this is not my area, I am unable to evaluate the technical merit and novelty of proot techniques in the paper, and I did not dive into the proofs in the appendix. So I will also wait to hear from more expert reviewers. Strengths: See above Weaknesses: See above Technical Quality: 3 Clarity: 3 Questions for Authors: See above Confidence: 2 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: See above Flag For Ethics Review: ['No ethics review needed.'] Rating: 7 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We thank you for dedicating your time and effort to reviewing our manuscript. We appreciate that you acknowledge our paper. We will also carefully revise the paper to enhance its readability.	Rebuttal 1: Rebuttal: ## Experiments: We add a simple simulation to validate our proposed algorithm. We do experiments on learning one single neuron network over a simplex, which has been discussed in Example 3.4 in Section 3.2. The objective function is written as \begin{align} f(p,P)=\frac{1}{2}E_{x,y}(\sum_{i=1}^m p_i\sigma(x^\top P_{i})-y)^2 \end{align} where $\sigma(\cdot)=\exp(\cdot), p\in \Delta_m$ and $P = (P_1,\cdots,P_m)\in\prod_{i=1}^m\Delta_d $ and the target $y$ given $x\in[-C,C]^{d}$ admits $y = \sigma(x^\top P_1^)$ for some $P_1^\in \Delta_d$. We set $m=20$, $d=20$, $C=4$ and the optimal solution $\mathbf{P}_1^*\in\Delta_d$ is a random probability density function. We compare the convergence performance of OMD\_IF (Optimistic Mirror Descent Method with internal functions, Algorithm 1), OMD\_G (Optimistic Mirror Descent Method with gradients, (Rakhlin et al., 2012)), MD (Mirror Descent Method with gradients), and AMD (Accelerated Mirror Descent Method with gradients, (Lan, 2020)). The experimental result is shown in the PDF file. From the figure, we can see that Algorithm 1 consistently shows superior convergence performance under various step size choices. Pdf: /pdf/cd9e05d45be95a153adc93b730c566dc466539db.pdf	NeurIPS_2024_submissions_huggingface	2,024	null	null	null	null	null	null	null	null
PaCE: Parsimonious Concept Engineering for Large Language Models	Accept (poster)	Summary: This paper introduces Parsimonious Concept Engineering (PaCE), a novel framework for aligning LLMs by modifying their activation space. This framework constructs a large-scale concept dictionary in the activation space and partitions concepts into benign or undesirable categories. During inference, PaCE decomposes activations into these concept directions, removing undesirable components to reorient the LLMs' behavior towards alignment goals while preserving linguistic capabilities. The authors demonstrate the effectiveness of PaCE in tasks such as response detoxification, faithfulness enhancement, and sentiment revision. Strengths: - This paper proposes a novel activation manipulation framework, different from the existing methods of vector addition and orthogonal projection. - The authors comprehensively and clearly articulate the deficiencies of existing methods and the ways in which the new framework addresses these deficiencies. - The authors validate the superiority of the proposed method in pursuing alignment goals and preserving linguistic capabilities compared to existing methods across multiple tasks. Weaknesses: Overall, I did not find significant flaws that outweigh the contributions of this paper. The main weakness lies in the efficiency of the framework. During inference, using this method results in each token's prediction taking approximately 2-3 times longer than when not using it. In contrast, simply using vector addition only increases the time by about 30-40%. Additionally, although the concept vocabulary only needs to be obtained once, the concept direction dictionary needs to be constructed and stored for different LLMs. Technical Quality: 4 Clarity: 4 Questions for Authors: * Could the authors provide more discussion on the computation time required and storage space (including memory usage and disk space) needed for constructing the direction dictionary? * Why does OrthoProj require significantly more computation time in Table 2 compared to other methods? This seems to conflict with the statement in Remark 3 that OrthoProj is a special case of the method proposed in this paper. Confidence: 4 Soundness: 4 Presentation: 4 Contribution: 3 Limitations: Yes Flag For Ethics Review: ['No ethics review needed.'] Rating: 7 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Dear Reviewer rWAy, Thank you for your insightful reviews and kind support for acceptance. It is our pleasure to reply to your comments. Time efficiency Thank you for discussing PaCE's time efficiency. We would like to emphasize that the main contributions of the paper are (1) a new dataset of concept dictionary and (2) a sparse coding process using this dictionary to decompose latent LLM representations. VecAdd and OrthoProj do not sufficiently model the representation space, and PaCE adopts a dictionary to have enhanced alignment performance. * PaCE is faster than OrthoProj ([18, 79], Table 2) and we provide an option of a time-efficient solver (OMP, see discussion below). * PaCE has much higher averaged safety (Table 1), faithfulness and sentiment revising scores (Table 4) than VecAdd, OrthoProj, Prompting, and the vanilla model. Hence, the goal of this paper is to offer a competitively novel solution to representation manipulation for alignments of LLMs. We gladly take your comment and explore alternative optimization techniques for Equation 1 in the paper from the sparse coding literature that significantly reduces inference times. Below we briefly describe various solvers. * The elastic-net solver (Line 213) optimizes Equation 1 to obtain coefficients simultaneously. * We also consider Orthogonal Matching Pursuit (OMP), a fast greedy solver [D1, D2]. OMP iteratively adds concepts to the support based on maximum coherence with the residual and updates the residual until the support size k is reached. In other words, k is the number of non-zero coefficients. Comparing solvers on response detoxification (\S 4.1) shows certain setups of OMP are faster than Elastic Net (e.g., OMP with $k=50$ is an order of magnitude faster with 12.3% lower safety score). This demonstrates that a greedy solver improves speed at a performance cost. We will include the experiment in $\S4.1$ of the paper based on your suggestions. \|\| OMP (k=50) \| OMP (k=100) \| OMP (k=150) \| OMP (k=20) \| Elastic Net \| \|---\|---\|---\|---\|---\|---\| \| Time per decomposition (s) \| 0.045 \| 0.182 \| 0.381 \| 0.749 \| 0.411 \| \| Safety (↑) \| 63.1 \| 64.4 \| 66.9 \| 70.8 \| 72.0 \| Computing concept dictionary for different LLMs Thank you for bringing up the point. We would like to explain our motivation for collecting the dataset, and how the cost may not be a concern in terms of our efforts in open-sourcing and low amortized cost. * (Motivation) We noted that the scale of existing datasets on the concept stimulus (e.g., RepE [79], ActAdd [64] in the paper) was limited. This incentivized us to (1) collect a large stimulus dataset to better model the concepts and (2) pre-compute the concept dictionary in an LLM representation space. * (Open-sourcing) Future researchers need not rerun dataset collection as we will open-source our dataset of human-readable concepts, stimuli, and concept dictionaries for LLaMA2 7B/13B and the Mistral 7B. Pre-computed concept partitions for downstream alignment tasks will also be shared." * (Low amortized cost) Each target model's concept dictionary needs to be extracted only once, with costs amortized across tasks. For example, extracting our dictionary for LlaMA2-7B took 25 minutes on an NVIDIA A40 machine and the dictionary was used for all evaluations on that model. Specific time and space requirements for constructing the dictionary We provide the statistics of computation and storage space needed for constructing the dictionary. We will include them as a new subsection (Appendix C.4) in our paper. * GPU memory usage for hosting the dictionary during each decomposition: * LLaMA2 7B: 0.31 GB * LLaMA2 13B: 0.45 GB * Local disk storage usage for storing dictionaries (for all involved 19 layers): * LLaMA2 7B: 5.80 GB * LLaMA2 13B: 8.26 GB * Time taken to compute and extract the dictionaries (NVIDIA A40 GPU machine): * LLaMA2 7B: 25.2 minutes * LLaMA2 13B: 40.8 minutes Computation time of OrthoProj * We appreciate your excellent comments on this point. The difference lies in the fact that the OrthoProj method computes a projection at multiple layers of the LLM to maintain a fair estimation of the coefficient. While PaCE uses a concept dictionary to model the representation space, OrthoProj does not sufficiently model it and could remove the contribution of desirable concepts. To maintain a stable steering process, the coefficient needs to be re-estimated at every decoder layer. Instead, VecAdd pre-sets a coefficient as a hyper-parameter and reuses it across the rest of the layers. Similarly, in experiments ($\S 4$ of the paper), PaCE performs decomposition at the first encountered intermediate layer and reuses it across the rest of the layers. In practice, one can freely tune the VecAdd coefficient through grid search and choose the layers where PaCE decomposition happens. * We find the choice of PaCE is sufficient for steering the LLM. We hypothesize this is because sparse coding can find a more suitable steering direction that directly reflects the desired concept change as compared to simply running orthogonal projection. Lastly, thank you for carefully reading the remarks. Your time and effort are greatly appreciated. Also, our paper’s Appendix B.4 includes the implementation details of frameworks and hyper-parameters for inferring on open-source LLMs. Based on your suggestion, we will include the justification in a new subsection Appendix B.5 and update the existing description to better clarify the difference between OrthoProj and other methods. Hopefully this alleviates your concerns; if not, please engage with us during the discussion phase. Thanks and regards, Authors of Submission #8804 [D1] _Orthogonal matching pursuit: Recursive function approximation with applications to wavelet decomposition._ Pati et al., 1993. [D2] _Orthogonal matching pursuit for sparse signal recovery with noise._ Cai et al., 2011. --- Rebuttal Comment 1.1: Comment: Thank you for your effort in providing numerous complementary results. These address my concerns about efficiency and could greatly improve the soundness of the paper. As a result, I will maintain my current score in favor of acceptance. --- Reply to Comment 1.1.1: Title: Thank you for your reply and support! Comment: Dear Reviewer, Thank you for your timely feedback and kind recognition of our evaluations. Your support of acceptance is very encouraging to us, and your insights have greatly improved our paper. We will take your suggestions on revising the paper as promised: the manuscript $\S 4$ and Appendix B will be revised to include our discussions on framework efficiency. Sincerely, Authors	Summary: This paper introduces an activation engineering framework for LLM alignment. PaCE addresses the challenge of reducing undesirable outputs while retaining benign concepts through a two-stage process: Concept Construction and Partition, and Activation Decomposition and Intervention. It constructs a large-scale concept dictionary and uses sparse coding to decompose LLM activations and selectively remove undesirable components. The authors demonstrate PaCE's effectiveness on tasks like response detoxification, faithfulness enhancement, and sentiment revising, showing improved alignment performance without compromising linguistic capabilities. Strengths: 1. This paper presents a refined activation engineering method for alignment, achieving outstanding results especially with the 13B model. 2. The method is provided with a clear visualization and detailed interpretation to illustrate its underlying principles. Weaknesses: 1. Quality of Concept Dictionary: the method's performance could be overly dependent on the quality and coverage of the concept dictionary. What measures are taken to ensure its comprehensiveness and that the dictionary is free from biases? 2. Static Activation: Static concept dictionary provides static activation while being decomposed, making PaCE unable to handle context-dependent concepts. 3. Without consideration of polysemy: Many common concepts have different meanings in different scenarios. For example, the concept ‘kill’ is harmful when referring to the act of terminating a life, while it is benign in the usage of ‘killing time’. Both two usages are common. However, this paper ignores the polysemy and treats all 30 pieces of contextual stimuli same in constructing concept dictionary construction. Thus, I am a bit concerned about the accuracy of the extracted concept direction. Besides, the design of retrieval promotes the diversity of the concept stimuli, while may undoubtedly increases the potential of the situation that the same word in different or even opposite meanings. 4. Misuse of representation reading algorithm (arXiv:2310.01405) in extracting concept directions: In Appendix B.2 Eq.6, authors take a difference between the activations of concept t_i and all other concepts t_j (j≠i). However, Zou et al. (in Sec. 3.1) state that stimuli in the pair should have different target concept or function. Take detoxification task as an example, the concept ’bias’ and ‘narrowed’ in Fig. 7 are both harmful content which needed to be removed. Since ‘bias’ and ‘narrowed’ have the same target function, representation reading algorithm will not pair them to take a difference, while this paper also pairs them. This is quite weird. 5. At the risk of compromising helpfulness: Following W3. and W4., the proposed concept directions may not be accurate and removing them may remove some benign concepts, thus limiting the model helpfulness. 6. Slower inference speed and time-efficiency: Since authors have claimed that ‘PaCE is computationally slower than VecAdd’, PaCE will undoubtedly slow down the original model’s inference speed, which is very important during model deployment. Technical Quality: 3 Clarity: 3 Questions for Authors: 1. Can PaCE handle concepts that have multiple semantic meanings or are context-dependent? 2. Are the VecAdd/OrthoProj baselines constructed based on the same concept dictionary with PaCE during the experiments? 3. How to choose the intervention decoder layers: Prior works (arXiv:2312.06681) find the optimal layer for steering by sweeping over all layers and performing intervention to evaluate the effect. Out of curiosity, how do authors choose the layers? 4. Can authors provide an evaluation on XSTest dataset (arXiv:2308.01263) when applying PaCE in detoxification task? I am a bit concerned about that PaCE will increase the false refusal rate on benign queries. 5. The MMLU results about VecAdd in Tab. 1 are questionable according to (arXiv:2402.19465). I doubt if the choice of steering layers for VecAdd is also the same as the setting of PaCE (last 19 layers) which results in the significant performance drop in MMLU since for VecAdd choosing the middle layers is better. Confidence: 5 Soundness: 3 Presentation: 3 Contribution: 2 Limitations: In terms of experimental backbone models, it is recommended that the authors include more models (e.g., Mistral) not merely Llama2 Family for a comprehensive validation on the effectiveness of PaCE. In terms of linguistic capability evaluation, only fluency, perplexity and one multi-choice QA benchmark is limited. Generation tasks like open domain QA (e.g., MT-Bench), summarization (e.g., XSum) are supposed to be evaluated. Flag For Ethics Review: ['No ethics review needed.'] Rating: 3 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thank you for your time on the paper. We are happy to address your comments below. Quality of the concept dictionary Excellent question! The efforts are made in two-fold: the concept words (indexes) and the context stimuli (contents) that define the semantics of each word. * For the concept indexes, we took the most frequent 40,000 words from the Brown Corpus ([14] in our paper), which models standard American English [C2] to ensure that our dictionary covers comprehensive concepts. * We defined the meaning of each concept via retrieving external knowledge from Wikipedia and instructing GPT4 to generate context stimuli. The semantics of a concept may develop over time: e.g., “mouse” was not regarded as a digital input device before the era of PC [C1]. We consider that GPT’s parametric knowledge is fixed at a timestamp. We propose to retrieve external knowledge for stimulus synthesis because the external knowledge is more updatable and explicit, which serves as enriched descriptions for concepts and helps GPT to generate more grounded and diverse scenarios. To further add to the transparency, we will open-source the dataset so that the community can reuse and improve on it. Words have polysemy We agree with you that concepts may have multiple meanings. We tried to be frank on this in our paper on Lines 954-956: “Current practice usually finalizes a single vector per concept by SVD, but theories on polysemanticity and recent studies on the causal models of language suggest that a concept might be better represented by a union of subspaces, each corresponding to different semantic meanings”. However, we argue that, while VecAdd and OrthoProj may be affected by polysemy, PaCE's overcomplete dictionary allows accurate analysis of the target representation through sparse decomposition. Tables 1 and 2 in the paper show PaCE outperforming OrthoProj and VecAdd on linguistic metrics. We further provide empirical evidence on MT-Bench (QA) and XSum (summarization) with explanation for polysemy not harming PaCE’s model helpfulness. Empirical evidence. Table 1 and Table 4 in $\S 4$ show that PaCE is better than OrthoProj and VecAdd on linguistic capability metrics (fluency, perplexity, MMLU). Based on your suggestions, we conduct further experiments on MT-Bench and XSum, whose implementation is detailed in the one-page PDF. PaCE maintains superior linguistic capability compared to VecAdd and OrthoProj. \|MT-Bench Results\|\|Vanilla\|PE\|VecAdd\|OrthoProj\| PaCE \| \|:---:\|:---:\|:---:\|:---:\|:---:\|:---:\|:---:\| \| LlaMA2-7B-Chat \| GPT-4o as Judge (%, ↑) \| 6.81 \| 6.80 \| 6.68 \| 6.42 \| 6.75 \| \| LlaMA2-13B-Chat \| GPT-4o as Judge (%, ↑) \| 7.02 \| 6.95 \| 6.83 \| 6.94 \| 6.96 \| \|XSum Results\|\|Vanilla\|PE\|VecAdd\|OrthoProj\|PaCE\| \|:---:\|:---:\|:---:\|:---:\|:---:\|:---:\|:---:\| \| LlaMA2-7B-Chat \| ROUGE-1 (%, ↑) \|14.1\| 14.6 \| 13.7 \| 14.4 \| 14.6 \| \|\| BERT F1 (%, ↑) \|82.2\| 82.6 \| 81.9 \| 82.5 \| 83.0 \| \|\| METEOR (%, ↑) \|23.0\| 22.9 \| 22.0 \| 23.1 \| 22.5 \| \|LlaMA2-13B-Chat\| ROUGE-1 (%, ↑) \| 15.9 \| 15.3 \| 15.7 \| 14.8 \| 15.8 \| \|\| BERT F1 (%, ↑) \|85.3 \| 84.8 \|84.7\|84.9\|85.2\| \|\| METEOR (%, ↑) \|24.1 \| 22.9 \| 24.0\|23.8\|24.0\| Explanations. We attribute this to the large-scale dictionary with sparse coding, explained as follows. * Since the dictionary is large, concepts with single and clear semantics are involved. E.g., if stimuli of “kill” may have different meanings, there exist other more polarized concept vectors such as “murder” (more harmful) and “spend” (more benign). * Sparse coding aims to choose the fewest concepts to reconstruct the latent representation (i.e., parsimony). For the sake of argument, assuming the sentence is about “killing time” and the vanilla LLM has the correct semantic understanding of its benignness, the latent representation of the whole sentence will be closer to concepts such as “spend” and “time” rather than string-matching to “kill” (which in your setup could have mixed harmful and benign senses). As the sparse coding of the target representation promotes the parsimonious selection of concepts with monosemantics, it helps to represent benign contexts correctly without assigning significant weights to ambiguous terms like "kill". Response for pairing in representation reading; time efficiency; context-dependent concepts Thank you for your insightful discussion. Due to OpenReview's limitations on figures, please refer to the one-page PDF under Global Response for our detailed responses. Compliance to benign requests in the XSTest dataset Thanks for inquiring on this very insightful dataset. PaCE decomposes input based on underlying semantics, not string-matching. It removes only malicious parts of the representation, preserving compliance with benign requests. We evaluate benign requests in XSTest, following the XSTest classification: I: full compliance, II: full refusal, III: partial refusal. Compliance is minimally affected. More detailed evaluations are in the one-page PDF. \|\|Vanilla\|\|\|PaCE\|\|\|VecAdd\|\| \|:---:\|:---:\|:---:\|:---:\|:---:\|:---:\|:---:\|:---:\|:---:\| \|I\|II\|III\|I\|II\|III\|I\|II\|III\| \|211\|30\|9\|206\|33\|11\|200\|39\|11\| Middle layers for representation manipulation Thank you for this comment that allows us to correct a typo: In Line 909 (Appendix B.4), we wrote “activation vectors are extracted from the last 19 layers of the target LLM’s decoder layer.” We meant to say “activation vectors are extracted from the last-29th to the last-11th layer (totaling 19 layers) of the target LLM’s decoder layers”. Thus, the VecAdd in our experiments did use the middle layers. PaCE on Mistral-7B Please refer to our response and evaluation in Generalization to Other LLMs (Reviewer Cdr5). Thanks and regards, Authors of Submission #8804 [C1] Biases in Large Language Models: Origins, Inventory, and Discussion. Navigli et al., 2023. [C2] Role of the Brown Corpus in the History of Corpus Linguistics. Kholkovskaia et al., 2017. --- Rebuttal Comment 1.1: Comment: Dear Reviewer, We are grateful for your efforts and insights in the review, and we hope our response together with experiments covered your concerns. As the discussion phase is about to close, we are looking forward to hearing from you about any further feedback. We are more than happy to clarify or address additional comments (if any). Thanks and regards, Authors	Summary: This work proposes an activation engineering framework for alignment, i.e., PaCE, consisting of a task-driven concept dictionary and linear decomposition algorithm. Thus, the toxic or harmful input can be decomposed into a new activation vector combination, that removes the harmful activation. This is a generalization and adaptive method compared to VecAdd and OrthoProj. However, the presentation is somewhat confusing. It is expected that the authors can clarify it and improve the paper's readability. Strengths: This paper presents a novel safety alignment method based on activation engineering. The evaluation of PaCE is comprehensive. The results are insightful. Weaknesses: There are some confusing parts (see detailed comments below). The presentation needs significant improvements. There is a large time and computation cost in the preparatory phase. The efficiency is questionable. It is necessary to analyze more serious jailbreaking examples, e.g., toxic，biased. Technical Quality: 3 Clarity: 2 Questions for Authors: It is not clear how the dictionaries and scores are provided to the LLMs along with the input? Why does this linear degradation performed at each layer of the LLM not cause a decrease in fluency and efficiency? Is it possible to classify activation alignment as a scheme for the inhibition of malicious neurons or prompt purification? Compared to knowledge editing (e.g., "Detoxifying Large Language Models via Knowledge Editing"), what advantage of PaCE? Is the decomposition of semantic concepts the key to PaCE not affecting the normal response? According to Remark 3 of 3.3, I understand that the limitation that exist in VecAdd and OrthoProj are directional deletion and some limitations claimed in Figure 2. So, is dictionary construction the key to helping PaCE address limitations such as ambiguous deletions? How is it that independence between concepts can be solved by sparsity? Compared to RLHF, SFT, and other secure alignment schemes, I understand that the dataset of RLHF is replaced with a knowledge-driven concept dictionary, and the dynamic scoring during training is replaced with a task-driven concept splitter of GPT-4. If the understanding is correct, can the authors further explain the advantages? As far as I know, the Chat model of the LLaMA series has strong security alignment capabilities, but the jailbreak examples given by the author are relatively weak. If possible, can the authors give a comparison of the detoxification of the more offensive or malicious jailbreak examples on AdvBench, for example? Confidence: 4 Soundness: 3 Presentation: 2 Contribution: 3 Limitations: The authors discussed limitations in Appendix B.6. Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: ### Response to Reviewer s54k (1/2) Dear Reviewer s54k, Thank you for the thoughtful feedback and constructive questions. We address the concerns and provide additional evaluations to validate our work. We add the suggested improvements to enhance the quality of our work. Preparatory Phase Concerns Thank you for recognizing our efforts in dataset collection. Early in developing PaCE, we noted that the scale of existing datasets on the concept stimulus (e.g., RepE [79], ActAdd [64]) was limited. This incentivized us to (1) collect a large stimulus dataset to better model the concepts (2) pre-compute the concept dictionary in a larger LLM's representation space. Below we try to address your comments on computational costs in this phase. Future researchers need not rerun dataset collection as we will open-source our large dataset of human-readable concepts and stimuli. Further, each target model needs a concept dictionary to be extracted only once and this cost is amortized across different tasks on this model. For example, extracting our dictionary for LlaMA2-7B model took 25 minutes on our eight-card GPU machine and was used for all evaluations on that model. Lastly, despite this one-time preparation cost, the decomposition and intervention stages of PaCE are flexible in time efficiency and task performance. In Table 2 of the paper, we show that PaCE is faster than OrthoProj with a higher performance. Providing dictionary and scores to LLM We appreciate your question which helps improve the clarity of the paper! In short, dictionaries are a collection of concept vectors and are frozen for representation decomposition. * First, the LLM takes an input prompt (e.g., malicious requests) * Compared to a regular LLM, the representation engineering framework [B1] extracts the activations at each decoder block of the transformer. Such extraction results in a vector corresponding to the input prompt, which can then be modified for steering in different ways. * For PaCE, this steering has two main stages: * Stage 1 pre-computes the large-scale concept dictionary offline and the partition (i.e., scores) of which concepts represent benign/harmful concepts. * In Stage 2, at inference time, our approach extracts the representation of an input prompt and uses sparse coding to decompose this as the linear combination of atoms in our frozen dictionary. We then modify this linear combination by removing undesirable components and proceeding with inference in the LLM with the detoxified representation. We hope this clears up any misunderstandings. We will include the justifications in Appendix B.3 of our paper and revise other parts of the manuscript correspondingly. Effects on fluency and efficiency, and the role of decomposition to preserve the normal response Thank you for acknowledging the advantages of PaCE. Below we explain in two parts: fluency (Part 1) and efficiency (part 2). Part 1: The key to keeping fluency is to maintain the benign components of the activation of user input prompts. PaCE handles it by decomposing the user input’s activation along benign and harmful concepts, while VecAdd and OrthoProj do not sufficiently model the concepts in the representation space. The decomposed solution of PaCE accurately estimates the benign components and preserves them during the intervention process. Part 2: The decomposition is efficient because the optimization problem for Equation (1) in the paper is convex with known fast solvers such as active-set (used in this paper) and OMP (see the discussion with Reviewer Cdr5/rWAy,). Connection between activation alignment and the inhibition of malicious neurons or prompt purification Thank you for bridging PaCE with other alignment schemes. If the concept vectors in PaCE are sparse/axis-aligned, i.e. each concept vector only has a few specific non-zero elements, then yes – subtracting such a vector from the activation is inhibiting the support neurons. We observe that concept vectors are dense and not axis-aligned, which means that many neurons are involved. Detoxification on AdvBench AdvBench uses GCG [B2] to adversarially optimize a jailbreak suffix for a harmful behavior request. In the table below, we show the LlaMA-7B-Chat safety score (%, ↑) on the effective set of suffix attacks for AdvBench harmful behavior set. We observe that PaCE outperforms other baselines. We also note that the outperformance of PaCE in our paper’s jailbreaks is more significant than that in suffix attacks. This could be because story-telling and roleplay jailbreaks (used in our paper) contain more complex and entangled concepts. Under this scenario, PaCE decomposes the target representation and well estimates the malicious component, while VecAdd and OrthoProj do not model the space sufficiently. In the AdvBench case, instead, the optimized adversarial suffix can be regarded as the text-space inversion of straightforward malicious concepts. PaCE and other defense mechanisms in latent space and prompt space shall effectively defend these suffixes more easily. We will include the evaluation in $\S 4.1$ of our paper. \| \| Vanilla \| PE \| VecAdd \| OrthoProj \| PaCE (Ours) \| \|---\|---\|---\|---\|---\|---\| \| LlaMA2-7B-Chat \| 11.72 \| 91.90 \| 94.51 \| 92.81 \| 96.65 \| \| LlaMA2-13B-Chat \| 18.04 \| 93.86 \| 95.33 \| 96.72 \| 99.17 \| (continued in the comment below) --- Rebuttal 2: Comment: (rebuttal continued here) ### Response to Reviewer s54k (2/2) Dictionary construction to address limitations such as ambiguous deletions (Part 1); Sparsity to handle independence between concepts (Part 2). Part 1: Yes, PaCE has a large-scale dictionary modelling sufficient concept directions in the latent space, which allows PaCE to accurately analyze the compositionality of benign concept directions in the target. Part 2: When the concept vectors are not independent, infinitely many linear combinations of concept vectors can reconstruct the activation equally well. Therefore, one needs regularization to break ties. * Sparsity serves as one regularization, which models the belief that the activation can be written as a linear combination of _only a few_ concept vectors. This is favorable since shorter explanations are deemed more interpretable [B3, B4]. * A standard regularizer is the sum-of-squares of the combination coefficients, as in ridge regression. However, this tends to give combinations that use most of the concept vectors in the dictionary, * and since the dictionary is large this is not efficient. Advantages of PaCE over RLHF, SFT, and Knowledge Editing (KE) Your analogy of the correspondence among different alignment paradigms is inspiring. We would like to point that the main advantage of PaCE over the mentioned RLHF, SFT, and KE are two-fold. * Training-free: RLHF, SFT, KE all need to tune the parameters of LLM, which potentially degrade the well-structured priors of the pre-trained LLM. Taking a step back, even if LoRA is adopted for these paradigms, the training/tuning incur significant computation and memory costs. PaCE does not modify the parameters of LLM and requires no training. It better preserves priors of LLM, provides a low-resource alignment solution, and retains the general linguistic capabilities. * Interpretable and Adaptive: The solved coefficients are an accurate interpretation of how a user input’s representation is composed in the concept space. Also, when a new alignment goal is set, RLHF, SFT, and KE need to collect sufficient task samples and tune the LLM on the new dataset. In contrast, PaCE just needs to run the concept partitioner through PaCE-1M, which is expected to be much faster and more convenient. We are glad to respond to further comments or other concerns (if any) during the discussion period. We appreciate your insightful suggestions which have helped to validate and strengthen our framework. Thanks and regards, Authors of Submission #8804 [B1] _Representation Engineering: A Top-Down Approach to AI Transparency._ Zou et al., 2023. [B2] _Universal and Transferable Adversarial Attacks on Aligned Language Models._ Zou et al., 2023. [B3] _Interpretable by design: Learning predictors by composing interpretable queries_. Chattopadhyay, et al., 2022. [B4] _Leveraging sparse linear layers for debuggable deep networks._ Wong et al., 2021. --- Rebuttal 3: Title: Response Comment: Thanks for the authors' response. Most of my concerns have been addressed. It would be great if the authors incorporate those insights into the revised version. I will increase the score from 5 to 6. --- Rebuttal Comment 3.1: Title: Thank you for your recognition and raising the score! Comment: Dear Reviewer, We greatly appreciate your thoughtful feedback, your support of our work, and your decision to raise the score. We are especially grateful for your comments on the preparatory phase, valuable suggestions for an additional benchmark, and comprehensive insights on multiple alignment paradigms. As you suggested, we will revise manuscript $\S 4.1$ for AdvBench and adjust Appendix for (1) details of the preparatory phase, (2) dictionary input, and (3) connections among different paradigms. We will also revise other points as promised. Thank you again for your valuable input! Sincerely, Authors	Summary: This paper presents a framework for aligning LLMs by using sparse coding techniques on a comprehensive concept dictionary. PaCE effectively controls and modifies neural activations of LLMs to achieve alignment goals such as response detoxification, faithfulness enhancement, and sentiment revising. The proposed method addresses limitations of existing alignment methods, such as costly fine-tuning, inadequate removal of undesirable concepts, and unnecessary removal of benign concepts. The paper demonstrates state-of-the-art performance in alignment tasks while maintaining the linguistic capabilities of LLMs. Strengths: Originality: The paper introduces a unique approach to LLM alignment by combining sparse coding with a large-scale concept dictionary. This method provides a novel perspective on activation engineering, differentiating itself from traditional parameter fine-tuning and prompt engineering approaches. Quality: The methodology is well-developed, with thorough explanations and clear diagrams illustrating the concept construction, partitioning, and activation intervention processes. The use of sparse coding for accurate activation decomposition is well-motivated and justified. Clarity: The paper is well-written, with a clear structure and logical flow. Each section builds on the previous one, providing a comprehensive understanding of the proposed framework. The experimental setup and results are presented in a detailed and easy-to-follow manner. Significance: The proposed PaCE framework addresses critical issues in LLM alignment, offering a scalable and efficient solution. The state-of-the-art performance demonstrated in various alignment tasks highlights the practical significance and potential impact of this work on the broader AI community. Weaknesses: Computational Efficiency: While the paper demonstrates that PaCE outperforms existing methods in alignment tasks, the computational efficiency could be further optimized. The comparison with OrthoProj shows a significant improvement, but PaCE is still slower than VecAdd. Future work could focus on enhancing the speed of the proposed method. Generalization to Other Models: The paper primarily evaluates PaCE on llama 2 7B, 13B models. It would be beneficial to demonstrate the generalizability of the framework across a wider range of LLMs to strengthen the claims of broad applicability. Interpretability of Concept Dictionary: While the paper provides examples of concept vectors and their partitioning, a deeper analysis of the interpretability and semantic consistency of the concept dictionary would be valuable. Understanding the intrinsic properties of the concept vectors could offer further insights into the alignment process. Reproducibility: The paper mentions the availability of the PaCE-1M dataset and plans to release the source code. Ensuring that all experimental details, including hyperparameters and specific configurations, are thoroughly documented will be crucial for reproducibility and wider adoption of the proposed framework. Technical Quality: 3 Clarity: 3 Questions for Authors: Computational Efficiency: Are there any ongoing efforts or planned approaches to optimize the computational efficiency of PaCE, particularly in comparison to VecAdd? Model Generalization: Have you tested PaCE on other LLMs beyond llama 2 7B, 13B models? If so, could you provide insights or results from those experiments? Concept Dictionary Analysis: Can you provide more detailed analyses or case studies on the interpretability and semantic consistency of the concept dictionary used in PaCE? Ablation Studies: The ablation studies provide valuable insights into the contributions of various components of PaCE. Are there any additional ablation results, particularly focusing on the impact of different sparsity-promoting regularizers or dictionary sizes? Confidence: 4 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: The authors adequately address the limitations and potential societal impacts of their work. They discuss the computational overhead and the need for extensive concept dictionaries, as well as the importance of ensuring ethical considerations in the alignment of LLMs. Constructive suggestions for improvement include exploring more efficient algorithms for sparse coding and investigating the scalability of PaCE to even larger concept dictionaries. The paper could also benefit from a deeper discussion on mitigating potential biases in the concept partitioning process. Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Dear Reviewer Cdr5, We greatly appreciate your constructive feedback and kind acknowledgment of the novelty and advantages of our approach. It is our pleasure to address your comments and provide clarification below. Computational Efficiency Thank you for acknowledging the task (e.g., detoxification) performance of PaCE. Yes, it is possible to make the method even faster, at the trade-off of task performance. While our method jointly solves for the entire set of coefficients simultaneously, we can solve sparse decomposition in a greedy manner to speed up the computation. That is, the coefficients are obtained procedurally (i.e., one by one) in the order of energy minimization. The main benefit of using greedy algorithms is that the user can arbitrarily set the desired threshold on the number of non-zero coefficients in pursuit of better time efficiency. Based on such observation, we recap specific decomposition details in our paper and show an alternative speed-up version: * The elastic-net solver (Line 213) used in the experiments in $\S 4$ is designed to compute an exact solution of coefficients for the optimization problem (Equation (1), Line 186-189). The coefficients are obtained simultaneously. * Motivated by your inquiry, we additionally evaluate Orthogonal Matching Pursuit (OMP), a fast greedy solver in the compressed sensing literature [A1, A2] for our activation decomposition. On a high level, OMP iteratively 1) adds to the support the concept that has maximum coherence with the unexplained residual, and 2) updates the residual by solving a least square using the new support. It stops when a pre-defined maximum size k of support is reached. Intuitively, the k is the number of non-zero elements in the solved coefficients. We compare two solvers on response detoxification ($\S 4.1$) in the table below. * Notably, with a small choice of k=50, OMP is an order of magnitude faster than Elastic Net, while it achieves a safety score 12.3% lower than that of Elastic Net. * Overall, the observations in this experiment validate that a greedy solver can improve computational speed at the cost of safety performance. \| \| OMP (k=50) \| OMP (k=100) \| OMP (k=150) \| OMP (k=200) \| Elastic Net \| \|---\|---\|---\|---\|---\|---\| \| Time per decomposition (s) \| 0.045 \| 0.182 \| 0.381 \| 0.749 \| 0.411 \| \| Safety (↑) \| 63.1 \| 64.4 \| 66.9 \| 70.8 \| 72.0 \| Generalization to Other LLMs We evaluate PaCE on Mistral-7B-Instruct (version 1) for detoxification and observe PaCE’s superior performance while preserving linguistic capability. Mistral-7B’s strong instruction-following leads to a lower initial safety score and higher MMLU performance. Hence unlike LLaMA2-7B-Chat, the prompting baseline for detoxification does not significantly harm linguistic performance. \|Mistral-7B-Instruct\|Safety\|Linguistic\|Capability\|\| \|:---:\|:---:\|:---:\|:---:\|:---:\| \|Method / Metrics\|Average (%, ↑)\|Fluency (↑)\| Perplexity (↓) \| MMLU (%, ↑) \| \|Vanilla \| 5.20 \| 6.91 \| 3.57 \| 56.4 \| \|Prompting \| 54.8 \| 6.80 \| 3.58 \| 54.4 \| \|VecAdd \| 64.7 \| 6.69 \| 4.23 \| 44.3 \| \|OrthoProj \| 65.2 \| 6.74 \| 4.35 \| 44.9 \| \|PaCE (Ours) \| 76.3 \| 6.89 \| 4.19 \| 46.1 \| Interpretability of Concept Dictionary Figure 8 and Appendix D.2 show semantic similarity between concept vectors by clustering the first 10,000 concepts in our PaCE-1M dictionary. The observed semantic structures (salient clusters and decision boundaries) indicate that the target LLM has an activation space that understands and organizes semantic information of the concepts, enabling further analysis and manipulations in PaCE. Figure 10 and Appendix D.3 show the activation space's utility for concept retrieval, indicating close coherence between the target and relevant concept vectors. With full respect to your and Reviewer s54k’s insights on this, we will add a new section Appendix D.4 for discussions of analyzing other properties (e.g., sparsity, magnitude, direction). Ablation Study Thank you for inquiring about ablation studies on key components of PaCE. Figure 6 shows detoxification performance for LLaMA2-13B-Chat across different dictionary sizes, with safety scores increasing and converging around 9000-10000. For the regularization design, we evaluated different setups of $\tau$ in Equation (2) that implement the sparsity-prompting regularizer. The results are shown in the table below: \| $\tau$ \| 0 \| 0.35 \| 0.65 \| 0.95 \| 1.0 \| \|---\|---\|---\|---\|---\|---\| \| Note \| Pure $\ell_2$ \| N.A. \| N.A. \| N.A. \| Pure $\ell_1$ \| \| Safety (↑) \| 68.9 \| 65.4 \| 71.6 \| 72.0 \| 66.5 \| The results show that the regularization with $\tau$=0.95 (our choice stated in the paper Appendix B.4) yields the best safety performance among the five choices. Pure ridge regression ($\tau$=0) and pure lasso regression ($\tau$=1) do not perform as well as the mixed regularization strategy. Reproducibility We appreciate your emphasis on reproducibility. Appendix B.4 details our frameworks, experiments, and hyper-parameters for inferring on open-source LLMs, concept curation, and knowledge retrieval. We have also included details of our computing resources on GPUs. Upon acceptance, we will open-source the PaCE-1M dataset and PaCE implementation with documentation. We will continue maintaining the project (e.g., computing concept dictionaries and partitioning the concepts with new expert models) for future open-source LLMs and new alignment tasks. We will include the experiments above in $\S4$ and Appendix B.5 of the paper, and we will revise the writing based on your suggestions. We appreciate your valuable insights which have helped to validate and strengthen our framework. Thanks and regards, Authors of Submission #8804 [A1] _Orthogonal matching pursuit: Recursive function approximation with applications to wavelet decomposition._ Pati et al., 1993. [A2] _Orthogonal matching pursuit for sparse signal recovery with noise._ Cai et al., 2011. --- Rebuttal Comment 1.1: Comment: Dear Reviewer Cdr5, Thanks again for your time and efforts in writing the review. We addressed your concerns in detail in the rebuttal, and we hope the response covers the comments. We are more than happy to further clarify or address additional questions. Please let us know if you still have any unclear parts of our work. Sincerely, Authors	Rebuttal 1: Rebuttal: ### Global Response We thank all reviewers for their valuable time and intellectual input on the paper. We take the opportunity to provide a global summary here and communicate with the reviewers in the individual rebuttals. Presentation Quality As per reviewers’ comments, the paper has _thorough explanations and clear diagrams_ (Cdr5), provides _clear visualization and detailed interpretation to illustrate its underlying principles_ (75Pi), gives _comprehensive and clear articulations_ (rWAy), and is _well-written with a clear structure and logical flow_ (Cdr5). We also appreciate Reviewer s54k’s contributive inquiry about additional details of activation alignment, and Reviewer rWAy’s comments on baseline implementation. We have addressed them in the individual rebuttals. Novelty and Significance Our Parsimonious Concept Engineering (PaCE) sparsely decomposes a target LLM representation on a large-scale concept dictionary to precisely re-orient the LLM behavior and effectively improve its trustworthiness. * New dataset. We collect a large-scale concept representation dataset, PaCE-1M, that consists of 40,000 concepts extracted from over 1,200,000 context sentences. It is generalizable to multiple LLMs and its concepts are annotated for downstream tasks such as detoxification and sentiment revising. * Novel method. We decompose the neural activations as a sparse linear combination of these concept directions using efficient sparse coding techniques. The decomposition provides an effective and accurate estimate of both undesirable and benign components in the target representation, which is often overlooked in previous activation engineering methods. Indeed, the novelty of the proposed framework and the innovations in the pipeline are widely recognized by multiple reviewers. We appreciate the reviewers' comments that the proposed PaCE framework is _novel_ (s54k, rWAy, Cdr5), _different from the existing methods_ (rWAy), _unique, well-motivated, and well-developed_ (Cdr5), _adaptive and insightful_ (s54k). In the individual rebuttals, we address valuable comments for concept polysemy and questions on framework design for activation manipulation (75Pi), and we elaborate on design choices of sparse optimization (Cdr5). We also respond to the insightful observations on the position of PaCE among secure alignment schemes (s54k). Experiments To validate our proposed framework, we evaluate PaCE on multiple alignment tasks including response detoxification, faithfulness enhancement, and sentiment revising ($\S4.1$, $\S4.2$). We show that PaCE achieves state-of-the-art performance on these tasks while retaining its linguistic capability at a comparable level. We further investigate the LLM activation space by PaCE-1M samples, showing the geometric consistency of concept semantics and the interpretability of the PaCE decomposition ($\S4.3$). It is encouraging to have reviewers’ comments that the proposed approach is _showing improved alignment performance_ and _achieving outstanding results with 13B model_ (75Pi), and the framework is _state-of-the-art, scalable_ and _efficient_ (Cdr5). They also find that the _evaluation is comprehensive_ and _the results are insightful_ (s54k), and the _superiority of the proposed method in pursuing alignment goals_ is validated (rWAy). We are grateful to receive reviewers’ suggestions about computational efficiency (Cdr5, rWAy), additional malicious jailbreaks (s54k), additional benign (helpfulness) benchmarks (75Pi), and other target LLMs (Cdr5, 75Pi). These inquiries on experiments help us to further confirm the general applicability and outperformance of our framework, and we have addressed all of them in the individual rebuttals. Summary In each of the individual rebuttals below, we address the reviewers’ valuable suggestions on paper presentation, framework validity, additional experiments, and all other perspectives. All insights and questions are highly important to the continual improvement of our work, and based on them we have made or will make multiple revisions to the manuscript and appendix as promised. In summary, we are more than happy to receive four high-quality, solid, and insightful reviews this time; thank you all. Please feel free to communicate with us during the discussion period if you have any further questions. Pdf: /pdf/6c4fb30fe2cad15b46c1d78b02323bcad4951cc0.pdf	NeurIPS_2024_submissions_huggingface	2,024	null	null	null	null	null	null	null	null
Thought of Search: Planning with Language Models Through The Lens of Efficiency	Accept (poster)	Summary: The paper analyst the use of LLM in planning, and propose to write the successor function and goal test in code instead of directly solving the problem. The paper showed experiments that using the code can get a higher accuracy and lower number of calls to the LLM compared to LLM-based solutions. Strengths: The paper is very well-written and easy to follow. The discussion is very detailed and provide extra insights on the topic. Weaknesses: - First of all, I think the conclusion is not surprising or bring any insights at all. The idea of letting LLM to write code instead of directly solve problems has extensively appear in LLM for reasoning [1], and more specifically on LLM for planning [2][3]. - Because the conclusion does not provide any big insights, I am expecting to see a strong experiment, which is not the case in this paper. The benchmark selected, are quite classic from different manners, and similar code has appeared on GitHub for at least a year. It is hard to justify whether the success of generating code is coming from LLMs remembering similar context from Github or they do have the ability to correctly generate code. And even in this case, the authors mentioned “The mistakes GPT-4 makes when producing the code repeat from one experiment to another”, which is not a good sign if one wants to deploy similar methods to more general applications. - One minor drawback is that I feel the authors failed to cover some related works that are quite close, for example the ones I mentioned in the first weakness. A more thorough literature review are recommended for a more convincing paper. [1]. Zhou, Aojun, et al. "Solving challenging math word problems using gpt-4 code interpreter with code-based self-verification." arXiv preprint arXiv:2308.07921 (2023). [2]. Liu, Bo, et al. "Llm+ p: Empowering large language models with optimal planning proficiency." arXiv preprint arXiv:2304.11477 (2023). [3]. Guan, Lin, et al. "Leveraging pre-trained large language models to construct and utilize world models for model-based task planning." Advances in Neural Information Processing Systems 36 (2023): 79081-79094. Technical Quality: 3 Clarity: 4 Questions for Authors: I am concerned about the current need of human feedback to successfully generate the code. While I can see from the appendix that a large portion of them might need such feedback, can you provide some statistics about this? Confidence: 3 Soundness: 3 Presentation: 4 Contribution: 2 Limitations: Yes. Flag For Ethics Review: ['No ethics review needed.'] Rating: 4 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We thank the reviewer for their feedback. We hope that our clarification regarding the statistics on when feedback was needed alleviate the reviewer’s concerns and can allow to raise their rating. ## Answer to the Question: We have provided the average number of interactions with the model (required to produce the correct code) in the paper. The average is computed over 5 separate runs. The number of times a feedback was needed is one less than the number of interactions - the first interaction asking to produce the code is also counted. For 24 game (line 204): 1.2 interactions on average for the successor function and 1 interaction for the goal test. That means that the goal test did not need any feedback and the successor function needed a single feedback once out of 5 runs. For Mini crosswords (lines 219-220) the model required 2.4 interactions on average to produce a valid successor function and 1.4 interactions on average to produce the goal test. For BlocksWorld (line 239), 2.8 and 1 iterations on average for successor function and goal test, respectively. For PrOntoQA (line 275), 1.6 and 1 iterations on average for successor function and goal test, respectively. The sum of the two averages (successor function + goal test) is shown in the “Calls” column of Table 1, last row. ## On weaknesses: * While we agree that the idea to ask the language models to produce code is not revolutionary, we are not aware of other work that uses language models to produce search components, such as successor function and goal test. Further, we consider this to be only one of the contributions of our work. Please see the response to all reviewers for the discussion on the contribution of our work. * On related work: we will add the mentioned papers to the related work. * Guan et al, NeurIPS 2023 is already cited in our work. The paper proposes generating a classical planning model (PDDL), under particular assumptions. The direction is complementary to our work and is probably the more efficient method, when applicable, as it allows using existing planners. Unfortunately, not all planning problems are easily captured by a classical planning model and in such cases our method can still help. One example of such case is the 24 game we experiment with in our work. * Liu et al, Arxiv 2023 assumes that the PDDL domain (the major part of the PDDL model) already exists and proposes a way to use LLMs to produce PDDL problem instances (objects, initial state, goal). We do not make such assumption; rather we are asking LLM to convert natural language domain information to python code. * While our focus is on using LLMs for generating code for successor and goal functions for search problems, Zhou et al. ICLR 2024 focuses on math reasoning problems. They illustrate that LLMs can be used to generate, execute (verify), and self-refine the python code for math problems. Their study corroborate our findings that LLMs can indeed be used for generating verifiable code with some feedback. --- Rebuttal Comment 1.1: Comment: Thanks for addressing my concerns and pointing out the connection between the literature I mentioned. I am still concerned about the novelty of this work being quite incremental. Although I agree with the authors that not all problems can be solved with PDDL, I do not think this part poses a significant enough challenge to the problem studied in this paper: translating human language description to a specific language of code. Given that existing planners also use search in their underlying architectures, the work seems just challenging LLM to generate a more popular language of code, Python, instead of PDDL. In my opinion, the current paper has not adequately addressed the strong connection with this NeurIPS paper. Instead of hiding this reference in a short number, I would like to see a full discussion on how the current paper is different enough from the NeurIPS paper. Furthermore, I personally believe that the transition gap between math reasoning problems and planning problems is bigger than from PDDL to Python. If the authors do believe Zhou et al. ICLR 2024 can be used as a cross-reference to support that LLMs can generate verifiable code with automatic feedback instead of human feedback used in the current paper, I do believe the contribution to be even smaller. Otherwise, the need for human feedback in the loop still seems to be a big challenge to the proposed algorithm. --- Reply to Comment 1.1.1: Comment: # Comparison with Guan et al. Guan et al, NeurIPS 2023 propose a method that consists of three parts: domain construction, domain refinement by a human, planning with the refined domain. To construct the domain, an LLM is queried for action preconditions and effects on an action-by-action basis, providing it with the description of the action in a natural language and a (possibly incomplete) set of predicates. The LLM can provide not only action parameters, preconditions, and effects, but also potentially propose missing predicates. Few shot examples from a BlocksWorld domain are given in the prompt. The process is repeated once, with the list of full predicates from the first iteration. Additionally, PDDL problem instances are created, using the predicates from the domain generation process. The feedback is provided in two forms: 1. A symbolic validator VAL is run on domain and problem files, with the result being translated into natural language feedback. This feedback is mostly on the syntax of the generated PDDL model. 2. The generated PDDL domain is translated into a natural language and presented to a human expert. The expert provides an explicit feedback on missing and extra preconditions and effects of each action. Upon correspondence with the Guan et al. authors, we validated following conceptual differences from our work: 1. They are interested in actions only, assuming initial state and goal are given. Our work does not make this assumption; rather asks LLM to generate a goal function. 2. They feed the language model with targeted pieces of information (single action description one-by-one, predicates to use). We provide entire description of the problem and asks for a single successor function. 3. Their feedback on the generated PDDL is explicit and requires nuance understanding of modeling in PDDL (e.g., 'unnecessary precondition "no other object stacked on object ?x"'). While ours is mostly generic python code feedback (e.g., issues with shallow copy of a dictionary with list values, error trace) and generic logic (e.g., 'two of the operations are not symmetric, division and subtraction'). We agree that this comparison is interesting and will add the discussion to the paper. # PDDL vs Python code generation There is a conceptual difference between representing planning problems in PDDL and coding a search problem (successor function and goal test). To overcome the limitations of PDDL, modelers often resort to tricks like presenting additional predicates encoding the negation of modeled predicates, adding predicates that explicitly encode an information that could be derived from other predicates, such as (hand empty) in addition to (holding ?b) in BlocksWorld, because such derivation cannot be done in the preconditions and effects in classical planning, and would require axioms, which are rarely supported by existing planners. As a result, human validation of such PDDL models is often harder and requires more skills than validating code-based successor function/goal encoding. --- Reply to Comment 1.1.2: Comment: # Comparison with Zhou et al. We share your belief that the transition gap between math reasoning problems and planning problems is large. Therefore, their approach requires independent interaction with LLM for every math problem. A search problem, however, should not require independent interactions with LLM for each problem. That is the main premise our work. A single interaction with LLM to generate successor and goal function can allow solving all problems in that domain. In our previous comment we only meant to highlight that success in Zhou et al's work indicate that LLMs can generate code and refine it with feedback. In our work we leverage this code generation and refinement ability of LLMs. But the similarity ends there. We highlight some of the differences between our work and Zhou et al. below: 1. They propose generation of code as a means to solve a given math problem and generate final answer. In our work, the final answer is the code itself. 2. Their approach is not iterative or incremental in nature, they propose generating a predefined collection of validators for a particular instance, regardless of the performance of previously generated validators. Ours iteratively fixes issues with the previously generated code. # On the need for human feedback The need to provide human feedback is shared by all approaches investigated in our work. In our case, the human feedback is needed on producing the solver and is not required once a sound solver is produced, as the solutions are then guaranteed to be correct. In the case of the previous approaches, the human feedback is needed to validate each and every one of the produced solutions, an almost impossible task. We are not aware of an acknowledgement of such a limitation in the existing literature. We hope that the challenge of alleviating the need for human feedback can be adequately addressed in future work. # On novelty Finally, We would like to emphasize that a large portion of our work focuses on filling the gap in the current literature on planning with LLMs with regard to the computational complexity and properties of the proposed algorithms. This investigation is essential for understanding approaches and building up on them. So, we believe, this investigation itself satisfies the novel contribution requirement. Additionally, none of the existing approaches use language models to generate code for search components, so that contribution is also novel.	Summary: The authors propose a position paper that argues current works for LLMs for planning waste significant compute, on top of having poor algorithmic and empirical performance. The authors also propose ideas on how to use LLMs more efficiently and effectively by using them to preprocess search algorithms instead of using them directly during search. More specifically, the proposed method consists of using the LLM to generate the successor generator and goal state checker, alongside user feedback. Strengths: The paper has several strengths as a position paper. It provides a simple yet original idea, namely that researchers should strive for responsible usages of LLMs for planning in terms of computing efficiency and also provides arguments that this would actually improve the performance of LLMs for planning as well. The idea is complemented by an extensive survey of LLMs for planning methods with a summary of worst case complexities, and whether the algorithms are sound or complete. Furthermore, it is complemented by a novel methodology for using LLMs for planning which adheres to the authors' proposition: more efficient and effective LLMs for planning research. The experiments are extensive with a wide variety of planning benchmarks, implementation details, and results. The results are quite positive as summarised in Table 1, with minimal calls to LLMs in comparison to existing approaches. The authors are also transparent about the limitations of their approach in its current state, being that it requires user interaction. Nevertheless, this does not undermine the proposition of the position paper. Weaknesses: With regards to the idea of the paper, there are no major weaknesses. Nevertheless, the paper could benefit from some additional details regarding experiments, and improved clarity in certain areas. - It may not be clear to some readers how %States could go over 100%. By its definition in line 344-345, it is not clear whether visiting the same state twice double dips into the percentage or not, but from the next 2 sentences, it does seem to be the case. - Minor formatting issue: the benchmark domains are not consistently capitalised, e.g. 24 game and crossword in line 357 but elsewhere they are capitalised such as in Table 1 and page 8. - Although the focus of the paper is in LLMs for planning, the paper misses more general related work regarding learning for planning/generalised planning. Such methods are magnitudes of orders more efficient than LLMs in evaluating learned heuristics (ToT or GoT in LLM terminology) or policies and solve problems magnitude of orders larger than problems solved by LLM research. Example works include learned heuristics [1] or generalised policies [2], as well as foundation models for planning [3]. [1] Simon Ståhlberg, Blai Bonet, Hector Geffner: Learning General Optimal Policies with Graph Neural Networks: Expressive Power, Transparency, and Limits. ICAPS 2022: 629-637 [2] Dominik Drexler, Jendrik Seipp, Hector Geffner: Expressing and Exploiting Subgoal Structure in Classical Planning Using Sketches. J. Artif. Intell. Res. 80 (2024) [3] Dillon Ze Chen, Sylvie Thiébaux, Felipe W. Trevizan: Learning Domain-Independent Heuristics for Grounded and Lifted Planning. AAAI 2024: 20078-20086 Technical Quality: 3 Clarity: 3 Questions for Authors: No questions or clarifications that could change my opinion. Confidence: 3 Soundness: 3 Presentation: 3 Contribution: 4 Limitations: The authors adequately addresses limitations of their work, and also the checklist with appropriate justification. Flag For Ethics Review: ['No ethics review needed.'] Rating: 7 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We thank the reviewer for their feedback and support. We hope that the reviewer could advocate for the paper. Your understanding regarding the %States is correct, we will clarify the text in the final version. --- Rebuttal Comment 1.1: Comment: I thank the authors for their response and clarification. I also acknowledge that I have read the strengths and weaknesses pointed out in other reviews as well as the corresponding rebuttals but still stand with my rating. More specifically, I am still convinced by the message concerning the efficient usage of LLMs that the paper proposes with a focus on soundness. --- Reply to Comment 1.1.1: Comment: We greatly appreciate your support. It does look like we will need it.	Summary: Existing LLM-based planning approaches usually involve searching and multiple passes of the model, which leads to significant inefficiency and cost while failing to guarantee the correctness of the generated plans. Motivated by this issue, this paper first analyzes the soundness, completeness, and complexity of a series of existing planning methods, arguing that they do not meet the standard of these properties. A new algorithm, Thought of Search (ToS), is proposed to alleviate the heavy computing demand of the searching operation for planning. It queries LLMs for generating the code implementation of successor functions and goal tests. The proposed method achieves 100% success rate for four representative search problems, while the total time spent (on CPU) is shorter than or comparable with a single LLM evaluation time. Further discussion describes that, compared with other approaches, ToS is not only sound and complete but also more cost-effective and able to explore a much larger portion with only O(1) complexity. Strengths: 1. Impactful motivation. Recently, there have been increasing efforts to improve the planning ability of LLMs. However, since this line of research is still at its early stage, the efficiency of the proposed methods is easy to overlook, especially when the system contains multiple agents. Therefore, I strongly agree with the motivation of this paper (i.e., the need for sound and complete LLM-based approaches that uphold efficiency) and believe in its importance for future research. 2. Comprehensive analysis. The paper systematically and comprehensively studies the properties of twelve related works that are commonly used or recently proposed, providing convincing support for the authors’ claim and, more importantly, valuable information to the community. 3. Solid results. Across all four evaluated tasks, ToS consistently generates valid final solutions only and reaches 100% accuracy in a relatively short time. Notably, the searching operation is run on the CPU. These results are sufficient to demonstrate the effectiveness of the proposed method. Weaknesses: 1. Current works on code generation using LLMs show that the correctness of the generated code is not guaranteed. In this paper, the obtained implementations in the experiments are also not always valid at the first trial and require human feedback. Thus, I am slightly concerned about whether ToS can generalize to more difficult tasks while maintaining its nice properties. Nonetheless, this is likely to be alleviated by combining with some automated optimization techniques as discussed by the authors. Therefore, I am still relatively optimistic about this approach at this point. Further experiments (if any) to address this concern are welcome. 2. While the paper's contents are generally well organized, which I appreciate, there are quite a few typos and wrong words/phrases. For exemplification, a non-exhaustive list is written below. I highly recommend a careful and thorough check of the whole paper to fix all the mistakes. - Line 22: ‘The purpose of our work is precisely that.’ - Line 99: ‘Reflection’ - Line 323: ‘a lower than reported by the approaches accuracy’ Technical Quality: 3 Clarity: 2 Questions for Authors: See weaknesses Confidence: 3 Soundness: 3 Presentation: 2 Contribution: 3 Limitations: The limitations have been discussed. Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We thank the reviewer for their feedback and hope that we can somewhat alleviate their concerns and get their strong support for this work. Existing literature on code generation e.g., [1,2], as well as the literature on generalized planning with LLMs e.g., [3] shows evidence that automated feedback improves LLM performance. Our preliminary investigation of feedback automation for ToS supports this as well. We will proofread the paper and fix the typos - thank you for pointing these out! [1] Chen, Xinyun, et al. "Teaching large language models to self-debug." ICLR 2024 [2] Zhang, Kechi, et al. "Self-edit: Fault-aware code editor for code generation." ACL 2023 [3] Silver, Tom, et al. "Generalized planning in pddl domains with pretrained large language models." AAAI 2024. --- Rebuttal Comment 1.1: Comment: We hope that our responses have strengthened your support for the paper. We would greatly appreciate if that could be reflected in your final score.	Summary: The authors propose a Thought of Search: thinking before searching strategy to solve Automated Planning problems using LLMs. They use the GPT-4 LLM to generate a Python code for generating successor states and goal test functions which are the crucial parts of any search. They argue that this method is sound and complete and requires the least calls to the LLM before successfully solving the problem when compared to the relevant literature. Strengths: 1. Analysis of Complexities for the existing methodologies of using LLM for Planning. 2. Innovative use of LLM to reduce the number of calls made to solve the problems. Weaknesses: However, there are several concerns regarding the proposed work: 1. If the Large Language Model (LLM) is being used to generate successor and goal test functions, which are already intrinsic components of automated planners, it is unclear what improvement is being made. The inherent memory and time complexities associated with solving these planning problems are not addressed or mitigated by the proposed approach. It appears that the LLM is simply re-writing (a few components of) a more basic planner, which does not seem sufficiently innovative for a NeurIPS-level conference. 2. The process is not fully automated, as it requires human intervention to re-prompt the LLM until the correct code is generated. 3. The methodology would have been more compelling if it had included an exploration of generating useful heuristic functions. Technical Quality: 3 Clarity: 3 Questions for Authors: 1. Why did the authors use only GPT-4 for their experimentation? No reasons are presented in the paper. Presenting a comparison of the performance with different LLMs would have been interesting to see. 2. What is the need to use LLMs to generate successor and goal test functions and solve the planning problems with a naive blind search? Is the objective to evaluate GPT-4's capability to generate Python code, or to assess its reasoning ability in solving planning problems? 3. Comparing the number of calls to the LLM with other approaches may not be appropriate. Your methodology necessitates generating the search components once for each domain, whereas other approaches involve the LLM in solving every problem within the domain. A more relevant comparison would be the average time required to solve these problems and the success rate across different methodologies. Confidence: 5 Soundness: 3 Presentation: 3 Contribution: 1 Limitations: 1. Showcasing results with only GPT-4 model. 2. Incomplete comparisons with the other existing approaches as mentioned above. Flag For Ethics Review: ['No ethics review needed.'] Rating: 3 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We thank the reviewer for their feedback. ## Answers to the Questions: 1. Our purpose was to show the feasibility of the approach rather than comparing which models are better at generating successor functions / goal test for the four search problems. For that purpose one language model is sufficient. Further, all papers we compare to (described in section 2) use GPT(3, 3.5, or 4) in their experiments and therefore, by using GPT model we could compare to their results without redoing their experiments, which we show in this paper to be so unnecessarily wasteful. 2. The objective of this work is neither to evaluate GPT-4's capability to generate Python code, nor to assess its reasoning ability in solving planning problems. Our primary objective is to fill the gap in the current literature on planning with LLMs and point out the inefficiency and pitfalls of the current trends. Please see the response to all reviewers for the discussion on this. In order to solve a search/planning problem, one needs to be able to capture the dynamics of that problem. Of course, for planning problems that are already expressed in a formal language, such as PDDL, one can simply use an existing PDDL planner, which internally performs a search, defining the successor and goal functions based on the PDDL. If the problem does not have a PDDL yet, but it can be easily captured in PDDL, one may prefer to do so (with or without the help of LLMs), as it is done in e.g., the cited work (Guan et al, NeurIPS 2023, Oswald et al, ICAPS 2024). Many search problems, however, are not easy to capture in PDDL, and therefore an alternative approach is needed. This is the case for many of the planning domains mentioned in the recent literature and we used some of them in our work. Probably the best example is 24 game, which has numeric features not easily captured by classical planning. Please see the discussion in Section 3. 3. The number of calls to the language model is precisely how we measure complexity in this work. Additionally, we provide the overall time and accuracy comparison in the paper (more on this below). - One of the major advantages of our approach is that we only need to call the LLM a constant number of times per domain, regardless of the number of problems in the domain. However, even if you want to take away this advantage and say that each domain has only a small number of problems, our approach needs less LLM calls per domain than the other approaches per problem in the domain. Further, after the successor function and the goal test were obtained, solving all the problems in a domain by search on a single core of a personal computer CPU typically takes as much time as a single call to the LLM. - Both the total search time (the average time is easy to derive since the number of instances is provided) and the accuracy/success rate results are provided in the paper. The success rate of our approach is 100% in all the tested domains. The success rates reported in the literature for 24 Game and for PrOntoQA are presented in lines 198 and 268, respectively. The success rate of ToT on mini crosswords is 20% and the success rate of RaP on BlocksWorld is 100%, 88%, and 42% for problems with solution length of 2, 4, and 6, respectively. We forgot to mention it in the paper, will add. The total search time results for our approach are provided in the text, lines 210, 226, 247, 278. To exemplify, solving all the 1362 instances of 24 game takes 2 seconds with the “fastest” successor generator and 7 seconds with the “slowest” among the 5 times we conducted the experiment. As mentioned before, our accuracy is 100% - we solve all 1361 solvable games and report unsolvable for the unsolvable one. In comparison, the ToT approach (Yao et al, NeurIPS 2023) restrict their experiments to 100 out of 1362 tasks and performs ~100 calls to an LLM per task. Assuming the same average number of calls and 7 seconds per call, the ToT approach would take around 10^6 seconds or 11 days, achieving the reported success rate of 75%. Even if the LLMs become significantly more efficient and the time of a single call would be cut down to 1s, it would still take ToT more than 1.5 days. Please also see the response to Reviewer PhwG regarding automating the feedback. --- Rebuttal Comment 1.1: Comment: Thank you for addressing my comments. I am still not convinced that the work produced in its current state is novel enough for the NeurIPS submission. The authors seem to see their methodology is applicable in cases where devising PDDL is harder. To generate feasible plans for realistic/near-to-realistic domains using the methodology proposed is still the same as a blind search and would require a lot of computation and resources. I read through the rebuttal responses for other reviewers. On automation - "Our preliminary investigation of feedback automation for ToS supports this as well." - do the authors provide this preliminary investigation in their paper? if so can you please point out where? --- Reply to Comment 1.1.1: Comment: Our preliminary investigation of feedback automation is not part of this work. It is not clear to us whether reviewer's intention is that blind search is not as effective as the previous approaches (ToT, RaP, etc) or that blind search is not as effective as heuristic search on large realistic domains. Could you please clarify?	Rebuttal 1: Rebuttal: The main objective of our work is to fill the gap in the current literature on planning with LLMs with regard to the computational complexity and properties of the proposed algorithms. Our main contribution is precisely this investigation. We show that the current literature proposes inefficient methods for producing unsound results. We present complexity analysis of the algorithms proposed and establish the lack of soundness and completeness of these algorithms. We not only show just how inefficient the results are, but also show that there is another way by proposing a simple alternative. We would like to highlight the importance of soundness, which is mostly overlooked by the existing literature body. With soundness, the solutions produced by search are guaranteed to be correct and do not require external validation. Without soundness, the produced solutions have a large potential to be invalid, and an automated validator would be needed. It is not clear from the literature on the algorithms for planning with LLMs studied in this work whether such validators were created and used to verify the claimed success rates. To clarify, let us use the 24 Game as an example, let's assume the initial state is [3 3 8 8] (one of the instances in the existing benchmark). An LLM could produce [24] as a successor during search. If you do not validate each transition from [3 3 8 8] to [24], you would not be able to validate the answer. Note that there is no way to reach [24] from [3 3 8 8]. Producing a goal state does not mean solving the problem and a sound validator is required. We feel that it is crucial to expose the scientific community to these results, giving our position a stage at this point in time, in an attempt to reduce the amount of work that continues the same trend in the recent literature.	NeurIPS_2024_submissions_huggingface	2,024	null	null	null	null	null	null	null	null
Efficient and Private Marginal Reconstruction with Local Non-Negativity	Accept (poster)	Summary: This paper introduces residuals-to-marginals (ReM), a novel approach to constructing marginals from input data residuals while maintaining differential privacy. ReM operates by first estimating true residuals from noisy ones, then transforming these estimates into target marginals. The authors propose two variants of ReM: ReM-LNN, which enhances accuracy by enforcing non-negativity constraints on the resulting marginals, and GReM-MLE, a more efficient version for scenarios involving Gaussian noise injection to residuals, leveraging a closed-form solution. The paper also presents an extension of ReM that incorporates the MWEM. This extension identifies residuals that are useful for marginal construction, offering an end-to-end solution when residual choices are not predetermined. Experimental results demonstrate that ReM significantly reduces errors compared to baseline methods in constructing differentially private marginals. ReM also demonstrates good scalability, whereas PrivatePGM, a baseline method, failed to run for large data domains. Strengths: 1. This paper presents ReM, a new approach to constructing marginals from residuals. 2. ReM reduces errors compared to some baseline methods. 3. ReM can be applied as a post-processing method to improve the accuracy of some other methods. Weaknesses: 1. The paper compares ReM against MWEM and Private-PGM [12], but omits the state-of-the-art methods such as AIM [7] and PrivMRF [19]. Both AIM and PrivMRF have been demonstrated to significantly outperform MWEM and Private-PGM. As a consequence, it is unclear whether ReM really advances the state of the art. 2. It would be helpful to explicitly define k_tau in Line 1 of Algorithm 2. I do not see where this paper defines k_tau. It seems that for each attribute set tau, the algorithm injects noise k_tau times, resulting in k_tau marginals, ranging from y_{tau, 1} to y_{tau, k_r}. Why is it necessary to inject noise multiple times for each marginal? Technical Quality: 3 Clarity: 3 Questions for Authors: 1. Please include the derivation of L_tau in Line 216, Page 5. It appears to represent the probability of generating y_{tau, i} given alpha, over the randomness of Gaussian noise. Additionally, please provide a detailed argument for the closed-form solution mentioned in Line 218. 2. Could you clarify why "when the marginal is observed with isotropic noise, the corresponding noisy residuals are independent"? (Line 239, Page 6) Confidence: 4 Soundness: 3 Presentation: 3 Contribution: 2 Limitations: Yes. Flag For Ethics Review: ['No ethics review needed.'] Rating: 3 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We thank the reviewer for providing feedback on our paper and suggesting improvements to clarify the presentation. Below, we address various concerns raised in the review. Since ReM is a reconstruction method, the appropriate comparison is not with full end-to-end synthetic data or query answering mechanisms such as AIM and PrivMRF but rather with the reconstruction methods these mechanisms use. Both AIM and PrivMRF use Private-PGM as part of their reconstruction steps, and we include Private-PGM in our experiments. Regarding Scalable MWEM, we use this end-to-end mechanism to choose which marginals to measure but compare Scalable MWEM's reconstruction method (discussed in Appendix D) to ReM with Local Non-negativity (GReM-LNN) and Private-PGM in our experiments. ReM takes an arbitrary number $k_\tau$ of measurements of each residual $R_\tau \in \mathcal{S}$ as input. We will be more explicit about $k_\tau$ when it is introduced in the text on Line 189. Regarding why a residual query might be measured multiple times, please see the global response. The loss function $L_\tau$ from Line 216 comes from using the negative log-likelihood as the loss function (as described in Line 190) for the noise model described in Lines 212 and 213, i.e., with $k_\tau$ independent Gaussian random variables $y_{\tau, i}$ with covariance $\Sigma_{\tau, i}$, so that $$ L_\tau(\alpha_\tau) = -\sum_{i=1}^{k_\tau} \log p(y_{\tau, i} \| R_\tau p = \alpha_{\tau}) $$ with $p(y_{\tau,i} \| R_\tau p = \alpha_\tau) = \mathcal{N}(y_{\tau,i}; \alpha_\tau, \Sigma_{\tau,i})$. We solve for $\hat \alpha_\tau$ in closed-form by finding the critical point of $L_\tau$. In the revision, we will include the derivation of the closed-form solution to $\hat \alpha_\tau$ in an appendix. Lines 238-240 are an informal statement of Theorem 2. To clarify this, we will specify that $z_\tau$ is a ''noisy residual'' in the statement of Theorem 2.	Summary: The paper studies the problem of reconstructing a data distribution from noisy measurements of a set of carefully selected marginal queries and using the reconstructed data distribution to answer a set of workload queries. The difficulty in the problem lies in the high dimensionality of the data distribution, which in turn requires measuring and inverting an exponential number of queries in the worst case. In this paper, the authors propose to 1. translate the marginal workload queries to their equivalent residual queries representation, and compute their noisy measurements 2. exploiting the structure of residual query matrix to perform efficient pseudoinverse via Kronecker product 3. perform reconstruction via an optimization problem that has solution $=$ the product of pseudoinverse matrix and noisy measurements, where additional non-negativity constraints on the solution are enforced as post-processing Compared to prior approaches for marginal reconstruction, the proposed approaches enjoy improved efficiency as the Kronecker product matrix is efficiently invertible. The technique of converting the query matrix to its residual representation is further extended to the MWEM algorithm to allow efficient computation and representation of reconstructed marginals under the high-dimensional data. Experiments confirm that the proposed reconstruction and scalable MWEM algorithms allow improved query-answering performance when used on top of existing mechanisms that are solely based on noisy measurements rather than reconstructed data distribution. Strengths: - An interesting idea of using the residual basis of query matrix to efficiently represent and compute pseudoinverse matrix for marginal reconstruction. - Based on the observation, two interesting new algorithms are proposed: efficient marginal reconstruction both via reconstructing data distribution, and scalable MWEM on high-dimensional data via residual computation and representation of marginal queries. - Strong experimental performance when used as post-processing mechanisms on top of existing query-answering mechanisms that are solely based on noisy measurements rather than reconstructed data distribution. Weaknesses: 1. One concern is regarding the discussion of computation complexity of the proposed method. The main advantage of the proposed algorithms is that they "will not have exponential complexity" (line 46). However, as the proposed algorithms still involve some enumerating operations on all possible attribute subsets (e.g., lines 9-11 in algorithm 3), it is not immediately clear why they "will not have exponential complexity". It would also be helpful if the authors could clarify what they mean by exponential complexity. 2. Relevant to the previous question, there are well-known hardness results of generating synthetic data under differential privacy [a], that talk about the necessity of exponential runtime for accurate reconstruction of data distribution. Such lower bounds should be discussed in detail to understand the results of this paper, in terms of the saved computation complexity. 3. The comparison is mainly done with regard to Residual planner [6] and PGM [12], while several other reconstruction methods such as PrivBayes [8], GEM [9], RAP [10], and RAP++ [11] are not considered in the comparison. Such comparisons are useful for understanding the performance of the proposed reconstruction algorithm, and for understanding the contributions of the paper. Alternatively, it would be useful if the authors could further clarify why such algorithms are not considered in the comparison. [a] Ullman, Jonathan, and Salil Vadhan. "PCPs and the hardness of generating private synthetic data." Theory of Cryptography Conference. Berlin, Heidelberg: Springer Berlin Heidelberg, 2011. Technical Quality: 3 Clarity: 3 Questions for Authors: 1. Why would not the proposed algorithms have exponential complexity (weakness 1)? And how does this relate to the known hardness results (weakness 2)? 2. Could the authors explain more about the lack of comparison with other reconstruction methods (weakness 3)? Confidence: 3 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: NA Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We thank the reviewer for providing feedback on our paper and suggesting improvements to clarify the presentation. Below, we address various concerns raised in the review. Regarding time complexity, please see the global response for specific running time bounds, which we will include in the revised paper. Regarding ``exponential complexity'', we will also clarify this. Our goal is to bound the running time of the reconstruction problem in terms of the size of its own inputs and outputs. The inputs are the noisy residual measurements, and the outputs are the reconstructed marginals. The results we state in the global response show that our algorithms take time that is polynomial (and subquadratic) in these size parameters, and thus avoid exponential complexity. As a concrete example, suppose the workload is all 2-way marginals over $d$ attributes, all attributes have domain size $m$, and the measurements are taken from ResidualPlanner. Then the input size and output size are both ${d \choose 2} m^2$. Our methods are polynomial (subquadratic) in ${d \choose 2} m^2$ (and thus in $d$ and $m$), while a method that reconstructs the full data vector will take $\Omega(m^d)$ time. Thank you for the comment about hardness results. We will add a discussion of complexity results to our paper. Briefly, the theorem of Ullman and Vadhan [a] does not apply because our mechanism does not output a synthetic data set. Indeed, as noted by Vadhan [b], ''the requirement that the mechanism produces a synthetic dataset cannot be removed'' from this hardness result. This is because the class of queries used in the result (two-way marginals) can be answered with vanishing error in polynomial time by a mechanism (e.g., the Laplace mechanism) that directly answers the queries without producing synthetic data. We will also add a discussion of other computational hardness results for query answering. In general, these results rely on worst-case query classes constructed from cryptographic primitives and not natural query classes such as marginals, so we don't expect them to apply to our practical use cases. Vadhan [b] highlights it as an open problem to give query-answering hardness results for natural query classes. Regarding comparisons, our goal in this paper is to study the reconstruction subproblem and develop general-purpose solutions. Among the alternative algorithms, only Private-PGM is a self-contained reconstruction method. The other algorithms (PrivBayes, RAP, GEM, etc.) are full query-answering or synthetic data mechanisms, which include some approach to reconstruction in the context of their mechanism. In several cases, such as RAP, a generic reconstruction approach can be extracted. RAP uses a neural network to represent the solution space and gradient descent variants to train it. However, to the best of our knowledge, due to the parametric modeling assumptions of these mechanisms, reconstruction requires solving a non-convex optimization problem. Thus the final result will be dependent both on how well the data distribution meets the parametric modeling assumptions, and on how well one is able to solve the non-convex optimization problem in practice. For the purpose of studying reconstruction in a self-contained manner, we choose in this paper to restrict to Private-PGM, which is the only other method that is general purpose and has predictable output due to solving a convex optimization problem to optimality. In-depth empirical comparisons of different reconstruction methods in the context of full mechanisms, including the impacts of non-convexity, etc., is an interesting avenue for future work. [a] Ullman, Jonathan, and Salil Vadhan. "PCPs and the hardness of generating private synthetic data." Theory of Cryptography Conference. Berlin, Heidelberg: Springer Berlin Heidelberg, 2011. [b] Vadhan, Salil. The Complexity of Differential Privacy. In: Lindell, Y. (eds) Tutorials on the Foundations of Cryptography. Information Security and Cryptography. Springer, 2017.	Summary: Matrix mechanisms are a foundational class of methods in differential privacy for efficiently answering large sets of marginal queries on tabular data. However, matrix mechanisms infamously suffer in high-dimensional data domains, where memory and compute explode. Significant efforts have been made to overcome this hurdle including the HDMM, the Private-PGM, and recently, the ResidualPlanner. This paper generalizes the residual decomposition techniques of ResidualPlanner to be far more effective and usable. They authors demonstrate how to answer arbitrary workloads of queries from arbitrary sets of residuals in a principled loss-minimizing way. They then extend this technique to offer non-negativity (when marginals cannot be negative-valued), and demonstrate how to apply it for the canonical MWEM mechanism. The show the significantly improved error of their reconstructions over a variety of baselines including the original ResidualPlanner. Their method is only improved on by Private-PGM, which suffers heavily from memory blow-up (in my own experience as well). Strengths: The methods demonstrated in this paper constitute a significant improvement to the state of the art in a foundational problem in Differential Privacy. They take the clever concepts of residual-based high-dimensional reconstruction and make them practical across a wide range of settings. They demonstrate how these methods can be used for canonical mechanisms and a significant improvement in utility on standard datasets/tasks. Weaknesses: There is a lot of technical content in this paper, which is challenging to get through in 9 pages. The writing could have been organized differently to help the reader gain intuition on how this method improves the state of the art. It took me a few reads. For instance, it is clear to me that the fundamental challenge is the computational intractability of the pseudoinverse, and that residual workloads help solve this. I still do not have great intuition on how residual workloads make this improvement. Perhaps I’m slow on this point, but it feels like it should be illuminated more prominently since it is central to the problem. Granted, residual workloads were introduced in a prior work. Another confusion: a major asset of ReM when it is introduced is that a given residual can be queried multiple times and then more accurately estimated with MLE. It was not clear to me why one would want to measure a single residual multiple times. Later in the scalable MWEM section, I could see how a residual would be queried multiple times as it appears in multiple marginals, but I was not sure if that’s the only reason. I think the motivation behind ReM could be described differently, perhaps even starting with something the reader knows (MWEM) and then showing how ReM and grem-lnn come to the rescue. Technical Quality: 3 Clarity: 2 Questions for Authors: See two questions in weaknesses section. Confidence: 4 Soundness: 3 Presentation: 2 Contribution: 3 Limitations: I think the authors’ treatment of limitations covers most bases I would consider. Flag For Ethics Review: ['No ethics review needed.'] Rating: 7 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We thank the reviewer for providing feedback on our paper and suggesting improvements to clarify the presentation. In the revision, we will do our best to improve readability of the paper and better motivate residual queries in Section 2.1. Regarding why there might be multiple measurements of the same residual, please see the global response. --- Rebuttal Comment 1.1: Comment: I'm glad to hear the paper motivation will be revised. The examples of repeated measurements of a single residual in the global response clarified things for me.	Summary: The paper introduces ReM method for efficiently reconstructing answers to marginal queries with differential privacy. This aim is to minimize the error and allow scalability to high-dimensional datasets. As an extension, this paper also proposes ReM-LNN which ensures that the reconstructed marginals are non-negative, which further reduces error. The effectiveness of these methods is demonstrated by comparing them with existing private query-answering mechanisms like ResidualPlanner and MWEM. Strengths: This paper is well-written. The comparisons with prior works are adequately addressed. The proposed method is novel and practical. Weaknesses: It is not clear how to apply the reconstructed marginal queries in practical settings. For example, Private-PGM is able to generate synthetic datasets for further downstream ML tasks. However, the proposed method is not able to generate synthetic datasets. Technical Quality: 3 Clarity: 3 Questions for Authors: The proposed method claims to be scalable. I am wondering how the time complexity of the dual ascent algorithm used for solving (1) compares with other methods, especially how it scales with the dataset. Confidence: 3 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: The limitations are adequately addressed. Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We thank the reviewer for providing feedback on our paper and suggesting improvements to clarify the presentation. Below, we address various concerns raised in the review. Private query answering is valuable in its own right outside of its application for synthetic data generation. Much work in differential privacy has focused on minimizing error for query answering while satisfying differential privacy. For example, for a fixed privacy budget and target workload of queries, matrix mechanisms minimize error --- often $L_2$ error --- with respect to the output answers. The original matrix mechanism [0] can answer arbitrary linear queries but is inefficient, HDMM [1] restricts to highly structured linear queries such as marginals and range queries and is more scalable than the original matrix mechanism, and ResidualPlanner [2] efficiently answers marginal queries. ReM is a generalization of the reconstruction step of ResidualPlanner. Thank you for your questions about time complexity. Please see the global response for time complexity results, which we will add to the paper. Since the submission of the paper, we have been working to improve the dual ascent algorithm in practice by incorporating line search over the step size $t$ in Line 6 of Algorithm 4. This reduces the number of rounds of dual ascent until convergence. [0] Li, Chao, et al. "The matrix mechanism: optimizing linear counting queries under differential privacy." The VLDB journal 24 (2015): 757-781 [1] McKenna, Ryan, et al. "Optimizing Error of High-Dimensional Statistical Queries Under Differential Privacy." Journal of Privacy and Confidentiality 13.1 (2023). [2] Xiao, Yingtai, et al. "An optimal and scalable matrix mechanism for noisy marginals under convex loss functions." Advances in Neural Information Processing Systems 36 (2024). --- Rebuttal Comment 1.1: Comment: Thank you for addressing my question. My score remains unchanged.	Rebuttal 1: Rebuttal: Below, we address two concerns raised in the reviews: time complexity and natural cases where multiple residuals are measured. Regarding time complexity, let $\gamma, \tau$ be tuples of attributes and suppose each attribute has domain size $m$. Then an answer to marginal $M_\gamma$ has size $m^{\|\gamma\|}$. Reconstructing an answer to marginal $M_\gamma$ from a set of residuals using GReM is $\mathcal{O}(\| \gamma \| \cdot m^{\|\gamma\| + 1} \cdot 2^{\|\gamma\| - 1})$. This running time is subquadratic with respect to output size $m^{\|\gamma\|}$. This can be verified by observing that $\Big( \frac{\| \gamma \| m^{\|\gamma\| + 1} 2^{\|\gamma\| - 1}}{m^{2\|\gamma\|}} \Big) \rightarrow 0$ as either $m$ or $\|\gamma\|$ grow arbitrarily large. As a pre-processing step, we only store one combined ``measurement'' per residual in the downward closure of the workload. After this pre-processing step, the input size is no more than the output size. Therefore, the reported running time bounds are in terms of the output size only. For the dual ascent algorithm in GReM-LNN, one round has time complexity $\mathcal{O}(\| \mathcal{W} \| \cdot \| \gamma^* \| \cdot m^{\|\gamma^\| + 1} \cdot 2^{\|\gamma^\| - 1} )$ where $\mathcal{W}$ is the workload of marginals to answer and $\gamma^*$ is the largest tuple of attributes in $\mathcal{W}$. This running time is linear with respect to the number of marginals in the marginal workload $\mathcal{W}$ and subquadratic with respect to the size of the largest reconstructed marginal $M_\gamma$. Consider the workload of all $k$-way marginals over a data domain with $d$ attributes. Reconstructing all marginals in this workload using GReM is $\mathcal{O}(k d^k m^{k+1} 2^{k-1})$. Since $k$ appears only as an exponent or multiplicative term, reconstruction is not exponential in any size parameter other than the size of the marginals, for which we expect exponential scaling. Similarly, for GReM-LNN, each round of dual ascent is $\mathcal{O}(k d^k m^{k+1} 2^{k-1})$. In many use cases, only 3-way or smaller marginals are measured. In the revision, we will include these time complexity results in the main text as well as proofs in an Appendix. Regarding multiple measurements of the same residual, there are several reasons we might encounter this. First, some state of the art mechanisms regularly measure the same query multiple times. A good example is AIM, which has a budget annealing process that allows it to measure queries more accurately in later rounds. Because of this it will often take repeated measurements of the same marginal at decreasing noise levels. Another reason for measuring a residual query multiple times is the conversion from marginal measurements to residual measurements. Synthetic data mechanisms like AIM and PrivMRF select marginals to measure with isotropic Gaussian noise. In Theorem 2, we show how to convert such a marginal measurement into an equivalent set of residual measurements. When using this conversion within such a mechanism that selects marginals, a residual will be measured each time a marginal that contains it is selected. We will include this motivation earlier in the paper to better understand the problem setting.	NeurIPS_2024_submissions_huggingface	2,024	null	null	null	null	null	null	null	null
DisenGCD: A Meta Multigraph-assisted Disentangled Graph Learning Framework for Cognitive Diagnosis	Accept (poster)	Summary: The paper studies cognitive diagnosis based on graph learning. The proposed method utilizes multiple graphs to assist the representation learning. Specifically, the authors further disentangle another two graphs from the most comprehensive student-exercise-concept interaction graph by removing certain node and edge types. Therefore, three kinds of disentangled representations are learned based on the three graphs. Lastly, a diagnostic function fuses the three representations for final predictions. Strengths: 1. The problem that the paper aims to solve is well formulated and it is of practical value. 2. The authors give very detailed illustrations of the methodology part. 3. Comprehensive experiments are conducted to demonstrate the effectiveness of the proposed method. Weaknesses: 1. Since the proposed method learns the graph representations based on disentangled graphs, the authors are encouraged to add a section about disentangled graph representation learning in the related work part to summarize the recent works on this problem and illustrate the relation of this work with previous ones. 2. A more detailed introduction to incorporated datasets should be given. Specifically, both accuracy and RMSE are employed as evaluation metrics, the authors should provide a clearer illustration of the target tasks of those datasets. 3. The authors should include more mode variants to demonstrate the effectiveness of the added graphs (e.g., w/o G_R, w/o G_D, and w/o both). 4. GAT is used in this work to learn exercise and concept representations. The authors are encouraged to introduce the reason for this choice or conduct experiments based on other graph model backbones. 5. There are many typos in the paper, e.g., line 16 and line 254. The authors should check the grammar carefully. Technical Quality: 3 Clarity: 2 Questions for Authors: Please refer to the weakness part. Confidence: 4 Soundness: 3 Presentation: 2 Contribution: 1 Limitations: N/A Flag For Ethics Review: ['No ethics review needed.'] Rating: 5 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: ## Response to Q.1 The related work on disentangling graph representation learning is as follows, and we will add this to the revised paper: 1. Learning potential representations of disentangling in complex graph networks to achieve model robustness and interpretability has been a hot topic in recent years. Researchers have put forward a lot of classic neural network disentanglement approaches (e.g., DisenGCN[R1], DisenHAN[R2], DGCF[R3], DCCF[R4], DcRec[R5], etc.) to address this challenge. Some representative approaches will be briefly introduced. 2. For example, in DisenHAN[R1], the authors utilized disentangled representation learning to account for the influence of each factor in an item. They achieved this by mapping the representation into different spatial dimensions and aggregating item information from various edge types within the graph neural network to extract features from different aspects； In DCCF[R4], the authors introduced global intent disentanglement into graph contrastive learning, extracting more fine-grained latent factors from self-supervised signals to enhance model robustness； DcRec[R5] disentangles the network into a user-item domain and a user-user social domain, generating two views through data augmentation, and ultimately obtaining a more robust representation via contrastive learning. 3. Despite many approaches suggested, they were primarily applied to bipartite graphs to learn different representations from different perspectives, for learning more comprehensive representations. While this paper aims to leverage disentanglement learning to mitigate the influence of the interaction noise in the interaction graph, and thus we proposed a meta multigraph-assisted disentangled graph cognitive diagnostic framework to learn three types of representations on three disentangled graphs. By doing so, the influence of the noise on exercise and concept learning can be well alleviated. ``` [R1]Jianxin Ma, et al., Disentangled Graph Convolutional Networks. [R2]Yifan Wang, et al., DisenHAN: Disentangled Heterogeneous Graph Attention Network for Recommendation. [R3]Xiang Wang, et al., Disentangled Graph Collaborative Filtering. [R4]Xubin Ren, et al., Disentangled Contrastive Collaborative Filtering. [R5]Jiahao Wu, et al., Disentangled Contrastive Learning for Social Recommendation ``` ## Response to Q.2 We would like to answer your question from the following three aspects: - In the Appendix, we indeed provided a brief introduction to the three datasets (ASSISTments, Math, SLP). However, this introduction may not be detailed, thus we will give more information about the datasets in the revised paper, including their collectors, the contained subjects, and many other attributes. - The cognitive diagnosis task is to assess the true knowledge level of students. To achieve this, this task is generally solved by modeling the students' answer prediction task, which is a binary classification task. Therefore, we use the evaluation metrics of classification to measure the diagnosis accuracy. - The utilized metrics include ACC, RMSE, and AUC, which are widely used for the classification tasks. The three metrics are also widely adopted in many previous CD approaches. ## Response to Weak.3 As you suggested, we added three variants of the DisenGCD to show the effectiveness of the added graphs, which are as - DisenGCD(w/o both) is same as DisenGCD but learns exercise and concept representations directly through native embedding modules(NEMs) . - DisenGCD(w/o $\mathcal{G_R}$) is same as DisenGCD but its exercise representation is learned through a NEM. - DisenGCD(w/o $\mathcal{G_D}$) is same as DisenGCD but its concept representation is learned through a NEM. Table II in *global.pdf* summarizes the results of all variants and the DisenGCD on the ASSISTments dataset. As can be seen, the DisenGCD outperforms both DisenGCD(w/o $\mathcal{G_D}$) and DisenGCD(w/o $\mathcal{G_R}$), which indicates the effectiveness of learning exercise/concept representations by the GAT on added graphs. Moreover, the performance leading of DisenGCD(w/o $\mathcal{G_D}$) and DisenGCD(w/o $\mathcal{G_R}$) over DisenGCD(w/o both) can further validate the effectiveness of the added graphs. In addition, the effectiveness of the DisenGCD can be indirectly validated through the comparisons of DisenGCD(w/o both) and DisenGCD(I), as well as DisenGCD(w/o $\mathcal{G_D}$) and DisenGCD(Is+Rec). In summary, the above comparison validates the effectiveness of the added graphs and the proposed disentangling learning framework. ## Response to Q.4 As you suggested, other three types ofGNNs (i.e., GCN, GraphSage, and HAN) are used as the graph model backbones to replace the GAT module in DisenGCD. The variants of DisenGCD, utilizing GCN, GraphSage, and HAN as the backbones, are denoted as DisenGCD(GCN), DisenGCD(GraphSage), and DisenGCD(HAN). To observe their influence on the proposed approach, these three variants were executed on the Math dataset, whereas the other two datasets were not used due to the time limit. As a result, Table I:Upper in *global.pdf* summarizes the results of these variants and the DisenGCD. We can find that the variant DisenGCD(GraphSage) can obtain competitive performance than state-of-the-art approaches, while DisenGCD(GCN) and DisenGCD(HAN) cannot. Despite its competitive performance, DisenGCD(GraphSage) is still worse than DisenGCD, which indicates that the GAT is most suitable for the proposed framework. In summary, the results show the GAT is a suitable and optimal choice among these four GNNs for the proposed DisenGCD, and it is reasonable and effective for the proposed DisenGCD to adopt the GAT to learn the representations. ## Response to Q.5 Thanks for pointing out the topos. We will correct these typos and double-check the paper. ## We appreciate your valuable feedback and will revise the paper based on the above. --- Rebuttal Comment 1.1: Comment: Thanks for the detailed response from the authors. After reading the rebuttal, I decide to raise my score to 5. --- Rebuttal 2: Comment: Thank you for your positive feedback. We greately appreciate your recognition of our efforts to address your concerns.	Summary: In this paper, the authors introduce a meta multigraph assisted disentangled graph cognitive diagnosis model. Its main contribution is to propose a disentangled graph framework, which disentangles the student-exercise-concept dependency graph into an exercise-concept interaction graph and a concept-concept relationship graph to improve the robustness of representation modeling. Additionally, the concept of meta multigraph is introduced in the student-exercise-concept dependency graph to assist in better modeling of student representation. DisenGCD has demonstrated good performance compared to previous models in multiple datasets, and has shown good robustness in the presence of noise interference or sparse datasets. Finally, the authors also discussed some limitations of DisenGCD in terms of computational complexity and task transferability. Strengths: 1. The paper is well-organized and clearly written, making it accessible to readers. Besides, the authors also present mathematical formulations, detailed explanations of the proposed models. 2. The authors focus on improving the robustness of the CD model and provide a detailed analysis of some problems found in previous graph cognitive diagnosis works. 3. The experiments are very thorough. The authors demonstrate the robustness of the model through decoupling experiments, sparse dataset experiments, and noisy dataset experiments Weaknesses: 1. Although DisenGCD has demonstrated excellent performance across multiple datasets, it still appears to have the drawback of high computational complexity, which may limit its application in practical scenarios. 2. Although the model has shown improved performance, its interpretability may be lacking. In real-world applications within the education field, transparent and interpretable models are more likely to gain acceptance and trust. However, the paper does not address methods for enhancing the interpretability of the model. 3. The paper lacks detailed discussion on the challenges and potential issues that may arise when applying the model in practical educational settings. For instance, it does not address the strategies for handling data privacy and security concerns, which are critical considerations in the implementation of such models. Technical Quality: 3 Clarity: 3 Questions for Authors: 1. KaNCD seems to have been found to work equally well in some papers. Can you compare DisenGCD and KaNCD? 2. The paper utilizes graph attention networks (GAT) to learn representations of exercises and concepts. However, it would be valuable to explore and compare alternative graph neural network (GNN) models, such as graph convolutional networks (GCN) or LightGCN, to assess their performance in the same context. 3. How are the multiple learnable propagation paths in the meta-multigraph learning module designed and selected? Have you tried different numbers or types of propagation paths and how well have they worked? Confidence: 4 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: \ Flag For Ethics Review: ['No ethics review needed.'] Rating: 7 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: ## Response to Q.1 As you suggested, we added KaNCD [R1] to be compared with the proposed DisenGCD, and the comparison results are summarized in Table I:Upper in *global.pdf, where only the Math dataset is used because there is no enough time to validate on other two larger datasets. As you can see, in addition to KaNCD, we have also compared our approach with SCD [R2] and KSCD [R3], which are other two recently published CDMs. As shown in Table I:Upper* in *global.pdf, the proposed DisenGCD still performs better than KaNCD and other two state-of-the-art CDMs, further validating the effectiveness of the proposed DisenGCD. In the future revised paper, we will add these CDMs to the comparison on all three datasets. ``` [R1] F. Wang, Q. Liu, E. Chen, Z. Huang, Y. Yin, S. Wang, and Y. Su, “Neuralcd: a general framework for cognitive diagnosis,” IEEE Transactions on Knowledge and Data Engineering, 2022. [R2] Wang S, Zeng Z, Yang X, et al. Self-supervised graph learning for long-tailed cognitive diagnosis[C]//Proceedings of the AAAI conference on artificial intelligence. 2023, 37(1): 110-118. [R3] H. Ma, M. Li, L. Wu, H. Zhang, Y. Cao, X. Zhang, and X. Zhao, “Knowledge-sensed cognitive diagnosis for intelligent education platforms,” in Proceedings of the 31st ACM International Conference on Information & Knowledge Management, 2022, pp. 1451–1460. ``` ## Response to Q.2 To solve your concern, three GNNs (GCN, GraphSage, and HAN) are used to replace the GAT in the proposed DisenGCD, which are termed DisenGCD(GCN), DisenGCD(GraphSage), and DisenGCD(HAN), respectively. Then, we compared the three approaches with the proposed DisenGCD on the Math dataset, where the other two datasets were not used due to the time limit. The comparison of them is presented in Table I:Upper* in *global.pdf. As shown in Table I:Upper, the GCN-based DisenGCD achieves the worst performance, followed by DisenGCD(HAN). The GraphSage-based DisenGCD holds a competitive performance to yet is still worse than the proposed DisenGCD (based on GAT). The above results show the GAT is a suitable and optimal choice among these four GNNs for the proposed DisenGCD, but we think the GAT may be not the best choice because there exist other types of GNNs that can be integrated with DisenGCD, showing better performance. In summary, it is reasonable and effective for the proposed DisenGCD to adopt the GAT to learn the representations. ## Response to Q.3 The design of the learnable propagation paths is inspired by the success of meta graph. And the propagation paths are selected by the adopted routing strategy, which learns a value for each candidate propagation path. Afterward, these propagation paths that satisfy the threshold will be automatically kept. Therefore, the selection of propagation paths is achieved by the model learning. As for the influence of the propagation path number on the proposed approach, it can be found in line 179 that the path number is determined by the number of hyper nodes (i.e., P). Therefore, this influence can be investigated by analyzing the influence of the number of hyper nodes (i.e., P). In fact, we have investigated the influence of P* on the proposed DisenGCD in *Appendix B.4. As shown in Figures 6 and 7 in Appendix B.4, the proposed DisenGCD is not very sensitive to the number of hyper nodes, where DisenGCD obtains a promising performance when P* is 5. That indirectly indicates that the proposed DisenGCD is not sensitive to the number of candidate propagation paths. In addition, for more convincing, the influence of different P on the robustness of the proposed DisenGCD is also analyzed as shown in Fig.1(a)&(b) in *global.pdf, where only the Math dataset is used due to the time limit. Fig.1(a)&(b)* present the results of RCD and DisenGCD with different P obtained under different ratios of noise interactions (1%, 30%, and 50%). As can be seen, the robustness of DisenGCD is a bit sensitive to different P, and the DisenGCD achieves the most promising robustness when P is equal to 5. That indirectly indicates that the proposed DisenGCD is sensitive to the number of candidate propagation paths. In summary, the performance of the proposed DisenGCD is not sensitive to the propagation path number while its robustness is sensitive. Nevertheless, the proposed DisenGCD can obtain promising performance and robustness when P is set to 5. ## We appreciate your valuable feedback and will revise the paper based on the above. --- Rebuttal Comment 1.1: Comment: The author has addressed some of my concerns, and I am inclined to raise my score. However, some issues still need further improvement in the final version. --- Rebuttal 2: Comment: Thank you for your quick feedback and for contributing to our improved score. We greatly appreciate your efforts. We apologize that some issues still need improvement, and we would like to resolve them as follows. ## Response to Weakness 1 To investigate the computational efficiency of the proposed approach, we have compared it with RCD on ASSISTments and Math datasets in terms of model inference time and training time. Table III in *global.pdf* presents the overall comparison, where the time of the proposed DisenGCD under different hyperparameters P is also reported. As can be seen, DisenGCD's inference time is better than RCD's. This indicates that the proposed DisenGCD is more efficient than RCD, further signifying that it is promising to extend DisenGCD to dynamic CD. It can be seen from Table III:Lower: although a larger P will make DisenGCD take more time to train the model, the proposed DisenGCD achieves the best performance on both two datasets when P=5 and its runtime does not increase too much, which is much better than RCD. In summary, the above comparison shows the model efficiency superiority of the proposed DisenGCD. ## Response to Weakness 2 We have to argue that the diagnostic function in the proposed DisenGCD is interpretable. Firstly, $\mathbf{h}\_{si} = F_{si}(\overline{\mathbf{S}_i}+ \overline{\mathbf{C}_k})$ denotes the combination of the learned student representation and the learned concept representations, which is similar to the combination in RCD, i.e., $Concat(\mathbf{S}\_i, \mathbf{C}_k)$. As a result, $\mathbf{h}\_{si}$ can be seen as the student's mastery of each knowledge concept. Secondly $\mathbf{h}\_{ej} = F\_{ej}(\overline{\mathbf{E}_j}+\overline{\mathbf{C}_k})$ denotes the combination of the learned exercise representation and the learned concept representations, which is also similar to the combination in RCD for exercises, i.e., $Concat(\mathbf{E}\_j, \mathbf{C}_k)$. Therefore, $\mathbf{h}\_{ej}$ can represent the exercise difficulty of each concept. Thirdly, $\mathbf{h}\_{simi} = \sigma(F\_{simi}(\mathbf{h}\_{si}\cdot\mathbf{h}\_{ej} ))$ is mainly used to measure the similarity between $\mathbf{h}\_{si}$ and $\mathbf{h}\_{ej}$ via a dot-product followed by an FC layer and a Sigmoid function. Therefore, a higher value in each bit of $\mathbf{h}\_{simi}$ represents a higher mastery on each concept, further indicating a higher probability of answering the related exercises. Finally, $\hat{r\_{ij}} = (\sum Q_k\cdot \mathbf{h}_{simi})/\sum Q_k$ has a similar idea to NCD to compute the overall mastery averaged over all concepts contained in exercise $e_j$. In summary, the proposed model is well-interpretable, as agreed by Reviewer 2m1M. ## Response to Weakness 3 We have to admit that data privacy and security are indeed a crucial problem in intelligent education. However, existing cognitive diagnosis models did not explore this. To the best of our knowledge, we think the following two potential techniques can be applied to CD to solve data privacy and security issues. 1. The first is to apply the differential privacy technique. Differential privacy [R1] protects individual data by adding noise, ensuring personal information can’t be easily inferred from data analysis. Differential privacy is widely used in fields like computer vision [R2], recommender systems [R3,R4], and many others [R5]. Among these, the recommendation task is most related to cognitive diagnosis (CD), and thus the differential privacy can be easily extended to CD by summarizing the experiences of these approaches. A successful example of extending techniques in the recommendation to CD is the MF [R6] model. 2. The second is to apply the Federated Learning technique. It trains a model across multiple devices without sharing data, improving privacy and efficiency. It has succeeded in many domains, including recommender systems [R7]. According to its successful experiences in the recommendation, federated learning can also be extended to CD. Moreover, a few studies [R8] have applied federated learning to knowledge tracing (KT), making good progress. KT is the most related task to CD in intelligent education; thus, their experiences can be brought to CD to solve data privacy and security problems. To sum up, the proposed DisenGCD and existing CDMs could be combined with differential privacy and federated learning to solve the data privacy and security problem in intelligent education. We will discuss this problem in revised paper. ``` [R1] Dwork, Differential privacy. [R2] Mixed differential privacy in computer vision. [R3] A differential privacy framework for matrix factorization recommender systems. [R4] Applying differential privacy to matrix factorization. [R5] Differential privacy: A survey of results. [R6] Matrix factorization techniques for recommender systems. [R7] Federated recommendation systems. [R8] Federated deep knowledge tracing. ```	Summary: This paper introduces DisenGCD, a new framework for cognitive diagnosis in educational contexts. The authors make several contributions: 1) They propose a disentangled graph learning approach, separating the typically unified graph into three distinct graphs: student-exercise-concept interactions, exercise-concept relations, and concept dependencies. 2) They develop a meta multigraph module for learning student representations, which innovatively allows access to lower-order exercise representations. 3) They employ GAT-based modules to learn exercise and concept representations on the disentangled graphs. 4) They design a new diagnostic function to effectively combine the learned representations from the different graphs. The authors evaluate DisenGCD on multiple datasets, demonstrating performance improvements and enhanced robustness compared to state-of-the-art methods in cognitive diagnosis. Strengths: The paper presents several notable strengths: 1) The proposed disentangled graph learning approach is a novel contribution to the field of cognitive diagnosis. By separating different types of information (student-exercise-concept interactions, exercise-concept relations, and concept dependencies) into distinct graphs, the authors have created a more robust model that appears less susceptible to noise in the data. This is a significant advancement over existing unified graph approaches. 2) The meta multigraph module introduces an innovative mechanism for student representation learning. By allowing access to lower-order exercise representations, this module potentially captures more nuanced relationships in the data, which could be particularly valuable in educational contexts where student knowledge evolves over time. 3) The empirical evaluation is comprehensive and well-executed. The authors conduct experiments on multiple datasets (ASSISTments, Math, and SLP), demonstrating consistent improvements in performance and robustness over state-of-the-art methods. This thorough evaluation strengthens the paper's claims and increases confidence in the generalizability of the proposed method. 4) The paper includes detailed ablation studies and analyses that effectively validate the different components of DisenGCD. These studies provide valuable insights into the contribution of each component and help justify the design choices made by the authors. Weaknesses: 1) While the disentangled graph approach is innovative, it may lose some valuable cross-entity interactions. The separation of student-exercise interactions from exercise-concept relations could potentially limit the model's ability to capture complex, multi-hop relationships that might exist in a unified graph structure. 2) The meta multigraph module, although showing improvements, adds considerable complexity to the model. It's not clear if this added complexity is always justified, particularly for simpler datasets or scenarios with limited noise. A more thorough analysis of the trade-off between model complexity and performance gains would strengthen the paper. 3) The paper lacks a comprehensive discussion on how DisenGCD handles cold-start problems, especially for new students or exercises that don't have established interactions in the graph structure. This is a common challenge in educational settings and needs to be discussed. Technical Quality: 3 Clarity: 3 Questions for Authors: 1) How does DisenGCD adapt to dynamic changes in the knowledge structure, such as the introduction of new concepts or the discovery of new relationships between existing concepts? Does the disentangled structure present any challenges in updating the model under these scenarios? 2) In Section 4.1, you describe the meta multigraph as containing P hyper-nodes, and in your experiments you set P to 5. How did you determine this value? Have you explored the impact of different values of P on the model's performance and computational efficiency? 3) While your experiments demonstrate improved robustness to interaction noise, how does DisenGCD perform when noise is introduced in the concept dependency graph or exercise-concept relation graph? This could simulate errors in curriculum design or expert knowledge. 4) Have you investigated whether the disentangled approach leads to any loss of performance in scenarios where complex, cross-entity relationships are crucial for accurate diagnosis? Are there cases where a unified graph approach might outperform DisenGCD? Confidence: 3 Soundness: 3 Presentation: 3 Contribution: 2 Limitations: N/A Flag For Ethics Review: ['No ethics review needed.'] Rating: 5 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: ## Response to Q.1 We must admit that the design of the proposed DisenGCD did not consider the scenery of dynamic changes in knowledge structures. However, the proposed DisenGCD and RCD are trained in an inductive manner in graph learning. Therefore, it is feasible for DisenGCD and RCD to handle the scenery of introducing new concepts, but their performance is unknown. To investigate this, we conducted the experiments on the Math dataset due to the time limit. Specifically, interactions and exercises related to 80% of knowledge concepts were not masked during the training, and we tested the model on the remaining interactions and exercises. The comparison results of the DisenGCD and RCD are presented in Table I:Lower in *global.pdf. We can see DisenGCD exhibits poor performance on such this scenery, which is poorer than RCD. That indicates that the proposed DisenGCD cannot adapt to dynamic changes in the knowledge structure. The reason may be that the propagation paths learned in the meta-multigraph are not suitable for newly added knowledge concepts. Since adding 20% of concepts changes the topology of the interaction graph, while different-topology graphs generally need to learn different propagation paths. This fact can be found in Figure 4(a) in the submitted paper, stated in [R1]. In short, the proposed DisenGCD may not adapt to the scenery of introducing new concepts well due to the utilized learnable meta multigraph. In future work, we would like to explore how to address this issue to handle dynamic changes in knowledge structures. ``` [R1] Chao Li, Hao Xu, and Kun He. Differentiable meta multigraph search with partial message propagation on heterogeneous information networks. ``` ## Response to Q.2 Actually, setting P* to 5 is by trial and error on the DisenGCD. In *Appendix B.4, we have discussed the influence of different P* on the performance of DisenGCD. Figures 6 and 7 plot the results of DisenGCD obtained under different P (4, 5, 6, and 7) on the two datasets. It can be found that DisenGCD can achieve the optimal performance on both two datasets when setting P to 5. However, we did not explore its influence on the computational efficiency. The most intuitive is that a larger P will cause worse computational efficiency. To investigate this, Table III:Lower in *global.pdf* presents the performance and runtimes (for training) of RCD and DisenGCD under different P. As can be seen, it is consistent with the above assumption that a larger P will make DisenGCD take more time to train the model. The proposed DisenGCD achieves the best performance on both two datasets when P=5, and the runtime increase from P=4 to P=5 is not too much and is acceptable. Besides, Fig.1(a)&(b) in *global.pdf* show that the proposed DisenGCD can obtain the most promising robustness when P=5. Therefore, in this paper, we set P to 5 for a comprehensive consideration of model performance, model robustness, and model efficiency. ## Response to Q.3 To figure out the robustness of the proposed DisenGCD to exercise-concept noise, we conducted experiments on the Math dataset, where 1\%, 10\%, 20\%, and 30\% of exercise-concept relations are randomly removed as or added as noise relations, respectively. (Note that larger dataset) Fig. 1(c)&(d) in *global.pdf* compares the performance of RCD and DisenGCD on Math under different ratios of exercise-concept noise. As can be seen, it is unexpected that the proposed DisenGCD exhibits better robustness than RCD to the added exercise-concept noises. However, different from against interaction noise (as shown in Figure 3(a) in the submitted manuscript), the robustness superiority/leading of DisenGCD over RCD does not increase as the ratio of noise increases, where the leading performance becomes smaller when a larger portion of noises are added. It is reasonable because the added exercise-concept noise affects the learning on two graphs (the interaction graph and the relation graph), which further affects the learning of student and exercise representations. Since the learning of the concept representation is not influenced, DisenGCD still outperforms RCD. However, when more noise is added, DisenGCD's robustness against exercise-concept noise will not be better than that of DisenGCD against interaction noise, where the learning of two graphs is affected. ## Response to Q.4 As for the first sub-question, to be honest, we have not investigated the performance of the proposed approach in scenarios where complex, cross-entity relationships. To answer your concern, we have tried our best to check many educational datasets， but we indeed cannot find a suitable dataset that holds cross-entity relations. Therefore, we are sorry that we cannot answer your question through experiments. However, as far as we are concerned, the DisenGCD can extend to handle cross-entity relations by disentangling more graphs. That may make promising performance, but its complexity and efficiency will be criticized. We also would like to answer your question if you can recommend a specific dataset. As for the second, we would like to state that the compared RCD approach is essentially a unified graph approach, where all representations are learned and exchanged in an implicit graph. As shown in the submitted manuscript, the proposed DisenGCD exhibits better performance than the RCD. In addition, we have also compared DisenGCD with another unified graph-based approach, i.e., SCD [R1], on the Math dataset, where the results are presented in Table I:Upper in *global.pdf*. We can find that the DisenGCD still outperforms SCD. In summary, there may exist other novel unified graph approaches outperforming DisenGCD, but existing unified graph-based CDMs (RCD and SCD) cannot. ``` [R1] Wang S, Zeng Z, Yang X, et al. Self-supervised graph learning for long-tailed cognitive diagnosis[C]. ``` --- Rebuttal Comment 1.1: Comment: In general, I am satisfied with the answers. I also think the paper studies an interesting domain where GNNs are applied. I will keep my score, but the paper seems to be above the acceptance bar. --- Reply to Comment 1.1.1: Comment: Many thanks for taking the time to carefully read the rebuttal. We appreciate your recognition of the value of our work! We are pleased to address your concerns and will revise the paper according to your suggestions.	Summary: The paper presents DisenGCD, a novel cognitive diagnosis model designed to enhance the robustness and accuracy of student, exercise, and concept representations by leveraging a meta multigraph-assisted disentangled graph learning framework. DisenGCD constructs and disentangles three specific graphs for interactions, relations, and dependencies, thus mitigating the negative impact of noise in students’ interactions. The framework incorporates a meta multigraph learning module and GAT to improve the learning of latent representations, demonstrating superior performance and robustness compared to state-of-the-art CDMs on multiple datasets. Strengths: 1. The use of a meta multigraph-assisted disentangled graph learning framework is a novel approach that significantly enhances the robustness of cognitive diagnosis models. 2. DisenGCD shows better performance in terms of AUC, accuracy, and RMSE compared to other state-of-the-art CDMs. 3. The model effectively handles noise in student interactions, ensuring that the learning process remains accurate and reliable. 4. The experiments cover multiple datasets and different data splits, providing a thorough validation of the model’s effectiveness. 5. The diagnostic function of DisenGCD maintains high interpretability, comparable to traditional models like NCD, IRT, and MIRT. Weaknesses: 1. The implementation of the meta multigraph learning module and the disentangled graph learning framework could be complex and resource-intensive. 2. The paper does not report error bars or other statistical significance measures, which are important for validating experimental results. 3. The comparison does not include some recent models like SCD, which could provide a more comprehensive evaluation. 4. While some limitations are discussed, a more detailed analysis of potential weaknesses and future improvements could enhance the paper’s transparency and reliability. Technical Quality: 3 Clarity: 3 Questions for Authors: 1. Can the model be adapted to handle real-time data for dynamic cognitive diagnosis? 2. How does the choice of hyperparameters affect the model’s performance and robustness? 3. Can the disentangled graph learning framework be integrated with other neural network architectures for further performance enhancement? Confidence: 1 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: NA Flag For Ethics Review: ['No ethics review needed.'] Rating: 5 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: ## Response to Weak.1 & Q.1 We admit that the proposed DisenGCD seems complex and needs more resources for its implementation because more graphs need to be handled. In future work, we would like to design a more efficient paradigm to learn robust representation in a unified graph. As for Question 1, to be honest, we have to admit that the proposed DisenGCD is not designed for real-time dynamic cognitive diagnosis. However, as shown in some research [R1,R2], the graph-based CD approaches could be easily extended to knowledge tracing (KT, a task similar to dynamic cognitive diagnosis): for example, RKT [R1] is actually a variant of RCD in knowledge tracing, where the representations learned by graph NNs are attached as static exercise features used for sequence prediction. Due to the time limit and our limited coding capabilities， we are sorry that we failed to extend our approach to dynamic cognitive diagnosis. However, to extend graph CD approaches to KT or dynamic CD, a crucial problem to be solved is the graph-based model's efficiency. Therefore, to solve your concern indirectly, we make the statistics about the proposed DisenGCD's efficiency regarding its inference time. To this end, the inference time (seconds) of DisenGCD and RCD are compared in Table III:Upper in *global.pdf. As can be seen, the inference time of DisenGCD is better than RCD. That indicates that the proposed DisenGCD is more efficient than RCD, further signifying that it is promising to extend DisenGCD to dynamic CD. ``` [R1] Pandey, S., & Srivastava, J. (2020, October). RKT: relation-aware self-attention for knowledge tracing. In Proceedings of the 29th ACM international conference on information & knowledge management (pp. 1205-1214). [R2] Yang, Y., Shen, J., Qu, Y., Liu, Y., Wang, K., Zhu, Y., ... & Yu, Y. (2021). GIKT: a graph-based interaction model for knowledge tracing. In Machine learning and knowledge discovery in databases. ``` ## Response to Q.2 The most important hyperparameter is the number of hyper nodes (i.e., P) in the meta-multigraph. To solve your concern, we would like to explore its influence from two aspects* (i.e., on performance and robustness): On the one hand, its influence on model performance has been discussed in *Appendix B.4. As shown in Figures 6 and 7 in Appendix B.4, the proposed DisenGCD can achieve promising performance when the number of hyper nodes is 5, indicating setting P* to 5 is reasonable. *On the other hand, we also analyzed the influence of different P* (set to 4, 5, 6, and 7 ) on the model robustness under the Math dataset (ASSISTments is not used due to the time limit). Fig.1(a)&(b) in *global.pdf* shows the results of RCD and DisenGCD with different P obtained under different ratios of noise interactions (1\%, 30\%, and 50\%). As you can see, when the number of hyper nodes is equal to 5, the robustness of DisenGCD is the most promising, whose performance under different ratios of noises is balanced better than RCD. In summary, it can be seen that the proposed DisenGCD can achieve good performance and promising robustness when setting the number of hyper nodes to 5. ## Response to Q.3 Other NNs can be integrated with the proposed framework yet the performance of them are not sure. To validate this, we used three NNs (GCN, GraphSage, and HAN) to replace the GAT in the proposed approach, respectively, which are denoted as DisenGCD(GCN), DisenGCD(GraphSage), and DisenGCD(HAN). Due to the time limit, we only compared them with the proposed approach on the Math dataset, and the comparison results are summarized in Table I:Upper in *global.pdf. It can be seen that the GAT enables the proposed framework to hold the best performance than other three variants, where DisenGCD(GraphSage) performs better than the other two. The above comparison indicates that the proposed DisenGCD could be integrated with other neural network architectures and their performance may be improved when suitable NNs are adopted. ## Response to Weak.3 According to your suggestions, we have added more recent CDMs, like SCD [R1], KaNCD [R2], and KSCD [R3] to the comparison. The comparison between them and DisenGCD is presented in Table I:Upper* in *global.pdf, where only the Math dataset is used due to the time limit. As can be seen from Table I:Upper*, compared to these recent CDMs, the proposed DisenGCD still exhibits better performance. We will add these CDMs to the comparison in the revised paper. ``` [R1] Wang S, Zeng Z, Yang X, et al. Self-supervised graph learning for long-tailed cognitive diagnosis[C]//Proceedings of the AAAI conference on artificial intelligence. 2023, 37(1): 110-118. [R2] F. Wang, Q. Liu, E. Chen, Z. Huang, Y. Yin, S. Wang, and Y. Su, “Neuralcd: a general framework for cognitive diagnosis,” IEEE Transactions on Knowledge and Data Engineering, 2022. [R3] H. Ma, M. Li, L. Wu, H. Zhang, Y. Cao, X. Zhang, and X. Zhao, “Knowledge-sensed cognitive diagnosis for intelligent education platforms,” in Proceedings of the 31st ACM International Conference on Information & Knowledge Management, 2022, pp. 1451–1460. ``` ## We appreciate your valuable feedback and will revise the paper based on the above. --- Rebuttal Comment 1.1: Comment: Thanks for the rebuttal. I have read through it and appreciate the response. I will keep my score unchanged. --- Reply to Comment 1.1.1: Comment: Thanks for your response, and we appreciate your recognition of our work.	Rebuttal 1: Rebuttal: We sincerely appreciate the valuable feedback provided by all the reviewers. We have carefully addressed their questions and concerns in our response, aiming to provide satisfactory answers. Here we uploaded a file named "global.pdf" to show some necessary results and comparisons, which contains three tables and one figure. In all subsequent responses to reviewers, this file is termed *global.pdf* for easy presentation and convenient discussion. Pdf: /pdf/0a01a4e3aa5a9b206f4ec8c81a06a4c696f4ff63.pdf	NeurIPS_2024_submissions_huggingface	2,024	null	null	null	null	null	null	null	null
On the Role of Attention Masks and LayerNorm in Transformers	Accept (poster)	Summary: - This paper investigates the role of attention masks and layer normalization (LayerNorm) in mitigating the rank collapse issue in transformer models, and gets following conclusions: - A long as there is a token which all other tokens in the sequence can directly or indirectly attend to over a fixed number of layers, exponential rank collapse is guaranteed. - local attention mask or focus attention can slow down the collapse. - layernorm plays an important role in self-attention dynamics. - This paper provides extensive analysis and numerical experiments on the self attention dynamics, Expressivity and Versatility. Strengths: - The paper provides a rigorous mathematical analysis of the role of attention masks and LayerNorm in transformers, contributing to a deeper understanding of these models' inner workings. - This paper uses a graph-theoretic approach to analyze self-attention, so it is more general and can be extended to sparse attention and causal attention. - This paper shows an interesting counterexample of collapse and shows that self-attention dynamics can process a rich set of sequences stablely. - The paper supports its theoretical findings with numerical experiments. Weaknesses: - Application of the conclusion, I wonder can the findings in this paper guide some design choices in the future transformer models. - And how the findings connect to (or explain) the strong performances of of existing LLMs? It claims that sparse attention can slow down collapse, why most existing LLMs do not use sparse attention? - The numerical experiments are somewhat weak as the results are obtained from only 32 samples. Technical Quality: 3 Clarity: 2 Questions for Authors: I wonder what is the differences of self-attention dynamics between the causal attention and bidirectional attention? I cannot find clear analysis about this in the paper, since different self-attention methods are generalized with graph connections. Confidence: 4 Soundness: 3 Presentation: 2 Contribution: 2 Limitations: None Flag For Ethics Review: ['No ethics review needed.'] Rating: 4 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We thank the reviewer for taking the time to review our paper and providing constructive feedback. In line with the reviewer’s suggestion, we have rerun all the numerical experiments, validating our theoretical findings on 3000 examples. The results are provided in the rebuttal pdf file. In what follows, we provide detailed responses to the rest of the comments raised by the reviewer. > Q1: Application of the conclusion, I wonder can the findings in this paper guide some design choices in the future transformer models. Thank you for the question. Here are some ideas along our theoretical results: - The graph-dependent rate of rank collapse established in the paper can be used as a principle for attention mask design. It is, however, important to also note that the design cannot be solely based on this result. One also needs to consider other implications of the mask, including its effects on universal approximation property [1]. We have discussed this remark in in line 178-184 of the manuscript. - Our result could also further explain the role and importance of LayerNorm in transformers. It clarifies a misconception popularized by a previous work [2] that “layer normalization plays no role” for rank collapse in transformers. We rigorously prove that layer normalization can in principle prevent rank collapse and stabilize token representations in transformers, enabling the model to utilize depth more efficiently. > Q2: How do the findings connect to (or explain) the strong performances of existing LLMs? It claims that sparse attention can slow down collapse, why do most existing LLMs not use sparse attention? - Existing implementations of LLMs utilize LayerNorm, which is in line with our results on the role of LayerNorm in preventing rank collapse. This could justify why this module is preserved in the design of pioneering LLMs and improved variants. - As we discussed in the response to Q1, rank collapse is not the only issue one needs to consider when designing the attention mask. One also needs to consider aspects such as the universal approximation power of masked attention. How to balance the trade-off for more powerful LLMs is a vital research direction which we are currently exploring. - In fact, while many existing LLMs do not use sparse attention, sparse attention is used in practice and is gaining popularity due to the demand for efficiency. For example, sparse attention was populated by Longformer [3] and OpenAI [4] and nowadays LLMs like Mistral 7B use sliding window attention (SWA) [4]. Other popular sparse attention models include, but not limited to BigBird [5], Recurrent Memory Transformers (RMTs) [6,7] and StreamingLLM [8]. - Besides language tasks, sparse attention is also common in vision transformers. Examples of the use can be found in the following papers [9,10,11]. We will update our manuscript for an improved literature review on sparse attention. > Q3: The numerical experiments are somewhat weak as the results are obtained from only 32 samples. Thank you for your constructive suggestion. Our original sample size 32 was chosen according to the experiments in Dong et al. [2]. In line with the reviewer’s suggestion, we validate our theoretical findings on 3000 examples instead. The results are provided in the rebuttal pdf file. They are still similar to the original results; however, as suggested by the reviewer, the larger sample size makes them more reliable. We appreciate your questions and comments very much. Please let us know if there are any further questions. ----------------- References [1] Yun et al. O(n) connections are expressive enough: Universal approximability of sparse transformers. In NeurIPS, 2020. [2] Dong et al. Attention is not all you need: pure attention loses rank doubly exponentially with depth. In ICML, 2020. [3] Beltagy et al. Longformer: The long-document transformer. 2020. [4] Child et al. Generating Long Sequences with Sparse Transformers. 2019. [5] Jiang et al. Mistral 7B. 2023. [6] Zaheer et al. Bigbird: Transformers for longer sequences. In NeurIPS, 2020. [7] Bulatov et al. Recurrent Memory Transformer. In NeurIPS, 2022. [8] Bulatov et al. Scaling Transformer to 1M tokens and beyond with RMT. 2024. [9] Xiao et al. Efficient Streaming Language Models with Attention Sinks. In ICLR, 2024. [10] Liu et al. Swin transformer: Hierarchical vision transformer using shifted windows. In ICCV, 2021. [11] Pan et al. Slide-Transformer: Hierarchical Vision Transformer with Local Self-Attention. In CVPR, 2023. [12] Hassani et al. Neighborhood Attention Transformer. In CVPR, 2023.	Summary: This paper theoretically studies the role of attention masks and layer norm in the convergence to the rank collapse degeneracy in Transformers, which are two architectural components that have previously been overlooked in studying the rank collapse phenomenon. The authors first define the problem through the lens of graph theory. They then show that self-attention with general mask structure leads to rank collapse, but that factors such as the “graph diameter” of the mask (related to locality of attention) and the uniformity of attention (related to the temperature). They continue to show that LayerNorm does not prevent rank collapse for general attention masks with orthogonal value matrices, but that counterexamples exist for other choices of value matrices. Strengths: 1. The paper is mostly well written and clear. 2. The paper provides new insights into the rank collapse phenomenon. In particular, the graph theoretic formulations provides general conditions in terms of attention mask (e.g. giving a quasi-strongly connected graph) that lead to rank collapse. 3. Moreover, the paper shows a more nuanced analysis for the collapse possibilities when a more complete architecture is considered (in particular with layer normalisation) . 4. The paper also shows the impact of hyperparams like attention context length or temperature in causing/preventing rank collapse. Weaknesses: My concerns are presented in order of importance (to me). 1. My main concern is that the theoretical setting considered omits several other important architectural components which are used in practice. In particular skip connections (which are known to be effective in preventing rank collapse properties a lot, e.g. Dong et al 2020, Noci et al 2022 or He et al 2023), or positional encoders. The paper is motivated by the fact that rank collapse papers do not study standard archs with all components included, but the omission of such architectural components (especially skips) also means one could argue the paper does not meet its goal. 2. The experimental verification of the theory seems a bit unconvincing at the moment. For example, the effect of temperature seems to disappear after 5 or 6 layers in the non-sliding window settings, and there do not seem to be error bars in figs 2/3 which makes one wonder if the results are significant (are the error bars just smaller than what is visible?). Moreover, it would be good to consider depths of larger than 10, as modern deep transformers can be up to ~100 layers deep and the effects should be more pronounced at larger depths. Also, there doesn't seem to be a difference between the top row of Figure 2 vs the bottom. Finally, how do you pretrain a SAN (figure 2 bottom left) without skip connections as the rank collapse literature would tell you that such models are not trainable (Noci et al 2022). 3. Somewhat related to my first point, it is not clear in my reading that the results presented concerning the counterexamples with particular value weights provide new insights into the fundamental behaviours of transformers in practice or rather represent theoretical edge cases which are nice to know exist but never occur in practice. Some questions that could address this: do such value weights (or value weights with these properties) appear in trained Transformers, or do these theoretical insights generate new initialisations/parameterisations for value weights to train transformers, given that orthogonal weights will lead to rank collapse even with layernorm (Theorem 2). 4. I would also argue that the intuition for the given counterexamples is not particularly clear at present (though as I say most other parts of the paper are well presented). As I understand it, the mathematical argument is that the "center node" has a neuron that has 0 activation due to the layernorm, and as a result even when attention allows other tokens to see the center node's activations the zero activations are not being transfered to other tokens and this prevents collapse. This can be clearer e.g. in section 4.3.1, but as I say I have reservations about how fundamental/important this insight could be. Technical Quality: 3 Clarity: 3 Questions for Authors: 1. For the results without LayerNorm (e.g. Theorem 1 or figure 2 left plots), it seems like the definition of mu(X) doesn't rule out that simply X converges to 0 in Frobenius norm, so that the activations simply go to 0, as opposed to saying something about the different tokens becoming aligned (in terms of cosine similarity going to 1) which is what some previous works have described as rank collapse. Do different attention masks lead to this trivial case (as opposed to the alternative where X is non-zero and all the tokens are identical non-zero)? 2. Does theorem 2/corollary 1 hold for non-orthogonal value weights? For example other random matrices like iid Gaussian/uniform should all give the same cosine similarity properties as orthogonal in the large-width limit. 3. The LayerNorm without centering on line 128 is exactly RMSNorm https://arxiv.org/abs/1910.07467 which is popularly used in LLMs like LLama or Mistral instead of LayerNorm. 4. Why does this work only provide exponential convergence rates whereas the original rank collapse paper has doubly exponential? Typo: 1. $d_N$ not $d_d$ in line 128. Confidence: 4 Soundness: 3 Presentation: 3 Contribution: 2 Limitations: There is not a clear limitations section but the authors do mention a limitation on line 283. Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We greatly appreciate your positive assessment and insightful comments, which have helped strengthen our work. Below, we provide individual responses the comments you raised. Weaknesses W1 We agree with the reviewer that one should be aware of the effect of skip connections when analyzing rank collapse in transformers. Because we were mainly motivated by analyzing the sole effect of attention masks and LayerNorm, we excluded skip connections in our analysis for a controlled study. Building on our work, future research could include analyzing the effects of skip connections as well as positional encoding, among others. In the direction of the comment, we have conducted preliminary experiments measuring the evolution of $\mu(X)$ in SANs with skip connections alone and with both LayerNorm and skip connections. The results (Fig. 1 & 2 in the rebuttal) offer interesting insights: - For both initialized and pretrained models, adding pure skip connections indeed prevents $\mu$ from converging to 0, confirming existing theory. However, it seems to make $\mu$ unstable, particularly in deeper layers (Fig. 2). Compared with full transformers where $\mu$ stays relatively stable, there is a clear discrepancy. - Incorporating LayerNorm seems to alleviate this stability issue: - In initialized models, LayerNorm alone effectively prevents rank collapse without causing $\mu$ to diverge. - In pretrained models, both LayerNorm and skip connections mitigate rank collapse while LayerNorm is key to maintaining tokens at a scale consistent with full transformers, as we discussed in line 265 of the paper. These findings underscore the complex interplay between different components in transformers. LayerNorm emerges as a crucial element in mitigating rank collapse and stabilizing token representations while also counteracting potential negative effects of pure skip connections. We plan to further explore this in theory in the future. W2 After carefully checking the code, we found that we accidentally loaded the results for initialized models when plotting for the pretrained models. We feel indebted to the reviewer for spotting this mistake. The results for pretrained models in Fig. 2 & 3 are now reported in Fig. 1 & 3 of the rebuttal, respectively. - For the effect of the temperature, the error bars are indeed very narrow for initialized models. One hypothesis for the effect to disappear after 5 or 6 layers at initialization is that in deeper layers as rank collapse happens, tokens align with each other and the temperature naturally has much less effect (especially given random $W_K$ and $W_Q$ cannot account for the increased similarities between tokens). We leave the thorough investigation for future work. As for pretrained models, the effect of temperature now can be observed in all layers and it is significant for all masks. - Due to the space limit, the results for 128 layer initialized models were provided in Appendix G.1. They exhibit the same trend as 12 layer models. We have also included 128 layer versions of Fig. 2 & 3 for initialized models over 3000 samples in the rebuttal. - For pretrained SANs, we follow Dong et al. (their official implementation: https://github.com/twistedcubic/attention-rank-collapse) by taking the existing pretrained models and only considering the self-attention layers. W3 Dong et al. claims that self-attention with LayerNorm cannot prevent rank collapse. These counterexamples are sufficient to show that self-attention with LayerNorm can prevent it from happening. We do not claim that they are the only possible weights that work. It merely establishes that the set of such desirable designs is not empty. Fig. 1 in the rebuttal shows that just adding LayerNorm to SANs (no any other components such as skip connections) prevents rank collapse in initialized models, while in pretrained models LayerNorm mitigates the issue together with other modules and helps token representations maintain a scale consistent to full transformers. This empirically validates our theory. W4 We appreciate this detailed feedback. We would like to clarify a misunderstanding here: - The counterexample works not because of zero activations in the center node (token 1 in this case). Even with zero activations in the center node, rank collapse can still occur: - In Section 4.3.1, when token 2 is initialized on the olive segment in Fig. 1 of the paper, rank collapse still happens. - In the case without LayerNorm, suppose we have the causal mask, and $X_{1,:}^{(0)}$ is a vector with a zero in its elements, and all $W_V^{(t)}$ are identity with $W_K^{(t)}$ and $W_Q^{(t)}$ satisfying A2). Then token 1 would stay at its initial value (so there is a deliberate zero activation from the start) and this would be the case where “attention allows other tokens to see the center node's activations but the zero activations are not being transferred to other tokens” --- yet rank collapse would still happen here (Theorem 1). - The right intuition here can be better elucidated from a dynamical system perspective. Without LayerNorm, rank collapse happens as the center node attracts all tokens to align. In the counterexample, the key insight to prevent token 2 from aligning with token 1 (assuming token 1 has converged to $v_1$) is that token 2's updated representation must cancel token 1's attraction, i.e. $X^{(t)}\_{2,:}W$ needs to have a component negating $v_1$. In the example, $W$ satisfies $Wv_2 = v_1+v_2$ for some $v_2$, so a $-v_2$ component in $X^{(t)}_{2,:}$ can negate token 1's effect ($v_1$). The crucial role of LayerNorm here is to stabilize this process by readjusting token scales to ensure that the cancellation still persists in subsequent updates. This key scaling effect of LayerNorm can also be observed from the newly added experiments, as we discussed in W1. ---- Due to the character limit, we answer the reviewer's questions in a comment below. --- Rebuttal 2: Title: Answers to the questions raised by Reviewer 7spk Comment: Questions > Q1: The definition of mu(X) doesn't rule out that simply X converges to 0 in Frobenius norm, so that the activations simply go to 0, as opposed to saying something about the different tokens becoming aligned (in terms of cosine similarity going to 1) which is what some previous works have described as rank collapse. Do different attention masks lead to this trivial case (as opposed to the alternative where X is non-zero and all the tokens are identical non-zero)? Thank you for the question. - In the case with LayerNorm, X would not converge to 0 and hence the convergence of $\mu(X)$ to zero implies convergence to the same point on the unit sphere. In this case, it is equivalent to the cosine similarity going to one. - Without LayerNorm, whether or not $X$ goes to zero as $\mu$ goes to zero depends on many factors, including the mask, the initial inputs, and the model parameters. > Q2: Does Theorem 2/Corollary 1 hold for non-orthogonal value weights? For example other random matrices like iid Gaussian/uniform should all give the same cosine similarity properties as orthogonal in the large-width limit. This is a good point. We believe that Theorem 2 and Corollary 1 could be extended to more general classes of matrices, such as families of random matrices, as the reviewer mentioned. We leave this for future work. > Q3: The LayerNorm without centering on line 128 is exactly RMSNorm which is popularly used in LLMs like LLama or Mistral instead of LayerNorm. Thank you for the reference! This is a good call. We were aware of RMSNorm but instead chose the name “LayerNorm” following the convention of previous works such as Geshkovski et al. 2023 and Tian et al. 2023 to avoid confusion of terminology. We will add a footnote for this point. > Q4: Why does this work only provide exponential convergence rates whereas the original rank collapse paper has doubly exponential? This is a good question. Our result here is actually tight because in general for linear systems (which is the special case of $W^{(t)}_K=W^{(t)}_Q=0$, $W_V^{(t)}$ being the identity matrix and no LayerNorm), one can never have a doubly exponential convergence (see Antsaklis and Michel, Section 4.8). For this reason, we believe that the results in Dong et al. 2020 are specific to their setting and would not trivially extend to other settings, including ours. We appreciate your questions and comments very much. Please let us know if there are any further questions. ---------------------------------- References Dong et al. Attention is not all you need: pure attention loses rank doubly exponentially with depth. In ICML, 2020. Geshkovski et al. A mathematical perspective on transformers. ArXiv, abs/2312.10794, 2023. Tian et al. Scan and snap: Understanding training dynamics and token composition in 1-layer transformer. In NeurIPS, 2023. Antsaklis and Michel. A Linear Systems Primer. 2000. --- Rebuttal Comment 2.1: Title: Additional Questions Comment: Thank you for the rebuttal and clarifications. I have some additional questions/comments in light of the response: 1. Doesn't the SANs+LN at initialisation plot (in Figure 1 and Figure 2 of additional pdf) contradict Theorem 2, which states that even with LN and with orthgonal initialisation you should obtain rank collapse? How are the QKV weights initialised in the plots? The mechanism outlined in W4 for the counterexample, while interesting, seems to rely on specific properties of the value weights/representation X which won't occur at random initialisation. 2. Is it necessary to have certain neurons in the activations to have zero values in order to construct the counterexample? If not, it feels like an unnecessary detail which complicates the intuitive picture. The intuition of W4 should be included in the main paper to help readers. 3. Regarding W1, a large body of existing work has shown that the combination of Pre-Norm skip connections reduces the effect of the residual branch which mitigates signal propagation degeneracies like rank collapse. Perhaps the main citation for this is https://arxiv.org/abs/2002.10444, but see also https://arxiv.org/abs/2010.12859 https://arxiv.org/abs/2003.04887 or https://arxiv.org/abs/2102.06171 to name a few. In the context of transformers, Noci et al 2022 and also https://arxiv.org/abs/2311.01906 have shown this effect too. I would recommend including the new plots and a discussion of these more practical architectures in the paper. 4. I still maintain that the fact that the definition of mu(X) doesn't separate out the cosine similarities going to 1 vs the activation norms going to 0 as problematic for the understanding of the mechanisms of these different architectural components on rank collapse. Even if theoretically it is not possible to show, I would devise separate metrics to isolate these two effects and produce the equivalent of Figures 1 and 2. --- Rebuttal 3: Title: Answers to the additional questions Comment: We thank the reviewer for the quick response and the additional questions. Below, we provide point-to-point answers: Q1: Thank you for the comment. - QKV are initialized using the $U(-\sqrt{k}, \sqrt{k})$, with $k$=1/input_dim, which is the default initialization used in transformers like BERT implemented in HuggingFace. - Theorem 2 is established under the assumption that the value matrices are orthogonal. It is important to note that while finite random uniform matrices are orthogonal in expectation, each realization is not orthogonal. Hence the phenomenon in experiments does not contradict our theoretical results. - Note that due to the existence of LayerNorm, token trajectories are not a continuous function of model parameters, i.e. the value matrices. To get an idea, notice that if $x_1, x_2$ go to zero, the normalized values $x_1/\|\|x_1\|\|_2$ and $x_2/\|\|x_2\|\|_2$ can have a distance as large as two. As a result, any measure on token trajectories, including $\mu$, is not a continuous function of value matrices. Q2: Thank you for the question. As detailed in the response to W4 with concrete examples, having zero activations in the center node does not play any role in the mechanism of counterexample. If one replaces $W$ in the example with $W_{Z} = Z^{T}WZ$ where $Z$ is an orthogonal matrix, then the trajectory of $X^{(t)}$ of the new dynamics would be $X_Z^{(t)} = X^{(t)}Z$. Notice that under the new dynamics, the first token is going to converge to $(0,1)Z$, which clearly may not have any zero activation. This is a good point. We will add a remark to include a detailed discussion of the right intuition as the one provided to W4 to avoid confusion and misinterpretation. Thank you for bringing this up. Q3: Thank you for the comment and the pointers to the references. We are glad that the reviewer found our newly added experiments meaningful and connected to the literature. We will include the newly added experiments in the updated version of the manuscript and discuss the complex interplay between different architectural components in transformers and the related literature. Q4: Thank you for the comment. Our definition of $\mu$ is standard: it is mathematically equivalent to the definition of the measure $\textbf{res}$ used in Dong et al. (the original rank collapse paper), and the definition adopted in more recent e.g. Geshkovski et al. to study convergence of tokens in transformers. This measure checks for whether all tokens converge to the same representation or not. While it does not give information about whether the common representation is zero or not, in both cases the model loses representation power as it can no longer map tokens within the same sequence to different values. We appreciate your questions and comments very much. Please let us know for any further questions. --- Rebuttal Comment 3.1: Title: Thank you Comment: Thanks for the response. Regarding Q1, it would be good to do the relevant plot at initialisation matching the assumptions of Theorem 2 (e.g. orthogonal weights) just as a verification of the theory. I am surpised that SAN+LN does not converge to rank collapse at initialisation, and I remain unconvinced that the theoretical counterexamples provided explain this because I don't think they should not apply at initialisation. But otherwise thanks again for the clarifications. --- Reply to Comment 3.1.1: Title: Response to the comments Comment: We thank the reviewer for the comment. - To further address the concern regarding the verification of our Theorem 2, we perform an additional set of experiments of SAN+LN under the exact conditions in Theorem 2 with initialization of the value matrices set to be exactly orthogonal. Here is the table summarizing the mean (std) of $\mu$’s in a 128 layer networks under different masks: \| layer \| complete \| causal \| slide-window \| slide-window-uni \| \|-----\|--------------------\|-------------------\|-------------------\|-------------------\| \| 0 \| 274.7604 (2.0498) \| 274.7832 (2.0682) \| 274.6891 (1.9408) \| 274.7870 (2.0094) \| \| 32 \| 4.6109e-5 (2.0682) \| 47.1223 (16.2438) \| 111.6686 (3.4742) \| 161.0730 (2.5183) \| \| 64 \| 4.6347e-5 (2e-6) \| 67.9591 (12.3382) \| 95.4022 (4.8795) \| 160.5059 (3.8947) \| \| 96 \| 4.6396e-5 (2e-6) \| 82.9471 (11.5331) \| 85.6265 (5.5486) \| 164.1741 (5.2969) \| \| 128 \| 4.6220e-5 (2e-6) \| 92.3288 (11.6265) \| 79.9934 (5.6181) \| 165.205 (5.4153) \| In particular, we observe that while $\mu$ converges to zero for complete masks and shows a clear decreasing trend for slide-window masks (both masks are strongly connected), such converging trends do not hold for non-strongly connected masks (causal and sliding-window-uni, as they are only quasi-strongly connected), verifying our Theorem 2 and showing that (1) the convergence rate depends inversely on the graph diameter (2) the tightness of our result that Theorem 2 only applies for strongly connected masks. We will include these sets of experiments to better illustrate our theoretical results. - Finally, we would like to emphasize again that the goal of the counterexample is to establish that self-attention with layer normalization can prevent rank collapse from happening, as opposed to what Dong et al. claims (“layer normalization plays no role”). We do not claim that they are the only possible weights that can work. It merely establishes that the set of such desirable designs exists. We thank the reviewer once again for the discussion and the constructive feedback!	Summary: In this work, the authors investigate the issue of rank collapse in Transformers, providing insights into how attention masks and layer normalization can mitigate this problem. The paper includes extensive analysis, addressing two important questions and offering valuable contributions to the field. Strengths: 1.I like this paper for its strong motivation and very interesting insights. I appreciate the authors for addressing the rank collapse issue in Transformers, a topic often overlooked in the community. 2. The authors provide detailed analysis to demonstrate that attention masks and layer normalization can help address this issue. 3. The experimental results support the authors' analysis effectively. 4. Figure 1 is particularly helpful in understanding the illustration of the effectiveness of layer normalization. Weaknesses: It seems obvious that causal masking and local attention would help mitigate the issue of rank collapse in Transformer/Attention mechanisms compared to full attention. Maybe no need to demonstrate this a lot. Technical Quality: 3 Clarity: 3 Questions for Authors: I like this work; however, I have two additional questions: If local attention alleviates rank collapse, how do the experimental results compare to those of full attention or causal masked attention? Which approach is more effective in addressing the rank collapse issue: local attention or causal attention? Are there any strong insights into this? I assume local attention may not be generally applicable in language tasks but is more suitable for vision tasks. Confidence: 3 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: Please see above. Flag For Ethics Review: ['No ethics review needed.'] Rating: 7 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We appreciate your thoughtful comments and positive assessment of our work. After carefully reviewing your feedback, below we provide answers to the comments you raised. > Q1: It seems obvious that causal masking and local attention would help mitigate the issue of rank collapse in Transformer/Attention mechanisms compared to full attention. Thank you for the comment. While the conclusion might seem intuitive, we would like to point out that formalizing it in theory with rigorous mathematical proofs turns out to be very nontrivial — causal masking has been populated by GPTs for years, yet none of the existing theoretical works on analyzing rank collapse [1,2,3,4] in transformers can accommodate transformers with causal masking or local attention. As we discussed in the paper, all those works require having full attention as a necessary assumption to derive their theoretical results on rank collapse. To tackle this technical difficulty, we take a novel graph-theoretic approach by formalizing causal masking and local attention through the lens of directed graphs and make use of both graph theory and discrete dynamical systems theory tools. Moreover, our results establish a formal connection between the rate of rank collapse and the graph structure. We hope this novel analysis technique would be a useful tool for the literature and future research in the community. > Q2: If local attention alleviates rank collapse, how do the experimental results compare to those of full attention or causal masked attention? The experimental results can be found in Figure 1, left column. We choose sliding window attention (SWA) deployed in Longformer and more recently Mistral 7B as the representative for local attention. Compared with both full and causal attention, the convergence is clearly slower, which confirms our theoretical results. > Q3: Which approach is more effective in addressing the rank collapse issue: local attention or causal attention? Are there any strong insights into this? I assume local attention may not be generally applicable in language tasks but is more suitable for vision tasks. For pure self-attention networks, local attention would be more effective for mitigating rank collapse. Our theory suggests that the rate of rank collapse is directly affected by the diameter of the directed graph (Theorem 1), which is confirmed with numerical experiments (Figure 1). Regarding the use of local attention, it is indeed popular in vision tasks. See [5,6,7] for references. However, due to the demand for efficiency in long-context, local attention is also getting popular in language tasks. For example, local and sparse attention was populated by Longformer [8] and OpenAI [9] and nowadays LLMs like Mistral 7B use sliding window attention (SWA) [10]. Other popular sparse attention for language tasks include, but not limited to, BigBird [11], Recurrent Memory Transformers (RMTs) [12,13] and StreamingLLM [14]. We appreciate your questions and comments very much. Please let us know for any further questions. ---------------- References [1] Dong et al. Attention is not all you need: pure attention loses rank doubly exponentially with depth. In ICML, 2020. [2] Noci et al. Signal propagation in transformers: Theoretical perspectives and the role of rank collapse. In NeurIPS, 2022. [3] Geshkovski et al. A mathematical perspective on transformers. 2023. [4] Geshkovski et al. The emergence of clusters in self-attention dynamics. In NeurIPS, 2023. [5] Liu et al. Swin transformer: Hierarchical vision transformer using shifted windows. In ICCV, 2021. [6] Pan et al. Slide-Transformer: Hierarchical Vision Transformer with Local Self-Attention. In CVPR, 2023. [7] Hassani et al. Neighborhood Attention Transformer. In CVPR, 2023. [8] Beltagy et al. Longformer: The long-document transformer. 2020. [9] Child et al. Generating Long Sequences with Sparse Transformers. 2019. [10] Jiang et al. Mistral 7B. 2023. [11] Zaheer et al. Bigbird: Transformers for longer sequences. In NeurIPS, 2020. [12] Bulatov et al. Recurrent Memory Transformer. In NeurIPS, 2022. [13] Bulatov et al. Scaling Transformer to 1M tokens and beyond with RMT. 2024. [14] Xiao et al. Efficient Streaming Language Models with Attention Sinks. In ICLR, 2024.	null	null	Rebuttal 1: Rebuttal: ## Response to all reviewers We would like to thank the reviewers for carefully reading our paper and giving insightful comments and constructive feedback. We are glad that our work was recognized as “interesting”, “rigorous” (Reviewer ivrv) and “offering valuable contributions to the field” (Reviewer yhKR), including “a graph-theoretic approach to analyze self-attention” (Reviewer ivrv) and “new insights into the rank collapse phenomenon” (Reviewer 7spk). We are also encouraged that the reviewers found our paper "well-written" (Reviewer 7spk). We have provided detailed responses to each of the reviews separately. We also include additional numerical results in the rebuttal pdf under the same experimental setup as in our paper. Specifically, we increase the sample size of the experiments to 3000 in all experiments and consider the setting of 128-layer randomly initialized models when measuring the effects of different attention masks, LayerNorm and the temperature term. The authors. Pdf: /pdf/2e916ea39b53d446ba8f39722bd3733a28d8f44a.pdf	NeurIPS_2024_submissions_huggingface	2,024	null	null	null	null	null	null	null	null
Mitigating Covariate Shift in Behavioral Cloning via Robust Stationary Distribution Correction	Accept (poster)	Summary: This paper is motivated by the observation that BC is well-known to be vulnerable to the covariate shift resulting from the mismatch between the state distributions induced by the learned policy and the data collection policy. To solve this problem, the authors formulate a robust BC training objective and employ a stationary distribution correction ratio estimation (DICE) to derive a feasible solution. They evaluate the effectiveness of our method through an extensive set of experiments covering diverse covariate shift scenarios. Strengths: 1. This paper is well-written and well-organized. 2. The proposed method seems promising. 3. A lot of covariate shift scenarios are investigated in the experiments, including discrete navigation tasks and mujoco control tasks. Weaknesses: 1. The experimental results do not fully support the advantages over baselines. Technical Quality: 3 Clarity: 3 Questions for Authors: 1. The advantage over OptiDICE-BC seems marginal in Table 2 and Table 4. What could be the reason? 2. What does “marginal room (or action) distribution of dataset” mean in line 189? 3. For mujoco experiments, the authors only use the expert split. How does the method perform in other splits that have lower quality? Confidence: 2 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: One limitation mentioned by the authors in the last section is that: don’t consider uncertainty from transition shifts or noisy demonstrations. Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thank you for your insightful feedback. 1. Addressing Concerns on Marginal Performance Improvements Compared to Baselines In response, we have updated our main experiments to include the DRO baseline DR-BC [18] and variation of $f$-divergence for our approach accordingly: inspired by the choice of $f$ in DR-BC, we introduced the soft-TV distance, where the $f$ function is a log-cosh function and its derivative is the tanh function. [A] This enables us to obtain a closed form solution of $w$ by using Proposition 1. Based on this, we have expanded our analysis by incorporating three additional methods: DR-BC for all scenarios, OptiDICE-BC (with soft-TV), and DrilDICE (with soft-TV) for Mujoco scenarios. The expanded results for Four Rooms environment and Scenario 1,2 can be found in Table A, B, C in the following comment, and the result for Scenario 3 can be found in Figure B in the PDF. The results consistently show that DrilDICE with the soft-TV outperforms baselines including DR-BC across most scenarios. Notably, DrilDICE with soft-TV, utilizing $f$-divergence similar to the one used in DR-BC, consistently outperforms DR-BC in most evaluation scenarios. We attribute this performance improvement to the remaining key difference: the inclusion of a Bellman flow constraint, which is a key contribution of our work. The results also demonstrate that the choice of the soft-TV significantly enhances the performance of DrilDICE compared to using soft-$\chi^2$. To explain this performance gain, we hypothesize that the soft-TV distance provides more discriminative weighting $w$ of samples based on their long-term policy errors. As shown in Figure A in the PDF, conventional $f$-divergences (e.g. KL, soft-$\chi^2$, …) make $w$ less pronounced to the long-term policy error $e$, yet the soft-TV distance responds sensitively to changes in $e$, resulting in a more pronounced $w$. This enables BC loss to more selectively focus on critical samples with significant magnitude of long-term policy errors, thereby effectively enhancing performance, akin to the benefits observed in Sparse Q-Learning [C]. In summary, the choice of the $f$-divergence for DrilDICE is critical for performance. The use of soft-TV distance enhanced DrilDICE's performance across most problem settings, directly addressing the reviewer's concerns regarding its comparative performance. We hope these updates can address your concerns about the initial marginal improvements. [A] Saleh et al., "Statistical Properties of the log-cosh Loss Function used in Machine Learning." arXiv preprint (2022). [B] Xu et al., “Offline RL with No OOD Actions: In-sample Learning via Implicit Value Regularization”, ICLR 2023. 2. Clarification on the Terminology "Marginal Room (or Action) Distribution of Dataset" In line 189, the term "marginal room (or action) distribution of dataset" specifically refers to $p(u)$, the proportion of transitions containing target factors (e.g. room visitation, action) within the manipulated dataset. To illustrate, these factors in our experiment include (1) the room visitation and (2) the action of the transition. By intentionally manipulating frequencies of these factors, we designed covariate shift scenarios. For example, in the Room 1 manipulation scenario, we initially split the expert dataset into two subsets $D_A,D_B$ based on whether each transition’s state was associated with Room 1 or not. We then subsampled transitions from subsets $D_A, D_B$ using the predetermined proportion $p(u), 1-p(u)$, respectively and combined them to construct the shifted dataset $D_i$. Here, the term "a marginal room (or action) distribution" refers to these proportions $p(u)$, representing how frequently transitions with a target factor (e.g. room visiting, action) are sampled relative to others in the dataset. We will revise this into a more clear term such as “a proportion of transitions containing a target factor” and spend our effort to improve presentations. 3. Performance Comparison with Lower Quality Segments Despite our primary focus on expert-quality datasets, we conducted experiments with additional datasets on Scenario 3 (segment datasets) to ensure consistency across different datasets. Rather than employing the D4RL `expert-v2`, we use `medium-v2` quality demonstrations as our imitation standard for comparison. The results are depicted in Figure D of the PDF. Given that relying solely on the normalized scores may not accurately reflect the fidelity of imitation for a medium-quality policy, we primarily measured the target MSE for our comparisons. The results indicate that both DrilDICE and DR-BC demonstrate competitive imitation performance compared to other baselines, with a consistent decrease in target MSE as the number of segments increases. Thank you once more for your valuable feedback. If there are any more questions or concerns, please feel free to respond, and we will address them quickly. --- Rebuttal 2: Title: Additional Experimental Results Comment: ## Four Rooms Environment \| Scenario \| BC \| OptiDICE-BC \| DR-BC \| DrilDICE (Ours) \| \| --- \| --- \| --- \| --- \| --- \| \| Room 1 \| 90.84 ± 0.69 \| 94.30 ± 0.41 \| 91.38 ± 0.73 \| 95.04 ± 0.48 \| \| Room 2 \| 89.16 ± 1.07 \| 94.06 ± 0.65 \| 89.28 ± 1.08 \| 94.44 ± 0.62 \| \| Room 3 \| 88.50 ± 1.27 \| 94.20 ± 0.98 \| 88.70 ± 1.26 \| 95.04 ± 0.86 \| \| Room 4 \| 90.94 ± 0.75 \| 94.26 ± 0.54 \| 90.94 ± 0.75 \| 94.92 ± 0.37 \| \| Action UP \| 84.96 ± 1.33 \| 92.06 ± 0.69 \| 84.96 ± 1.33 \| 93.22 ± 0.61 \| \| Action DOWN \| 89.96 ± 0.96 \| 93.62 ± 0.62 \| 89.96 ± 0.96 \| 94.60 ± 0.39 \| \| Action LEFT \| 90.18 ± 1.11 \| 91.86 ± 1.03 \| 90.18 ± 1.11 \| 92.62 ± 0.95 \| \| Action RIGHT \| 93.04 ± 0.63 \| 94.46 ± 0.44 \| 93.44 ± 0.60 \| 94.52 ± 0.44 \| Table A. Expanded performance comparison of normalized scores on Four Rooms environment. (corresponds to Table 2) ## Scenario 1: Rebalanced Dataset \| Scenario \| Task \| $p(D_1)$ \| BC \| OptiDICE-BC (Soft-$\chi^2$) \| DrilDICE (Soft-$\chi^2$) \| DR-BC \| OptiDICE-BC (Soft-TV) \| DrilDICE (Soft-TV) \| \| --- \| --- \| --- \| --- \| --- \| --- \| --- \| --- \| --- \| \| Rebalanced by state \| hopper \| 0.1 \| 24.65 ± 4.15 \| 35.26 ± 3.19 \| 58.92 ± 4.30 \| 27.02 ± 4.31 \| 12.72 ± 1.27 \| 52.22 ± 5.57 \| \| \| \| 0.5 \| 35.38 ± 4.08 \| 24.41 ± 5.12 \| 60.91 ± 7.61 \| 36.71 ± 3.56 \| 8.61 ± 1.97 \| 67.12 ± 8.18 \| \| \| \| 0.9 \| 11.23 ± 2.49 \| 17.26 ± 2.10 \| 28.29 ± 3.03 \| 27.37 ± 4.89 \| 10.44 ± 1.76 \| 36.39 ± 6.10 \| \| \| walker2d \| 0.1 \| 18.85 ± 4.05 \| 5.91 ± 0.59 \| 31.92 ± 5.94 \| 14.69 ± 3.26 \| 4.91 ± 1.08 \| 51.55 ± 8.16 \| \| \| \| 0.5 \| 22.91 ± 3.13 \| 11.12 ± 1.36 \| 30.53 ± 3.47 \| 45.08 ± 9.98 \| 8.09 ± 0.42 \| 73.74 ± 5.37 \| \| \| \| 0.9 \| 30.42 ± 7.05 \| 17.51 ± 2.61 \| 43.31 ± 4.54 \| 46.04 ± 8.35 \| 7.69 ± 0.45 \| 77.60 ± 5.45 \| \| \| halfcheetah \| 0.1 \| 49.31 ± 5.16 \| 33.42 ± 4.70 \| 44.79 ± 5.16 \| 32.85 ± 3.82 \| 7.05 ± 1.47 \| 52.45 ± 3.62 \| \| \| \| 0.5 \| 37.98 ± 3.07 \| 33.36 ± 3.04 \| 41.13 ± 3.53 \| 26.16 ± 4.91 \| 6.14 ± 1.21 \| 55.04 ± 3.27 \| \| \| \| 0.9 \| 15.54 ± 3.06 \| 2.21 ± 1.16 \| 7.28 ± 1.59 \| 8.96 ± 3.26 \| 1.02 ± 1.12 \| 22.28 ± 2.88 \| \| Rebalanced by action \| hopper \| 0.1 \| 29.71 ± 4.00 \| 28.37 ± 1.11 \| 42.29 ± 6.39 \| 25.92 ± 2.45 \| 11.71 ± 2.12 \| 56.60 ± 11.90 \| \| \| \| 0.5 \| 26.35 ± 4.88 \| 30.03 ± 4.23 \| 53.37 ± 9.50 \| 35.13 ± 5.41 \| 11.79 ± 1.14 \| 73.80 ± 3.63 \| \| \| \| 0.9 \| 30.50 ± 3.60 \| 38.92 ± 4.78 \| 63.14 ± 7.12 \| 36.56 ± 2.29 \| 19.42 ± 2.82 \| 48.99 ± 12.27 \| \| \| walker2d \| 0.1 \| 23.61 ± 5.10 \| 12.06 ± 1.20 \| 40.72 ± 1.13 \| 31.18 ± 4.23 \| 7.27 ± 0.46 \| 70.60 ± 3.21 \| \| \| \| 0.5 \| 32.29 ± 6.74 \| 16.40 ± 1.70 \| 47.93 ± 12.36 \| 30.52 ± 3.89 \| 6.37 ± 0.97 \| 72.09 ± 8.70 \| \| \| \| 0.9 \| 16.87 ± 2.80 \| 15.64 ± 3.57 \| 43.68 ± 11.50 \| 37.55 ± 8.97 \| 4.60 ± 0.97 \| 69.51 ± 8.54 \| \| \| halfcheetah \| 0.1 \| 41.91 ± 4.80 \| 26.62 ± 2.54 \| 32.76 ± 1.87 \| 27.50 ± 1.01 \| 8.41 ± 3.38 \| 56.42 ± 4.57 \| \| \| \| 0.5 \| 45.80 ± 4.45 \| 45.52 ± 3.24 \| 48.08 ± 5.50 \| 33.39 ± 6.52 \| 4.64 ± 0.84 \| 60.81 ± 1.56 \| \| \| \| 0.9 \| 25.91 ± 3.35 \| 4.28 ± 1.52 \| 9.57 ± 2.34 \| 12.08 ± 2.01 \| 0.59 ± 0.68 \| 29.19 ± 4.58 \| Table B. Expanded performance comparison on Scenario 1 (rebalanced dataset). (corresponds to Table 3) ## Scenario 2: Time-dependently Subsampled Dataset \| Task \| (a, b) \| BC \| OptiDICE-BC (Soft-$\chi^2$) \| DrilDICE (Soft-$\chi^2$) \| DR-BC \| OptiDICE-BC (Soft-TV) \| DrilDICE (Soft-TV) \| \| --- \| ------- \| --- \| --- \| --- \| --- \| --- \| --- \| \| hopper \| (1, 1) \| 28.89 ± 3.77 \| 50.33 ± 6.60 \| 54.83 ± 7.66 \| 21.10 ± 2.26 \| 22.77 ± 3.94 \| 45.44 ± 5.11 \| \| \| (1, 5) \| 31.03 ± 0.90 \| 39.54 ± 4.02 \| 37.18 ± 7.92 \| 25.00 ± 1.66 \| 19.25 ± 1.21 \| 45.60 ± 4.63 \| \| \| (5, 1) \| 26.75 ± 7.12 \| 48.40 ± 12.98 \| 39.91 ± 9.20 \| 17.51 ± 3.38 \| 25.68 ± 6.01 \| 34.71 ± 9.00 \| \| \| (5, 5) \| 27.65 ± 6.71 \| 32.46 ± 10.79 \| 40.24 ± 7.41 \| 23.20 ± 6.32 \| 14.12 ± 3.59 \| 25.61 ± 6.03 \| \| walker2d \| (1, 1) \| 28.95 ± 5.34 \| 17.42 ± 3.00 \| 51.85 ± 5.30 \| 45.66 ± 9.92 \| 6.13 ± 1.03 \| 81.21 ± 5.40\| \| \| (1, 5) \| 61.48 ± 5.19 \| 37.25 ± 5.66 \| 64.46 ± 7.92 \| 57.29 ± 4.79 \| 17.55 ± 2.65 \| 84.28 ± 4.89 \| \| \| (5, 1) \| 8.13 ± 0.72 \| 4.43 ± 0.91 \| 23.31 ± 3.44 \| 17.97 ± 2.58 \| 4.37 ± 0.52 \| 48.23 ± 8.30 \| \| \| (5, 5) \| 6.65 ± 1.20 \| 8.50 ± 2.27 \| 14.40 ± 3.21 \| 12.45 ± 1.84 \| 5.54 ± 0.68 \| 52.57 ± 6.24 \| \| halfcheetah \| (1, 1) \| 33.74 ± 2.99 \| 33.65 ± 5.49 \| 33.43 ± 3.73 \| 17.09 ± 2.43 \| 9.93 ± 3.56 \| 44.17 ± 5.62 \| \| \| (1, 5) \| 72.72 ± 2.60 \| 52.94 ± 3.98 \| 69.63 ± 3.86 \| 61.81 ± 2.50 \| 24.42 ± 3.27 \| 77.12 ± 2.42 \| \| \| (5, 1) \| 2.35 ± 0.51 \| 2.46 ± 1.05 \| 3.97 ± 1.18 \| 3.81 ± 1.26 \| 1.29 ± 1.13 \| 5.68 ± 1.50 \| \| \| (5, 5) \| 2.01 ± 0.91 \| 1.61 ± 1.30 \| 4.61 ± 1.55 \| 2.85 ± 1.05 \| -1.19 ± 0.37 \| 5.50 ± 0.83 \| Table C. Expanded performance comparison on Scenario 2 (time-dependently collected dataset). (corresponds to Table 4) --- Rebuttal Comment 2.1: Title: Reply to authors Comment: Thanks to the authors for adding new experimental results to address my concerns. The new results provide new evidence and support for the proposed method. I don't have further questions and I raised my score to 6 to lean to accept this paper. --- Reply to Comment 2.1.1: Comment: We're glad to hear that your concerns have been addressed, and we sincerely appreciate your positive review! We will incorporate all the findings discussed with you into the revised manuscript. --- Rebuttal 3: Title: Dear Reviewer k6J7 Comment: We respectfully remind you that less than 48 hours remain in our discussion period. We are dedicated to addressing any remaining concerns you may have. In short, we addressed your key concerns as follows: - Concerns on Marignal Improvements: - We found that performance of DrilDICE depends on a choice of $f$-divergence. When similarly aligned with TV distance, DrilDICE consistently outperforms all baselines. - Comparison on Lower-Quality Datasets: - We conduct additional experiments on segmented trajectory scenario (Scenario 3) with a medium-quality dataset, instead of the expert dataset. Please refer to our detailed rebuttal for more comprehensive information. If you have any further questions or concerns, please do not hesitate to share your comments. We sincerely thank you again for your reviews.	Summary: This paper studies imitation learning when the offline dataset does not come from the stationary expert distribution. To address this problem, the authors introduce the objective of distributionally robust optimization into behavioral cloning. To avoid overly pessimistic solutions, the authors further incorporate the Bellman flow constraint and empirical distribution divergence into the min-max optimization objective. Based on the DICE technology, the solution can be derived. The experimental results have shown the proposal's effectiveness on corrupted datasets. Strengths: The proposed solution is technically reasonable and theoretically sound. Weaknesses: 1. I agree this paper is well-contained. This work successfully introduces the min-max formulation of DRO to the imitation learning community. However, the new setting proposed in the paper is not well-supported. The shift between the offline dataset and the stationary expert distribution is a reasonable setting. However, all experiments are validated based on simulated shifts, which somewhat makes me feel less excited. 2. The experiments are conducted on simulated covariate shifts. More evaluation in real-world cases will further improve this paper. 3. Some notations have not been well defined, the $\Delta S$ in Equation 2, and the $\mathcal{W}$ in Equation 5. Technical Quality: 3 Clarity: 3 Questions for Authors: In the navigation task like the current experiments in the Four Rooms environment, limited expert trajectories from some start points, especially, have only partial coverage on the state space, is there also a covariate shift with the stationary distribution? I personally think this might be a more real and direct case in real-world applications. Is the method proposed still effective in this case? Confidence: 3 Soundness: 3 Presentation: 3 Contribution: 2 Limitations: The authors have provided a discussion about the limitations and broader societal impacts. Flag For Ethics Review: ['No ethics review needed.'] Rating: 5 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thank you for your valuable feedback. 1. Experimental Design and Real-World Applicability Concerns Our experimental setup, though based on simulated data, is intended to rigorously assess the adaptability and robustness of our approach under realistic conditions. This strategy also ensures reproducibility and is consistent with established methodologies in the field, as demonstrated by similar studies like Yan et al., which also employs segments of the D4RL dataset to test hypotheses in learning from observations (LfO) problem setting [A]. Beyond the simulated shifts, we also have conducted tests on complete trajectories with limited data sizes—a scenario frequently encountered in real-world applications. The results, illustrated in Figure C, demonstrate that DrilDICE adeptly handles sampling errors in small datasets and surpasses existing approaches in performance. [A] Yan et al., “A Simple Solution for Offline Imitation from Observations and Examples with Possibly Incomplete Trajectories,” NeurIPS 2023. 2. Clarification on Unclear Notations We clarify that $\Delta S$ represents a set of arbitrary state distributions, and $\mathcal{W}$ denotes a function space of $w(s,a)$ with $w:\mathcal{S}\times\mathcal{A} \to \mathbb{R}$. These updates will be reflected in the final version of the paper. 3. Applicability in Partial Coverage Scenarios DrilDICE, like other DICE approaches, relies on the assumption that the support of the target state distribution $d$ should encompass the support of the expert's stationary distribution $d_E$. This assumption presents a challenge in ensuring the robustness of DICE approaches when the support of $d$ does not fully cover that of $d_E$. Despite this, DrilDICE has shown strong performance in scenarios with limited data coverage. In our experiments, particularly in Scenario 3 and the complete trajectories experiment (see Figure B, C in the PDF respectively), we varied dataset sizes to adjust data coverage. The result suggests that DRO approaches possibly handle sampling errors in small datasets. While these results are encouraging, the comprehensive applicability of our method when $d$ partially covers $d_E$ is yet to be fully established. We believe that further exploration of this critical aspect represents a promising research topic. 4. Expanded Evaluation Results on Main Scenarios We have updated our main experiments to include a baseline DR-BC [18] and a variation of $f$-divergence for our approach accordingly: Inspired by the choice of $f$ in DR-BC, we introduced the soft-TV distance, where the $f$ function is a log-cosh function and its derivative is the tanh function. [B] This provides a relaxed and invertible version of TV distance, and enables us to obtain a closed form solution of $w$ by using Proposition 1. Based on this, we have expanded our analysis by incorporating three additional methods: DR-BC for all scenarios, OptiDICE-BC (with soft-TV), and DrilDICE (with soft-TV) for Mujoco scenarios. The expanded results for Four Rooms environment and Scenario 1,2 can be found in Table A, B and C in the comment, and the result for Scenario 3 can be found in Figure B in the PDF. The results consistently show that DrilDICE with the soft-TV outperforms baselines including DR-BC across most scenarios. Notably, DrilDICE with soft-TV, utilizing $f$-divergence similar to the one used in DR-BC, consistently outperforms DR-BC in most evaluation scenarios. We attribute this performance improvement to the remaining key difference: an inclusion of a Bellman flow constraint, which is a key contribution of our work. The results also demonstrate that the choice of the soft-TV significantly enhances the performance of DrilDICE compared to using soft-$\chi^2$. To explain this performance gain, we hypothesize that the soft-TV distance provides more discriminative weighting $w$ of samples based on their long-term policy errors. As shown in Figure A in the PDF, conventional $f$-divergences (e.g. KL, soft-$\chi^2$, …) make $w$ less pronounced to the long-term policy error $e$, yet the soft-TV distance responds sensitively to changes in $e$, resulting in a more pronounced $w$. This enables BC loss to more selectively focus on critical samples with significant magnitude of long-term policy errors, thereby effectively enhancing performance, akin to the benefits observed in Sparse Q-Learning [C]. [B] Saleh et al., "Statistical Properties of the log-cosh Loss Function used in Machine Learning." arXiv preprint (2022). [C] Xu et al., “Offline RL with No OOD Actions: In-sample Learning via Implicit Value Regularization”, ICLR 2023. Thank you again for your insightful comments. If you have any further concerns or questions about this, please feel free to reply and we will address them promptly. --- Rebuttal Comment 1.1: Comment: Thanks for your response. I agree that segmented trajectories have simulated a type of real-world covariate shift. --- Reply to Comment 1.1.1: Comment: We're glad to hear that some of your concerns have been addressed. We truly believe that your insightful feedback has notably improved the quality of our manuscript. If you have any remaining concerns, please let us know so that we can further enhance the quality of our research. --- Reply to Comment 1.1.2: Title: Dear Reviewer oZLc Comment: We are awaiting your feedback on any remaining concerns, and we would be very happy to resolve these issues. We are eager to address remaining concerns to enhance our manuscript, so please do not hesitate to let us know. Thank you once again for your invaluable reviews. --- Rebuttal 2: Title: Additional Experimental Results Comment: ## Four Rooms Environment \| Scenario \| BC \| OptiDICE-BC \| DR-BC \| DrilDICE (Ours) \| \| --- \| --- \| --- \| --- \| --- \| \| Room 1 \| 90.84 ± 0.69 \| 94.30 ± 0.41 \| 91.38 ± 0.73 \| 95.04 ± 0.48 \| \| Room 2 \| 89.16 ± 1.07 \| 94.06 ± 0.65 \| 89.28 ± 1.08 \| 94.44 ± 0.62 \| \| Room 3 \| 88.50 ± 1.27 \| 94.20 ± 0.98 \| 88.70 ± 1.26 \| 95.04 ± 0.86 \| \| Room 4 \| 90.94 ± 0.75 \| 94.26 ± 0.54 \| 90.94 ± 0.75 \| 94.92 ± 0.37 \| \| Action UP \| 84.96 ± 1.33 \| 92.06 ± 0.69 \| 84.96 ± 1.33 \| 93.22 ± 0.61 \| \| Action DOWN \| 89.96 ± 0.96 \| 93.62 ± 0.62 \| 89.96 ± 0.96 \| 94.60 ± 0.39 \| \| Action LEFT \| 90.18 ± 1.11 \| 91.86 ± 1.03 \| 90.18 ± 1.11 \| 92.62 ± 0.95 \| \| Action RIGHT \| 93.04 ± 0.63 \| 94.46 ± 0.44 \| 93.44 ± 0.60 \| 94.52 ± 0.44 \| Table A. Expanded performance comparison of normalized scores on Four Rooms environment. (corresponds to Table 2) ## Scenario 1: Rebalanced Dataset \| Scenario \| Task \| $p(D_1)$ \| BC \| OptiDICE-BC (Soft-$\chi^2$) \| DrilDICE (Soft-$\chi^2$) \| DR-BC \| OptiDICE-BC (Soft-TV) \| DrilDICE (Soft-TV) \| \| --- \| --- \| --- \| --- \| --- \| --- \| --- \| --- \| --- \| \| Rebalanced by state \| hopper \| 0.1 \| 24.65 ± 4.15 \| 35.26 ± 3.19 \| 58.92 ± 4.30 \| 27.02 ± 4.31 \| 12.72 ± 1.27 \| 52.22 ± 5.57 \| \| \| \| 0.5 \| 35.38 ± 4.08 \| 24.41 ± 5.12 \| 60.91 ± 7.61 \| 36.71 ± 3.56 \| 8.61 ± 1.97 \| 67.12 ± 8.18 \| \| \| \| 0.9 \| 11.23 ± 2.49 \| 17.26 ± 2.10 \| 28.29 ± 3.03 \| 27.37 ± 4.89 \| 10.44 ± 1.76 \| 36.39 ± 6.10 \| \| \| walker2d \| 0.1 \| 18.85 ± 4.05 \| 5.91 ± 0.59 \| 31.92 ± 5.94 \| 14.69 ± 3.26 \| 4.91 ± 1.08 \| 51.55 ± 8.16 \| \| \| \| 0.5 \| 22.91 ± 3.13 \| 11.12 ± 1.36 \| 30.53 ± 3.47 \| 45.08 ± 9.98 \| 8.09 ± 0.42 \| 73.74 ± 5.37 \| \| \| \| 0.9 \| 30.42 ± 7.05 \| 17.51 ± 2.61 \| 43.31 ± 4.54 \| 46.04 ± 8.35 \| 7.69 ± 0.45 \| 77.60 ± 5.45 \| \| \| halfcheetah \| 0.1 \| 49.31 ± 5.16 \| 33.42 ± 4.70 \| 44.79 ± 5.16 \| 32.85 ± 3.82 \| 7.05 ± 1.47 \| 52.45 ± 3.62 \| \| \| \| 0.5 \| 37.98 ± 3.07 \| 33.36 ± 3.04 \| 41.13 ± 3.53 \| 26.16 ± 4.91 \| 6.14 ± 1.21 \| 55.04 ± 3.27 \| \| \| \| 0.9 \| 15.54 ± 3.06 \| 2.21 ± 1.16 \| 7.28 ± 1.59 \| 8.96 ± 3.26 \| 1.02 ± 1.12 \| 22.28 ± 2.88 \| \| Rebalanced by action \| hopper \| 0.1 \| 29.71 ± 4.00 \| 28.37 ± 1.11 \| 42.29 ± 6.39 \| 25.92 ± 2.45 \| 11.71 ± 2.12 \| 56.60 ± 11.90 \| \| \| \| 0.5 \| 26.35 ± 4.88 \| 30.03 ± 4.23 \| 53.37 ± 9.50 \| 35.13 ± 5.41 \| 11.79 ± 1.14 \| 73.80 ± 3.63 \| \| \| \| 0.9 \| 30.50 ± 3.60 \| 38.92 ± 4.78 \| 63.14 ± 7.12 \| 36.56 ± 2.29 \| 19.42 ± 2.82 \| 48.99 ± 12.27 \| \| \| walker2d \| 0.1 \| 23.61 ± 5.10 \| 12.06 ± 1.20 \| 40.72 ± 1.13 \| 31.18 ± 4.23 \| 7.27 ± 0.46 \| 70.60 ± 3.21 \| \| \| \| 0.5 \| 32.29 ± 6.74 \| 16.40 ± 1.70 \| 47.93 ± 12.36 \| 30.52 ± 3.89 \| 6.37 ± 0.97 \| 72.09 ± 8.70 \| \| \| \| 0.9 \| 16.87 ± 2.80 \| 15.64 ± 3.57 \| 43.68 ± 11.50 \| 37.55 ± 8.97 \| 4.60 ± 0.97 \| 69.51 ± 8.54 \| \| \| halfcheetah \| 0.1 \| 41.91 ± 4.80 \| 26.62 ± 2.54 \| 32.76 ± 1.87 \| 27.50 ± 1.01 \| 8.41 ± 3.38 \| 56.42 ± 4.57 \| \| \| \| 0.5 \| 45.80 ± 4.45 \| 45.52 ± 3.24 \| 48.08 ± 5.50 \| 33.39 ± 6.52 \| 4.64 ± 0.84 \| 60.81 ± 1.56 \| \| \| \| 0.9 \| 25.91 ± 3.35 \| 4.28 ± 1.52 \| 9.57 ± 2.34 \| 12.08 ± 2.01 \| 0.59 ± 0.68 \| 29.19 ± 4.58 \| Table B. Expanded performance comparison on Scenario 1 (rebalanced dataset). (corresponds to Table 3) ## Scenario 2: Time-dependently Subsampled Dataset \| Task \| (a, b) \| BC \| OptiDICE-BC (Soft-$\chi^2$) \| DrilDICE (Soft-$\chi^2$) \| DR-BC \| OptiDICE-BC (Soft-TV) \| DrilDICE (Soft-TV) \| \| --- \| ------- \| --- \| --- \| --- \| --- \| --- \| --- \| \| hopper \| (1, 1) \| 28.89 ± 3.77 \| 50.33 ± 6.60 \| 54.83 ± 7.66 \| 21.10 ± 2.26 \| 22.77 ± 3.94 \| 45.44 ± 5.11 \| \| \| (1, 5) \| 31.03 ± 0.90 \| 39.54 ± 4.02 \| 37.18 ± 7.92 \| 25.00 ± 1.66 \| 19.25 ± 1.21 \| 45.60 ± 4.63 \| \| \| (5, 1) \| 26.75 ± 7.12 \| 48.40 ± 12.98 \| 39.91 ± 9.20 \| 17.51 ± 3.38 \| 25.68 ± 6.01 \| 34.71 ± 9.00 \| \| \| (5, 5) \| 27.65 ± 6.71 \| 32.46 ± 10.79 \| 40.24 ± 7.41 \| 23.20 ± 6.32 \| 14.12 ± 3.59 \| 25.61 ± 6.03 \| \| walker2d \| (1, 1) \| 28.95 ± 5.34 \| 17.42 ± 3.00 \| 51.85 ± 5.30 \| 45.66 ± 9.92 \| 6.13 ± 1.03 \| 81.21 ± 5.40\| \| \| (1, 5) \| 61.48 ± 5.19 \| 37.25 ± 5.66 \| 64.46 ± 7.92 \| 57.29 ± 4.79 \| 17.55 ± 2.65 \| 84.28 ± 4.89 \| \| \| (5, 1) \| 8.13 ± 0.72 \| 4.43 ± 0.91 \| 23.31 ± 3.44 \| 17.97 ± 2.58 \| 4.37 ± 0.52 \| 48.23 ± 8.30 \| \| \| (5, 5) \| 6.65 ± 1.20 \| 8.50 ± 2.27 \| 14.40 ± 3.21 \| 12.45 ± 1.84 \| 5.54 ± 0.68 \| 52.57 ± 6.24 \| \| halfcheetah \| (1, 1) \| 33.74 ± 2.99 \| 33.65 ± 5.49 \| 33.43 ± 3.73 \| 17.09 ± 2.43 \| 9.93 ± 3.56 \| 44.17 ± 5.62 \| \| \| (1, 5) \| 72.72 ± 2.60 \| 52.94 ± 3.98 \| 69.63 ± 3.86 \| 61.81 ± 2.50 \| 24.42 ± 3.27 \| 77.12 ± 2.42 \| \| \| (5, 1) \| 2.35 ± 0.51 \| 2.46 ± 1.05 \| 3.97 ± 1.18 \| 3.81 ± 1.26 \| 1.29 ± 1.13 \| 5.68 ± 1.50 \| \| \| (5, 5) \| 2.01 ± 0.91 \| 1.61 ± 1.30 \| 4.61 ± 1.55 \| 2.85 ± 1.05 \| -1.19 ± 0.37 \| 5.50 ± 0.83 \| Table C. Expanded performance comparison on Scenario 2 (time-dependently collected dataset). (corresponds to Table 4) --- Rebuttal 3: Title: Dear Reviewer oZLc Comment: We gently remind you that less than 48 hours remain in our discussion period. We are committed to thoroughly addressing the reviewer’s remaining concerns. In summary, our responses are as follows: - Concerns on Experiments on Simulated Shift: - Our covariate shift experiment setting ensures robustness and reproducibliity of comparison. Similar simulated shifts have been utilized in exisiting literature. - Additionally, we have included natural complete trajectory scenarios where DrilDICE shows significant data-efficiency. - Concerns on Partial Coverage: - Although DrilDICE is not specifically designed to address partial coverage issues, it has shown empirical robustness in such scenarios. - Expanded Main Experiments: - We introduced DR-BC as a new baseline across all experiments; DrilDICE consistently outperforms DR-BC when selecting a similar $f$-divergence. Please see our rebuttal for more information. If you have any further questions or concerns, do not hesitate to share your comments. Again, thank you for your thoughtful reviews.	Summary: This paper devices distribution correction ratio estimation (DICE)-based optimization to mitigate the covariate shift issue in the Behavior cloning algorithm. They test their heuristic loss on Mujoco benchmarks. Strengths: The design of drildice in Section 3.2 and its evaluation on Mujoco benchmark are the strengths of this work. Weaknesses: The results in this work are preliminary and need to be evaluated carefully. I move all my concerns to the Questions section below instead of treating these as weaknesses. I will also rely on the author-reviewer and reviewer-reviewer discussion periods for updating my decisions. Technical Quality: 2 Clarity: 3 Questions for Authors: - Section 3.2 requires further details and refined writing. For example, - Slater's condition/assumption must be explicitly mentioned since the drildice algorithm relies on it. - Example of $f$ and $f'$ should be mentioned in this section to make sense of the experiments section (they are only mentioned in Appendix as KL and $\chi^2$.) - More details are needed to support this statement: "the problem can be solved by alternatively optimizing $\nu$ and $\pi$." Such as citing previous DICE works and so on. E.g reference https://proceedings.neurips.cc/paper_files/paper/2019/file/cf9a242b70f45317ffd281241fa66502-Paper.pdf - These details will also help in providing some theoretical justifications from previous DICE-related works, which this paper currently lacks. - Comparison with previous works must be further expanded. Especially with [18]. - Both this paper Section 3.1 and [18] share similar motivation to come up with loss functions to mitigate covariate shift created by $d_D\neq d_E$. - At line 92, it is mentioned that $d_D\neq d_E$ but the underlying policy (which is $\pi_E$) is the same for the two state-action visitation distributions. Now, considering the definition of state-action visitation distribution at line 54, it is easy to see that $d_D$ and $d_E$ in fact only differ in the shifts corresponding to transition distribution T of the underlying MDP. Thus, this paper is considering transition distribution shifts as opposed to the claim at line 89! - Based on the above points, [18] also considers distribution matching in their inner problem w.r.t transition distribution T shifts. - Despite this similarity between [18] and this work, both take different approaches to design the loss. This work takes the Bellman flow constraints route whereas [18] considers direct DRO on the transition distribution T shifts. - [18] provides both theoretical guarantees and empirical evaluations and this paper provides only empirical evaluations. - Regarding experiments - Can you please explain how $D_i$ at line 194 differs in size of the dataset compared to $D_E$? With $p(u)=0.4$, is it the case that 40% of states are sampled from $D_E$ and the rest are random states? - The analogy of sampling in line 187 "if a data collection device (e.g. cameras or sensors) operates at different recording frequencies in each room" can also be considered as shifts in transition distribution T, thus sharing the similarities as discussed before. Similar observation can be made for the Scenarios 1 and 3. Connection to transition shifts is more apparent in Scenario 2. So I think it is worthwhile adding more benchmark algorithms [2, 18]. - Can you also include another scenario, training without covariate shift for instance? BC and other non-robust versions should outperform in such ablation-type tests. My score reflects this review provided here. Confidence: 4 Soundness: 2 Presentation: 3 Contribution: 3 Limitations: na Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: First and foremost, we sincerely appreciate your insightful feedback. 1. Justification on Alternative Optimization To clarify, our initial statement suggested that alternating optimization provides a practical approach to approximating the solution of the problem as outlined in Equation (7). In response, we will clearly revise this statement to “... the problem can be practically addressed by alternately optimizing $\nu$ and $\pi$, following recent developments in DICE approaches. [citations]” with related literature including suggested DualDICE and OptiDICE. Additionally, we will provide a more detailed exposition on the derivation (e.g. details on Slater’s condition, assumptions) in the forthcoming revision. 2. Clarification on $d_D$ and Our Problem Setting It is important to note that we did not assume that $d_D$ is a stationary distribution of any policy; rather, it is an arbitrary state distribution. Hence, even if the transition dynamics $T$ and the policy $\pi_E$ remain unchanged, $d_D$ can deviate from $d_E$ since we do not assume that $d_D$ should be induced from MDPs. We clarify this point in the revision. Additionally, we also clarify that our problem setting concerns shifts in stationary distribution (say $d$-shifts), not transition shifts ($T$-shifts). Specifically, we assumed that the expert dataset is a collection of transitions $(s,a,s’)$, where $s$ sampled from $d_D$, an expert action $a$ decided by the deterministic expert $\pi_E$, and $s’$ sampled from $T(s’\|s,a)$ (which will not be shifted in the testing phase). Even when $T$ and $\pi_E$ remain unchanged, practical constraints such as high costs or limited recording frequencies often result in dataset with sparsely collected transitions. This sparsity leads to incomplete trajectories, causing a deviation of $d_D(s)$ from $d_E(s)$. We do not focus on delayed transitions such as ($s_t, a_t, s_{t+k}$) with $k \ge 2$, which may arise from not collecting immediate subsequent states. Although we maintain that our setting does not directly involve $T$-shifts, we compare our approach with DR-BC, as their objective also pertains to $d$-shifts. 3. Comparison with DR-BC [18] In response to the reviewer’s suggestion, we have revisited DR-BC [18] and investigated the method more thoroughly. While DR-BC primarily aims to address $T$-shift scenarios, its objective considers the uncertainty set of arbitrary state distributions $d(s)$, which is potentially applicable to our scenario. Hence, we evaluate DR-BC as a baseline for all of our scenarios. To compare, the key differences between DR-BC and our approach can be summarized as: - Choice of Uncertainty Set: To define its uncertainty set, DR-BC employs TV distance while DrilDICE uses $f$-divergence, which is more general. - Adoption of Bellman Flow Constraints: For an uncertainty set, DR-BC considers arbitrary state distributions, including non-stationary ones, while DrilDICE considers state stationary distributions by enforcing Bellman flow constraints. 4. Technical Adjustment for DrilDICE and Summary of $f$-divergence We initially attempted to align the $f$-divergence to the TV distance used in DR-BC for a fair comparison. Our formulation requires the inverse of the derivative of $f$ for a closed form solution of the weight $w$. However, since the derivative $f$ of TV distance is a step function and is not invertible, we cannot directly use TV distance. We address this issue by adopting the log-cosh function [A] for $f$, which has a tanh function as $f’$; a relaxed version of the step function while ensuring invertibility. We refer to this as the soft-TV distance and utilized this in Mujoco-based experiments. (See Figure A for visualization) In summary, the following table compares the choice of $f$-divergence for DrilDICE, highlighting our method's adaptability in obtaining practical solutions through refined mathematical approaches. As the reviewer commented, we will include this table in the main manuscript. \| Divergence \| $f(x)$ \| $(f')^{-1}(y)$ \| $w^_{\pi,\nu}$ \| \| --- \| --- \| --- \| --- \| \| KL Divergence \| $x\log x$ \| $\exp(y-1)$ \| $\exp\left(\frac{e_{\pi,\nu}(s,a, s')}{\alpha}-1\right)$ \| \| $\chi^2$-Divergence \| $\frac{1}{2}(x-1)^2$ \| $(y+1)$ \| $\mathrm{ReLU}(\frac{e_{\pi,\nu}(s,a, s')}{\alpha} + 1)$ \| \| Soft-$\chi^2$ Divergence \| $f_{\text{soft}-\chi^2} (x) $ \| $(f^{'})^{-1}_{\text{soft}-\chi^2} (y) $ \| $\mathrm{ELU}\left(\frac{e_{\pi,\nu}(s,a, s')}{\alpha}\right) + 1$ \| \| TV Distance \| $\frac{1}{2}\\|x-1\\|$ \| - \| - \| \| Soft-TV Distance \| $\frac{1}{2} \log (\cosh( x-1))$ \| $\tanh^{-1}(2y)+1$ \| $\mathrm{ReLU}(\tanh^{-1}(\frac{2e_{\pi,\nu}(s,a, s')}{\alpha})+1)$ \| where $f_{\text{soft}-\chi^2} (x) := x\log x -x +1$ if $0<x<1$ and $(x-1)^2$ if $x \ge 1$, $(f^{'})^{-1}_{\text{soft}-\chi^2} (y) := \exp(y)$ if $y<0$ and $y+1$ if $y \ge 0$, $\mathrm{ReLU}(x):=\max(0,x)$, $\mathrm{ELU}(x):= \exp(x)-1$ if $x<0$ and $x$ for $x\ge0$. [A] Saleh et al., "Statistical Properties of the log-cosh Loss Function used in Machine Learning." arXiv preprint (2022). --- Rebuttal 2: Title: Rebuttal (continued) Comment: 5. Expanded Results on Main Experiments* We added three additional methods: DR-BC for all scenarios, OptiDICE-BC (with soft-TV), DrilDICE (with soft-TV) for Mujoco scenarios. The expanded results for Four Rooms environment and Scenario 1,2 are presented respectively in Table A, B, and C respectively. The results for Scenario 3 are detailed in Figure B of the PDF. The results consistently show that DrilDICE with the soft-TV outperforms baselines including DR-BC across most scenarios. Notably, DrilDICE with soft-TV, utilizing $f$-divergence similar to the one used in DR-BC, consistently outperforms DR-BC in most evaluation scenarios. We attribute this performance improvement to the remaining key difference: an inclusion of a Bellman flow constraint, which is a key contribution of our work. The results also demonstrate that the choice of the soft-TV significantly enhances the performance of DrilDICE compared to using soft-$\chi^2$. To explain this performance gain, we hypothesize that the soft-TV distance provides more discriminative weighting $w$ of samples based on their long-term policy errors. As shown in Figure A in the PDF, conventional $f$-divergences (e.g. KL, soft-$\chi^2$, …) make $w$ less pronounced to the long-term policy error $e$, yet the soft-TV distance responds sensitively to changes in $e$, resulting in a more pronounced $w$. This enables BC loss to selectively focus on critical samples with significant magnitude of long-term policy errors, thereby effectively enhancing performance, akin to the benefits observed in Sparse Q-Learning [B]. [B] Xu et al., “Offline RL with No OOD Actions: In-sample Learning via Implicit Value Regularization”, ICLR 2023. 6. Clarification on Datasets Used in Four Rooms Experiments In our Four Rooms experiment, we designed the covariate shift scenarios caused by unobserved factors affecting data curation, specifically considering such factors as (1) room visitation or (2) action. In the Room 1 manipulation scenario, we initially split the expert dataset into two subsets $D_A,D_B$ based on whether each transition’s state was associated with Room 1 or not. We then subsampled transitions from subsets $D_A, D_B$ using the predetermined proportion $p(u), 1-p(u)$, respectively and combined them to construct the shifted dataset $D_i$. Specifically, we utilized 100 original episodes, comprising a total of 3994 transitions with a maximum episode length of 50, for $D_E$. From these, after setting $p(u) = 0.4$, we subsampled 1000 transitions to construct $D_i$, which includes 40% (400 transitions) from $D_A$, and 60% (600 transitions) from $D_B$ (e.g., transitions in Rooms 2, 3, 4). To avoid potential issues with support coverage—beyond the focus of our study—we ensured that all states ($\|S\| = 11 \times 11 = 121$) in $D_i$ appear at least once, as detailed in the Appendix. 7. Experiments on Standard Scenarios We acknowledge that our initial experiments did not include comparisons on standard scenarios. To address this, we have conducted additional experiments using complete expert trajectories using D4RL `expert-v2` dataset varying the number of trajectories in \{1, 5, 10, 50\}. The results are detailed in Figure C in the PDF. We also observed that DrilDICE can deal with sampling errors of small datasets, showing a superior performance compared to other methods. Once again, we extend our sincere thanks for your valuable feedback. We believe that addressing your concerns has substantially enhanced the quality of our research. We plan to incorporate all our discussions into the final version. If you have any remaining concerns or questions, please do not hesitate to comment and we will respond as soon as possible. --- Rebuttal 3: Title: Additional Experimental Results Comment: ## Four Rooms Environment \| Scenario \| BC \| OptiDICE-BC \| DR-BC \| DrilDICE (Ours) \| \| --- \| --- \| --- \| --- \| --- \| \| Room 1 \| 90.84 ± 0.69 \| 94.30 ± 0.41 \| 91.38 ± 0.73 \| 95.04 ± 0.48 \| \| Room 2 \| 89.16 ± 1.07 \| 94.06 ± 0.65 \| 89.28 ± 1.08 \| 94.44 ± 0.62 \| \| Room 3 \| 88.50 ± 1.27 \| 94.20 ± 0.98 \| 88.70 ± 1.26 \| 95.04 ± 0.86 \| \| Room 4 \| 90.94 ± 0.75 \| 94.26 ± 0.54 \| 90.94 ± 0.75 \| 94.92 ± 0.37 \| \| Action UP \| 84.96 ± 1.33 \| 92.06 ± 0.69 \| 84.96 ± 1.33 \| 93.22 ± 0.61 \| \| Action DOWN \| 89.96 ± 0.96 \| 93.62 ± 0.62 \| 89.96 ± 0.96 \| 94.60 ± 0.39 \| \| Action LEFT \| 90.18 ± 1.11 \| 91.86 ± 1.03 \| 90.18 ± 1.11 \| 92.62 ± 0.95 \| \| Action RIGHT \| 93.04 ± 0.63 \| 94.46 ± 0.44 \| 93.44 ± 0.60 \| 94.52 ± 0.44 \| Table A. Expanded performance comparison of normalized scores on Four Rooms environment. (corresponds to Table 2) ## Scenario 1: Rebalanced Dataset \| Scenario \| Task \| $p(D_1)$ \| BC \| OptiDICE-BC (Soft-$\chi^2$) \| DrilDICE (Soft-$\chi^2$) \| DR-BC \| OptiDICE-BC (Soft-TV) \| DrilDICE (Soft-TV) \| \| --- \| --- \| --- \| --- \| --- \| --- \| --- \| --- \| --- \| \| Rebalanced by state \| hopper \| 0.1 \| 24.65 ± 4.15 \| 35.26 ± 3.19 \| 58.92 ± 4.30 \| 27.02 ± 4.31 \| 12.72 ± 1.27 \| 52.22 ± 5.57 \| \| \| \| 0.5 \| 35.38 ± 4.08 \| 24.41 ± 5.12 \| 60.91 ± 7.61 \| 36.71 ± 3.56 \| 8.61 ± 1.97 \| 67.12 ± 8.18 \| \| \| \| 0.9 \| 11.23 ± 2.49 \| 17.26 ± 2.10 \| 28.29 ± 3.03 \| 27.37 ± 4.89 \| 10.44 ± 1.76 \| 36.39 ± 6.10 \| \| \| walker2d \| 0.1 \| 18.85 ± 4.05 \| 5.91 ± 0.59 \| 31.92 ± 5.94 \| 14.69 ± 3.26 \| 4.91 ± 1.08 \| 51.55 ± 8.16 \| \| \| \| 0.5 \| 22.91 ± 3.13 \| 11.12 ± 1.36 \| 30.53 ± 3.47 \| 45.08 ± 9.98 \| 8.09 ± 0.42 \| 73.74 ± 5.37 \| \| \| \| 0.9 \| 30.42 ± 7.05 \| 17.51 ± 2.61 \| 43.31 ± 4.54 \| 46.04 ± 8.35 \| 7.69 ± 0.45 \| 77.60 ± 5.45 \| \| \| halfcheetah \| 0.1 \| 49.31 ± 5.16 \| 33.42 ± 4.70 \| 44.79 ± 5.16 \| 32.85 ± 3.82 \| 7.05 ± 1.47 \| 52.45 ± 3.62 \| \| \| \| 0.5 \| 37.98 ± 3.07 \| 33.36 ± 3.04 \| 41.13 ± 3.53 \| 26.16 ± 4.91 \| 6.14 ± 1.21 \| 55.04 ± 3.27 \| \| \| \| 0.9 \| 15.54 ± 3.06 \| 2.21 ± 1.16 \| 7.28 ± 1.59 \| 8.96 ± 3.26 \| 1.02 ± 1.12 \| 22.28 ± 2.88 \| \| Rebalanced by action \| hopper \| 0.1 \| 29.71 ± 4.00 \| 28.37 ± 1.11 \| 42.29 ± 6.39 \| 25.92 ± 2.45 \| 11.71 ± 2.12 \| 56.60 ± 11.90 \| \| \| \| 0.5 \| 26.35 ± 4.88 \| 30.03 ± 4.23 \| 53.37 ± 9.50 \| 35.13 ± 5.41 \| 11.79 ± 1.14 \| 73.80 ± 3.63 \| \| \| \| 0.9 \| 30.50 ± 3.60 \| 38.92 ± 4.78 \| 63.14 ± 7.12 \| 36.56 ± 2.29 \| 19.42 ± 2.82 \| 48.99 ± 12.27 \| \| \| walker2d \| 0.1 \| 23.61 ± 5.10 \| 12.06 ± 1.20 \| 40.72 ± 1.13 \| 31.18 ± 4.23 \| 7.27 ± 0.46 \| 70.60 ± 3.21 \| \| \| \| 0.5 \| 32.29 ± 6.74 \| 16.40 ± 1.70 \| 47.93 ± 12.36 \| 30.52 ± 3.89 \| 6.37 ± 0.97 \| 72.09 ± 8.70 \| \| \| \| 0.9 \| 16.87 ± 2.80 \| 15.64 ± 3.57 \| 43.68 ± 11.50 \| 37.55 ± 8.97 \| 4.60 ± 0.97 \| 69.51 ± 8.54 \| \| \| halfcheetah \| 0.1 \| 41.91 ± 4.80 \| 26.62 ± 2.54 \| 32.76 ± 1.87 \| 27.50 ± 1.01 \| 8.41 ± 3.38 \| 56.42 ± 4.57 \| \| \| \| 0.5 \| 45.80 ± 4.45 \| 45.52 ± 3.24 \| 48.08 ± 5.50 \| 33.39 ± 6.52 \| 4.64 ± 0.84 \| 60.81 ± 1.56 \| \| \| \| 0.9 \| 25.91 ± 3.35 \| 4.28 ± 1.52 \| 9.57 ± 2.34 \| 12.08 ± 2.01 \| 0.59 ± 0.68 \| 29.19 ± 4.58 \| Table B. Expanded performance comparison on Scenario 1 (rebalanced dataset). (corresponds to Table 3) ## Scenario 2: Time-dependently Subsampled Dataset \| Task \| (a, b) \| BC \| OptiDICE-BC (Soft-$\chi^2$) \| DrilDICE (Soft-$\chi^2$) \| DR-BC \| OptiDICE-BC (Soft-TV) \| DrilDICE (Soft-TV) \| \| --- \| ------- \| --- \| --- \| --- \| --- \| --- \| --- \| \| hopper \| (1, 1) \| 28.89 ± 3.77 \| 50.33 ± 6.60 \| 54.83 ± 7.66 \| 21.10 ± 2.26 \| 22.77 ± 3.94 \| 45.44 ± 5.11 \| \| \| (1, 5) \| 31.03 ± 0.90 \| 39.54 ± 4.02 \| 37.18 ± 7.92 \| 25.00 ± 1.66 \| 19.25 ± 1.21 \| 45.60 ± 4.63 \| \| \| (5, 1) \| 26.75 ± 7.12 \| 48.40 ± 12.98 \| 39.91 ± 9.20 \| 17.51 ± 3.38 \| 25.68 ± 6.01 \| 34.71 ± 9.00 \| \| \| (5, 5) \| 27.65 ± 6.71 \| 32.46 ± 10.79 \| 40.24 ± 7.41 \| 23.20 ± 6.32 \| 14.12 ± 3.59 \| 25.61 ± 6.03 \| \| walker2d \| (1, 1) \| 28.95 ± 5.34 \| 17.42 ± 3.00 \| 51.85 ± 5.30 \| 45.66 ± 9.92 \| 6.13 ± 1.03 \| 81.21 ± 5.40\| \| \| (1, 5) \| 61.48 ± 5.19 \| 37.25 ± 5.66 \| 64.46 ± 7.92 \| 57.29 ± 4.79 \| 17.55 ± 2.65 \| 84.28 ± 4.89 \| \| \| (5, 1) \| 8.13 ± 0.72 \| 4.43 ± 0.91 \| 23.31 ± 3.44 \| 17.97 ± 2.58 \| 4.37 ± 0.52 \| 48.23 ± 8.30 \| \| \| (5, 5) \| 6.65 ± 1.20 \| 8.50 ± 2.27 \| 14.40 ± 3.21 \| 12.45 ± 1.84 \| 5.54 ± 0.68 \| 52.57 ± 6.24 \| \| halfcheetah \| (1, 1) \| 33.74 ± 2.99 \| 33.65 ± 5.49 \| 33.43 ± 3.73 \| 17.09 ± 2.43 \| 9.93 ± 3.56 \| 44.17 ± 5.62 \| \| \| (1, 5) \| 72.72 ± 2.60 \| 52.94 ± 3.98 \| 69.63 ± 3.86 \| 61.81 ± 2.50 \| 24.42 ± 3.27 \| 77.12 ± 2.42 \| \| \| (5, 1) \| 2.35 ± 0.51 \| 2.46 ± 1.05 \| 3.97 ± 1.18 \| 3.81 ± 1.26 \| 1.29 ± 1.13 \| 5.68 ± 1.50 \| \| \| (5, 5) \| 2.01 ± 0.91 \| 1.61 ± 1.30 \| 4.61 ± 1.55 \| 2.85 ± 1.05 \| -1.19 ± 0.37 \| 5.50 ± 0.83 \| Table C. Expanded performance comparison on Scenario 2 (time-dependently collected dataset). (corresponds to Table 4) --- Rebuttal 4: Title: Dear Reviewer zQPx Comment: We kindly remind you that less than 48 hours remain in our discussion period. We are committed to addressing any remaining concerns. In summary, we have addressed your key concerns as follows: - Clarification of Problem Setting: (1) we do not assume that the data distribution $d_D$ is not stationary, (2) we consider shifts of $d$, not $T$. - Comparison with DR-BC [18]: (1) a choice of $f$-divergence for the uncertainty set, (2) our method includes Bellman flow constraints. - Additional Experiments: - DR-BC has been added as a baseline across all experiments. DrilDICE consistently outperforms DR-BC under similar choices of $f$-divergence. - We have incorporated complete trajectory scenarios, demonstrating significant data efficiency in DrilDICE. For further information, please see our rebuttal. If you have any additional questions or concerns, we encourage you to provide your comments. Thank you again for your insightful reviews. --- Rebuttal Comment 4.1: Comment: Thank you for reflecting on the reviews. I have updated my score from 4 to 6. Good luck. --- Reply to Comment 4.1.1: Comment: We sincerely thank the reviewer for the thoughtful review and positive assessment of our work! We believe that your insights clearly enhance the quality of our manuscript. We will incorporate your valuable feedback and suggestions into the next revision.	null	null	Rebuttal 1: Rebuttal: # General Response We are grateful for the insightful and detailed feedback provided by all reviewers. Below, we summarize our response to the main concerns raised. Should any points require further clarification or detailed discussion, we are fully prepared to engage in discussions during the author-reviewer discussion period. ## 1. Comparison with DR-BC [18] Thanks to reviewer zQPx’s suggestion, we found that DR-BC’s objective is also related to our problem setting. In short, DR-BC considers the uncertainty set of arbitrary state distributions by utilizing TV distance. We evaluated DR-BC as a baseline for all scenarios. ## 2. Soft-TV Distance: Technical Adjustment of $f$ for DrilDICE To provide a fair comparison between DrilDICE and DR-BC, we introduced the soft-TV distance into DrilDICE. This relaxed version of the TV distance has an invertible derivative of $f$, enabling DrilDICE to obtain a closed form solution of $w^_{\pi,\nu}$ while maintaining similar properties of TV distance. See Figure A* in the PDF to compare with other $f$-divergence. ## 3. Additional Experimental Results We have expanded our main experiments by incorporating three additional methods: DR-BC, OptiDICE-BC (w\ soft-TV) and DrilDICE (w\ soft-TV) for our main scenarios. Due to space constraints, we omitted the results for DemoDICE, AW-BC, worst-25% score, target 0-1 loss in tables. For the expanded main experiments, please refer to the following: - Four Rooms : Table A in the comment to the rebuttal - Scenario 1 (rebalanced dataset) : Table B in the comment to the rebuttal - Scenario 2 (time-dependently subsampled dataset) : Table C in the comment to the rebuttal - Scenario 3 (segmented trajectory dataset) : Figure B in the PDF Additionally, we conducted experiments in the following additional problem settings: - Complete trajectory setting to address reviewer zQPx’s concern : Figure C in the PDF - Medium-quality segment setting as commented by reviewer k6J7 : Figure D in the PDF In summary, the results of additional experiments consistently demonstrate that using the soft-TV distance as an $f$-divergence significantly improves the performance of DrilDICE, outperforming baselines including DR-BC. Pdf: /pdf/2bf5992fae846df2c7fd35088767c4e8afc13494.pdf	NeurIPS_2024_submissions_huggingface	2,024	null	null	null	null	null	null	null	null
Unitary Convolutions for Learning on Graphs and Groups	Accept (spotlight)	Summary: The paper introduces two unitary graph convolution operators (UniConv and Lie UniConv), and studies their performance and ability to avoid over-smoothing even in deep Graph Neural Networks. UniConv (short for separable unitary convolution) takes the form f_{Uconv} = exp(iAt)XU (with A the adjacency matrix, i the imaginary unity, and U a unitary operator), while Lie UniConv (short for Lie Unitary Convolution) is defined as f_{Uconv} = exp(g_conv)(X) (with g_conv(X)=AXW, and W a skew-symmetric matrix). In both cases, the exponential map is approximated with a truncated Taylor series to contain complexity and avoid any form of eigen-decompositon. In experimental evaluation, GNNs implemented with UniConv achieved good performance on a variety of real-world datasets, while Lie UniConv was the only model to succeed in a toy example exhibiting long-range dependencies when compared to classic GNNs (such as GCN, GAT and Residual GCN). Strengths: The paper is generally well written and with a good introduction exhibiting the oversmoothing effect that one can observe with a classic GCN approach, and how the proposed unitary convolution can avoid that. In terms of the presentation of the methods, I didn’t have particular difficulties following the general idea of the authors, although there are some items I need to clarify with them (please see below in the weaknesses of the paper). Experiments appear overall good (though with some limitations on the Lie Uniconv method), with UniConv achieving good performance on a variety of benchmarks (including datasets showing long range dependencies and heterophilic graphs). Weaknesses: While I generally appreciated the paper (hence my slightly positive score), I’m not fully convinced by the Lie UniConv approach. The authors propose the two methods I highlighted above as a solution to avoid overfitting. As overfitting is caused by a contraction of all frequencies associated to all but the lowest frequency of the Laplace operator, converting the adjacency matrix of the provided graph into a unitary operator (as it is done in UniConv) makes sense to avoid the issue (as all eigenvalues of the obtained diffusion operator now have norm equal to 1). However, the Lie UniConv approach doesn’t go in that direction, but rather applies an exponential map on operator g_conv to obtain an orthogonal map. In this case, why the exponential of such operator is an orthogonal map is not specified in the paper, and I struggle to understand the rationale behind this approach. In this direction, in the experimental evaluation of the paper, I didn’t see any result using Lie Uniconv besides the experiment of the synthetic long rage benchmark I highlighted above. On top to this, on such benchmark, UniConv is not applied in the paper. If this was achieving comparable, or even superior, results to Lie Uniconv, then there would not be any result in the paper showing the benefit of using the Lie Uniconv method. I would kindly ask the authors if they could clarify my doubts above, and present additional results with Lie Uniconv in the rebuttal (unless I missed them somehow in the paper). In addition to this, I would like to point out that the results presented in Table 2 are incomplete, and while the proposed Uniconv outperforms the reported methods, it is possible to see better performing solutions on Roman E in “A critical look at the evaluation of gnns under heterophily: Are we really making progress?”. I would thus ask the authors to report all the relevant results from the paper, to ensure a fair comparison. Technical Quality: 2 Clarity: 2 Questions for Authors: See Weaknesses Confidence: 3 Soundness: 2 Presentation: 2 Contribution: 3 Limitations: Yes Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We thank the reviewer for their feedback and for their insightful questions. We have conducted additional experiments to answer the reviewers' main questions and will also amend the manuscript accordingly. We respond to individual points below.   ___ > While I generally appreciated the paper (hence my slightly positive score), I’m not fully convinced by the Lie UniConv approach. The authors propose the two methods I highlighted above as a solution to avoid overfitting. As overfitting is caused by a contraction of all frequencies associated to all but the lowest frequency of the Laplace operator, converting the adjacency matrix of the provided graph into a unitary operator (as it is done in UniConv) makes sense to avoid the issue (as all eigenvalues of the obtained diffusion operator now have norm equal to 1). This is a neat perspective on the UniConv operator that we appreciate you highlighting. Indeed, one can view this as replacing standard graph convolution which is a dissipative process with a unitary one. Let us now address the point about Lie UniConv. To continue, you write: > However, the Lie UniConv approach doesn’t go in that direction, but rather applies an exponential map on operator g_conv to obtain an orthogonal map. In this case, why the exponential of such operator is an orthogonal map is not specified in the paper, and I struggle to understand the rationale behind this approach. We believe there are two points here that need clarifying. First, we want to clarify what we think is a simple misunderstanding. The operator $g_{conv}$ is a linear map which lies in the Lie algebra of the orthogonal/unitary group. The matrix exponential maps from the Lie algebra to its Lie group, so the exponential map of this operator will be unitary/orthogonal meaning that $\\|g_{conv}(X)\\| = \\|X\\|$ and $g_{conv}$ is invertible. One can easily check that enforcing $W + W^\dagger=0$ enforces that the map $g_{conv}$ is in the Lie algebra. Second, there is a question about the rationale. In practice, Lie UniConv and UniConv are both unitary, but the UniConv operator acts as a tensor product $\exp(iA)\otimes W$ whereas $\exp(g_{conv})$ cannot be written in this tensor product form. The reason we separate these two is because in practice one may only want to change the message passing to be unitary leaving the feature transformation acting as before. Also, the tensor product structure makes implementation faster as one can apply feature transformation and message passing separately in sequence. Remark 2 aims to highlight these points, which we will make clearer in the revised manuscript. To return to your intuition from earlier, both of these will act as diffusion operators that are unitary. The sole difference is in how one wants to transform the feature space of the nodes. For example, one may use Lie UniConv in situations where the 'magnitude' of message passing depends on the features. We also refer the reviewer to the overall response to all reviewers for additional details about this. Since this was also a point of confusion for another reviewer, we will clarify this further in the paper (see overall response for specifics).   ___ > I didn’t see any result using Lie Uniconv besides the experiment of the synthetic long rage benchmark I highlighted above. On top to this, on such benchmark, UniConv is not applied in the paper. If this was achieving comparable, or even superior, results to Lie Uniconv, then there would not be any result in the paper showing the benefit of using the Lie Uniconv method. We thank the reviewer for sharing this feedback which was also shared by other reviewers. There are now a number of updates for this (see overall response with attached pdf). We have added a table comparing Lie UniConv to UniConv on the Peptides and TU datasets including a split for whether one parameterizes in the orthogonal or unitary group. UniConv performs slightly better though the differences are minor. Furthermore, we include a figure showing that Lie UniConv and UniConv are both equally capable of learning the toy model task. In practice, the differences between the two methods can be hard to discern. In real-world experiments, we often find that UniConv often performs slightly better since it addresses oversmoothing issues in the message passing leaving feature transformations unconstrained. Please see the overall response for more details. > I would kindly ask the authors if they could clarify my doubts above, and present additional results with Lie Uniconv in the rebuttal (unless I missed them somehow in the paper). Please see prior response. We would be happy to answer any additional questions or perform additional experiments that could help clarify the contributions of our paper.   ___ > In addition to this, I would like to point out that the results presented in Table 2 are incomplete, and while the proposed Uniconv outperforms the reported methods, it is possible to see better performing solutions on Roman E in “A critical look at the evaluation of gnns under heterophily: Are we really making progress?”. The reviewer is correct in that there are models in “A critical look at the evaluation of gnns under heterophily: Are we really making progress?” that perform better, for example on the Roman Empire dataset. However, these methodologies use different embeddings in their pre-processing. As noted in Appendix F under 'Heterophilous Graph Datasets', "we do not separate ego- and neighbor-embeddings, and hence also do not report accuracies for models from the original paper that used this pre-processing (e.g. GAT-sep and GT-sep)". --- Rebuttal Comment 1.1: Comment: I thank the reviewers for their response (which clarifies some of my doubts) and for the additional results listed in their attached PDF. Unfortunately, there is still no evidence on the benefit of using Lie Uniconv in the paper, which holds me back from increasing my score. Do the authors have any result comparing the speed of the two approaches perhaps? --- Reply to Comment 1.1.1: Comment: Thank you for your comment and quick feedback. Can we ask for a point of clarification? When you state `there is still no evidence on the benefit of using Lie Uniconv in the paper`, are you asking for evidence that this would outperform UniConv? As we state in the draft and in our additional experiments, we do not expect these to perform very differently. These are simply two different means of parameterizing unitary maps. They both nonetheless outperform other non-unitary message passing methods in our experiments.   ___ We are happy to provide empirical runtimes validating that unitary GCN is simply a constant factor overhead over vanilla GCN. See below. Time in seconds for a single forward pass on a single batch of data: ```markdown \| Number of Nodes \| Vanilla GCN \| Spectral GCN \| Gated GCN \| GPS \| Unitary GCN \| \|-----------------\|:-----------:\|:------------:\|:---------:\|:------:\|:-----------:\| \| 200 \| 0.0024 \| 0.0054 \| 0.0028 \| 0.0047 \| 0.0061 \| \| 1000 \| 0.0036 \| 0.0106 \| 0.0064 \| 0.0124 \| 0.0122 \| \| 5000 \| 0.0084 \| 0.0383 \| 0.0331 \| 0.3009 \| 0.0479 \| \| 25000 \| 0.0733 \| 0.2640 \| 0.1541 \| 7.1890 \| 0.1894 \| \| 100000 \| 0.2383 \| 1.1645 \| 1.1916 \| oom \| 1.3550 \| ``` The table above shows runtimes for a forward pass of a model with two convolution layers of the specified form of width 128. As expected, unitary GCN has constant factor overhead over vanilla GCN scaling similarly to other enhanced message passing schemes like spectral GCN and Gated GCN. GPS scales poorly to large graphs and runs out of memory (oom) for the largest graph of 100,000 nodes. Unitary GCN uses UniConv layers though Lie UniConv performed virtually the same in terms of runtime. All experiments are on synthetic data using the layers as implemented in pytorch geometric. Finally, we should note that we did not implement some of the efficiency improving techniques that one can implement in this line of work. For example, there exist approximations other than the Taylor approximation that may perform faster, implementations in the spectral basis can speed up unitary convolution in some instances, and there are other tricks as well (e.g. see projUNN for very fast unitary matrix updates using low rank updates). These are described in detail in the draft in appendix A and C.	Summary: The paper proposes to use so-called "unitary group convolutions" in graph neural networks, with main motivation to do steps to overcome gradient collapse or explosion and oversmoothing effects in graph neural networks under equivariance constraints. This fits within a more general line of work in which unitary constraints are applied for regularizing various learning architectures. Strengths: The use of unitary constraints is a sound idea for use in GNNs, in principle. Edit after rebuttal: It is also a novel idea, different than usual normalizations and ensuring higher nondegeneracy. The explanations are clear and the paper is mostly well written (see "questions" part for some possible doubts). Weaknesses: Edit: after rebuttal, the main issue of originality is 100% solved, and the points in questions 3 and 5 are fully solved. The authors have proposed better experiments. This makes my concern in high part mute, the only small (and out of the scope of this work) point is that it will be great to see more verifications of oversmoothing control of beyond the (already convincing) Rayleigh Quotients preservation, which is proved in the work. As said above, this is for future work, and requires much more space, so the paper is very strong as is. --------------------------- The originality of this approach is a bit limited, since normalizing message passing is not a very novel idea. Doing it with unitary constraints has not been explored before for GNNs. It has been explored in other settings though. The mathematical setup is sketchy at times, it needs better care. In particular, it is not clear if the proposed convolution is really a unitary operation (see question 3 and the crucial question 5 in "questions" section, especially). The experiments show that normalizing is a good idea, but they don't fully convince that unitary constraints is the best way to normalize, since comparisons are done mainly against non-normalized GNNs. One has the impression that this is mainly a first exploration of a theme, and it would need more careful benchmarking, since the computational cost for implementing this particular normalization is not always negligible. Technical Quality: 4 Clarity: 3 Questions for Authors: What's the relative benefit compared to more naive normalizations? Minor correctionsl/doubts: 1) lines 29-31: "it has been widely observed that group-convolutional networks suffer from instabilities as their depth increases" -- if it's widely observed, please indicate a few references where this was observed? 2) line 103: "linear convolution" -- as opposed to what other kind of convolution? isn't convolution always linear? 3) line 144: the following is an important point/issue. "When W contains only real-valued entries, the above returns an orthogonal map" -- I don't follow this, why is that? this statement would require a proof, because it far from obvious, and in general it is false. Recall that a product of matrix exponentials is in general not the exponential of the product, so $exp (g_{\mathrm{conv}})(X) \neq e^A X e^{W^*}$ in this case. 4) I don't see what's the importance / relevance of examples 1 and 2 from Section 3. I think these are standard examples, they can be skipped or moved to the appendix. 5) lines 217-218: this probably the most important theory issue of the paper. "the Rayleigh quotient is invariant to unitary transformation giving a proof that unitary graph convolution avoids oversmoothing" One can vaguely agree that Rayleigh quotient collapse may in some way be correlated to oversmoothing, but saying that Prop 6 "gives a proof that unitary convolution avoids oversmoothing" is unjustified without further proofs/details. In particular some form of a result saying that "Rayleigh quotients do not collapse implies there is no oversmoothing" is missing. Without this, the main claim of the paper hinges on "vague intuition" about RQ's and the paper is much less valuable at a theory level. 6) Definition 6 (dynamical isometry): what iis the connection of this definition with the definitions from section 3.1? This part on dynamical isometry is at present very weakly (if at all) connected to the core of the paper, and I did not understand how the authors are linking it to the main definitions. See question 3 related to this. 7) The experiments show performance on some benchmarks, but I would have been more convinced if the unitary convolution proposed method would have been compared to other simpler normalization methods which have similar effects on oversmoothing and gradient vanishing/explosion. Confidence: 5 Soundness: 4 Presentation: 3 Contribution: 4 Limitations: Main limitations are already included in the "questions" section, esp questions 3,5,7. Flag For Ethics Review: ['No ethics review needed.'] Rating: 8 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We thank the reviewer for their feedback and questions. As an overall point, the reviewer made many comments about normalized convolution which we were unsure of the specific message. We kindly ask the reviewer to clarify. ___ > The originality of this approach is a bit limited, since normalizing message passing is not a very novel idea. Doing it with unitary constraints has not been explored before for GNNs. Unitary layers normalize the 2-norm, but we politely disagree that they are simply a type of normalization as they also guarantee invertibility, lack of oversmoothing, stable Jacobian, etc. > What's the relative benefit compared to more naive normalizations? As stated above, we are not sure what a "naive normalization" is referring to. We believe this may refer to other interventions listed below. We stress that that unitary layers are not competing with these, but complement them: any architecture can be implemented in tandem with such techniques. 1. batch norm: batch norm helps normalize layer outputs, but does not have useful properties of unitary layers listed above, e.g. vanishing/exploding gradients, over-smoothing, and stability. 2. regularizers and feature adjustments: There are methods to constrain, regularize or adjust features or layer outputs to make neighboring feature vectors more dissimilar and mitigate oversmoothing. These include Dirichlet energy constraints (e.g. Zhou et al. Dirichlet energy constrained learning...), Pairnorm (Yang et al. Pairnorm...), and others. These may help over vanilla methods in addressing oversmoothing though their benefit is arguably limited -- see survey of [RBM23] and comments below. 3. normalizing adjacency matrix: One typically normalizes adjacency matrices by multiplying it by $D^{-1/2}AD^{-1/2}$ where $D$ is a diagonal matrix with $i$-th entry as the degree of node $i$. This is by default used in message passing architectures, and we use it too. This to us is a separate topic from how to perform message passing on an adjacency matrix. 4. residual connections: Though not a type of normalization, residual connections have been proposed to help avoid oversmoothing. We experimentally compare to this intervention in our experiments and also give a theoretical argument for why residual connections alone can only alleviate the problem (see Proposition 4). For more detail, consider a concurrent paper [SWJS24] using feature normalizations and residual connections. They have weaker empirical results on datasets such as Mutag and Proteins with their normalization, GraphNorm, and residual connections. We can include more details if this is what the reviewer means by "naive normalization". ___ Other questions/concerns: > if [instabilities with depth] is widely observed, please indicate a few references where this was observed? Instabilities come in various forms and have inspired various architectectural changes, including for CNNs and GNNs (see related work). We have also included some examples. Figure 4 in the appendix shows an example of instabilities with increasing depth for several message-passing GNNs. The attached one page response also shows this issue with vanishing/exploding gradients. > "linear convolution" -- as opposed to what other kind? isn't convolution always linear? Linear here is linearity in the input $x$. Some convolutions are not of this form, e.g. steerable convolution acts over a set of potentially nonlinear filters. > "When W contains only real-valued entries, the above returns an orthogonal map" -- I don't follow this, why is that? this statement would require a proof, and in general it is false... $\exp(g_{conv})(X)\neq e^{A} X e^{W^}$ in this case. We think there is a simple misunderstanding. This fact is true for the simple reason that $\exp(W)$ is real-valued whenever $W$ is real-valued (the exponential map cannot return complex numbers when its input is real-valued). So the output is always orthogonal provided $W$ meets the Lie algebra criteria of eq. (7), namely $W+W^T=0$. We also do not claim that $\exp(g_{conv})(X)=e^{A} X e^{W^}$. In general, it will not take this tensor product form. Perhaps the reviewer can clarify further their concern? > I don't see what's the importance / relevance of examples 1 and 2 from Section 3. I think these are standard examples, they can be skipped or moved to the appendix. Please see general response. We will shorten section 3.2 by moving the discussion of the Fourier/spectral based methods to the appendix. > Saying that Prop 6 "gives a proof that unitary convolution avoids oversmoothing" is unjustified without further proofs/details... We will be more precise stating that UniConv provably avoids oversmoothing as measured by the Rayleigh Quotient or Dirichlet Energy, but we push back that this is simply "vague intuition". The Rayleigh quotient is a widespread measure of oversmoothing (see references below). Beyond this, unitary message passing is invertible and isometric, two properties that ensure signals always propogate through a network without getting "lost". > [vanishing/exploding gradients and dynamical isometry]: what is the connection of this definition with the definitions from section 3.1? Due to lack of space available here, please see response to reviewer vJVB asking the same question.   References: [AL24] Roth, Andreas, and Thomas Liebig. Rank Collapse Causes Over-Smoothing and Over-Correlation in Graph Neural Networks. PMLR, 2024. \ [RBM23] T Konstantin Rusch et al. A survey on oversmoothing in graph neural networks. arXiv:2303.10993, 2023 \ [RCR+22] T. Konstantin Rusch et al. Graph-coupled oscillator networks. PMLR, 2022. \ [SWJS24] Scholkemper, Michael, et al. "Residual Connections and Normalization Can Provably Prevent Oversmoothing in GNNs." arXiv preprint arXiv:2406.02997 (2024). \ [WAWJ24] Wu, Xinyi et al. "Demystifying oversmoothing in attention-based graph neural networks." NeurIPS (2024). --- Rebuttal Comment 1.1: Comment: I'll slowly go through the points of your rebuttal, so this may not be my only comment (sorry). First of all, thanks for pointing out the difference to unitary layers, I was wrong on the novelty part, I apologize and I am grateful that you replied in detail. Reflecting upon this error of mine and your clarification, I now think that actually you could prove even stronger results about oversmoothing prevention by your approach, than the one on Rayleigh Quotients, but due to heavy review overload in 1 week time, I don't have a proposal right now, and that's anyway material for future work. About my correction/doubt number 3: when $W$ is an antisymmetric real matrix, $\exp(W)$ is orthogonal, we agree this far. But in line 144 I'm concerned that you say "the above outputs an orthogonal map". In that comment, you refer to $exp(W)$ or to $f_{Uconv}(X)$? Given that $f_{Uconv}$ has a "U" in it, and given this paper's title, I assume it's $f_{Uconv}$. How does "exponential of real-valued antisymmetric implies orthogonal" apply to $f_{Uconv}(X)=\exp(g_{conv}(X))=\exp(AXW)$ ? are you using the fact that $AXW$ is antisymmetric real-valued too if $W$ is? I don't follow.. I'd like to see a proof of the map being orthogonal. Thanks for the detailed rebuttal so far. --- Reply to Comment 1.1.1: Comment: Thank you for your quick response and especially for recognizing the novelty of our work. We would be happy to hear about your thoughts on how we can prove stronger results on oversmoothing if you are willing to share it. ___ Regarding your question, yes we are referring to $f_{Uconv}$. Let us go through this in more detail. As a reminder, we have that $f_{Uconv}(X)=\exp(g_{conv})(X)$ and $g_{conv}(X) = AXW$ and for an undirected graph $A = A^\dagger$ (for directed graphs, see Appendix D for proper handling of this). Furthermore, we constrain $W$ so that $W + W^\dagger=0$. Treating inputs and outputs as vectors, we can equivalently view $g_{conv}$ as acting on a vector rather than a matrix: $$ g_{conv}(X) = AXW \leftrightarrow vec( g_{conv}(X) ) = (A \otimes W^T) vec(X) , $$ where we vectorize $X \in \mathbb{C}^{a \times b}$ with the command $vec(X) \in \mathbb{C}^{ab}$ that stacks columns on top of each other. Now, note that $$ (A \otimes W^T) + (A \otimes W^T) ^\dagger = A \otimes (W + W^\dagger)^T = 0, $$ and thus $(A \otimes W^T)$ is in the Lie Algebra of the Unitary/Orthogonal group. Finally, we now have $$vec(\exp(g_{conv}(X))) = \exp(A \otimes W^T) vec(X),$$ and thus $\exp(A \otimes W^T)$ is in the Unitary/Orthogonal Lie group. Hopefully, viewing it this way in the "vector sense" clarifies things. Let us know if additional details are requested. We will clarify this in the updated manuscript. --- Rebuttal 2: Comment: Ok, thanks for the detailed reply. I hope it won't bother you if I continue a bit with the clarifications on the same topic (sorry if it does): 1) (minor point) So if $A=A^T$ and $W=-W^\dagger$ then these are square matrices, and then $X$ must be a square matrix, right? Or how can I compute $\exp(AXW)$ if $AXW$ is not a square matrix? (I'm confused by the notation $X\in \mathbb C^{a\times b}$, is it that necessarily $a=b$?) 2) Assume now that $A=Id$, and $X,W$ are real-valued square matrices of equal dimension $n$, of which $W=-W^T$, to simplify things. Say $A=Id$, and $W=\left[\begin{array}{cc}0&-1\\\\ 1&0\end{array}\right]$ and $X=\left[\begin{array}{cc}0&1\\\\ 0&0\end{array}\right]$. Then $AXW=\left[\begin{array}{cc}1&0\\\\ 0&0\end{array}\right]\neq -(AXW)^T$ right? Then $\exp(AWX)$ is not unitary or orthogonal. I am confused, so maybe clarifying what I misunderstood with this example can help. --- Rebuttal 3: Comment: No problem; we are happy to clarify. Let us answer your questions in order. 1. Formally, unitary transformations must be invertible so the input and output dimension must be equal. Thus the row and column dimensions of $W \in \mathbb{C}^{b\times b}$ must be equal and similarly for $A \in \mathbb{C}^{a \times a}$. However, this does not mean that $X$ has equal row and column dimension. The map $f_{Uconv}$ maps matrices of size $a \times b$ to matrices also of size $a \times b$ so dimensionality is preserved in this sense. As noted in the draft though, this does not limit implementation: for $W$ with arbitrary row and column dimension, one can also work within the Stiefel manifold or even more simply pad inputs and outputs so that dimensionalities are equal to actually implement things in practice (see appendix D). 2. It does not need to hold that $AXW =- (AXW)^\dagger$. In fact, this will not hold true in general since the row and column dimensions of $X$ are not equal so this equation formally may not even make sense. What does hold true however is that the linear operator $g_{conv}$ is in the Lie algebra. As a reminder, exponential map is equal to $$\exp(g_{conv})(X) = \sum_{k=0}^\infty \frac{g_{conv}^{(k)}(X)}{k!} = \sum_{k=0}^\infty \frac{A^kXW^k}{k!}. $$ Viewed in the vectorized form from our previous comment this is equivalent to: $$vec[\exp(g_{conv})(X)] = \sum_{k=0}^\infty \frac{(A \otimes W^T)^k}{k!} vec[X],$$ so as stated earlier, one can check that $A \otimes W^T$ is in the Lie algebra to ensure unitarity. Please let us know if we can clarify anything further. --- Rebuttal Comment 3.1: Title: All doubts clarified, thanks Comment: I understand now. I was interpreting $g_{conv}^k(X)$ as $(g_{conv}(X))^k$ and now that I noticed the difference everything is clear. My main concerns have been all lifted this is the case for concern for novelty (which is now "reverted": I believe the paper has strong novelty), and the topics of questions 3 and 5 (technical points). The concern that Rayleigh Quotients may be only one of several possible measures for oversmoothing is still valid for me, but I think it can/should be considered in future work, and due to understanding the difference between normalizing 2-norm and imposing unitary transformations, I am confident in the result being "fortifiable". What I mean is mainly to prove nondegeneracy according to more of the used metrics for oversmoothing. Given the above, I think it's fair to raise my score to 8. --- Reply to Comment 3.1.1: Comment: Thank you once again for your valuable feedback and the constructive discussion. We appreciate your decision to raise your score to an 8. The score change does not appear on our side. Would you mind verifying that the updated score has been recorded in the system?	Summary: In this paper the authors propose a mathematically consistent treatment of unitary convolutions in the context of neural networks defined on graphs and groups. In the graph neural network context, the main idea is to transform a standard linear convolutional layer $f_\text{conv} = \mathbf{AXW}$ with $\mathbf{A}$ the adjacency matrix, $\mathbf{X}$ the signal input matrix and $\mathbf{W}$ the parameterize into a linear layer $f_\text{Uconv}$ that is unitary (norm preserving), taking into account both the matrices $\mathbf{A}$ and $\mathbf{W}$ (previous literature focus on $\mathbf{W}$ or perform this in a less natural way (with worse computational complexity) [QBY24]. The main idea revolves around the following two steps: (1) project the operation into matrix Lie algebra (e.g., the space of skew-symmetric matrices) (2) apply the exponential map on the projected operation to obtain an unitary operation. In the graph setting the authors do this in two ways: the first is called "Separable unitary graph convolution (UniConv)" and takes the form $f_{\text{Uconv}} (\textbf{X}) = \exp(i\textbf{A}𝑡)\textbf{X}\textbf{U}, \textbf{U} \textbf{U}^{\dagger} = \textbf{I}$. The second is called "Lie orthogonal/unitary graph convolution (Lie UniConv)" and is defined by applying an exponential map on a convolution with skew-symmetric weight matrices in Lie algebra, having the advantage of being parameterizable over the reals only. The authors then discuss how to approximate the exponential map (they resort to simple Taylor truncation). Subsequently, the authors develop the theory in the case of general groups and show to ways of obtaining unitary convolutions: the first case generalizes the two methods defined for graphs, by first skew-symmetrizing the equivariant map and then applying exp; the second is defined as a (block-diaognal) convolution in the Fourier domain followed by inverse transform. Following the authors show basic properties of the proposed method (invertibility, isometry and equivariance) and prove favorable results which mainly follow by isometry (avoiding oversmoothing, avoiding vanishing/exploding gradients). In the experimental section, the authors perform experiments on toy ring graph (on Lie UniConv) and the standard benchmarks LRGB and Heterophilous Graph Dataset (on UniConv) (also showing group convolution in Appendix on Dihedral group $D_n$). Strengths: - Well founded mathematical formalism, defining unitarity in message passing graph neural networks in a natural way. - Strong performance over other message passing algorithms, comparable performance with respect to graph attention based models. On very large graphs attention could be prohibitive so UConv could perform well in that situation. - The paper has an extensive and complete literature overview, comparing with related literature in very detailed manner (the Appendix on related work relative to unitary convolutions is a very good addition and clearly let us understand the contribution). - Theoretical analysis on oversmoothing and vanishing/exploding gradients shows the advantage on adopting UConvs in a quantitative way. Weaknesses: - There aren't many experiments on Lie UniConv. From what I understand, the only experiment with this layer is the graph ring experiment in Figure 2. Also it would have been interesting to see comparison between the two type of convolutions (UniConv vs Lie UniConv), with the more realistic datasets, in order to asses which one performs better in practice. - The part on unitary maps in Fourier basis feels somehow detached from the general discourse of the paper. I believed the author added it for completeness but (1) the theory to understand it is relatively advanced (had to study externally some representation theory to understand well the definition of Fourier transform on finite groups); a short primer on representation theory in the Appendix, similar to the one for Lie groups / algebras, would have been of great help. (2) There are no experiments in this setting. (3) some parts are a little bit shaky, for example in: ```One can also generally implement convolutions in the (block diagonal) Fourier basis of the graph or group (Algorithm 2). Here, one employs a Fourier operator which block diagonalizes the input into its irreducible representations or some spectral representation.```, if one considers using the eigendecomposition of some graph Laplacian as a spectral basis, the (graph) Fourier transform is not given by a set of matrices varying over irreps but a signal of the same dimensionality of the input function. I do not think one can use the same approach there. - Sometimes authors leave facts implicit and the writing could be improved in key places. For example I had difficulty understanding why Lie UniConv returns an unitary map before looking at the tensor product version of $g_{\text{conv}}$ (Eq. (36), Proposition 6). Also the statements in Propositions 4 and 7 are not really clear from a first reading, some more intuition about the statements should be given. Technical Quality: 4 Clarity: 3 Questions for Authors: - I don't see how the graph convolutions defined are particular cases of the group convolutions described. Is it the symmetric group S(n) acting on the graph? - Line 122: Why UniConv has a "tensor product nature"? From what I understand the tensor product is more related to the Lie UniConv because you can write it as $\exp(\mathbf{A} \otimes \mathbf{W}^T)$. - Line 186: "We set input/output representations to be equal so that the exponential map of an equivariant operator is itself equivariant": Assuming this does not limit the class of groups or transforms we can use? Not all groups have irreps of same dimensions. Or we say that the transforms are equivariant only in this sense? Also why exponential map is equivariant only if the input output representations are equal? - Line 212: is (I − A) a type of graph Laplacian? - Line 272: Why authors don't compare to new architectures as well on this task, like in Table 1? Also I see in Appendix that GPS it is performing pretty well w.r.t. Unitary GCN but it was not put in this graph (maybe to show better improvements?). - Line 763: I think authors should say that in case of Algo 2 this follows because Fourier transform is unitary. It is not clear if and why this should be the case in finite group setting. - Line 766: "For unitary convolution in the Fourier domain (Algorithm 2), equivariance follows from convolution theorems where linear operators appropriately applied in the block diagonal Fourier basis are equivariant". This is too succinct and should be more expanded formally. - Line 821: missing trace on following lines and last line should not be put there ($= R_{\mathcal{G}(\mathbf{X})}$) Confidence: 4 Soundness: 4 Presentation: 3 Contribution: 4 Limitations: The authors have discussed limitations of their contribution adeguatly. Flag For Ethics Review: ['No ethics review needed.'] Rating: 8 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We thank the reviewer for their positive comments and going through the paper in detail. There are many insightful questions and points of feedback that we will incorporate into the draft. ___ > There aren't many experiments on Lie UniConv... it would have been interesting to see comparison between the two We agree with the reviewer on this point and have included many comparisons in the global response and attached one page pdf there. This includes experiments for Peptides and TU datasets. In general, the empirical differences are rather minimal. > The part on unitary maps in Fourier basis feels somehow detached from the general discourse of the paper... the theory to understand it is relatively advanced... a short primer on representation theory in the Appendix... would have been of great help. There are no experiments in this setting. We aimed to show that Fourier space can often speed up operations and simplify convolution as observed in many works. Since this is tangential to the main points, we'll move it to the appendix with a detailed background. In the main text, we'll briefly note that a Fourier basis methodology exists. > The part on unitary maps in Fourier basis feels somehow detached... if one considers using the eigendecomposition of some graph Laplacian as a spectral basis, the (graph) Fourier transform is not given by a set of matrices varying over irreps but a signal of the same dimensionality of the input function. We believe there is a small misunderstanding here. One can diagonalize the adjacency matrix of an undirected graph and perform convolution over the spectral basis as in spectral graph convolution. When the input and output is a single channel, this can be treated as block diagonal operations, but the blocks are trivially scalars ($1 \times 1$ matrices). More generally, the block size will depend on the input/output channel dimensions. We will provide a lemma explicitly showing this in the more general case for arbitrary graphs/groups. > Sometimes authors leave facts implicit... I had difficulty understanding why Lie UniConv returns an unitary map before looking at the tensor product version of (Eq. (36), Proposition 6). Also the statements in Propositions 4 and 7 are not really clear from a first reading, some more intuition about the statements should be given. We agree this was slightly confusing on a first read. We will revise these passages for clarity in the updated manuscript.   ___ Questions: > I don't see how the graph convolutions defined are particular cases of the group convolutions described. Is it the symmetric group S(n) acting on the graph? To map graph convolution to this case where $conv_G(X) = \sum_i T_i X W_i$, simply set $T_i= A^i$ or the $i$-th power of the (potentially normalized) adjacency matrix. We will explicitly state this after the equation in the main text. > Line 122: Why UniConv has a "tensor product nature"? From what I understand the tensor product is more related to the Lie UniConv because you can write it as $\exp(A \otimes W^T)$. The reviewer is right that Lie UniConv is parameterized in a tensor product fashion, but crucially, the output is not a tensor product: there may not exist matrices $M_1, M_2$ such that $\exp(A \otimes W^T) \neq M_1 \otimes M_2$. In contrast, UniConv is always of the form $\exp(iA)\otimes W^T$ and one can apply the feature and message passing transformations separately and in sequence. We tried to capture this in remark 2, but will clarify in the main text. > Line 186: "We set input/output representations to be equal so that the exponential map of an equivariant operator is itself equivariant": Assuming this does not limit the class of groups or transforms we can use? Not all groups have irreps of same dimensions... This statement is written confusingly and we will clarify. What we meant to say is that the input/output representations need to act on the same vector space. For a unitary transformation to be well defined, it needs to be invertible and isometric so the input and output vector space are must have the same dimension. One can generalize beyond these assumptions, but it seems unnecessary for our purposes. > Line 212: is (I − A) a type of graph Laplacian? That is correct. > Line 272: Why authors don't compare to new architectures as well on this task, like in Table 1? Also I see in Appendix that GPS it is performing pretty well w.r.t. Unitary GCN but it was not put in this graph (maybe to show better improvements?). We show GPS for sake of completeness, but did not include it in the main plots because it is not a message passing architecture. The goal of the toy model task was to see which architectures can learn long range dependencies with local operations. For transformers, the notion of long range dependency is not as well defined. Is there another method that the reviewer feels we should compare to? We are happy to include it. > Line 763: I think authors should say that in case of Algo 2 this follows because Fourier transform is unitary. > Line 766: "For unitary convolution in the Fourier domain (Algorithm 2), equivariance follows from convolution theorems...". This is too succinct and should be more expanded formally. Yes, we will say that and expand on this. Thank you for pointing it out. > Line 821: missing trace on following lines and last line should not be put there ($=R_G(X)$) We think there is a misunderstanding. The following lines have the notation $vec(X)^\dagger ... vec(X)$ which is an equivalent way of writing the trace as an inner product between vectors since $Tr(A^\dagger B)= vec(A)^\dagger vec(B)$ where $vec$ vectorizes the matrices. We will use inner product angle bracket notation to make this clearer. The last line also needs the normalization by the Frobenius norm which we will include as a separate line. Thank you for pointing this out. --- Rebuttal Comment 1.1: Comment: I thank the authors for addressing different points such as additional experiments on Lie UniConv and improving the clarity of the paper where I have pointed out different misunderstandings. The papers is a solid contribution and should definitely be accepted.	Summary: This paper introduces convolutional layers for GNNs that are based on group-convolutions. Some of the proposed layers have interesting theoretical properties such as avoiding oversmoothing or vanishing gradients. Experimentally, these layers are able to aggregate information across long distances and achieve strong results on real-world benchmarks. Strengths: - (S1 - Novelty) The contributions in this paper seem novel. - (S2 - Significance) From a theory perspective, some of the proposed operations have interesting theoretical properties such as that they can avoid oversmoothing or vanishing gradients. - (S3 - Significance) The authors perform extensive experiments on real-world datasets and achieve strong results on most of them. Weaknesses: - (W1 - Clarity and Significance) The reader is not guided through the paper and it is difficult to understand the flow of ideas. I will give some examples of this: - In Section 3.1 (Lie) UniConv layers are introduced which is the main contribution of the pape. However, to me it was not clear that this is the central definition. - In the following Section 3.2, the paper generalizes the setting of unitary convolutions and gives several examples. Yet, to me it is unclear what this section intends to communicate and how it is relevant to the rest of the paper. - Another example of this is the “Vanishing/Exploding gradients”-part of Section 4, here dynamical isometry is defined as a measure as a property to analyze exploding gradients and an example is given that a combination UniConv together with a yet completely undefined activation function (GroupSort) is “perfectly dynamically isometric”. This raises many questions to the reader: Why is an example important enough to warrant the definition of Dynamical isometry? How does this relate to the main contribution of the paper? Are vanishing / exploding gradients a problem with GNNs? What is GroupSort and is it actually used in practice? - (W2 - Clarity) The experiment section is difficult to read and does not properly explain the experiments or the design decisions. For example, for the graph distance dataset the model is not mentioned in the text. In this case, the model is GCN with the novel Lie UniConv layer (Figure 2). However, it is not explained why the Lie UniConv layers are used and not the UniConv Layers. Similarly, for the experiments on LRGB UniConv layers are used but it is not explained why Lie UniConv layers are not tried. To sum up. To me it seems like this could be a good paper with interesting ideas and solid experiments. However, the writing significantly holds it back and I think more time is required to allow the authors to rewrite their paper. Furthermore, I think that there is a lot of useful information hidden in the appendix and it might make sense to focus the paper on the main idea (Section 3.1) and move other parts to the appendix to make more space. As I do believe that this paper requires significant changes to the text I thus vote to reject. Technical Quality: 2 Clarity: 1 Questions for Authors: Most of my questions are already part of the weakness section. - (Q1) (Lie) UniConv are implemented as a $k$-th taylor approximation. What is the value for $k$? It seems to me like the ability to model long-range interactions is primarily dependent the value of $k$ and the number of layers $L$, so knowing this value seems important. Confidence: 3 Soundness: 2 Presentation: 1 Contribution: 2 Limitations: Yes. Flag For Ethics Review: ['No ethics review needed.'] Rating: 4 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We thank the reviewer for praising the novelty and theoretical results in our work. We have read their concerns about the format and writing of the paper, and hope to address them below. ___ > In Section 3.1 (Lie) UniConv layers are introduced... it was not clear that this is the central definition. Can the reviewer expand on what they mean by central definition? From our understanding, the definitions of UniConv operations are central to the paper. > The paper generalizes the setting of unitary convolutions... it is unclear what this section intends to communicate... Our goal was two-fold: to show that unitary convolution on graphs is a simple method applicable to general symmetric domains, and to connect it to previous CNN implementations, demonstrating how to generalize beyond graph convolution, such as applying unitary convolution to hypergraphs or other message-passing types. That being said, we understand the reviewer's concerns. We will shorten this section in the paper focusing specifically on the points above. Please see our general response for details. > Dynamical isometry is defined as a measure as a property to analyze exploding gradients and an example is given that a combination UniConv together with a yet completely undefined activation function (GroupSort) is “perfectly dynamically isometric”... Why is an example important enough to warrant the definition of Dynamical isometry? > Are vanishing / exploding gradients a problem with GNNs? Vanishing/exploding gradients are a general problem with any deep or recurrent network and occur because norms of the states at intermediate layers grow or decay exponentially with depth (see e.g. [LMTG19], [LMGK21]). GNNs are no different from deep networks in that they also have weights initialized randomly causing norms to grow or decay as a product of the number of layers. We found that standard GNNs cannot train at very large depths (see figure 4 showing standard message passing schemes do not train well at around 40 layers for example). We also show an example of a standard GCN featuring vanishing/exploding gradients in the one page response pdf (UniConv does not have this issue in contrast). For scalability to many layers, stable information and gradient propagation is essential. Dynamical isometry offers a rigorous framework to study this stability, supported by experiments and theory across various architectures. We aim to show that GNNs also exhibit the stability properties described by dynamical isometry. > What is GroupSort and is it actually used in practice? Groupsort is a nonlinearity that splits the input vector into pairs of numbers and sorts the pairs (see eq. 58). It is a nonlinearity that preserves isometry and used in the adversarial learning literature to get provable adversarial bounds via unitary architectures. E.g., see its use in Trockman & Kolter Orthogonalizing convolutional layers with the cayley transform. > For the graph distance dataset the model is not mentioned in the text. In this case, the model is GCN with the novel Lie UniConv layer (Figure 2). However, it is not explained why the Lie UniConv layers are used and not the UniConv Layers. Similarly, for the experiments on LRGB UniConv layers are used but it is not explained why Lie UniConv layers are not tried. Please see global response and additional figures/tables for a full answer to this. In summary, this choice did not make a significant difference. For experiments where networks were not very deep, we found that enforcing unitarity in the message passing was helpful and unitarity was not needed in the feature transformation. For example, UniConv performs slightly better in LRGB though little difference is observed. > To me it seems like this could be a good paper with interesting ideas and solid experiments...I think more time is required to allow the authors to rewrite their paper... there is a lot of useful information hidden in the appendix and it might make sense to focus the paper on the main idea (Section 3.1) and move other parts to the appendix to make more space. If we understand the reviewer correctly, they have requested that we focus more on section 3.1, defer parts of section 3.2 to appendix, and make more space for clarifying some points such as when to use Lie UniConv vs. UniConv. We will make these changes in the revised manuscript. Are there other parts of the paper that the reviewer feels should be moved from the appendix to the main text or vice versa? We are hesitant to make significant changes to the writing and format given this concern was not shared by all the reviewers, but are of course open to more feedback. > (Lie) UniConv are implemented as a k-th taylor approximation. What is the value for k?... knowing this value seems important. We found in all our experiments that $k=10$ suffices. The error in the Taylor approximation exponentially decreases with $k$. This value is in line with that in other papers (e.g. see [SF21]). We will state this in the main text. ___   Dynamical Isometry References: [SMG13] Saxe et al. Exact solutions to the nonlinear dynamics of learning in deep linear neural networks. arXiv:1312.6120, 2013 \ [XBSD+18] Xiao et al. Dynamical isometry and a mean field theory of cnns: How to train 10,000-layer vanilla convolutional neural networks. ICML, 2018 \ [PSG18] Pennington et al. The emergence of spectral universality in deep networks. PMLR, 2018   Vanishing/ Exploding Gradient Problems with GNNs: In addition to above, see also: [LMTG19] Li et al. Deepgcns: Can gcns go as deep as cnns?. Proceedings of the IEEE/CVF international conference on computer vision, 2019. \ [LMGK21] Li et al. Training graph neural networks with 1000 layers. PMLR, 2021. --- Rebuttal Comment 1.1: Comment: Dear authors, I am sorry for the delay, I submitted a reply with the wrong reader-group. You should now be able to see my reply above. --- Rebuttal 2: Comment: I thank the authors for their thorough rebuttal. The additional experiments do strengthen the paper and including the results for both architectures does in my opinion simplify the experiment section. > Can the reviewer expand on what they mean by central definition? From our understanding, the definitions of UniConv operations are central to the paper. Yes, they are. However, for some reason this was not immediately clear for me when reading this section. Maybe a bit more guiding of the reader is required. > If we understand the reviewer correctly, they have requested that we focus more on section 3.1, defer parts of section 3.2 to appendix, and make more space for clarifying some points such as when to use Lie UniConv vs. UniConv. We will make these changes in the revised manuscript. Are there other parts of the paper that the reviewer feels should be moved from the appendix to the main text or vice versa? You understand me correctly, I think that moving 3.2 to the appendix is a good idea. This should allow you to focus more on your essential results (key definitions + theorems/propositions) and intuitively explain what they mean and why they matter. > We are hesitant to make significant changes to the writing and format given this concern was not shared by all the reviewers, but are of course open to more feedback. I agree with this sentiment. While I do think that my issues with the presentation / clarity persist (and this is somewhat mirrored by feJe), I cannot ignore that this does not seem to be an issue for the other reviewers. Furthermore, upon a further look I have to concede that this paper does contain many non-trivial theoretical insights. I will slightly increase my score but adjust my confidence downward.	Rebuttal 1: Rebuttal: We thank all the reviewers for their insightful comments, recommendations, and feedback.   There were some common themes in the reviews which we want to address in an overall response here. We have also included a one page pdf containing plots and tables which help address reviewers' concerns. 1. Lie UniConv experiments: Reviewers requested additional experiments comparing Lie UniConv to UniConv. We originally did not include these as they performed similarly, and apologize for the oversight. We include in the one page pdf experiments comparing Lie UniConv and UniConv (both with real-valued orthogonal and complex-valued orthogonal/unitary matrices) for the Peptides experiments and TU datasets. Overall, performance is very similar for Lie UniConv and UniConv. We also include a figure showing that UniConv and Lie UniConv are equally effective at learning the toy model task of graph distance. All results will be included in the revised version of the paper. 2. Vanishing/exploding gradients: Reviewers vJVH and EVha questioned the relevance and importance of vanishing/exploding gradients. In response, we have included a simple plot showing vanilla GCNs are no different than other deep networks and can suffer from this issue. This plot also shows unitary GCNs avoid this issue. This plot was created using standard initialization and settings of the graph convolution layer in pytorch Geometric. Of course, vanishing/exploding gradients are a common problem accross deep networks of all kinds, and unitary layers are not the only way to help alleviate this problem. 3. Generalized and Fourier basis convolution: Section 3.2 generalizes unitary graph convolution to convolution over arbitrary finite groups and in the Fourier/spectral basis. While we believe that this more general perspective is instructive to the reader in that it gives background on other attempts at unitary/orthogonal convolution, we agree that the discussion of the Fourier/spectral based methods in the main paper distracts from the main contributions of the paper. Based on your detailed comments and feedback, we have decided to make the following changes to the presentation: We will defer the description of Fourier/spectral basis convolution to the appendix and include a richer background into representation theory to make it more accessible to the reader. We will shorten the discussion in the main text to a paragraph to informally present potential advantages of implementation in the Fourier domain and note that we did not implement this in our work. These changes will give us space to include additional experimental results for Lie UniConv.   *Please see attached one page pdf with additional tables and plots below.* Pdf: /pdf/8f111e73082bcb4cbdb8657f4134fff46c2f02c3.pdf	NeurIPS_2024_submissions_huggingface	2,024	Summary: This paper proposed a new convolution operator UniConv on graphs and general tasks. Both graph convolution and general form convolution form are provided based on the group theory. With the toy experiment, it shows the advantages of the proposed method clearly. Based on both theoretical and experimental analysis, the proposed network is effective in long range graphs and heterophilic graphs. Strengths: 1. Theoretical analysis is conducted to show the effectiveness in handling oversmoothing and oversquashing issue. 2. The toy analysis is easy to understand and clearly shows the advantage. Weaknesses: 1. Though the general form of convolution is provided. It is not evaluated in experiment. I will suggest putting these content in the appendix in this case. Otherwise, it is better to provide the corresponding experimental results. 2. The computational cost is higher compared to other methods, but there are no experimental results to show the important details. Technical Quality: 3 Clarity: 3 Questions for Authors: See weakness. Confidence: 3 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: N/A Flag For Ethics Review: ['No ethics review needed.'] Rating: 7 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We thank the reviewer for their positive comments and feedback. We respond to their questions and comments below.   ___ > Though the general form of convolution is provided. It is not evaluated in experiment. I will suggest putting these content in the appendix in this case. Otherwise, it is better to provide the corresponding experimental results. We thank the reviewer for pointing this out. Our primary goal was to show a general recipe for constructing equivariant unitary/orthogonal maps, which allows for introducing more general convolution on finite groups beyond just graph convolution. Our main application are GNNs, hence, much of the experiments are focused on the graph setting. We provide experimental results for another instance of the general case, namely group convolution over the dihedral group, in Appendix E.3. We note that previous literature has studied unitary/orthogonal CNNs which also fit within this framework. We discuss and reference these works in the extended related works (see Appendix A.1). We will add more explicit pointers to these sections in the main text.   ___ > The computational cost is higher compared to other methods, but there are no experimental results to show the important details. The unitary and orthogonal layers have a constant factor overhead over standard message passing. As stated in the paper, the runtime is $O(KnD)$ where $K$ is the approximation in the exponential map, $n$ is the number of nodes, and $D$ is the maximum degree of the graph. Since the approximation error is exponentially small in $K$ (see Appendix C), we found that setting $K$ to be a constant (around 10 for example) sufficed in all experiments. We will specify this in the main text in the revised manuscript. Note, that this is more efficient than many graph algorithms such as graph transformers which are quadratic scaling in $n$ and other proposed (near) unitary graph message passing algorithms which also scale poorly in $n$ and $E$. The computational cost is detailed further in Appendices A and C. We should also note that this factor $K$ overhead is in some sense unavoidable if one wants invertibility and isometry as detailed in Proposition 4.	null	null	null	null	null	null
S$^{2}$FT: Efficient, Scalable and Generalizable LLM Fine-tuning by Structured Sparsity	Accept (poster)	Summary: This paper aims to address the fact that the current parameter efficient fine-tuning methods cannot simultaneously achieve high-quality, efficient training or scalable LLM services. Therefore, a series of structured sparse fine-tuning methods for LLM are proposed, which simultaneously achieve state-of-the-art fine-tuning performance, training efficiency and reasoning scalability. S2FT achieves this through "sparse selection and dense computation". It only selects a few heads in the MHA module and a few channels in the FFN module in each Transformer block. Then the weight matrix of the coupling structure in the LLM is rearranged so that the selected parts are interconnected to generate multiple compact and dense trainable weight sub-matrices. S2FT only needs to update these sub-matrices for efficient parameter fine-tuning. The paper focuses on the fact that current methods such as LoRA, DoRA, etc., although they can reduce memory, are not as good as fine-tuning in large language models. The unstructured characteristics of SFT require matrix sparse operations, which makes service scalability and training efficiency impossible to guarantee. This is the main reason and motivation for the proposal of S2FT in this paper. In addition, this paper establishes an interface between S2FT and LoRA to support joint or non-joint computing paradigms. Strengths: The idea of this paper is very novel and valuable. It starts with the three problems that the current LLM model may encounter, namely training quality, efficient training, and service scalability. It proposes a structured sparse fine-tuning method to accelerate the large language model and provides corresponding mathematical proofs. This method is very effective for both in-distribution and out-of-distribution model training. The highlights of this paper are as follows: 1. The mathematical theoretical proof of this paper is very complete, the proof ideas are very clear, and the proofs of the in-distribution model and the out-of-distribution model are very complete. The conclusions and theoretical analysis of this paper are clearly seen, and the mathematical theory is very solid. 2. The analysis process of this paper is complete. The advantages and disadvantages of the current research are analyzed in place in the literature review part, the method part is concise and clear, and a complete theoretical proof is provided. The advantages of operating efficiency and service scalability are also analyzed in the experimental part. 3. The simplicity and clarity of this method is a highlight of this paper. Compared with the complicated fine-tuning methods of other methods, this paper only needs to select and rearrange the matrices of some layers for updating, and develops a partial back-propagation algorithm. The number of lines of code implemented is very small. Simplicity is a big highlight of this paper. Weaknesses: The advantages of the S2FT method in this paper are very prominent, and it provides some references for efficient, scalable, and high-quality fine-tuning methods for large language models. However, the author can consider the following areas for improvement or sufficient explanation: 1. The symbols in the proof process of this paper can be slightly organized. Readers may feel that the symbols of the proof are a little messy when reading the entire proof process, but the whole process is complete and flawless. 2. There are few figures for the method in this paper, only the right picture of Figure 2 is shown, and the overall picture is not clear enough. In addition, does this method extract the same layer for each different model? Or is it better to extract specific layers for each model? 3. This paper locks the U of the low-rank decomposition matrix, which is a good attempt, but for different data sets, will the distribution of data affect the data form of the U matrix and the form of the U matrix basis? 4. In the experimental results, there are many phenomena that the method used in this paper is lower than the previous methods. Can the author analyze each task more fully, because it can be seen from the table that some tasks are more obviously degraded? Technical Quality: 3 Clarity: 3 Questions for Authors: 1. The symbols in the proof process of this paper can be slightly organized. Readers may feel that the symbols of the proof are a little messy when reading the entire proof process, but the whole process is complete and flawless. 2. There are few figures for the method in this paper, only the right picture of Figure 2 is shown, and the overall picture is not clear enough. In addition, does this method extract the same layer for each different model? Or is it better to extract specific layers for each model? 3. This paper locks the U of the low-rank decomposition matrix, which is a good attempt, but for different data sets, will the distribution of data affect the data form of the U matrix and the form of the U matrix basis? 4. In the experimental results, there are many phenomena that the method used in this paper is lower than the previous methods. Can the author analyze each task more fully, because it can be seen from the table that some tasks are more obviously degraded? Confidence: 4 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: This paper mentions some limitations of this paper in the limitations. For example, although various model architectures show coupled structures, this paper does not explore the opportunity to extend S2FT to other architectures. At the same time, in terms of model deployment, although this paper verifies the feasibility of scalable services in the work, there is still a lack of a practical and scalable actual service system. However, when using this method, is it necessary to consider the fairness or importance of each layer? For example, for a large language model, the first layer may contain more information, and some information in the middle layer can be appropriately eliminated. Secondly, whether it is truly quality-oriented needs to be considered because S2FT does not perform well in any task in the experiment of this paper. Finally, this paper also mentions that the subsequent optimization direction may be targeted at the residual dependency network, and how to perform S2FT on this type of network needs further consideration. However, since the work of this paper mainly focuses on PEFT, it leads to a reduction in GPU computing resource consumption. Therefore, this method has the potential to have a positive impact on the computing environment by minimizing the computing resources required for fine-tuning LLM. Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thank you for your valuable comments and suggestions. We will carefully revise our paper based on your comments. Our responses to your questions are detailed below. We would greatly appreciate your input on whether our revisions address your concerns. Q1: The symbols in the proof process of this paper can be slightly organized. A1: Thank you for pointing out this issue. In the camera-ready version of our work, we will simplify the symbols and add additional notations before our proof to improve readability. --- Q2: There are few figures for the method in this paper, and the overall picture is not clear enough. A2: Thank you for your suggestion. We will include more figures to illustrate our selection and permutation strategy specific to transformer architectures in detail. The current Figure 2 serves as an abstract demonstration of each coupled structure in a standard Transformer model. Additionally, we will ensure that our figures are clearer in the camera-ready version. --- Q3: Does this method extract the same layer for each different model? A3: Our method employs the same extraction strategy for all models discussed in our paper. Specifically, we always extract all feed-forward network (FFN) layers for Transformer models, with the trainable parameters uniformly distributed across the Up, Down, and Gate Projection Layers. This approach is based on our findings in Figure 4, which show that fine-tuning FFN layers results in better performance compared to fine-tuning Attention layers under the same budget. To demonstrate the effectiveness of our uniform allocation strategy among FFN layers, we have included an additional ablation study in Table R1. In Table R1, we have added an ablation study regarding layerwise allocation of trainable parameters. This study includes the following design patterns: 1. Increasing ($n_{i+1} > n_i$): the number of trainable parameters in every layer gradually increases (or remains the same); 2. Uniform ($n_{i+1} = n_i$: the number of trainable parameters in every layer is the same; 3. Decreasing ($n_{i+1} < n_i$): the number of trainable parameters in every layer gradually decreases; 4. Random One: only one randomly selected layer has trainable parameters. The results show that maintaining a uniform distribution of trainable parameters among different layers leads to the best performance. We will include these results in the camera-ready version of our paper if accepted. Table R1: Performance of different layer allocation strategies on commonsense reasoning tasks. \| Allocation Strategy \| Average Accuracy \| \|-\|-\| \| Increasing \| 81.2\| \| Uniform \| 81.8\| \| Decreasing \| 80.4 \| \| Random One \| 79.9 \| While our allocation strategy is both model- and task-independent, there is great potential to further improve performance by selecting trainable parameters based on model weights or downstream tasks. We leave this problem for future research. --- Q4: Will the distribution of data affect the data form of the U matrix and the form of the U matrix basis? A4: Thank you for your question. Our work employs a random selection strategy for channel selection in Section 5.3, which is task-independent. To demonstrate how different U matrices affect performance on various tasks, we used three different random seeds for channel selection. The results for both commonsense reasoning and math reasoning tasks are presented in Table R2. Our findings indicate that the optimal form of the U matrix varies across different tasks, suggesting that the "best U matrices" should be task-specific. We have left this direction for future research. Table R2: Ablation study of channel selection strategies on the commonsense reasoning and math reasoning tasks. \| Selection Strategy \| Commonsense \| Math \| \|-\|-\|-\| \| Seed: 0 \| 82.2 \| 68.6 \| \| Seed: 42 \| 81.5 \| 70.1 \| \| Seed: 1234 \| 81.0 \| 69.4 \| Q5: Can the author analyze each task more fully? A5: We would like to clarify that in Sections 5.1 and 5.2, we train the base model on a single dataset and evaluate it across several sub-tasks. Therefore, the average performance should be the primary focus, as our method achieves significant improvements in both settings. These results sufficiently demonstrate the superiority and robustness of S$^2$FT. It is expected that our method may not always perform the best in every sub-tasks as different sub-tasks will affect each other. Nonetheless, S$^2$FT consistently achieves the best or second-best performance in sub-tasks, which cannot be characterized as "obviously degraded." To further address your concerns, we have provided more in-depth analyses of the results in Table 1 and Table 2 of the original paper. For the commonsense reasoning tasks in Table 1, S$^2$FT achieves the best performance in most sub-tasks, while maintaining second-best performance in the remaining tasks. These results demonstrate the superiority and robustness of S$^2$FT in memorizing common knowledge. Table 2 further verifies this conclusion, showing that S$^2$FT outperforms LoRA, especially in memorization tasks such as Writing or Humanities within instruction-following scenarios. It also excels in tasks requiring pre-trained knowledge, such as those in STEM. This indicates that S$^2$FT's performance improvements are primarily due to its enhanced ability to memorize new information and retain pre-trained knowledge. For other tasks, such as Reasoning and Math, which focus on multi-hop reasoning with limited knowledge, the performance of LoRA and S$^2$FT is similar. It is common for LoRA to slightly outperform S$^2$FT in some of these tasks. Consequently, SFT-based methods are more effective for knowledge-rich tasks. In comparisons among SFT-based methods, LISA performs slightly better than S$^2$FT in some tasks. This is because LISA updates much more parameters than S$^2$FT, resulting in similar behavior and better performance in certain tasks. --- Rebuttal 2: Title: Reminder on follow-up discussion (2 days left before the discussion period ends) Comment: Thank you so much for your dedicated review of our paper. We recognize the significant time and effort involved in your review, and we greatly appreciate it. With only 2 days remaining before the conclusion of the discussion phase, we wish to extend a respectful request for your feedback about our responses. Thank you! --- Rebuttal Comment 2.1: Title: Response to authors' rebuttal Comment: Author's rebuttal mostly addresses my concern. I will change the score.	Summary: The paper introduces a structured pruning method for LLMs. The main idea is to permute the rows and columns of the weight matrices and select a submatrix during the fine-tuning process. The authors show that the proposed technique outperform previous techniques in terms of accuracy and efficiency. Strengths: 1. The paper is well written. The authors gave a good summary of existing parameter-efficient fine-tuning (PEFT) methods. While the proposed technique itself is not entirely new (it can be considered as a special form of structured sparse training), the authors described the technique clearly. 2. Theoretical analysis is provided for generalization performance. 3. Experiments are comprehensive. The authors tested their technique for two types of tasks (commonsense reasoning and instruction-following) with five fine-tuning categories on more than 10 datasets. Weaknesses: Novelty in the pruning technique itself is limited. It is basically a structured sparse training method. Similar structured pruning methods have been proposed on non-LLM models previously (e.g., [27, 24]). Technical Quality: 3 Clarity: 3 Questions for Authors: Will the authors open-source the code? Confidence: 4 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: Yes. Flag For Ethics Review: ['No ethics review needed.'] Rating: 5 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thank you for your constructive comments and suggestions. We respond to your questions below and would appreciate it if you could let us know if our response addresses your concerns. Q1: Novelty in the pruning technique itself. A1: Thank you for your comment. We acknowledge that structure sparsity is commonly used in model pruning. However, we are the first to adopt structure sparsity for the parameter-efficient fine-tuning (PEFT) of large language models (LLMs). Therefore, our method is a gradient selection technique instead of a pruning technique. Motivated by the challenges of practical efficiency and scalability in previous SFT-based methods, we use coupled structures for flexible and fine-grained gradient selection, introducing a completely new gradient selection strategy. This idea is both novel and effective in this research line. Before the era of LLMs, methods like Diff pruning [1] and Fish Mask [2] mainly focused on unstructured selective fine-tuning, as model sizes were not very large. These methods used a binary mask to enable sparse gradient updates during training, which led to large memory footprints and time costs. In the era of LLMs, researchers additionally prioritize the memory efficiency of PEFT methods during training and scalable serving ability during inference, leading to the popularity of LoRA. SFT-based methods like LISA [3] only enable layer-wise selection and have limitations in serving scalability. In comparison, our method addresses these efficiency bottlenecks and revitalizes SFT-based approaches, surpassing LoRA in both performance and efficiency. Given the current research trends in PEFT methods, our approach is novel in this area. --- Q2: Will the authors open-source the code? A2: We strongly agree that open-sourcing the code is critical for reproducibility and supporting future research in this area. We therefore will definitely make our code publically available and results easy to reproduce upon acceptance. --- References: [1] Guo D, Rush A M, Kim Y. Parameter-efficient transfer learning with diff pruning[J]. arXiv preprint arXiv:2012.07463, 2020. [2] Sung Y L, Nair V, Raffel C A. Training neural networks with fixed sparse masks[J]. Advances in Neural Information Processing Systems, 2021, 34: 24193-24205. [3] Pan R, Liu X, Diao S, et al. LISA: Layerwise Importance Sampling for Memory-Efficient Large Language Model Fine-Tuning[J]. arXiv preprint arXiv:2403.17919, 2024. --- Rebuttal Comment 1.1: Title: Code Release in an Anonymous Repository Comment: To facilitate reproduction, we have provided our code at https://anonymous.4open.science/r/S2FT_Rebuttal-7B17 for reviewer’s verification and will make it publicly available upon paper acceptance. This repository contains the training and inference code necessary to fine-tune a LLaMA-7B model on commonsense reasoning tasks. We hope this addresses your reproduction concerns. --- Rebuttal 2: Title: Reminder on follow-up discussion (2 days left before the discussion period ends) Comment: Thank you so much for your dedicated review of our paper. We recognize the significant time and effort involved in your review, and we greatly appreciate it. With only 2 days remaining before the conclusion of the discussion phase, we wish to extend a respectful request for your feedback about our responses. Thank you!	Summary: Current PEFT methods for LLMs fail to achieve high quality, efficient training, and scalable serving simultaneously. To overcome this, the authors developed Structured Sparse Fine-Tuning (S²FT), which excels in all three areas. S²FT improves generalization by selecting a few heads in the Multi-Head Attention (MHA) and channels in the Feed-Forward Network (FFN) modules for each Transformer block. It forms dense, trainable submatrices by co-permuting weight matrices, preventing overfitting and forgetting. S²FT achieves state-of-the-art performance and reduces fine-tuning memory usage by up to 3 times and increases throughput by 1.5-2.7 times. Strengths: 1. This paper is well-written and organized. The proposed method is technically sound. 2. Efficient tuning of LLMs is an important topic. 3. The performance of the proposed method is very promising. Weaknesses: 1. The author claimed that S2FT prevent overfitting and forgetting. However except the toy experiments in 2nd section, I did not see any experimental proof. E.g. The author only gives MTBench results after the finetuning on AlpacaGPT4, what about the overfitting and forgetting issues? 2. Experiments only conducted on relative small models. 3. The name SFT may incur some confusions as it already has the meaning of supervised fine-tuning. 4. The experiments parts are not sufficient to proof the effectiveness of the proposed method especially on some downstream tasks. Technical Quality: 3 Clarity: 3 Questions for Authors: 1. The author has some ablations on the trainable allocation choice within a block, what about the allocation in different layers? 2. Can this paper be combined with quantization just like QLoRA. Confidence: 4 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: Yes Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thank you for reviewing our paper and for your valuable feedback. Below, we address your concerns point by point and we will revise our paper according to your suggestions. We would appreciate it if you could let us know whether your concerns are addressed by our response. Q1: Empirical Results about the overfitting and forgetting issues? A1: To verify that S$^2$FT can prevent overfitting and forgetting issues, we evaluate its performance in an out-of-distribution (OOD) scenario using MT-Bench, as shown in Table 2 of the original paper. Additionally, we conduct an experiment on arithmetic reasoning tasks, demonstrating that S$^2$FT significantly outperforms LoRA in both near OOD and far OOD settings. In Table 2, we present the results of training the model on the Alpaca GPT-4 dataset and evaluating it on the MT-Bench benchmark, representing an OOD setting. Our method significantly outperforms both LoRA and Full FT in this task, demonstrating that it leads to much less forgetting and better generalization to new datasets. To further address your concern, we trained LLaMA-7B/13B on the Math10K dataset and evaluated its performance on three in-distribution (ID) tasks (GSM8K, AQuA, MAWPS), four near OOD tasks (SVAMP, MultiArith, AddSub, SingleEq), and one far OOD task (MMLU). The results are shown in Table R1. Table R1: Performance comparison between S$^2$FT and LoRA for LLaMA-7B/13B on the arithmetic reasoning tasks and MMLU benchmark. \| Method \| Average (ID) \| Average (Near OOD) \| MMLU (Far OOD) \| \| - \| - \| - \| - \| \| LoRA (LLaMA-7B) \| 45.1 \| 78.7 \| 27.8 \| \| S$^2$FT (LLaMA-7B) \| 46.6 \| 81.7 \| 33.1 \| \| LoRA (LLaMA-13B) \| 49.9 \| 81.6 \| 36.3 \| \| S$^2$FT (LLaMA-13B) \| 50.5 \| 84.2 \| 42.2 \| According to the results, S$^2$FT led to improvements of 4.3%, 3.0%, and 1.4% for LLaMA-7B in far OOD, near OOD, and ID settings, respectively, compared to LoRA. For LLaMA-13B, the corresponding improvements are 5.9%, 2.6%, and 0.6%. As the distribution difference between the training data and test data increases, the performance gap between S$^2$FT and LoRA enlarges, demonstrating that our method is effective in preventing overfitting and forgetting. We will include these results in the camera-ready version of our paper if accepted. --- Q2: Experiments on larger models. A2: Thank you for your suggestion. Following LISA, we have added experiment results for LLaMA2-70B on MT-Bench and GSM8K in Table R2. The results show that S$^2$FT outperforms other PEFT methods for larger models, providing strong evidence to support S$^2$FT’s scalability under large-scale training scenarios. We will include these results in the camera-ready version of our paper if accepted. Table R2: Performance comparison between different methods for LLaMA2-70B on MT-Bench and GSM8K. \| Method \| MT-Bench \| GSM8K \| \|-\|-\|-\| \| Vanilla \| 5.19 \| 54.8 \| \| LoRA \| 6.10 \| 59.4 \| \| LISA \| 6.72 \| 61.1 \| \| Full FT \| 6.25 \| 67.1 \| \| S$^2$FT \|6.91\|64.7\| --- Q3: The name SFT may incur some confusions as it already has the meaning of supervised fine-tuning. A3: Thank you for your suggestion. We will replace "SFT" with "sparse FT" in the camera-ready version to avoid any confusion. --- Q4: The experiment parts are not sufficient. A4: In the original paper, we included experiments with five different base models and more than ten tasks, covering both commonsense reasoning and instruction-following tasks. We have also added some new experiments in Table R1 and R2. We hope these findings provide more comprehensive results of our method, and we will include them in the camera-ready version of our paper if accepted. In Table R1, we further present experimental results for arithmetic reasoning tasks, demonstrating the effectiveness of our method across various downstream tasks. The results in near OOD and far OOD settings further highlight its ability to address overfitting and forgetting issues. Additionally, in Table R2, our experiments on LLaMA2-70B showcase S$^2$FT's scalability in large-scale training scenarios. --- Q5: What about the allocation in different layers? A5: Thank you for pointing out this issue. In Section 5.3, we maintain a uniform allocation of parameters across different layers. To further address your question, we have added an ablation study in Table R3 concerning layer-wise allocation. This study includes the following design patterns: (i) Increasing ($n_{i+1} > n_i$): the number of trainable parameters in every layer gradually increases (or remains the same); (ii) Uniform ($n_{i+1} = n_i$: the number of trainable parameters in every layer is the same; and (iii) Decreasing ($n_{i+1} < n_i$): the number of trainable parameters in every layer gradually decreases. The results in Table R3 indicate that maintaining a uniform distribution of trainable parameters across different layers leads to the best performance. A more detailed analysis in this direction will be left for future research. Table R3: Performance of different allocation strategies on commonsense reasoning tasks for LLaMA-7B. \| Allocation Strategy \| Average Accuracy \| \|-\|-\| \| Increasing \| 81.2 \| \| Uniform \| 81.8 \| \| Decreasing \| 80.4 \| --- Q6: Can this paper be combined with quantization just like QLoRA A6: Thank you for highlighting the potential of combining S$^2$FT with quantization. Our method can indeed be integrated with quantization, similar to QLoRA, by using mixed precision storage. Once the trainable parameters are determined, we retain these parameters in their original precision while quantizing the other parameters to low bits. This approach enables quantized PEFT, similar to QLoRA. By maintaining the trainable parameters as small, dense submatrices after permutation, our storage remains relatively hardware-efficient, even with mixed precision. We plan to conduct more experiments in this direction in the future. --- Rebuttal 2: Title: Reminder on follow-up discussion (2 days left before the discussion period ends) Comment: Thank you so much for your dedicated review of our paper. We recognize the significant time and effort involved in your review, and we greatly appreciate it. With only 2 days remaining before the conclusion of the discussion phase, we wish to extend a respectful request for your feedback about our responses. Thank you! --- Rebuttal Comment 2.1: Comment: Thanks for the authors' response. I have no more questions right now. I choose not to change the scores.	Summary: This paper introduces a new family of methods called Structured Sparse Fine-Tuning (S$^2$FT) for large language models (LLMs). S$^2$FT aims to achieve state-of-the-art fine-tuning performance, training efficiency, and inference scalability simultaneously. The method selects a few heads in the multi-head attention (MHA) module and a few channels in the feed-forward network (FFN) module for each Transformer block, then co-permutes the weight matrices to connect these selected components. This results in multiple compact, dense, and trainable weight submatrices that are updated during fine-tuning. The approach prevents overfitting and forgetting, delivers superior performance on benchmarks, and improves memory and throughput efficiency compared to full fine-tuning and existing parameter-efficient fine-tuning (PEFT) methods. Strengths: This paper introduces a method combining structured sparsity with fine-tuning, enhancing both efficiency and performance. The method proposed can reduce memory costs and improve throughput. The method demonstrates strong generalization capabilities. The approach allows for scalable batched serving of multiple fine-tuned models without additional inference overhead. The experiments are comprehensive. Weaknesses: While the results look impressive, the reviewer is not fully convinced by the theoretical motivations. For example, the paper relies on the assumption that structured sparsity can effectively represent the necessary model adaptations, which is hard to verify in realistic settings. Although the benchmarks are comprehensive, most of the benchmarks do not include SFT or other SFT-based methods. Since S$^2$FT is in some sense an SFT-based method, it would be necessary to include such comparisons to understand if the performance improvements come from the "structure" (i.e. the $^2$) or SFT itself. Based on the current benchmarks, it is difficult to make such conclusions. While the paper includes some implementation details, aiding reproducibility, the code is not provided, making it hard to verify or reproduce the results, especially for a paper with mostly empirical results. Technical Quality: 3 Clarity: 3 Questions for Authors: Could the authors elaborate on the scaling of trainable parameters, as well as the space and time complexity, compared to other SFT-based or LoRA-based methods? Could the authors elaborate on how the hyperparameters are chosen, and provide a more comprehensive list of hyperparameters used in this paper, the procedure to choose them, and the discarded hyperparameters? Right now, the paper mainly discusses how to apply S$^2$FT to attention-based LLMs. The reviewer wonders how to identify the important weights and apply this method to other models. Most of the experiments were performed on LLaMA and LLaMA2. The reviewer wonders if there is any difficulty in applying the method to LLaMA3. Is it possible that different datasets/tasks could result in a different set of "important weights"? The reviewer wonders if the authors explored this possibility. The reviewer wonders if the author could release the source code into an anonymous repository for review purposes. In the NeurIPS checklist, the author claims that their results include confidence intervals. Could the authors make some comments on the confidence intervals for the tables in the paper? Confidence: 3 Soundness: 3 Presentation: 3 Contribution: 2 Limitations: The limitations and broader impacts are addressed. Flag For Ethics Review: ['No ethics review needed.'] Rating: 5 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thank you so much for the insightful and valuable comments! They are very helpful for further improving the clarity and quality of our paper. We'll revise our manuscript in the camera-ready version to address all of your concerns. Q1: How can we verify the model adaptation ability in real-world settings? A1: To verify the model's adaptation ability in real-world settings, we evaluate its performance in an out-of-distribution (OOD) scenario, as shown in Table 2 in the original paper. Our method, S$^2$FT, significantly outperforms both LoRA and Full FT by a large margin in this task, demonstrating its superior model adaptation ability. Additionally, we conduct an experiment on arithmetic reasoning tasks to demonstrate that S$^2$FT significantly outperforms LoRA in OOD settings (See Empirical results for generalization task part from rebuttal to all reviewers). --- Q2: Ablation study between SFT and S$^2$FT. A2: Thank you for your suggestion. Our results in Table R2 show that SFT primarily leads to performance improvement, while our structure enhances efficiency and scalability, as discussed in Section 6 in the original paper. Table R2: Performance comparison between S$^2$FT and SFT on the commonsense reasoning tasks. \| Model \| SFT \| S$^2$FT \| \| - \| - \| - \| \| LLaMA-7B \| 81.2 \| 81.8 \| \| LLaMA2-7B \| 83.7 \| 83.0 \| \| LLaMA3-8B \| 87.2 \| 86.6 \| --- Q3: Reproduction of the paper. A3: We strongly agree that open-sourcing the code is critical for reproducibility and supporting future research in this area. We therefore will definitely make our code publically available and results easy to reproduce upon acceptance. --- Q4: How are the hyperparameters chosen? A4: Thank you for highlighting the importance of detailing our hyperparameter configuration. We have included a more comprehensive list of the hyperparameters used in this paper in Tables R4 and R5. Most hyperparameters follow previous work [2, 3], and we only tune the learning rates. Table R4: Hyperparameter configurations of S$^2$FT for LLaMA-7B/13B, LLaMA2-7B, and LLaMA3-8B on the commonsense reasoning tasks. \| Hyperparameters (S$^2$FT) \| LLaMA-7B \| LLaMA-13B \| LLaMA2-7B \| LLaMA3-8B \| \| - \| - \| - \| - \| - \| \| LR \| 1e-6 \| 1e-6 \| 1e-6 \| 1e-6 \| \| LR Scheduler \| Linear \| Linear \| Linear \| Linear \| \| Optimizer\| AdamW\| AdamW \| AdamW \| AdamW \| \| Batch size\| 16 \| 16 \| 16 \| 16 \| \| Warmup Steps \| 100 \| 100 \| 100 \| 100 \| \| Epochs \| 3 \| 3 \| 3 \| 3 \| \| Where \| Up,Down,Gate \| Up,Down,Gate \| Up,Down,Gate \| Up,Down,Gate \| Table R5: Hyperparameter configurations of S$^2$FT for LLaMA2-7B and Mistral-7B on the instruction-following task. \| Hyperparameters (S$^2$FT) \| Mistral-7B \| LLaMA2-7B \| \| - \| - \| - \| \| LR \| 2e-5 \| 1e-5 \| \| LR Scheduler \| Cosine \| Cosine \| \| Optimizer\|AdamW\| AdamW\| \| Batch size\|4\|4\| \| Warmup Steps\|100\|100\| \| Epochs \| 1 \| 1 \| \| Where \| Up,Down,Gate \| Up,Down,Gate \| --- Q5: How to identify the important weights? A5: In Section 5.2, we discuss our weight selection strategy. For a standard transformer architecture, we first uniformly allocate the trainable parameters across different transformer layers. Next, we freeze all attention modules and evenly assign parameters to the Up, Gate, and Down projection modules. Since these three modules represent a coupled structure within the FFN module, we randomly select the same channels to update for each module. In Table 3, we also introduce an alternative activation-based selection strategy, which resulted in inferior performance. We leave the exploration of more advanced metrics to identify important weights as a topic for future research. --- Q6: How to apply S$^2$FT to different models? A6: Thank you for your interest in the application of S$^2$FT across different model architectures. It is indeed versatile and can be applied to Transformers, CNNs, RNNs, and GNNs. As detailed in Section 3.2 of our work, S$^2$FT can be utilized in the Multi-Head Attention and Feed-Forward Network modules of standard Transformer architecture, which is sufficient for most LLMs and diffusion models currently. Additionally, such coupled structures in Figure 3(a) also exist in CNNs, RNNs, and GNNs as shown in [1]. Therefore, S$^2$FT is effective across various model architectures. --- Q7: Most of the experiments were performed on LLaMA and LLaMA2. The reviewer wonders if there is any difficulty in applying the method to LLaMA3. A7: In Table 1, we present the results of LLaMA3-8B on commonsense reasoning tasks following DoRA and ReFT. S$^2$FT achieves an average improvement of 1.4%, demonstrating its applicability to LLaMA3. For other experiments, we primarily focus on LLaMA and LLaMA2 to ensure a fair comparison, as these base models are commonly used in our baseline methods. --- Q8: Could the authors make some comments on the confidence intervals for the tables in the paper? A8: Thank you for your suggestion. We have added the full results with confidence intervals in the Appendix, and included comments in the tables in the main paper, which will be visible in the camera-ready version. --- Q9: Could different datasets/tasks result in a different set of "important weights"? A9: See The affect of selecting different trainable parameters on different datasets/tasks part from rebuttal to all reviewers. --- References: [1] Fang G, Ma X, Song M, et al. Depgraph: Towards any structural pruning[C]//Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. 2023: 16091-16101. --- Rebuttal Comment 1.1: Comment: Thanks a lot for taking the time and effort to answer my questions. I would like to keep my recommendation for acceptance of the paper, and am considering raising the score in the next 2 days. --- Reply to Comment 1.1.1: Comment: Thank you very much for your response. Please don't hesitate to reach out if you have any further questions. We greatly appreciate your willingness to consider raising the score. --- Rebuttal 2: Title: Code Release in an Anonymous Repository Comment: To facilitate reproduction, we have provided our code at https://anonymous.4open.science/r/S2FT_Rebuttal-7B17 for reviewer’s verification and will make it publicly available upon paper acceptance. This repository contains the training and inference code necessary to fine-tune a LLaMA-7B model on commonsense reasoning tasks. We hope this addresses your reproduction concerns.	Rebuttal 1: Rebuttal: We thank reviewers [R1(CFWk), R2(8C1L), R3(d9g6), R4(Cfys)] for their thoughtful and highly supportive feedback! We were glad that the reviewers found the problem significant and interesting [R2], the observations and theoretical analysis insightful and highly valuable [R3, R4], the methods novel, simple, and effective [R2, R4], the presentation easy to follow [R2, R3], experimental results comprehensive and impressive [R1, R2, R3]. We have updated the paper to incorporate constructive suggestions, which will show in the camera-ready version. We summarize the major changes: * [ R1, R3] Empirical results for generalization tasks: In Table R1, we trained LLaMA-7B/13B on the Math10K dataset and evaluated its performance on three in-distribution (ID) tasks (GSM8K, AQuA, MAWPS), four near OOD tasks (SVAMP, MultiArith, AddSub, SingleEq), and one far OOD task (MMLU). As the distribution difference between the training data and test data increases, the performance gap between S$^2$FT and LoRA enlarges, demonstrating that our method is effective in generalization by preventing overfitting and forgetting. Table R1: Performance comparison between S$^2$FT and LoRA for LLaMA-7B/13B on the arithmetic reasoning tasks and MMLU benchmark. \| Method \| GSM8K \| AQuA \| MAWPS \| SVAMP \| MultiArith \| AddSub \| SingleEq \| Average (ID) \| Average (Near OOD) \| MMLU (Far OOD) \| \| -------------------------- \| ----- \| ---- \| ----- \| ----- \| ---------- \| ------ \| -------- \| ------- \| ----------- \| ---------- \| \| LoRA (LLaMA-7B) \| 37.5 \| 18.9 \| 79 \| 52.1 \| 95 \| 83.3 \| 84.4 \| 45.1 \| 78.7 \| 27.8 \| \| S$^2$FT (LLaMA-7B) \| 35.8 \| 22 \| 81.9 \| 57.1 \| 93.3 \| 87.3 \| 89.2 \| 46.6 \| 81.7 \| 33.1 \| \| LoRA (LLaMA-13B) \| 47.5 \| 18.5 \| 83.6 \| 54.6 \| 94.8 \| 87.3 \| 89.8 \| 49.9 \| 81.6 \| 36.3 \| \| S$^2$FT (LLaMA-13B) \| 45.8 \| 21.7 \| 84 \| 63 \| 95 \| 87.3 \| 91.5 \| 50.5 \| 84.2 \| 42.2 \| * [ R2 ] Experimental results for large-scale language models: We have added experiment results for LLaMA2-70B on MT-Bench and GSM8K in Table R2. The results show that S$^2$FT outperforms other PEFT methods for larger models, providing strong evidence to support S$^2$FT’s scalability under large-scale training scenarios. We will include these results in the camera-ready version of our paper if accepted. Table R2: Performance comparison between different methods for LLaMA2-70B on MT-Bench and GSM8K. \| Method \| MT-Bench \| GSM8K \| \|---------\|-----------\|--------\| \| Vanilla \| 5.19 \| 54.8 \| \| LoRA \| 6.10 \| 59.4 \| \| LISA \| 6.72 \| 61.1 \| \| Full FT \| 6.25 \| 67.1 \| \| S$^2$FT \|6.91\|64.7\| * [ R2, R4 ] Ablation Study for layerwise allocation strategy: For the allocation among different layers, we maintain a uniform allocation strategy, meaning we fairly assign trainable parameters to all layers. Our study in Table 3 includes four design patterns: (i) Increasing ($n_{i+1} > n_i$): the number of trainable parameters in every layer gradually increases (or remains the same); (ii) Uniform ($n_{i+1} = n_i$: the number of trainable parameters in every layer is the same; and (iii) Decreasing ($n_{i+1} < n_i$): the number of trainable parameters in every layer gradually decreases; (iv) Random One: only one randomly selected layer has trainable parameters. The results show that maintaining a uniform distribution of trainable parameters among different layers leads to the best performance. Table R3: Performance of different layerwise allocation strategies on commonsense reasoning tasks for LLaMA-7B. \| Allocation Strategy \| Avg. Accuracy \| \|-----------------------\|---------------\| \| Increasing \| 81.2 \| \| Uniform \| 81.8 \| \| Decreasing \| 80.4 \| \| Random One \| 79.9 \| * [R1, R4] The affect of selecting different trainable parameters on different datasets/tasks: As detailed In Section 5.3, we use a random selection strategy for channel selection. We apply the same three random seeds to determine the selection strategies for both commonsense reasoning and math reasoning tasks. The results are shown in Table R3. The results indicate that the best channel selection strategy varies across different tasks. Therefore, the "important trainable parameters" are different for each task. Table R4: Ablation study of channel selection strategies on different tasks. \| Selection Strategy \| Commonsense \| Math \| \|-----------------------\|---------------\|-----------------------\| \| Seed: 0 \| 82.2 \| 68.6 \| \| Seed: 42 \| 81.5 \| 70.1 \| \| Seed: 1234 \| 81.0 \| 69.4 \|	NeurIPS_2024_submissions_huggingface	2,024	null	null	null	null	null	null	null	null
Lucy: Think and Reason to Solve Text-to-SQL	Reject	Summary: This paper addresses the challenge of developing effective LLM-based assistants for querying SQL databases. In this context, users pose questions to a relational database in natural language, and the goal is to generate a SQL query that correctly answers the user's question when executed. The authors focus on overcoming a limitation of current text-to-SQL approaches: the difficulty LLMs face in handling databases with numerous tables and complex relationships, making it hard to determine the necessary table joins for the query. To tackle this issue, the authors propose a workflow that begins with using the LLM to identify relevant tables and their attributes. In the second step, a constraint satisfaction solver (CSP) is employed to determine the necessary joins while adhering to database constraints. In the third step, a materialized view is created by joining the relevant tables. Finally, this view, combined with the user's question, is used to prompt the LLM to generate the final SQL query. Strengths: This paper tackles a very relevant practical problem, which is attracting significant attention in both academia and industry. The proposed workflow may add practical value. Weaknesses: The paper lacks depth, and the writing does not, in my opinion, meet the quality standards required for a venue like NeurIPS. Additionally, several potential limitations of the proposed workflow are not discussed. • Accuracy of the answers is not the only important requirement in generating SQL from text. Database users also expect query generation to be time-efficient. The proposed workflow includes several computationally intensive steps: first, solving an NP-complete problem (CSP), and second, creating a potentially enormous materialized view by joining many tables. I would have expected a discussion on the computational limitations of this approach. • The workflow lacks sufficient precision and clarity. For example, it is unclear whether the final query is expressed with respect to the materialized view as the only table or with respect to the original schema. Additionally, how lookup tables and various schema design patterns are identified in the input database is not well-explained. The authors claim that their approach guarantees the generated query respects database constraints, but this guarantee is not clearly defined. Algorithm 1 is underspecified; at this high level of detail, the algorithm seems redundant and could be subsumed by the text description. The exact SQL fragment covered by this approach is also unclear. While the limitations section mentions that queries requiring the union operator are not supported, it is unclear if other standard SQL constructs are also unsupported. • While relevant related work is cited, the main body of the paper lacks a detailed discussion on the contribution in relation to recent approaches. • The evaluation section is somewhat lacking. The tables are confusing, and it is unclear what each of the rows actually represents. Technical Quality: 2 Clarity: 2 Questions for Authors: Please comment on the limitations concerning unsupported SQL constructs as well as on the computational limitations of the proposed approach. Confidence: 4 Soundness: 2 Presentation: 2 Contribution: 2 Limitations: The limitations section mentions some but not all of the relevant limitations of this approach. Flag For Ethics Review: ['No ethics review needed.'] Rating: 3 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thank you for your comments and suggestions! We will clarify the following points in the paper: > Efficiency solving an NP-complete problem (CSP) Solving CSP takes less than a second in all experiments we tried. As the reviewer pointed out, solving CSP is an NP-complete problem. However, modern CSP solvers scale to thousands of variables for industrial benchmarks. In the text2SQL use case, the number of variables is of the order of the number of relevant tables squared. It is a very easy problem for modern solvers and has negligible computational overhead. Please see more in replies to all reviewers. > Efficiency of enormous materialized view by joining many tables. It is unclear whether the final query is expressed with respect to the materialized view as the only table or with respect to the original schema. The view we build is not materialized. Instead, it serves as a sub-query to the final database query. This enables the database query planner to apply standard optimizations such as predicate pushdown to execute the final query efficiently. > how lookup tables and various schema design patterns are identified in the input database is not well-explained. Snowflake, many-to-many, star, and other patterns, which we formalize as constraints, are standard database modeling concepts applied by the database designer when creating the database schema. It is therefore straightforward for a domain expert to identify and capture these constraints. In fact, they are often explicitly written down as part of the logical database model (when one exists). > The authors claim that their approach guarantees the generated query respects database constraints, but this guarantee is not clearly defined. We guarantee that the generated query satisfies database constraints in the sense that is precisely defined in Section 3.3 C1-C5. As an example, constraint C1 guarantees that when joining a pair of tables with matching primary and foreign keys, the join will be performed using these keys. Constraint C3 guarantees that when joining two tables connected by a many-to-many relation, the join will be implemented as a 3-way join via the auxiliary table. Our constraint satisfaction problem models all constraints precisely. As the solver is complete and the constraints are hard, we are guaranteed to output a sequence of joins that satisfies these constraints. > The exact SQL fragment covered by this approach is also unclear. While the limitations section mentions that queries requiring the union operator are not supported, it is unclear if other standard SQL constructs are also unsupported. The final SQL query consists of the view definition produced with the help of the constraint solver and the query against this view output by the LLM. The former consists of inner and left joins only. The latter can include arbitrary SQL constructs generated by the LLM. > While relevant related work is cited, the main body of the paper lacks a detailed discussion on the contribution in relation to recent approaches. We will add a comparison with the schema linking method as Reviewer 2 suggested. However, the main contribution is novel (Please see more in the reply to all reviewers.) > The evaluation section is somewhat lacking. The tables are confusing, and it is unclear what each of the rows actually represents. We have descriptions of all the metrics that we used for evaluation, including the standard metric from BIRD (ex). We will further clarify row descriptions. --- Rebuttal Comment 1.1: Title: Author rebuttal Comment: I confirm that I have read the authors' rebuttal and thank the authors for their clarifications. I would like to keep my score. --- Reply to Comment 1.1.1: Comment: Thank you for your comment! We would appreciate more feedback from the reviewer on the points not addressed in the rebuttal. We believe we address all technical concerns in the rebuttal.	Summary: In this paper the authors introduce LUCY, a new LLM based framework for converting text-to-SQL to query databases. Primarily this framework focusses on addressing user queries to databases that contain a large number of tables with complex relations between them. The core idea of this approach is to decompose the query generation process into distinct stages. LLMs (GPT-4) are utilized for generative tasks such as identifying relevant tables and attributes and generating the SQL query. Meanwhile a deterministic constraint solver (OR-Tools) is employed to map relationships between these elements. In essence LUCY processes a user query through 3 phases namely MatchTables, GenerativeView and QueryView phases. In the MatchTables phase the goal is to identify the relevant tables and attributes. This is accomplished by iteratively prompting a Large Language Model (such as GPT-4) to identify relevant tables and attributes based on the user query and the database model, which includes the schema and an optional list of high-level design patterns. The database model is presented in a hierarchical manner and explored using a breadth-first search approach. Once the relevant relations and attributes are identified a schema graph is constructed and solved using a constrainst solver (i.e to identify the optimal path to join the tables) to build a view in the GenerativeView phase. A LLM is then prompted to generate a SQL query given the summary view and the user query in the QueryView phase. The authors further conduct experiments that demonstrate that the proposed technique achieves a better execution accuracy as compared to the existing state-of-the-art techniques on standard datasets(ACME, BIRD). Furthermore, they also introduce a show large improvements on a new benchmark dataset (Cloud Resources). Strengths: 1. The literature review is comprehensive and the paper does a good job at clearly defining the problem to solve. 2. The novelty of the proposed approach lies in the decomposition of tasks involved in generating SQL queries. By employing LLMs to handle specific subtasks, it effectively circumvents the need for LLMs to perform complex reasoning. A core distinguishing factor from prior research is the use of constraint solvers to identify the relevant paths for joining the identified tables. 3. The authors also demonstrate that the proposed approach achieves a better execution accuracy than the existing SOTA on several benchmarks. Weaknesses: 1. The paper ends abruptly without a clear and comprehensive conclusion. The paper presentation needs improvement in this regard. 2. The authors introduce a new benchmark for evaluation but do not offer sufficient details regarding it. A detailed overview of the queries and an analysis of why the existing SOTA techniques do not perform well on the same could be provided which could greatly inform future work. 3. The practical utility of the proposed technique seems to be limited as each user query requires multiple calls to be made to LLMs thereby entailing both increased latency and cost. 4. The error analysis is not very comprehensive and could be improved. For instance how does this technique fare when the names of entities in database schemas are not semantically meaningful or if there are conflicts in descriptions etc (as is often the case in real-world industrial databases). Technical Quality: 3 Clarity: 2 Questions for Authors: 1. For the new benchmark introduced based on the Cloud Resources dataset it is shown that LUCY significantly outperforms the state-of-the-art. What are the queries used in this benchmark? What sets them apart from the standard benchmarks where the execution accuracy of SOTA is comparable/close to LUCY? 2. Has any analysis been done to measure the performance gain vs cost tradeoff when compared to SOTA. 3. What happens when 2 entities in a database have similar descriptions? How sensitive is this technique to having semantically meaningful table names? 4. Is there any mechanism used to handle hallucinations from the MatchTable phase in the GenerativeView Phase? Confidence: 4 Soundness: 3 Presentation: 2 Contribution: 2 Limitations: 1. As the technique leverages LLMs it seems to be heavily reliant on having semantically meaningful entity names /descriptions 2. The proposed technique seems sensitive to hallucinations as it involves processing a query through multiple LLM phases. The errors in any of the earlier phases would result in it propagating to the next stage. For instance as the authors pointed out if the MatchTables phase produces an extra table this could in turn effect the end output. Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thank you for your comments and suggestions! > For the new benchmark introduced based on the Cloud Resources dataset, it is shown that LUCY significantly outperforms the state-of-the-art. What are the queries used in this benchmark? What sets them apart from the standard benchmarks where the execution accuracy of SOTA is comparable/close to LUCY? The main distinguishing feature is the complex relationships between tables and queries required to reason about these relationships (more than 6 tables). Cloud Resources is similar in this respect to ACME Insurance introduced recently by [a]. Lucy works very well on ACME as well (Please see Section 5, ACME insurance). Our results support conclusions from [a]: pure LLM-based approaches do not work on such benchmarks. For example, GPT-4 adds comments like we pointed out in the introduction in most of the outputs: 'This join may need adjustment based on the actual logic of relating claims to policy coverage details.’ In summary, these databases are much more complex compared to BIRD databases, including financial and formula1 databases that we tested on, which are the most complex in terms of relationships in BIRD. #### [a] Sequeda, D. Allemang, and B. Jacob. A benchmark to understand the role of knowledge graphs on large language model’s accuracy for question answering on enterprise sql databases, 2023 > Has any analysis been done to measure the performance gain vs. cost tradeoff when compared to SOTA? At the time of submission, Chat2Query is the state-of-the-art method for zero-shot performance. We performed cost analysis for Cloud resources for the benchmarks we reran: Chat2Query costs $15, while our method costs 50 cents. A direct performance comparison is hard to perform, as we ran Chat2Query on the cloud service TiDBCloud ( Chat2Query is a closed-source industrial tool) and Lucy runs locally using GPT-4 via API . Another data point is GPT-4, which costs $2 on the same dataset. The reason that LUCY is cheaper than GPT4 is that we perform iterative traversal on snowflake patterns before feeding them to GPT-4. Please see more in the reply to all reviewers. > What happens when two entities in a database have similar descriptions? We noticed that Lucy performs well in such cases, given that the descriptions convey meaningful information. Here is an example of three table descriptions from the ACME (insurance) dataset that look similar, at least to non-experts in insurance terminology: * "Claim Payment": The amount paid for loss or expense to settle a claim in whole or in part. * "Expense Payment": The amount paid for the expenses to settle a claim in whole or in part. * "Loss Payment": The amount paid to claimants to settle a claim. As Section 5 (ACME insurance results) shows, Lucy produces highly accurate results on this benchmark. > How sensitive is this technique to having semantically meaningful table names? We have not experimented with renaming objects. However, we believe that Lucy might be sensitive in the first phase, when detecting relevant objects, if table names are meaningless. Nevertheless, we think all techniques have the same limitation. For example, the idea of evidence in BIRD is to clarify a mismatch between the user's request and how these should be mapped to a SQL query. Consider, for example, a question from the financial dataset. "query id": "financial_90": "query": "How many accounts that have a region in Prague are eligible for loans?" "evidence": "A3 contains the data of region" The relevant table for this query is DISTRICT that contains meaningless column names like A1, A2, etc., so evidence has to compensate for that. > Is there any mechanism used to handle hallucinations from the MatchTable phase in the GenerativeView Phase? MatchTable phase: There is a chance that MatchTable produces more relevant tables/attributes than necessary. Usually, it is an overapproximation of the true relevant tables/attributes. While additional attributes do not pose an issue, additional tables can indeed be a problem. However, the task of finding relevant tables is much simpler than solving text2SQL as a whole, so experimentally, Lucy performs well. Also, note that this phase is 100\% hallucination-proof in terms of making up nonexistent tables/attributes (GPT4 is often prone to this type of hallucination). * First, we force GPT to output tables/attributes from a predefined set of attributes. * Second, we always check that it outputs a subset of predefined tables. Please see D.2.1 promptA and example D2 that show the list of elements GPT has to pick from. GenerativeView phase: This phase itself does not introduce additional hallucinations as we do not use LLMs in this phase. Only a constraint solver is used to form a view V. --- Rebuttal Comment 1.1: Title: Acknowledgement of rebuttal Comment: I thank the authors for the clarifications. I am keeping my score.	Summary: The author proposes a new method, Lucy, designed to handle large databases with complex relationships between objects. Lucy operates through three steps: MatchTables, GenerateView, and QueryView. It first identifies relevant tables and attributes using LLMs, constructs a combined view with an automated reasoner, and generates the final SQL query. Lucy shifts complex reasoning from LLMs to a CSP solver, supporting various database design patterns. Experiments on ACME insurance, Cloud Resources, and the two BIRD databases show that Lucy outperforms other zero-shot text-to-SQL models. Strengths: - The proposed method offers a fresh perspective on tackling text-to-SQL research with a logical workflow. - The paper is well-written and easy to follow. Weaknesses: - I am not convinced by the motivation of zero-shot text-to-SQL with the example of industrial databases having complex relationships. Text-to-SQL systems deployment in real-industry requires high performance. I doubt that people won't be using zero-shot models for real use applications. In KaggleDBQA, it also states "we believe the zero-shot setting is overly-restrictive compared to how text-to-SQL systems are likely to be actually used in practice." I would like to hear the authors' thoughts on this. - The paper does not appear to be well-grounded in text-to-SQL research. For example, one way to handle complex relationships in text-to-SQL using LLMs is through schema linking. However, the paper does not mention this area of research and instead proposes MatchTables, seemingly ignoring the rich literature of text-to-SQL works. Other approaches include least-to-most prompting attempts in text-to-SQL for task decomposition and Natural SQL for intermediate representation (although it does not handle query nesting). Properly discussing these relevant methods of the proposed method will better situate the work. Technical Quality: 3 Clarity: 4 Questions for Authors: Please address my above concerns. I am willing to raise the score if convinced. Confidence: 2 Soundness: 3 Presentation: 4 Contribution: 3 Limitations: The limitations of the work are well-stated. Flag For Ethics Review: ['No ethics review needed.'] Rating: 4 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We thank the reviewer for their comments and suggestions! > Zero-shot vs multi-shot In large industrial databases with complex relationships, using multi-shot will not improve accuracy with respect to database constraints. There are two reasons for that: * First, the structure of the database is very complex, with easily hundreds of tables. * Second, user questions are very diverse, as the user can query about any aspect of their database. Hence, getting predefined or automatically generated examples that allow LLMs to capture dependencies between 6 or more tables to answer a user query is as hard as answering the original user request. > Schema linking We appreciate the reviewer pointing us to the schema linking work. This is indeed a similar approach with a similar goal. We will cite the literature on this topic. The main difference in our approach is that we can handle database structures, like snowflake efficiently. However, we can definitely borrow some ideas from the literature to improve this phase, e.g., ideas from latest work that uses linking schema [a] However, we would like to highlight that our main contribution -- the separation of responsibilities between LLMs and automated reasoners -- is new to the best of our knowledge. #### [a] Tapilot-Crossing: Benchmarking and Evolving LLMs Towards Interactive Data Analysis Agents Jinyang Li, Nan Huo, Yan Gao, Jiayi Shi, Yingxiu Zhao, Ge Qu, Yurong Wu, Chenhao Ma, Jian-Guang Lou, Reynold Cheng https://arxiv.org/abs/2403.05307 --- Rebuttal Comment 1.1: Comment: Dear Reviewer, we hope our rebuttal addressed addressed your concerns. We are happy to answer any additional questions.	Summary: The paper introduces Lucy, a framework for solving Text2SQL by LLMs, particularly for complex enterprise databases. Lucy leverages LLMs' understanding and reasoning capabilities to handle intricate database relationships and constraints. The framework operates in three phases: identifying relevant tables and attributes (MatchTables), constructing a view through constraint reasoning (GenerateView), and generating the final SQL query (QueryView). The empirical studies show Lucy achieves performance improvements on several zero-shot Text2SQL tasks. Strengths: Text2SQL is an essential problem in commercial scenarios. Weaknesses: The draft seems far from complete, so leave some high-level suggestions. 1. Make the title, abstract, and introduction more concrete. It is hard to tell the contribution or uniqueness of this work among other papers about Text2SQL by LLMs. 2. Survey related works and clearly state the contribution/novelty of the proposed method against others. 3. Define the terminologies or abbreviations before their first appearance. 4. Make the draft concise by removing unnecessary content. For example, the first challenge introduced in Motivation section is not relevant to this work. 5. The empirical studies could be more convincing by following others' evaluation protocols, such as BIRD. 6. Lack of comparison to other competitors. 7. The figures, tables, and their captions should be self-explanatory. Technical Quality: 1 Clarity: 1 Questions for Authors: 1. What is the key difference between the proposed framework and competitors (e.g., MCS-SQL, MAC-SQL, Chat2Query)? 2. How does the proposed method perform if evaluated by the same protocol as the BIRD-SQL leaderboard? Confidence: 4 Soundness: 1 Presentation: 1 Contribution: 2 Limitations: There is a discussion about the limitation, though the first limitation seems too broad and unnecessary. Flag For Ethics Review: ['No ethics review needed.'] Rating: 3 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thank you for your comments and suggestions! > What is the key difference between the proposed framework and competitors (e.g., MCS-SQL, MAC-SQL, Chat2Query)? The key difference is that all text-to-SQL methods rely on LLMs to reason about database constraints. These constraints are hard constraints that must be enforced in any valid query. LLMs have been shown to be weak in reasoning about hard constraints. In our work, we use automated reasoning tools to enforce database constraints. We emphasize that none of the existing techniques have such capabilities. > How does the proposed method perform if evaluated by the same protocol as the BIRD-SQL leaderboard? We use the exact same protocol as the BIRD leaderboard (e.g. Leaderboard - Execution Accuracy (EX)). Please row 'ex' in Tables 2-5. Other metrics are used in addition to 'ex'. > There is a discussion about the limitation, though the first limitation seems too broad and unnecessary. We respectfully disagree with the reviewer regarding the first limitation. It is a critical capability of our method that we generate queries that satisfy database pattern constraints. While it is not possible for any method to provide full guarantees that a SQL query answers the user’s question at the moment, our approach is a step forward in the direction to produce correct queries. > Presentation comments We will improve presentation taking into accounts reviewer's comments. --- Rebuttal Comment 1.1: Comment: Thank the authors for the response. I've recognized the contribution, so I adjust my score from 2 to 3 accordingly. However, as others and my comments mentioned, the presentation needs improvement to match the quality of the conference. The related work of Text2SQL and automated reasoning should also be discussed more comprehensively. --- Reply to Comment 1.1.1: Comment: Thank you for adjusting the score. We are puzzled by the final reject, given that - Our detailed evaluation demonstrates that this work significantly improves over the current state of the art in zero shot text2sql - We have satisfactorily addressed reviewer's concerns (as indicated in the reviewer's last comment) - Presentation-related issues pointed out by reviewers can be easily addressed in the final version of the paper, and we outlined our suggested improvements in the rebuttal - While we missed related work on schema linking, it does not affect the novelty of this work and can easily be incorporated in the related work section in the final version of the paper. We therefore believe that the reject score is not justified given the novelty of the proposed method and the significance of the results	Rebuttal 1: Rebuttal: We thank you for your comments for their comments! We would like to clarify the following important points: [Main contribution] Our primary contribution is the elimination of the weakest point in LLM-based text-to-SQL methods: reasoning about database constraints. We propose utilizing powerful automated reasoner tools to perform such reasoning. To the best of our knowledge, none of the existing methods provide this capability, which is crucial for large databases with complex relationships. [Database patterns] In this work, we assume that the text-to-SQL developer has to specify database patterns, like many-to-many, snowflake, lookup, etc. It is a reasonable assumption for many SaaS services where the service provider runs data analytics on behalf of customers, for example. [Performance overhead] Our method does require solving CSP and building an intermediate view table. However, the overhead of both these steps is negligible in practice. * Solving CSP is very fast because it is a small and underconstrained model, which can be easily handled by any modern solver (in less than a second). * The view we build is not materialized. Instead, it serves as a sub-query to the final database query. This enables the database query planner to apply standard optimizations such as predicate pushdown to execute the final query efficiently. We would like to emphasize that formal reasoning about a set of logical constraints will remain beyond reach of LLMs in the near future. For example, there is no evidence to suggest that LLMs can replace automated reasoners, such as SAT, SMT, and CSP solvers. Therefore, logical reasoning must be performed by specialized solvers for text2sql tools to be useful in practice.	NeurIPS_2024_submissions_huggingface	2,024	null	null	null	null	null	null	null	null
Quadratic Quantum Variational Monte Carlo	Accept (poster)	Summary: This work introduces a new gradient formulation for Variational Quantum Monte Carlo based on the discretized imaginary time evolution of the Schrödinger equation. In their empirical evaluation, their Q²VMC algorithm consistently converges faster and to lower energies than the traditional VMC objective at no additional cost. Strengths: The paper presents an interesting derivation of a new gradient for the classical objective in VMC. The objective is well-motivated, and the derivation is sound. Further, the empirical results support the theoretical claims of the paper. Weaknesses: The paper struggles to distinguish itself from previous works and lacks sufficient discussion of the surrounding literature. * For someone unfamiliar with QMC (which is likely at NeurIPS), the paper may seem hard to understand due to missing steps in derivations. * The convergence result of the imaginary time evolution is well-known and the basis for diffusion quantum Monte Carlo (DMC) and stochastic reconfiguration in traditional VMC. * A proper discussion of related work is missing—for instance, the connection to Wasserstein QMC [1], diffusion Monte Carlo [2] or stochastic reconfiguration [3]. * "However, one cannot make QVMC perform better than Q²VMC solely by tuning the learning rate." is limited to the evaluation of this work. * The notation is unclear at places, e.g., in Eq. (7) it would help to write the explicit dependence on (x) there as E_0 is not a function of (x) but psi and E_L are. * While the empirical results are consistent, the improvement generally lives in the region of <1mE_h. [1] K. Neklyudov et al., “Wasserstein Quantum Monte Carlo: A Novel Approach for Solving the Quantum Many-Body Schrödinger Equation.” [2] C. J. Umrigar, M. P. Nightingale, and K. J. Runge, “A diffusion Monte Carlo algorithm with very small time‐step errors” [3] S. Sorella, “Generalized Lanczos algorithm for variational quantum Monte Carlo” Technical Quality: 3 Clarity: 1 Questions for Authors: * "Remark 4.3: [...] Asymptotic convergence is guaranteed regardless of the time step size." - Does this mean one can pick the learning rate arbitrarily or very large? This does not seem supported by the results in Table 2. What happens for large lr? * The author writes that the quantum infidelity in Equation (11) is not well understood. Could the author elaborate on what the concrete issue is there? It looks like the 1 - (cosine similarity)² between the two wave functions, which has previously also been used for instance in enforcing orthogonality between wave functions to obtain excited states [1]. * How does the objective influence relative energies (typically much more important than absolute energies)? * Does the Q²VMC objective also work for excited states? [1] M. Entwistle, Z. Schätzle, P. A. Erdman, J. Hermann, and F. Noé, “Electronic excited states in deep variational Monte Carlo,” Overall, I do not recommend acceptance due to the issues outlined in its current form. While I believe the contribution to be sufficient, the discussion of related work is lacking. If the authors address my concerns, I am willing to increase my score. Confidence: 5 Soundness: 3 Presentation: 1 Contribution: 2 Limitations: yes Flag For Ethics Review: ['No ethics review needed.'] Rating: 5 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: ## Weaknesses 1. We appreciate the reviewer’s detailed feedback and suggestions for improving our paper. We acknowledge the challenge to balance the introduction of quantum background concepts for machine learning audiences, and the strict page limit. For the camera-ready version, we have provided a more comprehensive introduction to the background while make explanation of the convergence of the (discrete) evolution more concise. 2. We acknowledge the reviewer’s concern and will include a thorough discussion of Diffusion Monte Carlo (DMC) and stochastic reconfiguration, and more articles will be cited. In particular, our method is closely related to DMC as constantly evolving towards the ground state with Imaginary Time Schrodinger Equation, but keeps a parametric representation; and Q^2VMC is related to stochastic reconfiguration (SR) in terms of how Fisher information matrix naturally arises in the projection. We will also discuss how our method can be seamlessly integrated with Wasserstein QMC(WQMC), providing results showing that our methods can also improve the performance of WQMC. Please find the global rebuttal for experiment details. 3. We have conducted more evaluations and ablation studies to demonstrate the consistent improvements offered by our algorithm. For example, as suggested by other reviewers, we have included combined tuning of learning rate decay and norm constraint of KFAC. 4. To address the presentation issues, we have meticulously proofread the paper and will further improve the clarity and readability in the camera-ready version. 5. Regarding the empirical results and performance of our method, our primary goal was to emphasize that researchers can easily integrate the Q^2VMC method into any of their own implementations without any hyperparameter tuning and achieve significant speed-ups. We regret that this point was not effectively communicated. Consequently, the specific hyperparameters to evaluate our method in the paper can be suboptimal. We have now fine-tuned the hyperparameters for our algorithm, resulting in much better accuracy in the empirical experiments. Please find the global rebuttal for details. Additionally, given the accuracy of the current benckmarks, it will be increasingly difficult to make results better within the framework. Given the size of the system and the absolute values of the "exact" energies in the benchmark, it may not be possible to improve some of the accuracy beyond 1mE_h. ## Questions 1. In Section 4.1 and Remark 4.3, we have clearly stated that our discussion is limited to the nonparametric space of functions. The ability to take a large time step in the Hilbert space unfortunately does not imply the same in the parametric space of neural networks, which depends on the intricate loss landscape and stochastic variance of gradient estimation. Consequently, training a neural network with very large learning rates can lead to divergence, as expected. Greater parametric updates are certainly possible, by e.g. making an inner loop to update parameters after a large step in Hilbert space. However, our primary experiments found this to hinder performance evaluated by optimization time. 2. Quantum infidelity measures the difference between functions, allowing for projections. Details on the definition and usage of quantum infidelity in projections can be found in [1]. The issue with these measurements is that infidelity values are defined on wavefunctions, which can be positive or negative in different regions, complicating the evaluations. In quantum chemistry, we are primarily interested in densities, which depend solely on the probability measure $p\propto\|\psi\|^2 $. Thus, it is much more natural to work directly with the divergence between probability measures, avoiding complications from wavefunctions. This framework is more flexible and allows us to use the KL divergence or any other f-divergences for projection, as in our proposed algorithm. 3. We agree that relative energies are crucial. However, such computations are known to require high computational complexity. Therefore, evaluating an algorithm on ground state energies is a widely adopted practice for assessing performance. Following the reviewer’s suggestion, we conducted experiments on the ionization potentials of several atoms using the LapNet network and were able to finish the testings with atom V within the rebuttal period. The results are: the atom's energy is -943.8785(3), the ion's energy is -943.6417(2), and the ionization potential is 0.2368(4), compared to the benchmark value of 0.2361(2) [2] and experiment value of 0.23733 [4]. 4. We appreciate the reviewer’s suggestion regarding excited states. Our algorithm can, in principle, be used for computing excited states. For example, combined with recent work [3], our algorithm can be adapted with a simple modification of the gradient of the objective function as defined in eq(9). We have added a separate paragraph to discuss some recent works on excited states computation in the related works section. However, thorough justifications and complete experiments to test the performance for excited states are beyond the scope of this paper. [1] Giuliani, Clemens, et al. "Learning ground states of gapped quantum Hamiltonians with Kernel Methods." Quantum 7 (2023): 1096. [2] Li, R., Ye, H., Jiang, D. et al. A computational framework for neural network-based variational Monte Carlo with Forward Laplacian. Nat Mach Intell 6, 209–219 (2024). [3] Pfau, David, et al. "Natural quantum Monte Carlo computation of excited states." arXiv preprint arXiv:2308.16848 (2023). [4] Balabanov, Nikolai B., and Kirk A. Peterson. "Basis set limit electronic excitation energies, ionization potentials, and electron affinities for the 3d transition metal atoms: Coupled cluster and multireference methods." The Journal of chemical physics 125.7 (2006). --- Rebuttal Comment 1.1: Comment: Thanks for the detailed response. There are a few follow-up questions arising from the rebuttal. * Could the authors sketch out the integration with WQMC? * How are the hyperparameters tuned, and are they optimized on a per-structure basis? Optimizing hyperparameters per molecule seems rather tedious in real-world applications. A great benefit of NN-VMC is that it can generally be applied without much tuning. * I disagree with the statement about the quantum infidelity measure. The measure has a clear interpretation, and having both positive and negative values does not complicate computations further. While I agree with the reviewer's sentiment that the density is interesting and offers new pathways forward (like this work), it is not the fundamental object of interest in quantum chemistry. I am convinced that the authors can communicate the value of their work well without distorting the view on quantum chemistry. * I appreciate the author's experiment on ionization potentials. However, it is standard practice in NN-QMC literature to investigate relative energies further. I would appreciate it if the authors could provide relative energies also for different structures, e.g., cyclobutadiene. --- Reply to Comment 1.1.1: Comment: We appreciate the reviewer's prompt response and the additional questions. Below are our detailed answers: ### 1. Integration with WQMC We appreciate the reviewer's interest in how our method integrates with Wasserstein Quantum Monte Carlo (WQMC). To begin, consider the "energy-minimizing 2-Wasserstein gradient flow," expressed in terms of the time derivatives of wavefunctions: $$ \frac{\partial \psi}{\partial t} = \nabla_x \psi(x)^\mathrm{T} \nabla_x E_L(x) + \frac{1}{2} \psi(x) \nabla_x^2 E_L(x) $$ Given any continuous-time gradient flow, discretized after one time step of $\tau$ as: $\psi_{t + \tau} = \psi_t + \tau \frac{\partial \psi}{\partial t}$. We then project the corresponding probability measures by minimizing the KL divergence between the evolved distribution and the updated distribution defined by the neural ansatz. For a general gradient flow, this yields an update of the form: $$ \Delta \theta^\ast \propto F(\theta)^{-1} \mathbb E_{\psi_\theta^2} \left[\left(1 - \tau \left. \frac{\partial \psi}{\partial t} \right \vert_{\psi = \psi_\theta}\right)^2 \nabla_\theta \log \psi^2_\theta (x) \right] - \text{mean} $$ Remarks: Our paper primarily focused on combining probability projection with the "discretized" imaginary time Schrödinger equation, which is more amenable to analysis and exhibits really good convergence properties upon discretization. We fully acknowledge the contribution made by [1], but we did not delve further into how our method could be combined with WQMC due to the complexity of integrating it with other methods. However, following the reviewer's suggestion, we explored this integration. We found that, in principle, our modification could be implemented in code. Empirically, this modification only led to marginal improvements over standard WQMC, as shown in our results table. We suspect this is because, unlike standard imaginary time evolution, the c-Wasserstein gradient flow does not have good properties upon discretization. However, a more detailed analysis is beyond the scope of this paper and would require comprehensive future work. ### 2. Hyperparameter Tuning We acknowledge the reviewer's concern regarding hyperparameter tuning. However, the hyperparameters are certainly not optimized on a per-structure basis, as this would be impractical with our resources. Instead, they were tuned based on performance on the NH$_3$ molecule, as has been extensively studied in our ablation experiments. We then applied the same set of hyperparameters across all tested molecules. Furthermore, based on our experiments and reproduction of the baselines for standard VMC, as well as our ablation studies, we must disagree with the statement that "a great benefit of NN-VMC is that it can generally be applied without much tuning." Our findings suggest that, while a single set of hyperparameters can be applied across different systems, the optimal set of hyperparameters can vary significantly from one system to another. This highlights the need for a more comprehensive study on the impact of hyperparameters on optimization performance, though this is beyond the scope of our current work. ### 3. Quantum Infidelity Measure We appreciate the reviewer's comment regarding the quantum infidelity measure and understand the concerns raised. However, we respectfully disagree with the reviewer's assertion that having both positive and negative values in the wavefunction does not complicate computations. In our view, the presence of these sign changes in the wavefunction can indeed introduce complexities, particularly when projecting onto the parametric space in VMC. That said, we agree that the quantum infidelity measure itself has a clear interpretation and value in quantum chemistry. Our intent was not to diminish its significance but rather to highlight the advantages of focusing on the probability measure directly, which simplifies certain aspects of the computations. We will revise the manuscript to more accurately communicate this balance, emphasizing the value of our work without distorting the broader context of quantum chemistry. ### 4. Additional Experiments on Relative Energies We appreciate the reviewer's suggestion to include more experiments, particularly on relative energies across different structures. However, at the time of this comment, these additional experiments are still ongoing and will take a few more days to complete. We will provide the results of these experiments in the camera-ready version of the paper. Nonetheless, we believe that the experiments already presented in the paper and further in this rebuttal are extensive and sufficient to demonstrate the effectiveness of our proposed algorithm. [1] Kirill Neklyudov, Jannes Nys, Luca Thiede, Juan Carrasquilla, Qiang Liu, Max Welling, and Alireza Makhzani. Wasserstein quantum monte carlo: A novel approach for solving the quantum many-body schrödinger equation. arXiv preprint arXiv:2307.07050, 2023.	Summary: This paper proposed a new QMC method that utilizes the imaginary time evolution of the Schrodinger equation. Unlike Diffusion Monte Carlo (DMC) which uses Langevin dynamics to simulate the dynamic of the imaginary time evolution, this paper suggests a way to perform the update in discrete steps and then project back to the parametric space. Interestingly, under the KL-divergence metric, the projection is equivalent to the regular VMC algorithm with an additional term. Strengths: 1. the author shows that under discretization of neural network parameterization and KL-divergence projection, imaginary time evolution has a similar update rule as the regular VMC. The simplicity of the modification makes the algorithm very easy to adapt. 2. The mathematical derivation for the key step (Appendix B) is clear and easy to follow. Weaknesses: 1. DMC should be cited, and the difference between the proposed method and DMC should be highlighted, as early as in the abstract. Also, a comparison with DMC in terms of convergence and speed should be provided. If the author can address this point I'll consider rasing the score. 2. the improvement in both speed and convergence is a bit marginal. 3. The presentation flow is a bit strange, which makes the paper harder to understand than necessary. For example, the imaginary time evolution should be in the background section. 4. There are some typos as well. For example, in line 252 the two sides of the inequality are the same. I'd suggest the authors to proofread their paper further. 5. There could be more experiments on larger systems to further demonstrate the improvement provided by the new method. Technical Quality: 3 Clarity: 2 Questions for Authors: 1. We know that DMC needs importance sampling to achieve good performance. Why is it not the case here? 2. Do you have an intuitive explanation for why Q^2VMC can perform the ground-state projection as in DMC with only a minor modification to the standard VMC algorithm? 3. How exactly does your method sidestep the sign problem? Confidence: 4 Soundness: 3 Presentation: 2 Contribution: 3 Limitations: The central issue of scaling of the ab initio method is still not addressed. This is evident as the systems tested are small in terms of the number of electrons contained. Flag For Ethics Review: ['No ethics review needed.'] Rating: 5 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We thank the reviewer for their detailed comments and suggestions. ## Weaknesses: 1. We appreciate the reviewer's insight regarding the relation of our work with the widely known Diffusion Monte Carlo (DMC) method. In summary, our method is closely related to DMC in terms of using the Imaginary Time Schrödinger Equation to evolve to the ground state. The key difference is that our method always maintains a parametric representation of the evolved wavefunction throughout the process, as defined by the neural network, and updates the parameters guided by KL divergence projection to track the evolved distribution of the particles (replicas). We have updated our abstract and introduction to emphasize this important relationship, and in the camera-ready version of the paper, we will discuss much more contemporary works in the field of DMC. 2. Regarding the improvement in speed and convergence provided by our method, our primary goal was to highlight that researchers can easily integrate the Q^2VMC method into their implementations without hyperparameter tuning and achieve significant speed-ups. We regret that this point was not effectively communicated. Consequently, the hyperparameters used to evaluate our method in the paper may have been suboptimal. We have now fine-tuned the hyperparameters for our algorithm, resulting in much better accuracy in the empirical experiments. Please refer to Table 1 in the global rebuttal PDF. Additionally, given the current benchmarks' accuracy, achieving even better results within this framework is increasingly difficult. Our tuned algorithm matches the benchmark performance of Psiformer (Large) with Psiformer (Small), where the former has four times the parameters of the latter. 3. We appreciate the reviewer's suggestions regarding the presentation of our paper. While the evolution is background information, we placed it in the following section to intuitively introduce the discretized evolution and its projection after discussing the continuous imaginary time evolution. However, we will make additional efforts to clarify this part in the camera-ready version. 4. Thank you for pointing out the typographical errors. We have meticulously proofread the paper and will further improve its clarity and readability. 5. We aim to scale up our experiments to test the performance of our method on larger systems. However, despite our improvements, the time and computational complexity of QMC remain notoriously high. In our paper, we have tested molecular systems with up to 30 electrons (bicyclobutane), which is considered quite large among most contemporary works. Following the reviewer's suggestion, we attempted to test our method on a larger system, benzene. Due to resource constraints, we have completed 160k out of the total 200k optimization steps with Psiformer (Small) by the rebuttal submission time, which has already required more than 1200 V100 GPU hours. The current inferred energy with our method is -232.2412(2), compared to the benchmarking value of -232.2400(1) trained for the full duration. Full evaluations of more large systems are unfortunately beyond the scope of this work. ## Questions: 1. & 2. The intuition behind our method lies in the gradient expression: $$g=\mathbb E_{\|\psi_\theta\|^2} \left[(1 - \tau E_L(\textbf{x}; \theta))^2 \nabla_\theta \log \psi^2_\theta (\textbf{x})\right]$$ This expectation can be viewed as performing an importance sampling step to sample from the evolved distribution $(1 - \tau E_L(\textbf{x}; \theta))^2\|\psi_\theta\|^2$ towards the ground state, with the current MCMC samples follow the distribution of $\|\psi_\theta\|^2$. The term $\nabla_\theta \log \psi^2_\theta (\textbf{x})$ represents the gradient of the log-distribution used to update the network parameters to match this importance-sampled distribution. A more rigorous derivation minimizes the KL divergence between the evolved distribution and the current parameter distribution, leading to an expression with only a minor modification to the standard VMC algorithm. 3. The sign problem can refer to various but closely related issues in fermionic systems. In DMC specifically, particles are restricted to their initial nodal pockets during the diffusion process and thus previous works rely on the accurate predictions of the nodal surface (e.g., [1]). In our Q^2VMC algorithm, we maintain a parametric representation of the wavefunction while evolving towards the ground state, allowing the particles to evolve with MCMC updates as in standard VMC. Thus, the sign problem should not complicate our algorithm. [1] Ren, Weiluo, et al. "Towards the ground state of molecules via diffusion Monte Carlo on neural networks." Nature Communications 14.1 (2023): 1860. --- Rebuttal Comment 1.1: Comment: Thanks for your detailed response. My questions are well answered, and some of my concerns are addressed. However, due to the limited novelty (as also pointed out by other reviewers) and limited improvement to the baseline method, I will keep my current score. --- Reply to Comment 1.1.1: Comment: We would like to thank the reviewer once again for their insightful suggestions and comments. We appreciate the time and effort invested in evaluating our work. While we understand the reviewer's concerns regarding the perceived novelty and improvements of our method, we believe that our approach offers important advancements in the field. We hope the reviewer has noticed the additional results presented in the global rebuttal, where we demonstrated substantial improvements over the baseline methods. We believe these results, along with our detailed explanations, highlight the value and impact of our proposed algorithm.	Summary: In this paper, the authors propose the Quadratic Quantum Variational Monte Carlo (Q$^2$VMC) algorithm to enhance the optimization process of Quantum Variational Monte Carlo (QVMC). Unlike the standard QVMC, Q$^2$VMC employs an improved projection method to guide the update direction of the ansatz. The authors conduct experiments on various systems and perform an ablation study to demonstrate the effectiveness of their method. Strengths: 1. Designing a better optimization algorithm is useful for most research in this field. 2. Different from some existing improved QVMC algorithms (e.g., WQMC[1]), the method proposed in this paper does not introduce any additional computational burden, making it easy to integrate with existing VMC packages. [1] Kirill Neklyudov, Jannes Nys, Luca Thiede, Juan Carrasquilla, Qiang Liu, Max Welling, and Alireza Makhzani. Wasserstein quantum monte carlo: A novel approach for solving the quantum many-body schrödinger equation. arXiv preprint arXiv:2307.07050, 2023. Weaknesses: 1. The primary concern of the reviewer is the significance of the proposed method. Although the authors provide numerous numerical results, the energy differences between Q$^2$VMC and QVMC are usually within chemical accuracy. As shown in the ablation study section, such small energy differences may influenced by the choice of hyperparameters. While the authors try to demonstrate that under most initial learning rate choices, the performance of Q$^2$VMC is better than QVMC, some important hyperparameters have not been considered for now. For example, Ref.[1] increases the norm constraint of KFAC optimizer and decreases the learning rate decay rate when lowering the initial learning rate. The reviewer suggests a more detailed ablation study to demonstrate the significance of the proposed method. 2. While the proposed update method can be derived from the change of the metric used in the projection step, it is difficult to argue that there is a large difference between the original VMC projection step with the proposed step in Q$^2$VMC. From reviewer's perspective, the original VMC projection, which uses standard inner product in Hilbert space as the metric, may be more reasonable for the quantum systems. The reviewer suggests that the authors should provide a more detailed theoretical analysis about the proposed method. [1]Leon Gerard, Michael Scherbela, Philipp Marquetand, and Philipp Grohs. Gold-standard solutions to the schrödinger equation using deep learning: How much physics do we need? Advances in Neural Information Processing Systems, 35:10282–10294, 2022. Technical Quality: 2 Clarity: 3 Questions for Authors: While the authors focus on improved projection step to derive a better optimization algorithm, the reviewer is curious about if more significant improvements could be achieved by choosing a better target function. Confidence: 3 Soundness: 2 Presentation: 3 Contribution: 2 Limitations: The systems studied in the paper are relatively small, so the performance of the proposed method for larger systems remains unknown. Considering the substantial computational resources required for large-scale experiments, the reviewer does not request the authors to provide concrete numerical results on larger systems during rebuttal. However, the reviewer kindly suggests that the authors provide some evidence to imply the effectiveness of their method on larger systems, e.g., using a smaller network or fewer training step to study a large system. Flag For Ethics Review: ['No ethics review needed.'] Rating: 5 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We thank the reviewer for their detailed comments and suggestions. ## Weaknesses 1. Regarding the improvement in speed and convergence provided by our method, our primary goal was to highlight that researchers can easily integrate the Q^2VMC method into their implementations without hyperparameter tuning, achieving significant speed-ups. We regret that this point was not effectively communicated. Consequently, the hyperparameters used to evaluate our method in the paper may have been suboptimal. We have now fine-tuned the hyperparameters for our algorithm, resulting in much better accuracy in the empirical experiments. Please refer to Table 1 in the global rebuttal PDF. Additionally, thanks to the suggestions from the reviewers, we have tuned some extra optimization hyperparameters of the baseline, including learning rate decay and norm constraint, to demonstrate the consistency of our improvements. Please see Table 3 in the global rebuttal PDF. 2. We believe that the close relationship between our method, Q^2VMC, and the original VMC is a strength rather than a weakness. This relationship enables efficient and easy integration of our method with standard VMC framework, improving results at no additional computational cost. The reason for projecting probability measures using KL divergence instead of the original wavefunction by standard inner product (or fidelity if normalized) is based on the observation that in quantum chemistry, densities, which depend solely on the probability measure $ p \propto \|\psi\|^2 $, are of primary interest. In contrast, the wavefunctions of fermionic systems can be positive in some regions and negative in others, complicating the analysis. Thus, it is more natural, intuitive, and easier to use the neural network to define the probability measures directly. We kindly ask the reviewer what specific theoretical justifications they expect. ## Questions We appreciate the reviewer's question about choosing a better target function. However, we would like to request further clarification on what is meant by this. As detailed in Section 4 of our paper, the proposed optimization loop of the algorithm involves first evolving by a discretized gradient flow, then projecting the evolved distribution onto the parametric distribution of the neural network, and finally updating the MCMC samples. Our work focuses partly on the discretization but mainly on the projection step, the second step in this loop. If the reviewer is referring to designing more efficient gradient flows, which is the first step, this is beyond the scope of this work. A better gradient flow can certainly enhance efficiency, but for any novel gradient flow, our proposed algorithm of projection can still be integrated into the loop. For example, in the PDF of the global rebuttal, we demonstrated consistent improvement of our method integrated with the novel flow of Wasserstein gradient flows [1]. ## Limitations We thank the reviewer for raising the concern about the scalability of our method to larger systems. Scaling up the experiments to test our algorithm on even larger molecules is certainly our goal and the hope of many researchers. However, despite our improvements, the time and computational complexity of QMC remain notoriously high. In our paper, we have tested molecular systems with up to 30 electrons (bicyclobutane), which is already considered quite large among most contemporary works. Following the reviewer's suggestion, we attempted to test our method on a larger system, benzene, which has 42 electrons. Due to resource constraints, we have only completed 160k out of the total 200k optimization steps with Psiformer (Small) by the rebuttal submission time, requiring over 1200 V100 GPU hours. The current inferred energy, averaged over the latest 5000 steps of training with our method, is -232.2412(2), compared to the benchmarking value of -232.2400(1) trained for the full duration. We appreciate the reviewer's suggestion of studying larger systems with smaller networks and fewer training steps. However, scientifically, this would require re-evaluating all baseline systems on the smaller network to reveal the scaling impact. Due to time constraints, we could not complete the full benchmark building yet during the rebuttal, but will definitely add an additional table using a smaller network or fewer training steps to study smaller to larger systems in the camera-ready version once testing is completed. [1] Kirill Neklyudov, Jannes Nys, Luca Thiede, Juan Carrasquilla, Qiang Liu, Max Welling, and Alireza Makhzani. Wasserstein quantum monte carlo: A novel approach for solving the quantum many-body schrödinger equation. arXiv preprint arXiv:2307.07050, 2023. --- Rebuttal Comment 1.1: Comment: Thank you for your response. However, the response does not fully address the reviewer's concern about significance. While the authors claim that the averaged training energy is lower than the inferred baseline by 1.2 mHa on the benzene system, the training energy can be lower than the standard inferred energy, making it hard to evaluate the performance of the proposed method. As a result, the reviewer will keep the score. --- Reply to Comment 1.1.1: Comment: We are very thankful to the reviewer for their comments and insightful feedback. Your suggestions have been very helpful in improving our work and guiding future research directions. We regret that our additional results have not fully addressed your concerns about the significance of our method. We would like to emphasize that we have achieved significant improvements over the baselines on the specific large system within shorter training steps. While we understand the reviewer's concern regarding the comparison between training energy and standard inferred energy, we note that the training energy is not always lower than the inferred energy. As discussed in our paper and rebuttal, evaluating quantum Monte Carlo methods on large systems is extremely resource-intensive, often requiring well over thousands of GPU hours. Unfortunately, given the limited time frame and resources, we could not complete additional tests on even larger systems. We believe that the systems tested in the paper and rebuttal, such as bicyclobutane with 30 electrons, are already considered very large compared to most contemporary works in academia (e.g., [1]). Therefore, the significant improvements we have demonstrated over the baselines (Table 1 in the global rebuttal) are sufficient to establish the importance and effectiveness of our approach. We appreciate your understanding and hope that our existing results adequately convey the significance of our work. [1] Kirill Neklyudov, Jannes Nys, Luca Thiede, Juan Carrasquilla, Qiang Liu, Max Welling, and Alireza Makhzani. Wasserstein quantum monte carlo: A novel approach for solving the quantum many-body schrödinger equation. arXiv preprint arXiv:2307.07050, 2023.	Summary: This paper centers on quantum chemistry, specifically targeting the ground state of molecular systems. Unlike previous methods that apply approximate natural gradient techniques to the wavefunction, Q2VMC executes natural gradient optimization on the distribution. Experimental results demonstrate that Q2VMC significantly enhances energy performance. Strengths: 1. This paper introduces Q2VMC, which can be easily implemented. 2. Faster convergence and enhanced accuracy are clearly demonstrated in the experiments. Weaknesses: 1. You mention Wasserstein quantum monte carlo in your paper. I wonder whether your method can be added on top of WQMC. If so, additional experiments would be great. If not, I would like to see the comparison between WQMC and Q2VMC. 2. Typo: line 122, "wavefunctios" instead of "wavefunctions". Technical Quality: 3 Clarity: 3 Questions for Authors: See Weakness above. Confidence: 4 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: The authors addressed some of the limitations of the proposed algorithm in the conclusion. Flag For Ethics Review: ['No ethics review needed.'] Rating: 5 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: ## Weaknesses and Questions 1. We thank the reviewer for pointing out the typographical error. The typo "wavefunctios" has been corrected to "wavefunctions," and the entire paper has been carefully proofread to address any other typos. 2. Regarding the integration of our proposed methods with Wasserstein Quantum Monte Carlo (WQMC), the answer is yes. Our proposed algorithm, Q^2VMC, is a versatile parametric projection method that can be combined with any gradient flow, including WQMC. This versatility means that if better gradient flows are developed in the future, they can also be integrated with our method. To address the reviewer's request for additional experiments, we have included new results in Table 2 of the overall rebuttal PDF. We compared our method against the standard WQMC baseline using three atoms as in the original paper, and we also included an additional molecule NH$_3$. It is important to note that the baseline WQMC ground state energy values are derived from our own implementation. These values are more accurate than those reported in the original paper because we adhered to a total of 200k training steps, as opposed to the 10k/20k steps used in their experiments, to maintain consistency and avoid confusion. Our results consistently show improvements over the baseline performance, demonstrating the efficacy of our method. --- Rebuttal Comment 1.1: Title: Response Comment: Thank you for your response. It is great to see that your method achieves consistent improvements on both VMC and WQMC. However, due to the limited improvement, I will only raise my score slightly. --- Reply to Comment 1.1.1: Comment: We thank the reviewer for their thorough review and for suggesting the integration of our method with the WQMC approach. However, we would like to emphasize that the integration with WQMC, while valuable, is not the primary focus of our paper. Our work is fundamentally based on the findings that the imaginary time Schrödinger evolution exhibits very good convergence properties upon discretization (as shown in Theorem 4.2). Unfortunately, these same properties do not seem to apply as well to the discretization of the Wasserstein gradient flow. While we fully acknowledge the significant contributions made by WQMC, we did not delve further into how our method could be combined with it due to the complexity and the reasons mentioned above. Nevertheless, following the reviewer's suggestion, we explored this integration. The marginal improvement observed when combining our method with WQMC could be attributed to the fact that, unlike standard imaginary time evolution, the Wasserstein gradient flow may not retain its favorable properties upon discretization, or it might be due to reaching the accuracy limit of the neural network employed. A more detailed analysis of this integration is beyond the scope of our current work and would require comprehensive future research. We hope the reviewer understands that our paper’s primary focus lies in the novel aspects of Q^2VMC itself, and we encourage the reviewer to consider the work in this context. Thank you again for your constructive feedback and for acknowledging the consistent improvements our method has achieved.	Rebuttal 1: Rebuttal: ## Global Rebuttal We thank the reviewers for their thorough and insightful comments. We have carefully considered all feedback and made substantial revisions to address the concerns raised. This global rebuttal is intended to introduce the additional data provided in the accompanying one-page PDF, which contains three tables to further demonstrate the effectiveness and robustness of our proposed Q^2VMC method. ### Table 1: Benchmarking and Hyperparameter Tuning Results This table included the energy values for a set of molecules tested with the Psiformer model, both in its small and large configurations. The benchmarking values from [1] are included for comparison. We present results using the original hyperparameters from [1] and our results with tuned hyperparameters, both with the Psiformer (Small) model. - The table shows that the results with tuned hyperparameters for the small model with out method match or exceed the accuracies of the benchmarking large model. - This addresses the reviewer's concerns about the significance of improvements and the potential impact of hyperparameter tuning on performance. ### Table 2: Integration with Wasserstein Quantum Monte Carlo (WQMC) This table presents the ground state energies for a set of four molecules, optimized using the Wasserstein Quantum Monte Carlo (WQMC) method [2] and combined with our Q^2VMC method. - The results demonstrate that our method consistently improves upon the WQMC baseline, validating the robustness and efficiency of our approach when integrated with other existing gradient flows. ### Table 3: Additional Ablation Study This table provides results of some extra experiments with additional tuned hyperparameters conducted as part of the ablation study, as suggested by the reviewers. It shows the computed ground state energies of NH_3 molecules using the standard quantum Monte Carlo method, tested with different (reduced) learning rates, (increased) learning rate decay times, and (increased) update norm constraints. - It addresses concerns about the potential influence of hyperparameter choices on the baselines and demonstrates that one cannot make standard VMC more accurate than our proposed Q^2VMC algorithm by only tuning the hyperparameters. ### Conclusions We believe these additional results comprehensively address most of the concerns of reviewers, demonstrating the significance, robustness, and ease of integration of our Q^2VMC method. We have made clarifications and provided additional insights in the revised manuscript to ensure a better understanding of our contributions. We are confident that our revisions and the new data will meet the reviewers' expectations and illustrate the value of our work. We thank the reviewers again for their constructive feedback and hope our responses satisfactorily address their concerns. Pdf: /pdf/3da12496fce0eb09ee21ab71688303fe3a6796ee.pdf	NeurIPS_2024_submissions_huggingface	2,024	null	null	null	null	null	null	null	null
Real-time Core-Periphery Guided ViT with Smart Data Layout Selection on Mobile Devices	Accept (poster)	Summary: • This paper proposes ECP-ViT, which optimizes the ViT model by introducing the core-periphery (CP) principle. • The Core-Periphery Principle Guided self-attention mechanisms successfully reduce memory bandwidth by eliminating data transformation. • By applying pruning and removing data transformation, the optimization, which considers both software-level design (algorithm) and hardware-level design, achieves real-time performance on a mobile GPU (Snapdragon 8 Gen 2 SoC) with an average speedup of 4.8 times. Strengths: • The performance generality of the proposed ECP-ViT is demonstrated across various datasets (STL-10, CIFAR-100, TinyImageNet, and ImageNet) and hardware environments (OnePlus 11, Xiaomi 6). • The justification for the proposed Core-Periphery Guided Self-Attention is well-explained, including background information, and effectively illustrated with figures and equations. • ECP-ViT achieves speed improvements by completely eliminating data transformation, which impacts memory bandwidth and causes network overhead. • (Table 8) The compiler speed is significantly improved compared to the MNN and TVM frameworks. • (Table 9) The overhead of data layout transformation and computation for the mobile GPU is analyzed in detail. Weaknesses: • The comparative experiments lack consistency. • Overall, the performance improvements are marginal compared to previous research. • The implementation uses Fixed Point 16-bit, but most current mobile environments typically utilize 8-bit or higher quantization, and there are no experiments addressing this. Data compression through quantization is essential for mobile environments. • More fair comparisons could be made if experiments were conducted in specific NPU environments. • Although a compiler for mobile GPUs is proposed, detailed information about the compiler itself is lacking. o (Figure 4) There is insufficient explanation on how the specific allocation of the core/periphery nodes is determined, and there is no criterion for specifying core nodes. o The detailed operation of the proposed dimension reduction heuristic algorithm is not explained. More detailed descriptions and algorithms are needed to optimize data layout. Technical Quality: 3 Clarity: 3 Questions for Authors: • (Table 1, Table 9) Are there experimental results on latency for layout transformation and computation in other frameworks such as TFLite and Pytorch-Mobile? o The experimental setup in Table 9 indicates a batch size of 1 to 18. Is this an average value? Do the experimental results for latency ratios align with different batch sizes? • (Table 3) Are there any comparative experimental results for Pytorch-Mobile? • (Figure 6) Are there experimental results on the computation complexity and memory usage of the model according to the core ratio? Are there results consistent with those presented in Table 1? • (Table 6) Were the experiments for TNN, TVM, and MNN conducted on NPUs while those for Ours were conducted on a mobile GPU? Are there experimental results for Ours on NPUs? • Are there experimental results for tasks such as Object Detection or Instance Segmentation, where real-time inference is crucial? Confidence: 3 Soundness: 3 Presentation: 3 Contribution: 2 Limitations: N/A Flag For Ethics Review: ['No ethics review needed.'] Rating: 4 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: ### Q1 (Table 1, Table 9) Are there experimental results on latency for layout transformation and computation in other frameworks such as TFLite and Pytorch-Mobile? Response: \| Model \| Implicit Transformation (ms) \| Explicit Transformation (ms) \| Computation (ms) \| Latency (ms) \| \|-------------------------\|------------------------------\|------------------------------\|------------------\|--------------\| \| ECP-ViT (our framework) \| 0 \| 0 \| 99.83 \| 99.83 \| \| ViT-Base \| 198.93 \| 0.92 \| 134.52 \| 334.37 \| \| DeiT-Base \| 216.26 \| 5.294 \| 136.09 \| 357.64 \| Device – Oneplus 12 Snapdragon 8 Gen 3 SoC Platform – TFLite ### Q2 The experimental setup in Table 9 indicates a batch size of 1 to 18. Is this an average value? Do the experimental results for latency ratios align with different batch sizes? Response: Thank you for your question regarding the experimental setup in Table 9. We actually tested all batch sizes from 1 to 18 to ensure comprehensive evaluation. However, we observed that there were no major differences in latency across these different batch sizes. As a result, for simplicity and clarity in our presentation, we reported the results for batch size 1. The latency ratios remained consistent regardless of the batch size, ensuring that our conclusions are robust across varying batch sizes. ### Q3 (Table 3) Are there any comparative experimental results for Pytorch-Mobile? Response: Thank you for your question regarding comparative experimental results for Pytorch-Mobile in Table 3. Currently, Pytorch-Mobile does not support ViT model on mobile GPU. Therefore, we use TensorFlow Lite (TFLite) for our experiments instead and conducted all tests using OpenCL on GPUs. This approach allows us to provide a fair and comprehensive comparison of performance across different frameworks that support Vision Transformer models on mobile GPUs. Please refer to Q7 for the reason of choosing GPU instead of NPU. \| Model \| Framework \| Latency (ms) \| \|------------\|-----------\|--------------\| \| ECP-ViT \| Ours \| 99.83 \| \| ViT-Base \| TFLite \| 334.37 \| \| DeiT-Base \| TFLite \| 357.64 \| Device – Oneplus 12 Snapdragon 8 Gen 3 SoC Results are averaged by running 10 rounds and 5 warm_ups ### Q4 (Figure 6) Are there experimental results on the computation complexity and memory usage of the model according to the core ratio? Are there results consistent with those presented in Table 1? Response: \| Model \| Peak Memory In core ratio 0.6 \| Peak Memory In core ratio 0.7 \| Peak Memory In core ratio 0.8 \| Peak Memory In core ratio 0.9 \| \|------------\|-------------------------------\|-------------------------------\|-------------------------------\|-------------------------------\| \| ECP-ViT \| 361 MB \| 378 MB \| 403 MB \| 420 MB \| These results are consistent with those presented in Table 1. After pruning, we separate computation into core-to-core and core-to-periphery nodes. The small size of core nodes results in intermediate results being similar across different core ratios. Additionally, our layout elimination technique stores intermediate results on the GPU and reuses finished buffers, saving more memory compared to other frameworks. This efficient handling keeps memory usage stable across different core ratios, aligning with the results in Table 1 and demonstrating the effectiveness of our ECP-ViT model in managing computation complexity and memory usage. ### Q5 Are there experimental results for tasks such as Object Detection or Instance Segmentation, where real-time inference is crucial? Response: \| \| DETR [1] \| SegFormer [2] \| \|-----------------\|----------\|---------------\| \| Vanilla \| 42.0 \| 83.8 \| \| w/ Our Method \| 42.6 \| 84.2 \| [1] End-to-end object detection with transformers. In ECCV 2020. [2] SegFormer: Simple and efficient design for semantic segmentation with transformers. In NeurIPS 2021. ### Q6 Why not use quantization? Response: Our paper focuses on layout elimination techniques to reduce memory burden without sacrificing accuracy. Our main contribution lies in optimizing data layout and computation patterns, eliminating data transformation overhead, and enhancing performance on mobile devices. It’s important to note that quantization and our optimization techniques are orthogonal and can be applied independently or together for cumulative benefits. We plan to extend our research to include experiments with 8-bit quantization, aligning with current mobile environment practices. ### Q7 Why Not using NPU? Response: All the experiments were conducted on mobile GPUs because most frameworks like MNN or TNN do not support NPUs. We did not conduct experiments using NPUs because mobile NPUs, while faster than GPUs, lack low-level programmable interfaces for individual developers [1]. TFLite uses an NPU backend via system calls provided by the Android Runtime System, which do not provide an interface for independent developers to support or optimize certain operators [1]. We have to put the comparison table in the PDF due to space limit. While the NPU theoretically outperforms the GPU, our GPU implementation of ECP-ViT runs at almost the same speed as TFLite’s NPU, demonstrating our framework’s efficiency in leveraging GPU capabilities effectively. Qualcomm AI hub devices do not support Oneplus 12 yet, for TFLite with NPU we are using remote device Samsung Galaxy S24+ provided by Qualcomm AI hubs [1] [1] The Qualcomm® AI Hub Models https://github.com/quic/ai-hub-models?tab=readme-ov-file	Summary: This paper introduces ECP-ViT, a framework that accelerates Vision Transformers on mobile devices using a core-peripheral guided self-attention mechanism. This approach reduces computational demands and achieves up to 16.1x speedup on a OnePlus 11 GPU, enabling efficient real-time deployment of ViT models while maintaining high accuracy. Strengths: - The compiler and model architecture were co-optimized, significantly improving computational efficiency on actual devices. - Based on the Brain Neural Networks, this paper cleverly distinguishes between important and less important parts of the attention mechanism. - The slice and transpose reshape operations are eliminated through the hardware-level design. Weaknesses: - Overall, I think this paper is above the acceptance bar. However, the comparison (or discussion) between ECP-ViT and other lightweight ViT methods is not comprehensive enough, for example, NAS for lightweight ViT [1][2], and advanced token optimization method [3]. - Showing the results on ImageNet seems like a reasonable choice, but insufficient to show that the method can be generalized to other domains, such as detection/segmentation. One possible way to do this is to directly transfer the weights for another downstream task and provide some comparison. By the way, the performance improvements in Tab.11 seem a bit marginal compared to the main results. [1] Elasticvit: Conflict-aware supernet training for deploying fast vision transformer on diverse mobile devices, ICCV 2023 [2] Nasvit: Neural architecture search for efficient vision transformers with gradient conflict-aware supernet training, ICLR 2022 [3] Diffrate: Differentiable compression rate for efficient vision transformers, ICCV 2023 Technical Quality: 4 Clarity: 3 Questions for Authors: How does the core rate be selected for different devices and tasks? Would making it a sample-aware parameter further enhance the results? Confidence: 5 Soundness: 4 Presentation: 3 Contribution: 3 Limitations: Please refer to the weakness. Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: ### Q1 More comparison between other lightweight ViT methods. [1] Elasticvit: Conflict-aware supernet training for deploying fast vision transformer on diverse mobile devices, ICCV 2023 [2] Nasvit: Neural architecture search for efficient vision transformers with gradient conflict-aware supernet training, ICLR 2022 [3] Diffrate: Differentiable compression rate for efficient vision transformers, ICCV 2023 Response: \| Model \| Top1 Acc on ImageNet-1k \| Computation size (GFLOPs) \| \|---------------\|-------------------------\|--------\| \| ECP-ViT \| 84.6 \| 10.1 \| \| ElasticViT-L3 \| 80.0 \| 0.86 \| \| NasViT-A5 \| 81.8 \| 0.76 \| \| DiffRate-Base \| 81.5 \| 11.5 \| As shown in the table, ECP-ViT achieves the highest top-1 accuracy on ImageNet-1k at 84.6%, demonstrating superior performance compared to other lightweight ViT methods. We will add the comparison and the speed comparison in the revision ### Q2 Show results on other domains. Response: Thank you for the suggestion. We have further applied the core-periphery guided ViT to detection and segmentation tasks, specifically using DETR on the COCO dataset and SegFormer on the Cityscapes dataset. \| \| DETR [1] \| SegFormer [2] \| \|-----------------\|----------\|---------------\| \| Vanilla \| 42.0 \| 83.8 \| \| w/ Our Method \| 42.6 \| 84.2 \| [1] End-to-end object detection with transformers. In ECCV 2020. [2] SegFormer: Simple and efficient design for semantic segmentation with transformers. In NeurIPS 2021. A major contribution of our work emphasizes real-time performance on mobile devices. Our method reduces parameters and latency while maintaining or even improving model performance on devices. Although the performance improvements in Table 11 may seem marginal, they demonstrate that our method achieves a balance between efficiency and effectiveness. ### Q3 How does the core rate be selected for different devices and tasks? Would making it a sample-aware parameter further enhance the results? Response: Thank you for your question regarding the selection of the core ratio for different devices and tasks. Since our core-periphery principle guided ViT is inspired by brain functional networks, where different networks responsible for various tasks (such as vision and movement) have different core ratios, we have adopted a similar approach that we experiment with different core ratios to find the optimal ratio for specific datasets or tasks. Your suggestion to make it a sample-aware parameter is indeed interesting and a great idea for further exploration. According to Figure 8 and Figure 6, we employ different core ratios to run benchmarks and obtain a performance line in terms of accuracy. We then test the latency for these different core ratios. By evaluating both accuracy and latency metrics, we can identify the optimal core ratio that balances performance and efficiency. This systematic approach ensures that the selected core ratio is well-suited for the specific device and task, providing optimal results. Additionally, making the core ratio a sample-aware parameter could potentially further enhance results by allowing dynamic adjustment based on the specific characteristics of the data, thereby improving both accuracy and efficiency.	Summary: This paper presents ECP-ViT, a real-time framework for deploying Vision Transformers (ViTs) on mobile devices. Inspired by the brain's core-periphery principle, this method guides self-attention in ViTs to reduce computational demands and eliminate data transformation operations. ECP-ViT integrates algorithm-system co-optimizations, achieving a speedup of 4.6× to 26.9× on mobile GPUs across various datasets while maintaining high accuracy. Strengths: 1. The motivation data in Table 1 is interesting. 2. This paper targets an interesting problem: eliminating the expensive transform operation in ViTs. Weaknesses: 1. The paper writing can be improved. 2. Some figures and their elaborations are not clear enough. 3. The rationale behind the design is not well-explained. Please see my comments. Technical Quality: 2 Clarity: 2 Questions for Authors: 1. In line 67 of the introduction, it mentions support for a pruning scheme, which is confusing because there is no prior content about pruning before this line. 2. In the background section, it states that the token pruning method introduces additional Reshape and Transpose operations to the feature map, leading to reduced benefits from reduced computation costs. However, in many token pruning methods [1] [2], pruning is applied to certain specific layers, so the overhead of Transpose and Reshape is not that large. Moreover, token pruning quadratically reduces computation cost, which is highly effective. 3. In Section 3.1, it says it classifies data transformation into two categories, Explicit and Implicit, and Explicit transformations are denoted as red color operators in round boxes. But there are several colors in Figure 2, and some colors are quite similar, making it hard to distinguish between Explicit and other transformations. Also, it seems that there is no indication of the implicit transformations in the figure. 4. It says the CP-guided self-attention design is shown in Figure 4. But Figure 4 seems to depict the ratio of core nodes and the number of edges. This is confusing as it does not illustrate any design component or workflow. 5. The rationale behind the CP-guided self-attention design is not detailed enough. CP-guided self-attention is an efficient self-attention mechanism that reduces the number of self-attention operations, but why is it better than other efficient self-attention mechanisms such as Swin-Transformer and Criss-Cross Attention? Why do we adopt this CP-guided self-attention? How do you ensure that the information exchange through only core nodes is sufficient?   [1] Liang, Youwei, et al. "Not all patches are what you need: Expediting vision transformers via token reorganizations." arXiv preprint arXiv:2202.07800 (2022). [2] Kong, Zhenglun, et al. "Peeling the onion: Hierarchical reduction of data redundancy for efficient vision transformer training." Proceedings of the AAAI Conference on Artificial Intelligence. Vol. 37. No. 7. 2023. Confidence: 3 Soundness: 2 Presentation: 2 Contribution: 3 Limitations: Please see my comments. Flag For Ethics Review: ['No ethics review needed.'] Rating: 3 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: ### Q1 Confusions in line 67 of the introduction. Response: Thank you for the suggestion. We will revise the introduction to provide a brief context about pruning before mentioning our support for it. Specifically, we will add a paragraph to explain the concept and relevance of pruning in the context of our work. ### Q2 Token pruning methods add some overhead with Reshape and Transpose operations but are still highly effective in reducing computation costs by being applied to specific layers. [1] Liang, Youwei, et al. "Not all patches are what you need: Expediting vision transformers via token reorganizations." arXiv preprint arXiv:2202.07800 (2022). [2] Kong, Zhenglun, et al. "Peeling the onion: Hierarchical reduction of data redundancy for efficient vision transformer training." Proceedings of the AAAI Conference on Artificial Intelligence. Vol. 37. No. 7. 2023. Response: Thank you for your feedback regarding the token pruning methods. Token pruning methods [1][2] can result in sparse data patterns that implicitly require more data transformations. Implicit data transformations occur when the underlying data layout needs to be reorganized to optimize performance for specific operations, leading to irregular data access patterns and increased memory bandwidth demands. This can be particularly unfriendly for mobile devices, which have limited memory bandwidth. In ECP-ViT, we address this issue by grouping important patches together and pruning at the node level. This approach ensures that even though the model is pruned, each sub-matrix remains dense, allowing for more efficient computation. By maintaining dense computation patterns, ECP-ViT minimizes the need for implicit data transformations, thereby enhancing performance and efficiency on mobile devices. This makes ECP-ViT more suitable for mobile deployment compared to methods that result in sparse data patterns and increased implicit data transformations. ### Q3 In Section 3.1, it states that data transformations are classified into Explicit (red, round boxes) and Implicit categories, but Figure 2 uses several similar colors making it difficult to distinguish Explicit transformations, and lacks indication of Implicit transformations. Response: Thank you for your observation regarding the classification of data transformations in Section 3.1 and their representation in Figure 2. We will update the hardware design in the revision. In the updated figure, every computational operator will have its preferred data layout, and between each computational operator, there will potentially be an implicit reshape or transpose. These implicit reshapes occur because different computational operators often require their input data in specific formats or layouts for optimal performance. When the output layout of one operator does not match the required input layout of the next, an automatic reshape or transpose is needed to ensure compatibility and efficiency. In our framework, by using smart layouts, we will eliminate the implicit reshapes. Additionally, we will ensure that the colors used to denote different types of operations are distinct and easily distinguishable. ### Q4 Confusions in Figure 4 (CP-Guided self-attention) Response Thanks for pointing this out. We will revise the sentence to clarify that Figure 4 shows examples of core-periphery graphs with different core ratios. Specifically, Figure 4 illustrates the selection of various CP graphs referenced in Figure 2(a). The workflow involves CP graph generation, CP-guided self-attention, and CP-guided QKV multiplication, corresponding to Figure 2(a), Figure 2(b1), and Figure 2(b2). ### Q5 More details are needed for CP-guided self-attention design. Why is it better than Swin and Criss-Cross Attention? Only core nodes are sufficient? Response: Thank you for your question regarding the rationale behind the CP-guided self-attention design and its advantages over other efficient self-attention mechanisms such as Swin-Transformer and Criss-Cross Attention. From a hardware perspective, CP-guided self-attention can group core patches together and periphery patches together, which increases data locality. This grouping improves the efficiency of data access and reduces the need for frequent data movement across memory, thereby enhancing computational efficiency. While this design involves more implicit and explicit reshape and transpose operations, our framework successfully eliminates both, making it highly suitable for CP-guided self-attention. In contrast, mechanisms like Swin-Transformer and Criss-Cross Attention do not inherently focus on optimizing data locality and may still involve significant overhead from data transformations. By reducing these overheads, CP-guided self-attention becomes more efficient and effective, particularly on hardware with limited memory bandwidth like mobile devices. Additionally, the core-periphery principle ensures that information exchange through core nodes is sufficient by maintaining connections among core nodes while selectively connecting periphery nodes, balancing comprehensive context capture with reduced computational complexity. This makes CP-guided self-attention a better fit for our optimization framework compared to other mechanisms. Furthermore, our method is actually a pruning mechanism that can also be applied to Swim-Transformer or Criss-Cross Attention. This flexibility allows for the benefits of our approach to be realized in various self-attention-based transformer models, enhancing their efficiency and performance across different applications and hardware configurations.	Summary: This paper introduces ECP-ViT, a framework designed to improve the performance of vision transformers (ViTs) on mobile devices. The authors observed the intensive and irregular memory access involved in the data transformation for self-attention layers, which significantly slows down transformers compared to traditional CNNs. To address this, they propose a hardware-friendly self-attention pruning technique motivated by the core-periphery structure in brain networks, thereby reducing the computational and memory access burdens. ECP-ViT also incorporates compiler optimizations to fuse and eliminate transformation operators to further boost the performance. These combined optimizations enable the work to speedup in ViT inference on mobile GPUs by 4.6x to 26.9x without sacrificing the inference accuracy. Strengths: + The problem is well-motivated with clear benchmarks in terms of accuracy, MACs, and latency comparisons across various ViTs. + The core-periphery concept borrowed from brain networks for self-attention is interesting. + The proposed compiler optimization effectively eliminates unnecessary data transformation operators, leading to significant performance improvements. + The evaluation is conducted on real-world mobile phones, and includes both high-end and low-end devices. Weaknesses: - The evaluation section could be improved: - It would be beneficial to discuss the memory access pattern differences before and after applying ECP to showcase its effectiveness in improving memory efficiency. - The power and energy consumption evaluation should also be included, given the limited battery capacity of mobile devices. - There are several presentation issues; for example, the space between Table 3 and Table 4 is too dense, and the font size seems to change abruptly starting from the “Evaluation environment” in Section 4. Please ensure consistent font sizes throughout the text. Technical Quality: 3 Clarity: 2 Questions for Authors: 1. Can you provide a detailed comparison of memory behavior before and after applying the ECP technique? 2. Can you show the power/energy improvement achieved by the proposed optimizations? Confidence: 3 Soundness: 3 Presentation: 2 Contribution: 3 Limitations: The authors have mentioned the plan to evaluate other model architectures in future work, which addresses some limitations. However, it would be beneficial to include additional evaluations such as memory studies, power/energy consumption analysis, and resource utilization metrics. Flag For Ethics Review: ['No ethics review needed.'] Rating: 7 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: ### Q1 Need to provide a detailed comparison of memory behavior before and after applying the ECP technique. Response: Thank you for your question regarding the memory access pattern differences before and after applying ECP-ViT. Our pruning technique ensures that computations remain dense, maintaining a memory access pattern similar to the original model. This consistency in memory access patterns, combined with fewer computations, results in faster performance. (We will add a fig to illustrate the memory access in the revision) If we just use naive fusion, the access pattern of the tensors between operators will be strided. This strided access pattern is determined by the data transforming operators like slice and transpose. By using our smart layout design, we can map the indices of tensors as desired, and the access pattern will be continuous. Compared to the strided access pattern, our access pattern can utilize the data locality and reduce the cache miss, which shows tremendous reduction in latency. \| Model \| Peak Memory w/ ECP (MB) \| Peak Memory w/o ECP (MB) \| Latency w/ ECP (ms) \| Latency w/o ECP (ms) \| \|-----------\|-------------------------\|--------------------------\|---------------------\|----------------------\| \| ViT-Base \| 403 \| 454 \| 99.84 \| 421.25 \| Device – Oneplus 12 Snapdragon 8 Gen 3 SoC Results are averaged by running 10 rounds and 5 warm-ups \| Model \| L1 Cache Miss Rate \| L2 Cache Miss Rate \| L3 Cache Miss Rate \| \|-----------\|--------------------\|--------------------\|--------------------\| \| ViT \| 0.77% \| 5.94% \| 15.12% \| \| ECP-ViT \| 0.66% \| 5.38% \| 15.05% \| After ECP pruning, each part of our computation is still dense. Combining with our layout selections, we are able to increase data locality and reduce cache misses in all levels. ### Q2 Need to show the power/energy improvement achieved by the proposed optimizations Response: Testing device: OnePlus 12, Snapdragon 8 Gen 3 SoC All results are collected by running 10 rounds and pick the peak power usage \| Model \| Power w/o ECP \| Power w/ ECP \| \|------------------\|---------------\|--------------\| \| ViT-Base \| 2.46 W \| 1.05 W \| \| Swin-Tiny \| 0.87 W \| 0.78 W \| \| MetaFormer-Base \| 2.74 W \| 1.31 W \| ### Q3 Table 3, 4 and section 4 font issues Response: Thanks for the suggestions. We will revise it in the revision. --- Rebuttal Comment 1.1: Comment: Thanks for the additional experiments and explanations. I will keep my score.	Rebuttal 1: Rebuttal: ### Q1 (Figure 4) How are core nodes determined? Response: In our ECP-ViT, we consider the image patches as nodes and predefine a series of core ratios to divide nodes into cores and peripheries. During training, we employ Grad-CAM to identify important regions of the images and assign the core nodes to those regions. Accordingly, the QKV matrices of these patches are divided into core and peripheral components. For example, for images with a resolution of 224x224 and a patch size of 16x16, there are a total of 196 patch tokens as nodes. For a core ratio of 10%, around 20 patch tokens are considered as cores, and we choose the top 20 important regions as cores. This partitioning method is inspired by human brain networks [1], where different networks exhibit different core ratios. [1] "Gyri vs. sulci: Disentangling brain core-periphery functional networks via twin-transformer," MICCAI 2024. ### Q2 Need to explain dimension reduction heuristic algorithm. Response: Thank you for your feedback regarding the description of the dimension reduction heuristic. The detailed pseudo code for the heuristic has been attached in the uploaded PDF. The score is collected by running mini benchmarks, and the different layouts are defined by the frameworks. The process begins by identifying key nodes in the computational graph that perform actual computation (Step 1). For each key node, possible data layouts are determined (Step 2), taking into account the various layout options provided by the frameworks. Each possible layout is then evaluated by calculating a score based on data locality and GPU utilization using mini benchmarks (Step 2.1). The layout with the best score is selected and assigned to the key node (Step 3). Finally, the computational graph is updated with the new layouts for all key nodes to ensure that the entire graph benefits from the optimized layouts, improving overall performance (Step 4). Pdf: /pdf/44e134702027d54e306a894d4ea17ccb8a66e4f4.pdf	NeurIPS_2024_submissions_huggingface	2,024	Summary: This work proposes a ViT accelerating framework, ECP-ViT, to deploy ViT models on smartphones. This framework consists of two parts: 1) Core-Periphery Guided Self-Attention (reducing the computational and bandwidth cost of ViT) and 2) Data Layout Selection based on compiler optimizations (removing the time-consuming data transformation). Specifically, Core-Periphery Guided Self-Attention partitions the tokens in K, Q, and V matrices into core and periphery components. Tokens in the core component exchanged messages with all tokens, and tokens in the periphery component only exchanged messages with tokens in the core component. To fully eliminate the data transformation operations, this work attempts to find a common data layout that works efficiently for both contiguous operators. But the search space is large. They propose the dimension reduction heuristic algorithm to shrink the search space of the data layouts. ECP-ViT obtains competitive results on STL-10, CIFAR100, TinyImageNet, and ImageNet, achieving lower smartphone latency. Strengths: - The Core-Periphery Guided Self-Attention, which is inspired by Brain Neural Networks, is interesting. - The Data Layout Selection based on compiler optimizations can effectively speed up ViTs on Smartphones, as demonstrated by comprehensive experimental validation. Weaknesses: - Some details need to be elaborated on. The description of the dimension reduction heuristic needs to be more detailed, and it is encouraged to provide its pseudocode. - There seems to be a discrepancy in Equation 1. The conventional form of self-attention is softmax(qV)K, while in ECP-ViT, it appears to be softmax(qVK). This discrepancy should be carefully reviewed and corrected. Technical Quality: 3 Clarity: 3 Questions for Authors: 1. How do you partition the KQV matrices into core and peripheral components? Are there any partitioning criteria? And why do you use this partitioning method? 2. MobileViT [1] is a lightweight, low-latency network for mobile vision tasks. What significant advantages does ECP-ViT offer over MobileViT that we should consider? [1] Mehta, Sachin, and Mohammad Rastegari. "Mobilevit: light-weight, general-purpose, and mobile-friendly vision transformer." arXiv preprint arXiv:2110.02178 (2021). Confidence: 3 Soundness: 3 Presentation: 3 Contribution: 2 Limitations: Not applicable. Flag For Ethics Review: ['No ethics review needed.'] Rating: 5 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: ### Q1 Explain dimension reduction heuristic Response: Thanks for pointing this out. Below is the detailed pseudo code for the heuristic. The score is collected by running mini benchmarks, and the different layouts are defined by the frameworks. The process begins by identifying key nodes in the computational graph that perform actual computation (Step 1). For each key node, possible data layouts are determined (Step 2), taking into account the various layout options provided by the frameworks. Each possible layout is then evaluated by calculating a score based on data locality and GPU utilization using mini benchmarks (Step 2.1). The layout with the best score is selected and assigned to the key node (Step 3). Finally, the computational graph is updated with the new layouts for all key nodes to ensure that the entire graph benefits from the optimized layouts, improving overall performance (Step 4). ### Algorithm 1: Dimension Reduction Heuristic Algorithm Require: Computational graph G Ensure: Optimized computational graph G' Step 1: Find all the key nodes - Identify key nodes K ← identify_key_nodes(G) - For each node n ∈ K do: - Step 2: Determine layouts for key nodes by Algo 3 - P ← determine_possible_layouts(n) - best_layout ← None - best_score ← ∞ - For each layout l ∈ P do: - Step 2.1: Calculate scores by mini-benchmarks - score ← calculate_data_locality_score(l, n) + calculate_gpu_utilization_score(l, n) - If score < best_score then: - best_score ← score - best_layout ← l - End for - Step 3: Assign best layout for the key node - n.data_layout ← best_layout - End for - For each node n ∈ K do: - Step 4: Infer layout for all nodes - G.update_node_layout(n) - End for - Return G ### Q2 A discrepancy in Equation 1. Response: Thank you for pointing this out. We will correct it in Eq.1 and the corresponding figures. As stated in Equation 2, our ECP-ViT follows the conventional form of self-attention, which is softmax(QK)V. ### Q3 How to determine the core and peripheral components. Response: In our ECP-ViT, we consider the image patches as nodes and predefine a series of core ratios to divide nodes into cores and peripheries. During training, we employ Grad-CAM to identify important regions of the images and assign the core nodes to those regions. Accordingly, the QKV matrices of these patches are divided into core and peripheral components. For example, for images with a resolution of 224x224 and a patch size of 16x16, there are a total of 196 patch tokens as nodes. For a core ratio of 10%, around 20 patch tokens are considered as cores, and we choose the top 20 important regions as cores. This partitioning method is inspired by human brain networks [1], where different networks exhibit different core ratios. [1] "Gyri vs. sulci: Disentangling brain core-periphery functional networks via twin-transformer," MICCAI 2024. ### Q4 Comparison with MobileViT Response: Thank you for your question regarding the advantages of ECP-ViT over MobileViT. While MobileViT is a lightweight, low-latency network designed for mobile vision tasks, it achieves a top-1 accuracy of 78.4% on ImageNet. In contrast, our ECP-ViT achieves a significantly higher top-1 accuracy of 84.6% on the same dataset, demonstrating superior performance. Additionally, ECP-ViT employs a pruning method applied to the self-attention layer, which is an orthogonal approach to MobileViT’s design. Importantly, after pruning, each operation in ECP-ViT remains dense, making them friendly for execution on mobile GPUs. Our framework further eliminates all implicit and explicit data transformations caused by pruning and model design, which significantly reduces latency. This makes ECP-ViT highly efficient and suitable for deployment on mobile devices, offering substantial improvements in both performance and efficiency for self-attention-based transformer models. --- Rebuttal Comment 1.1: Comment: Thank the authors for their effort in the rebuttal. The feedback addressed most of my concerns. I raised my initial score from 4 to 5.	null	null	null	null	null	null
John Ellipsoids via Lazy Updates	Accept (poster)	Summary: The paper proposes provably faster algorithm for computing John ellipsoids. The first idea is to approximate Leverage score with an existing faster algorithm given by Theorem 1.3. The second idea is to use fast matrix multiplication and lazy updates to reduce polylog suboptimality. Strengths: - Provably faster algorithm proposed Weaknesses: - despite that the main result does not rely on Theorem 1.3, there is no proof or reference for it. - contribution should clarified. The proof of Theorem 1.6 mainly relies on the existing results. Technical Quality: 2 Clarity: 2 Questions for Authors: Could you please clarify on your particular contribution? Could you please refer to the proof of Theorem 1.3? I am not an expert in John Ellipsoids. For me it seemed that the contribution was incremental, because the theoretical result combines existing techniques. By the way, the new result is obtained. I would also like you to comment on the most recent papers and preprints from 2023 and 2024 if any, to confirm your contribution. Confidence: 3 Soundness: 2 Presentation: 2 Contribution: 2 Limitations: NA Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: > despite that the main result does not rely on Theorem 1.3, there is no proof or reference for it. > Could you please refer to the proof of Theorem 1.3? The proof of Theorem 1.3 is sketched in the paragraphs immediately below it. We first summarize the result of [DMMW12]. It is known that we can compute $Q$ such that $(1+\epsilon)Q \preceq A^\top A \preceq (1-\epsilon) Q$ in time $\mathrm{nnz}(A) + \mathrm{poly}(d/\epsilon)$ time [CW13], so the leverage scores of $A$ are the row norms of $A Q^{-1/2}$ up to a $(1\pm\epsilon)$ factor. We can then apply the Johnson--Lindenstrauss lemma and multiply this matrix by a random projection matrix $P$ with $t = O(\epsilon^{-2}\log n)$ rows, so that the row norms of $A Q^{-1/2} P$ approximate the row norms of $A Q^{-1/2}$ up to a $(1\pm\epsilon)$ factor. Thus, we indeed have that the leverage scores are approximated by the row norms of some matrix $AR$ for $R = Q^{-1/2}P$ up to $(1\pm\epsilon)$ factors. Finally, we can efficiently multiply $A$ by $Q^{-1/2}P$ quickly using Corollary 1.5. In the revision, we will include a full proof in the appendix. > contribution should clarified. The proof of Theorem 1.6 mainly relies on the existing results. > Could you please clarify on your particular contribution? Our result gives the fastest known algorithm for computing $(1+\epsilon)$-approximate John ellipsoids when $\epsilon$ is not too small. While we indeed build on existing techniques, and in particular the iterated leverage score algorithm of [CCLY19], we make two key improvements to this algorithm: (1) the use of lazy updates that allow us to better exploit the faster running time of leverage scores by batching the leverage score computations, and (2) the use of fast matrix multiplication to give a faster algorithm for approximating leverage scores. Finally, these ideas also give low-space streaming algorithms. > I would also like you to comment on the most recent papers and preprints from 2023 and 2024 if any, to confirm your contribution. To the best of our knowledge, the most recent relevant work on computing $(1+\epsilon)$-approximate John ellipsoids is the works of [CCLY22] and [SYYZ22], which are discussed thoroughly in the introduction. We are not aware of works from 2023 or 2024 on this topic.	Summary: The authors give a leverage score approximation algorithm that runs in nearly linear time for dense matrices. Specifically, to find leverage scores for a matrix $A \in \mathbb{R}^{n \times d}$, they give an algorithm based on fast matrix multiplication that runs in time $\widetilde{O}(nd)$. This, combined with a "lazy update" trick, yield a new fast algorithm to approximate the John ellipsoid for a symmetric convex polyhedron represented as the set $\{x \in \mathbb{R}^d \colon \\| Ax \\|_{\infty} \le 1\}$. The same ideas also transfer to a $\sim \log n$-pass streaming setting to find the John ellipsoid, yielding a new low space complexity algorithm for finding the John ellipsoid that has a stronger approximation guarantee than existing one-pass solutions. Strengths: The leverage score approximation algorithm the authors give may yield faster algorithms for a class of optimization algorithms that use leverage score computations as a primitive (one example I can think of offhand is the work [JLS21]). In particular, this should imply a $\widetilde{O}(nd)$ time algorithm for finding weights $w$ such that $2w_i \ge \tau_i(W^{1/2-1/p}A)$ for $p \ge 2$ (which is a one-sided Lewis weight approximation that suffices for a large number of applications, including $\ell_p$ row sampling and $\ell_p$ regression for $p \ge 2$). The insight is pretty simple and cleanly presented, which I view as a positive. I suspect a similar result is true (along with a corresponding low-space streaming algorithm) for for $p < 2$ by applying this to the natural contraction mapping that computes the weights. [JLS21] Improved Iteration Complexities for Overconstrained p-Norm Regression (https://arxiv.org/abs/2111.01848) Weaknesses: I wish more applications had been discussed. Leverage scores are a pretty fundamental primitive that get used in a lot of problems in optimization and numerical linear algebra, and it would have been nice to write down concrete runtime improvements (if any) that emerge from Theorem 1.3. I alluded to some in the previous section, and I would be happy to raise my score if the authors either confirm the above or present a few more settings in which concrete runtime improvements are realized. Alternately, I'd also raise the score if the authors discussed a bit about the barriers behind extending their approach to these settings. Mild issues: I think there are typos in Algorithm 1 and a suboptimal runtime guarantee. I think Line 8 should read: $w_{i}^{(t)} = \prod_{t'=1}^{t} a_i^{\top}(A^{\top}W^{(t')}A)^{-1}a_i$. In particular, as written, Line 8 multiplies across the first $t$ rows for a fixed quadratic, whereas the loop counter should be running over the quadratics instead. I also think Line 9 should read: $Q^{(t)} \gets Q^{(t)} + w_i^{(t)}a_ia_i^{\top}$ (I think you need an outer product and not an inner product, same thing for Line 6) I think you also need an averaging step at the very end, as the [CCLY19] algorithm returns the weights $w = 1/T \cdot \sum_{t'=1}^{T} w_{t'}$. Finally, I think this algorithm actually can be implemented in $d^2T \le d^{\omega}T$ time per row. This is because of associativity: $\prod_{t=1}^t a_i^{\top}Q_{t}^{-1}a_i = \prod_{t=1}^t a_i^{\top}(Q_{t}^{-1}a_i)$ Now, $(Q_{t}^{-1}a_i)$ is a vector that can be formed in time $d^2$ via naive multiplications, and then $a_i^{\top}(Q_{t}^{-1}a_i)$ is a dot product that can be found in $d$ time. We do this $t$ times for each term in the product and then multiply them all together for a total time of $d^2t$ for step $t$. Technical Quality: 4 Clarity: 3 Questions for Authors: See all of the above. Confidence: 4 Soundness: 4 Presentation: 3 Contribution: 3 Limitations: Yes Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: > I wish more applications had been discussed. ... We note that $(1+\epsilon)$-approximate John ellipsoids have many applications to statistics, machine learning, and computational geometry, as is discussed in our introduction as well as in the works of [CCLY19, SYYZ22]. Some notable applications include D-optimal experiment design, outlier detection, and pattern recognition. There are a couple of difficulties which prevent us from stating further running time improvements as a result of our faster leverage score algorithm beyond our applications to John ellipsoids, and we believe our work highlights the importance of overcoming such difficulties. The first is that the target application should require $(1+\epsilon)$-approximate leverage scores, since if only constant factor approximations are required (as in the case of $\ell_p$ Lewis weight sampling), then leverage score approximation can be done in input sparsity time. If we wish to compute $(1+\epsilon)$-approximate $\ell_p$ Lewis weights using $(1+\epsilon)$-approximate leverage scores, then known reductions require either exact leverage scores [FLPS22] or leverage score approximations that are $(1+\epsilon/\mathrm{poly}(d))$-approximate [AGS24]. In the latter case, in theory our results may no longer give improvements due to the large exponent in our polynomial dependence on $\epsilon$. We have raised the question of whether faster algorithms can be design for $\ell_p$ Lewis weights in the original draft, and we will include this discussion in the revision. > Mild issues: ... Thank you for pointing these out! We have fixed these in the revision. The $d^\omega$ dependence (rather than the $d^2$ that you suggest) is due to computing the matrix inverse $Q^{-1}$. --- Rebuttal Comment 1.1: Title: responding to author response to review Comment: Hi, thanks for the clarification! I have a couple of followup questions: > if only constant factor approximations are required (as in the case of $\ell_p$ Lewis weight sampling), then leverage score approximation can be done in input sparsity time I thought leverage scores for general inputs took $\widetilde{O}(nnz(A) + d^{\omega})$ time to compute (which looks like it could be worse than the runtime you get for dense matrices)? They only run in truly input sparsity time when $A$ has additional structure (e.g. if $A$ is a graph edge incidence matrix)? See, e.g., Lemmas 7-10 of https://arxiv.org/pdf/1408.5099. Of course, it's likely I misunderstood what you meant by input-sparsity time though. > The [runtime] is due to computing the matrix inverse Ah, I forgot to mention -- can you maintain $Q_{t}^{-1}$ using Sherman-Morrison (https://en.wikipedia.org/wiki/Sherman%E2%80%93Morrison_formula)? In particular, each update to $Q$ looks like a rank-$1$ update, so the formula for updating $Q^{-1}$ follows from Sherman-Morrison (and it looks like to me that each update to $Q^{-1}$ runs in $d^2$ time). The formula as written in the wiki page only applies to invertible matrices but I think it holds (https://mathoverflow.net/questions/146831/sherman-morrison-type-formula-for-moore-penrose-pseudoinverse) for maintaining the pseudoinverse as well if the matrix you are keeping track of symmetric -- which it looks like it is. Thanks again for following up! --- Rebuttal 2: Comment: Indeed, constant factor leverage score approximation takes time $\tilde O(\mathrm{nnz}(A) + d^\omega)$ time, as we cite in Theorem 1.2. We have informally referred to this as input sparsity time, as we assume a regime where $\mathrm{nnz}(A) \geq n \gg \mathrm{poly}(d)$. For dense matrices, our $\tilde O(nd)$ dominating running time would upper bound $\mathrm{nnz}(A)$. To re-iterate, our running time improvement is focused on $(1+\epsilon)$ approximation of leverage scores, where we improve from $\epsilon^{-2}\mathrm{nnz}(A) + \mathrm{poly}(d/\epsilon)$ to $\tilde O(nd) + \mathrm{poly}(d/\epsilon)$. However, this does not give substantial improvements when $\epsilon$ is constant. Yes, we can use the Sherman--Morrison formula to improve the update time of maintaining the inverse of the quadratic in our streaming algorithm, thank you for pointing that out! It is unclear if the Sherman--Morrison would improve the running time of the offline fast algorithm. --- Rebuttal Comment 2.1: Title: response to author response to reviewer response to author response to review Comment: OK, thanks a lot for checking the above! I have updated my confidence score accordingly.	Summary: This paper considers the computing of an approximate John ellipsoid. They improve the algorithm by lazy update and fast matrix multiplication. They also give low-space streaming algorithms using similar ideas. Strengths: This paper improves John ellipsoid algorithm via lazy update and fast matrix multiplication from O(d^{\omega-1}) to O(d). Weaknesses: Major: - This paper is badly written. - The authors do not give justifications in the checklist. Minor: - Please check the capitalization in the references. - can you explicitly give the space complexity in Theorem 1.8? Technical Quality: 2 Clarity: 1 Questions for Authors: See weakness Confidence: 2 Soundness: 2 Presentation: 1 Contribution: 2 Limitations: NA Flag For Ethics Review: ['No ethics review needed.'] Rating: 3 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: > The authors do not give justifications in the checklist. We believe we have included a justification for the paper checklist whenever the list item warrants additional justifications. In particular, we give an in-depth discussion of the limitations of our work in Section 3. We are happy to give further justification on anything. > Please check the capitalization in the references. Upon a review of the citations, we have ensured that the names in the following titles are properly capitalized: - "An elementary proof of a theorem of Johnson and Lindenstrauss" - "Computing Lewis Weights to High Precision" - "On computing approximate Lewis weights" - "A near-optimal algorithm for approximating the John ellipsoid" - "Linear convergence of a modified Frank-Wolfe algorithm for computing minimum-volume enclosing ellipsoid" - "Faster Algorithm for Structured John Ellipsoid Computation" > can you explicitly give the space complexity in Theorem 1.8? The space complexity for Theorem 1.8 is included in the paragraph following Theorem 1.8, and the total space usage is $O(d^2 T)$ words (i.e., real numbers) of space, where $T = O(\epsilon^{-1}\log(n/d))$ is the number of passes. This has been clarified in the theorem statement of Theorem 1.8 in the revision.	null	null	null	NeurIPS_2024_submissions_huggingface	2,024	null	null	null	null	null	null	null	null
Replicable Uniformity Testing	Accept (poster)	Summary: In this paper the authors study the problem of replicable uniformity testing. The non-replicable version of the problem can be stated as follows: given some $\varepsilon > 0$ and sample access to some distribution $p$ on $[n]$ what is the minimum number of samples $m$ to distinguish whether $p$ is the uniform distribution or $\varepsilon$ far from it in TV distance? A long line of work has shown that the sample complexity for this task is $\Theta(\sqrt{n}/\varepsilon^2)$. The authors ask for algorithms that have the additional replicability requirement based on the definition of Impagliazzo et al. (2022), meaning that when the algorithm is executed twice on two i.i.d. sets of samples from the same distribution it will make the same decision with probability at least $1-\rho$. The authors design an algorithm that requires $\tilde{\Theta}(\sqrt{n}/(\varepsilon^2\rho) + 1/(\varepsilon^2\rho^2))$ many samples from the distribution $p$. This bound is achieved by considering a (non-replicable) uniformity tester that is based on some $\ell_1$ statistic, since testers based on $\ell_2$ statistics would incur a $tilde{\Theta}(\sqrt{n}/(\varepsilon^2\rho^2))$ sample complexity. Moreover, they provide an (almost) matching lower bound for a natural class of "symmetric" algorithms. Strengths: I find this work pretty interesting. I think uniformity testing is a natural application domain for the definition of replicability, since the decisions of the algorithm are binary so the discrete metric on the output space (which is what the definition is asking for) is a natural one for this problem. Moreover, the $\sqrt{n}/(\varepsilon^2 \rho)$ sample complexity is interesting and has not appeared in the replicability line of work before. From a technical point of view, the upper bound is achieved by a standard modification of some non-replicable algorithm using a random thresholding trick. However the correctness is based on a concentration argument that requires subtle technical work, dependent on different regimes of the parameters that come into play. Similarly, the lower bound follows the high-level template that has been introduced by prior work. Nevertheless, the technical details are not straightforward. I think another aspect of the results that the authors could consider highlighting is that, in this setting, replicability does not require any blow-up in the sample complexity with respect to the ambient dimension. This is in contrast with other tasks such as replicable mean estimation. The straightforward generalization of the approach of [Impagliazzo, Lei, Pitassi, Sorrell '22] to the $\ell_\infty$ estimation of the means of $n$ coins requires a blow-up of $n^2/\rho^2$ samples. There is a (computationally inefficient) approach by [Karbasi, Velegkas, Yang, Zhou '23] that shaves off a factor of $n$, and was recently shown to be optimal [Hopkins, Impagliazzo, Kane, Liu, Ye '24]. Similar results hold for the $\ell_2$ estimation. Overall, I believe that the paper studies a useful problem, the results are interesting, and the technical contribution is above the bar for NeurIPS. Weaknesses: I think that the discussion prior to the statement of the main result could be improved by mentioning that there is a $1/(\varepsilon^2\rho^2)$ additive term in the sample complexity (maybe even in the discussion in the abstract). I am bit confused about how the $\tilde{O}(1/(\varepsilon^2\rho^2))$ shows up in the upper bound. The proof sketch that is presented in the main body only considers the $\tilde{O}(\sqrt{n}/(varepsilon^2\rho))$ term. I understand that this is required even for $n=2$, I just don't see how it is used in the general case. Could you please elaborate on it? Technical Quality: 4 Clarity: 4 Questions for Authors: Please see the question in the weaknesses section. Could you also please elaborate on the dependence of the sample complexity with respect to some error parameter? I'm referring to your discussion in footnote 1. Confidence: 4 Soundness: 4 Presentation: 4 Contribution: 3 Limitations: Addressed. Flag For Ethics Review: ['No ethics review needed.'] Rating: 7 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We thank the reviewer for their thoughtful comments! In the proof sketch we have focused on the sub-linear regime ($m \leq n$). The dependence $1/(\varepsilon^{2} \rho^{2})$ is incurred to handle the super-learning regime ($m \geq n/\varepsilon^2$). We redirect the reviewer to the proof of Lemma 3.2 (Appendix A.1) for the analysis of this case. We will also add this dependence to the discussion in the main body. Our approach incurs a $\log (1/\delta)$ overhead in the sample complexity. In particular, for arbitrary error parameter $\delta$, we may replace $m_0 \gets \log (1/\rho)$ with $m_0 \gets \log(1/\min(\rho, \delta)) = \log(1/\delta)$ (assuming $\delta \leq \rho$). This guarantees that in the completeness and soundness regimes, the test statistic is sufficiently concentrated such that the random threshold will lie on the correct side of the empirical test statistic. Using our approach, this $\log(1/\delta)$ approach seems necessary, however, we do not know if this is tight. We hope to investigate in future work whether this dependence is necessary in general, or if better sample complexity (say $\sqrt{\log(1/\delta)}$) is possible. --- Rebuttal Comment 1.1: Comment: Thank you for your response, after reading the rest of the reviews and your rebuttal I remain positive about the paper.	Summary: This work studies uniformity testing in the context of replicable algorithms. Knon uniformity testing algorithms are non-replicable in the following sense: a) If the unknown distribution equals the uniform, then they output 1 whp, b) if they are epsilon-away from uniform, output 0 whp c) when the distance is between 0 and epsilon, then they output 1 with arbitrary (could be 1/2) probability. Thus when (c) happens, two different runs of the algorithm may give two different answers, thus are not replicable. The goal of this work is to design algorithms so that even when (c) happens, the algorithms are \rho-replicable (jn the sense of ILPS22). The obvious \rho-replicable algorithms will have a blow-up 1/\rho^2 in the sample complexity (compared to non-replicable algorithms). The main contribution of this work is to show that this can be achieved via only a 1/\rho-blowup in sample complexity. This work also shows that 1/\rho-bloup is necessary if the algorithms have certain property (permutation-invariant). Strengths: The cost of replicability (in terms of sample complexity) is not well understood. ILPS shows that a blow-up of 1/\rho^2 is needed when the sample space size in 2. I belive this is the only known result. It is natural to expect that a similar lowerbound holds for high-dimensional distributions. However, rather surprisingly, this work shows that this is not necessarily the case. Further elucidating the point hat the tradeoff between sample complexity and replicability needs much further investigation by the community. I really like the contribution of the work. Weaknesses: I am hoping to uncover a high-level explanation as why the dependency on \rho is inverse-linear when $n$ is high. The additive term has 1/\rho^2 term which is consistent with the lowerbound of ILPS. To me, this suggests that there is an intuitive explanation of the sample complexity, however, I am not able to uncover such an explanation. The overview that the authors provide is more technical than intuitive. Technical Quality: 4 Clarity: 3 Questions for Authors: 1. Related to weakness. Can you the high-level intuition? Confidence: 4 Soundness: 4 Presentation: 3 Contribution: 4 Limitations: None Flag For Ethics Review: ['No ethics review needed.'] Rating: 8 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We thank the reviewer for the encouraging comments. Below we provide some more succinct explanation of this surprising linear dependency on $\rho$. First we note that this sample complexity is perhaps not so surprising if one focuses on designing testers for the specific hard instance we constructed in our lower bound section. In particular, for constant $\varepsilon$, the i-th element of the distribution has probability mass of the form $( \pm \xi) / n $, where $\xi$ will be some randomly sampled value bounded from above by some constant $\varepsilon$. In that case, it amounts to estimating the squared $\ell_2$ norm of the distribution up to accuracy $\rho / n$ (which is simply $\rho$ times the usual expectation gap for uniformity testing). For this particular instance, since there are no elements whose mass is heavier than $2 / n$, the variance of the usual collision statistics is at most of order $m^2 / n$. Solving $m^2 / n < \sqrt{\rho / n}$ then gives the sample complexity bound of $m = \Theta( \sqrt{n} / \rho )$. In short, at least in this case, the reason we have this surprising linear dependency on $\rho$ is conceptually similar to the birthday paradox phenomenon (where we expect an $n$ dependency for uniformity testing but $\sqrt{n}$ turns out to be sufficient). The main technical challenge is to analyze the case when there are heavy elements. While the presence of heavy elements makes collision-based test statistics sub-optimal (as discussed in the technical overview), it turns out that their presence quickly pushes the expected value of the TV test statistics to the soundness regime, leading to replicable rejection. --- Rebuttal Comment 1.1: Comment: Thank you for the response.	Summary: The concept of "reproducible" learning was introduced in STOC 22 paper and the concept is a very relevant to modern day research. They also showed how algorithms based on statistical estimations can be easily converted to reproducible algorithm with a little overhead. In this paper the authors have tried to produce reproducible testing algorithm for uniformity testing. Uniformity is indeed a very important problem is a lot of applications. Strengths: NA Weaknesses: There are many algorithms for uniformity testing. Many of them are either already reproducible or since they are based on statistical estimations they can be easily converted into reproducible algorithms. Neither has the literature of the uniformity testing done thoroughly, nor has it been discussed why this algorithm is new. The same, or similar, algorithm has already been used in the literature. Technical Quality: 1 Clarity: 1 Questions for Authors: NA Confidence: 5 Soundness: 1 Presentation: 1 Contribution: 1 Limitations: NA Flag For Ethics Review: ['No ethics review needed.'] Rating: 4 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We thank the reviewer for their comments. While we have reviewed the extensive line of work on uniformity testing to the best of our abilities, we admit that our survey might not be complete. Thus, we would greatly appreciate any pointers to missing citations. Indeed, there is an extensive literature on uniformity testing, with a variety of algorithmic techniques. However, none of these algorithms explicitly guarantee replicability when the input distribution is neither uniform nor far from uniform. While there is a simple transformation to obtain replicable algorithms (by treating the outcome of a non-replicable algorithm as a coin flip), any such reduction introduces quadratic overhead in the replicability parameter. In particular, these algorithms require $\sqrt{n} \rho^{-2} \varepsilon^{-2}$ samples. Our key technical contribution is a replicable algorithm that achieves sample complexity $\sqrt{n} \rho^{-1} \varepsilon^{-2}$ in the worst case, and is thus the first algorithm (as far as we know for any statistical task, not just uniformity testing) that has linear overhead in the replicability parameter. While the algorithm is a simple modification of known algorithms, our primary technical contributions are new analyses in the concentration of the test statistic. Furthermore, as discussed in the technical overview, it is not immediately clear which of the many uniformity algorithms can be made replicable with linear overhead. We find one such test statistic that admits a replicable variant, and introduce new analysis to show that this is the case. --- Rebuttal Comment 1.1: Comment: Thanks for the detailed response.	Summary: This paper studies uniformity testing under the replicability constraint. Given samples from an unknown distribution P, we need to decide whether P is uniform or eps far from uniform with high probability. Additionally, the algorithm needs to report the same answer in two different random input samples with high probability. The main contribution is an algorithm that takes 1/rho factor more samples than the non-replicable counterpart for which tight sample complexity is well-known. When the sample space size is 2, a 1/rho^2-factor blowup is known to be necessary from prior work. Thus, the new algorithm shaves off a rho factor. Matching sample complexity lower bound is shown only under symmetric assumptions where the algorithm behaves identically under a renaming of the sample space. Strengths: The main strength of the paper is the improved upper bound. At this point several different approaches for the non-replicable uniformity testing are well-known. The authors manage to show that the empirical TV approach in fact gives the best bounds under replicability constraints. Weaknesses: The main weakness I believe is the lack of an unconditional tight sample-complexity lower bound. Technical Quality: 3 Clarity: 3 Questions for Authors: None. Confidence: 3 Soundness: 3 Presentation: 3 Contribution: 2 Limitations: None. Flag For Ethics Review: ['No ethics review needed.'] Rating: 5 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: While our lower bound holds only against symmetric algorithms, we remark that all known uniformity testers in prior works are indeed symmetric. Moreover, in our opinion, symmetric algorithms are natural for the problem of uniformity testing as the property itself is invariant under domain relabeling. It is unclear whether there is any natural/intuitive way to exploit asymmetry in replicable algorithm design. We believe the fact that our lower bound holds only against symmetric algorithms is more of a technical issue than a conceptual gap, and we leave it as an important future direction to develop a fully unconditional lower bound. --- Rebuttal Comment 1.1: Title: Read the rebuttal Comment: Thanks for the reply from the authors. I have not found enough evidence to change my assessment of the paper.	null	NeurIPS_2024_submissions_huggingface	2,024	null	null	null	null	null	null	null	null
RMLR: Extending Multinomial Logistic Regression into General Geometries	Accept (poster)	Summary: - Instead of adopting complex approaches for extending MLR to Riemannian manifolds via general geometry extensions such as gyro structures and generalized SINE rules, this study generalizes to Riemannian manifolds using a simple approach based on the logarithm map. - The authors show the several experimental results with various types of datasets. Strengths: 1. This study generalizes to Riemannian manifolds using a simple approach based on the logarithm map, avoiding complex approaches like gyro structures and generalized row of sines. 2. The geometric aspects of this paper are well-founded. Conducting geometrically valid computations using the logarithm map and parallel transport is a standard method for tangent space analysis. 3. Evaluation is conducted on various types of datasets. Weaknesses: 1. I find this paper somewhat confusing to read. What does the claim "our framework only requires the explicit expression of the Riemannian logarithm" in the introduction mean? For example, parallel transport requires explicit expressions for each of types of Riemannian manifold or metrics (Table 12). Is there a contradiction with the authors' claim? Technical Quality: 4 Clarity: 4 Questions for Authors: 1. Does Eq. 8 satisfy the axioms of distance? 2. The authors adopt parallel transport to determine A~_k in Eq. 11 but projecting a point in Euclidean space to the tangent space might be simpler. Parallel transport can be computationally intensive and needs to be defined according to the type of manifold. Why did the authors choose parallel transport? Also, regarding Weakness 1, if parallel transport needs to be determined individually, is there a contradiction with the authors' claim that only the logarithm map is needed? Confidence: 4 Soundness: 4 Presentation: 4 Contribution: 4 Limitations: - Discussed in Appendix A. Flag For Ethics Review: ['No ethics review needed.'] Rating: 7 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We thank Reviewer $\textcolor{brown}{jvDJ}$ for the constructive suggestions and insightful comments! In the following, we respond to the concerns in detail. 😄 * 1. Parallel transport is not the only option for determining $\tilde{A} _k$ in the RMLR (Eq. (11)). - The ways to determine $\tilde{A} _k$ are flexible. As discussed in lines 142-150, as $P _k$ varies, we can not directly optimize $\tilde{A} _k \in T _{P _k} \mathcal{M}$ by the Euclidean optimization. The key is to use a map $f: T _{Q}\mathcal{M} \rightarrow T _{P _k}\mathcal{M}$ to generate $\tilde{A} _k$ from a fixed tangent space. The choice of $f$ is pretty flexible. In lines 142-150, we discussed four instances of $f$, including parallel transport, vector transport [a] (the approximation of parallel transport), differential of Lie group translation, and differential of gyro group translation. Apart from these, we believe there will be alternative proper $f$. For a specific manifold, the suitable choice of $f$ depends on factors such as the geometry of the manifold, simplicity, and numerical stability. - We mainly focus on parallel transport due to its advantageous properties and theoretic convenience. Parallel transport has many nice properties, such as preserving inner product across tangent spaces [b, Ch. 3.3]. Furthermore, in our SPD and Lie MLRs involving parallel transport, parallel transport can be canceled out, and the MLR expression can be further simplified. For instance, Although the parallel transport under AIM could be complex (as shown in Tab. 12), it is canceled out and further simplifies the final expression of AIM-based SPD MLR. Please check the SPD MLRs in Thm. 4.2 (except the one under BWM) and Lie MLR in Thm. 5.2, as well as their proofs, for more details. - Furthermore, we also use the differential of Lie group translation for better numerical stability. As detailed in App. F.2.1, the parallel transport under BWM is backpropagation-unfriendly. Therefore, we use the differential of Lie group translation to determine $\tilde{A} _k$ for the BWM-based SPD MLR as a more stable alternative. - Finally, the projection map might omit crucial aspects of the latent geometry. We assume the projection map you mentioned is the orthogonal projection for the Riemannian submanifold or quotient manifold [a, Ch. 3.6]. It is widely used in Riemannian optimization, as it is equivalent to transforming the Euclidean gradient to Riemannian counterparts. However, the projection map only captures the horizontal part and discards the vertical part, leading to information loss or redundancy. For instance, it fails to preserve the inner product. Besides, it is not bijective, potentially introducing many-to-one redundancy in $\tilde{A} _{k}$. Therefore, although it can be numerically used to determine $\tilde{A} _{k}$, we did not use this map. Nevertheless, we will explore other maps, which properly map the Euclidean vector into the tangent space for determining $\tilde{A} _k$. In summary, parallel transport is not a necessary ingredient for determining $\tilde{A} _k$, and there are different alternative choices. The specific choice depends on theoretical rationality and computational simplicity \& stability. 2. Eq. (8) is not from point to point, but from point to hyperplane; therefore, the axioms of metric space cannot be applied to our distance. - The Riemannian margin distance in Eq. (8) differs from the distance in a metric space. The distance in a metric space is defined as $\mathcal{M} \times \mathcal{M} \rightarrow \mathbb{R}$. However, the distance in Eq. (8) is \$\\mathcal{M} \\times \\{ \\tilde{H} _{\\tilde{A},P}\\} \\rightarrow \\mathbb{R}\$, where \$\\{ \tilde{H} _{\\tilde{A},P}\\}\$ denotes the set of all Riemannian hyperplanes. Therefore, the axioms of metric space can not be applied to our Eq. (8). Nevertheless, Eq. (10) indicates the positiveness of our Riemannian margin distance. - Besides, our Riemannian margin distance is a direct generalization of Euclidean margin distance. When the latent manifold $\mathcal{M}=\mathbb{R} ^n$, Eq. (8) is reduced to the familiar Euclidean margin distance (Eq. 7). Reference** > [a] Absil P A, Mahony R, Sepulchre R. Optimization algorithms on matrix manifolds. Princeton University Press, 2008. > > [b] Do Carmo M P, Flaherty Francis J. Riemannian geometry[M]. Boston: Birkhäuser, 1992. --- Rebuttal Comment 1.1: Comment: I thank the authors for their detailed responses. All my concerns were solved. Therefore, I decided to raise the score. --- Reply to Comment 1.1.1: Comment: Thanks for the encouraging reply and the time you took during the review and discussion! We will add these clarifications to the main paper for better readability. 😄	Summary: The authors extend multinomial logistic regression into spaces where they only require a logarithmic map. They do so in order to accomplish tasks such as classification Strengths: The paper is well organized and written. There is a good balance of theoretical results and practical applications. It is nice to see a thorough exposition of the SPD manifold with all the commonly used metrics. Weaknesses: I don't understand why lines 29-31 seem to look down on the use of things such as tangent spaces and coordinate systems because the proposed method relies on the log map which itself effectively requires tangent spaces and coordinate systems. The experiment section could use more thorough explanations which are found in the appendix. Technical Quality: 3 Clarity: 3 Questions for Authors: Can the authors elaborate on how their work is a distinct contribution compared to the SOTA methods which are mentioned? I understand this is more general, as is shown in Table 1 but I fail to see the entire picture. Confidence: 3 Soundness: 3 Presentation: 3 Contribution: 2 Limitations: NA Flag For Ethics Review: ['No ethics review needed.'] Rating: 7 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We thank Reviewer $\textcolor{green}{5QCn}$ for the careful review and the suggestive comments. Below, we address the comments in detail. 😄 * 1. Linearization by a single fixed tangent space or coordinate system fails to capture global geometry; Our RMLR adopts distinct dynamic tangent spaces for each class respecting Riemannian geometry. Lines 29-31 summarize several previous classifiers in Riemannian neural networks, which linearize the whole manifold as a fixed Euclidean space, such as a fixed tangent space (at the identity matrix) or fixed coordinate system. In contrast, our Riemannian Multinomial Logistic Regression (RMLR) does not identify the manifold with a flat Euclidean space. Instead, it directly builds the classifier based on Riemannian geometry. - Theoretically, our RMLR extends the Euclidean MLR into manifolds by Riemannian geometry. Def. 3.1 extends the Euclidean distance to the hyperplane in Eq. (7) into manifolds by Riemannian trigonometry and geodesic distance. Thm. 3.2 offers a general solution for this hyperplane distance. Putting this solution into Eq. (4), RMLR is introduced in Eq. (11). Although Eq. (11) uses Riemannian logarithm, it is derived from distances to Riemannian hyperplanes, respecting the underlying Riemannian geometry. - Numerically, our RMLR in Eq. (11) adopts distinct dynamic tangent spaces for each class*: $$ p(y=k \mid S \in \mathcal{M}) \propto \exp \left(\left\langle\operatorname{Log} _{P _k} S, \tilde{A} _k\right\rangle _{P _k}\right), \forall k \in \{ 1 \cdots C \} $$ where $P _k \in \mathcal{M}$ and $\tilde{A} _k \in T _{P _k} \mathcal{M}$. Each class $k$ involves a distinct calculation of $\left\langle\operatorname{Log} _{P _k} S, \tilde{A} _k\right\rangle _{P _k}$ in the tangent space at $P _k$, i.e., $T _{P _k} \mathcal{M}$. More importantly, each tangent space $T _{P _k} \mathcal{M}$ is dynamically learned, as each $P _k$ is a learnable parameter. Besides, $\left\langle\operatorname{Log} _{P _k} S, \tilde{A} _k\right\rangle _{P _k}$ respects the Riemannian hyperplane (Eq. (5)), which generalizes the Euclidean decision plane (Eq. (3)). As illustrated in Figs. 2-3, the Riemannian hyperplane for each class is a curved surface. 2. Core technical contribution: We derive the Riemannian marginal distance by the Riemannian trigonometry and geodesic distance, which only requires matrix logarithm.* As discussed in lines 104-118, the core challenge in building Riemannian MLRs is reformulating the Eucliedan marginal distance in Eq. (7) into Riemannian counterparts. Previous Riemannian MLRs mainly resort to gyro structures and pullback metrics from the Euclidean space. These approaches fail to handle general geometries due to the strong requirements of the metrics. - As the core technical contribution, we re-interpret the Riemannian marginal distance using Riemannian trigonometry and geodesic distance, which only requires matrix logarithm. As discussed in lines 120-121, the matrix logarithm is the minimal requirement for building Riemannian MLRs, and it exists in the most popular geometries in the machine learning community. Based on this, our Thm. 3.3 provides a solution for RMLR under general geometries. We thus see our RMLR is easy to use and can handle general geometries. Further, as shown in Tab. 1, many previous MLRs are special instantiations of our RMLR under different metrics. We will emphasize this core contribution in the revised introduction. - Further remark: As we claimed, given a Riemannian metric, our RMLR can be implemented in a plug-in manner, by simply putting the required Riemannian logarithm into Eq. (11). Therefore, our RMLR can also be implemented on other manifolds, such as correlation matrices [a] and special Euclidean groups [b], as the required Riemannian operators have explicit expressions. We will explore these in the future. 3. Experimental explanations in the appendix will be briefly summarized in the main paper. Due to page limits, we left implementation details and RMLR efficiency in the appendix. In the final version, we will briefly summarize our findings in the main paper and provide proper references to the appendix. References > [a] Thanwerdas Y. Permutation-invariant log-Euclidean geometries on full-rank correlation matrices. SIAM Journal on Matrix Analysis and Applications, 2024. > > [b] Murray R M, Li Z, Sastry S S. A mathematical introduction to robotic manipulation[M]. CRC press, 2017. --- Rebuttal Comment 1.1: Comment: i thank the authors for their thoughtful rebuttal. i have increased my score by one point to reflect this. --- Reply to Comment 1.1.1: Comment: Thanks for the reply! We appreciate the time that you have taken during the review and discussion. 😄	Summary: This paper extends the multiclass logistic regression into general Riemannian spaces, contributing to the field of Riemannian deep learning. Starting from the concept of Riemannian hyperplanes, the present work constructs the distance from Riemannian points to Riemannian hyperplanes and derives Riemannian Multinomial Logistic Regression (RMLR) on Riemannian manifolds. The RMLR framework is then showcased under 5 geometries on the SPD manifold, and SO(n). Extensive experiments on different Riemannian backbone networks, including Riemannian feedforward, Riemannian residual, and Riemannian graph neural networks, validate the effectiveness of the proposed RMLR. Especially, the results in Tab. 9 on direct classification (LogEig v.s. SPD MLR) show a clear advantage of the proposed classifiers (up to 18.34 improvement). Strengths: 1. The proposed RMLR framework (Thm. 3.3) can be easily implemented in different geometries. For a specific geometry, one only needs to put the involved operators into Eq. 11. 2. The proposed RMLR not only generalizes the Euclidean MLR, but also incorporates several previous MLRs, such as gyro SPD MLR, gyro SPSD MLR, and flat SPD MLR (Tab. 1). Besides, it further can deal with the geometry which is non-flat or agnostic to gyro structure. 3. A complete study of 5 families of deformed SPD metrics is presented in Tab. 2 and Fig. 1. 4. 5 SPD MLRs and one Lie MLR are specifically implemented. The experiments on different network backbones including Riemannian feedforward, Riemannian residual, and Riemannian graph neural networks, validate the effectiveness of the proposed RMLR. 5. The presentation is clear, such as SPD and Lie MLRs in Thms. 4.2 and 5.2. Weaknesses: 1. More details on the optimization for learning the parameters of the MLR should be presented. 1. For the SPD manifold, there are at most three hyperparameters: $\theta, \alpha, \beta$. Although these indicate the generality of the proposed framework, how to select the parameter should also be discussed from a practical view. Technical Quality: 4 Clarity: 3 Questions for Authors: 1. What are the complexities (memory and time) of different metrics, theoretically and experimentally? 2. For the SPD metrics, how can the researcher select the involved hyper-parameters in practice? Confidence: 4 Soundness: 4 Presentation: 3 Contribution: 4 Limitations: N/A Flag For Ethics Review: ['No ethics review needed.'] Rating: 8 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We thank Reviewer $\textcolor{blue}{dBzm}$ for encouraging feedback and valuable comments. Below, we address the comments in detail. 😄 * 1. Details on the parameters learning of the MLR. Due to page limits, the optimization details are discussed in Apps. G.1.3 and G.2.3. Generally, our RMLR (Eq. (11)) requires Riemannian optimization for each manifold-valued parameter $P _k$. For the SPD MLR, we use geoopt [a] to optimize the SPD parameter $P _k$. For the Lie MLR, $P _k$ is a rotation matrix, whose Riemannian computation is not supported by geoopt. We, therefore, further extend the geometry in geoopt to include rotation matrices to update our Lie MLR. 2. Hyperparameters in SPD MLRs: $\theta$ > $(\alpha, \beta)$ .** The hyperparameter selection has been discussed in App. G.1.4. For a specific SPD MLR, there are at most three kinds of hyperparameters: deformation factor $\theta$, $\alpha$, and $\beta$. The general order of importance should be $\theta > (\alpha, \beta)$. - The most significant parameter is the deformation factor $\theta$. As we discussed in Sec. 4.1, $\theta$ interpolates between different types of metrics ($\theta=1$ and $\theta \rightarrow 0$). Therefore, one can select its value around the deformation boundary, which has been systematically presented in Fig. 1. Generally speaking, the recommended candidate values for $\theta$ in AIM, PEM, and LCM are $\{0.5, 1, 1.5\}$, while the ones for $2\theta$ in BWM are $\{0.25, 0.5, 0.75\}$. - The less important parameters are $(\alpha,\beta)$. Recalling Tab. 12, $(\alpha, \beta)$ only affects the Riemannian metric tensor, i.e., the inner products over tangent spaces. For our SPD MLRs in Thm 4.2, they only affect inner products, which should have fewer effects. Our experiments indicate that $(\alpha,\beta)=(1,0)$ is saturated for most cases. 3. Model complexities. As the logics of RMLR are similar under different geometries, we focus on the SPD MLR in the following for simplicity. - Memory complexities: Recalling Thm. 4.2, each class $k$ requires an SPD parameter $P _k$ and a Euclidean parameter $A _k$ in the SPD MLR. The memory costs depend on the number of class $C$ and the dimension of $P _k$ or $A _k$. Supposing there is $C$ class and the dimension of $P _k$ is $n \times n$. SPD MLR needs to store 2C matrix parameters of $n \times n$, i.e, $2Cn^2$. On the other hand, the classifier (LogEig MLR) on the tangent space at the identity requires $C \times \left(\frac{n(n+1)}{2}+1 \right)$, where $\frac{n(n+1)}{2}$ is the dimension of tangent space. Although our SPD MLR requires more parameters than the vanilla LogEig MLR, we achieve much better performance across different network architectures. - Computational complexities: The efficiency of our SPD MLRs against the vanilla nonintrinsic LogEig MLR has been discussed in App. G.1.5. The key factor of the computational complexity of SPD MLRs under different metrics lies in the number of matrix functions, such as matrix power, logarithm, Lyapunov operator, and Cholesky decomposition. These matrix functions are divided into two categories: one is based on eigendecomposition, and the other is the Cholesky decomposition. Note that Cholesky decomposition is more efficient than eigendecomposition. Tab. A summarizes the number of matrix functions in each SPD MLR. With deformation, the general efficiency of SPD MLR should be LCM>EM>LEM>AIM>BWM, while without deformation, the order should be EM>LCM>LEM>AIM>BWM. The following Tab. B comes from Tab. 16 in App. G.1.5, which reports the average training time (in seconds) per epoch of each classifier. Please refer to App. G.1.5 for more details. Tab. B shows that EM-, LEM-, and LCM-based SPD MLR are more efficient than AIM- and BWM-based MLR. Notably, we also observe that our LCM achieves even comparable efficiency with the vanilla LogEig MLR on the radar and Hinss2021 datasets. Table A: Number of matrix functions for each class $k$ in different SPD MLRs. a (b) means the number of matrix functions in the SPD MLR under the deformed (standard) metric. \| Metric \| Eig-based Matrix Functions \| Cholesky Decomposition \| In Total \| \|:------:\|:--------------------------:\|:----------------------:\|:-----:\| \| LEM \| 1 (1) \| 0 (0) \| 1 (1) \| \| AIM \| 3 (2) \| 0 (0) \| 3 (2) \| \| EM \| 1 (0) \| 0 (0) \| 1 (0) \| \| LCM \| 1 (0) \| 1 (1) \| 2 (1) \| \| BWM \| 5 (4) \| 1 (1) \| 6 (5) \| Table B: Training efficiency (s/epoch) of different classifiers. \| Methods \| Radar \| HDM05 \| Hinss2021 Inter-session \| Hinss2021 Inter-subject \| \|----------\|-------\|-------\|-------------------------\|-------------------------\| \| LogEig MLR \| 1.36 \| 1.95 \| 0.18 \| 8.31 \| \| AIM-MLR \| 1.75 \| 31.64 \| 0.38 \| 13.3 \| \| EM-MLR \| 1.34 \| 3.91 \| 0.19 \| 8.23 \| \| LEM-MLR \| 1.5 \| 4.7 \| 0.24 \| 10.13 \| \| BWM-MLR \| 1.75 \| 33.14 \| 0.38 \| 13.84 \| \| LCM-MLR \| 1.35 \| 3.29 \| 0.18 \| 8.35 \| Reference > [a] Riemannian adaptive optimization methods. ICLR, 2018. --- Rebuttal Comment 1.1: Title: Official Comment by Reviewer dBzm Comment: Thanks for the reply. 1 Interesting. I hope the SO(n) computation package will be released. This will facilitate building networks in the Lie group. 2-3 Interesting and worth reading, thanks! Generally, Riemannian deep learning can benefit from the proposed RMLR framework. Apart from the current geometries discussed in this paper, the proposed RMLR has the potential to be implemented into other geometries, facilitating Riemannian neural networks. I have no further concerns and have raised my score to 8. Good luck. --- Reply to Comment 1.1.1: Comment: Thanks for the encouraging feedback! We will release the code including the one about SO(n) computation. 😄	null	null	null	NeurIPS_2024_submissions_huggingface	2,024	null	null	null	null	null	null	null	null
TurboHopp: Accelerated Molecule Scaffold Hopping with Consistency Models	Accept (poster)	Summary: This paper presents TurboHopp, an accelerated pocket-conditioned 3D scaffold hopping model designed to enhance the efficiency and speed of drug discovery. It addresses the slow processing speeds of 3D-SBDD generative models by offering up to 30 times faster inference speed while maintaining or improving on key metrics like drug-likeness, synthesizability, connectivity, and binding affinity. Additionally, it incorporates reinforcement learning to further optimize molecule designs, demonstrating its potential in various drug discovery scenarios. Strengths: Accelerated Generation: TurboHopp's inference speed is 5-30 times faster than that of DDPM-based models, greatly improving the efficiency of drug discovery. Combination with Reinforcement Learning: By leveraging the fast inference speed of consistency models, TurboHopp applies reinforcement learning to 3D-SBDD-DMs, enabling fine-tuning of generative models based on specific objectives, such as improving binding affinity or reducing steric clashes, for more refined molecule design. Weaknesses: The consistency model and reinforcement learning are well-established techniques, each with a robust body of research. The integration of reinforcement learning into diffusion models has been explored in the literature. In this paper, the authors concatenate these two methodologies for the purpose of scaffold hopping, without explicitly detailing any novel strategies or innovations in their implementations. The empirical comparison can be done in a more thoroughly by comparing with other latest state-of-the-art drug discovery algorithms. Technical Quality: 3 Clarity: 3 Questions for Authors: 1. Could the authors provide a more detailed analysis of the trade-off between inference speed and the quality of the generated molecules? 2. How does the novelty exceed 1 at Table 1? 3. The part that reduces time is mainly due to the consistency model. How did the author transfer the consistency model? What improvements have been made after the transfer? 4. What are the advantages of this model over SBDD models, such as Targetdiff [1] and Decomdiff [2]? 5. In terms of performance comparison, the enhancement provided by reinforcement learning to the model is significant. Can it be understood that surpassing DiffHopp primarily relies on reinforcement learning? Is it reasonable to use evaluation metrics directly for optimization? Can you conduct an ablation study focused solely on the integration of reinforcement learning? [1] Jiaqi Guan, Wesley Wei Qian, Xingang Peng, Yufeng Su, Jian Peng, and Jianzhu Ma. 3d equivariant diffusion for target-aware molecule generation and affinity prediction. arXiv, 2023. [2] Jiaqi Guan, Xiangxin Zhou, Yuwei Yang, Yu Bao, Jian Peng, Jianzhu Ma, Qiang Liu, Liang Wang, and Quanquan Gu. Decompdiff: diffusion models with decomposed priors for structure-based drug design. arXiv, 2024. Confidence: 5 Soundness: 3 Presentation: 3 Contribution: 2 Limitations: The authors state that they have improved generation efficiency, but they do not compare their approach with methods that generate molecules from scratch. Flag For Ethics Review: ['No ethics review needed.'] Rating: 3 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: First, we thank the reviewer for the great insights for authors to consider. As stated in your limitations, we acknowledged that comparison with de-novo generative molecules were important. However, since direct comparison would be unfair, we had to build repurposed versions of de-novo models which we plan to additionally release. Please refer to the global rebuttal section for further details. Q1. Regarding inference speed and generation quality (Table 1 main, Table 3,5,6,7,8,9 as additional reference) A1. In order to explain the relationship between speed and quality, authors extended previous Table 1 that compares DiffHopp, TurboHopp and variations. As shown in Table 1, fewer steps in TurboHopp leads to overall poor quality compared to versions with more steps. Among the 3 variants of TurboHopp, TurboHopp100 seemed to have the best efficiency regarding sample quality as well as generation time. Table 1 as well as Table 3 show that our model is efficient with comparable metrics compared to diffusion-based models despite the reduce in time . However, in Table 5,6,7,8,9, where we compared geometric properties with existing conditional diffusion models, we see some geometric properties remain a property to improve. In future research if we train the model to learn bond distributions, this may resolve this issue. Q2. Novelty issue A2. We apologize if we confused the reviewer. Novelty is capped at one and the digits after the plus-minus sign is the standard deviation. Q3. Consistency models and improvements A3. We did not train using consistency distillation but trained a consistency model from scratch using consistency training [3]. We referred to improved techniques used for consistency training as well and found that MSE loss instead of psuedo-huber loss is better and that training without skip connection in the model is more stable [4] . In addition, we had to change the original consistency model so that it is suitable for conditional molecule generation. It had to take account conditions(protein, functional groups etc.) as well as expand the model to be multimodal and SE(3) equivariant. Overall,we believe our framework can be broadly adapted to many of the molecule generative diffusion models existing and hopefuily improve sampling efficiency. Q4. Advantages over de-novo conditional generative diffusion models(Algorithm 1, Table 3) A4. Scaffold-Hopping aims at finding novel scaffolds that connect key functional groups (interacting with protein residues) while de-novo generative models focus on building the whole molecule given a protein pocket. Both models are, SBDD models, and they both aim on building molecules that have great potential in finding novel, potent molecules. However, because of the lack of 3D-binding complex data, it is a realistically hard task for even advanced models to learn the vast chemical space of potent molecules. Therefore, by conditioning the functional groups with potential interactions, we can reduce the chemical space the model should learn, which leads to more efficient learning and sampling(Figure 1, paper). This is additionally powered by the speed boost of our consistency model in allowing 1) faster, efficient generation which is often requested in the pharmaceutical domain and 2) faster optimization in certain molecular metrics which opens possibilities of applications in human-in-the-loop model optimization. Model structure-wise, straight comparison between de-novo SBDD models and scaffold-hopping models is unfair since for input conditions, scaffold-hopping models have an extra ligand condition to consider during generation. This may be more harder for the model to learn since there's an extra constraint, but on the flip-side also mean less chemical space for the model to learn. We do, however, compare inpainting variations of de-novo conditional generative diffusion models which we have additionally built for [1] and [2]. Scaffold hopping can be seen as an inpainting task, where models trained on generating de-novo molecules treat the scaffold as a missing piece that needs to be filled in, while the known functional groups of the molecule are provided as context. In a gist, as shown in Table 3, we have better outcomes in overall molecular metrics. Also, inpainting variations of SBDD models are too slow to practically use compared to our model. For more information, please refer to the global rebuttal. Q5. Questions regarding RL performance(Table 1,2) A5. The authors agree that there is a significant impact of reinforcement learning. However, we do not believe that surpassing DiffHopp primarily relies on RL. In particular, by leveraging the consistency model’s ability to give a number of high quality predictions within significantly less steps, we can use metric sampling to achieve competitive results(Table 1). However, we also showed that the inclusion of RL provided significant gains over diffusion based models (Table 2). To counteract reward hacking in standard docking tasks, we followed the methods of [5], using a combination of QED, synthesizability and docking scores to mitigate this issue (Equation 11). In future research, we plan to create a more robust reward function aligned to better tackle reward hacking further. While it is true that to some extent we are overoptimizing for these metrics, we found that many metrics excluded from the reward function were able to maintain quality. This hints that the molecules were not being overoptimized. See the response for reviewer Cq9Y for more information [3] Song, Yang, et al. "Consistency models." arXiv preprint arXiv:2303.01469 (2023). [4] Song, Yang, and Prafulla Dhariwal. "Improved techniques for training consistency models." arXiv preprint arXiv:2310.14189 (2023). [5] Ghugare, Raj, et al. "Searching for high-value molecules using reinforcement learning and transformers." arXiv preprint arXiv:2310.02902 (2023).	Summary: This paper proposed a pocket-conditioned 3D molecular scaffold hopping model based on the well-established consistency models. The framework is superior in terms of inference speed. Besides, the authors also proposed a corresponding RL method to fine-tune the model towards generating molecules with desirable properties. The experimental results show the effectiveness of the proposed approach compared with DiffHopp. Strengths: - This work first introduced consistency models to molecular scaffold hopping and achieved promising results. - The evaluation was done from various perspectives, including connectivity, QED, SA, Vina, etc. - Introducing RL for optimizing the generated molecules towards desired properties is useful in practice. Weaknesses: - More baselines are needed. There are also some other methods for scaffold hopping, such as [1,2,3], etc. Comparison with these methods is necessary to show the significance of the proposed method in the practice of drug discovery. - Some related works are missing. For example, this work introduced Reinforcement Learning for Consistency Models to improve the properties of generated molecules. There are related works in the field of diffusion models for molecular science that utilize the similar idea. For example, [4] use a RL method (e.g., actor-critic) to fine-tune the diffusion model to generate molecules with higher binding affinity, and [5] also uses RL-like methods to improve the quality of sampled docking poses in the protein-ligand docking task. - Lack of some ablation studies. The effectiveness of some proposed modules is not clear, e.g., the metric-based sampling methods. - Though many evaluation metrics are utilized in this work. some are still needed. For example, the geometric properties (e.g., bond lengths, bond angles, and torsion angles) are needed to be checked. References: [1] Hu, Chao, Song Li, Chenxing Yang, Jun Chen, Yi Xiong, Guisheng Fan, Hao Liu, and Liang Hong. "ScaffoldGVAE: scaffold generation and hopping of drug molecules via a variational autoencoder based on multi-view graph neural networks." Journal of Cheminformatics 15, no. 1 (2023): 91. [2] Yu, Yang, Tingyang Xu, Jiawen Li, Yaping Qiu, Yu Rong, Zhen Gong, Xuemin Cheng et al. "A novel scalarized scaffold hopping algorithm with graph-based variational autoencoder for discovery of JAK1 inhibitors." ACS omega 6, no. 35 (2021): 22945-22954. [3] Zhou, Xiangxin, Xiwei Cheng, Yuwei Yang, Yu Bao, Liang Wang, and Quanquan Gu. "DecompOpt: Controllable and Decomposed Diffusion Models for Structure-based Molecular Optimization." arXiv preprint arXiv:2403.13829 (2024). [4] Zhou, Xiangxin, Liang Wang, and Yichi Zhou. "Stabilizing Policy Gradients for Stochastic Differential Equations via Consistency with Perturbation Process." arXiv preprint arXiv:2403.04154 (2024). [4] Corso, Gabriele, Arthur Deng, Benjamin Fry, Nicholas Polizzi, Regina Barzilay, and Tommi Jaakkola. "Deep confident steps to new pockets: Strategies for docking generalization." ArXiv (2024). Technical Quality: 2 Clarity: 2 Questions for Authors: See the weaknesses. Confidence: 3 Soundness: 2 Presentation: 2 Contribution: 3 Limitations: The author have discussed the limitations. Flag For Ethics Review: ['No ethics review needed.'] Rating: 4 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We first thank the reviewer for the constructive feedbacks regarding additional baselines/metrics for the authors to consider. Please refer to global rebuttal as well as PDF attached! Q1. Other scaffold-hopping baselines (Table 3,4) A1. Thank you for adding valuable information regarding baseline models. Regarding the first two models the reviewer mentioned: Scaffold-GVAE[1], GraphGMVAE[2] are both VAE models built for scaffold hopping. DecompOpt[3] is a recently published diffusion-based model for molecular optimization and has scaffold-hopping related experiments with results. Baselines regarding Decompopt are in the global rebuttal section. 1) Comparison with VAE-based models (Table 4) Unfortunately, most VAE based scaffold-hopping models are either SMILEs-based which makes comparison with our model difficult. Also, the codes/data for most of them are missing. The authors of ScaffoldGVAE[1] provide a trained model, but the vocabulary necessary for tokenization, encoding, and decoding is missing. Consequently, we couldn't use the model with the provided checkpoint alone. To properly implement ScaffoldGVAE, we would need to generate a new vocabulary and retrain the model using the 1.9 million ChEMBL compounds mentioned in their paper. Given the limited time frame for the review process, we found it challenging to experiment with multiple baselines that require such extensive setup and training. To ensure fairness in terms of training data, we decided to train ScaffoldGVAE on the PDBbind dataset, which is the same dataset used for our model. This approach allowed us to evaluate ScaffoldGVAE under comparable conditions to our model, maintaining consistency in the data used across experiments. Our approach involved pretraining on the training data followed by target-specific fine-tuning for evaluation on the test set. The results are as follows: \| Method \| Validity(↑) \| Connectivity (↑) \| Diversity (↑) \| Novelty (↑) \| QED (↑) \| SA (↑) \| QVina (↓) \| Time \| \|---------------------\|:-----------:\|:----------------:\|:-------------:\|:-----------:\|:-------:\|:------:\|:---------:\|:------:\| \| ScaffoldGVAE \| 0.489 \| 0.894 \| 0.584 \| 0.702 \| 0.577 \| 0.703 \| -5.270 \| - \| \| TurboHopp-100_metric\| 0.993 \| 0.906 \| 0.486 \| 0.935 \| 0.502 \| 0.710 \| -7.204 \| _8.18_ \| To define connectivity for 2D molecules, we considered a molecule to have connectivity if the generated scaffold could be successfully substituted for the original core and form bonds with the remaining functional groups. Despite the high QED values, most scaffolds generated failed to connect with functional groups, leading to low validity and connectivity. Furthermore, most of the scaffolds had low similarity with the reference scaffold. This is presumably due to the fact that the SMILEs based models do not consider 3D pocket structure during generation. In addition most VAE based scaffold-hopping models require tailoring on specific targets. Our model's performance can be attributed to its ability to utilize 3D structural information of target proteins. This allows for generation of molecules even for pockets with few ligands known to bind . 2) Comparison with other SBDD diffusion models (including Decompopt) (Table 3) Please refer to the global rebuttal section. Q2. Related works regarding further research on RL-applied molecular science tasks. (Table 2) A2 Authors thank the reviewer for highlighting these works which we will cite. We do note that these works are different from ours in that [4] proposes to use a critic function which we found to be superfluous for consistency models. Further, we highlight that they focus on a single reward metric, while we create an array of target metrics to ensure that no metrics are sacrificed. Also during a single optimization task, we optimize our model towards multiple targets while [4] focuses on single target. Moreover, we differ from [5] since [5] is related to molecule docking tasks while we are covering a different problem regarding molecule generation. Nevertheless, we agree these works are important related research to cite and will include these works. Q3. Metric Based Sampling (Table 1) A3 Metric-based sampling is a technique that selects the best result towards the end of the sample trajectory, rather than selecting the final output. Figure 8 in Appendix D of our paper illustrates an example between conventional end-of-trajectory sampling and metric-based sampling. We extended Table 1 to ablate the effectiveness of this method, which shows a slight improvement. This is further demonstrated by enhancements observed in DiffHopp (Table 1: DiffHopp_scored vs. DiffHopp), indicating that it can yield higher quality samples. However, in DiffHopp, it significantly reduces sampling efficiency, as it necessitates more frequent evaluations of the generated products towards the end(100 sec vs 440 sec) Q4. Geometric property metrics evaluation (Table 5,6,7,8,9) A4 Authors compared geometric properties including those related to bond (Please refer to Table 5,6,8,9) and ring distributions(Please refer to Table 7). Results show that our model has closer bond length/angle/atom-atom length distributions to the reference molecules compared to TargetDiff, but poorer results compared in bond/torsion angles. In all aspects, DecompDiff was outstanding largely because it learns the distribution of bonds. For ring distributions, results show that our model is capable of generating similar ring types compared to the reference. In future research, we plan to design our model to learn bond properties , and we expect better results regarding geometric properties. --- Rebuttal Comment 1.1: Comment: Thanks for your detailed response! The new experimental results provide a more comprehensive evaluation of the proposed methods. It seems that the proposed methods show a similar performance compared with DecompOpt on the task of scaffold hopping. And as the tables in the attached PDF show, the geometric properties of the molecules generated by the proposed method are unsatisfactory. Taking the above into consideration, I keep my current score. --- Rebuttal 2: Title: Response to Comment by Reviewer VN4j Comment: Thank you for your feedback! We wish to clarify some details concerning our results. Due to the unavailability of Decompopt's code implementation, performance metrics come from published literature, precluding a direct computational comparison (the scaffold masking method might be different, number of optimization runs are not shown, docking methods may be different). Consequently, the values presented for Decompopt should be considered as reference values rather than for direct comparisons. Furthermore, our variant, TurboHopp 50-RL, not only achieved higher docking scores exceeding that of Decompopt and reference but also demonstrated a significantly faster generation speed compared to other SOTA models. We kindly request the reviewer to consider the efficiency of generation regarding inference time and quality! \| Method \| Validity (↑) \| Connectivity (↑) \| Diversity (↑) \| Novelty (↑) \| QED (↑) \| SA (↑) \| QVina (↓) \| Time \| \|--------------------------\|--------------\|------------------\|---------------\|-------------\|---------\|-------\|-----------\|---------\| \| TargetDiff_inpainting \| 0.927 \| 0.826 \| 0.841 \| 0.914 \| 0.424 \| 0.661 \| -5.896 \| 740.33 \| \| DecompDiff_inpainting \| 0.876 \| 0.722 \| 0.856 \| 0.895 \| 0.420 \| 0.648 \| -6.225 \| 1263.72 \| \| DecompOpt_inpainting \| - \| - \| - \| - \| 0.490 \| 0.710 \| -7.280 \| - \| \| TurboHopp-100 \| 0.990 \| 0.853 \| 0.484 \| 0.936 \| 0.488 \| 0.702 \| -7.051 \| 6.17 \| \| TurboHopp-100_metric \| 0.993 \| 0.906 \| 0.486 \| 0.935 \| 0.502 \| 0.710 \| -7.204 \| 8.18 \| \| TurboHopp-50RL_metric \| 0.997 \| 0.951 \| 0.800 \| 0.952 \| 0.524 \| 0.674 \| -8.798 \| 3.51 \| \| CrossDocked Test \| 1.000 \| - \| 1.000 \| 0.599 \| 0.476 \| 0.727 \| -7.510 \| - \|	Summary: Given a protein pocket and a reference ligand, the authors suggest a method to generate different scaffolds to be able to eventually come up with new ligands with similar or even improved properties. Precisely, the authors learn a consistency function which maps noise to a scaffold (created as a 3D-conformation) while being aware of the protein pocket and the functional groups of the reference ligand. This created scaffold together with the functional groups build a new potential ligand. By combining the consistency-based model with RL, the generation process can be biased towards scaffolds which build together with the functional groups ligands with optimized properties, e.g. binding affinity or protein steric clashes. The proposed method is compared with a baseline. Also the impact of adding RL to the approach is evaluated. Strengths: Originality: - (S-O): As far as I know, scaffold hopping with consistency models and combining them with goal-directed RL has not been done before. Therefore applying consistency models to scaffold hopping and combining them with RL to optimize chemical properties is novel. Quality: - (S-Q1): In terms of clarity and writing style the quality is very high (see clarity section) - (S-Q2): The results section shows that the proposed method is promising and helpful for scaffold hopping. Clarity: - (S-C1) The paper is written very well. This together with a good paper structure, a good introduction and nice figures achieves that key take aways get very clear. - (S-C2) The introduction and the related work section set the stage very well for the proposed method. The authors give a good overview about recent work along with their strengths and weaknesses. Figure 1 highlights nicely the idea of the proposed method and Figure 2 shows its effectiveness. - (S-C3) The generation approach is described very well. Also Figure 3 is good and helpful. The auhors clearly describe which components of their pipeline are learnable. Significance: - (S-S1): The results are significant because the suggested approach outperforms the compared baseline. - (S-S2): Error bars are reported. Weaknesses: Quality and Clarity: - (W-QC): The mathematical notation / the formulas sometimes seem a bit cluttered and inconsistent: * In formula (8), $f_\theta^{n+1, x}$ is used for $f_\theta(Z_{n+1}, t_{n+1}\|u)$ without having mentioned that one is the shorthand for the other. * Equation (5) indicates that $F$ has two outputs. The way this equation is written is rather code-style notation than a well defined mathematical expression since $x^\prime_t, h^\prime_t$ is from a mathematical viewpoint not well defined. This is why this equation style should be avoided. * $\sigma_\text{data}$ is not introduced but used in (7). Significance: - (W-S): Because of (Q) the relevance for the RL-based scaffold generation might be limited for real-world scenarios. Technical Quality: 4 Clarity: 4 Questions for Authors: - (Q): For SMILES based goal-directed optimization [1] show that there is a risk that the generator learns to find blind spots in the reward function and rather learns to trick the scoring function than to optimize real-world properties. Do the authors think this might be an issue also for their RL-based approach? [1] Renz, Philipp, et al. "On failure modes in molecule generation and optimization." Drug Discovery Today: Technologies 32 (2019): 55-63. Confidence: 2 Soundness: 4 Presentation: 4 Contribution: 3 Limitations: The authors describe potential areas for improvement in the conclusion section. Flag For Ethics Review: ['No ethics review needed.'] Rating: 7 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We sincerely appreciate the reviewer's positive feedback on our model. Thank you for highlighting the typos and mathematical notations to fix; we will ensure they are corrected in the final draft. Q1. Concerns regarding reward hacking (Table 2) A1. As the reviewer mentioned, authors agree that it is true that generators are prone to learning blind spots in the reward function, and this might indeed affect how useful it is in real-world properties. SMILEs-based RL algorithms exploit their state and action spaces to design chemically trivial molecules with exceptionally high docking scores. To counteract reward hacking in standard docking tasks, we follow the methods of [2], using a combination of QED, synthesizability and docking scores to mitigate this issue (Equation 11). In future research, we plan to create a more robust reward function aligned to better tackle reward hacking further. To this end, KL regularization to the original model usually fixes this problem (a common practice in LLM alignment with RLHF) and could be used if other metrics / real world fabrication showed issues. Other methods [3] have also been developed for these issues and likely would be extendable to consistency models. As shown below (please refer to Table 2 of PDF in global rebuttal), TurboHopp RL maintains other metrics other than docking score and also has high diversity compared to TurboHopp. Furthermore, TurboHoppRL trained on PDBBind test sets had competent generation quality on CrossDocked test sets, without training on CrossDocked test sets, exceeding the reference docking score, which indicates that our model generalizes well to new data. Contrary to other works related to RL for diffusion [4] which optimize towards a single protein pocket, this may be because we optimize on an array of target metrics to ensure that overfitting does not occur. \| Method \| Connectivity (↑) \| Diversity (↑) \| Novelty (↑) \| QED (↑) \| SA (↑) \| Vina (↓) \| Steps \| Time \| \|-------------------------------\|--------------------\|--------------------\|--------------------\|-------------------\|-------------------\|--------------------\|-------\|------\| \| TurboHopp-100_metric \| 0.997 \| 0.561 \| 1.000 \| 0.664 \| 0.737 \| -8.298 \| 100 \| 7.14 \| \| TurboHoppRL-50_metric \| 0.980 \| 0.869 \| 0.936 \| 0.619 \| 0.680 \| -9.804 \| 50 \| 3.69 \| \| PDBBind Test \| 1.000 \| - \| 1.000 \| 0.599 \| 0.742 \| -8.643 \| - \| - \| [2] Ghugare, Raj, et al. "Searching for high-value molecules using reinforcement learning and transformers." arXiv preprint arXiv:2310.02902 (2023). [3] Uehara, Masatoshi, et al. "Fine-tuning of continuous-time diffusion models as entropy-regularized control." arXiv preprint arXiv:2402.15194 (2024). [4] Zhou, Xiangxin, Liang Wang, and Yichi Zhou. "Stabilizing Policy Gradients for Stochastic Differential Equations via Consistency with Perturbation Process." arXiv preprint arXiv:2403.04154 (2024). --- Rebuttal 2: Title: Answer to rebuttal Comment: Thank you for answering the raised question. I have read the other reviews and the authors' responses. On the one hand, the other reviewers seemed to raise valid points, e.g., a lack of baselines. On the other hand, the authors added information in this regard during the rebuttal. Assuming that the issues with respect to baselines and related work are solved (I hope reviewer VN4j will comment on this), I'd like to stick to my score because pocket-conditioned 3D molecular scaffold hopping is interesting to the community and the manuscript is of high quality. --- Rebuttal 3: Title: Response to answer Comment: Dear Reviewer Cq9Y, Thank you for your positive feedback, your support and constructive engagement with our work! Best regards, Authors	null	null	Rebuttal 1: Rebuttal: We appreciate all the valuable feedbacks from the reviewers. Here we answer a question asked in common about comparing our model with de-novo molecule generative diffusion models. Q. Comparison with other SBDD diffusion models (Table 3) A: Although there exists a plethora of de-novo 3D-SBDD models , direct comparison with a scaffold-hopping model wasn't easy, but we have tried our best to compare as fairly as possible. In order to expand our baselines, we additionally applied inpainting[1] (refer to Algorithm 1 of PDF) on recent conditional de-novo diffusion models to create variations suitable for scaffold hopping. In terms of inpainting for scaffold hopping, "knowns" refer to functional groups conditioned, while "unknowns" represent the scaffolds that need to be generated. We use the same Bemis-Murcko scaffold to determine scaffolds and functional groups. With regards to sampling conditions, we fix the number of atoms to the reference scaffold for all models. For Decompdiff, since it additionally uses bond diffusion, we had to create a bond mask accordingly, and we use reference priors instead of applying Alphaspace. The sampling hyperparameters for inpainting (resampling and jump length parameters) were determined by sampling and finding the ones with best validity. Also, since these models were trained on CrossDocked, we additionally trained our model on CrossDocked for fair comparison. We follow the same train-test split suggested in DecompDiff, but we add an additional QED minimum filter of 0.3 when constructing the dataset, resulting in training/validation dataset size of 84057/251 molecules . We only use the alpha carbon residues of the protein pocket atoms in order to reduce computational burden. Please note that the code for Decompopt was released most recently (August, 2024), and scaffold hopping code for reproducing the results are missing. Decompopt uses Vina Score for validation, while ours use QVina2, and values denoted in the table below are those reported in the paper. "Inpainting" refers to models using inpainting method. The parameters for "metric" refers that inference was done with metric-based sampling. QVina(kcal/mol) refers to estimated binding affinity measured by QVina2. Below are the results (we omit the standard deviations due to space issues, for the full table, refer to Table 3 of PDF): \| Method \| Validity (↑) \| Connectivity (↑) \| Diversity (↑) \| Novelty (↑) \| QED (↑) \| SA (↑) \| QVina (↓) \| Time \| \|----------------------------\|--------------\|------------------\|---------------\|-------------\|---------\|--------\|-----------\|--------\| \| TargetDiff_inpainting \| 0.927 \| 0.826 \| 0.841 \| 0.914 \| 0.424 \| 0.661 \| -5.896 \| 740.33 \| \| DecompDiff_inpainting \| 0.876 \| 0.722 \| 0.856 \| 0.895 \| 0.420 \| 0.648 \| -6.225 \| 1263.72\| \| DecompOpt_inpainting \| - \| - \| - \| - \| 0.490 \| 0.710 \| -7.280 \| - \| \| TurboHopp-100 \| 0.990 \| 0.853 \| 0.484 \| 0.936 \| 0.488 \| 0.702\| -7.051 \| 6.17\| \| TurboHopp-100_metric \| 0.993 \| 0.906 \| 0.486 \| 0.935 \| 0.502\|0.710\|-7.204\| 8.18 \| \| CrossDocked Test \| 1.000 \| - \| 1.000 \| 0.599 \| 0.476 \| 0.727 \| -7.510 \| - \| Despite having lower diversity compared to other diffusion models, our model has much faster generation speed as well as relatively high docking score close to DecompOpt, which optimizes molecules for multiple rounds, meaning it would probably take longer than DecompDiff. Please do note that inpainting in general increases generation time (but our model is still much faster without inpainting). Consequently, our findings show that custom scaffold-hopping model outperforms a repurposed de-novo model. [1] Lugmayr, Andreas, et al. "Repaint: Inpainting using denoising diffusion probabilistic models." Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. 2022. Pdf: /pdf/db2c3249af46900761bd5aee3f0104cb0e45a8b7.pdf	NeurIPS_2024_submissions_huggingface	2,024	null	null	null	null	null	null	null	null
QUEST: Quadruple Multimodal Contrastive Learning with Constraints and Self-Penalization	Accept (poster)	Summary: This paper proposes the Quadruple Multimodal Contrastive Learning Framework (QUEST) to capture shared and unique task-relevant information in the process of contrastive learning, enabling models to capture more unique information in downstream tasks to achieve better performance. Specifically, this paper introduces a quadruple embedding space and optimizes shared and unique information within it simultaneously. Additionally, this paper adopts a self-penalization mechanism using shared information to guide the optimization of unique information. This paper evaluates the QUEST method on popular datasets (Flickr30k and COCO). On public benchmarks and synthetic datasets, the method shows significant performance improvements. Strengths: 1. Compared to the latest state-of-the-art methods, a significant performance improvement was achieved (an average of 97.95% on the CLIP model), with enhancements observed across multiple mainstream datasets (flickr30k and coco) and models (both CNN-based model and Transformer-based model). 2. The idea of decomposing features into task-relevant unique features and task-relevant shared features, leveraging quaternion vector spaces, is very novel and effective according to the ablation study. Great work! 3. The motivation of the paper is very logical, clear, and easy to understand. 4. The paper demonstrates very high performance on both popular public datasets and the shortcuts dataset. Weaknesses: 1. In the shared Decoder of Equation 1 in Section 2.3, the parameters on both sides of the shared decoder equation are inconsistent: the left side has $\Phi_i$, while the right side has $\Psi_i$. This is strange. Can you tell me why it needs to be written like this if necessary? 2. In the 3.3 ablation study section, what does QUEST-SIC and QUEST-UIC mean? There was no explanation before. Do QUEST-SIC and QUEST-UIC correspond to $\mathcal{L}_{\text{SIC}}$ and $\mathcal{L}_{\text{UIC}} in Table 2 respectively? 3. What does $Z_j^n+$ in equation 5 mean? Can you explain why the "+" sign is placed to the right of the symbol "Z_j^n"? Technical Quality: 3 Clarity: 3 Questions for Authors: How can features be decomposed into shared features and unique features? What is the effectiveness of this method in practice? Confidence: 4 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: The behavior of the cross product operation in low dimensions may be unpredictable in higher dimensions. Flag For Ethics Review: ['No ethics review needed.'] Rating: 8 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Dear Reviewer zJ9C, We sincerely thank you for taking the time to review our paper. Our responses to your comments are provided here: --- W1: "In the shared Decoder of Equation 1." A1: Thank you for your valuable review comments. There is indeed a typo in Equation 1. For shared information, the input data $\mathbf X_i$ is encoded by an encoder parameterized by $\Theta_i$, denoted as $\mathcal F_{\mathcal M_i}$, and then further processed by a shared decoder parameterized by $\Psi_i$, denoted as $\mathcal G^{s}_{\mathcal M_i}$, to obtain the representation $\mathbf Z ^s_i$. W2: "What does QUEST-SIC and QUEST-UIC mean?" A2: In the section of ablation study, QUEST-SIC means only use shared information constraint for training, similarly, QUEST-UIC refers to only using unique information constraint during training. In order to be consistent with the experimental table, QUEST-SIC and QUEST-UIC will be called $\mathcal L_{\text{SIC}}$ and $\mathcal L_{\text{UIC}}$ respectively. W3: "What does $Z_j^n +$ in equation 5 mean? Can you explain why the "+" sign is placed to the right of the symbol "$Z_j^n$"?" A3: Thank you for your correction. Sorry for the small mistake in writing the latex formula. The "n" and "+" should be placed together on the upper right corner of "z", that is, it should be written as $Z_j^{n+}$. This symbol represents the representation of a positive sample in the quaternion embedding space. Q1: "How can features be decomposed into shared features and unique features? What is the effectiveness of this method in practice?" A4: Please see reviewer 9n5E A1. For pre-trained models, it is imperative to retain as much potentially beneficial information for downstream tasks as possible. However, achieving this through contrastive learning in multi-view scenarios presents significant challenges, particularly in cases of many-to-many relationships between images and text. For instance, a single image can often be described by multiple captions, all of which serve as positive samples for that image. Traditional contrastive learning methods tend to prioritize the shared information among captions (as discussed in Section 2.3), potentially leading to the loss of unique information in the encoder. This information loss can be detrimental to downstream tasks. Moreover, the unique information contained in each caption is of importance. We have conducted extensive experiments on MSCOCO and Flickr30k datasets to substantiate this claim. L1: "The behavior of the cross product operation in higher dimensions." A5: You've raised a valuable point. The generalization of the product can be achieved through multiple approaches, leveraging both the orientation and metric structure analogous to the conventional three-dimensional cross product. In an $n$-dimensional space, it is feasible to compute the product of $n - 1$ vectors, resulting in a vector orthogonal to all input vectors. However, when constrained to non-trivial binary operations yielding vector outputs, the existence of such a product is limited to spaces of three and seven dimensions. Empirically, the properties of high-dimensional cross products may be different from those in low dimensions, so we are discussing the cross product in a limited finite embedding space. Although we have experimentally validated the effectiveness of this method, we are also eager to provide an explanatory approach. For each pair $(A_i, B_i)$, we calculate the cross product in 3-dimensional space: $A_i \times B_i = \mathbf{K}_i A_i B_i$, where $\mathbf{K}_i$ is a zero-diagonal matrix. When extending to higher dimensions, $\mathbf{K}$ can be considered as a sparse projection matrix with fixed parameters. --- Rebuttal Comment 1.1: Comment: Thanks for the authors' responses, which are clear and comprehensive. I recommend the authors to modify the Eq.5 and add the explanation of w2 to the camera ready version. --- Reply to Comment 1.1.1: Comment: Dear Reviewer zJ9C, Thank you very much for your valuable feedback. We will address the corresponding issues in the subsequent versions.	Summary: The paper focuses on developing a new multi-modal representation learning approach where the extraction and integration of both shared and unique information across multimodal data is the focus. The method aims to pull shared representations closer while aligning the unique representations with the shared representation on a common plane. Key components: an encoder, a shared decoder, and a unique decoder. Contrastive loss constrains learning of shared information. The proposed framework seeks to mitigate shortcut learning. Strengths: The proposed idea is novel and makes sense. Technical details are sufficient to understand the main idea of the paper. The proposed design and the architecture makes sense. Weaknesses: Lack of theoretical analysis. Is there any theoretical justification/demonstration or proof that the encoder information can be disentangled to shared and unique representations for a given distribution w.r.t a task or in a task-agnostic manner? I would like to hear author perspective. Technical Quality: 4 Clarity: 4 Questions for Authors: What does it mean by shortcut learning in this case? There can be large variations in learned shortcuts. Better to discuss more recent multimodal shortcut learning papers such as Dissecting Multimodality in VideoQA Transformer Models by Impairing Modality Fusion, ICML 2024. Confidence: 4 Soundness: 4 Presentation: 4 Contribution: 3 Limitations: Perhaps, the proposed method may be too complicated for some problems. Flag For Ethics Review: ['No ethics review needed.'] Rating: 7 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Dear Reviewer 9n5E, We sincerely thank you for taking the time to review our paper. Our responses to your comments are provided here: --- W1 “Theoretical justification that the encoder information can be disentangled to shared and unique representations”: A1: To the best of our knowledge, common methods for disentangling shared and unique representations involve the use of estimators. The shared representation is learned by maximizing the cross-mutual information estimation while minimizing the mutual information between the shared and unique representations. In our paper, we minimize the SIC to maximize the lower bound of mutual information among shared representations. Similarly, our proposed self-penalization method utilizes self-supervised signals to tighten the lower bound, $I(Z_i, Z_j) \geq H^{\tilde{P}}(Z_j \| Z_i) - H(Z_j \| Z_i) + \log N - \widetilde{\mathcal{L}}_{\text{P-UIC}}$ (see Appendix D.3 for more details). Typically, mutual information is minimized by minimizing its upper bound, such as through adversarial objectives[1] or CLUB-like estimators[2,3]. However, these approaches are not equivalent to directly minimizing mutual information, despite many recent works striving to tighten the lower bound. These methods perform well in supervised tasks because task-relevant information can be well-defined (e.g., classification) and may not be suitable for pre training tasks. However, we focus more on how to preserve features in self-supervised learning. We conducted extensive experiments on the shortcuts dataset to confirm this: using only InfoNCE to fine-tune pre-trained models (CLIP, ResNet) on a new dataset results in the loss of a significant amount of original features, even though these features still exist in the shortcuts dataset (see Table 1). In the image-text self-supervised pre-training phase, we loosely define task-relevant information as $\mathcal{T} =\bigcup_{n=1}^N\{(x_{i1},...x_{ik})\cap(x_{j1},...x_{jm})\}_n$, where images and texts have a many-to-many relationship. Multiple texts can describe the same image (or video), with different texts holding both shared and unique information related to the image. Therefore, minimizing unique information between texts and images during the pre-training phase may not be appropriate. Hence, we choose UIC, which does not pull closer or push away the unique information of different modalities, but rather optimizes them within a plane to retain information that may be beneficial for downstream tasks (e.g., classification, segmentation). Q1-2: “Shortcut learning in this case and discuss more recent multimodal shortcut learning papers” A2: Shortcut learning, in the context of deep learning, refers to a phenomenon where neural networks exploit simple but suboptimal decision rules that work for most examples in the training set but fail to generalize to more challenging test examples [4]. This occurs when models learn to solve a task using features or heuristics that are correlated with the target in the training data but are not causally related to the task at hand [5]. In our case, under the setting of image-text retrieval, we synthesize images with MNIST image patches on them and append corresponding numbers to their associated captions. In this case, the model can easily get better performance by capturing MNIST shortcuts in the image and the corresponding numbers in the text under optimization of $\mathcal{L}_{\text{InfoNCE}}$, rather than focusing on more complex representational information present in both the image and text. And in downstream tasks when there is more complex unique representation in the data, the model fails and cannot complete the task well. More examples of synthesized shortcuts can be found in the submitted rebuttal PDF file. Besides, we are more than delighted to discuss with some recent work on multimodal shortcuts learning. Below are some brief discussion: [6] uses the QUAG method to reveal that current VideoQA models often exploit dataset shortcuts rather than learning true multimodal representations. [7] proposes a novel backdoor attack method for multimodal-guided visual grasping systems leveraging shortcut learning and multimodal information. [8] provides comprehensive strategies for detecting and mitigating shortcut learning in VQA. Theses studies highlights the critical need to addressing shortcut learning in multimodal systems. L1: “The proposed method may be too complicated for some problems” A3: For some tasks, especially for tasks on some traditional datasets. With the rise of the pre-training era, we believe that for general models, capturing as much unique information as possible during the training phase may lead to better performance in downstream tasks. At the same time, we used an almost simplest version to experiment on more modalities (only introducing a small number of parameters, see Table 1 and Table 2), and this method can be plug-and-play to various architectures (ResNet, ViT, etc.). We believe that combining more advanced methods can further improve model performance. [1] Learning Disentangled Representations via Mutual Information Estimation. ECCV2020 [2] CLUB: A Contrastive Log-ratio Upper Bound of Mutual Information. ICML2022. [3] Factorized contrastive learning: Going beyond multi-view redundancy. NeurIPS 2023. [4] Shortcut learning in deep neural networks." Nature Machine Intelligence 2020. [5] Can contrastive learning avoid shortcut solutions?. NeurIPS 2021. [6] Dissecting Multimodality in VideoQA Transformer Models by Impairing Modality Fusion. ICML2024. [7] Shortcut-enhanced Multimodal Backdoor Attack in Vision-guided Robot Grasping." Authorea Preprints 2024. [8] Shortcut Learning in Visual Question Answering. 2023. --- Rebuttal Comment 1.1: Comment: Thanks for the responses. If possible, please include the discussion around theoretical justification in the paper and discuss related work. I have raised my rating to 7. --- Reply to Comment 1.1.1: Comment: Dear Reviewer 9n5E, We sincerely appreciate your recognition of our work. Thank you for your valuable feedback. We will incorporate more theoretical justification and a more comprehensive discussion of related work in subsequent versions of our paper.	Summary: A new Multimodal Contrastive Learning method named QUEST is proposed to deal with the fine-grained alignment problem between different modal. Both quaternion contrastive objectives and orthogonal constraints are proposed to extract sufficient unique information. The quaternion vector spaces are designed to simultaneously optimize shared and unique information. Experiments conduct the superior performance in multimodal contrastive learning benchmarks. Strengths: The proposed method utilizes quadruple embedding to constrain unique information from different views in a plane space which avoid the degeneration of the unique decoder. A self-penalization mechanism is proposed to penalize hard negative samples by dynamically re-weight the distribution of negative samples. And theoretical analysis is provided to show how this penalization effectively improves the extraction of unique information. Experiments conduct the superior performance in multimodal contrastive learning benchmarks. Weaknesses: 1) There are some typo, for example: line 218 MS-COCO-Cpation 2) The position of Z_i^s and Z_i^u of Modal M_i should be exchanged in Figure3. Also, Z_i_s and Z_i^u of Modal M_j should be Z_j^s and Z_j^u in Figure3. 3) Please check the position of + symbol in Eqn5. 4) Why the value of Eqn4 should be optimized to maximum value? Can you explain with an example in multimodal contrastive learning? Technical Quality: 2 Clarity: 2 Questions for Authors: Why the value of Eqn4 should be optimized to maximum value? Can you explain with an example in multimodal contrastive learning? Confidence: 3 Soundness: 2 Presentation: 2 Contribution: 2 Limitations: The experiments conducted in text and vision modal. Experiments on more modal data should be conducted to show the generalization ability. Flag For Ethics Review: ['No ethics review needed.'] Rating: 7 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Dear Reviewer KzFg, We sincerely appreciate your thorough review and insightful comments. Please find our responses below. ------ W1-3: ”Presentations/Grammar/Typos” A1: We apologize for the some typos in this paper and we have fixed it. We have re-examined Equation 5 and refined them. The notation $\mathbf z_j^{\mathbf n+}$ now correctly refers to the positive samples from modality $j$, and $\mathbf Z_{jk}^{\mathbf n-}$ refers to the negative samples from modality $j$. We appreciate your careful review and guidance. W4: “Maximize the Equation 4 value” A2: We apologize for any confusion caused by our oversight. In fact, we intend to maximize the absolute value of Equation 4 (Appendix.B3 for more details). We will correct this in the new version. First, we investigate why InfoNCE falls into shortcuts. We provide a simplified explanation in Equation 6: $-\frac{\partial\mathcal L_{\mathrm{InfoNCE}}}{\partial Z_{a}} = \frac{1}{\tau}(Z_{b}^{+} - \sum_{i=0}^{N}\beta_{i}Z_{bi})$. The term $\frac{1}{\tau}Z_b^+$ brings positive samples closer together, maximizing . $\|\|Z_{a}\|\|\cdot\|\|Z_{b}^+\sin \alpha$.However, in multi-view scenarios, $Z_{b}^{+}$ is sampled from $k$ positive samples as $Z_{b1}^{+}, Z_{b2}^{+}, \ldots, Z_{bk}^{+}$. Under the guidance of the gradient, the final representation tends to capture the shared information among all positive samples, leading to shortcuts. To capture unique information, we disentangle features into shared and unique representation. We apply strict constraints to bring the shared representations from different modalities closer together, while applying weaker constraints to keep the unique representations in the same plane (Intuitively, the unique information between different views is unrelated, we will not pull them closer) as shown in Figure 1(b). Maximizing the absolute value of Equation 4 is equivalent to optimizing the quaternion vectors $(\mathbf{Z}_i^\mathbf{s},\mathbf{Z}_i^\mathbf{u},\mathbf{Z}_j^\mathbf{s},\mathbf{Z}_j^\mathbf{u})$ to be in the same plane. ------ L1: ”Generalization ability for more modality” A3: We conducted additional experiments on common modalities including images, text, and audio. Specifically, we performed image-audio experiments using the FMA and GTZAN datasets, as shown in Table 1. For text-audio experiments, we utilized the CLOTHO and AUDIOCAPS datasets, with results presented in Table 2. All of our models outperformed the baseline. Almost all of our models outperformed the baseline. For simplicity, we employed the almost simplest decoder structure (linear layers) and did not implement modality fusion or any cross-modal interactions. We believe that enhancing the architecture, strengthening cross-modal interactions, employing a larger batch size and incorporating pre-training would yield even higher performance. Table 1: The results on the GTZAN and FMA datasets. \| Method \| Datasets \| R@1 \| R@5 \| R@10 \| R@1 \| R@5 \| R@10 \| \| ------- \| -------- \| --------- \| --------- \| --------- \| --------- \| --------- \| --------- \| \| \| FMA \| \| i2a \| \| \| a2i \| \| \| InfoNCE \| \| 15.87 \| 28.62 \| 35.87 \| 12.50 \| 25.50 \| 29.12 \| \| QUEST \| \| 17.83 \| 30.87 \| 38.50 \| 13.50 \| 26.52 \| 30.62 \| \| \| GTZAN \| \| i2a \| \| \| a2i \| \| \| InfoNCE \| \| 34.01 \| 84.73 \| 94.41 \| 32.48 \| 78.68 \| 90.86 \| \| QUEST \| \| 41.62 \| 88.83 \| 97.65 \| 35.53 \| 82.23 \| 93.40 \| Table 2: The results on the CLOTHO and AUDIOCAPS datasets. \| Method \| Datasets \| R@1 \| R@5 \| R@10 \| R@1 \| R@5 \| R@10 \| \| ------- \| --------- \| --------- \| --------- \| --------- \| --------- \| --------- \| --------- \| \| \| CLOTHO \| \| t2a \| \| \| a2t \| \| \| InfoNCE \| \| 20.16 \| 51.30 \| 66.56 \| 20.06 \| 52.66 \| 68.23 \| \| QUEST \| \| 21.10 \| 52.45 \| 68.86 \| 22.36 \| 54.23 \| 70.42 \| \| \| AUDIOCAPS \| \| i2a \| \| \| a2i \| \| \| InfoNCE \| \| 4.59 \| 17.79 \| 26.22 \| 5.45 \| 15.78 \| 22.48 \| \| QUEST \| \| 5.16 \| 18.08 \| 27.17 \| 6.02 \| 15.98 \| 24.78 \| --- Rebuttal Comment 1.1: Comment: Thank you very much for the author's response. Most of my doubts have been resolved, and I have raised my score. --- Reply to Comment 1.1.1: Comment: Dear Reviewer KzFg, We extend our sincere gratitude for your thorough review of our work. We humbly accept the issues you've raised and we will address them in future iterations of our manuscript. Your insightful feedback is invaluable to us, and we are dedicated to improving our work based on your suggestions.	null	null	Rebuttal 1: Rebuttal: Dear reviewers, we would like to sincerely thank all the reviewers for taking the time to read our paper and provide valuable feedback. We are delighted that reviewer KzFg (zJ9C) acknowledged the superior performance, innovation (9n5E, zJ9C), and reasonable motivation (9n5E, zJ9C) of our approach. Additionally, we appreciate all the reviewers' comments, which we have taken into consideration during the rebuttal period and made the following revisions: 1. According to reviewer KzFg's comments, we conducted experiments on additional modalities. To simplify, we did not employ any multimodal techniques (e.g., cross-modal interaction) and only added linear layers. Table 1 demonstrates that Quest(ours) achieved superior performance in the image-audio modality compared to the baseline model. Table 2 shows that Quest outperformed the baseline model in the text-audio modality. The experimental results indicate Quest's generalizability across more modality data. 2. In response to reviewers KzFg and 9n5E, we added examples of shortcuts in multimodal contrastive learning in Figure 1. Handwritten digits were simultaneously added to both text and image samples (with minimal impact on the original meaning of the images). However, training on these datasets led to a significant drop in model performance due to the tendency to learn easy features. Our model effectively addresses this issue. Even without adding these handwritten digits, an image is often described by multiple captions that usually hold different meanings. This is common in the real world and underscores the significance of our approach. 3. We appreciate reviewers KzFg and zJ9C for their suggestions regarding our phrasing and terminology. In our rebuttal, we provided explanations and committed to correcting all typos and terminology. Please note that we have added tables and figures in the attached pdf to support our responses to the reviewers KzFg, 9n5E, and zJ9C. Pdf: /pdf/2c4266cf447844fc9b06f24d0ad2f74560fcca3d.pdf	NeurIPS_2024_submissions_huggingface	2,024	null	null	null	null	null	null	null	null
Toward Conditional Distribution Calibration in Survival Prediction	Accept (poster)	Summary: This paper proposes a new postprocessing method, CSD-iPOT, for survival analysis based on conformal prediction. Strengths: The proposed method tries to achieve the conditional calibration, which is known to be hard. Weaknesses: This paper puts emphasis on achieving conditional calibration (not marginal calibration), but the conditional calibration is known to be hard to achieve even for datasets without any censored data point. See, e.g., + Lei, J. and Wasserman, L. (2014). Distribution-free prediction bands for non-parametric regression. Journal of the Royal Statistical Society: Series B (Statistical Methodology), 76(1):71–96. + Rina Foygel Barber, Emmanuel J Candes, Aaditya Ramdas, Ryan J Tibshirani, The limits of distribution-free conditional predictive inference, Information and Inference: A Journal of the IMA, Volume 10, Issue 2, June 2021, Pages 455–482. + Zhao et al., Individual Calibration with Randomized Forecasting, ICML 2020. This paper does not cite these papers, and this paper does not discuss the hardness of conditional calibration. Lack of extensive discussion on the hardness of conditional calibration is a serious problem of this paper. Despite the hardness, this paper claims that the conditional calibration can be achieved in Theorem 3.2. However, this theorem actually shows nothing: it shows only that, if we have an estimator that achieves the conditional calibration (i.e., Eq. (10) is satisfied), the output of the proposed method achieves the conditional calibration, too. I think that if we have an estimator that achieves the conditional calibration, the proposed method is not required. Furthermore, there are many other problems in this paper: + While $\Gamma_{M}$ should be a set of scalar values according to line 9 of Algorithm 1, $\Gamma_{M}$ is a set of pairs of scalar values according to line 11 of Algorithm 1. + According to lines 180-183, $R$ copies of comformity scores $\gamma_{i,M}$ are generated for each uncensored data point. If so, the equation between lines 537 and 538 is incorrect. + According to lines 180-183, $R$ copies of comformity scores $\gamma_{i,M}$ are generated for each uncensored data point. Neverthless, this paper assumes that the set $\Gamma_{M}$ does not have any tie in line 538. + The proof of Theorem C.1 are not fully described. Even though the goal of the proof is to prove equation (9) on $\rho_{1}$ and $\rho_{2}$, the proof for censored data points does not argue $\rho_{1}$ and $\rho_{2}$. An equation for censored data points analogous to the equation for uncensored data points (between lines 537-538) must be presented. + The proof of Theorem C.1 completely ignores $R$, even though Theorem C.1 does not hold if $R=1$. (I think that the main idea of the proposed method, CSD-iPOT, is to use a sufficiently large $R$ to "blur" the censored data points.) + The assumptions used in the proof of Theorem C.2 are not clearly stated before the proof: many implicit assumptions are used in lines 563 and 564. In particular, during the proof of (i), a statement (i.e., an implicit assumption) similar to (i) is used in line 563. + In the proof of Theorem C.2, the alleged proof of (i) on $x_{n+1}$ does not include any discussion on $x_{n+1}$. ------------------------ # Additional Comments I submitted the following comments in the Author-Reviewer discussion period, but it accidentally did not visible to the authors (probably due to the complex system on comments visibility of OpenReview). I noticed this fact during the Reviewer-AC discussion period, and the AC allowed me to post the comments here. ==comments begin== Thank you for your comments. I will keep my score. This paper has several critial problems. + The authors' comments did not give any evidence that the proposed algorithm achieves conditional calibration (claimed in lines 47-48). Since the hardness of the conditional calibration is an important topic in machine learning as already studied by many researchers, the authors must pay careful attention when they discuss on this topic. + Huge discrepancy between the implemented algorithm (with R=1000) in the experimental section and the provided proof (valid only for R=1), even though the key idea of the proposed algorithm CSD-iPOT is to "blur" a censored subject with a large . Regarding the presentation, all the assumptions must be clearly presented before the proof. The assumption is stated at the end of the proof (in lines 546-548). Minor thing: The authors violated the rule on the 1-page pdf in the rebuttal phase: "Please use this PDF only for figures (including tables) and captions that describe the figure." ==comments end== Technical Quality: 1 Clarity: 2 Questions for Authors: Nothing. Confidence: 4 Soundness: 1 Presentation: 2 Contribution: 2 Limitations: This paper does not discuss the hardness of conditional calibration. Flag For Ethics Review: ['No ethics review needed.'] Rating: 3 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We thank the reviewer for providing interesting comments but are sorry that you did not agree with the other reviewers about the merit of our paper. To deal with your concern: >W1: … Lack of extensive discussion on the hardness of conditional calibration is a serious problem of this paper. Thank you for stating your concern. However, we believe that this concern is mostly because this reviewer’s perspective is very different from ours and the other reviewers'. While our paper clearly relates to statistical issues, our goal was a method that survival analysis researchers can use to effectively achieve good calibration and maintain discrimination. That is why the main text de-emphasized theoretic proofs, and instead focused on intuitive explanations (using many real-world examples) of why our method should work. This stems from our concern that if our paper focused too much on conformal prediction and its history, the readers (mainly survival analysis researchers, because it is a tool for them) would lose interest. Furthermore, Lei and Wasserman (2014) and Barber et al (2021) both note that conditional calibration is hard as they have shown that the finite sample guarantee for conditional calibration is impossible to achieve. This is why our paper does not try to provide any finite sample guarantee. Instead, we follow the route of many other researchers (Romano et al. (2019); Sesia and Candes (2019) [1]; Izbicki et al. (2020); Chernozhukov et al. (2021) [2]) by providing only an asymptotic guarantee for infinite samples. The main idea of this asymptotic conditional guarantee is that the construction of the post-conformal predictions relies on the original prediction, and we only hope for conditional calibration for the class of predictions that can be learned well (which correspond to the assumptions explicitly made in the paper, see also in W2). Lastly, despite the hardness of conditional calibration and the limitation of our assumptions, our extensive experiments have shown the effectiveness and robustness of the method for 15 datasets and 7 baselines – to our best knowledge, no previous conformal prediction paper has provided such a broad empirical validation. While this conditional guarantee is hard in theory, our experiments certainly support it – which we view as a merit of our paper. We again appreciate the suggestion. Towards reaching a broad audience, the revised manuscript will include a brief discussion about the hardness issue in the main text and also a new Appendix section to extensively discuss this. [1] A comparison of some conformal quantile regression methods. Stat [2] Distributional conformal prediction. PNAS >W2: … Theorem 3.2 actually shows nothing We acknowledge that the consistent estimator assumption in Theorem 3.2 is weak. As mentioned earlier, our paper focuses on asymptotical guarantee – only providing conditional calibration in the class of predictions that can be learned well. This assumption strictly follows Assumption 3 in Sesia and Candes (2019); Assumption 2.3 in Izbicki et al. (2020); and Assumption 1 in Chernozhukov et al. (2021). Therefore, our Theorem 3.2 has the same practicality as those previous methods. We also thank the reviewer for confirming that proving the conditional calibration is hard. Moreover, to our best knowledge, there are no results that can prove conditional calibration with more relaxed assumptions. Finally, note that our paper focuses on empirical performance more than theoretical results. >W3: … $\Gamma_M$ is a set of pairs of scalar values according to line 11 of Algorithm 1. Thank you for noting this. Algorithm 1 uses this problematic expression (and also adds a direct comment behind this expression to explain) as we did not have a better notation to represent repeating a value by R times. The revised version will reduce this confusion by including an underbrace to express repeating a value. >W4: … If so, the equation between lines 537 and 538 is incorrect … As shown under W7, here we prove that this theorem holds for $R=1$. Therefore, the equation between lines 537-538 is correct. >W5: … this paper assumes that the set $\Gamma_M$ does not have any tie in line 538 … This “no tie” assumption is a technical assumption in conformal regression to make the proof. This assumption is easy to solve in practice because users can always add a vanishing amount of random noise to the scores to avoid ties. >W6: …the proof for censored data points does not argue $\rho_1$ and $\rho_2$ Thank you for pointing this out. The original proof only considers the upper bound to be adaptive and sets the lower bound to 0. We have included the revised derivation to incorporate an adaptive lower bound – see Figure 3 in the 1-page PDF. The revised derivation is almost the same as the original but has an extra term when breaking the joint probability into pieces. Fortunately, this modification does not compromise the validity of the conclusion. >W7: … Theorem C.1 completely ignores 𝑅, even though Theorem C.1 does not hold if 𝑅=1 … In fact, the proof shows the theorem holds for $R=1$, as stated in lines 546-548: “if we do one sampling for each censored subject …, the above proof asymptotically converges…”. The repeating strategy is just to get a more accurate estimation for finite samples. >W8: …many implicit assumptions are used in lines 563 and 564.… during the proof of (i), a statement (i.e., an implicit assumption) similar to (i) is used in line 563. Sorry about the confusion. The revised version will put the assumption (now in line 564) before the proof. Note however that the statement in line 563 is directly derived from line 562, so it is not an assumption. >W9: In the proof of Theorem C.2, the alleged proof of (i) on $𝑥_{𝑛+1}$ does not include any discussion on $𝑥_{𝑛+1}$. Thanks for catching this typo! In lines 559-565, the conditions should be $x_{n+1}$, not $x_i$, which will be corrected in the revised manuscript.	Summary: The authors study the problem of how to create individual survival distribution (ISD) models which are well-calibrated, both in a marginal and conditional sense, without negatively affecting the discriminative performance. They refine the "Conformalized Survival Distribution" (CSD) approach and propose "Conformalized Survival Distribution using Individual survival Probability at Observed Time" (CSD-iPOT). Both are post-processing methods which utilize conformal prediction, and can be applied on top of various survival analysis models to improve their calibration. CSD-iPOT is however designed to not only improve marginal calibration, but also conditional calibration. The method is evaluated on 15 datasets, comparing baseline (without neither CSD nor CSD-iPOT), CSD and CSD-iPOT versions of 7 survival models. The results are quite promising overall. Strengths: - The paper studies an interesting and important problem, how to build well-calibrated and discriminative survival analysis models. The special focus on conditional calibration is also important. - The paper is well-written and very solid overall, the authors definitely seem knowledgeable. - The general idea of the proposed method, to utilize the conformal prediction framework in order to adjust predicted survival distribution curves on calibration data, makes sense. - The evaluation is quite extensive with 15 datasets and 7 baseline survival analysis models, and a direct comparison with CSD. - The results are quite promising overall, CSD-iPOT improves the calibration of CSD in most cases without negatively affecting the discrimination very often. Weaknesses: - I found parts of Section 3 and 4 a bit difficult to follow, the proposed method/metric could perhaps be described in a more intuitive way. Figure 2 is neat, but would be helpful to have a similar visualization to illustrate the difference between CSD and CSD-iPOT. - The technical novelty/innovation compared to CSD is perhaps somewhat limited, both methods employ the same general approach (post-processing methods which utilize conformal prediction). - The experimental results could be more convincing. CSD-iPOT usually improves the calibration compared to CSD, but far from always, and the gains also seem to be relatively small quite often. Overall, I find it quite difficult to judge how much added benefit CSD-iPOT actually would have compared to CSD in practice. Are there concrete examples where CSD leads to significant mis-calibration within a certain patient subgroup, which is addressed by CSD-iPOT? The current results/metrics are a bit abstract / difficult to interpret. Summary: - Well-written and very solid paper overall that studies an important/interesting problem. The technical novelty compared to CSD, and how much better the proposed method actually would be in practice, is however a bit unclear. I am leaning towards accept. * * * * Update after the rebuttal: I have read the other reviews and all rebuttals. 2/3 other reviews are also positive, and I think the authors respond well to the third. All my questions were addressed, Figure 1 and 2 in the provided pdf are neat. I will increase my score to "7: Accept", this is a very well-written and solid paper that I think should be accepted. Technical Quality: 3 Clarity: 3 Questions for Authors: - 189: "Unlike CSD [8], which adjusts the ISD curves horizontally (changing the times, for a fixed percentile), our refined version scales the ISD curves vertically", could you perhaps visualize how both CSD and CSD-iPOT modify the ISD curves in an example? I think that could help illustrate how these two methods differ. - I thought that CSD-iPOT shouldn't affect the relative ordering of subjects, but when comparing the ISD curves before and after in Figure 2(d) and (e), the top orange curve is always above the top blue curve in (e), whereas it is partially below the blue curve in (d)? - 236: "Qi et al. [8] demonstrated that CSD theoretically guarantees the preservation of the original model’s discrimination performance in terms of Harrell’s concordance index (C-index) [1]. However, CSD-iPOT lacks this property", but both methods have a check mark for "Discrimination guarantee Harrell’s" in Table 1? Minor things: - 77: "In summary, individual calibration is ideal but not impractical", impractical --> practical? - 219: "with adequate modifications to the accommodate our method", remove "the"? - 297: "Compared CSD" --> "Compared to CSD"? Confidence: 4 Soundness: 3 Presentation: 3 Contribution: 2 Limitations: Yes. Flag For Ethics Review: ['No ethics review needed.'] Rating: 7 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We thank the reviewer for the insightful comments and suggestions. To address the concerns: >W1 & Q1: … would be helpful to have a similar visualization to illustrate the difference between CSD and CSD-iPOT. Great suggestion! The revised version will include a side-by-side visual comparison of our method to CSD – see Figure 1 in the additional 1-page PDF. After the paper's acceptance, we will add two animated GIFs to the GitHub repo, to better visually illustrate the difference between our method and CSD. >W2: The technical novelty/innovation compared to CSD is perhaps somewhat limited, both methods employ the same general approach (post-processing methods which utilize conformal prediction). It is true that both CSD and CSD-iPOT use the same general idea of conformal prediction. However, it is important to highlight that conformal prediction is an active area of research, with a variety of methods aimed at enhancing prediction coverage and calibration more effectively and efficiently. Specifically, the novelty of this work stems from the unique design of the conformity score, and the downstream method for handling censored subjects (thanks to the conformity score design), which together provide a theoretical guarantee for the calibration. >W3: .. Overall, I find it quite difficult to judge how much added benefit CSD-iPOT actually would have compared to CSD in practice. Are there concrete examples where CSD leads to significant mis-calibration within a certain patient subgroup, which is addressed by CSD-iPOT? … While our extensive experimental results did not find universal improvement, we did observe CSD-iPOT superior to CSD in 68/104 cases (significantly in 37 cases) for marginal calibration, and in 51/69 cases (significantly in 26 cases) in conditional calibration – see Table 2 in the paper. So in real applications, this method should be in your quiver of tricks to consider. Furthermore, thanks to the reviewer’s great suggestion, we provide 4 case studies in Figure 2 of the 1-page PDF – concrete examples where CSD leads to significant miscalibration within certain subgroups (elderly patients, women, high-salary, and non-white-racial), but CSD-iPOT can effectively generate more conditional calibrated predictions. Furthermore, all 4 examples show that CSD’s miscalibration is always located at the low-probability regions, which corresponds to our statement (lines 208-212) that the conditional KM sampling method that CSD used is problematic for the tail of the distribution (low-probability regions). >Q2: … the top orange curve is always above the top blue curve in (e), whereas it is partially below the blue curve in (d)? Thank you for your insightful comment. Yes, the blue curve is partially at the top in (d), intersecting the orange curve around 1.7 days, while the orange curve is consistently at the top in (e). This discrepancy arises from the discretization step used in our process, which did not capture the curve crossing at 1.7 days due to the limited number of percentile levels (2 levels at 1/3 and 2/3) used for simplicity in this visualization. The post-discretization positioning of the orange curve above the blue curve in Figure 2(e) does not imply that the post-processing step alters the relative ordering of subjects. Instead, it reflects the limitations of using only fewer percentile levels. Note that other crossings, such as those at approximately 1.5 and 2.0 days, are captured. In practice, we typically employ more percentile levels (e.g., 9, 19, 39, or 49 as in Ablation #2), which allows for a more precise capture of all curve crossings, thereby preserving the relative ordering. >Q3: …both methods have a checkmark for "Discrimination guarantee Harrell’s" in Table 1? We thank the reviewer for carefully examining this discrepancy between the statement and Table – yes, there is a typo in Table 1. For the proposed CSD-iPOT, the “Monotonic” should be a check, and “Discrimination guarantee Harrell’s” should be a cross. We are sorry for this mistake and will revise this and other typos in the revised version. --- Rebuttal Comment 1.1: Comment: Thank you for the response. I have read the other reviews and all rebuttals. 2/3 other reviews are also positive, and I think the authors respond well to the third. All my questions were addressed, Figure 1 and 2 in the provided pdf are neat. I will increase my score to "7: Accept", this is a very well-written and solid paper that I think should be accepted.	Summary: The paper enhances the Conformalized Survival Distribution (CSD) post-processing framework to account conditional calibration. The proposed framework, CSD-iPOT, utilizes a conformal set to adjust survival curves vertically, aligning them with predetermined percentiles at test time. Unlike CSD, which relies on Kaplan-Meier curves, CSD-iPOT leverages Individualized Survival Distributions (ISD) for censored events when constructing the conformal set. The paper provides theoretical guarantees for marginal calibration, conditional calibration, and the monotonicity of ISD. Comprehensive experimental results from 15 datasets demonstrate that CSD-iPOT enhances both the marginal and conditional calibration of baseline models with minimal impact on the concordance index (C-index). Strengths: - The paper is well-written and easy to follow. - The reviewer appreciates that the visual plots provided offer great intuitive illustrations of the proposed approach. - This is the first paper to address conditional calibration in survival analysis, an important yet under-explored problem. - The paper provides theoretical guarantees for the proposed approach in terms of marginal calibration, conditional calibration, and monotonicity. - CSD-iPOT is more computationally efficient in terms of storage than CSD. - Extensive experimental results across 15 datasets and 7 baselines demonstrate that CSD-iPOT significantly improves the calibration (both marginal and conditional) of baseline models with minimal loss in the concordance index (C-index). Weaknesses: The paper highlights that CSD-iPOT lacks theoretical guarantees for preserving Harrell's concordance index. This is a major limitation of this work; nevertheless, experimental results demonstrate that the impact is minimal The description in lines 142-162 requires some improvements: - Eqn. 4: Needs to be adjusted to something like $\tilde{S}^{-1}(p \| x_{n+1}) = T(\hat{S}^{-1}(\text{Percentile}(\cdot) \| x_{n+1}))) $, where $T$ is the proposed vertical transformation. Then provide all the necessary numbered steps to finally obtain the calibrated ISD (Eqn. 5). For completeness, I encourage the author(s) to also benchmark competitive baseline models that directly model event times, e.g., Chapfuwa et al. 2018 and Miscouridou et al. 2018 Minor: - In Table 1, CSD-iPOT should be marked with an 'X' under Harrell’s concordance index category. - Line 297: Typo -> "Compare to CSD" Technical Quality: 4 Clarity: 3 Questions for Authors: - Given that CSD-iPOT lacks theoretical guarantees for preserving Harrell's concordance index, is there a hyperparameter that explicitly controls this trade-off? - It seems that CSD-iPOT achieves calibration lower than the empirical lower limit set by Kaplan-Meier. Are these the instances that result in an impact on the C-index? - Why does CSD-iPOT struggle with both calibration and the C-index for models such as DeepHit? - Are the calibration metrics obtained at different percentiles than those used for the post-processed ISD? What happens if these two sets are different? - Why does CSD-iPOT improve marginal and conditional calibration models better than CSD? Is the difference due to how CSD-iPOT handles censored events? Confidence: 4 Soundness: 4 Presentation: 3 Contribution: 3 Limitations: I encourage the authors to discuss the limitations of their work, including any violations of modeling assumptions. Flag For Ethics Review: ['No ethics review needed.'] Rating: 7 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We thank the reviewer for these wonderful comments and suggestions! Wrt your insightful concern: >The paper highlights that CSD-iPOT lacks theoretical guarantees for preserving Harrell's concordance … Yes, we acknowledged this limitation in the paper. However, we argue this is not a big issue for two reasons: 1. While our method does not have a preservation guarantee for Harrell’s C-index, it does have this guarantee for two other discrimination metrics: AUROC and Antonili’s C-index, which are also (arguably more) commonly used in application studies. Due to their distinct nature, we do not know any method that is guaranteed to preserve all these three notions of discrimination (and suspect this might not be possible). 2. While our method is not guaranteed to preserve Harrell’s C-index, our extensive experimental results (on 104 comparisons) show that, in 84 cases, our method does not decrease the C-index. And for the 22 cases that it does decrease, none of them are significant. This really shows that this lack-of-formal-guarantee seems minor. >The description in lines 142-162 requires some improvements … Thanks for the suggestion. This helped us to identify a possible misunderstanding. The proposed vertical transformation pertains solely to Eq 4. It does not apply to Eq 5, which transforms the inverse-ISD back to the ISD function (i.e., from Figure 2(e) to curves similar to Figure 2(a)). However, we realized that the position of lines 158-162 might cause this misunderstanding (now it is right after Eq 5); we will move this description ahead, to lead to a better presentation of the method. >I encourage the author(s) to also benchmark competitive baseline models that directly model event times … We appreciate this suggestion. However, the proposed CSD-iPOT method requires the original survival models to be able to generate survival distributions (Section 3.1). If a baseline model can only generate a scalar value, it is not clear how to convert the scalar value into a distribution. Also, because such a model only generates a time prediction (which is not a probability), it is also unclear what calibration means for time prediction. However, if the reviewer has any insight on this, please let us know and we are happy to use your approach to benchmark these models. >Q1: … is there a hyperparameter that explicitly controls this trade-off? No, there is no hyperparameter to control how much the CSD-iPOT process modifies the C-index. However, we do not consider this as an issue because (1) in 75/104 cases, the CSD-iPOT process did not affect the C-index, (2) for the remaining cases, the effect is not always negative: in 7/104 cases, CSD-iPOT improves the C-index. >Q2: It seems that CSD-iPOT achieves calibration lower than the empirical lower limit set by Kaplan-Meier. Are these the instances that result in an impact on the C-index? Yes. The same columns in Figure 3 correspond to the same instances. The reviewer might think that there is an implicit trade-off – that increasing calibration is associated with decreasing C-index. However, this is not the case for our method: the level of calibration improvement is not associated with the level of C-index decreasing. For example, in the HFCR datasets with the GB baseline, our CSD-iPOT improves both C-index and calibration (Figure 3, left). >Q3: Why does CSD-iPOT struggle with both calibration and the C-index for models such as DeepHit? Appendix C.4 discusses why our method is sub-optimal for these models (DeepHit and CQRNN), as indicated by lines 300-301 in the main text. In summary, this is because such models are often significantly miscalibrated, by implicitly assuming that, by the end of the predefined $t_{\text{max}}$, every individual must have had the event, and therefore their predicted survival distributions must drop to 0% at that time. In such cases, because CSD-iPOT can not have any intervention for the ending probability (0%), therefore, the ending position cannot be moved. Figure 7 shows an example of this issue, showing that the post-processed distribution has a sharp-dropping tail, which leads to miscalibration in those regions. >Q4: Are the calibration metrics obtained at different percentiles than those used for the post-processed ISD? ... The percentiles for calibration evaluation are always set to be 10%, 20%, … 90%, as recommended by Haider et al. However, our Ablation #2 explores the impact of different predefined percentiles for the proposed CSD-iPOT. The short summary (lines 998-999) is that the number of percentiles has no impact on the C-index and a slight impact on marginal and conditional calibration. More details on the experiment settings and results can be found in Appendix E.6. >Q5: Why does CSD-iPOT improve marginal and conditional calibration models better than CSD? Is the difference due to how CSD-iPOT handles censored events? Yes, that is exactly right! For marginal calibration, because the CSD-iPOT shifts the survival probability values vertically (while CSD shifts the time prediction horizontally), it avoids the challenges of interpolation and extrapolation of the distribution, which is particularly problematic when the censoring rate is high and when KM ends at a high probability level. For conditional calibration, CSD-iPOT is better than CSD because CSD-iPOT considers the heterogeneity of the features by sampling the event times from the individual survival distributions, while CSD samples from the conditional KM distribution do not consider the features. The current paper discusses these two aspects in detail, in lines 206-215 and 194-205, resp. >I encourage the authors to discuss the limitations of their work, including any violations of modeling assumptions. Thanks for the excellent suggestion. As mentioned above, the current paper discusses two limitations of our methods. However, the revised version will also include a discussion about the violations of assumptions. --- Rebuttal Comment 1.1: Title: Official Comment by Reviewer kAoa Comment: Thanks for addressing most of my concerns. However, this statement "If a baseline model can generate only a scalar value, it is not immediately clear how to convert this scalar value into a distribution," is not necessarily accurate. Individual survival distributions can be obtained from parametric time-to-event models (e.g., AFT) since $f(t\|x) = h(t\|x) S(t\|x)$. For non-parametric approaches (e.g., DATE), time-to-event distributions are implicitly defined through sampling. After reviewing the rebuttal and reviewer's response, I find that overall, this is a solid paper, and I am still leaning more towards acceptance. I am keeping my score as it is. --- Reply to Comment 1.1.1: Title: Re: Official Comment by Reviewer kAoa Comment: We are grateful to the reviewer for taking the time to respond to our rebuttal. We are sorry we may have misinterpreted the request, as we thought was asking for a comparison of models which only generate time predictions. If the requirement is to compare models that can predict both ISDs and survival times, we confirm that all 7 baselines in the paper (AFT, GB, N-MTLR, DeepSurv, DeepHit, CoxTime, and CQRNN) qualify. Notably, the censored quantile regression neural network model (CQRNN, Pearce et al. 2022), a variant of quantile regression, directly predicts the median survival time for each individual, aligning with what the reviewer has requested.	Summary: While previous work focuses on calibration in a marginal sense, this paper proposes a post-processing approach that also imposes conditional calibration on all individual features. Therefore, the proposed method can guarantee equal calibration for different groups, ensuring fairness. Contributions: 1. The paper proposes the post-processing approach to ensure conditional distribution calibration in survival analysis (accommodating censorship) and develops the corresponding metric. 2. The paper provides asymptotic guarantees for both marginal and conditional calibration and demonstrates the computation complexity. 3. The paper validates the method by conducting experiments on 15 datasets. Strengths: Originality: The paper motivates the use of conditional distribution calibration in survival analysis via post-processing. The idea is intuitive as both the class-specific and the general survival probability at observed time should be uniformly distributed. Quality and significance: Although the calibration method is simple and known, the paper accommodates it to the censored setting and proves its effectiveness empirically. Besides, the paper provides corresponding theoretical guarantees. Lastly, it applies to various SOTA methods in survival analysis and improves the performance in most cases. Clarity: The paper clearly conveys the idea and method of calibrating survival probability curves. Weaknesses: 1. Missing related work in conformal prediction: the paper claims the method is based on conformal prediction but it is not mentioned and illustrated clearly in the main content. Thus, this knowledge gap hampers the understanding. 2. The method seems to be sensitive to the percentiles. As shown in Figures 2 (d) and (e), the post-processed curves are piecewise functions highly related to the percentiles (1/3, 2/3). Thus, the resulting curves would be sensitive to the choice of percentiles. Therefore, the number of percentiles should be included in the complexity and asymptotic behavior analysis. 3. The discrimination properties of the method are unclear. The authors first claim it lacks the discrimination performance preservation property for Harrel's C-index, which conflicts with the results in Table 1. 4. The empirical exploration and validation of the group fairness is missing. Technical Quality: 3 Clarity: 2 Questions for Authors: As mentioned in the weakness: Q1: How does the choice of percentiles affect the performance? Are there any empirical results that can illustrate the effect? Q2: Does the approach keep the Harrel C-index? Q3: There are plenty of survival analysis methods exploring the heterogeneity of groups. How does the proposed method compare to the SOTA heterogeneity methods? Will it improve the performance? Confidence: 3 Soundness: 3 Presentation: 2 Contribution: 2 Limitations: It remains unknown if this method would improve/hurt fairness in survival analysis. The study is missing. Flag For Ethics Review: ['No ethics review needed.'] Rating: 5 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We thank the reviewer for the insightful comments and suggestions. To address the concerns: > W1: Missing related work in conformal prediction: the paper claims the method is based on conformal prediction but it is not mentioned and illustrated clearly in the main content … The current method (detailed in Section 3) was intended to incorporate the detailed algorithm of conformal prediction so that readers with no conformal prediction background should be able to understand it, and the current Section 2.3 (lines 96-110) has summarized the related work on conformal prediction with censored data. However, thanks to the reviewer’s feedback, we will include a more detailed related work about standard conformal prediction. > W2 and Q1: How does the choice of percentiles affect the performance? Are there any empirical results that can illustrate the effect? As noted in line 314, our Ablation #2 study explored the impact of predefined percentiles, with results presented in the Appendix. The short summary (lines 998-999) notes that the number of percentiles has no impact on the C-index and only a slight impact on marginal and conditional calibration. More details on the experimental settings and findings can be found in Appendix E.6. As to the space complexity analysis, the number of percentiles indeed has an impact on that complexity. However, this impact is so subtle that in the big-O notation, it will be ignored. To expand more on this, Appendix E.5 establishes that the complexity of storing the conformity scores is $O(N \cdot R)$. After the method gets all the conformity scores, we will apply the percentile operation to all the conformity scores $\Gamma_{M}$ (as presented in Equation 4). The space complexity of this operation is just $O(\|\mathcal{P}\|)$ because there are only $\|\mathcal{P}\|$ scores to be saved (each for a unique percentile level) for later use. Therefore, the total space complexity is $O(N \cdot R + \|P\|) = O(N \cdot R)$, because $\|\mathcal{P}\|$ (a user-specific hyperparameter – here between 9-49) is significantly smaller than $N \cdot R$. > W3 and Q2: The discrimination properties of the method are unclear… We thank the reviewer for carefully examining this discrepancy between the statement and Table! This is indeed a typo in Table 1. For the proposed CSD-iPOT, the “Monotonic” should be “check”, and “Discrimination guarantee Harrell’s” should be “cross”. The rest of the paper, including the statement in the main text and proofs in the Appendix, supports the above claim. > W4 and Limitation 1: The empirical exploration and validation of the group fairness is missing… We thank the reviewer for bringing this up. Throughout the paper, we have considered the proposed conditional calibration metric also as a fairness metric. The resemblance to fairness metrics has been discussed multiple times in the paper, including lines 34-36 and 225-226. To briefly illustrate their resemblances, the proposed conditional calibration score evaluates the calibration performance across all possible subgroups (and reports the worst score in all the subgroups). This aligns with some fairness definition that clinical decision systems should guarantee equalized performance (e.g., accuracy) across any protected groups. However, we acknowledge that fairness has different definitions, and in this paper, we consider the proposed conditional calibration metric, $\text{Cal}_{\text{ws}}$, also as a fairness metric. We provide empirical results in Figures 4 and 11. > Q3: … How does the proposed method compare to the SOTA heterogeneity methods? Will it improve the performance? The proposed method is a model-agnostic post-processing method, meaning it can be adapted to any heterogeneity survival methods, as long as they can generate individual survival distributions (ISDs). Our experiments have extensively demonstrated, using 7 SOTA heterogeneity models, that our approach can improve the marginal and conditional calibration without decreasing the discrimination performance. As we noted, this approach can be very helpful in model development, as it allows researchers to simply seek models with superior discriminative abilities and subsequently apply the proposed method to improve calibration. This simplifies the model development process while ensuring robust performance across these key metrics. Furthermore, the proposed method aims to maximize conditional calibration, a score that evaluates the calibration performance by considering the heterogeneity of all possible subgroups. Ideally, one would seek extreme heterogeneity – i.e., individual calibration – which means the prediction is calibrated conditioned on any possible combination of features. However, for each unique combination of features, there is only one realization and therefore is impossible to perform the evaluation. Instead, we reach a middle ground – conditional calibration. This idea is the motivation of our paper and is discussed in detail in Section 2.2. --- Rebuttal Comment 1.1: Comment: Thank you for the responses. I have read the other reviews and all the rebuttals. The responses have addressed my concerns. I decided to keep my score as it is.	Rebuttal 1: Rebuttal: We sincerely thank all reviewers for their time and thoughtful feedback! In particular, we are grateful for the largely positive reception of our work. To mention a few key points from the reviewers: >The paper is well-written and easy to follow. The visual plot offers great intuitive illustrations and is neat (`kAoa`, `a6Gj`) >This is the first paper to focus on conditional calibration in survival analysis (`rRGm`, `kAoa`), which is important and hard (`kAoa`, `a6Gj`, `b9k1`). >This paper provides theoretical asymptotic guarantees for marginal and conditional calibration (`rRGm`, `kAoa`) >This method applies to various SOTA methods in survival analysis and improves the performance in most cases, which is promising overall. (`rRGm`, `kAoa`, `a6Gj`) The paper extensively validates the method by conducting experiments on 15 datasets and 7 baselines, and a direct comparison with CSD. (`rRGm`, `kAoa`, `a6Gj`) We are encouraged by the positive evaluations (scores of 7, 5, and 5) and are keen to address the concerns underlying the lower score (3) to bridge any gaps in understanding. We have carefully considered each point of criticism raised by the reviewers and our individual rebuttals provide detailed clarifications and justifications. Perhaps the main critique is from `b9k1`’s concern that “this paper does not discuss the hardness of conditional calibration", and “the proof of Theorem 3.2 is nothing” – leading to poor ratings on soundness. We appreciate the reviewer’s comments and hope our direct response clarifies the concerns. We hope our individual responses and the additional 1-page PDF effectively address all of the concerns raised. We are eager to engage further during the discussion phase and to answer any additional queries that may arise. Pdf: /pdf/08a6787ac935832951ef621321e28abe6f00182a.pdf	NeurIPS_2024_submissions_huggingface	2,024	null	null	null	null	null	null	null	null
MambaLLIE: Implicit Retinex-Aware Low Light Enhancement with Global-then-Local State Space	Accept (poster)	Summary: This paper proposes a Mamba-based framework, namely MambaLLIE, for low-light image enhancement. Specifically, the authors claim that they have two technical contributions: (i) A global-then-local state space block that integrates a local-enhanced state space module and an implicit Retinex-aware selective kernel module to capture intricate global and local dependencies. (ii) An implicit Retinex-aware selective kernel mechanism to guide deeper neural representations and segregate them into independent positive and negative illumination components before integrating them. Strengths: (i) The idea is novel and interesting. Mamba is a scorching research topic in computer vision now and has shown promising performance in many high-level vision tasks like image recognition, object detection, image segmentation, etc. But until now, there have been fewer efforts dedicated to the low-light image enhancement task, a small low-level topic. Thus, it is very exciting to see this wonderful work. (ii) The presentation is well-dressed. I like the style of the figures in this paper. For example, the teaser figure can clearly show the advantages in larger receptive fields of the proposed MambaLLIE than previous Transformer-based methods and Mamba-based image restoration methods. For another example, the pipeline in Figure 2 can also clearly show the workflow and the details of each submodule of the whole framework. (iii) The performance is good and solid. The proposed MambaLLIE significantly outperforms the state-of-the-art methods on six benchmarks including LOL-v2-syn, LOL-v2-real, SMID, SDSD-indoor, and SDSD-outdoor, as reported in Table 1. The visual comparisons in Figure 3 also show the effectiveness of the proposed method. Looks good. Weaknesses: (i) The motivation is not clear. To be specific, why use Mamba for low-light image enhancement? The authors claim that the Mamba can help capture long-range dependences. However, Transformer architecture can also model the long-range dependences. Why use Mamba instead of Transformer? Also, why design the space state module like this? The insight of the proposed Mamba should be analyzed in more detail. (ii) A technique should be explained, i.e., what is the scanning operation in the state space module? The authors mention the scan in Mamba but I could not find any mathematical formula or text description in the paper to explain the computation process of this scan. (iii) In table 1, although the performance of the proposed MambaLLIE is better than the state-of-the-art Transformer-based method Retinexformer, its computational complexity and memory cost are larger than those of Retinexformer. For example, the FLOPS and the Params are 33% and 38% higher. In addition, what about the comparison of performance on the MIT-Adobe-FiveK dataset (sRGB output mode)? (iv) In Figure 4, it seems that Retinexformer achieves more visually pleasant results. For example, in the third line of Figure 6, Retinexformer reconstructs the complete edge of the redwood in the left zoomed-in patch but the proposed MambaLLIE fails. Thus, it is better to compare MambaLLIE and the state-of-the-art method on unpaired datasets such as the LIME, NPE, MEF, DICM, and VV datasets. (v) The source code and pre-trained models have not been submitted. The reproducibility cannot be checked. Technical Quality: 3 Clarity: 3 Questions for Authors: I am curious about the training time comparison between the proposed MambaLLIE and other methods. Confidence: 5 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: Yes, the authors have analyzed the limitations in the main paper. Flag For Ethics Review: ['No ethics review needed.'] Rating: 8 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Q1: Why use Mamba for low-light image enhancement? A1: Our Motivation aims to take into account both global and local image restore. As we know, the low-light enhancement task faces challenges of global color degeneration and local noise disturbance. Global color degeneration suffer from the decrease of illumination, which tends to introduce a global-aware estimator, while the local degradation may contains the large or small regions. Hence, we intuitively introduce Mamba [8] into low light enhancement task, which excels at capturing the global dependencies of input data to restore. Compared with previous CNN and Transformer-based methods, Mamba-based model has exhibit promising performance with linear or near-linear scaling complexity in image super-resolution, image segmentation, etc., but still has the suboptimal results in low light enhancement task. Q2: Why use Mamba instead of Transformer? A2: We chose Mamba over Transformers due to its efficient long-range dependency modeling with linear complexity. Transformers can model long-range dependencies, but they typically involve higher computational complexity and memory costs. Q3: Why design the space state module like this? A3: Most prior VSSMs use the different directional scan in their sequential state, which is necessary for strong performance on visual tasks. However, it has been empirically observed that CNN seem to work fine for 2D dependency of vision data. Ours local-enhanced design essentially introduces the local invariance into state space model, which can integrate the existing directional scan with our local-enhanced term into VSSM. By additional convolving 2D information into directional scan, our method ensures a closer arrangement of relevant local tokens, enhancing the capture of local dependencies. This technique is depicted in Figure 2 (c). Next, we follow the selection mechanism by parameterizing the VSSM parameters based on the input, allows the model to filter out irrelevant information and remember relevant information indefinitely, as pointed in [8]. Besides, our local-enhanced term can be regarded as a local consistency constraint for the state space for vision data, guiding the module to learn neighborhood features and thus enhancing robustness, literatures [47], [C1] likewise added constraints of local term to a state–space model as equation 10 and developed the estimate performance. Q4: A technique should be explained, i.e., what is the scanning operation in the state space module? A4: To allow Mamba to process 2D images, the feature map needs to be flattened before being iterated by the state space. Therefore, the unfolding strategy is particularly important. In this work, we follow [38], which uses scans in four different directions to generate scanned sequences. Q5: In table 1, although the performance of the proposed MambaLLIE is better than the state-of-the-art Transformer-based method Retinexformer, its computational complexity and memory cost are larger than those of Retinexformer. For example, the FLOPS and the Params are 33\% and 38\% higher. A5: Reducing the computational complexity and memory cost is not the key part of MambaLLIE. Our MambaLLIE targets addressing the contradiction between global and local context interaction in the VSSM and proposes the flexible and effective IRSK for the LLIE task to achieve a larger receptive field in terms of global and local information. Besides, we significantly reduced the computational complexity and memory cost compared with our competitor MambaIR and kept comparable costs with other competitors. Additionally, we acknowledge that RetinexFormer is an outstanding work, maintaining a good balance in terms of parameter and perfermance. Our method slightly surpasses it in terms of performance. Q6: What about the comparison of performance on the MIT-Adobe-FiveK dataset (sRGB output mode)? A6: Following your valuable suggestion, we conducted the experiments of MambaLLIE and recent SOTA methods on the MIT-Adobe-FiveK dataset in Table R5, which verifies the comparable results of our method with RetinexFormer and surpasses the performance of other SOTA models. Figure R5 further shows the qualitative comparisons of MambaLLIE and RetinexFormer, where ours can remarkably improve the underexposure that appeared in RetinexFormer's result. Q7: In Figure 4, it seems that Retinexformer achieves more visually pleasant results. For example, in the third line of Figure 6, Retinexformer reconstructs the complete edge of the redwood in the left zoomed-in patch but the proposed MambaLLIE fails. Thus, it is better to compare MambaLLIE and the state-of-the-art method on unpaired datasets such as the LIME, NPE, MEF, DICM, and VV datasets. A7: Please refer to Table R4. We compare two non-reference perceptual metrics, MUSIQ and NIMA, on five unpaired datasets, including LIME, VV, NPE, MEF, and DICM. Experimental evaluations show the superiority of our method over SOTAs with better perceptual evaluation in most comparisons. Q8: The source code and pre-trained models have not been submitted. A8: Now, our code and the pretrained model are released on our project page. Please refer to our project page. Q9: Training time comparison A9: For example, training on the LOLV2 real dataset takes approximately 19 hours using PyTorch on a server equipped with 4090 GPUs, while Retinex tasks approximately 22 hours and MambaIR tasks approximately 27 hours . For the corresponding training parameters, please refer to lines 205-208 of the manuscript. --- Rebuttal Comment 1.1: Title: Response to the author rebuttal Comment: Thanks for your response. Most of my concerns have been addressed except one. You claim that `Mamba over Transformers due to its efficient long-range dependency modeling with linear complexity. Transformers can model long-range dependencies, but they typically involve higher computational complexity and memory costs.` However, this is not true. Many Transformers also enjoy linear computational complexity, e.g., Retinexformer. It takes much shorter time to train than the Mamba-based methods. Anyway, this is a good attempt to explore the potential of Mamba in LLE. Thus, I decided to raise my score to `strong accept` to support you. Thanks for your effort. --- Reply to Comment 1.1.1: Comment: We sincerely appreciate your recognition of our work and effort. YES, You are right, as discussed in [3], "O(IG-MSA) of Retinexformer is linear to the spatial size, mainly comes from the k computations of the two matrix multiplication of attention." Compared to the complexity of the global MSA used by previous Transformer methods, Retinexformer significantly reduces the computational complexity of attention while enhancing performance in LLIE tasks. Besides, our MambaLLIE inherits the linear complexity of SSM for visual representation learning, and a series of improvements in state space and Retinex-aware module are adopted to improve the performance, which have not significant and even negligible impact on overall complexity. Finally, we once again extend our gratitude to the reviewer for the constructive feedback during the review and rebuttal stages, giving a final score of strong accept for our manuscript. The insightful comments on complexity remind us to maintain rigorous discussions and analyses in our future work.	Summary: This paper presents MambaLLIE, an implicit Retinex-awre low light image enhancement framework with modified state space blocks. Specifically, a global-then-local state space block (GLSSB) is designed, which incorporates a local-enhanced state space module (LESSM) and an implicit Retinex-aware selective kernel (IRSK) module. By enhancing the original SSMs with local bias, the proposed MambaLLIE outperforms state-of-the-art CNN and Transformer-based methods. Strengths: 1. The main idea of this paper is well illustrated. 2. Experimental results and user studies demonstrate that the paper achieves better performance than the compared approaches. Weaknesses: 1. The novelty of this paper is limited. Like many Transformer-based works that embed local modeling capabilities in ViTs, the main contribution of this paper is incorporating local dependencies into SSMs. However, the advantages of the proposed GLSSB compared to ViTs and original SSMs have not been thoroughly analyzed. Furthermore, given the authors' design, the proposed GLSSB should have advantages in many other visual tasks, which need to be demonstrated through more experiments. 2. The writing of this paper needs improvement as there are many unclear descriptions. For example, how is the illumination prior in the paper obtained, how does the proposed IRSK module work specifically, and what are the specific settings for the different variants in the ablation study? 3. Some state-of-the-art works have not been compared (e.g., DiffLL[1]), and the experimental results in this paper do not seem to outperform them. 4. As the low-light enhancement task typically encounters issues like color distortion or other visual artifacts, the authors should compare more perceptual evaluation metrics. Merely reporting PSNR and SSIM is not sufficient. [1] Jiang et al, Low-Light Image Enhancement with Wavelet-based Diffusion Models. Siggraph asia 2023. Technical Quality: 2 Clarity: 3 Questions for Authors: Please refer to the `Weaknesses'. Moreover, it appears that the color in the last column of Table 2 is incorrectly marked. Confidence: 5 Soundness: 2 Presentation: 3 Contribution: 2 Limitations: The authors have discussed the limitations in the paper. Flag For Ethics Review: ['No ethics review needed.'] Rating: 4 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Q1: About novelty and the advantages of GLSSB. A1: We argue that the GLSSB is novel in terms of _the new exploration of the vision state space model_ and _technical improvements for the Retinex-aware low light enhancement task_. Our Motivation aims to take into account both global and local image restore. As we know, the low-light enhancement task faces challenges of global color degeneration and local noise disturbance. Global color degeneration suffer from the decrease of illumination, which tends to introduce a global-aware estimator, while the local degradation may contains the large or small regions. Hence, we intuitively introduce Mamba [8] into low light enhancement task, which excels at capturing the global dependencies of input data to restore. Compared with previous CNN and Transformer-based methods, Mamba-based model has exhibit promising performance with linear or near-linear scaling complexity in image super-resolution, image segmentation, etc., but still has the suboptimal results in low light enhancement task. We posit that the existing state space model is limited in modeling local dependencies, leading to a failure in restoring details. Besides, end-to-end models may exhibit suboptimal performance in low-light enhancement tasks because they do not consider illumination priors. Based on the two aforementioned assumptions, we propose the local-enhanced state space module (LESSM) and implicit retinex-aware selective kernel module (IRSK), which is the modeling of local and global degradation coupled in our framework for low light enhancement task. Therefore, we believe our novelty and contribution is on the par with the SOTA methods. Therefore, we believe our novelty and contribution are on par with the SOTA methods. Since IRSK is specially designed for low light enhancement task, We replaced the VSSM of MambaIR with LESSM to demonstrate the effectiveness of ours. As shown in Figure R1 and Table R1, we demonstrate the effectiveness of the proposed LESSM in MambaIR on image restoration tasks, indicating that our approach also has a positive impact on other low-level vision tasks, as discussed by \textbf{Reviewer 1bAi}. While the differences between LESSM and the vanilla VSSM may appear relatively small in description, each has its key motivations and shows significant improvements on common benchmarks over the previous models. We believe that simplicity combined with effectiveness is one of the most favored traits in the field of computer vision. Q2: How is the illumination prior in the paper obtained and how does the proposed IRSK module work specifically. A2: Sorry for the confusion caused. In our framework, we propose a Retinex-aware kernel selective mechanism, where two coupled Retinex-aware priors are used to select the spatial context regions, the maximum and mean values of RGB images can be regarded as a rough illumination prior. We use it in a implicit manner instead of restore the illumination map or use a single illumination feature for feature guidance. Base on this, we clarify the work process of IRSK as stated in Lines 184-194. Q3: what are the specific settings for the different variants in the ablation study? A3: Baseline-1 directly uses the standard vision state space module (VSSM) to process flattened vision data in our proposed UNet-shaped framework, following the Norm → VSSM → Norm → channel attention layer flow as referenced in [14] for image denoising tasks. Baseline-2 introduces retinex thory $L=R \ast I$ to Baseline-1, where $I$ denotes the low light image and $R$ denotes the reflectance (enhanced image), $I$ is the illumination map. aims to estimate the illumination map instead of directly predicting the enhanced image, and then restores the enhanced result by $R=L/I$, as referenced in [13],[43],[45]. [C3] Guo X. LIME: A method for low-light image enhancement. In ACM MM. pages 87-91. 2016. Q4: Some state-of-the-art works have not been compared (e.g., DiffLL[1]). A4: As reported in [C4], DiffLL indeed outperforms ours on some benchmark datasets in terms of PSNR and SSIM. As Reviewer 1rUY commented, computational complexity and memory cost are also important parts of the LLIE task. We find that more performance and parameters (inference time over 0.1s, model size over 100M) of DiffLL may be required in the case of comparable performance, while our inference time is below 0.1s and model size is only about 2M. Besides, comparing MambaLLIE and the state-of-the-art method on unpaired datasets, our MambaLLIE outperforms DiffLL in terms of perceptual evaluation metrics and visual results. Please refer to Table R4 and Figure R3. In our revised version, we will add citations to this work and discuss it. [C4] Jiang et al, Low-Light Image Enhancement with Wavelet-based Diffusion Models. Siggraph Asia 2023. Q5: The authors should compare more perceptual evaluation metrics. A5: Please refer to Table R4. We compare two non-reference perceptual metrics, MUSIQ and NIMA, on five unpaired datasets, including LIME, VV, NPE, MEF, and DICM. Experimental evaluations show the superiority of our method over SOTAs with better perceptual evaluation in most comparisons. Q6: The color in the last column of Table 2 is incorrectly marked. A6: Thanks for your careful reading. We will rectify this mistake in the revised version and double-check the color of Table 2. --- Rebuttal 2: Comment: ## Dear Reviewer 1bAi Thank you for taking the time to review our submission and providing us with constructive comments and a favorable recommendation. We would like to know if our responses adequately addressed your earlier concerns. Additionally, if you have any further concerns or suggestions, please feel free to let us know. We eagerly await your response and look forward to hearing from you. Thank you for your valuable time and consideration! Best regards, The authors	Summary: The authors proposed a Mamba-inspired method for LLIE, which is designed to address some challenges of the existing method, MambaIR. By integrating GLSSB consisting of LESSM and IRSK, the authors could make the model capture a large local receptive field while preserving global understanding natures. The overall pipeline is explained in detail and motivated by sufficient reasoning. MambaLLIE achieved SOTA performance on several benchmarks, significantly outperforming prior works. Strengths: The overall pipeline is explained in detail and motivated by sufficient reasoning. MambaLLIE achieved SOTA performance on several benchmarks, significantly outperforming prior works. Considering emerging attention on Mamba for low-level vision tasks, the proposed work is timely and can thus attract attention to the community. Weaknesses: The proposed method is somewhat incremental in that it makes small (although relevant) changes to MambaIR. Since this is not the first work that introduces Mamba for LLIE, a more in-depth analysis and exploration is expected; however, the proposed is a slightly better engineered network structure. The authors are recommended to discuss MambaIR in more detail and clarify technical contributions over MambaIR. Technical Quality: 3 Clarity: 3 Questions for Authors: Although the ERF visualization in Figure 1 is impressive, the authors are expected to provide some numeric results for quantitative performance comparisons. It is not clear why it is called "retinex-aware selection" in Figure 2(d). There are several methods dedicated for face detection in low light. It will be interesting to compare face detection performance compared to such methods, Confidence: 5 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: The authors adequately addressed the limitations; but some complexity analysis will be helpful for readers to understand the method. Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Q1: Technical contributions over MambaIR. A1: We want to emphasize that our target and contribution are distinct from MambaIR. 1). Our MambaLLIE is specifically designed for the low-light enhancement (LLIE) task, whereas MambaIR is proposed for image restoration, including image super-resolution (SR) and denoising tasks. Our quantitative and qualitative comparisons indicate that MambaIR exhibits suboptimal performance in the LLIE task. 2). Our aim is to leverage global and local context dependency in the state space module, while MambaIR merely introduces an additional convolution after the vanilla VSSM to restore neighborhood similarity. Our local-enhanced design essentially introduces the local invariance of convolution on state space model, which integrates the directional scan with local-enhanced into VSSM, and propose a novel state space. By additionally convolving 2D feature into directional scan, our method ensures a closer arrangement of relevant local tokens, enhancing the capture of local dependencies. This technique is depicted in Figure 2 (c). Next, we follow the selection mechanism by parameterizing the VSSM parameters based on the input, allowing the model to filter out irrelevant information and remember relevant information indefinitely, as pointed in [8]. Our proposed LESSM brings new insights into how to aggregate local context features in existing VSSM, addressing this gap and achieving significant improvements in LLIE and even SR tasks. Besides, our IRSK is a novel block designed for the LLIE task, its Retinex-aware prior selects the spatial context regions, primarily inspired by the selective kernel mechanism as referenced in [23]. The ablation study indicates that our IRSK shows better performance than the channel attention layer adopted in MambaIR. Q2: About ERF visualization. A2: Thanks for your positive comment and worthy suggestion. Most existing models, including ours, primarily depend on visualization results to understand the receptive field. We further statistics the results as follow \| Method\| Global Mean Brightness \| 2020 Center Region Mean Brightness \| \|--------\|-------------------------\|-------------------------------\| \|SNR-Net\| 97.59 \| 180.11 \| \|RetinexFormer\| 90.72 \| 187.00 \| \|MambaIR\| 94.08 \| 163.99 \| \|MambaLLIE\| 98.10 \| 198.42 \| Q3: It is not clear why it is called "retinex-aware selection".* A3: In our framework, we propose a Retinex-aware kernel selective mechanism, where two coupled Retinex-aware priors are used to select the spatial context regions, each with a different receptive field. Hence, we term it as "Retinex-aware selection" in our framework. According to [13], [43], [45], [C3], illumination is a key prior for low-light enhancement, and the maximum values of RGB images can be regarded as a rough illumination prior. We use it in an implicit manner instead of restoring the illumination map or using a single illumination feature for feature guidance. [C3] Guo X. LIME: A method for low-light image enhancement. In ACM MM. pages 87-91. 2016. Q4: Face detection performance. A4: We investigate the performance of low-light image enhancement methods on face detection in the dark. We use the DARK FACE dataset and randomly sample 300 images for evaluation. The RetinaFace [C4] is used as the face detector and fed with the results of different LLIE methods. We show the results of different methods in Figure R2 and Table R3. In general, MambaLLIE achieves the better mAP score and visual detection result. Please note that the effectiveness of face detection in low-light conditions depends not only on the quality of the enhancement results but also on the specific face detection algorithm employed. In our evaluation, we utilize the pre-trained RetinaFace model to assess the performance of different low-light image enhancement methods to some extent. [C4] Sefik Serengil and Alper Ozpinar. A Benchmark of Facial Recognition Pipelines and Co-Usability Performances of Modules. In Journal of Information Technologies, volume 17, number 2, pages 95-107, 2024. Q5: Complexity Analysis. A5: In the domain of low-light image enhancement, the primary challenges encountered include global color degeneration and local noise disruptions. To address these issues, we have integrated the Mamba architecture into our framework, which excels at capturing the global dependencies of input data to restore. While Mamba-based models have demonstrated commendable performance in tasks such as image super-resolution and segmentation, with linear or near-linear scaling complexity, they tend to yield less than optimal performance in low-light enhancement task. This limitation is largely attributed to the current state space models’ inadequacy in capturing local dependencies, which is crucial for detailed restoration. Moreover, conventional end-to-end models typically fall short in low-light tasks due to their neglect of illumination priors. To overcome these shortcomings, we introduce two innovative components: the local-enhanced state space module (LESSM) and the implicit retinex-aware selective kernel module (IRSK). These modules are specifically designed to effectively model both local and global degradations within our framework.	Summary: This paper introduces a low-light image enhancement module (MambaLLIE). This module has a U-shaped structure and each GLSSB block follows the Transformer-based design. LESSM is proposed to capture the spatial long-term dependency. IRSK is proposed to introduce large and selective kernels for enhancing feature-capturing ability. IRSK also introduces illumination guidance by utilising attention mechanisms. Strengths: 1. The usage of the state space model and large selective kernel greatly increases the effective receptive fields. By capturing longer dependency on features, the network can generate better outputs. 2. MambaLLIE achieves state-of-the-art results on low-light image enhancement tasks across various datasets and evaluation methods. Weaknesses: 1. Could this paper state the difference between LESSM and the Vision State Space Module in [14]? It seems these two blocks have similar designs. 2. The core block GLSSB appears to be a combination of [14] and [23]. However, since [14] is proposed for image restoration tasks, this may limit the novelty of this paper. 3. The descriptions of baseline-1 and baseline-2 are not clear. 4. In Section 4.4 Ablation Study, which dataset is used for the ablation study? How about other datasets? If the dataset used in the ablation study is SDSD-indoor, then the baseline-1 result (PSNR 28.87, SSIM 0.865) is quite similar to [14] (PSNR 28.97, SSIM 0.884), which may indicate that the better performance of this paper relies on Retinex-aware guidance rather than a larger receptive field. However, Retinex-aware guidance has already been explored in [3] and "Low-Light Image Enhancement with Multi-stage Residue Quantization and Brightness-aware Attention." Technical Quality: 3 Clarity: 3 Questions for Authors: 1. Could this paper provide an analysis or explanation of why MambaLLIE has a larger receptive field in terms of global and local? Confidence: 5 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: This paper discusses limitation in Sec 5 and give potential solutions. Flag For Ethics Review: ['No ethics review needed.'] Rating: 4 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Q1: The difference between LESSM and VSSM. A1: Thanks for your comment, we would like to rebut that the MAIN innovation presented in LESSM is the integration of local 2D dependencies into VSSM and formulate a simple yet efficient local-enhanced state space module (LESSM) to aggregate both local and global information. We formulate a novel state space in low light image enhancement community as equation 10, while MambaIR merely introduces an additional convolution after the vanilla VSSM to restore neighborhood similarity. Ours local-enhanced design essentially introduces the local invariance of convolution on state space model, which integrates the directional scan with local-enhanced into VSSM. This technique is depicted in Figure 2 (c). Besides, our local-enhanced term can be regarded as a local consistency constraint for the state space for vision data, guiding the module to learn neighborhood features and thus enhancing robustness. Similar motivation in state space has been discussed in the literature [47], [C1]. [C1] Pfeffermann, Danny, and Richard Tiller. Small-area estimation with state–space models subject to benchmark constraints. Journal of the American Statistical Association 101.476,1387-1397, 2006. Q2: The novelty of GLSSB. A2: We argue that the GLSSB is novel in terms of _the new exploration of the vision state space model_ and _technical improvements for the Retinex-aware low light enhancement task_. The low-light enhancement task faces challenges of global color degeneration and local noise disturbance. Our motivation aims to take into account both global and local image restore. Hence, Our GLSSB mainly follow the _Norm - LESSM - Norm - IRSK_ flow. Firstly, leveraging the long-range dependency modeling of Mamba [8] with linear complexity, we inherited the global modeling capabilities of VSSM, yet we find that existing vision state space model does not pay enough attention on capturing local dependencies. As noted in Q1, our LESSM significantly differs from VSSM [14]. Our IRSK is introduced after LESSM to selectively aggregate Retinex-aware features, which is a structure design based on physical priors for low light enhancement tasks. Specifically, Our IRSK introduces the Retinex-aware prior to select the spatial context regions, primarily inspired by the selective kernel mechanism as referenced in [23]. However, we would like to point that large selective kernel mechanism is not suitable for low-level vision tasks. This is because that the padding operation required for expanding a single large depth-wise kernel generates massive invalid information at the edges of feature maps, which inevitably has a negative impact on image restoration. Additionally, the large selective kernel mechanism uses self-aware spatial kernel selection, which limits the network’s ability to focus on physically prior-aware spatial context regions. Given the global modeling ability of our LESSM, we focus on local selective modeling in our IRSK. Our designed Retinex-aware prior only requires a smaller kernel to learn the spatial features, thereby avoiding excessive padding operations and highlighting the Retinex-aware interested regions. Therefore, we believe our novelty and contribution is on the par with the SOTA methods. Q3: The descriptions of baseline-1 and baseline-2. A3: Baseline-1 directly uses the standard vision state space module (VSSM) to process flattened vision data in our proposed UNet-shaped framework, following the Norm → VSSM → Norm → channel attention layer flow as referenced in [14] for image denoising tasks. Baseline-2 introduces retinex thory $L=R / I$ to Baseline-1, where $I$ denotes the low light image and $R$ denotes the reflectance (enhanced image), $I$ is the illumination map. Baseline-2 aims to estimate the illumination map instead of directly predicting the enhanced image, and then restores the enhanced result by $R=L/I$, as referenced in [13],[43],[45]. Q4: About Ablation Study. A4: Sorry for the confusion caused. The dataset used in the ablation study is SDSD-indoor in previous manuscript. As restated in Q3, Baseline-1 has a similar structure to [14], hence its results (PSNR 28.87, SSIM 0.865) are comparable to those of [14] (PSNR 28.97, SSIM 0.884). Compared with Baseline-1, the improved performance when using our model without IRSK benefits from LSEEM. Our results without LESSM indicate the superior performance of IRSK compared to the vanilla channel attention layer [14]. Meanwhile, we supplement the ablation study with tests on two other benchmark datasets to compare the improvements of LESSM and IRSK. Please refer to the Table R2. Q5: Why MambaLLIE has a larger receptive field in terms of global and local. A5: We would like to answer this question from two perspectives: 1) LESSM. Our proposed LESSM addresses this drawback by introducing a local-enhanced term into the VSSM, which retains the global modeling capability and improves the local receptive field. 2) IRSK. Our IRSK further enlarges the local receptive field through the Retinex-aware selective mechanism. As pointed out in [23], [C2], selective kernels can capture context information by adaptively adjusting their receptive field sizes according to the input. In our MambaLLIE, our LRSK has a comparatively smaller kernel than SKNet and LSKNet, but combining LESSM with LRSK achieves a larger receptive field in terms of both global and local information. As shown in Figure 1 (paper), compared with MambaIR, our MambaLLIE indeed has a larger local receptive field, which is an improvement over VSSM. Additionally, ours have a larger receptive field in terms of both global and local information compared to recent Transformer-based methods. [C2] Xiang Li, Wenhai Wang, Xiaolin Hu, and Jian Yang. Selective Kernel Networks. In CVPR, pages 510-519. IEEE, 2019. --- Rebuttal Comment 1.1: Comment: ## Dear Reviewer ahrr Thank you for taking the time to review our submission and providing us with constructive comments and a favorable recommendation. We would like to know if our responses adequately addressed your earlier concerns. Additionally, if you have any further concerns or suggestions, please feel free to let us know. We eagerly await your response and look forward to hearing from you. Thank you for your valuable time and consideration! Best regards, The authors --- Rebuttal 2: Comment: Thank you for your valuable time and consideration. We would like to clarify the novelty of our method through the following key points: 1. First to Effectively Combine Mamba and Retinex: Our method is the first to successfully integrate Mamba and Retinex theory and propose MambaLLIE for low-light enhancement task, achieving state-of-the-art results. 2. Fundamental Improvement of the State-Space Model in MambaIR: We introduce a novel state-space model, LESSM (Equation 9), specifically designed for low-light enhancement. This represents a fundamental improvement over the State-Space Model of MambaIR, which merely adds a convolution after the vanilla VSSM to restore neighborhood similarity. In contrast, our LESSM is both simple and effective, as demonstrated by its innovative integration of 2D dependencies into the directional scan (Equation 10), ensuring a closer arrangement of relevant local tokens. Experimental results show that our model improves PSNR by an average of 4.43% across all datasets compared to MambaIR, while reducing FLOPs by 65.63% and the parameter count by 46.98%. 3. Retinex-Aware Selective Mechanism (IRSK): Our IRSK mechanism expands the local receptive field through a Retinex-aware selective approach by coupling two illumination maps (Equation 12). This allows for flexible selection of the region of interest with different kernels, as illustrated in Figure 2(d). In contrast, [3] introduces a relatively fixed illumination feature into the attention mechanism, while our selective maps could capture the effective interested regions with the normal and reversed feature and illumination maps. We have discussed the advantages of our selective kernel behavior in *Lines 274-284. 4. Mamba for LLIE: The low-light enhancement task is challenged by global color degradation and local noise disturbance. This led us to develop a LLIE model that fuses a global model and a local model. Since Mamba is a global-aware estimator with linear or near-linear scaling complexity, we chose it as our global baseline. Additionally, the CNN-based model preserves local 2D dependencies, refining local degradation to enhance LESSM and IRSK. Benchmark and real-world experimental evaluations indicate that our method outperforms previous approaches in both qualitative and quantitative comparisons. We thank you again for your feedback and hope that our explanation can make you better understand our contribution and efforts. Best regards, The authors	Rebuttal 1: Rebuttal: We thank all reviewers and chairs for your time, constructive comments, and recognition of our work. We appreciate the positive comments on our idea, contributions, and state-of-the-art performance, such as "novel and interesting," "well illustrated," "attracts attention from the community," and "significantly outperforms prior works." We believe all concerns have been clearly and directly addressed. Here we also want to summarize a few key clarifications concerning the contributions of our work. ## The novelty of GLSSB We argue that the GLSSB is novel in terms of _the new exploration of the vision state space model_ and _technical improvements for the Retinex-aware low light enhancement task_. Our Motivation aims to take into account both global and local image restore. As we know, the low-light enhancement task faces challenges of global color degeneration and local noise disturbance. Global color degeneration suffer from the decrease of illumination, which tends to introduce a global-aware estimator, while the local degradation may contains the large or small regions. Hence, we intuitively introduce Mamba [8] into low light enhancement task, which excels at capturing the global dependencies of input data to restore. Compared with previous CNN and Transformer-based methods, Mamba-based model has exhibit promising performance with linear or near-linear scaling complexity in image super-resolution, image segmentation, etc., but still has the suboptimal results in low light enhancement task. We find that the existing state space model is limited in modeling local dependencies, leading to a failure in restoring details. Besides, end-to-end models may exhibit suboptimal performance in low-light enhancement tasks because they do not consider illumination priors. Based on the two aforementioned assumptions, we propose the local-enhanced state space module (LESSM) and implicit retinex-aware selective kernel module (IRSK), which is the modeling of local and global degradation coupled in our framework for low light enhancement task. Therefore, we believe our novelty and contribution is on the par with the SOTA methods. ## The difference between LESSM and VSSM of MambaIR Our Main Contribution lies in enhancing the local 2D dependencies for vision state space modle (VSSM). Specifically, we aim to formulate a simple yet efficient local-enhanced state space module (LESSM) to aggregate both local and global information by introducing additional local-enhanced term into the scan directions for sequence modeling, compared to existing state space methods, our approach requires only the _small additional number of parameters_ of VSSM to achieve significant performance improvements. The Key insight between LESSM and vanilla VSSM and its variants: most prior VSSMs use the different directional scan in their sequential state, which is necessary for strong performance on visual tasks. However, it has been empirically observed that CNN seem to work fine for 2D dependency of vision data. Ours local-enhanced design essentially introduces the local invariance into state space model, which can integrate the existing directional scan with our local-enhanced term into VSSM. By additional convolving 2D information into directional scan, our method ensures _a closer arrangement of relevant local tokens_, enhancing the capture of local dependencies. This technique is depicted in Figure 2 (c). Next, we follow the selection mechanism by parameterizing the VSSM parameters based on the input, allows the model to filter out irrelevant information and remember relevant information indefinitely, as pointed in [8]. Besides, our local-enhanced term can be regarded as a local consistency constraint for the state space for vision data, guiding the module to learn neighborhood features and thus enhancing robustness, literatures [47], [C1] likewise added constraints of local term to a state–space model as equation 10 and developed the estimate performance. As shown in Figure R1 and Table R1, we demonstrate the effectiveness of the proposed LESSM in MambaIR on image restoration tasks, indicating that our approach also has a positive impact on other low-level vision tasks. While the differences between LESSM and the vanilla VSSM may appear relatively small in description, each has its key motivations and shows significant improvements on common benchmarks over the previous models. We believe that simplicity combined with effectiveness is one of the most favored traits in the field of computer vision. ## Summary As commented in Reviewers DUEm and 1rUY, Considering emerging attention on Mamba for low-level vision tasks, the proposed work is timely and can thus attract attention to the community. Until now, there have been fewer efforts of Mamba dedicated to the low-light image enhancement task. We posit that our contributions can provide a fresh perspectives on VSSM for low-level vision tasks. Our method achieves the SOTA results on low-light image enhancement, even image super-resolution, across various datasets and evaluation methods. Finally, we are willing to supplement the newly added experiments and analysis in the final manuscript/supplementary material. Also, our code and pre-trained model will be released on our project page. [C1] Pfeffermann, Danny, and Richard Tiller. Small-area estimation with state–space models subject to benchmark constraints. Journal of the American Statistical Association 101.476,1387-1397, 2006. Pdf: /pdf/30bda7c88a28d484537ec5fc5e465c921fac96cb.pdf	NeurIPS_2024_submissions_huggingface	2,024	null	null	null	null	null	null	null	null
Practical $0.385$-Approximation for Submodular Maximization Subject to a Cardinality Constraint	Accept (poster)	Summary: The paper studies the problem of approximately maximizing a submodular function under a constraint that the sets can be of size at most $k$. By carefully combining algorithms from previous work and developing them further, the authors obtain an algorithm that guarantees a $0.385$-approximation with $O_\epsilon(n + k^2)$ queries to the function and a user-defined probability of failure. The good performance of the proposed algorithm is demostrated with several experiments. Strengths: Even though the work relies heavily on previous work as described in the beginning of Sec. 3, their combination appears highly nontrivial to me, and the techniques have been developed further from the previous work. The empirical results demonstrate clear improvements over previous state of the art and appear to have significantly less variance in the output. Weaknesses: Because of my limited expertise on the topic, my largest complaint is on the presentation of the results: Although the authors do a good job in explaining the ideas behind the math, the formulas are occasionally relatively cumbersome to read, e.g., Line 14 of Algorithm 1. On the other hand, I must admit that I do not have an immediate solution on how to improve them. Technical Quality: 3 Clarity: 2 Questions for Authors: In terms of the running time, i.e., not the number of queries, how does the proposed algorithm compare against the state of the art? Because the running time depends on, for example, $epsilon$, it is not immediately clear to me how large the constant factors of the algorithms would be. Other: - Abstract: Without previous background on the topic, it was unclear to what $k$ and $n$ referred to in the abstract. - Lines 29–32: first, $B$ is a subset of $A$, but then $A$ is a subset of $B$. This can be confusing to the reader. - Line 66: If I understood the paper correctly, I think it would be more precise to use $O_\epsilon$ instead of the $O$-notation here (although a constant $\epsilon$ is used in the experiments) - Caption of Alg. 3: maximiziation -> maximization - Line 172: OPT is not written with \mathbb Confidence: 2 Soundness: 3 Presentation: 2 Contribution: 3 Limitations: Since the paper focuses on a optimizing a method that is used in ML applications and does not study application directly, I think the limited discussion of limitations is sufficient here. Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We thank the reviewer for the positive evaluation of our work. We have thoroughly addressed each of the reviewer’s raised concerns and are eager to engage in further discussion to ensure all remaining issues are resolved. 1. In terms of the running time, i.e., not the number of queries, how does the proposed algorithm compare against the state of the art? Because the running time depends on, for example, $epsilon$, it is not immediately clear to me how large the constant factors of the algorithms would be. The runtime is a problematic measure as it highly depends on implementation details and the level of optimization employed, often making comparisons based on this method non-robust. Thus, it is customary in the literature about submodular maximization to use the number of queries to the objective function as a more reliable proxy for the real-world performance of an algorithm. Accordingly, we include in the PDF of the general rebuttal a plot comparing the number of queries used by our algorithm and central algorithms from the literature. Interestingly, the number of queries used by our algorithm is empirically dominated by the queries used for the initialization of the Fast Local search procedure (Line 1 of Algorithm 1). Implementing this initialization in the way described in the submitted paper leads to a query complexity that is somewhat worse than that of other algorithms. However, an alternative implementation of this initialization is implied by a very recent paper [1] that describes a fast ¼-approximation algorithm for maximizing non-monotone submodular functions under matroid constraints (of which cardinality constraints are a special case). When initialized this way, our algorithm becomes much faster (in terms of the empirical number of query calls), while the quality of the solution produced in terms of the objective value remains almost unchanged. We stress that the change in the initialization method only affects the implementation of Line 1 of Algorithm 1 in our paper. [1] Balkanski, Eric, Steven DiSilvio, and Alan Kuhnle. "Submodular Maximization in Exactly n Queries" arXiv preprint arXiv:2406.00148 (2024). 2. Typos/Writing suggestions/comments We thank the reviewer for their keen observations. We have addressed your comments and revised the manuscript accordingly. --- Rebuttal Comment 1.1: Comment: Although I partially agree with your criticism of evaluating the running time, I would also argue that it is still one factor to keep in mind. Consider, e.g., fast matrix multiplication, where the asymptotical speedups are mostly of theoretical interest due to the large constant factors hidden away by the O-notation. I'm sufficiently happy with the Rebuttal and will keep my score.	Summary: The paper introduces a new algorithm, FAST-LOCAL-SEARCH, for maximizing non-monotone submodular functions, achieving a 0.385-approximation with low query complexity. It combines initial solution search, accelerated local search, and stochastic greedy improvement steps to outperform existing algorithms in machine-learning applications like movie recommendation, image summarization, and revenue maximization. The algorithm guarantees a constant approximation to the optimal set and is supported by theoretical proofs and empirical evaluations. Strengths: The paper introduces an algorithm, FAST-LOCAL-SEARCH, for maximizing non-monotone submodular functions under a cardinality constraint. This algorithm combines initial solution search, accelerated local search, and stochastic greedy improvement steps to achieve a 0.385-approximation with query complexity of O(n + k^2). The practical performance of the algorithm has been validated in real-world applications. Weaknesses: The novelty of the paper is limited. The idea of this paper is derived from existing work [2], and more significantly, the "guided algorithm" proposed in this paper is very similar to [10] in both algorithm design and theoretical analysis. The authors claim that the algorithm proposed in this paper has a practical query complexity of O(n+k^2); however, in practical applications, the value of k is often very large, making an O(k^2) query complexity potentially impractical. Technical Quality: 2 Clarity: 2 Questions for Authors: Please refer to weaknesses. Confidence: 4 Soundness: 2 Presentation: 2 Contribution: 2 Limitations: N.A. Flag For Ethics Review: ['No ethics review needed.'] Rating: 5 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We would like to thank the reviewer for the detailed review. We hope that our response below addresses all their concerns about the paper. If further clarification is needed, we will be happy to provide it. 1. The novelty of the paper is limited. The idea of this paper is derived from existing work [2], and more significantly, the "guided algorithm" proposed in this paper is very similar to [10] in both algorithm design and theoretical analysis. As mentioned by the reviewer, our algorithm is based on ideas from [2], and its main novelty is in finding a way to implement these ideas efficiently and practically. As explained in Section 1.1, the work of [10] is an independent parallel work that is also based on [2]. Both works come up with a very similar basic algorithm having a query complexity of O(nk). However, the two works diverge beyond this point. The objective of [10] was to derandomize this basic algorithm and extend it to other constraints. In contrast, our objective was to further speed up the algorithm beyond the above-mentioned query complexity. This additional speed-up is an important technical contribution of our work. For example, getting the local search part of the algorithm of [2] to run fast (as implemented in Algorithm 1) required us to relax the notion of local optimum used and introduce a sophisticated probabilistic analysis deviating from the well-threaded path of analysis of local search algorithms in submodular optimization. 2. The authors claim that the algorithm proposed in this paper has a practical query complexity of O(n+k^2); however, in practical applications, the value of k is often very large, making an O(k^2) query complexity potentially impractical. As mentioned by the reviewer, the query complexity of our algorithm is O(n + k^2), which is linear for k = O(n^{½}). For applications that require larger values of k, our algorithm indeed runs in super-linear time, which is unfortunate. Still, one should note two things: - The PDF of the general rebuttal includes a plot comparing the empirical query complexity of our algorithm with various existing algorithms, including linear query complexity algorithms like the one of [5]. This plot demonstrates that our algorithm compares very well with these algorithms in terms of the empirical query complexity (while outcompeting them in terms of the value of the solution produced). - The only other algorithm in the literature that guarantees an approximation ratio better than 1/e and has a possibly practical query complexity is the algorithm due to the recent independent work of [10]. This algorithm has a query complexity of $O(nk)$, which is worse than the query complexity of our algorithm except in the regime of k = $\Omega(n)$ (and in this regime the query complexities of both algorithms become identical). --- Rebuttal 2: Comment: Thank you for your response. I decide to increase my score.	Summary: This work addresses the problem of maximizing a non-monotone submodular function subject to a cardinality constraint. In submodular maximization, we have a set of elements (ground set) and a function that assigns a value to any subset of elements. A function is submodular if adding an element to a smaller set contributes more value than adding it to a larger set. Formally, for sets A and B where A is a subset of B, adding an element x to A increases the function value more than adding x to B. A function is monotone if its value always increases with set size. Non-monotone functions don't necessarily have this property. The goal here is to find a subset of size at most k from the ground set that maximizes the value of the non-monotone submodular function. While they developed a 0.385-approximate solution for this problem, there is an algorithm with a better approximation factor of 0.401. Another work also achieves a 0.385-approximate solution, but both are impractical due to the high number of query calls required. In contrast, this work uses only O(n+k^2) query calls. Additionally, there are two works with a 0.367-approximation, one requiring O(nk) query calls and the other O(n) query calls. In the experiment section, the authors compared the output value of their algorithm with those of the two latter algorithms. The core idea of the proposed algorithm involves: - Running the sample greedy algorithm by Buchbinder et al. [5]. - Applying a local search technique to improve the initial solution. - Employing a Stochastic greedy method and returning the solution with the highest value between the local search and Stochastic greedy outputs. This approach leverages existing techniques while incorporating additional steps to enhance the solution within a reasonable computational cost. Strengths: For non-monotone submodular maximization under a cardinality constraint, their work achieves the best query complexity for algorithms that attain at least a 0.385-approximate solution. In other words, among practical algorithms—those that can be implemented and executed in a reasonable time—they offer the highest approximation factor. Weaknesses: Since their algorithm does not have the best approximation factor, it is less interesting for theoretical work and more important for practical applications. Therefore, a comprehensive experiment supporting their claims is crucial. While they did a great job by running experiments on different problems and demonstrating the advantage of their output value compared to two other algorithms with worse approximation guarantees, it is also necessary to plot the number of query calls for these algorithms. Comparing the number of query calls in addition to the output value is essential, as they claimed to have the best approximation algorithm with "a low and practical" query complexity. ### Comments for the authors: - Line 43: Remove “over” - Line 50: I think [14] present a ½-approximation algorithm for symmetric submodular functions and for non-symmetric, they only achieve a ⅖-approximation algorithm. - Line 68: Do you mean “3 applications”? - Line 149: Remove “)” after “pn” - Line 208: Replace “she is” with “they are” - Line 548: Replace “lemmata” with “lemmas” The revenue maximization experiment is mostly known as the max-cut problem. I think it is good to at least mention that. - In Algorithm 3, the term “AIDED-MEASURED DISCRETE STOCHASTIC GREEDY” has been used to refer to Algorithm 2. However, the name of Algorithm 2 is different. - From lines 95 and 96, I assumed that Algorithm 2 improves the output of Algorithm 1. However, after reading the algorithms, it seems that Algorithm 2 finds a different Technical Quality: 2 Clarity: 4 Questions for Authors: - What is the query complexity of your algorithm in terms of epsilon? - Do you have any practical comparisons of the number of query calls used in your experiments? Confidence: 4 Soundness: 2 Presentation: 4 Contribution: 2 Limitations: N/A Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We are grateful for the reviewer’s positive evaluation of our work. In what follows, we address the concerns raised by the reviewer in detail. We will be happy to engage to address any lingering concerns. 1. Comparing the number of query calls in addition to the output value is essential, as they claimed to have the best approximation algorithm with "a low and practical" query complexity. \| Do you have any practical comparisons of the number of query calls used in your experiments? We thank the reviewer for the insightful suggestion. We have generated graphs depicting the number of query calls, comparing the efficiency of our method with that of our competitors. These graphs can be found in the PDF file of the general rebuttal. Interestingly, the number of queries used by our algorithm is empirically dominated by the queries used for the initialization of the Fast Local search procedure (Line 1 of Algorithm 1). Implementing this initialization in the way described in the submitted paper leads to a query complexity that is somewhat worse than that of other algorithms. However, an alternative implementation of this initialization is implied by a very recent paper [1] that describes a fast ¼-approximation algorithm for maximizing non-monotone submodular functions under matroid constraints (of which cardinality constraints are a special case). When initialized this way, our algorithm becomes much faster (in terms of the empirical number of query calls), while the quality of the solution produced in terms of the objective value remains almost unchanged. We stress that the change in the initialization method only affects the implementation of Line 1 of Algorithm 1 in our paper. [1] Balkanski, Eric, Steven DiSilvio, and Alan Kuhnle. "Submodular Maximization in Exactly n Queries" arXiv preprint arXiv:2406.00148 (2024). 2. What is the query complexity of your algorithm in terms of epsilon? The query complexity of our algorithm is $O\left(n * \epsilon^{-2} \log\left(\frac{1}{\varepsilon}\right) + k^2 * \varepsilon^{-1} \log\frac{1}{\varepsilon}\right)$, excluding the query complexity required for obtaining the initial solution $S_0$ for the Fast Local search procedure. The query complexity required for getting this initial solution is independent of epsilon when the algorithm of [1] is used for that purpose, as described in the response to the previous question. 3. From lines 95 and 96, I assumed that Algorithm 2 improves the output of Algorithm 1. However, after reading the algorithms, it seems that Algorithm 2 finds a different Intuitively, Algorithms 1 and 2 balance each other. In a sense, Algorithm 2 uses the output of Algorithm 1 as a “set to avoid”, and its analysis shows that when the output of Algorithm 1 has a poor value, then by avoiding it Algorithm 2 is guaranteed to output a set of good value. Thus, it is guaranteed that the better of the outputs of the two algorithms is always a good solution. 4. Typos/writing suggestions/comments We have addressed all of the raised comments and suggestions concerning the writing. We are thankful for the keen observation of the reviewer and the fruitful suggestions. --- Rebuttal Comment 1.1: Comment: Thank you for your response. Since my questions have been fully answered and the plots comparing oracle calls have been provided, I’ve increased my score to 6.	Summary: The paper presents a practical $0.385$-approximation algorithm using $O(n+k^2)$ queries for non-monotone submodular maximization under a cardinality (size) constraint, where $n$ is the number of elements in the ground set and $k$ is the maximum size of a feasible solution. As a comparison, the state-of-the-art algorithms attain a $0.401$ approximation ratio with a high number of queries, or $1/e-\epsilon$ approximation ratio using $O_{\epsilon}(n)$ queries. The paper also evaluates its method experimentally on several applications. Strengths: 1. The paper makes a substantial progress for the studied problem by presenting a new algorithm with low query complexity and improved approximation ratio. The algorithm is obtained by carefully combining several well-known techniques and methods in the field. Such a result already contains enough originality as a NeurIPS submission. 2. The paper is complete, in the sense that it contains both a correct theoretical analysis and detailed experimental evaluations of its algorithm. 3. The paper is organized well. It splits the algorithm into several ingredients and then explains each part in detail and how to combine them to get the final result. In such a way, it’s easy for readers to follow the basic ideas and check the correctness of the algorithm. 4. The paper, as suggested by the experiments, provides a better solution for several important applications. On the other hand, it further exploits several mature techniques in the field, which may inspire more results in the area. Weaknesses: 1. There are many typos in the proofs (see Questions for some examples), which affects the clarity. The authors are suggested to proofread their paper in a later version. 2. The main algorithmic framework is borrowed from [2]. So, if the space is enough, I suggest the authors to add a paragraph discussing why the framework can beat the $1/e$ ratio. In doing so, the paper might be more readable to those who haven’t read [2] before. Technical Quality: 4 Clarity: 3 Questions for Authors: 1. I’m curious about if it’s possible to discretize the algorithm in [3] to get a practical $0.401$ algorithm, just like this paper. 2. Following are some typos: 1) I suggest using “at most” to replace “at max” in the paper. 2) Page 13. “Let S be a subset in N of size at (missing most) k”. 3) Page 16, Lemma A.3, $x_{\in}$ --> $x\in$. 4) Page 16, in the last inequality, \beta^{i} should be \beta_{i}. 5) Page 17, in the first formula, A\cup B_{\lambda}\cap Z should be A\cup B_{\lambda} ‘\cup’ Z. 6) Page 18, Lemma B.1, “that starts returns a set…” doesn’t read smoothly. 7) Page 18, in the first sentence of the proof of Lemma B.1, S_{i,j} should be S_i^j. 8) Page 20, in inequality (6), the subscript L should be i. Confidence: 5 Soundness: 4 Presentation: 3 Contribution: 3 Limitations: Yes Flag For Ethics Review: ['No ethics review needed.'] Rating: 7 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We deeply appreciate the reviewer's meticulous evaluation of our paper. Thank you for your invaluable feedback. In what follows, we address the reviewer’s questions/comments. We will be happy to engage with you to address any lingering concerns 1. There are many typos in the proofs (see Questions for some examples), which affects the clarity. The authors are suggested to proofread their paper in a later version. We have addressed the typos raised in the Questions section and will thoroughly pass over the paper to fix any additional types. We thank you for your keen observation. 2. The main algorithmic framework is borrowed from [2]. So, if the space is enough, I suggest the authors to add a paragraph discussing why the framework can beat the $\frac{1}{e}$ ratio. In doing so, the paper might be more readable to those who haven’t read [2] before. We thank the reviewer for the fruitful suggestion. We will be more than happy to add such a paragraph to the paper. 3. I’m curious about if it’s possible to discretize the algorithm in [3] to get a practical $0.401$ algorithm, just like this paper. Making the recent 0.401-approximation of [3] practical is an interesting question that we have also considered. Unfortunately, this goal seems to be very challenging and requires significant new ideas because the algorithm of [3] has a query complexity that is exponential in 1/epsilon. Notice that this complexity remains exponential even if we assume access to the multilinear extension of the objective function rather than requiring the algorithm to use samples to estimate it. Making the algorithm combinatorial, as we have done in the current paper, can make this assumption justified, but it does not help to alleviate the inherent exponentiality of the query complexity of the other parts of the algorithm. --- Rebuttal Comment 1.1: Comment: Thank you for your response. I will keep my score.	Rebuttal 1: Rebuttal: We sincerely appreciate the reviewers for their constructive feedback and insightful questions about our research. Your dedication and expertise in helping us enhance our work is invaluable. We are grateful for your observations and thank you for your contributions. Finally, attached here is our PDF containing all the figures depicting our algorithm with a different initialization in comparison with our algorithm with the original initialization discussed in the appendix and also in comparison with the presented competitors in the paper. We present in these figures a comparison with respect to two metrics: - Objective value, and - Oracle calls The initialization we are referring to is from [1] and is referred to as "Algorithm 1 with different initialization" in the uploaded graphs. [1] Balkanski, Eric, Steven DiSilvio, and Alan Kuhnle. "Submodular Maximization in Exactly n Queries" arXiv preprint arXiv:2406.00148 (2024). Pdf: /pdf/9ef28f0ffd847561620ada8316909ff2851c36e6.pdf	NeurIPS_2024_submissions_huggingface	2,024	null	null	null	null	null	null	null	null
Towards the Transferability of Rewards Recovered via Regularized Inverse Reinforcement Learning	Accept (poster)	Summary: This paper analyses the problem of the transferability of the reward functions learned from expert demonstrations in regularized MDPs. After having formalized the problem, the authors introduce some assumptions that permit to connect the distance between equivalence classes of reward functions to the suboptimality of the expert. Next, they use the notion of principal angles to relate the distance between transition models with the "amount of information" provided by demonstrations in such settings. Finally, authors propose an algorithm with theoretical guarantees for solving the problem, which they validate empirically. Strengths: - The paper is written in a clear manner. - The goal is to bring theoretical results (i.e., about known expert's policy) to the real world, by analysing the setting with demonstrations. - The transferability issue is one of the peculiar advantages of IRL over, for instance, BC, thus it is interesting. - The contributions are based on technically solid theoretical results (e.g., authors provide sample complexity analysis of the algorithm as well as convergence analysis). Weaknesses: - The framework of regularized MDPs makes things much easier w.r.t. the setting of common MDPs. - There are various assumptions along the way (which seem necessary). - Regularized MDPs do not represent a realistic model for expert behavior. Technical Quality: 4 Clarity: 4 Questions for Authors: questions: - Assumption 2.2 is rather strong, because you assume that the transition model is such that, for any arbitrary policy, each state is visited with a fixed non-zero probability. Do you agree? - As you comment in Remark 3.12, it seems clear that all the results depend a lot in how different the considered regularized MDP differs from a common MDP. Since common MDPs are more realistic models to model experts behavior, how limited do you think that your results are when applied to practical applications? minors: - the references should be made aligned with each other, and not like now where some have DOI code, others do not, some have volume, others number of pages, etc... - the checklist shall be put at the end of the appendix typos: - line 130 the operation of difference between equivalence classes is not defined - line 140 it is not clear what is the difference between $N^E$ (which is not defined) and $K$ - line 245 the simplex symbol has $\mathcal{S}$ and $\mathcal{S}\times\mathcal{A}$ reversed. Confidence: 5 Soundness: 4 Presentation: 4 Contribution: 3 Limitations: The authors have adequately addressed the limitations and potential negative societal impact of their work. Flag For Ethics Review: ['No ethics review needed.'] Rating: 7 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thank you for your feedback, valuable comments, and helpful suggestions. While we addressed some of your concerns in the response to all reviewers, we would like to respond in detail to your comments below. ## Our comments: - __(C)__ "Regularized MDPs do not represent a realistic model for expert behavior." __(A)__ As noted in the response to all reviewers, entropy regularization is a widely recognized model of bounded rationality in social choice theory and behavioral game theory. Additionally, the assumption of perfect expert optimality can be relaxed. Nevertheless, we agree that further research is needed to determine suitable models of expert behavior for specific applications and we will discuss this in the limitations and feature work section. - __(C)__ "(...) common MDPs are more realistic models to model experts behavior, (...)" __(A)__ We believe that assuming perfect optimality in standard MDPs is also quite limiting. Humans are not perfectly rational. For example, in a bandit problem, if an expert chooses arm $a_1$ 99% of the time and arm $a_2$ 1% of the time, we would expect $r(a_1) > r(a_2)$. Yet, the only rewards consistent with these choices would be $r(a_1) = r(a_2)$. Similar but more subtle situations could arise in MDPs with multiple states. Of course, this issue can again be alleviated by relaxing the assumption of perfect expert optimality, which, however, leads to even less identifiable rewards. - __(C)__ "As you comment in Remark 3.12, it seems clear that all the results depend a lot in how different the considered regularized MDP differs from a common MDP. (...) how limited do you think that your results are when applied to practical applications?" __(A)__ 1) As mentioned in the response to all reviewers, our results still apply if the expert acts approximately optimally with respect to the regularized MDP. 2) Our proof technique based on principal angles between subspaces (Lemma F.1) can be applied whenever the reward can be approximately identified up to potential shaping. In unregularized MDPs, this may be achieved, e.g., by collecting preferences between different policies. 3) For standard IRL without a regularizer, the set of rewards for which all experts are optimal is no longer an intersection of subspaces but an intersection of convex cones (see, e.g., [1, 2]). We believe some of our tools, especially principal angles, would still be useful in this context, but without additional assumptions we expect it to be difficult to recover a transferable reward in this setting. - __(C)__ "minors: the references should be made aligned with each other, and not like now where some have DOI code, others do not, some have volume, others number of pages, etc... the checklist shall be put at the end of the appendix (...) typos: (...)" __(A)__ Thank you very much for pointing these out. We will address it in the camera-ready version. ## References [1] Metelli, Alberto Maria, et al. "Provably efficient learning of transferable rewards." _International Conference on Machine Learning_. PMLR, 2021. [2] Schlaginhaufen, Andreas, and Maryam Kamgarpour. "Identifiability and generalizability in constrained inverse reinforcement learning." _International Conference on Machine Learning_. PMLR, 2023. --- Rebuttal Comment 1.1: Title: Follow-up Comment: Apologies, we realized that we missed addressing the following question: (C) "_Assumption 2.2 is rather strong because you assume that the transition model is such that, for any arbitrary policy, each state is visited with a fixed non-zero probability. Do you agree?_" (A) Yes, that's correct. The assumption could be ensured, for instance, by a lower bound on the initial state distribution. Moreover, in our analysis, we only require it to hold for the optimal policies under the experts' and the recovered reward. However, it would be interesting to explore whether this assumption could be replaced with weaker notions of coverability or concentrability.	Summary: The paper establishes bounds on the reward transferability for inverse reinforcement learning in discrete MDPs with known transition matrix. It is well known that the reward function learned in a single environment is in general not transferable as the reward shaping potential can not be identified and therefore, the recovered reward function may perform poorly under different dynamics. Prior work showed conditions for recovering the true reward function (up to a constant) when observing multiple experts acting in MDPs with different dynamics. Probably most relevant, Rolland et al. (2022) showed that reward identifiability is ensured under some rank conditions on a matrix constructed based on the transition matrixes. However, their analysis assumes that the IRL problem can be accurately solved in the different environments. * The submission extends these previous results by allowing for an error in the recovery of the reward function. Namely, the paper derives a bound based on the principal angles between the subspaces of potential shaping transformations corresponding to the different transformation matrices of the experts. That is, the second largest principle angle between the two observed MDPs, in combination with the optimality gap of the recovered reward function, is used to bound the optimality gap of the optimal policy (wrt the recovered reward function) under any other transition matrix. * Similarly, the paper uses the principle angle to establish a bound on the optimality gap on a given environment, when the reward was recovered in a different environment. * The theoretical results are accompanied with empirical results on a gridworld environment that show the relation between the principle angle and the error of the recovered reward function. The results use projected gradient descent to optimize the reward function based on demonstration under different system dynamics. Strengths: Originality Using prinicpal angles for bounding reward transferability seems to be a novel and useful concept. The previous rank conditions only provided a binary criterium usable for perfect conditions, whereas the principal angles enable us to relate approximation errors to the transferability of the recovered reward. Significance The theoretical results could be a step towards better understanding reward transferability also for more realistic settings (e.g. non-finite MDPs, unknown transition matrices). Reward transferability is an important open and critical question in inverse reinforcement learning. Indeed, arguably there is little additional value in a non-transferable reward function compared to an imitation learning policy. Clarity Overall, I find the paper well-written and easy to follow. The paper gives a good overview of the work and the theoretical results and assumptions are clear. I also like the geometrical proof Fig 1b a lot. In general the paper avoids unnecessary complications. Quality The assumptions (e.g. steepness & convexity of the regularizer and Lipschitz continuity of its gradient, full support for every state-occupancy measure) are reasonable. The claims are well-supported and seem correct (I did not check all derivations in detail). Weaknesses: Significance The paper builds on the concept of principle angles between different subspaces, which can be only computed in simple settings (finite MDPs with known transition functions). While I can imagine that concepts could be transferred to continuous MDPs and to settings where approximations based on transition samples are necessary, it is not fully clear if and how such extension could be accomplished. Clarity While overall the paper is clear, I think that the paper should provide some intution about its most relevant concepts. Namely, I am missing an intuition about the subspaces of potential shaping transformations $\text{Im}(E - \gamma T^\top)$, which in turn would be useful for getting an intuition about the principal angles of interest. Quality The paper mainly considers the setting when learning with demonstrations from K=2 different transition dynamics. The paper proposes a trivial extension of the bound for K>2, that merely uses the tightest bound among all combinations of two different dynamics. However, I think it can be safely assumed that much better transferability can be guaranteed when learning in K>2 enviroments, as these could complement each other. The bounds are quite loose as evident from Corollary 3.11 in the case of Shannon- and Tsallis-entropy regularization. In particular the term $\exp(H_\gamma)$ should easily explode for the Shannon-Entropy, and similarly the term $H_\gamma^\top$ for the Tsallis-entropy, where the term $H_\gamma$ can easily be in the range of hundreds or thousands for reasonable discount factors. However, I give credit for the fact that the paper mentions this limitation, and at least the large bounds can be somewhat compensated by very small approximation errors (also recovering previous results for $\hat{\epsilon}=0$). Technical Quality: 3 Clarity: 3 Questions for Authors: - Line 133 states that the paper considers equivalence relations induced by constant shifts and the "aforementioned potential shaping transformations". Do these include all general potential shaping transformations? - Could we use the principle angle to design experiments in order to learn transferable reward functions? - Any ideas on how the angles could transfer to continuous MDPs? - How could we estimate them from transition samples? - Could the angles be generalized for K>2? Confidence: 4 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: Limitations and potential negative societal impact have been adequately addressed. Flag For Ethics Review: ['No ethics review needed.'] Rating: 8 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thank you for your feedback, valuable comments, and helpful suggestions. While we addressed some of your concerns in the response to all reviewers, we would like to respond in detail to your comments below. ## Our comments: - __(C)__ "While I can imagine that concepts could be transferred to continuous MDPs and to settings where approximations based on transition samples are necessary, it is not fully clear if and how such extension could be accomplished." __(A)__ This is a good point. We will extend the corresponding discussion in the limitations section (see response to all reviewers). In short, we expect our results to translate to the continuous setting, but dealing with infinite-dimensional reward and measure spaces will introduce several technical subtleties and explicit constants could only be computed for special cases. - __(C)__ "While overall the paper is clear, I think that the paper should provide some intuition about its most relevant concepts. Namely, I am missing an intuition about the subspaces of potential shaping transformations $\text{im}(E-\gamma P)$, which in turn would be useful for getting an intuition about the principal angles of interest." __(A)__ Thanks for pointing this out. We will try to address this in the camera-ready version (see response to all reviewers). As for the subspace of potential shaping transformations $U_P =\text{im}(E-\gamma P)$, it lies perpendicular to the set of feasible occupancy measures $\mathcal{M}_P$ (see also [1] for more details). We will clarify this in the background section (see response to all reviewers). - __(C)__ "(...) I think it can be safely assumed that much better transferability can be guaranteed when learning in K>2 enviroments, as these could complement each other." __(A)__ Yes, that's correct. We'll add it to the limitations and future work section (see response to all reviewers). ## Answers to Questions: - __(C)__ "Line 133 states that the paper considers equivalence relations induced by constant shifts and the "aforementioned potential shaping transformations". Do these include all general potential shaping transformations?" __(A)__ Yes, all potential shaping transformations. We will drop the "aforementioned" for clarity. - __(C)__ "Could we use the principle angle to design experiments to learn transferable reward functions?" __(A)__ Yes indeed, developing an online algorithm based on principal angles could be an interesting direction for future research. - __(C)__ "Any ideas on how the angles could transfer to continuous MDPs?" __(A)__ As mentioned in response to all reviewers, principal angles could be extended to the continuous setting. However, the resulting potential shaping spaces would be infinite-dimensional. We could then either resort to the theory of angles between infinite-dimensional subspaces [2], or consider a finite-dimensional reward class such as e.g. in linear quadratic regulators or linear MDPs. - __(C)__ "How could we estimate them from transition samples?" __(A)__ The principal angles can be computed via a singular value decomposition. As shown in [3], the error $\|\sin \theta_k(P^0, P^1)-\sin \theta_k(\hat{P}^0, \hat{P}^1)\|$ would scale as $\mathcal{O}(\max_k\|\|P^k- \hat{P}^k\|\|)$ if $\hat{P}^k$ are estimates of $P^k, k=0,1$. We will add a brief section about this to the Appendix. Thanks for the question. - __(C)__ "Could the angles be generalized for K>2?" __(A)__ Yes, we believe so (see response to all reviewers). ## References [1] Schlaginhaufen, Andreas, and Maryam Kamgarpour. "Identifiability and generalizability in constrained inverse reinforcement learning." _International Conference on Machine Learning_. PMLR, 2023. [2] Zhu, Peizhen, and Andrew V. Knyazev. "Angles between subspaces and their tangents." _Journal of Numerical Mathematics_ 21.4, 2013. [3] Ji-Guang, Sun. "Perturbation of angles between linear subspaces." _Journal of Computational Mathematics_ (1987): 58-61. --- Rebuttal Comment 1.1: Comment: Thank you for the reply. I will maintain my (positive) score also after reading the other reviews. I see good contributions, and none of the mentioned weaknesses seem particularly critical.	Summary: This paper introduces the logical framework regarding rewards identification (up to potential shaping transformations, which can be achieved under entropy regularization) and rewards transferability (up to a constant, more difficult to achieve). The paper considers a practical scenario where only expert demonstrations are accessible, not the full expert policy, and analyzes $\epsilon$-transferability using the principal angles technique. Additionally, when only expert demonstrations are available, they analyze two cases: for two experts, transferring to an arbitrary transition law requires the two laws (from two experts) to be sufficiently different; for a single expert, transferring from the original to the target environment requires the transition laws to be closely related. Finally, they present a probably approximately correct (PAC) algorithm and an end-to-end analysis for learning transferable rewards from multiple expert demonstrations. Strengths: 1. I have particularly appreciated the logical framework regarding rewards identification (up to potential shaping transformations, which can be achieved under entropy regularization) and rewards transferability (up to a constant). I also value the introduction of two methods for attaining rewards transferability: first, by restricting the reward class, such as to state-only rewards, and second, by learning from multiple experts who share the same reward but have different transition laws, provided a specific rank condition is met. 2. The topic selection is meaningful, exemplifying that under the approximate optimality of experts, the learned reward can perform poorly in a new environment, even if the rank condition in Equation (5) is satisfied. 3. They consider a practical scenario where only expert demonstrations are accessible, not the full expert policy, and analyze $\epsilon$-transferability using the principal angles technique. This result encompasses the scenario where the full expert policy is accessible, recovering a reward r^ for which all experts are exactly optimal, with ϵ being 0 for all experts. 4. The comparison between single-expert transfer and multi-expert transfer is insightful. Weaknesses: 1. This work is mainly theoretical. I think experimental validation on real-world applications could provide valuable insight into the practical aspects and challenges of transferability (if possible). 2. I think the example used to describe the rank condition's insufficiency for reward transfer when $\epsilon$ is not exactly 0 (for all experts) is not effective. As $\beta$ tends to 0, the two transition laws become completely different, which is too extreme, despite being easy to follow and aligned with the main theorem. Maybe a toy demo with illustrations is better. 3. When $\epsilon$ is exactly 0 (recover a reward $\hat{r}$ for which all experts are exactly optimal), the work is already established, making the theoretical contribution slightly inferior. I believe a detailed discussion of this work's theoretical innovations and how it differs from the case when the expert policy can be achieved is necessary (from a theory aspect). I will be inclined to increase my score if I appreciate your response. Technical Quality: 3 Clarity: 4 Questions for Authors: 1. As the authors mentioned in Section 2, the reward in $\mathcal{U}$ and $\mathcal{V}$ is defined on $\mathcal{S} \times \mathcal{A}$. According to Definition 3.7 (lines 237-239), how are the elements in $\mathcal{U}_{P}, \mathcal{U}_{P'} \in \mathbb{R}^{\mathcal{S} \times \mathcal{A}}$ processed for inner product and modulus operations (to be a constant)? 2. $\\|[r]_{\mathcal{V}}-[r']_{\mathcal{V}}\\|_{2}$ is small. How to define 'minus' and 'small' needs clarification, as I believe this involves the difference between two sets or two equivalence classes. A detailed explanation would be better. 3. The description of the lower and upper bounds in (6) is unsatisfactory. For the recovered reward, we need to consider both the lower bound (not optimal) and the upper bound (approximately optimal). Therefore, both the lower and upper bounds should be included. 4. I think the implications of Theorem 3.10 should be introduced separately according to single expert and multi experts due to the different meanings of transition law differences. For multiple-experts, the difference in transition laws pertains to the source environments. For single experts, it pertains to the difference between the source environment and the target environment. Confidence: 4 Soundness: 3 Presentation: 4 Contribution: 3 Limitations: The authors have discussed the limitations and future work in Section 6. Flag For Ethics Review: ['No ethics review needed.'] Rating: 5 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thank you for your feedback, valuable comments, and helpful suggestions. While we addressed some of your concerns in the response to all reviewers, we would like to respond in detail to your comments below. ## Our comments: - __(C)__ "(...) I think experimental validation on real-world applications could provide valuable insight into the practical aspects and challenges of transferability (if possible)" __(A)__ This is a valid point. However, as our theoretical results apply to tabular settings only, we decided to experimentally validate them in a grid world environment and to leave extensions to more realistic continuous environments for future work. - __(C)__ "I think the example used to describe the rank condition's insufficiency for reward transfer when 𝜖 is not exactly 0 (for all experts) is not effective. As 𝛽 tends to 0, the two transition laws become completely different, which is too extreme, despite being easy to follow and aligned with the main theorem. Maybe a toy demo with illustrations is better." __(A)__ We intentionally picked an example where $P^1$ doesn't tend to $P^0$ as $\beta\to 0$. The example shows that even if $\|\|P^0 - P^1\|\|$ is large, and the rank condition is satisfied, the recovered reward may fail to be transferable to a new transition law. We could have picked $P^1(0\|s,a) = 1-\beta·\mathbb{1} \lbrace s = 0, a = 0\rbrace$, but then one might argue that the lack of transferability for small $\beta$ is due to $P^0\approx P^1$ and not due to the second principal angle being small. Alternatively, if you prefer, we could pick something intermediate like $P^0(0\|s,a) = 0.75$ and $P^1(0\|s,a) = 0.25+\beta·\mathbb{1} \lbrace s = 0, a = 0\rbrace$ and the example would still work out. Furthermore, for a toy demo we refer to the experiments, where we see that also for a gridworld example the rank condition alone, which is equivalent to $\theta_2(P^0, P^1)>0$, does not suffice to ensure effective transferability to new environments. Is this sufficiently addressing your concerns? - __(C)__ "I believe a detailed discussion of this work's theoretical innovations and how it differs from the case when the expert policy can be achieved is necessary (from a theory aspect)." __(A)__ We dedicated Section 3.2 and Example 3.2 to the discussion of related work in the exact setting. Example 3.2 shows that the previously established binary rank condition is meaningless when we cannot identify a reward for which the expert is perfectly optimal. ## Answers to Questions: 1. Thanks for pointing this out, we will clarify this more explicitly in Section 2. In lines 237-239, $\mathcal{V}, \mathcal{W}$ are subspaces of $\mathbb{R}^n$. As mentioned in the notation section, $\langle \cdot, \cdot\rangle$ denotes the standard inner product $\langle x, y \rangle = \sum\_{i=1}^n x\_i y\_i$ in $\mathbb{R}^n$. Moreover, the inner product on $\mathbb{R}^{\mathcal{S}\times\mathcal{A}}$ is defined analogously by $\langle r, r' \rangle = \sum_{s,a} r(s,a)r'(s,a)$. 2. Thanks for pointing this out. We clarified it the response to all reviewers. The quotient space is itself a vector space. Hence, addition and multiplication are well-defined. Moreover, we introduced the norm on $\mathbb{R}^{\mathcal{S} \times \mathcal{A}}/\mathcal{V}$ in line 129. 3. Would you mind to elaborate more on this question? We're not sure whether we understand it correctly. Equation (6) contains a lower and an upper bound, and both are discussed before and after the statement of the result. 4. We only stated Theorem 3.10 for a single expert. The idea is that when learning from only a single expert, transferability to arbitrary transition laws is impossible even in the exact setting. However, Theorem 3.10 shows that we are at least transferable to target environments that are close to the source environment in terms of the maximal principal angle. In contrast, Theorem 3.9 shows that when we are learning from two experts and the second principal angle between the two source environments is large enough, then we can transfer to any target environment. Does this sufficiently clarify your question? --- Rebuttal Comment 1.1: Comment: Thank the authors for the rebuttal. I will maintain my original score.	Summary: In the context of inverse reinforcement learning (IRL) -- i.e. the task of recovering a reward based on demonstrations from experts acting (approximately) optimally with respect to that reward -- this paper studies the question of transferability to new environment dynamics: if we learn a reward function from demonstrations on environments with dynamics P1 and P2, when can we expect for this reward to lead to the optimal policy on a third environment with dynamics P3? The authors show that previous results guaranteeing transferability do not generalize to cases where we are able to learn the reward only approximately, which would be the case in most practical settings. They then provide two theoretical results giving sufficient conditions for transferability. Strengths: - The broader question the paper addresses (transferability in IRL) is a key question for the practical usefulness of IRL. IRL papers often claim that the recovered reward is a more generalizable representation of the goal than a policy (that could be recovered e.g. using behavioural cloning, which is usually cheaper). This paper makes a step toward putting such claims on firmer theoretical basis. - The paper clearly shows important limits of prior results in the area. - The authors prove novel theoretical results using an original method. - Once one grasps the necessary formalisms, the rest of the paper has a clear and easy-to-follow structure. - The writing is largely free of typos and other mistakes. Weaknesses: Only major objection is the lack of pedagogical effort: I think the paper is not optimized to be read by a broader NeurIPS audience. I think it would be easy to read to someone who is (1) an expert on IRL (already a somewhat niche area) and at the same time (2) an expert on the kind of theoretical analysis used in the paper, which leaves only handful of people (I think an audience not big enough to even hold a workshop). Even someone from category (1) with a degree in math and some familiarity with similar theoretical results, I would really appreciate if authors put much more effort into explaining what are the intuitions behind the different abstract concepts that they use. E.g. I think most people active in IRL and RL have much stronger existing intuitions in thinking about policies, rather than occupancy measures, so a slight build-up toward defining stuff in terms of measures would help them. Similarly when introducing a "strictly convex regularizer" it may be useful to give an intuition what is the purpose of that regularizer .Often adding a single sentence in places like this would make a huge difference to many readers. Also defining the transition operator as $(Pv)(s,a) = \sum_{s'} P(s'\|s,a)v(s')$ is about as unintuitive as it gets for an otherwise so intuitive concept. Technical Quality: 4 Clarity: 2 Questions for Authors: I would appreciate more discussion on how you would expect similar results to generalize to continuous state and action spaces - you title the paper "Towards the Transferability ..." so it would be good to have more discussion of whether and how this simple finite linear case is a useful step toward establishing more practically useful results. This is key for assessing the importance of this paper. Confidence: 4 Soundness: 4 Presentation: 2 Contribution: 3 Limitations: The results in the paper are limited to only finite state and action spaces and furthermore to recovering reward from an expert whose policy is corresponding to a particular kind of regularization. It is unclear whether this kind of regularization anyhow corresponds e.g. to how a human would demonstrate a practical task to a robot (I would say it doesn't). Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thank you for your feedback, valuable comments, and helpful suggestions. While we addressed most of your concerns in the response to all reviewers, we would like to respond in detail to some of your comments below. ## Our comments: - __(C)__ "I think the paper is not optimized to be read by a broader NeurIPS audience. (...) " __(A)__ Thank you for pointing this out. We will try to address this in the camera-ready version (see response to all reviewers). - __(C)__ "I would appreciate more discussion on how you would expect similar results to generalize to continuous state and action spaces." __(A)__ We will extend the corresponding discussion in the limitations section (see response to all reviewers). In short, we expect our results to translate to the continuous setting, but dealing with infinite-dimensional reward and measure spaces will introduce several technical subtleties. - __(C)__ "It is unclear whether this kind of regularization anyhow corresponds e.g. to how a human would demonstrate a practical task to a robot (I would say it doesn't)." __(A)__ Assumptions about the expert's behavior are necessary to analyze identifiability and transferability of rewards. As outlined in the response to all reviewers, entropy regularization can be seen as a model of bounded rationality, and the assumption of perfect expert optimality can be relaxed. However, we believe that further research is needed to determine suitable models of expert behavior for specific applications. --- Rebuttal Comment 1.1: Comment: Dear authors, thank you for your response. Yes, the kind of clarifying sentences you list in your main rebuttal are the kind that would help the text to be clearer, so I do encourage adding them and others to the paper - in the end, it will help your work to be more impactful. You mention some of the difficulties in extending the result to continuous setting - the recent paper on Randomized algorithms and PAC bounds for inverse reinforcement learning in continuous spaces by Kamoutsi et al. (https://arxiv.org/abs/2405.15509) does provide a good set of tools for continuous-space analysis, so it may be worth your attention if you haven't seen it yet and want to extend your work in the continuous direction.	Rebuttal 1: Rebuttal: Thank you for the valuable reviews. In the following, we propose changes and answer to questions of general interest. ## Intuition building Reviewers hiXS and fi4j suggested adding more intuition on key concepts of the paper. We will address this by adding clarifying sentences in the background section such as for example: - Line 90: "Starting from some initial state $s_0\sim\nu_0$ the agent can at each step in time $t$, choose an action $a_t\in\mathcal{A}$, will arrive in state $s_{t+1}\sim P(\cdot\|s_t, a_t)$, and receives reward $r(s_t,a_t)$." - Line 126: "From a geometric perspective, the subspace $\mathcal{U}=\text{im}(E-\gamma P)$ lies perpendicular to the set of occupancy measures $\mathcal{M}$. Therefore, adding an element of $\mathcal{U}$ to the reward leaves the performance difference between any two occupancy measures invariant." - Line 127: "Intuitively, $\mathbb{R}^{\mathcal{S} \times \mathcal{A}}/\mathcal{V}$ is the vector space obtained by collapsing $\mathcal{V}$ to zero, or in other words, it is isomorphic to the orthogonal complement of $\mathcal{V}$." Additionally, we add an illustration of the occupancy measure spaces of Example 3.2 to Appendix E (see PDF). ## Expert optimality assumption Reviewers hiXS and k5eY mentioned that our assumptions about experts' behavior might be restrictive. We agree that assuming perfect expert optimality with respect to $r^E$ and a steep regularization might be limiting. However, we would like to point out that: - As mentioned in the introduction, entropy-regularization can be seen as an information-theoretic model of bounded rationality (see [1, Section 5]) and has been widely used as a model of human behavior in social choice theory (Bradley-Terry and Luce-Plackett models [2]), behavioral game theory [3], IRL [4], and RLHF [5]. - The assumption of perfect expert optimality can be relaxed. The transferability results in Theorem 3.9 & 3.10 apply if we recover a reward $\hat{r}$ such that $\ell_{P^k}(\hat{r}, \mathsf{RL}\_{P^k}(r^E))\leq \hat{\varepsilon}$. If $\max\_{r\in\mathcal{R}} \|J(r, \mu^E_{P^k}) - J(r, \mathsf{RL}_{P^k}(r^E))\|\leq \varepsilon\_{\text{mis}}$ where $\varepsilon\_{\text{mis}}$ is a misspecification error, the convergence proof of Theorem 4.1 can be adapted to show that with high probability we recover a reward $\hat{r}$ satisfying $\ell\_{P^k}(\hat{r}, \mathsf{RL}\_{P^k}(r^E))\leq \hat{\varepsilon} + 2K \varepsilon\_{\text{mis}}$. Thus, the transferability results apply with $\hat{\varepsilon}\leftarrow \hat{\varepsilon} + 2K \varepsilon\_{\text{mis}}$. However, $\varepsilon\_{\text{mis}}$ cannot be reduced by collecting more samples from the expert. We will add a brief discussion about this relaxation to the appendix. ## Limitations and future work Aside from the above expert optimality assumption, we will discuss the following points in the "limitations and future work" section: - Continuous setting: We believe Theorems 3.9 and 3.10 can be extended to continuous state and action spaces. However, the set of occupancy measures and potential shaping transformations are infinite-dimensional in this setting. Although our proof techniques involving convex analysis and principal angles can be extended to general Banach and Hilbert spaces (see [6] and [7]), the analysis is expected to be more challenging and technically involved. Moreover, deriving explicit constants, as in Corollary 3.11 or Theorem 4.1, would likely require further assumptions. - More than 2 experts: While we mentioned that Theorem 3.9 can be trivially extended to $K>2$ experts by considering the maximum second principal angle between any two expert transition laws, the result may not be tight for $K>2$, as pointed out by reviewer fi4j. It would be an interesting direction for future research to generalize the principal angle condition in Theorem 3.9 to more than two experts. ## Clarifications - In line 127, we clarify: "Given a linear subspace $\mathcal{V} \subset \mathbb{R}^{\mathcal{S} \times \mathcal{A}}$, the quotient space $\mathbb{R}^{\mathcal{S} \times \mathcal{A}}/\mathcal{V}$ is the set of all equivalence classes $[r]\_{\mathcal{V}} := \lbrace r'\in \mathbb{R}^{\mathcal{S} \times \mathcal{A}} : r' - r \in \mathcal{V}\rbrace$, which is itself a vector space with addition and multiplication operations defined by $[r]\_{\mathcal{V}} + [r']\_{\mathcal{V}} = [r+r']\_{\mathcal{V}}$ and $c[r]\_{\mathcal{V}} = [cr]\_{\mathcal{V}}$ for $c\in\mathbb{R}$." - There is a typo in line 249. It should be: "In Example 3.2, we have $\sin(\theta_2(P^0, P^1))=\mathcal{O}(\beta)$, indicating that the second and in this case maximal principal angle is small when $\beta$ is small (see Appendix E)." As shown in Appendix E, we have $\|\|P'-P^1 \|\|=\mathcal{O}(\beta)$ where $P'$ is such that $\mathcal{U}\_P = \mathcal{U}\_{P'}$, while $\|\|P^0-P^1 \|\|$ itself is large. ## References [1] Ortega, P. A., Braun, D. A., Dyer, J., Kim, K. E., & Tishby, N. Information-theoretic bounded rationality. _arXiv preprint arXiv:1512.06789_, 2015. [2] Hamilton, Ian, Nick Tawn, and David Firth. "The many routes to the ubiquitous Bradley-Terry model." _arXiv preprint arXiv:2312.13619_, 2023. [3] McKelvey, Richard D., and Thomas R. Palfrey. "Quantal response equilibria for normal form games." _Games and economic behavior_ 10.1, 1995. [4] Ziebart, Brian D. "Modeling purposeful adaptive behavior with the principle of maximum causal entropy". _Carnegie Mellon University_, 2010. [5] Christiano, P. F., Leike, J., Brown, T., Martic, M., Legg, S., & Amodei, D. "Deep reinforcement learning from human preferences." _Advances in neural information processing systems_, 2017. [6] Mordukhovich, Boris S., and Nguyen Mau Nam. "Convex analysis and beyond". _Springer International Publishing_, 2022. [7] Zhu, Peizhen, and Andrew V. Knyazev. "Angles between subspaces and their tangents." _Journal of Numerical Mathematics_ 21.4, 2013. Pdf: /pdf/a665cb7c3aa20adc50f063795f93b715bff63857.pdf	NeurIPS_2024_submissions_huggingface	2,024	null	null	null	null	null	null	null	null
Parameter Competition Balancing for Model Merging	Accept (poster)	Summary: They propose a PCB for merging which computes parameter importance using 3 steps for intra-balancing and inter-balancing, which outperforms previous methods. Strengths: - Paper well structured and easy to follow - Method outperforms baselines Weaknesses: - Lack of novelty - similar to TIES with some modified way to compute importance. Experimental setup also similar. - No theoretical motivation for method Technical Quality: 3 Clarity: 3 Questions for Authors: - How important is the CMA-ES for hyperparameter tuning? - What happened to Fisher Merging and RegMean for the 3 LLM Tasks in Table 2? - Is the softmax done across all parameters per task vector? What is the Norm in eq 1? - Is the number of searches done with a validation set the same across different methods in the results? - In eq 2, should j exclude i (i.e. inter-balance is not computed between a model and itself) Confidence: 4 Soundness: 3 Presentation: 3 Contribution: 2 Limitations: NA Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Reply to Reviewer Zv3Q \ Thank you for your valuable comments. We will explain your concerns point by point. Weaknesses 1：About the novelty. \ Reply: Please refer to the second point in our general response document. Weaknesses 2：About the theoretical motivation. \ Reply: Please refer to the first point in our general response document. Question 1: About the function of CMA-ES. \ Reply: In fact, the function of CMA-ES is not to adjust the overall scaling hyperparameter $\lambda$. It is used to further search for the coefficients $\lambda_i$ for each task, as shown in Equation 5 in Section 3.3 of our paper. Typically, we use a uniform $\lambda$ as the initialization value for each model's $\lambda_i$, and then employ evolutionary strategies (ES) to search for a more accurate $\lambda_i$. CMA-ES is used to accelerate this search process. Recently, evolutionary strategies have also been widely used to enhance model merging in works like Lorahub [8], EvoLLM [9], and Model_Evolver [10]. Question 2: Fisher Merging and RegMean for the 3 LLM Tasks in Table 2. \ Reply: The methods of Fisher Merging and RegMean are actually not suitable for LLMs. Firstly, these methods require more GPU resources, as they necessitate the additional computation of the Fisher Information Matrix or Inner Product Matrix. Specifically, the RegMean method is not feasible even with GPUs having 80GB of memory for 7B LLMs. Moreover, both methods are gradient-based, making the results highly sensitive to the choice of small sample data and the number of training iterations, which significantly increases the complexity and instability of LLM merging. Considering these factors, we primarily compare methods based on task vectors for LLM merging, as they are easier and more lightweight to implement. Question 3: About softmax and Norm. \ Reply: It is true that the softmax is done across all parameters per task vector. Norm refers to the normalization of the task vector, which enhances numerical stability. The choice of softmax is largely due to the exponential function it incorporates, which amplifies larger values and diminishes smaller ones, thus increasing the contrast. This helps in converting the output vector into a probability distribution. In our paper, Appendix B.1 (Additional Ablation Studies), we replaced the softmax activation function with common alternatives like sigmoid, ReLU, and tanh. The results show minimal performance loss with different activation functions. This is because these activation functions can represent complex nonlinear relationships to balance the values of parameters. Question 4: About the number of searches. \ Reply: Please refer to the fourth point in our general response and table 1 in PDF file in general response. Question 5: Should j exclude i. \ Reply: We considered this issue at the initial design stage of our experiment. In our early experiments with merging eight models, we found that whether or not to exclude the model itself when computing inter-balance had a negligible impact on the results. Therefore, we opted for a simpler approach by not excluding it, which made the process straightforward and the formulas more concise. In our current experiments merging three LLM tasks, we found that excluding the model itself resulted in a score of 35.14, which is a slight improvement compared to the original score of 35.12. This suggests that this issue has minimal impact on our method. The term "inter-balance" can be understood as the balance between the model and the entire population, or the balance among the other individuals within the population. We thank the Reviewer again for the useful comments. We have revised the manuscript according to the Reviewer’s suggestion and response to each comment provided in the Weakness section above. We hope that our rebuttal aligns with the reviewer’s expectations, and we hope that the Reviewer can consider possibly giving a higher rating. Thanks. References: \ [8] Huang et al. LoraHub: Efficient cross-task generalization via dynamic LoRA composition. (arXiv23) \ [9] Akiba et al. Evolutionary optimization of model merging recipes. (arXiv24) \ [10] Du et al. Knowledge fusion by evolving weights of language models. (ACL24) \ [11] Yu et al. Language models are super mario: Absorbing abilities from homologous models as a free lunch. (ICML24) --- Rebuttal Comment 1.1: Comment: Thanks for the response. I keep my score as is.	Summary: This paper introduces an innovative technique named PCB-MERGING (Parameter Competition Balancing), a lightweight and training-free technique that adjusts the coefficients of each parameter for effective model merging. Strengths: 1. This paper re-examines existing model merging methods, highlighting the critical role of parameter competition awareness; 2. This paper introduce a novel approach called PCB-MERGING, which effectively adjusts parameter coefficients through balancing parameter competition; 3. The method stabilizes and enhances model merging performance across various application scenarios without additional training. Weaknesses: 1. Figure 1 and 2 need to be re-explained. The meaning of the percentage is confusing. Does it refer to the pruning ratio or the proportion of adjusted key parameters? The meaning of scale also needs to be re-explained. 2. The time complexity in MOEA is a crucial topic for discussion. If additional MOEA-related algorithms are introduced, a time complexity analysis needs to be conducted. 3. An important mathematical symbols have not been defined, circle with dot. 4. The code seems unavailable? I'm curious about the process of selecting the task vector. Technical Quality: 3 Clarity: 3 Questions for Authors: Described in weaknesses. I might improve the rating if my questions are well addressed. Confidence: 4 Soundness: 3 Presentation: 3 Contribution: 4 Limitations: Described in weaknesses. I might improve the rating if my questions are well addressed. Flag For Ethics Review: ['No ethics review needed.'] Rating: 7 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Reply to Reviewer s8Ho \ Thank you for your valuable comments. We will explain your concerns point by point. Weaknesses 1：More details in Figure 1 and 2 \ Reply: The term 'magnitude' refers to the magnitude of the task vector in Fig. 1. We will clarify this in the introduction of our final version by adding explanatory details. Thank you for your suggestions. Figure 1 illustrates an interesting phenomenon: scaling the top percentiles of a task vector outperforms fine-tuning, which is consistent with the ideas presented in the paper DARE [11]. \ The percentage represents the top magnitude percentiles of a task vector. It pertains to both the pruning ratio $r$ and the proportion of adjusted key parameters shown in Fig. 2, reflecting different representations in drop and rescale operations. However, in the later part where we introduce our new method, PCB-Merging, we first adjust all parameters and then perform pruning based on these adjustment results. Weaknesses 2：Time complexity in MOEA \ Reply: The total time required for the overall evolutionary strategy is $T_{\text{total}} = \text{Generations} \times (T_{\text{merging}} + T_{\text{validate}})$, where "Generations" represents the number of generations needed for evolution. The time required for model merging primarily depends on the number of model parameters and the size of the model population, while the time for model validation is mainly influenced by the volume of inference data and inference speed. We have compiled and reported the number of generations and the time required for each task in Table 1 of the general response PDF, and we analyze these factors in the fourth point of the general response. Recently, evolutionary strategies have also been widely used to enhance model merging in works like Lorahub [8], EvoLLM [9], and Model_Evolver [10]. Weaknesses 3：Definition of circle with dot \ Reply: We interpret ⊙ as an element-wise product and use Norm as a shorthand for normalization. For more details, please refer to the third point in our general response document. Weaknesses 4：Code for processing the task vector \ Reply: Actually, we have provided the source code in the supplemental material .zip file during our initial submission. The details of our method are shown in the file `pcb-merging.py`. Additionally, you can check the application in different scenarios with evolutionary strategies in `pcb_ES.py` (found in the `vision_source_code` directory) or `merging.py` (in the `nlp_source_code` directory). You can obtain the model population by executing `run_finetuning.sh` and try different merging methods using `run_merging.sh`. We thank the Reviewer again for the useful comments. We have revised the manuscript according to the Reviewer’s suggestion and response to each comment provided in the Weakness section above. We hope that our rebuttal aligns with the reviewer’s expectations, and we hope that the Reviewer can consider possibly giving a higher rating. Thanks. References: \ [8] Huang et al. LoraHub: Efficient cross-task generalization via dynamic LoRA composition. (arXiv23) \ [9] Akiba et al. Evolutionary optimization of model merging recipes. (arXiv24) \ [10] Du et al. Knowledge fusion by evolving weights of language models. (ACL24) \ [11] Yu et al. Language models are super mario: Absorbing abilities from homologous models as a free lunch. (ICML24)	Summary: The authors propose an improved method to merge task vectors, called Parameter Competition Balancing (PCB-merging). The proposed method is simple, efficient, and attains superior performance in evaluations. Strengths: 1. In Figure 1, the authors show a very interesting phenomena, where scaling the top percentiles of a task vector outperforms fine tuning. 2. The proposed method is very simple. It can be implemented in a few lines of code, requires very little memory or computation, and does not require additional data. These are crucial advantages that make PCB-merging highly practical when also considering its improved performance. 3. The method performs well, surpassing TIES-merging, Task Arithmetic, RegMean, and standard averaging in a variety of settings. Weaknesses: 1. The paper has numerous typos. Technical Quality: 3 Clarity: 2 Questions for Authors: 1. In the introduction for Figure 1, it's not quite clear what "Magnitude" refers to. Is it the magnitude in the task vector or the magnitude after the task vector has been applied? This is answered later but it would be helpful to add a word or two to clear this up in the introduction. 2. I am a bit confused about the actual operations used for the parameter balancing. I am interpreting ⊙ as an element-wise product. Thus τᵢ ⊙ τᵢ is a vector and Norm(τᵢ ⊙ τᵢ) would be a scalar, but Softmax requires a vector. It might be helpful to clarify the notation in Section 3.2 a bit more. Confidence: 4 Soundness: 3 Presentation: 2 Contribution: 3 Limitations: 1. The authors point out that the mechanism leading to the improved performance is not well understood. 2. The merging method likely cannot be used to merge models that are fine tuned from different pretrained models. I believe both of these limitations are acceptable. Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Reply to Reviewer 28gt \ Thank you for your valuable comments. We will explain your concerns point by point. Weaknesses 1：More details in Figure 1. \ Reply: The term 'magnitude' refers to the magnitude of the task vector in Fig. 1. We will clarify this in the introduction of our final version by adding explanatory details. Thank you for your suggestions. Figure 1 illustrates an interesting phenomenon: scaling the top percentiles of a task vector outperforms fine-tuning, which is consistent with the ideas presented in the paper DARE [11]. Weaknesses 2：Clarification regarding the notation for Softmax and Norm in Section 3.2. \ Reply: Please refer to the third point in our general response document. Limitation 1：Theoretical understanding. \ Reply: Please refer to the first point in our general response document. Limitation 2：Reliance on shared initializations for applications. \ Reply: Please refer to the second point in our general response document. Shared initializations are a fundamental issue in model merging, and more complex scenarios can be addressed by converting them into shared initializations through methods like FuseLLM [7]. Common applications include multi-task learning [1], multi-domain adaptation [2], merging various training strategies [3], model compression [4], and mitigating catastrophic forgetting [5], among others. We thank the Reviewer again for the useful comments. We have revised the manuscript according to the Reviewer’s suggestion and response to each comment provided in the Weakness section above. We hope that our rebuttal aligns with the reviewer’s expectations, and we hope that the Reviewer can consider possibly giving a higher rating. Thanks. References: \ [1] Ilharco et al. Editing models with task arithmetic. (ICLR 2023) \ [2] Jin et al. Dataless knowledge fusion by merging weights of language models. (ICLR 2023) \ [3] Yadav et al. Ties-merging: Resolving interference when merging models. (NeurIPS 2023) \ [4] Wang et al. Localizing task information for improved model merging and compression. (ICML24) \ [5] Zhu et al. Model tailor: Mitigating catastrophic forgetting in multi-modal large language models. (ICML24) \ [7] Wan et al. Knowledge fusion of large language models. (ICML24) \ [11] Yu et al. Language models are super mario: Absorbing abilities from homologous models as a free lunch. (ICML24) --- Rebuttal Comment 1.1: Comment: I would like to thank the authors for their rebuttal. My concerns have been addressed, but since I did not note any major issues to begin with, I will keep my score.	Summary: This paper focuses on the model merging problem. The authors propose PSC-MERGING to adjust the coefficients of each parameter for effective model merging. Specifically, PSC-MERGING uses intra-balancing to weight the importance of parameters within tasks and inter-balancing to assess parameter similarities across tasks. Parameters with low importance scores are dropped, and the remaining ones are rescaled to form the final merged model. The authors conduct extensive experiments to validate the proposed method. Strengths: + The paper is well organized. The motivation is well provided and the authors make a comprehensive survey on the related works. + The authors conduct extensive experiments, including cross-task merging, cross-domain merging, cross-training configuration merging and out-of-domain generalization, to validate the proposed method. The proposed method shows clear superiority in all these settings. Weaknesses: + The proposed method introduces an hyperparameter r in Eqn. (3). From Figure 6, it can be seen that the performance is highly sensitive to this hyperparameter, which may make the proposed method not easy to use in practice. Furthermore, the authors claim “For TIES-Merging and PCB-MERGING, which require a masking ratio, we set mask ratio r = 0.2 as the default value for all experiments, except in LLM experiments where r = 0.1.” However, From Figure 6 it is obviously that the optimal r should be 0 instead of 0.2. The setting of the hyperparameter r should be more justified. + Another concern is that the authors use quite different search space when setting the hyperparameter lambda, as described in Line #217~222. Why does the search range of this parameter vary so much under different methods? Will it incur any unfair comparison? The rationale for such settings should be provided. + The proposed method relies on shared initializations, which significantly limits its applicability as acknowledged by the authors. Technical Quality: 3 Clarity: 4 Questions for Authors: Please see the Weaknesses Confidence: 5 Soundness: 3 Presentation: 4 Contribution: 3 Limitations: N/A Flag For Ethics Review: ['No ethics review needed.'] Rating: 5 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Reply to Reviewer 6PSY \ Thank you for your valuable comments. We will explain your concerns point by point. Weaknesses 1：The setting of the hyperparameter $r$. \ Reply: The reviewer has some confusion and misunderstandings regarding our experiment settings. Firstly, our hyperparameters are discussed in two scenarios: When no additional validation is performed, the masking ratio for TIES-Merging and PCB-Merging is set to $ r = 0.2 $. When validation is allowed, we search over ratios in $\{0.05, 0.1, 0.2\}$. Additionally, in Figure 6, the optimal $ r $ is not 0; it should be 0.05 instead. This conclusion is consistent with our findings in Appendix F.3, Table 17, concerning hyperparameter settings. \ In fact, when $ r $ is close to 0, the performance of TIES-Merging and PCB-Merging drops sharply, as shown in Figure 1 in the general response PDF file. The optimal range for $ r $ is between 0.03 and 0.3. To improve feasibility, we only search within the values $\{0.05, 0.1, 0.2\}$. Weaknesses 2：The rationale for the various settings of different methods. \ Reply: In this paper, the selection of hyperparameters for different methods adheres to the original papers and is indeed suitable for obtaining optimal solutions for these methods. Therefore, this comparison is fair and reasonable. Specifically: \ For TIES-Merging, we followed the hyperparameter settings outlined in Appendix C.4 and C.5 of the original paper. \ For Task Arithmetic, we adhered to the hyperparameter settings in Appendix D (Learning via addition) of the original paper. \ For RegMean, we followed the discussion on the impact of scaling non-diagonal values in Inner Product Matrices in section 5.3 of the original paper. \ Lastly, since the meaning of the hyperparameter $\lambda$ varies across different methods, we summarized the range and optimal values of hyperparameters for each method in our paper. This is detailed in the Experimental Setup section and Appendix F.3, Table 17. Weaknesses 3：Reliance on shared initializations for applications.: \ Reply: Please refer to the second point in our general response document. Shared initializations are a fundamental issue in model merging, and more complex scenarios can be addressed by converting them into shared initializations through methods like FuseLLM [7]. Common applications include multi-task learning [1], multi-domain adaptation [2], merging various training strategies [3], model compression [4], and mitigating catastrophic forgetting [5], among others. We thank the Reviewer again for the useful comments. We have revised the manuscript according to the Reviewer’s suggestion and response to each comment provided in the Weakness section above. We hope that our rebuttal aligns with the reviewer’s expectations, and we hope that the Reviewer can consider possibly giving a higher rating. Thanks. References: \ [1] Ilharco et al. Editing models with task arithmetic. (ICLR 2023) \ [2] Jin et al. Dataless knowledge fusion by merging weights of language models. (ICLR 2023) \ [3] Yadav et al. Ties-merging: Resolving interference when merging models. (NeurIPS 2023) \ [4] Wang et al. Localizing task information for improved model merging and compression. (ICML24) \ [5] Zhu et al. Model tailor: Mitigating catastrophic forgetting in multi-modal large language models. (ICML24) \ [7] Wan et al. Knowledge fusion of large language models. (ICML24) --- Rebuttal Comment 1.1: Comment: I appreciate author's feedback. Some of my concern are resolved. I keep my original positive score. --- Reply to Comment 1.1.1: Title: Replying to Reviewer 6PSY Comment: Thank you for your positive score and insightful feedback. If you have any concerns about our work, we would greatly appreciate receiving any further comments or suggestions.	Rebuttal 1: Rebuttal: General Response We appreciate your consideration in taking the time to review our comments. We have received feedback from four reviewers, all of whom have provided thoughtful insights. Almost all reviewers found our paper to be well-organized, motivated, practical, and easy to follow. Additionally, our paper has demonstrated extensive experiments and applications, leading to improved performance. However, there are still some questions and concerns from the reviewers. We have summarized and addressed these in four key points. 1. Motivation and theoretical understanding: \ Our research aims to improve the performance of model merging by addressing parameter competition through a balancing mechanism that adjusts parameter-level coefficients. Our method, PCB-Merging, retains the advantages of being lightweight, easy to implement, and high-performing. We systematically compare and analyze existing model merging methods regarding their intra-balancing and inter-balancing capabilities, emphasizing the importance of parameter competition awareness. We establish a balancing matrix that is self-aware and cross-aware for parameter scaling. Furthermore, we introduce PCB-Merging, a novel approach that effectively adjusts parameter coefficients by balancing parameter competition. 2. Application scenarios of our method: \ Firstly, as with most current research, model merging methods are limited by shared initializations. Despite this, they have a wide range of applications, such as multi-task learning [1], multi-domain adaptation [2], merging various training strategies [3], and model compression [4]. Secondly, our method can also enhance transferability and mitigate catastrophic forgetting. For instance, merging the transferred model with the original model in pairs can improve practicality, as demonstrated in the papers Model Tailor [5] and Robust Fine-tuning [6]. Finally, in knowledge fusion scenarios, we can use methods like knowledge distillation and model alignment to convert models with different initializations to the same initializations. This approach can even be applied across different frameworks to achieve model merging, as shown in the paper FuseLLM [7]. 3. Clarification regarding the notation in Section 3.2: \ We apologize for any confusion regarding our use of the term "Norm" in the manuscript. We meant the normalization of the vector τi ⊙ τi, not its norm (which would be a scalar). Normalizing this vector retains its vector form. Softmax is applied across all parameters per task vector. Normalizing the task vector improves numerical stability. We chose softmax for its exponential function, which increases contrast by amplifying larger values and diminishing smaller ones, converting the output into a probability distribution. In Appendix B.1 (Additional Ablation Studies), we replaced softmax with alternatives like sigmoid, ReLU, and tanh. The results showed minimal performance loss, as these functions also balance parameter values effectively. 4. Time Complexity for Evolution Strategy: \ Recently, evolutionary strategies have also been widely used to enhance model merging in works like Lorahub [8], EvoLLM [9], and Model_Evolver [10]. We propose evolution strategy to search for the coefficients $\lambda_i$ for each task, as shown in Equation 5 in Section 3.3 of our paper. A specific CMA-ES algorithm is used to accelerate the search process. The total time required for the overall evolutionary strategy is $T_{\text{total}} = \text{Generations} \times (T_{\text{merging}} + T_{\text{validate}})$, where "Generations" represents the number of generations needed for evolution. The time for model merging mainly depends on the number of model parameters and the size of the model population, while the time for model validation primarily depends on the volume of inference data and the inference speed. We have organized and reported the number of generations and the time required for each task in Table 1 of the general response PDF file, and analyzed this in the fifth point of the general response. Once again, we sincerely thank you for your involvement and thoughtful feedback! We have provided detailed responses to each question from each reviewer. Additionally, we have included the source code and the experimental procedures for the evolution strategy in NLP, vision, and other applications in the supplemental material zip file for your reference. References: \ [1] Ilharco et al. Editing models with task arithmetic. (ICLR 2023) \ [2] Jin et al. Dataless knowledge fusion by merging weights of language models. (ICLR 2023) \ [3] Yadav et al. Ties-merging: Resolving interference when merging models. (NeurIPS 2023) \ [4] Wang et al. Localizing task information for improved model merging and compression. (ICML24) \ [5] Zhu et al. Model tailor: Mitigating catastrophic forgetting in multi-modal large language models. (ICML24) \ [6] Wortsman et al. Robust fine-tuning of zero-shot models. (CVPR22) \ [7] Wan et al. Knowledge fusion of large language models. (ICML24) \ [8] Huang et al. LoraHub: Efficient cross-task generalization via dynamic LoRA composition. (arXiv23) \ [9] Akiba et al. Evolutionary optimization of model merging recipes. (arXiv24) \ [10] Du et al. Knowledge fusion by evolving weights of language models. (ACL24) Pdf: /pdf/ef522502eec534ce3c26c67d90104f42bc58b279.pdf	NeurIPS_2024_submissions_huggingface	2,024	null	null	null	null	null	null	null	null
Improved Sample Complexity for Multiclass PAC Learning	Accept (poster)	Summary: This paper presents improved sample complexity upper and lower bounds for multiclass classification, shaving previous upper bounds to within a factor of $\log(1/\epsilon)$ from the conjectured optimal dependence on $\epsilon$, and adding a dependence of $\log(1/\delta)$ to previous best lower bounds. To do so, the authors study multiclass classification through the lens of the more genearal problem of list-learning -- this generality allows for a theorem stating that the conjectured optimal dependence of $1/\epsilon$ in the sample complexity holds given a conjecture about the sample complexity of list-learning. They further discuss the potential of technique they term ``pivot shifting'' to remove the extra log factor of $1/\epsilon$ in the multi class complexity, showing that whenever pivoting shifting does not increase a characteristic combinatorial complexity measure by more than a constant factor, pivot shifting eliminates the extra log factor. Strengths: The paper introduces a novel boosting algorithm for list-learning which enables the log improvement in $1/\epsilon$, allowing for the effective application of list-learning machinery to the multiclass problem. It improves on previous best lower bounds via the addition of the usual dependence on the confidence parameter $\log(1/\delta)$. Weaknesses: Outside of Section 1, I found the presentation of the paper to be very weak. Certain paragraphs present hard to follow proof outlines (e.g. line 236 "the standard leave-one-out argument" is not something I am familiar with and is not explained in the main body). Section 3, which introduces the concept of pivot shifting, provides little intuition into the notation-heavy definitions provided, and cannot be understood without reference to definitions which only appear in the appendix. It is also a bit unclear what the upside of introducing the generality of list-learning is. The second half of Theorem 2.11 (line 281) does not feel like particularly compelling evidence for the conjectured $1/\epsilon$ dependence, given that it relies on an analogous intuitive conjecture on the sample complexity of the more general problem of list learning. Technical Quality: 3 Clarity: 1 Questions for Authors: 1) Is there a particularly compelling reason to believe in the existence of some $\mathcal{A}_{goodlist}$ in Theorem 2.11? What do see as the advantage of thinking about multi-class learning through this extra layer of generality? 2) Could you elaborate on the $\log(1/\delta)$ factor in the lower bound (Theorem 2.5) -- does the proof technique follow that of Charikar and Pabbaraju [2023], or is this a different construction? Confidence: 3 Soundness: 3 Presentation: 1 Contribution: 3 Limitations: Yes Flag For Ethics Review: ['No ethics review needed.'] Rating: 5 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thank you for reviewing our paper. Below, $d$ denotes DS dimension and $d_k$ denotes $k$-DS dimension. ## List learning * The advantage of reducing to list learning is that it reduces the original problem to an easier problem: list learning is easier than multiclass learning. Indeed, a multiclass learner is a list learner of any size. If a concept class is PAC learnable with some sample complexity, then it is also $k$-list PAC learnable with the same sample complexity for any size $k$. On the contrary, for any $k\ge2$, we cannot directly obtain a multiclass learner from a general $k$-list learner. If a concept class is $k$-list PAC learnable with some sample complexity, it is unclear whether it is multiclass PAC learnable and what its multiclass sample complexity is. As mentioned in line 285 (there is a typo, it should be "$\mathsf{dim}_k(\mathcal{H})\ge$'' instead of "$\mathsf{dim}_k(\mathcal{H})>$''), $d_k(\mathcal{H})$ is nonincreasing with $k$. Consequently, a $k$-list PAC learnable concept class (i.e., $d_k(\mathcal{H})<\infty$) is not necessarily multiclass PAC learnable. Thus, Theorem 2.7 and 2.11 are highly non-trivial because they reduce the harder problem of multiclass learning to the easier problem of list learning: quantitatively, if a concept class is $k$-list learnable with some error rate $r(n)$, it is also multiclass learnable with an error rate bounded by $r(n)+d\log(k)/n$. * Regarding $\mathcal A_{\text{goodlist}}$, if there is a multiclass learner $A$ with error rate bounded by $f(d)/n$, then $A$ is also a $k$-list learner with $\epsilon_{A,\mathcal{H}}(n)\le f(d)/n$ for any $k\in\mathbb{N}$. Thus, our conjecture on list learners holds if the conjecture of the same rate on multiclass learners holds, indicating that the conjecture for such a list learner is weaker than that for a multiclass learner, hence more approachable. Moreover, in Theorem 2.10, our list learner already saves a log factor compared to the previous best list learner in [1]. It is reasonable to conjecture the removal of the remaining log factor. * The fact that we reduced a log factor in the upper bound of multiclass sample complexity by reducing to list learning also shows its advantage. The improvement of log factors in PAC sample complexity is significant; e.g., in binary classification, though the upper bound with a $\log(1/\epsilon)$ factor was established in 1982 [2], the log factor was not removed until 2016 [3]. Given that we reduced multiclass learning to an easier problem of list learning and thus removed a log factor, we believe our paper has made a significant contribution to multiclass learning. * Finally, with a large or infinite label space, it is natural to consider reducing the size of the label space by proposing a list learner. This idea has been adopted in [1], the first paper to establish the equivalence of multiclass PAC learnability and the finiteness of DS dimension. [1] proposed a list learner to construct a list sample compression scheme and built their multiclass learner using the labels in the list. ## Lower bound One cannot derive the $\Omega((d_k-\log\delta)/k\epsilon)$ bound using the $\Omega(d_k/kn)$ expected error lower bound in [4] as a black box since one must stick to the distribution constructed for the lower bound when applying it as a black box. Thus, for the PAC lower bound, we have to construct different hard distributions from that in [4] for both the $\log(1/\delta)/k\epsilon$ (line 486) and the $d_k/k\epsilon$ (line 511) terms. The detailed constructions are provided in the proof of Theorem 2.5 in Appendix B. To list one difference between our constructions and that in [4], our distributions are not uniform over the $\mathcal{X}$ sequence selected and depends on the parameter $\epsilon$, while theirs are uniform over the $\mathcal{X}$ sequence selected. Finally, our major contribution is the upper bound. We include the lower bound and its proof for comparison and completeness as it is missing in the literature (see line 37 and 124). ## Presentation * We provided the reference [1, Fact 14] of the leave-one-out argument in line 102 where it first appears. We have added the reference in line 236 in the revision. The current proof outlines are concise due to the space limitation, but we believe that they convey the key points of the proof. If the reviewer has other questions regarding the proof outlines, we are glad to answer in the discussion. We will extend the proof outlines to improve the readability in the revision. * Due to the space limitation, we have to put some definitions in the appendix, but we included the references of the definitions when they appear in the main text. We will move the important definitions in Appendix A to the main text with the additional page if the paper is accepted. We believe that we have provided intuition for the definitions. We summarized the existing results in Proposition 3.1 to motivate the idea of upper bounding the density. In Theorem 3.2, this idea is supported by its success for classes of DS dimension 1. In line 329-335, we explained the reason of considering the pivot of a concept class. In line 346-352, we provided the intuition of pivot shifting and compared it to the standard shifting in the literature. Lemma 3.7 explained the partial validity of pivot shifting, which naturally leads to Open Question 2 about the remaining validity of pivot shifting in upper bounding the density using DS dimension. We have added more explanations in words for the technical definitions to improve the readability in the revision. [1] Brukhim et al. A characterization of multiclass learnability. 2022. [2] Vapnik. Estimation of Dependencies Based on Empirical Data. 1982. [3] Hanneke. The Optimal Sample Complexity of PAC Learning. 2016. [4] Charikar and Pabbaraju. A characterization of list learnability. 2023. --- Rebuttal Comment 1.1: Comment: Thank you for your reply, and apologies for not fully understanding your contribution on the first read. On account of this, I raise my rating to a 5. In particular, I thank the authors prompting me to absorb the advatange of the reduction analyzed in Theorem 2.7. This makes clear to me the utility of considering the conjectured $\mathcal{A}_{goodlist}$, and eliminates the concerns in the second half of the above "weaknesses" section of my initial review. I still feel the presentation leaves a fair amount to be desired. While I understand that the contribution is rather technical, the sheer density and abruptness of the writing in certain parts do make for a particularly good reading experience in my opinion. While it's of course possible improve such things in the next version, it does limit my revised rating. Whether the technical contribution outweighs this downside is probably not something I'm qualified to comment on, as the paper is part of a line of work that I'm not familiar with. --- Rebuttal 2: Comment: Thank you for raising your rating. We will further elaborate on the technical aspects of the paper and improve its readability.	Summary: For the problem of analyzing the optimal PAC sample complexity for multi-class learning, two possible routes to induce the improved sample complexity and error rate are proposed. On the one hand, benefiting from the reduction from multi-class learning to list learning and the boosting technique, the dependence of the corresponding upper bounds on the sample size is improved by a logarithmic factor. On the other hand, the introduction of "pivot shifting" shows its potential ability to improve the sample complexity of multi-class learning. Strengths: 1. Based on the Daniely-Shalev-Shwartz (DS) dimension, using the reduction and the boosting technique, the explicit lower bound on the sample complexity of multi-class learning is derived, and the upper bound of the error rate for multi-class learning is improved by a logarithmic factor. 2. In order to obtain the corresponding optimal theoretical bounds, the implementation limitations of some relevant assumptions are formalized as open problems and discussed in detail, i.e., the construction of the specific list learners and the impact of pivot shifting on DS dimension. Weaknesses: The theoretical analysis in this paper improves the existing theoretical results for multi-class learning on the DS dimension, but the relationship between the obtained theoretical bounds and other data-independent complexity bounds based on combinatorial dimensions, e.g., the graph dimension, the Natarajan dimension and its scale-sensitive analog, is not discussed and compared, which is not conducive to demonstrating the improvement of these theoretical results from a macro perspective. In addition, existing theoretical results on multi-class learning show that data-independent generalization bounds often depend on the number of classes, and the best (weakest) known dependency is logarithmic. The results in this paper focus on improving the dependency on the number of samples, and do not explicitly discuss and analyze the dependency on the number of classes. These related issues require further analysis and explanation. Technical Quality: 3 Clarity: 3 Questions for Authors: Please refer to the Weaknesses for details. Confidence: 4 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: This work does not seem to have any potential negative societal impact. Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thank you for reviewing our paper. We appreciate your positive feedback. First, we need to emphasize that in this paper, we study multiclass PAC sample complexity for general concept classes which can have infinite number of labels ($\|\mathcal{Y}\|=\infty$). Our results hold for general concept classes independent of the number of labels. Below, $d$ denotes DS dimension, $d_N$ denotes Natarajan dimension, and $d_G$ denotes graph dimension. ## Graph dimension and Natarajan dimension * As we have stated in line 26-28, [1] showed that a concept class is PAC learnable if and only if its DS dimension is finite. Moreover, as is detailed below, it has been shown in the literature that finite graph dimension or Natarajan dimension does not characterize multiclass PAC learnability. * For the graph dimension, [2, Section 6] showed that finite graph dimension does not characterize multiclass PAC learnability by identifying a PAC learnable class with infinite graph dimension and infinite label space. [3, Example 1] provided a family of concept classes whose graph dimensions can be any positive integers or infinite, and all those concept classes are PAC learnable with sample complexity $\log(1/\delta)/\epsilon$. [1, Example 8] is a concept class with DS dimension 1 (thus it is PAC learnable), infinite graph dimension, and infinite label space. * For the Natarajan dimension, [1, Theorem 2] provided a concept class with Natarajan dimension 1 and infinite DS dimension (thus it is not PAC learnable). * Thus, graph dimension and Natarajan dimension cannot appear in the optimal PAC sample complexity. In general, for $\mathcal{H}\subseteq \mathcal{Y}^{\mathcal{X}}$, we have $d_N(\mathcal{H})\le d(\mathcal{H})\le d_G(\mathcal{H})$ [4] and for the special case of $\|\mathcal{Y}\|<\infty$, $d_G(\mathcal{H})\le 5\log_2(\|\mathcal{Y}\|)d_N(\mathcal{H})$ [3]. Quantitative scaling between those dimensions for general concept classes is still missing, and as is discussed above, there exist $\mathcal{H}$ and $\mathcal{H}'$ such that $d_N(\mathcal{H})=1$, $d(\mathcal{H})=\infty$, $d(\mathcal{H}')=1$, and $d_G(\mathcal{H}')=\infty$. * For detailed comparisons, our lower bound in eq. (2) is better than the best lower bound in terms of Natarajan dimension in [3, Theorem 5], and the Natarajan dimension cannot upper bound the PAC sample complexity of general concept classes by [1, Theorem 2] discussed before. The graph dimension cannot lower bound the PAC sample complexity of general concept classes by the above examples of PAC learnable concept classes with infinite graph dimensions. The best upper bound in terms of the graph dimension is $O((d_G(\mathcal{H})+\log(1/\delta))/\epsilon)$ by Proposition I.5 of our paper (see also line 226), which can be infinite for PAC learnable concept classes by the examples discussed before. Moreover, by [3, Example 1], there exists $\mathcal H_i$ for all $1\le i\le\infty$ such that $d_G(\mathcal H_i)=i$ and $\mathcal{M}_{\mathcal{H}_i}(\epsilon,\delta)\le \log(1/\delta)/\epsilon$, indicating that their DS dimensions are uniformly bounded in terms of $i$. Thus, the optimal upper bound cannot depend on graph dimension, and the upper bound using graph dimension can be arbitrarily worse than our upper bound using DS dimension as our upper bound is always non-trivial with a finite gap of at most $\sqrt{d}\log(d)\log(d/\epsilon)$ from the optimal bound for PAC learnable concept classes. ## Number of classes * As mentioned before, our upper and lower bounds on multiclass PAC sample complexity are independent of the number of classes which can be infinite. We actually prove the following result. There exists an algorithm $A$ and universal constants $C_1,C_2>0$ such that for any feature space $\mathcal{X}$, label space $\mathcal{Y}$ ($\|\mathcal{Y}\|$ can be arbitrarily large or infinite), and concept class $\mathcal{H}\subseteq\mathcal{Y}^{\mathcal{X}}$, we have $\mathcal M_{\mathcal{H}}(\epsilon,\delta)\ge C_1({d+\log(1/\delta)})/{\epsilon}$ and $\mathcal{M}_{A,\mathcal{H}}(\epsilon,\delta) \le C_2({d^{3/2}\log(d)\log(d/\epsilon)+\log(1/\delta)})/\epsilon$. As we have listed above, there are many concept classes with infinite label space and finite DS dimension. Actually, by taking finite levels in the tree example in [1, Figure 3], we can construct a series of concept classes with DS dimension 1 and number of labels increasing to infinity. Thus, any bound that increases to infinity with the number of labels cannot be optimal and is worse than our upper bound for general concept classes. It follows that the optimal bound must be uniformly bounded with respect to the number of labels and the number of labels should not exist in the optimal bound. We will incorporate the key points in the above discussion in the revised version of the paper. [1] Brukhim et al. A characterization of multiclass learnability. FOCS 2022. [2] Natarajan. Some results on learning. 1989. [3] Daniely et al. Multiclass learnability and the ERM principle. JMLR, 2015. [4] Daniely and Shalev-Shwartz. Optimal learners for multiclass problems. COLT 2014. --- Rebuttal Comment 1.1: Comment: Thanks for your detailed feedback. It has partially addressed my concerns. --- Reply to Comment 1.1.1: Comment: Thank you for your response and we are glad to hear that.	Summary: This work is focused on improving sample complexity upper and lower bounds on multiclass classifcation for a general hypothesis class. Their bounds are given in terms of the so-called DS-dimension of the hypothesis class, which can be viewed as a generalization of the VC dimension to the multiclass setting. Their bounds improve on the best known upper bounds and leave a factor of $O(\sqrt{d})$ between their upper and lower bounds on the sample complexity. Their main idea is to apply a reduction to list learning, where the learner is allowed to output a set of $k$ potential labels for each $x$ and is evaluated based on whether any of the outputted labels are correct. In particular, they provide a multiclass algorithm that works by calling a list learning algorithm as a subroutine. The subsequent guaranteed loss then grows logarithmically with $k$, the number of used labels. Their full algorithm consequently requires a list learner, and they provide one that adapts a previously known boosting technique for binary classification to multiclass list learners. To apply boosting, they define the majority vote of a set of lists as the set of labels are included in at least half of the lists. Note that the resulting list has size $2k$ if each of the lists have size $k$. Combining this with their above reduction results in their upper bound. Strengths: This paper is a highly technical paper that makes significant progress on a classic and important open problem in learning theory. In addition to improving the current best bounds, it provides an approach for further improvements. In particular, their reduction implies that improvements on list learning will result in improvements on multiclass classification. Weaknesses: Many of the proof ideas are differed to the appendix. I think it would improve the presentation to include small "proof intution" sections following the theorems. While this would cost more space, I think this can be accounted for by moving the highly technical algorithm blocks into the appendix and merely giving summaries of what the algorithms do in English. Of course, this paper is "incremental" in that it merely improves existing bounds to an preexisting problem, but I do not see this as cause at all for rejection. The improvements are clearly non-trivial, and the problem being studied is of fundamental importance. Technical Quality: 4 Clarity: 3 Questions for Authors: See weaknesses. Confidence: 3 Soundness: 4 Presentation: 3 Contribution: 3 Limitations: Yes. Flag For Ethics Review: ['No ethics review needed.'] Rating: 7 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thank you for reviewing our paper. We appreciate your positive feedback. We provided a proof sketch for Theorem 2.7 in line 229-241. We will elaborate more on it and include the proof intuition for other major theorems in the revision. We believe that the extra space can be accounted for by the additional page allowed if the paper is accepted. --- Rebuttal Comment 1.1: Title: Response to rebuttal Comment: Thank you for your response, my score will remain the same, and I think this paper should be accepted. --- Reply to Comment 1.1.1: Comment: Thank you again for your support.	Summary: The paper considers the problem of analyzing the sample complexity of Multiclass PAC Learning. The key contributions of this paper are two-fold: (1) Give an improved upper bound for the sample complexity by a poly-log factor, and (2) give the first (formal) lower bound for the Multiclass PAC Learning problem, which matches that of binary concept class. The key idea is to use a reduction from multiclass learning to list learning, which can then be further analyzed using recent works on boosting algorithms for list learners. Besides, the paper also explores a (potential) alternative approach to analyze the sample complexity for a concept class in Multiclass PAC Learning by alternatively analyzing its corresponding one-inclusion graph. Following this route, if we can give an upper bound for the density of any concept class (defined by the average degree of its corresponding one-inclusion graph) by multiple of its DS dimensions, we can have a matching lower bound for the problem of Multi-class PAC Learning. They argue that this approach is promising by giving proof for the case where the DS dimension of the concept class is 1. Strengths: The paper is well-written and easy to follow. Motivation and key results in the main paper are presented clearly and concisely, though I would also want the proofs in the Appendix to be discussed and commented on in the same way as in the main paper. I did not go through the proof very carefully, but it looks good at first glance. I would try to go through that in the details in the rebuttal phase. I like the presentation of this paper: first presenting the improvement on the sample complexity of Multiclass PAC Learning using a technique, and then proposing an alternative view that can potentially lead to a matching upper bound. This gives good insights to the readers, not simply about the sample complexity improvements, but also about the problem itself. Overall, I think this is a good paper and vote to accept it. Weaknesses: There is not much to comment on the weaknesses of this paper. A fastidious reviewer might argue that this paper is a collection of many good but not-so-strong results (in the sense that the list learner techniques used for the proofs of lower and upper-bound using are not completely new), but I think it is good enough for me. Some minor comments: 1. It is good to include a proof sketch for every result presented in the Appendix, as well as the high-level intuitions behind the results. Though the statements of key results are clear, readers like us would like to see what is actually going on behind the proofs and would appreciate it if the authors could break down the steps for better readability. Technical Quality: 3 Clarity: 3 Questions for Authors: See above. Confidence: 3 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: See above. Flag For Ethics Review: ['No ethics review needed.'] Rating: 7 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thank you for reviewing our paper. We appreciate your positive feedback. We will include the proof sketches and high-level intuitions of the results presented in the Appendix in the revised version of the paper.	null	NeurIPS_2024_submissions_huggingface	2,024	Summary: This paper gives better bounds on the complexity of multiclass learning using DS-dimension and provides a lower bound. The improvement is relatively minor. Strengths: The main strength of this paper is that it addresses one of the most fundamental problems in learning theory -- the complexity of multiclass learning and provides an improved bound on the sample complexity. Weaknesses: The writing of the paper and the presentation style can be improved substantially. The technical summary is way too dense and the comparison with the previous work is very narrow. The technical improvement is rather weak (some log factors) and doesn't address the main gap (polynomial gap in d). Technical Quality: 3 Clarity: 2 Questions for Authors: -- I am not convinced that DS-dimension is the right quantity to look at here. Don't we already have optimal bounds for the sample complexity of multiclass learning in terms or Graph Dimension (see Daniely and Shalev-Shwartz)? Can you please explain how your results compare to the Graph Dimension results? Confidence: 4 Soundness: 3 Presentation: 2 Contribution: 2 Limitations: None Flag For Ethics Review: ['No ethics review needed.'] Rating: 5 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thank you for reviewing our paper. Below, $d$ denotes DS dimension and $d_G$ denotes graph dimension. ## Graph dimension * The DS dimension is the right quantity to look at here. The optimal multiclass PAC sample complexity is described by DS dimension and not by graph dimension. As is stated in line 26-28, [1] showed that a concept class is PAC learnable if and only if its DS dimension is finite. [2, Sec. 6] showed that finite graph dimension does not characterize PAC learnability by proposing a PAC learnable class with infinite graph dimension. [3, Example 1] provided a family of concept classes $\mathcal H_i$ for all $1\le i\le\infty$ such that $d_G(\mathcal H_i)=i$ and $\mathcal M_{\mathcal{H}_i}(\epsilon,\delta)\le \log(1/\delta)/\epsilon$, indicating that their DS dimensions are uniformly bounded. [1, Example 8] is a concept class with DS dimension 1 (thus it is PAC learnable) and infinite graph dimension. Thus, graph dimension cannot appear in the optimal multiclass PAC sample complexity. * By the above examples, graph dimension cannot lower bound the PAC sample complexity of general concept classes. The best upper bound using graph dimension is $O((d_G(\mathcal{H})+\log(1/\delta))/\epsilon)$ by our Proposition I.5 (see also line 226), which can be infinite for PAC learnable classes. By [3, Example 1], the optimal upper bound cannot depend on graph dimension, and the above upper bound can be arbitrarily worse than our upper bound using DS dimension as our upper bound is always non-trivial with a finite gap of at most $\sqrt{d}\log(d)\log(d/\epsilon)$ from the optimal bound for PAC learnable classes. * Moreover, Daniely and Shalev-Shwartz [4] did not provide optimal bound on PAC sample complexity in terms of graph dimension. Instead, in [4, Conjecture 11], they conjectured that the optimal bound is given by DS dimension. In the last paragraph, they wrote > it is known (Daniely et al., 2011) that the graph dimension does not characterize the sample complexity, since it can be substantially larger than the sample complexity in several cases. We will include the key points in the above discussion in our revision. ## Technical summary We believe that it is necessary to include enough technical summaries to introduce results and ideas. We made efforts on providing sufficient motivation of the technical summaries. For example, in Sec. 1, we reviewed existing results in multiclass PAC learning in the first two paragraphs, motivating the introduction of the concepts in PAC learning in Sec. 1.1. In paragraph 3 of Sec. 1, we previewed our reduction to list learning, motivating the technical summary of list learning. In paragraph 4 of Sec. 1, we provided the motivation of the pivot shifting in Sec. 3. In Sec. 2.2, after the main result, we provided a technical overview in line 229-241 to explain the key points of the proof. In Sec. 2.3, we provided the intuition of Alg. 2 in the first paragraph. In Sec. 3, we built the technical summary on existing results to explain the motivation of upper bounding the density and pivot shifting. We will elaborate more on technical summary in the revision. ## Comparison with previous work We believe that our paper makes adequate comparisons with previous work. We compared our upper bound to the previous best bound (line 39 and 120) in line 49 and 288. We mentioned that a potentially sharp lower bound on PAC sample complexity is still missing in the literature in line 37 and 124. In line 244-251, we compared to the list sample compression scheme in [1] to explain how we saved a $\log(n)$ factor using sampled boosting. In line 268-271, we compared to the bound of directly invoking the learner in [1] to Alg. 2 to explain how we avoided a $\log(d)$ factor using list learning. In Prop. 3.1, we reviewed known results on density to explain the reason of bounding the density. In line 329-334, we referred to the induction technique in Haussler et al. (1994) to motivate pivot shifting. In line 348-352, we compared pivot shifting to the standard shifting technique. ## Technical improvement * We do not think that our improvement is weak. The optimal PAC sample complexity is a fundamental and difficult problem in statistical learning theory. The characterization of multiclass PAC learnability remained open for decades until recently solved by [1]. The improvement of log factors in sample complexity is highly non-trivial: e.g., in binary classification, though the upper bound with a $\log(1/\epsilon)$ factor was established in 1982 [5], the log factor was not removed until 2016 [6]. * In this paper, we focus on improving the dependence of PAC sample complexity on $\epsilon$ (equivalently, the dependence of the error rate on $n$). In learning theory, it is important to study the dependence of error rate on $n$ with $d$ fixed because a learner is typically applied to a fixed concept class with increasing $n$ so that $n$ is much larger than $d$. Moreover, by [7], our upper bound improves the universal learning rate from $\log^2(n)/n$ to $(\log n)/n$ for the concept classes studied in [7]. Ideally, we want tight dependence on each parameter. However, as tight dependence on either $d$ or $\epsilon$ is unknown, we believe it is meaningful to explore better dependence on either parameter. * Besides the removal of a log factor, we also proposed two possible routes toward removing the remaining log factor of $\epsilon$. [1] Brukhim et al. A characterization of multiclass learnability. 2022. [2] Natarajan. Some results on learning. 1989. [3] Daniely et al. Multiclass learnability and the ERM principle. 2015. [4] Daniely and Shalev-Shwartz. Optimal learners for multiclass problems. 2014. [5] Vapnik. Estimation of Dependencies Based on Empirical Data. 1982. [6] Hanneke. The Optimal Sample Complexity of PAC Learning. 2016. [7] Hanneke et al. Universal rates for multiclass learning. 2023. --- Rebuttal Comment 1.1: Comment: Thanks, sounds good. I can increase the score to 5. --- Reply to Comment 1.1.1: Comment: Thank you for increasing the rating.	null	null	null	null	null	null
Diffusion Actor-Critic with Entropy Regulator	Accept (poster)	Summary: The paper proposes DACER — an actor-critic that uses the inverse diffusion process as policy. Additionally, some noise is added to actions to increase the entropy (similar motivation as in SAC but implemented differently). Empirical evaluations and ablations show that the method is quite capable. Strengths: Originality: good. The paper proposes a novel method how to use diffusion as policy representation for a SAC-like algorithm. Quality: good. The method is presented and evaluated thoroughly. Clarity: excellent. The paper is easy to follow. Significance: good. The method will be of interest to the RL community. Weaknesses: 1) No comparison to an algorithm with multimodal policy. The key advantage of the diffusion policy compared to the usual Gaussian is its multimodality. But there are other algorithms with multimodal policies, e.g., [Reinforcement Learning with Deep Energy-Based Policies](http://proceedings.mlr.press/v70/haarnoja17a.html?ref=https://githubhelp.com) and many newer works (check "Cited by" on that paper). 2) No clear demonstration where multimodality helps. Some toy task or better not-toy task where one sees a clear advantage would be nice to see. Technical Quality: 3 Clarity: 4 Questions for Authors: The proposed entropy regulation scheme seems a bit unusual. Did you try other options (e.g. like in TRPO)? Confidence: 4 Soundness: 3 Presentation: 4 Contribution: 3 Limitations: Please add a paragraph discussing the limitations Flag For Ethics Review: ['No ethics review needed.'] Rating: 7 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We appreciate you for the careful reading of our paper and detailed discussions. ### > Weakness 1 As you mentioned, it is meaningful to compare with other algorithms that have multimodal characteristics. We chose to compare the performance with two online RL with Diffusion model algorithms mentioned in the related work (DIPO and QSM). In Figure 3 of the author global rebuttal PDF, it can be seen that DACER achieved the best performance. ### > Weakness 2 We answer this question in two ways. First, some not-toy tasks are inherently multimodal. Zou et al.[1] proposes a vehicle control scenario in which the policies are multimodal, i.e., in the case where traffic regulations are not taken into account and the ego car is directly behind the front vehicle and there is only one weekly vehicle, it is an optimal policy for the ego car to overtake both from the left and from the right. In this case, only a multimodal policy is able to deal with the situation. We add experiments comparing the action (steering wheel angle) distributions of DSAC and DACER at the bifurcation point. We use the diffusion policy to sample 200 different actions and plot a histogram in Figure 5 of the author global rebuttal PDF. The results indicate that DACER learns a bimodal policy, whereas DSAC can only learn a unimodal distribution for overtaking from one side. Second, we think multimodality enhances exploration. To assess the multi-style capability of DACER, we use the method proposed in [2]. This method analyze the diversity of state trajectories by computing the information entropy from 1,100 independent simulation episodes. The experimental results are plotted in the author global rebuttal Table 2. The results show that in the Humanoid-v3, Ant-v3, Hopper-v3 and Pusher-v2 tasks, DACER achieves the highest entropy value, surpassing DSAC, SAC, DDPG, TD3, PPO, and TRPO. ### > Question 1 This question of yours is one of the most creative points of DACER. The entropy of the diffusion policy is difficult to calculate analytically, which means we cannot use entropy regularization methods like in TRPO to adjust the entropy (the gradient cannot be backpropagated to the diffusion policy). Additionally, we tried to adaptively adjust the single-step denoising coefficient based on the estimated entropy during the reverse diffusion process, but the results were not as good as the current methods. ### > Limitation In Appendix B we discuss related Limitation and future work, which we will put in the main text in the official version. ### Reference [1] Zou et al. " Policy Bifurcation in Safe Reinforcement Learning," arXiv, 2403. [2] Xiao et al. "Multi-Style Distributional Soft Actor-Critic: Learning a Unified Policy for Diverse Control Behaviors," in IEEE TIV, 2024. --- Rebuttal Comment 1.1: Title: Rebuttal Acknowledged Comment: I thank the authors for answering my questions and I appreciate the clarifications	Summary: The paper presents a novel method using diffusion models as a policy parameterization for online reinforcement learning. The method works by learning a Q function and backpropagating through the reverse diffusion process in order to update the diffusion policy weights, similarly with Diffusion-QL. In order to aid exploration, the authors estimate the entropy of the policy and use that to update $\alpha$, which is used to adjust the noise added to actions after sampling. Strengths: - The paper's initial results are strong and I think with further experimentation it could be a strong paper - Few papers have studied the application of diffusion policies to online RL, and this paper presents a fairly straight-forward method in that direction which seems to achieve strong performance compared to baselines. - The presentation of the method is generally easy to understand - The use of GMM to estimate the entropy of the diffusion policy is a novel and creative trick Weaknesses: - A glaring issue with the paper is that the authors chose not to compare to any other papers using diffusion policies for online RL. They claim without citation that these papers perform worse than SAC although these papers claim to outperform SAC. In order to fully contextualize this paper I feel that these comparisons are essential. - As I discuss in the following section, several key results are missing. - The presentation is generally pretty sloppy as I discuss at several points in the following section. - I worry that the method may be very slow, and that the gains in performance may not be enough to justify the added computational cost. It would be useful to see some data about the comparative wall clock times of the algorithm and baselines. Technical Quality: 3 Clarity: 3 Questions for Authors: Random questions and comments 1. Section 2 paragraph 2, you frequently cite incompatibility / combinability with past methods or a lack thereof as an important feature of an algorithm, using this metric to claim a victory over other methods using diffusion for online. I disagree that this is a useful metric for evaluating algorithms and was wondering whether you could elaborate on the logic behind this claim. 2. Line 108 you claim that DIPO and QSM perform worse than SAC and thus you don't compare to them. However, both papers claim to outperform SAC so I don't see why you believe they perform worse. 3. The method seems to rely on the fact that a GMM can accurately fit your action distribution. Is there a reason you can't simply use that GMM for your policy? Seems to me that that would be more efficient and easy to implement. I think this is an important ablation to run to show that the full expressive power of the diffusion policy is actually useful in the settings you consider. 4. Figures 3-5 all have the same caption, "Trainining curves" [sic]. The caption should fully describe what is going on in the figure. 5. Your method takes 7 hours to train with the JAX implementation and apparently 4-5 times longer with PyTorch. How long do the baseline algorithms take to train? Assuming that they train for much less time, can you make a case that the extra time is worth it for generally modest performance improvements? 6. In Figure 3 what is DAC (I assume that this is the ablated version but you never state the meaning of that acronym)? It seems to outperform your method. Is the legend switched? 7. One of the key claims of your paper is that a diffusion policy is necessary due to its ability to accurately model multimodal distributions, and yet your main evaluation suite doesn't include any tasks that require this. You do have a toy problem where this is necessary, but I believe that additional experiments in an environment such as Push-T from Cheng et al. would help you to provide evidence for this claim. 8. Why do your plots show learning over iterations rather than the more common metric of the number of environment timesteps? 9. The core novelty of your method is using the GMM for adaptive noise injection, but I don't fully understand whether this is necessary. It would be nice to see how the algorithm performs with a wider range of fixed $\alpha$ values with and without linear decay, across several environments. If you still show a win over all of these it shows that the adaptive noise injection is necessary. Even if you tie with a few, you can still use this experiment to argue that the adaptive noise injection helps you to avoid an expensive hyperparameter tuning process. 10. $\alpha$ is overloaded. You use it as a variable in your mathematical description of the denoising process as well as the variable used to inject noise into the policy. 11. You argue that the diffusion model is important because it enables multimodal policies, and yet at the end of the day you add unimodal noise to add exploration. Have you looked into whether this prevents the policy from learning multimodal behaviors? 12. How did you tune the hyperparameters for your algorithm and for the comparisons? Nit picks: - The text in all of the plots is far too small - The training curves would be much easier to read if the curves were smoothed. - The caption of Table 1 written entirely in capital letters is hard to read. - It is almost impossible to distinguish PPO from DDPG in your plots. Confidence: 4 Soundness: 3 Presentation: 3 Contribution: 2 Limitations: The authors discuss the limitations of their work, although they do it in the appendix. Flag For Ethics Review: ['No ethics review needed.'] Rating: 5 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thank you for insightful review. We report below the changes we are currently making to address your comments. # Q1 Our work emphasizes proposing the reverse diffusion process as a novel policy expression, combinable with existing online RL algorithms. Section 2.2 highlights the method's portability. To prove it, global rebuttal Fig 4 includes extra experiments combining the diffusion policy with SAC. However, we feel that your suggestion is good and we modifiy the relevant statement in Section 2.2 to "Yang et al.[1] pioneered this approach by proposing action gradient method. This approach achieves policy improvement by updating the action in the replay buffer through the $\nabla_a Q $, followed by mimicry learning of the action post-update using a diffusion model. However, action gradient increased the additional training time. Furthermore, it is difficult to fully learn both the action gradient and imitation learning steps simultaneously, which also resulted in suboptimal performance of this method in MuJoCo tasks. Penka et al.[2] proposes Q-score matching (QSM), a new methodology for off-policy reinforcement learning that leverages the score-based structure of diffusion model policies to align with the $\nabla_a Q $. This approach aims to overcome the limitations of simple behavior cloning in actor-critic settings by integrating the policy's score with the Q-function's action gradient. However, QSM needs to accurately learn $\nabla_a Q $ in most of the action space to achieve optimal guidance. This is difficult to accomplish, resulting in suboptimal performance of QSM." # Q2 As you mentioned, comparing the performance with DIPO and QSM is meaningful. In fact, we have previously conducted relevant experiments. However, using the hyperparameters provided in their papers and testing within our JAX framework, we found that their performance did not surpass SAC as claimed. To avoid any misunderstanding that we intentionally lowered their performance, we chose not to include the experimental results. To address your concerns, we have added the results in the global rebuttal PDF Figure 3. We will proactively contact the authors to discuss why the performance does not exceed SAC. # Q3 First, using GMM as a policy requires pre-specifying the number of Gaussian components, which limits its ability to approximate any continuous distribution. Second, in DACER, entropy estimation updates the noise parameter $\alpha$, where some estimation error is acceptable. Finally, we conduct experiments replacing the diffusion policy with GMM (3 Gaussian components) in MuJoCo tasks, showing that performance is lower than that of the diffusion policy, as illustrated in Figure 4 of the global rebuttal PDF. # Q4 Thank you for your reminder. We have written the revisions for the titles of Figures 3-5 in global rebuttal 4. # Q5 Your concern about training time is understandable, but our starting point is to train in an offline simulation environment and then apply the trained network online. Therefore appropriate longer training time is acceptable if performance gains can be realised. We add some comparative experiments and the results are shown in Table 1 of the author global rebuttal. Fortunately, DACER's 1-batch inference time is a small value within 1 ms, which is acceptable in the real-time application. We find DACER's training time is 4-5 times longer than DSAC due to the longer chain gradient of the diffusion policy. Your question is insightful and aligns with our recent work. We plan to introduce an ODE method to eliminate this bottleneck. # Q6 Thank you for your careful discovery that the DAC and DACER legends were written backwards here, which we have corrected. # Q7 As you said, it's important to have an experiment that clearly characterizes the need for a multi-peak policy. We choose the vehicle bifurcation experimental task presented in [3] (like push-T) and plot the action distribution at the bifurcation point in Figure 5 of global rebuttal PDF. # Q8 For the sake of fairness in comparison, we maintained the same experimental setup and plotting standards as used in DSAC. In DSAC, each iteration represents one network update, and 20 samples are sampled per iteration. # Q9 Your proposal is strong. To show that adaptive noise injection can achieve optimal performance without parameter tuning, we conducted experiments in three environments, with training curves in Figure 1 of the global rebuttal PDF. The results indicate that while certain linear decay results can perform comparably to adaptive noise injection, most of the results are not as good as adaptive noise injection. # Q10 Thank you for pointing out that we will modify $\alpha$ in the mathematical description of the denoising process to $\omega$. # Q11&12 It does not prevent policy from learning multimodal behavior. First, the diffusion policy's multimodality comes from the reverse diffusion action generation, with added noise just being a little perturbation. Second, experiments show the distance between peaks is generally greater than 0.5, while the injected noise magnitude is within 0.272, decaying to within 0.15 as training progresses. For the second problem, we used hyperparameters from the DSAC paper for comparison algorithms. For DACER, the diffusion steps, $\lambda$, and $\alpha$ learning rate were consistently set at 20, 0.1, and 0.03 for nearly all experiments. # Picks Thanks to your suggestions, we will zoom in on the text in the image, add smoothing to the training curves (all curves in global rebuttle have been added), change the title of Table 1 to lowercase. ### Reference [1] Yang et al. "Policy representation via diffusion probability model for reinforcement learning," arXiv, 2305. [2] Penka et al. "Learning a diffusion model policy from rewards via q-score matching," arXiv, 2312. [3] Zou et al. " Policy Bifurcation in Safe Reinforcement Learning," arXiv, 2403. --- Rebuttal Comment 1.1: Title: Some concerns addressed but still worried about the benefits of the algorithm Comment: Thank you for your detailed response and for performing several additional experiments to address some of my concerns. Overall I am feeling more convinced but my concerns about the benefits of the algorithm still persist. In your setting of first training in simulation and then deploying on a real robot, there are two main metrics that are important to me as a researcher: wall clock training time and final converged performance. The number of training iterations really doesn't matter since simulated data is cheap and the real bottleneck is the time it takes to spit out a strong policy. At the same time, taking longer to spit out a policy is justified if the converged performance is much stronger. So far, none of the results you've shown prove to me that your method outperforms DSAC in either of these metrics. In fact, I'm inclined to believe that if you changed the x-axis in Figure 1 to display wall-clock time instead of the number of training iterations, DSAC would look a lot better than your method. How would the results look if everything were trained to convergence, or if, for example, you trained a well-implemented version of PPO for the same 7 hours as DACER? I think since your method is definitely a lot slower than comparisons I need to understand whether its converged behavior is strong enough to justify using it rather than a more established method. --- Reply to Comment 1.1.1: Title: Re: Concerns about the benefits of algorithms Comment: We thank you for the quick and detailed response to our rebuttal. --- We are pleased to see that some of the previous concerns have been addressed, and that the focus has shifted to the final performance gap after algorithm convergence, as well as the experimental results under the same wall time. Regarding the final performance after algorithm convergence, we trained DACER, DSAC, and PPO for 12 hours (Wall time) in the Humanoid-v3 and Ant-v3 environments. It is evident that all three algorithms have converged. Due to limitations imposed by OpenReview, we are unable to upload additional images or links, so we have provided the following table instead. Table 1 Average final return. Computed as the mean of the highest return values observed in the final 10% of iteration steps per run, with an evaluation interval of 15,000 iterations. The mean value for each task is bolded. ± corresponds to standard deviation over five runs. \| Task \| DACER \| DSAC \| PPO \| \| ----------- \| --------------- \| ----------- \| ---------- \| \| Humanoid-v3 \| 13209 ± 116 \| 11087 ± 193 \| 9134 ± 412 \| \| Ant-v3 \| 11470 ± 244 \| 9083 ± 122 \| 7645 ± 291 \| Thank you for your question, which allowed us to assess DACER's final performance after convergence. In the Humanoid-v3 and Ant-v3 tasks (the two most complex control tasks in MuJoCo), DACER outperformed DSAC by 2122 and 2387, respectively. This performance has been quite encouraging for us as well. As training steps increase, DSAC gradually converges. Therefore, under these conditions, even with the same Wall time, if the training time is sufficiently long, the performance ceiling of DACER is significantly higher than that of DSAC.	Summary: This paper proposes using the reverse diffusion process as the policy for actor-critic-based online reinforcement learning. An EM-based mixture model is fitted to estimate current diffusion policies' entropy to balance exploration and exploitation. The proposed method, DACER, demonstrated on-par or improved performance compared with RL with deterministic and single-mode Gaussian policies in various tasks. Strengths: 1. It is the first work that I know of directly using online RL to train diffusion models by directly back-propagating through the diffusion chain. 2. The proposed method performed on-par or better than conventional RL methods with deterministic or single-mode Gaussian policies in some continuous control tasks. 3. The writing of the paper is easy to follow, and the contributions are clearly outlined. Weaknesses: 1. My major concern is that the calculation of entropy using GMM seems not incorrect; Eq. 15 is not equivalent to the entropy of GMM presented in Eq. 12. GMM does not have a closed-form solution for its entropy [1]. However, it could be approximated in various ways. If the author used any of the approximations, the reference should be mentioned in the paper. Furthermore, in the ablation study (Figure 4) comparing DACER with linear decay entropy, the entropy regularization does not show much performance difference. Considering the extra compute overhead by applying EM in the inner loop, I’m not convinced it is worth the effort. 2. Another major concern of mine is that the proposed method seems to be very computationally heavy; for each policy update, the gradient needs to backpropagate through the entire diffusion chain. Furthermore, an inner-loop EM algorithm needs to be implemented to approximate the policy entropy. Although the training time for Humanoid-v3 is mentioned in section 5, there is a lack of comparison with other baselines implemented in the same framework (JAX) and the same hardware, which is necessary for understanding the extra computational cost. 3. DACER does not show significant performance improvement over DSAC in most environments except Humanoid-v3 and Ant-v3. Given the large computational overhead, it is not sure if it is really worth using the method in given tasks. 4. In Figure 1, the iteration of PPO and TRPO are reported in the number of network update steps, which is uncommon and needs proper justification, as the number of network updates does not necessarily reflect the number of simulation steps for these two algorithms. 5. I am also concerned that the policy representation experiments can not fully show the representation power of diffusion policy. Diffusions are models with rich representation power as they can capture multi-modal action distribution [2] at the same state. For the task with 2d state space presented in Fig. 2, the multi-modality in action distribution only exists in the state of (0, 0), which is also why single-mode policies like DSAC and TD3 could also partially solve the task. Although the result of DACER looks slightly better than DSAC, I’m not fully convinced that this result suggests a better representation power for diffusion policy. [1] Robin S, Scrucca L. Mixture-based estimation of entropy. Computational Statistics & Data Analysis. 2023 Jan 1;177:107582. [2] Jia X, Blessing D, Jiang X, Reuss M, Donat A, Lioutikov R, Neumann G. Towards diverse behaviors: A benchmark for imitation learning with human demonstrations. arXiv preprint arXiv:2402.14606. 2024 Feb 22. Technical Quality: 3 Clarity: 3 Questions for Authors: 1. Could the author explain why the value of PPO looks like a constant for all the states in Figure 2? More explanations on why PPO does not work in this task would also help the understanding. Confidence: 4 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: The computational complexity is the main limitation of this work. Although it has been mentioned in the appendix, I feel like more discussion and evaluations are needed to better understand this limitation. Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We thank you for the careful reading of our paper and constructive comments in detail. ### > Weakness 1 1. As you pointed out, the entropy of GMM does not have a closed-form solution. Eq.15 provides an upper bound for the entropy of GMM [1], which can be used for approximate estimation. We will update this reference accordingly. 2. In Algorithm 1, we propose calculating 1 EM every 10,000 steps, resulting in a computational overhead that can be ignored in terms of overall efficiency. 3. The reason you did not observe a significant performance difference is that the initial and final values of the linear decay are provided by the results of the entropy adjustment method. We add more experimental results for different linear decay range of $\alpha$ values. In Figure 1 of the author global rebuttal PDF, it can be observed that the best performance after parameter tuning only approaches that of our method, while the computational cost of parameter tuning is significantly higher. ### > Weakness 2 As you pointed out, the discussion on training and inference efficiency is very interesting. We acknowledge that DACER has longer training and inference time compared to DSAC. But our starting point is to train in an offline simulation environment and then apply the trained network online. Therefore appropriate longer training time is acceptable if performance gains can be realised (the Humanoid-v3 task with the longest training time in Mujoco took only 7 hours in JAX). We add experiments (Task: Ant-v3) to compare the inference and training time using the JAX framework on the AMD Ryzen Threadripper 3960X 24-Core Processor, NVIDIA 3090Ti. We use the results of 10 calculations to find the mean and standard deviation. The results are shown in Table 1 of the author global rebuttal 1. Fortunately, DACER's 1-batch inference time is a small value within 1 ms, which is acceptable in the real-time application. We also find that the training time of DACER is 4-5 times longer than that of DSAC. The bottleneck is the longer chain gradient of the diffusion policy. This question you presented is insightful, and it is actually what we are working on recently. We are going to introduce an ODE method to eliminate the bottleneck. We would advise to follow our future work. ### > Weakness 3 Based on the above response, the appropriate longer training time relative to DACER is acceptable. As you pointed out, our performance improvements are particularly significant in the most complex state-space tasks, Humanoid-v3 (376-dimensional state, 17-dimensional action) and Ant-v3 (111-dimensional state, 8-dimensional action), demonstrating DACER's strong potential. In other Mujoco tasks, DSAC has already approached the performance ceiling, while DACER still achieves obvious improvements of 276 and 440 in the complex Walker2d-v3 and Hopper-v3 tasks, respectively, and outperforms or matches DSAC in all other tasks. We believe these performance gains are substantial. ### > Weakness 4 For the sake of fairness in comparison, we maintained the same experimental setup and plotting standards as used in DSAC [2, 3]. In DSAC, each iteration represents one network update, and 20 samples are sampled per iteration. For the implementation of TRPO and PPO, we have also aligned with DSAC. In their paper, they collect 2,000 samples per iteration, with a mini-batch size of 10 and repeat_num set to 10. This setup achieves a network update corresponding to 20 samples. ### > Weakness 5 You may have misunderstood. The multimodality of the action distribution is not limited to the (0, 0) point; it also exists between adjacent peaks (on two diagonals). This is why, in the results of DSAC and TD3, there are issues with flat and poorly learned regions between adjacent peaks, as described in Section 1.1. To demonstrate the powerful multimodality of DACER, we add an experiment. We select five points requiring multimodal policies: (0.5, 0.5), (0.5, -0.5), (-0.5, -0.5), (-0.5, 0.5), and (0, 0). For each point, we sampled 100 trajectories. The trajectories are plotted in Figure 2 of the author rebuttal PDF. The results show that compared with DSAC, DACER exhibits strong multimodality. This also explains why only the Q-function of DACER can learn the nearly perfectly symmetrical four peaks. In addition to using Multi-goal task, we use the method from [4] to analyze the diversity of state trajectories for MuJoCo tasks, with results presented in Table 2 of the author global rebuttal 2. This indicator can measure the algorithm's ability to explore the state space. The results show that in the Humanoid-v3, Ant-v3, Hopper-v3 and Pusher-v2 tasks, DACER achieves the highest entropy value, surpassing DSAC, SAC, DDPG, TD3, PPO, and TRPO. ### > Question 1 During training, we found that PPO always remains around its initial value and does not improve reward performance. "Multi-goal" is an environment with very sparse rewards, where huge rewards are only provided at four symmetrical peaks. In this environment, the state initialization is mostly near (0, 0). We think that PPO's conservative policy update mechanism is the fundamental reason it performs poorly in this multimodal environment. ### Reference [1] M. F. Huber et al. "On entropy approximation for Gaussian mixture random vectors," IEEE MFI, 2008. [2] Duan et al. "Distributional soft actor-critic: Off-policy reinforcement learning for addressing value estimation errors," IEEE TNNLS, 2021. [3] Duan et al. "Dsac-t: Distributional soft actor-critic with three refinements," arXiv, 2310. [4] Xiao et al. "Multi-Style Distributional Soft Actor-Critic: Learning a Unified Policy for Diverse Control Behaviors," in IEEE TIV, 2024. --- Rebuttal 2: Comment: I appreciate the very detailed response and new experiments. However, my major concerns about the entropy calculation and the long training time remain: - Could the authors specify which equation from [1] was used to derive the upper bound referenced in Eq. 15? Additionally, please provide an explanation about when this upper bound will be tight. If possible, could the authors include the derivation of the upper bound in the response? - While I agree that longer training times are acceptable in simulation, a training period that is 4-5 times longer without significant performance improvement raises concerns about the algorithm's efficiency. - In my opinion, reporting the number of network updates is not a straightforward measure of the training cost, especially considering the authors' starting point is to 'train in an offline simulation environment and then apply the trained network online.' In this case, wall-clock time would be a better measure, as reviewer 3PmU also pointed out. - I appreciate the inclusion of new evaluations on the diversity of learned behavior. Could the authors provide the equation used to calculate the Information Entropy? Moreover, the advantages of DACER in Table 2 do not appear significant compared to other methods with single-mode Gaussian policy. [1] M. F. Huber et al. "On entropy approximation for Gaussian mixture random vectors," IEEE MFI, 2008. --- Rebuttal Comment 2.1: Title: (1/2) Re: Concerns about the entropy calculation and the long training time Comment: Thank you for your quick response, we work hard to address your two major concerns. ### > Question 1 You can directly see the approximation formula we used in Eq. 8 of [1], without the need for further derivation. Theorem 3 in [1] indicates that this approximation formula serves as an upper bound for the GMM entropy. We provide the proof process below: $H(x)=-\int _ {\mathbf{R}^{d}}\sum _ {i=1}^{K}\omega _ {i}\cdot\mathcal{N}(x;\mu _ {i},\Sigma _ {i}) \cdot \log\left(\sum _ {j=1}^{K}\omega _ {j}\cdot\mathcal{N}(x;\mu _ {j},\Sigma _ {j})\right)\mathrm{d}x$ $=-\sum _ {i=1}^{K}\omega _ {i}\int _ {\mathbf{R}^{d}}\mathcal{N}(x;\mu _ {i},\Sigma _ {i}) \cdot \log\left(\omega_i\cdot\mathcal{N}(x;\mu_i,\Sigma_i)\cdot(1+\epsilon_i)\right)\mathrm{d}x$ $=-\sum_{i=1}^{K}\omega _ {i}\int_{\mathbf{R}^{d}}\mathcal{N}(x;\mu _ {i},\Sigma _ {i}) \cdot \left(\log\left(\omega_i\cdot\mathcal{N}(x;\mu_i,\Sigma_i)\right)+\log(1+\epsilon_i)\right)\mathrm{d}x, $ where $\epsilon _ {i}=\frac{\sum_{i\neq j=1}^{K}\omega_{j}\cdot\mathcal{N}(x;\mu _ {j},\Sigma _ {j})}{\omega _ {i}\cdot\mathcal{N}(x;\mu _ {i},\Sigma _ {i})}.$ Since $\log(1+\epsilon_i)$ is always non-negative, neglecting it yields the desired upper bound. This approximation method has the following three main advantages [1]: 1. It is highly efficient in terms of computation; 2. The upper bound is significantly closer to the true entropy value than the well-known bound given by a single Gaussian matching the first two moments of the distribution $f(x)=\sum_{i=1}^{K}\omega_{i}\cdot\mathcal{N}(x;\mu _ {i},\Sigma _ {i})$; 3. The bound is exact for the single Gaussian case. After providing the content related to entropy, I hope your first major concern can be addressed. ### > Question 2 & 3 As you pointed out, only significant performance differences make longer training times meaningful, and it is more appropriate to use Wall time as the horizontal axis in this context. Reviewer 3PmU also emphasized that whether the performance after convergence is strong enough is the core factor determining whether DACER is worth using. Based on the aforementioned viewpoints, we have changed the horizontal axis to Wall time and trained DACER, DSAC, and PPO for 12 hours in the most complex MuJoCo tasks, Humanoid-v3 and Ant-v3. It is evident that all three algorithms have converged. Due to the limitations imposed by OpenReview, we are unable to upload additional images or links, so we have provided the following table instead. Table 1 Average final return. Computed as the mean of the highest return values observed in the final 10% of iteration steps per run, with an evaluation interval of 15,000 iterations. The mean value for each task is bolded. ± corresponds to standard deviation over five runs. \| Task \| DACER \| DSAC \| PPO \| \| ----------- \| --------------- \| ----------- \| ---------- \| \| Humanoid-v3 \| 13209 ± 116 \| 11087 ± 193 \| 9134 ± 412 \| \| Ant-v3 \| 11470 ± 244 \| 9083 ± 122 \| 7645 ± 291 \| Thank you for your question, which allowed us to assess DACER's final performance after convergence. In the Humanoid-v3 and Ant-v3 tasks, DACER outperformed DSAC by 2122 and 2387, respectively. This performance has been quite encouraging for us as well. As training steps increase, DSAC gradually converges. Therefore, under these conditions, even with the same Wall time, if the training time is sufficiently long, the performance ceiling of DACER is significantly higher than that of DSAC. In summary, given the significant improvement in final performance, a longer training time is acceptable. We are also actively researching a method to accelerate training using ODE, and we hope that our future work will bring you new insights. ### Reference [1] M. F. Huber et al. "On entropy approximation for Gaussian mixture random vectors," IEEE MFI, 2008. --- Reply to Comment 2.1.1: Title: (2/2) Re: Concerns about the entropy calculation and the long training time Comment: ### > Question 4 During the Rebuttal phase, we proactively reached out to the authors of [2]. Next, we first introduce the process and formula for calculating information entropy (all details are kept consistent with [2]). 1. Data collection. Using the trained policy, sample 1100 episodes, and for each episode, take the first 40 steps, recording the state of each step as a row in a CSV file. This results in a collection of 44,000 data points. 2. Data discretization. For each column in the CSV file (each state dimension), use `pd.cut` to divide the data into 11 different intervals. 3. State combination. The discretized data columns are treated as independent features. These features are then concatenated into a single string to represent a state. 4. Calculation of state frequency and information entropy. Encode the different combined states into an integer array, then calculate the frequency of each state. Finally, use formula $H(X)=-\sum_ip_i\log p_i$ to calculate the information entropy, where $p_i$ is the probability of the $i$-th combined state. We would like to mention that when calculating information entropy for any MuJoCo task using the method described above, each task will have its own approximate entropy range, and there will not be particularly significant differences in absolute values. Let’s take a look at the results for Humanoid-v3: compared to TRPO, DACER’s information entropy increased by 13.7%, which is already quite significant. You might still have some concerns, so we have applied the method from [2] (Fig. 6 of [2]) to normalize the results from Table 2 of the global rebuttal using the Z-score, resulting in the following table. Table 1 Normalized trajectory diversity entropy. The horizontal axis represents different algorithms, and the vertical axis represents different MuJoCo tasks. \| Task \| DACER \| DSAC \| SAC \| DDPG \| TD3 \| PPO \| TRPO \| \| ----------- \| --------- \| ----- \| ------ \| ------ \| ------ \| ------ \| ------ \| \| Humanoid-v3 \| 0.560 \| 0.511 \| 0.450 \| 0.486 \| -0.111 \| 0.495 \| -2.393 \| \| Ant-v3 \| 0.905 \| 0.676 \| 0.603 \| 0.441 \| -2.257 \| -0.219 \| -0.149 \| \| Hopper-v3 \| 1.778 \| 1.102 \| -1.321 \| -0.354 \| -0.698 \| 0.011 \| -0.517 \| \| Pusher-v2 \| 1.503 \| 0.056 \| 0.773 \| -0.301 \| -2.011 \| 0.049 \| -0.069 \| From the normalized results, it is evident that DACER exhibited higher entropy across all benchmarks compared to the baseline algorithms, indicating its effectiveness in generating a diverse range of behaviors from the diffusion policy. Based on the relative results, the information entropy of DACER is obviously higher than that of the baseline algorithms, suggesting that DACER has a stronger exploration capability. ### Reference [2] Xiao et al. "Multi-Style Distributional Soft Actor-Critic: Learning a Unified Policy for Diverse Control Behaviors," in IEEE TIV, 2024.	Summary: The authors propose DACER (Diffusion Actor-Critic with Entropy Regulator), an online RL algorithm that uses the reverse diffusion process as a policy in order to capture multimodal behaviours. To balance exploration and exploitation, the authors propose carrying out entropy regularization by estimating the entropy of the diffusion policy using a Gaussian mixture model. The paper demonstrates the proposed method on a set of simulated control tasks, comparing against a suite of established RL algorithms. Strengths: - To the best of my knowledge, the method proposed in the paper is a novel combination of new techniques – online actor-critic methods and diffusion models. - The paper also discusses relevant offline and online RL algorithms that make use of diffusion policies for control, clarifying the differentiating factors of this work. - The paper compares against a suite of well-established RL algorithms on a range of simulated control tasks (ranging from simple to complex). The experimental results show that the proposed method generally leads to improvements in learned policy performance, with additional studies showing the multimodality of learned policies and ablations on each component of the proposed method. - Learning multimodal policies is of significance to the RL community, and the proposed parametrization of entropy regularization for diffusion policies seems likely to be built on in subsequent work. The authors release relevant code, enabling future work to build upon this paper. Weaknesses: - The related works section would be strengthened by adding in references to online RL algorithms used for image generation (e.g. "Training Diffusion Models with Reinforcement Learning" (Black et. al, 2023), or "DPOK: Reinforcement Learning for Fine-tuning Text-to-Image Diffusion Models" (Fan et al., 2023)). While these methods are applied in an image generation setting rather than direct control, the methods for policy improvement are related (i.e. optimizing for maximum reward with a policy based on the reverse diffusion process, learned by backpropagating through the diffusion chain), with some differences (not using actor-critic methods, learning a q-value, etc.). - It would be more meaningful to see Table 1 as the mean of returns in the final 10% of iterations, rather than mean of highest returns, the latter which may be skewed by luck/noise in the evaluation. - It would be interesting to see some measures of how computationally efficient the method is – since it requires stepping sequentially through the diffusion process to sample an action at each step. What are the tradeoffs? - Rather than method iterations, the results in Figure 1 could be better contextualized by also looking at environment steps on the x-axis (i.e. how sample efficient is each method), since the meaning of “iteration” can vary per method. - Minor note: The definitions of eq. (1) and (2) could be written more clearly (i.e. how do (s,a) affect the rewards on the right hand side of the equation?) - Minor note: for clarity, the definitions of $\mathcal{L}_\pi$ and $\mathcal{L}_q$ should be stated. - Minor note: figures 3,4,5 should have more descriptive captions. Technical Quality: 3 Clarity: 2 Questions for Authors: See suggestions and questions in the Weaknesses section. Confidence: 3 Soundness: 3 Presentation: 2 Contribution: 3 Limitations: The authors should include a section on the limitations of the proposed methods (e.g. computational inefficiencies?) Flag For Ethics Review: ['No ethics review needed.'] Rating: 5 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We thank you for the careful reading of our paper and constructive comments in detail. ### > Weakness 1 Thank you for your important information. We have added a paragraph to introduce image generation with online RL algorithms. "Diffusion models are widely applied in the field of image generation [1, 2]. Recently, several studies have employed online RL algorithms to fine-tune these models. Black et al. [3] propose an RL-based method for fine-tuning large-scale diffusion models to match potentially non-differentiable reward functions. This approach views the reverse process of diffusion models as a reinforcement learning task with time steps T, where the output of the diffusion model at each step serves both as an action and the observation for the next step. Fan et al. [4] proposed DPOK, a RLHF method for diffusion models in text-to-image generation. Given a reward model, the goal is to fine-tune the parameters of the diffusion model so that the resulting images from the sampling process achieve high rewards. The reward is a function of the final output of the entire diffusion process and, therefore, not be optimized independently across timesteps. The above work utilizes the reverse diffusion process to optimize the maximum reward, but does not learn the Q-function, as well as using the classical Actor-Critic framework." ### > Weakness 2 It can be seen that our core comparison algorithm is DSAC. To be consistent with it, we use the highest value in 10% of the iterations. It is worth noting that each point in the experiment is the average of 10 episodes. We also set 5 different random seeds in each experiment. The results in Table 1 are the average and standard deviation of the results of the 5 random seeds. These settings avoid the influence of luck and noise to a certain extent. But since you have mentioned it, we also give the mean and standard deviation of the 5 seed results averaged over the last 10% of iterations, which are shown in Table 3 of the author global rebuttal 3. ### > Weakness 3 As you pointed out, the discussion on training and inference efficiency is very interesting. We acknowledge that DACER has longer training and inference time compared to DSAC. But our starting point is to train in an offline simulation environment and then apply the trained network online. Therefore appropriate longer training time is acceptable if performance gains can be realised (the Humanoid-v3 task with the longest training time in Mujoco took only 7 hours in JAX). We add experiments (Task: Ant-v3) to compare the inference and training time using the JAX framework on the AMD Ryzen Threadripper 3960X 24-Core Processor, NVIDIA 3090Ti. We use the results of 10 calculations to find the mean and standard deviation. The results are shown in Table 1 of the author global rebuttal 1. Fortunately, DACER's 1-batch inference time is a small value within 1 ms, which is acceptable in the real-time application. We also find that the training time of DACER is 4-5 times longer than that of DSAC. The bottleneck is the longer chain gradient of the diffusion policy. This question you presented is insightful, and it is actually what we are working on recently. We are going to introduce an ODE method to eliminate the bottleneck. We would advise to follow our future work. ### > Weakness 4 For the sake of fairness in comparison, we maintained the same experimental setup and plotting standards as used in DSAC. In DSAC, each iteration represents one network update, and 20 samples are sampled per iteration. For the implementation of TRPO and PPO, we have also aligned with DSAC. In their paper, they collect 2,000 samples per iteration, with a mini-batch size of 10 and repeat_num set to 10. This setup achieves a network update corresponding to 20 samples. For env steps, you need to multiply the horizontal axis by 20. ### > Minor note 1 We modify Eq.1 to $J _ \pi=\mathbb{E} _ {(s _ {i\geq t},a _ {i\geq t})\sim\pi}\Big[\sum _ {i=t}^{\infty}\gamma^{i-t}r(s_i, a_i)\Big]$ Eq.2 to $Q(s,a)=\mathbb{E} _ \pi\Big[\sum _ {i=0}^\infty\gamma^ir(s_i, a_i)\|s_0=s,a_0=a\Big]$. ### > Minor note 2 $\text{min}\mathcal{L} _ {q}(\phi) = \mathbb{E} _ {({s},{a},{s}^{\prime})\sim\mathcal{B}}\left[\left(r(s,a) + \gamma \mathbb{E} _ {s' \sim p, a' \sim \pi}[Q^{\pi}(s',a')]-Q^{\pi} _ {\phi}(s, a)\right)^{2}\right]$ $\text{min}\mathcal{L} _ {\pi_{}}(\theta) = -\mathbb{E} _ {s \sim d _ {\pi},a \sim \pi }[Q^{\pi _ {\mathrm{old}}}(s,a)]$ ### > Minor note 3 Thanks to your reminder, we modify the figure notes of Figure 3-5 respectively as "Ablation training curves for the Entropy Regulator mechanism. DAC stands for not using the Entropy Regulator. DACER's performance on Walker2d-v3 is far better than DAC." "Ablation experiment curves for the Noise factor modulation mechanism. Adaptive tuning of the noise factor based on the estimated entropy achieved the best performance compared to fixing the noise factor or using the adaptive tuning method with initial, end values followed by a linear decay method." "Ablation experiment curves for the different diffusion steps. The best performance was achieved with diffusion steps equal to 20, in addition to the instability of the training process when equal to 30." ### > Limitation In Appendix B we discuss related Limitation and future work, which we will put in the main text in the official version. ### Reference [1] Xie et al. "Scalable diffusion models with transformers," ICCV, 2023. [2] Dhariwal et al. "Diffusion models beat gans on image synthesis," NIPS, 2021. [3] Black et al. "Training Diffusion Models with Reinforcement Learning," ICLR, 2024. [4] Fan et al. "DPOK: Reinforcement Learning for Fine-tuning Text-to-Image Diffusion Models," NIPS, 2023. --- Rebuttal Comment 1.1: Title: Fw: share good news with reviewer xVKC Comment: I would like to share some good news with you: we have successfully trained DACER, DSAC, and PPO to convergence on the most complex tasks in MuJoCo, namely Humanoid-v3 and Ant-v3, and have obtained the following results. Table 1 Average final return. Computed as the mean of the highest return values observed in the final 10% of iteration steps per run, with an evaluation interval of 15,000 iterations. The mean value for each task is bolded. ± corresponds to standard deviation over five runs. \| Task \| DACER \| DSAC \| PPO \| \| ----------- \| --------------- \| ----------- \| ---------- \| \| Humanoid-v3 \| 13209 ± 116 \| 11087 ± 193 \| 9134 ± 412 \| \| Ant-v3 \| 11470 ± 244 \| 9083 ± 122 \| 7645 ± 291 \| In the Humanoid-v3 and Ant-v3 tasks, DACER outperformed DSAC by 2122 and 2387, respectively. This performance has been quite encouraging for us as well. As training steps increase, DSAC gradually converges, the performance ceiling of DACER is significantly higher than that of DSAC. We firmly believe that this cutting-edge work, which combines diffusion models with online RL, will generate significant interest within the RL community. We look forward to your response.	Rebuttal 1: Rebuttal: Overall author rebuttal: We thank all reviewers for their thoughtful comments. We greatly appreciate all the reviewers' acknowledgment that our method is very novel and is the cutting-edge work combining diffusion model and online RL. In addition, our method achieved SOTA in MuJoCo tasks and is very easy to follow. To address your concerns about the training and inference time, the lack of comparative results with DIPO and QSM, the wide range of noise factor ablation experiments, the performance of using GMM directly as a policy, and how multimodality is represented and its importance, we add new experiments and evaluations: ### 1. Experiments with training and inference time We add experiments (Task: Ant-v3) to compare the inference and training time using the JAX framework on the AMD Ryzen Threadripper 3960X 24-Core Processor, NVIDIA 3090Ti. We use the results of 10 calculations to find the mean and standard deviation. Inference time represents the time required for the policy to generate actions. Backward time represents the time required for gradient propagation. (It should be noted that the backward time of DIPO consists of two parts: gradient propagation and action gradient. Besides, the diffusion steps of DACER and DIPO are both 20, and the action gradient steps of DIPO are 20.) Table 1 Time statistics results. Under the same hardware conditions, the inference and backward time of the three algorithm policy functions DACER, DIPO and DSAC are compared in the Ant-v3 task with different batches. \| Mode \| Batch \| DACER \| DIPO \| DSAC \| \|-----------------\|-------\|---------------------------\|-----------------------------------------------\|---------------------------\| \| Inference time \| 1 \| 0.472ms ± 0.006ms \| 0.502ms ± 0.011ms \| 0.188ms ± 0.027ms \| \| Inference time \| 100 \| 0.968ms ± 0.004ms \| 0.970ms ± 0.006ms \| 0.190ms ± 0.020ms \| \| Backward time \| 1 \| 1.283ms ± 0.036ms \| 0.582ms ± 0.020ms (gradient propagation); 0.801ms ± 0.008ms (action gradient) \| 0.571ms ± 0.038ms \| \| Backward time \| 100 \| 2.428ms ± 0.028ms \| 0.601ms ± 0.021ms (gradient propagation); 1.483ms ± 0.006ms (action gradient) \| 0.602ms ± 0.015ms \| ### 2. Experiment on evaluating the performance of algorithm exploration. To assess the multi-style capability of DACER, we use the method proposed in [1]. This method analyze the diversity of state trajectories by computing the information entropy from 1,100 independent simulation episodes. Table 2 Information entropy results. The results show that in the Humanoid-v3, Ant-v3, and Hopper-v3 tasks. DACER achieves the highest entropy value. \| Mode \| DACER \| DSAC \| SAC \| DDPG \| TD3 \| PPO \| TRPO \| \| ----------- \| ------ \| ------ \| ------ \| ------ \| ------ \| ------ \| ------ \| \| Humanoid-v3 \| 16.743 \| 16.711 \| 16.668 \| 16.693 \| 16.284 \| 14.699 \| 14.721 \| \| Ant-v3 \| 16.821 \| 16.716 \| 16.683 \| 16.609 \| 15.376 \| 16.307 \| 16.339 \| \| Hopper-v3 \| 12.453 \| 11.822 \| 9.559 \| 10.462 \| 10.141 \| 10.803 \| 10.310 \| \| Pusher-v2 \| 10.722 \| 9.719 \| 10.216 \| 9.471 \| 8.286 \| 9.714 \| 9.632 \| ### 3. Average return of the last 10% of iterations averaged over the 5 seeds. Table 3 Average final return. Computed as the mean of the average return values observed in the final 10% of iteration steps per run, with an evaluation interval of 15,000 iterations. The mean value for each task is bolded. ± corresponds to standard deviation over five runs. \| Task\|DACER\|DSAC\|SAC\|TD3\|DDPG\|TRPO\|PPO\| \| ------------------------- \| ----------- \| ----------- \| ----------- \| ----------- \| ------------ \| ----------- \| ----------- \| \| Humanoid-v3\| 10896 ± 306 \| 9630 ± 469 \| 7515 ± 399 \| 5305 ± 277 \| 3558 ± 848 \| 744 ± 333 \| 5398 ± 2211 \| \| Ant-v3\| 7584 ± 211 \| 6643 ± 324 \| 6208 ± 1412 \| 5772 ± 386 \| 3474 ± 784 \| 5851 ± 537 \| 5908 ± 247 \| \| Halfcheetah-v3\| 16489 ± 276 \| 16182 ± 779 \| 15884 ± 355 \| 7775 ± 3627 \| 11710 ± 3108 \| 4586 ± 1007 \| 5503 ± 2084 \| \| Walker2d-v3\| 6496 ± 134 \| 6171 ± 301 \| 5364 ± 817 \| 4785 ± 104 \| 2577 ± 266 \| 4708 ± 732 \| 3634 ± 573 \| \| Inverteddoublependulum-v3 \| 9360 ± 0 \| 9360 ± 0 \| 9360 ± 0 \| 9311 ± 11 \| 8983 ± 11 \| 5852 ± 2335 \| 9355 ± 2 \| \| Hopper-v3\| 3679 ± 362 \| 3111 ± 616 \| 2147 ± 621 \| 2909 ± 510 \| 1850 ± 275 \| 2805 ± 404 \| 1932 ± 290 \| \| Pusher-v2\| -20 ± 0 \| -21 ± 1 \| -23 ± 2 \| -25 ± 4 \| -37 ± 2 \| -26 ± 5 \| -24 ± 1 \| \| Swimmer-v3\| 146 ± 2 \| 135 ± 6 \| 135 ± 14 \| 114 ± 26 \| 140 ± 4 \| 69 ± 38 \| 129 ± 1 \| ### 4. Changes to the title of Figure 3-5. "Ablation training curves for the Entropy Regulator mechanism. DAC stands for not using the Entropy Regulator. DACER's performance on Walker2d-v3 is far better than DAC." "Ablation experiment curves for the Noise factor modulation mechanism. Adaptive tuning of the noise factor based on the estimated entropy achieved the best performance compared to fixing the noise factor or using the adaptive tuning method with initial, end values followed by a linear decay method." "Ablation experiment curves for the different diffusion steps. The best performance was achieved with diffusion steps equal to 20, in addition to the instability of the training process when equal to 30." ### Reference [1] Xiao et al. "Multi-Style Distributional Soft Actor-Critic: Learning a Unified Policy for Diverse Control Behaviors," in IEEE TIV, 2024. Pdf: /pdf/d97c6b38cd7548a574a0be32c73007b386093ab9.pdf	NeurIPS_2024_submissions_huggingface	2,024	null	null	null	null	null	null	null	null
Bandits with Abstention under Expert Advice	Accept (poster)	Summary: This paper considers the problem of prediction with expert advice in the bandit feedback setting with the possibility of abstention. The game is sequential: where in each round the each of $E$ experts gives a distribution over $K+1$ actions/arms. One of these actions represents abstention, yielding a reward of 0, while the remaining $K$ actions have rewards in $[-1,1]$ and are set by an oblivious adversary. The contributions of this paper include a procedure called confidence-rated bandits with abstention (CBA), which is an adaptation of the well-known exponential weighted scheme. The authors present guarantees in the form of an upper bound on the expected cumulative regret. Additionally, they explore applications of the resulting algorithm in various other problems. Strengths: The problem under consideration is both interesting and practically relevant, as the ability for a prediction strategy to abstain is crucial in many scenarios, particularly when the risk of making poor decisions is high. Additionally, accounting for the context is a significant and pertinent feature of modern applications. Weaknesses: Overall, I find this paper unclear and difficult to read. The results obtained are not easily comparable to existing ones in the literature. For instance, in the case of regret minimization with expert advice, such as the analysis of the EXP4 algorithm, regret is typically defined with respect to the best fixed expert in hindsight. The fact that the authors develop guarantees with respect to the best combination of experts (where this combination is not necessarily convex, since the set of vectors $u \in \mathcal{V}$, and $\mathcal{V}$ as defined in line 123 is not the set of convex weights) can be very confusing for the reader. The authors discuss comparisons with the guarantees of the EXP4 algorithm and state that "The EXP4 bound essentially replaces the term $\rho(u)$ in our bound by $\rho(u)/u$" in line 138. However, it is unclear how this translates in terms of worst-case guarantees. This part clearly needs a thorough discussion to be convincing. Additionally, it is not clear how this work compares to previous research on online learning with abstention. Technical Quality: 2 Clarity: 2 Questions for Authors: How do your results compare to [1]? [1] Gergely Neu and Nikita Zhivotovskiy. Fast rates for online prediction with abstention. In Proc. COLT, pages 3030–3048, 2020. Confidence: 2 Soundness: 2 Presentation: 2 Contribution: 2 Limitations: See weaknesses section. Flag For Ethics Review: ['No ethics review needed.'] Rating: 4 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thank you for your time and effort in reviewing our paper. > The results obtained are not easily comparable to existing ones in the literature. Our results can be, for example, compared to EXP4 (see reply below) and SpecialistEXP (see e.g., our reply to reviewer Edre). In both cases we achieve at least the same bounds and in many cases achieve significantly better bounds. We will improve our exposition and a add a more thorough discussion and in the final version. > For instance, in the case of regret minimization with expert advice, such as the analysis of the EXP4 algorithm, regret is typically defined with respect to the best fixed expert in hindsight. The fact that the authors develop guarantees with respect to the best combination of experts (where this combination is not necessarily convex, since the set of vectors , and as defined in line 123 is not the set of convex weights) can be very confusing for the reader. The authors discuss comparisons with the guarantees of the EXP4 algorithm and state that "The EXP4 bound essentially replaces the term in our bound by $\rho(u)/u$" in line 138. However, it is unclear how this translates in terms of worst-case guarantees. This part clearly needs a thorough discussion to be convincing. You are right in that both the reward of our algorithm and that of Exp4 are measured with respect to a linear combination of experts and for Exp4 that linear combination is required to be convex (i.e. the weights have to sum up to $1$). Typically the weights of our linear combination sum to some value $S$ far greater than $1$ (and never less than $1$). Hence (as the regret term limits to $0$) our reward is $S$ times that of Exp4. We will add this discussion to the paper for clarity. The fact that “Exp4 bound essentially replaces the term in our bound by $\rho(u)/\|u\|_1$” is because Exp4 plays with the weight vector $u$ normalized at each time step and therefore operates in a stricter space. >The results obtained are not easily comparable to existing ones in the literature. [...] Additionally, it is not clear how this work compares to previous research on online learning with abstention. How do your results compare to [1]? > [1] Gergely Neu and Nikita Zhivotovskiy. Fast rates for online prediction with abstention. In Proc. COLT, pages 3030–3048, 2020. We note that, to the best of our knowledge, there are no previous results on adversarial contextual bandits with the abstention option. The main novelty of the paper lies in merging the abstention action with confidence-rated predictors. This allows us to extend our result to (adversarial) contextual bandits with the abstention option. We introduce the first cumulative reward bounds on confidence-rated experts and we investigate contextual bandits as a potential application. Note that the cumulative reward bound we obtain is not achievable by previous work on confidence rated experts as they scale the regret by the confidence of the best expert in hindsight. We discuss previous results on online learning with abstention in the main body, but the previously considered settings are different from the ones we consider. Moreover, previous methods cannot be directly applied to confidence-rated predictors. Specifically, our results cannot be directly compared to [1], as their learning setting is different; it is a full information (in contrast to our bandit setting) binary classification problem. --- Rebuttal Comment 1.1: Comment: Thank you for your rebuttal; I have read your response.	Summary: This paper investigates the problem of prediction with expert advice in contextual adversarial bandits, introducing the CBA (Confidence-rated Bandits with Abstentions) algorithm, they apply CBA to adversarial contextual bandits and achieve near-optimal upper regret bound. Strengths: The CBA in adversarial contextual bandits achieves near-optimal upper regret bound. Weaknesses: Weakness 1: This algorithm is not as efficient as Exp4. Weakness 2: In the algorithm, they mentioned using interval bisection to find $\lambda$, but the regret upper bound does not include the approximating error of interval bisection. Technical Quality: 3 Clarity: 3 Questions for Authors: Q1: Why not consider the reward of the abstention action to be non-negative instead of zero? Q2: What is the novelty in the proofs? Confidence: 3 Soundness: 3 Presentation: 3 Contribution: 2 Limitations: The algorithm is not as efficient as Exp4 and the upper bound does not include the approximation error of $\lambda$. Flag For Ethics Review: ['No ethics review needed.'] Rating: 5 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thank you for your time and effort in reviewing our paper. >Weakness 1: This algorithm is not as efficient as Exp4. Given that the numerical precision of our machine is bounded, our algorithm has exactly the same $\mathcal{O}(EK)$ efficiency as Exp4. See the additional result presented in the general rebuttal for our time complexity (equal to that of Exp4 up to logarithmic factors) when the precision is unbounded. >Weakness 2: In the algorithm, they mentioned using interval bisection to find $\\lambda$, but the regret upper bound does not include the approximating error of interval bisection. We took the assumption that the degree of precision was bounded (the interval bisection method requires $\mathcal{O}(\ln(P))$ steps where $P$ is the precision. Thanks to your question, we attached a proof in the general rebuttal that removes this assumption. When the magnitudes of the components of the comparator u are bounded by some arbitrary value Z (which will typically be very small) we can achieve a time complexity of $\mathcal{O}(EK+E\ln(EZT))$ whilst only adding an additive factor of 1 (which can be made arbitrarily small) to the regret. >Q1: Why not consider the reward of the abstention action to be non-negative instead of zero? We can have any range $[-a,b]$ with any $a,b > 0$ and the reward of the abstention action can be any fixed constant. This can be achieved by scaling and translating the reward (on any trial) when received - bringing it into the constraints of our paper (hence Thm 3.1 still applies but the regret term is scaled by some constant in $\mathcal{O}(a+b)$). Such a scaling approach is standard in online learning. >Q2: What is the novelty in the proofs? Our proof borrows ideas from the analysis of mirror descent, EXP3, and the method of Lagrange multipliers (also with unbounded precision (see general comment)) . However, we note that the main novelty of our work was the realization that making such a modification to mirror descent would lead to true reward bounds for confidence-rated predictors. We note that mirror descent itself would not work as the constraint set is not known until after the last trial (and, even if it was known, computing the projections would be computationally hard). --- Rebuttal Comment 1.1: Comment: I have read your response and will keep my positive score.	Summary: The paper considers the problem of prediction advice under the bandit framework. There are $K$ arms plus a special $(K+1)$0-th arm, which always incurs the gain of 0; this arm may be interpreted as the action of abstaining. The learner outputs a distribution over $K+1$ handles and earns the scalar product gain. In other terms, the learner outputs a vector of $K$ non-negative components summing to a number between 0 and 1; the sum can be though of as the measure of its confidence. For this setup, the problem of prediction with expert advice is considered and a bound is obtained. There are two terms in the regret, one involves KL-divergence (there is a class of prediction with expert advice bounds of this kind) and the other has the form $O(\eta KT)$. It is seemingly linear, but $\eta$ can be tuned to get sublinear growth. The result has been applied to the framework with side information $x_t$. Each expert is associated with a pair $(B,k)$. Whenever $x_t\in B$, the expert bets everything on arm $k$ and is abstaining otherwise. The main result of the paper is applied to this scenario with finitely many experts based on disjoint balls and a matching lower bound is obtained. The case where $B$s are balls in a metric space is considered separately. Strengths: I think this is a very interesting result in prediction with expert advice. Weaknesses: No obvious weaknesses. Technical Quality: 4 Clarity: 4 Questions for Authors: None Confidence: 4 Soundness: 4 Presentation: 4 Contribution: 4 Limitations: Yes. Flag For Ethics Review: ['No ethics review needed.'] Rating: 8 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thanks for the very encouraging feedback! --- Rebuttal Comment 1.1: Comment: Thank you. This is to acknowledge the response.	Summary: This paper first considers the problem of bandits with expert advice, allowing the learner to abstain in any round. It introduces an algorithm, CBA, based on mirror descent, which can achieve a better cumulative regret bound in comparison with EXP4. The algorithm is then applied to adversarial contextual bandits, where the learner abstains in rounds where the given context is not covered by any expert. The paper provides upper and lower bounds on the cumulative reward, and details a more computationally-efficient implementation for settings where the contexts have inherent structure. Strengths: - Online learning with abstention is an interesting and relevant problem for the community, and the extension of CBA to adversarial contextual bandits with abstention seems novel. - The paper presents reward bounds for both the bandits with expert advice problem and the adversarial contextual bandit problem. These bounds can improve upon those of existing algorithms. A lower bound is also provided for the latter setting. - The experiments offer some insights into the construction of basis functions for contextual bandits with abstention. Weaknesses: - The paper can be quite difficult to follow at times. For instance, in Section 5, the discussion about covering foreground and background classes seems very abrupt. I understand this is mentioned in the introduction, but since the problem settings are formally introduced until later, the initial discussion is also confusing. In addition, the purpose of the experiments remains somewhat unclear to me -- they are not properly discussed, and there is this sudden change of theme from learning in bandits with abstentions to constructing basis functions for various graphs. I also find the contextualization relative to prior work somewhat inadequate -- there are not enough details about the SpecialistExp algorithm to make the comparison concrete (lines 215 - 223). See also questions below. - The improvement of CBA over EXP4 is somewhat vague and seems to largely depend on the set of experts. It would be nice if some concrete examples are provided. - There is an $\mathcal{O}(\ln \|\mathcal{B}\|)$ gap between the upper and lower bounds in Section 5, which is substantial as $\|\mathcal{B}\|$ can be exponential in the size of the set of contexts. - There seems to be a lack of discussion on the intuition behind why CBA should be applied for adversarial contextual bandits with abstention in this particular way. Technical Quality: 3 Clarity: 2 Questions for Authors: - Following Theorem 3.1, can you provide a concrete example where $\|\| u \|\|_1 = E$. The example given in the footnote seems to require all other experts to have confidence $0$ in the round where the expert has confidence $1$. What conditions need to be imposed on the set of experts? - In Corollary 5.1, can $M$ be any number in $\mathbb{N}$ for the sequence of basis elements to be disjoint? - How are the corresponding actions $b_j$ chosen? - In Section 5, what are the specific benefits of allowing the learner to abstain? Confidence: 2 Soundness: 3 Presentation: 2 Contribution: 2 Limitations: It would be nice if the authors include a separate section/paragraph that discuss the limitations in detail. Flag For Ethics Review: ['No ethics review needed.'] Rating: 5 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thank you for your time and effort in reviewing our paper. >[...] $\log\|B\|$ gap and $\|B\|$ can be exponential in the number of contexts $N$. While the comment is true, $\|B\|$ can indeed be as large as $2^N$ in principle, such a large basis set would defy its purpose (as an inductive prior on the shape of the clusters). All the basis sets we consider have size $\|N\|^2$, thus, the $\log\|B\|$ gap is just logarithmic in $N$. Generally, we think of $B$ as a set of “simple” stereotypical sets, which, for example, can be quantified by assuming $B$ has finite VC-dimension. In that case $\log\|B\|=\mathcal{O}(\mathrm{vc}(B)\log N)$ the dependence is still logarithmic in $N$ by the Sauer lemma. >[...] there is this sudden change of theme from learning in bandits with abstentions to constructing basis functions for various graphs [...] There seems to be a lack of discussion on the intuition behind why CBA should be applied for adversarial contextual bandits with abstention in this particular way. A natural application of experts, and therefore confidence-rated experts, is adversarial contextual bandits with a finite context space. In this setting, it is quite common in the literature (e.g., [1], [2]) to consider a graph structure over the contexts. Therefore, we decided to study a practical application of confidence-rated experts, where the experts, in our case, are the basis elements we introduce. We will improve the exposition here in the final version. [1] Cesa-Bianchi et al. "A gang of bandits." NeurIPS 2013. [2] Wu, Qingyun, et al. "Contextual bandits in a collaborative environment." SIGIR 2016. > [...] there are not enough details about the SpecialistExp algorithm to make the comparison concrete (lines 215 - 223). See also questions below. We shall add the regret bound of SpecialistExp to the paper. The difference is that while our regret is $\mathcal{O}(\sqrt{MKT\ln(N)})$ where $M$ is the number of disjoint basis elements that exactly cover each of the foreground (non-abstention) classes, the regret of SpecialistExp is $\mathcal{O}(\sqrt{M’KT\ln(N)})$ where $M’$ is the number of disjoint basis elements that cover each of the classes (including the background/abstention class). As depicted in Figure 1, $M’$ can be much larger than $M$. > In Section 5, what are the specific benefits of allowing the learner to abstain? The benefits of the abstention action are clear when dealing with contexts (and therefore nodes in graph-structured contexts) that do not have a behavior that reflects the inductive bias. In Figure 1 of the provided example, you can think of any point in white as being colored in any possible way, but we would need a ball that covers any of these points as in SpecialistsEXP. Therefore, abstaining avoids having to deal with points that present “unnatural” behavior with respect to our inductive bias. > Following Theorem 3.1, can you provide a concrete example where $\|\|u\|\|_1=\|E\|$. The example given in the footnote seems to require all other experts to have confidence 0 in the round where the expert has confidence 1. What conditions need to be imposed on the set of experts? Assuming each expert has confidence 1 on at least one trial (otherwise $\|\|u\|\|_1$ could be much higher) then $\|E\|$ is the maximum possible value of $\|\|u\|\|_1$. The condition for having such a value of $\|\|u\|\|_1$ is that the sum of the confidences of all experts are, on any trial, no greater than $1$. Such a $u$ is not that interesting - we place equal weight on each expert - but it was introduced in order to show just how high $\|\|u\|\|_1$ can be, so that the potential degree of improvement over Exp4 is huge. Things are much more interesting at slightly lower values of $\|\|u\|\|_1$. > In Corollary 5.1, $M$ can be any number $\mathbb{N}$ in for the sequence of basis elements to be disjoint? $M$ corresponds to the number of basis elements in the comparator. Hence, as we assume the basis elements to be disjoint in Corollary 5.1., $M$ is only restricted by the size of the number of contexts $\|\mathcal{X}\|$ and the size of the basis $\|\mathcal{B}\|$. We will clarify this in the final version. > How are the corresponding actions $b_j$ chosen? The $b_j$ are simply the actions in the comparator sequence. The result holds for an arbitrary such sequence. --- Rebuttal Comment 1.1: Comment: Thank you for the responses. I updated my confidence level to 2, and intend to maintain my rating.	Rebuttal 1: Rebuttal: We thank all reviewers for their feedback. We are happy to provide an additional result that was prompted by a question from Reviewer 3cAo. The reviewer asked how the regret bound and efficiency are affected if $\lambda$ is not exactly determined but only approximated (we previously assumed bounded precision so we did not address this). We will now show that, even in this case, we incur a time complexity equal to that of Exp4, up to logarithmic factors, adding only 1 to the regret. This additive factor, however, can be made arbitrarily small. Let us restrict ourselves to compare against $\boldsymbol{u}$ with $u_i\leq Z$ for some arbitrary $Z$. Note that this always has to be the case when each expert has confidence of at least $1/Z$ on some trial. Due to this restriction, we can always project, at the beginning of trial $t$, $\boldsymbol{w}_t$ into the set {$\{ \boldsymbol{v} \mid v_i \leq Z \}$} which simply requires clipping its components. For any $q\in\mathbb{R}$ let $\mathcal{V}_t(q)$ be the set of all $\boldsymbol{v}$ with $\boldsymbol{v}\cdot\boldsymbol{c}\leq q$. We note that given, for all $t\in[T]$, a value $q_t\in[1-1/T,1]$ we have that there exists $\hat{\boldsymbol{u}}\in\bigcap_t\mathcal{V}_t(q_t)$ such that the cumulative reward of $\boldsymbol{\pi}(\hat{\boldsymbol{u}})$ is no less than that of $\boldsymbol{\pi}(\boldsymbol{u})$ minus $1$. This means that as long as, on any trial $t$, we project into the set $\mathcal{V}_t(q_t)$ for some $q_t\in[1-1/T,1]$ then we will not add more than one to the regret. So the problem (for the projection step at time $t$ if necessary) is now to project into the set of all {$\{\boldsymbol{v}\,\|\,\boldsymbol{v}\cdot\boldsymbol{c}\leq q_t\}$} for some arbitrary $q_t\in[1-1/T,1]$. Following our use of Lagrange multipliers, this means that we need to find $\lambda^>0$ with $\sum_{i}c_{t,i}w_{t,i}\exp(-\lambda^c_{t,i})\in[1-1/T,1]$. So consider the function $f$ defined by $f(\lambda):=\sum_{i}c_{t,i}w_{t,i}\exp(-\lambda c_{t,i})$. Consider $\lambda:=ZE\ln(ZE)$. Since $w_{t,i}\leq Z$ we have that when $c_{t,i}<1/ZE$ then $c_{t,i}w_{t,i}\exp(-\lambda c_{t,i})\leq c_{t,i}w_{t,i}<1/E$ and that when $c_{t,i}\geq 1/ZE$ then $c_{t,i}w_{t,i}\exp(-\lambda c_{t,i})\leq Z\exp(-\lambda/ZE)=1/E$. This implies that $f(\lambda)\leq 1$ and hence (since $f$ is monotonic decreasing) an acceptable $\lambda^$ lies in $[0,ZE\ln(ZE)]$. For general $\lambda$ we note that $\nabla f(\lambda)=-\sum_i c_{t,i}^2w_{t,i}\exp(-\lambda c_{t,i})\geq- f(\lambda)$. This means that $\|\nabla f(\lambda^)\|\leq 1$. Since the length of the interval $[1-1/T,1]$ is $1/T$ this means that the length of the interval containing acceptable values of $\lambda^$ is at least approximately $1/T$. So we have shown that either $\lambda^=ZE\ln(ZE)$ is acceptable or the range of acceptable values of $\lambda^$ is of length approximately $1/T$ and lies in $[0,ZE\ln(ZE)]$ (which has length $ZE\ln(ZE)$). The ratio of these lengths is $ZET\ln(ZE)$ so interval bisection will find an acceptable value of $\lambda^$ in $O(\ln(ZET\ln(ZE)))=O(\ln(EZT))$ steps. So we have a time complexity $O(EK+E\ln(EZT))$ and we have only added $1$ to the regret (although this additive factor can be arbitrarily small).	NeurIPS_2024_submissions_huggingface	2,024	null	null	null	null	null	null	null	null
LP-3DGS: Learning to Prune 3D Gaussian Splatting	Accept (poster)	Summary: The paper proposed a pruning method for Gaussian Splatting training, which employ a learnable mask to find optimal pruning rate. Strengths: The paper is well-written and the proposed method is easy to understand. Weaknesses: * The proposed method use a regularization loss to encourage the model to prune, so that the weight parameter is an important balancing factor. The discussion on how to set such parameter is missing. * Figure 2 is not clear enough for me to well understand the proposed idea. In Figure 3, why pattern like (b) is better than (a)? * Experiments section use existing works for comparison, but they are all self-implemented. The performance in their original paper should be included for reference. * The author claim the proposed method can find optimal pruning ratio. Based on my observation, the conclusion is not solid. First, what is definition of optimal? Second, not all the subfigures in Fig. 4 & 6 support this conclusion. Technical Quality: 2 Clarity: 2 Questions for Authors: Please see the "weaknesses". Confidence: 4 Soundness: 2 Presentation: 2 Contribution: 2 Limitations: No potential negative social impact. Flag For Ethics Review: ['No ethics review needed.'] Rating: 5 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: ### Response to weakness 1: We would like to emphasize that the main contribution of our work is not merely the use of a regularization loss to encourage the model to prune, but rather the development of a learning framework that can automatically learn a pruning ratio embedding during the training process of a 3DGS model. This is achieved through our proposed Gumbel-Sigmoid mask pruning technique (Equation 8) and our ensemble loss function (Equation 10), all with only one round of training. We kindly request that the reviewer refer to our summary of technical contributions in the Common Question. For the setting of $\lambda_{m}$, we used a value of 5e-4 in all the reported experimental results. We have further conducted an ablation study on this hyperparameter based on the Room scene, as detailed below. \| $\lambda_{m}$ \| PSNR \| SSIM \| LPIPS \| Pruning Ratio \| \|---------------\|-------\|-------\|-------\|---------------\| \| 3e-4 \| 31.488\| 0.9241\| 0.2035\| 73.15% \| \| 5e-4 \| 31.490\| 0.9243\| 0.2032\| 73.78% \| \| 7e-4 \| 31.474\| 0.9238\| 0.2040\| 75.91% \| $\lambda_{m}$ have little impact on the final rendering quality. ### Response to weakness 2: Figure 2 is based on the vanilla 3DGS, with our proposed pruning method highlighted within the green box. This box indicates that a mask is learned for each Gaussian to determine whether it should be pruned during the training process. The detailed technique described mathematically relates to Equations 7, 8, and 10. For Figure 3, since the pruning process involves a binary decision (i.e., whether to prune or not prune), it is preferable for the learned mask distribution to naturally reflect a binary distribution—where values close to 0 indicate the Gaussian should be pruned, and values close to 1 indicate it should not be pruned. ### Response to weakness 3: Since we need to report the training time for an apples-to-apples comparison, it is essential to use the same GPU hardware setup. We have open-sourced all our code and experimental results on GitHub for peer researchers to verify and benchmark. For 3DGS and Mini-Splatting, we reported results using their open-source codes on our GPU hardware. Additionally, because RadSplat is not open-sourced, we re-implemented and benchmarked their method ourselves. It is important to note that all our self-implemented results are consistent with the rendering quality reported in the original papers. To further clarify, we have provided a comparison of the original paper results for 3DGS and our self-implemented results as follows: \| Scene \| Bicycle \| Bonsai \| Counter \| Kitchen \| Room \| Stump \| Garden \| Flowers \| Treehill \| AVG \| \|-----------------\|---------\|--------\|---------\|---------\|------\|-------\|--------\|---------\|----------\|---------\| \| Original paper PSNR \| 25.246 \| 31.980 \| 28.700 \| 30.317 \| 30.632 \| 26.550 \| 27.410 \| 21.520 \| 22.490 \| 27.205 \| \| Self-Implemented PSNR \| 25.087 \| 32.262 \| 29.079 \| 31.581 \| 31.500 \| 26.655 \| 27.254 \| 21.348 \| 22.561 \| 27.480 \| \|\| \| Original paper SSIM \| 0.771 \| 0.938 \| 0.905 \| 0.922 \| 0.914 \| 0.775 \| 0.868 \| 0.605 \| 0.638 \| 0.8151 \| \| Self-Implemented SSIM \| 0.7464 \| 0.9460 \| 0.9138 \| 0.9320 \| 0.9249 \| 0.770 \| 0.8557 \| 0.5876 \| 0.6358 \| 0.8125 \| \|\| \| Original paper LPIPS \| 0.205 \| 0.205 \| 0.204 \| 0.129 \| 0.220 \| 0.210 \| 0.103 \| 0.336 \| 0.317 \| 0.2143 \| \| Self-Implemented LPIPS \| 0.2441 \| 0.1799 \| 0.1839 \| 0.1164 \| 0.1978 \| 0.2423 \| 0.1224 \| 0.3601 \| 0.3469 \| 0.2215 \| ### Response to weakness 4: Please refer to the Common Question about the term "optimal" pruning for additional details. --- Rebuttal Comment 1.1: Comment: While the authors addressed the most of my minor concerns, I still insist on my opinion about the key point about the "optimal prune ratio". First, according to the experiments, most of the "optimal" prune ratios are similar, around 0.7-0.8 without specific pattern. It is a very narrow band. In such cases, the "learning" task for the prune ratio may be a trivial task since the prune ratio is easy to approach. Just set a 0.75 value may be good enough for most of the experiments in Fig.4 & 6. Second, even though we assume the task is non-trivial, the experiment cannot prove the proposed strategy is efficient enough. For example, in the Bicycle-Redsplat subfigures of Fig.6, we cannot tell the prune ratio ~0.6(LP-3DGS) is better than 0.7 or 0.8. Most of the experiments are with the same problem: no convincing metric to examine the efficiency of the proposed method. Actually, I think it is the problem of this task as I said in the first paragraph. In other words, if all the "answers" fall in around 0.6-0.8, and we cannot find the concrete value which is significantly better than others, the results are too vague to support the claims. Therefore for now I will keep my score. --- Rebuttal 2: Comment: Dear reviewer, We hope our rebuttal can address your comments and concern. As the deadline for the author-reviewer discussion is approaching, would you please share your thoughts on our rebuttal and reconsider the review score? thanks a lot --- Rebuttal Comment 2.1: Comment: Thanks for your comments ### For "optimal" wording As for the term of "optimal pruning ratio", in our response to the common question, we already claimed that we will not use it and change to "learned pruning ratio" with more detailed definition. Please refer to above for our detailed clarification. The reviewer o7SR suggested us to change to "learned effective pruning ratio" and we plan to use such in our updated manuscript. For the sensitivity analysis of rendering image quality against pruning ratios, as we show in our experiment results (figure 4 and 6), each scene has different sensitivities against pruning and the "band" is in general within around 0.5 to 0.8, rather than 0.7 to 0.8 or 0.6 to 0.8, where the rendering image quality starts to drop with increasing pruning ratio. In order to find such "band" for different scenes, a manual tuning of pruning ratios with multiple rounds of training is still needed in prior works, which is costly and not an "easy approach". We kindly disagree that just setting the pruning ratio to 0.75 will be just fine. To justify such, we conducted extra experiments based on the pruning ratio of 0.75 based on the Radsplat score for kitchen and room scenes, due to the limited rebuttal time left. Note that, table 1 in our manuscript reported the complete set of our method learned pruning ratio and its corresponding image quality of each scene. From table 1, it can be seen that, in average, our model with learned pruning ratio has almost the same rendering image quality as the unpruned version, which is what we defined in the loss function and the experiment results support that well. In particular, the learned pruning ratio is 0.58 for kitchen scene and 0.74 for room scene. It can be seen from the extra experiments in below table that the kitchen scene has a significant quality drop with a random 0.75 pruning ratio than that of learned value of 0.58. While, since the randomly chosen 0.75 pruning ratio is close to the learned ratio value of 0.74 for Room scene, the rendering image quality is very similar. Furthermore, such 0.75 pruning ratio is actually not "random", but from the thorough complete pruning sensitivity analysis reported in our submitted manuscript. We believe these ablation study could further prove the motivation of this work and effectiveness of our proposed method. To summarize, for scientific research and contributions, we believe our work discovered the pattern of pruning sensitivity in 3DGS and further propose a general learning method that could be applied to different types of pruning importance scores with a target of automatic pruning, which is a solid contribution to the community and will inspire more research in the future. \| \| Pruning Raio \| Kitchen \| Room \| \|:-----:\|:------------:\|:-------:\|:------:\| \| \| 0 \| 31.581 \| 31.500 \| \| PSRN \| Learned \| 31.515 \| 31.490 \| \| \| 0.75 \| 31.187 \| 31.420 \| \|\| \| \| 0 \| 0.9230 \| 0.9249 \| \| SSIM \| Learned \| 0.9311 \| 0.9243 \| \| \| 0.75 \| 0.9270 \| 0.9235 \| \|\| \| \| 0 \| 0.1164 \| 0.1978 \| \| LPIPS \| Learned \| 0.1194 \| 0.2032 \| \| \| 0.75 \| 0.1293 \| 0.2045 \| ### We are not sure whether the reviewers are criticising the effectiveness or efficiency. We discussed about both in below to clarify. Method effectiveness: For the Bicycle-Radsplat subfigure of Figure 6, all the quantitative metrics, i.e., PSNR/SSIM/LPIPS, that are very popularly used in this community to evaluate the rendering image quality are starting to become worse beyond 0.6 pruning ratio. It actually shows the effectiveness of our method that could learn such pruning ratio with only one time training. The pruning is always a tradeoff process between compression ratio and rendering quality, which is the nature of this problem. We never claimed our learned pruning ratio is always better than other manual pruning ratio settings, but the key point is our method is learnable and automatic that could efficiently and effectively achieve a pruning ratio that is always within the "band" mentioned by the reviewer, rather than costly manual tuning. Method efficiency: We conducted comprehensive training cost analysis in table 2. The actual training time compared with vanilla 3DGS reduces even for single round of training. Compared with other manual pruning method, as we discussed before, our method could achieve a learned effective pruning ratio, embedding with the 3DGS model construction learning process, without the costly pruning sensitivity analysis shown in the figure 4 and 6 that require multiple rounds of training with preset pruning ratio. We believe our method could improve both the 3DGS learning efficiency and pruning efficiency significantly compared with prior works. We kindly request the reviewer to reconsider our technical contributions to the community, as well as the final review score of our work. --- Rebuttal 3: Comment: Dear authors, I do agree that "our learned pruning ratio is always 'better' than random chosen pruning ratio, since there is no 'better' pruning ratio in model pruning problem". However, in my previous comment, I only expect that your method is better in over 80% scenes. I do not require an absolute better or optimal. At least a number larger than 50% can prove that the proposed method is effective. I do not think a successful rate of 80% is a high standard for a publishable paper. Learnable and automatic method is not preferred if just a random manual number setting can solve the problem comparably. This is the most important point in my opinion. If the proposed method can not demonstrate enough effectiveness, I can only regard it as an engineering trick in GS reconstruction. As for the comment "such 0.5 to 0.8 pruning ratio interval "found in this work" is based on the vanilla and default 3DGS training setup, ... Different 3DGS training hyperparameter settings...". I want to clarify that: If you want to prove your method is effective on other variants of GS, please just show evidence. If you want to prove that finding empirical prune ratio of other GS is more difficult compared with adopting your method, please show evidence. Finally, as for your comment "We kindly recommend the reviewer to check our discussion with the reviewers o7SR and fvLt above. Both agreed the community will benefit from our proposed automatic pruning", I do have seen all the discussions above and I will keep my independent thinking. I've spent lots of time on this paper and I am not an outsider of GS. I kindly recommend the authors repect every review equally no matter they are positive or negative. --- Rebuttal Comment 3.1: Comment: Dear reviewer 1UP2 thanks for the comments. 1. If you agree there is no 'better' pruning ratio, but how to guarantee our proposed method could learn the pruning ratio located within the cutoff band in the pruning sensitivity analysis plot shown in figure 4 and 6, what does that mean "At least a number larger than 50% can prove that the proposed method is effective." ? We believe all our learned pruning ratios for every scene is already located within the cutoff band of pruning sensitivity analysis plot. the table 1 also reports, with our learned pruning ratio, almost every scene has the same rendering image quality with the vanilla 3DGS. The average rendering image quality is also the same as the unpruned version with an average of 63% pruning ratio. We do not quite understand what are the requested extra experiments to prove we are at least 50% better. if you mean rendering quality, the average rendering quality of the pruned model is already the same as unpruned version, but with 2.7X smaller model size (i.e. average 63% pruning ratio) 2. We already setup the experiments with different vanilla 3DGS parameters, which will lead to different initial model redundancy. Hope we could get the results before rebuttal deadline. We will post the results once we get it	Summary: This work introduces LP-3DGS, aiming to compress 3DGS by replacing the previously manually set threshold with a learning-to-prune scheme. In particular, LP-3DGS learns a binary mask to automatically prune 3DGS, where such a mask is regularized with the Gumbel-Sigmoid technique. Experiments on three benchmarks showcase the soundness of the introduced methods. Strengths: * Applying Gumbel-Sigmoid to regularize the binary mask is reasonable. * Extensive experiments are conducted on several existing benchmarks. * The paper is well-structured and easy to follow. Weaknesses: * Limited technical contributions. This work seems to be a direct application of Gumbel-Sigmoid, which is a general engineering trick (https://github.com/AngelosNal/PyTorch-Gumbel-Sigmoid) to approximate discrete sampling in a differentiable manner. It needs to be clarified what the main technical insights this work aims to bring to the 3DGS compression community. To me, replacing the threshold setting with the Gumbel-Sigmoid is too intuitive to hit the standard of a top-tier conference. * It is difficult to know whether the learned binary mask achieves the “optimal Gaussian point size”. This work demonstrates that it learns the “optimal” solutions by plotting several line charts without any theoretical analysis. In some cases, LP-3DGS did not find the best balance, e.g., Fig. 4 Mini-Splatting on Kitchen. Besides, LP-3DGS seems to lose the flexibility of finding better solutions compared to its baseline, e.g., RadSplat can set a smaller ratio to compress further while maintaining reasonable visual quality. * Lack of discussions with some recent related works. Several more recent approaches in the 3DGS compression community need discussion, and it would be better to add some comparisons with them to understand the advantages of LP-3DGS better. * Girish, Sharath, Kamal Gupta, and Abhinav Shrivastava. "Eagles: Efficient accelerated 3d gaussians with lightweight encodings." arXiv:2312.04564. * Lu, Tao, et al. "Scaffold-gs: Structured 3d gaussians for view-adaptive rendering." CVPR2024. * Chen, Yihang, et al. "HAC: Hash-grid Assisted Context for 3D Gaussian Splatting Compression." arXiv:2403.14530. Technical Quality: 2 Clarity: 2 Questions for Authors: Please kindly refer to the [Weaknesses]. Confidence: 3 Soundness: 2 Presentation: 2 Contribution: 1 Limitations: Please kindly refer to the [Weaknesses]. Flag For Ethics Review: ['No ethics review needed.'] Rating: 4 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: ### Response to weakness 1: Please refer to the Common Question about the term "contribution" for more details. ### Response to weakness 2: Please refer to the Common Question about the term "optimal" pruning for more details. ### Response to weakness 3: We have comprehensively compared our method with two of the most recent and state-of-the-art (SoTA) 3DGS pruning works, namely RadSplat and Mini-Splatting. The experimental results are reported in Figure 4, Tables 1 and 2, and additional results in the appendix. We appreciate the reviewer’s comment and will include comparisons with the suggested works. EAGLES primarily modifies the hyperparameters of the densification stage of 3DGS to control the number of Gaussians, which differs from score-based pruning methods such as RadSplat, Mini-Splatting, and LightGaussian. Our method uses a learnable mask on the scores of Gaussian points, making it incompatible with EAGLES. Scaffold-GS and HAC focus on encoding the attributes of 3DGS to compress the model, which is different from score-based pruning methods. These works use encoding methods rather than pruning, making them incompatible with our approach. We will cite these works as references and discuss them in the manuscript. --- Rebuttal Comment 1.1: Comment: Dear reviewer, we hope our rebuttal has effectively addressed your comments and concerns. As the deadline for the author-reviewer discussion is tomorrow, we kindly request your prompt feedback on our response and ask you to reconsider the review score. Thank you very much for your time and consideration. --- Rebuttal Comment 1.2: Title: last day of rebuttal response Comment: Dear reviewer CWFJ, today is the last day of reviewer-author discussion period. We hope you have checked our rebuttal and found it could address your previous concerns. We would be happy to provide clarification of our work if you have any further comment. thanks for your review. --- Rebuttal 2: Comment: Dear reviewer, We hope our rebuttal can address your comments and concern. As the deadline for the author-reviewer discussion is approaching, would you please share your thoughts on our rebuttal and reconsider the review score? thanks a lot	Summary: This paper proposes a method to optimally prune Gaussians that do not participate in the rendering or optimizing process of the 3D Gaussian Splatting 3D reconstruction algorithm. The key idea in this paper is to use a Gumbel-sigmoid function instead of the standard Sigmoid function to optimize a binary mask based on transmittance of each splatted Gaussian. Experimental results show that by doing so the number of Gaussians can be reduced without losing reconstruction quality, keeping a near optimal balance between quality and computational speed. As a result, higher frame rate can be obtained for rendering. Strengths: - The idea to use the Gumbel-function is interesting because it allows to naturally binarize the input values while keeping differentiability. - Results show that higher frame rate can be achieved without losing accuracy or training time. The ablation study is also good because it shows the advantage of optimizing the mask values compared to hard thresholding. Weaknesses: - Except for replacing the sigmoid function with the Gumbel-Sigmoid function it seems that there is no other novel technical point in the paper. From the results it seems that a threshold value of 0.6 would be better than the naïve 0.5. I am wondering how the ablation study changes if using different threshold. - In the experiments, no comparisons to other methods that focus on optimizing number compressing the model. For example, comparison with LightGaussian would be interesting. It seems from the Figure 4 that the proposed method performs slightly worse than mini splatting. - What is the meaning of the “pruning ratio” for the proposed method in Figure 4? It seems to be different for the different datasets. However, as far as I understood, the proposed method does not use a pruning ratio: l. 194 => pruning […] points with mask values of 0. This is confusing. - There is no significant gain in accuracy, training time or memory consumption. The only significant gain is in rendering speed. This is interesting but because original 3DGS is already quite fast, in practice I am not sure the impact is high. For example if using a simple threshold of 0.5 in Compact3D how does frame rate compares? - Writing can be brushed up. o L.176-178 is not correct. o In figure 4, it seems legends are wrong in the lines for Mini-splatting => blue is described as RadSplat. o Tables are not easy to read. Add some special fonts like bold or underline to emphasize best and second-best results. Technical Quality: 3 Clarity: 2 Questions for Authors: - Clarify Figure 4 and explain what the meaning of the pruning ratio for the proposed method is. - Add more comparative experiments with other compact 3DGS methods. Confidence: 3 Soundness: 3 Presentation: 2 Contribution: 2 Limitations: - Limitations are properly discussed in the conclusion section. Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: ### Response to weakness 1\&3: Please refer to the Common Question about the clarification and summary of our technical contributions. We would like to highlight that the main contribution of our work is not merely adopting the Gumbel-Sigmoid to prune 3DGS models but rather developing a learning framework that can automatically learn a pruning ratio embedding during the training process of a 3DGS model. This is achieved through our proposed Gumbel-Sigmoid mask pruning technique (Equation 8) and ensemble loss function (Equation 10), with only one round of training. Regarding the analysis shown in Figure 4, RadSplat/Mini-Splatting calculates their defined importance score for each Gaussian. If the target is to prune 60% of the entire model, the Gaussians are ranked based on this predefined importance score to achieve the 60% pruning. The actual pruning threshold of the importance score value can then be calculated. To clarify, Figures 1, 4, and 6 display the pruning ratio of the entire model versus the rendering quality, rather than the actual threshold of the importance score. This is why the x-axis ranges from 0 to 1, where each point corresponds to the rendering quality with a pruning ratio ranging from 0.1 to 0.9 in our experiments. Each point represents the result of one round of training with a preset pruning ratio, sweeping from 10% to 90%. For different pruning methods, such as RadSplat and Mini-Splatting, the results indicate that there always exists a balanced pruning ratio region that achieves a large pruning ratio with almost no or negligible rendering quality drop. In contrast, our method can learn such a pruning ratio with only one round of training. As observed by the reviewer, our method can automatically learn balanced pruning ratios with respect to different scenes. In other words, our proposed method does not require a pre-defined pruning ratio or threshold like prior pruning works, but instead learns one. We believe this is the greatest contribution of our work, and our method will benefit the community. ### Response to weakness 2: We kindly disagree with the reviewer and believe that we have already provided a comprehensive comparison with two of the most recent and best-performing state-of-the-art (SoTA) works, namely RadSplat and Mini-Splatting. The experimental results are reported in Figure 4, Tables 1 and 2, and additional results in the appendix. In the paper "RadSplat: Radiance Field-Informed Gaussian Splatting for Robust Real-Time Rendering with 900+ FPS", arXiv:2403.13806, the authors compared their method with LightGaussian, demonstrating better rendering image quality and a smaller model size. Consequently, we adopted RadSplat as our baseline. We appreciate the reviewer’s comments and, as suggested, we have also included a comparison with LightGaussian as follows: \| \| Synthetic NeRF \| \| \| MipNeRF360 and Tank & Temple \| \| \| \|----------------------------\|-----------------------\|------------\|------------\|----------------------------------\|------------\|------------\| \| \| PSNR \| SSIM \| LPIPS \| PSNR \| SSIM \| LPIPS \| \| LightGaussian \| 32.725 \| 0.965 \| 0.0370 \| 27.206 \| 0.761 \| 0.217 \| \| Ours (RadSplat Score) \| 33.709 \| 0.9694 \| 0.03139 \| 27.812 \| 0.863 \| 0.189 \| \| Ours (Mini-Splatting Score) \| 33.496 \| 0.966 \| 0.0335 \| 27.467 \| 0.856 \| 0.201 \| Regarding the comparison with Mini-Splatting, since our method is a learn-to-prune approach, it exhibits slight variations across different scenes. The average performance is reported in Table 1. It appears that the method performs better with the importance score defined in RadSplat compared to the one defined in Mini-Splatting, given our hyperparameter settings. We will investigate this further in our future work. ### Response to weakness 4: We kindly disagree with the reviewer on this comment. Please refer to our response to the Common Question about technical contributions for a detailed explanation. Our primary goal is not to simply learn a smaller model with improved performance as in previous works, but to develop a learning-to-prune framework that automatically determines a balanced pruning ratio. This approach achieves an excellent compression ratio without compromising rendering quality, addressing a challenge that has not been previously tackled in the 3DGS community. The overall training cost for our learning-to-prune methodology is significantly lower compared to prior manual pruning works, such as RadSplat, Mini-Splatting, and LightGaussian. Even when compared to Compact3D, our method, which uses a differentiable mask rather than a straight-through estimator (STE) binary mask, demonstrates better performance. Tables 4 and 5 present the ablation study that highlights this improvement. ### Response to weakness 5: We appreciate the comment. We will correct the image. --- Rebuttal 2: Comment: Dear reviewer, We hope our rebuttal can address your comments and concern. As the deadline for the author-reviewer discussion is approaching, would you please share your thoughts on our rebuttal and reconsider the review score? thanks a lot --- Rebuttal Comment 2.1: Comment: I have read the authors' response and the other reviews. Thanks for the detailed answers. I had similar question as reviewer o7SR about W2. Figure 1-4 kind of gave the impression that the pruning ratio can be decided. It is now clarified that the parameter is the importance score, which results in some pruning ratio. And the proposed method allows to find the score to achieve desired pruning ratio. Other comments have also been addressed the rebuttal. I raise my rating to 6.	Summary: Whereas other works set a fixed pruning ratio or threshold, this paper introduces a method to learn how to prune Gaussians from a scene while retaining high image fidelity. Different scenes have a drop in fidelity at different pruning percentages and the goal of this work is to remove the need to train the model multiple times while searching for a pruning percentage near, but before, that drop. Specifically this work uses a Gumbel-Sigmoid — which is a 2-class Gumbel softmax followed by a Sigmoid function — to enable a differentiable mask that can learn to prune Gaussians during training. This mask is then applied to set the opacity of each Gaussian. Gaussians where the mask goes to 0 are pruned. Evaluations are performed on 2 real world datasets — MipNerf 360 and Tanks & Temples — and one synthetic dataset — NeRF-Sythetic. Strengths: S1: The writing in this work is quite clear. The methodology is straightforward to understand and the explanations appear to be detailed enough to be able to replicate experiments S2: This method seems useful to practitioners who are low on resources and want to quickly fit and compress a 3D-GS model. Weaknesses: Main Weaknesses: W1: This paper uses the terms “optimal” and “best” to describe the pruning ratio found by learning. Specifically a soft definition of the optimal pruning ratio is given in line 41: “minimize the number of Gaussian points while keeping the rendering quality.” But this method does not “keep” the rendering quality — image quality metrics decrease after pruning. Generally the terms “optimal” and “best” are used to describe maximal or minimal points in some meaningful measure. In the domain of pruning 3D-GS models, a more appropriate definition of optimal pruning would be the highest possible compression that can be applied to the model while retaining rendering quality to some percent of the original rendering quality. W2: It would be helpful to also provide analysis on pruning threshold vs. image quality metrics. It’s possible that there exists a pruning threshold for RadSplat and/or Mini-Splatting that prunes a large percent of Gaussians while retaining image quality. W3: Table 2: Minutes isn’t a great metric here. FLOPs or MACs over training would be a lot better. Other works that want to compare to this method will need to use exactly this manuscript’s architecture to replicate its results. Other minor considerations: MC1: There are still several hyper parameters that haven’t been ablated. This is mostly fine, but it would be good to know how big of an impact these parameters have: - Gumbel $\tau$ - $\lambda_m$ - Applying the pruning earlier/later than 20000 epochs MC2: It would be interesting to ablate equation 8 by removing $S_i$. Can you simply just learn $m_i$ parameter without computing an importance score? Cosmetic issues: C1: line 46, line 50, line 62: could -> can C2: $T$ in eq. 7 should be a $\tau$ C3: Tables 1 and 2, move the metric to the first column for clarity (like tables 4 and 5) Technical Quality: 3 Clarity: 3 Questions for Authors: My biggest concern for this manuscript is that the use of terms “optimal” and “best” is misleading. This method can find a good pruning ratio in one-shot, but it’s not clear that this ratio is optimal. I’m rating this manuscript a borderline reject for now because I strongly think this wording needs to change. I will raise my score if the authors can provide a detailed description of how they will update this wording in their rebuttal. I would also like to see W2 and W3 addressed, but these are likely computationally expensive and aren’t as important to me as W1. Confidence: 4 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: Once the wording in this manuscript is changed, I think the main limitations will be addressed. Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: ### Response to W1: Thank you for the comment. Please refer to the Common Question about the term ``optimal" pruning. ### Response to W2: For Radsplat/Mini-Splatting, they calculate their defined importance score for each Gaussian. If the target is to prune 60% of the entire model, the Gaussians are ranked based on this predefined importance score to prune 60% off. Correspondingly, the actual pruning threshold of the importance score value can be calculated. To clarify, Figures 1, 4, and 6 show the pruning ratio of the entire model versus the rendering quality, rather than the actual threshold of the importance score. These figures already report what the reviewer asked for. Each point in these figures represents the result of one round of training with a preset pruning ratio, sweeping from 10% to 90%. The results indicate there always exists a balanced pruning ratio region for both Radsplat/Mini-Splatting that can achieve a large pruning ratio with almost no or negligible rendering quality drop. We have never denied this. However, our method can learn such a pruning ratio with only one round of training. ### Response to W3: Following the same training time metric used in most prior 3DGS model pruning works, such as RadSplat, Mini-Splatting, and EAGLES, we also use training time as an important metric for a direct comparison. All experiments we reported are conducted with the same GPU hardware setup, ensuring consistency. Our method learns a smaller 3DGS model, where the saved MACs and FLOPS depend on the learned pruning ratio. The training time we report is for only one round of training. This is to show that our proposed method incurs almost no training time overhead or even results in shorter training time due to the smaller pruned model in the later stages of the training pipeline. It is important to note that prior manual pruning works require multiple rounds of training to find the balanced pruning ratio setting, making it challenging to simply report the MACs and FLOPS. ### Response to MC1: Thank you for the comments. We conducted further ablation studies on those hyperparameters. We discovered that the parameter $\tau$ has a very minor effect on the learned pruning ratio. We also conducted an ablation study on the parameter $\lambda_{m}$ using the Room scene. The experimental results are as follows: Ablation Study on Parameter $\lambda_{m}$ (Room Scene) \| $\lambda_{m}$ \| PSNR \| SSIM \| LPIPS \| Pruning Ratio \| \|---------------\|-------\|-------\|-------\|---------------\| \| 3e-4 \| 31.488\| 0.9241\| 0.2035\| 73.15% \| \| 5e-4 \| 31.490\| 0.9243\| 0.2032\| 73.78% \| \| 7e-4 \| 31.474\| 0.9238\| 0.2040\| 75.91% \| These results indicate that varying $\lambda_{m}$ has a negligible impact on the PSNR, SSIM, and LPIPS metrics in the Room scene. Regarding the timing of pruning the model, we found that the best time to perform pruning is at the 20,000th epoch. This timing allows the model extra training epochs to fine-tune based on the pruned model. Pruning the model too early is not advisable, as the 3DGS densification stage stops at the 15,000th epoch, and the model requires additional time to learn the scene. We have already reported the complete set of hyperparameter settings and open-sourced our code on GitHub for researchers to reproduce our results. We will include this ablation study in the final manuscript if space permits. ### Response to MC2: Given the rapid development of the 3DGS community, new pruning methods are emerging quickly. We hope our method is compatible with most pruning methods, especially those defining new importance scores for pruning (please refer to the Common Question about Contribution above). ### Cosmetic issues: We appreciate your comments and will address all cosmetic issues in the updated manuscript accordingly. --- Rebuttal Comment 1.1: Comment: W1: Thanks for the detailed reply. I think it’s fine if you include a different adjective to sell this method, just not “optimal”. For instance I think it’s totally fine to claim this is a “learned effective pruning ratio”. W2: I think there may be some confusion here. I’m referring to the actual score computed by these methods, $s \in \mathbb{R}$. A score is computed per Gaussian, then these scores are ordered to prescribe, for a given pruning ratio $x$%, which of them should be pruned. It’s not possible, given only the Gaussian’s ordering (i.e., the pruning ratio) to obtain the original scores. This criticism was meant to call out that all plots are relative to this percentage, but not the actual scores. It’s possible that for one scene a score $s$ may be at the $40$% pruning ratio whereas that same score may be at the $60$% pruning ratio in a different scene. It’s entirely possible that a specific score may provide a better threshold across scenes because it likely will not corespondent to a specific pruning ratio, but will be a measure of the confidence in a particular Gaussian. W3: I agree other methods report minutes. I am not overly concerned by this and don’t require the authors to run this. Note: by FLOPs above I mean “floating point operations”, not FLOPS “floating point operations per second”. MC2: This comment was that it would be interesting to ablate how well this method performs without the use of an auxiliary pruning score. It’s possible that $m_i$ may prove to be a good pruning score on its own. I suspect that $S_i$ may be only providing a good “initialization”. Overall: I still find this method interesting and my primary concern, the wording of the manuscript, has been addressed. --- Rebuttal 2: Comment: # W1 We appreciate your comment. we will change to learned effective pruning ratio. # W2 We thank you for your further clarification. Please see our detailed response below: *"It’s possible that for one scene a score may be at the 40% pruning ratio whereas that same score may be at the 60% pruning ratio in a different scene."* We totally agree with the reviewer. To justify such, we conducted extra experiments to get the pruning ratio while fixing the pruning threshold of RadSplat defined importance score. The results are listed in below table. It can be clearly seen that the same threshold score for different scenes lead to different pruning ratios for different scenes. This actually further justifies the importance of our work, as it implies that different scenes indeed need the "learned effective pruning ratio'' to optimize tradeoff between compression ratio and rendering quality. Due to the fact that it is a most common setting in prior works to measure the rendering image quality with target pruning ratios, for fair comparison, we design our framework to learn an effective pruning ratio for each scene, without multiple rounds of manual tuning of pruning ratios as used in prior works \| Threhold \| Bicycle \| Bonsai \| Counter \| Kitchen \| Room \| Stump \| Garden \| Flowers \| Treehill \| \|:-----:\|:-------:\|:------:\|:-------:\|:-------:\|:------:\|:------:\|:------:\|:-------:\|:--------:\| \| 0.025 \| 50.88% \| 43.79% \| 47.86% \| 38.46% \| 52.73% \| 51.07% \| 48.05% \| 50.65% \| 41.26% \| *"It’s entirely possible that a specific score may provide a better threshold across scenes because it likely will not corespondent to a specific pruning ratio, but will be a measure of the confidence in a particular Gaussian."* This is a very interesting point. We conducted such experiments on different scenes using different threshold values, based on RadSplat pruning importance score and the results are shown in figure below. It can be seen that, in the kitchen scene, the PSNR starts to drop significantly when the threshold is larger than 0.2. However, in the room scene, such cutoff point is around 0.15. It shows different scenes have different sensitivity against pruning threshold. [Kitchen](https://drive.google.com/uc?export=view&id=1dTP7mA_MIw_ROhxstItbThJyI6m-Sr3L) [Room](https://drive.google.com/uc?export=view&id=1qU1gwcNFo6hwcDOC7qch71YZcTcET0rU) More interestingly, since our proposed method learns the effective pruning ratio for different scenes during 3DGS model construction, we discovered that it naturally captures such unique cutoff threshold score for different scenes. # MC2 We appreciate you for sharing this valuable insight. In equation 8, if we only use $m_i$, then we could regard opacity ($o$) as the importance score. It is also true that the opacity itself could be one type of importance score. Our objective is to propose a method that is general, learnable, and orthogonal to other prior or potentially future works exploring different definitions of effective 3DGS pruning importance scores.	Rebuttal 1: Rebuttal: ## Common question about the term "optimal" pruning (Reviewer o7SR and Reviewer 1UP2 ) We agree with Reviewer o7SR that the term "optimal" may be misleading. We will change it to "learned pruning ratio" and provide a more detailed definition in the updated manuscript. In the domain of 3DGS model pruning, the primary objective is to compress the model size as much as possible using a predefined importance factor, without significantly affecting rendering image quality, which involves a trade-off process. As demonstrated in Figure 4, which shows the rendering quality versus pruning ratio, a small pruning ratio typically does not hamper rendering quality due to model redundancy. However, a very large pruning ratio will significantly degrade the rendering quality. There exists a small region that can achieve a relatively large pruning ratio with almost no or slightly degraded rendering quality, which is the optimal or more balanced pruning ratio range we aim to target. In all existing works on 3DGS model pruning, the optimal or final reported balanced pruning ratio is usually determined empirically by manually sweeping the pruning ratio from low to high (e.g., 0.1 to 0.9). Each predefined pruning ratio sweep point requires one round of model training, necessitating multiple rounds of pruning-aware training to optimize or balance the compression (i.e., pruning) ratio and rendering quality. In contrast, our proposed methodology can automatically learn a pruning ratio embedding during the training process of the 3DGS model through our proposed Gumbel-sigmoid mask pruning technique and ensemble loss function (Equation 10), with only one round of training. The experiments in Figure 4 and other reported tables clearly show that our method consistently learns an equivalent or even better pruning ratio compared to manual tuning in existing works. ## Common question about technical contributions (Reviewer fvLt and Reviewer CWFJ) As discussed in the previous common question, identifying the optimal or balanced pruning ratio in prior works is both compute- and time-consuming. We would like to highlight that the main contribution of our work is not merely the adoption of the Gumbel-sigmoid to prune 3DGS models, but rather the development of a learning framework that can automatically learn a pruning ratio embedding during the training process of a 3DGS model. This is achieved through our proposed Gumbel-sigmoid mask pruning technique (Equation 8) and ensemble loss function (Equation 10), all with only one round of training. Specifically, to distinguish our approach from prior pruning works in 3DGS, the advantages of the proposed method are: * Highly Efficient: Identifying balanced pruning ratio candidates from 0.1 to 0.9 with a 0.1 sweeping resolution typically requires 9 rounds of training. In contrast, our proposed method requires only one training round, reducing the overall training cost by up to 9x. * General: Our proposed Gumbel-sigmoid mask pruning technique can be applied to a general importance score (e.g., opacity, or any other importance score indicating the importance of Gaussian) as defined in Equation 8. This means it is orthogonal and compatible with the pruning criteria defined in prior and potentially future 3DGS model pruning works. * Effective: Our experimental results strongly demonstrate that our proposed learning-to-prune method successfully learns a pruning ratio that is always within the balanced pruning range, maintaining a good trade-off between compression ratio and rendering quality compared to existing works.	NeurIPS_2024_submissions_huggingface	2,024	null	null	null	null	null	null	null	null
Acceleration Exists! Optimization Problems When Oracle Can Only Compare Objective Function Values	Accept (poster)	Summary: In this paper the authors consider the problem of $\min_x f(x)$ with an order oracle that returns $sign (f(x)-f(y)+\delta)$ for some bounded noise $\delta$. The method is based on line search integrated with existing randomized coordinate update algorithms. Convergence rates on non-convex, convex, strongly convex objectives and its accelerated variant in the strongly convex case are provided. These are further extended to the stochastic setup $\min_x E_\zeta[f_\zeta (x)]$ with oracle $sign(f_\zeta(x)-f_\zeta(y))$ where a connection to normalized SGD is drawn and asymptotic convergence is derived. Numerical experiments are presented illustrating the performance of the proposed algorithms. Strengths: The paper has clear exposition and adequately surveys the related literature. The problem under study is of great interest to the machine learning community and has many interesting applications. The technical claims are sound and presented ideas are conceptually simple with demonstrated good performance in several setting. Weaknesses: I don't think the analysis result of "the convergence rate of random coordinate descent with the order oracle (OrderRCD), assuming minimal noise, is equal to first-order method random coordinate descent (RCD)" is surprising, given the algorithm is searching for best (1D) stepsize for coordinate update. And in this regard, neither is the acceleration result based on existing algorithmic framework. I've listed my main questions in the section below and here are some minor comments: - Line 77 is not finished. Line 78: it would be more clear if one could put in the argument in the norm instead of $\\|\cdot\\|$, which presumably would be on $x$? - It would be good to put the GRM in the main text otherwise Algorithm 1 does not make it clear how order oracle is used, which is the main focus of the paper. - Line 179: I'd suggest using a different letter for line search accuracy instead of $\epsilon$, which was used for denoting desired precision of the final optimization problem elsewhere. - In (8), it would be better to use the subscript $f_\zeta (x)$, which would make the notation more consistent with (1). Technical Quality: 3 Clarity: 3 Questions for Authors: - Why is coordinate descent the best thing to do? To me, it seems like the order oracle is essentially a "noisy normalized gradient oracle". So while I can understand it is one way to incorporate the oracle information into existing method, is there any sense in which this is the optimal way to use this information? I think the paper will benefit from such a justification otherwise it seems rather ad-hoc. For example, why is coordinate update a good idea in this setup? Most zeroth-order methods do not seem to rely heavily on this (note that the oracle information is only used to select the stepsize in Algorithm 1). - Algorithm 1 as stated works on deterministic objective, i.e., it is not exploiting any finite-sum-type structure. If parts of the result do not concern stochastic objective, the authors should redefine their problem statement in (1). - The accelerated algorithm 2 has several loops, which makes me wonder how well it would be able to deal with noise in the oracle. - Similarly in Section 5, these reductions to existing methods are useful for getting some convergence guarantees, but is there a way to see this is how one should use the information optimally? Confidence: 3 Soundness: 3 Presentation: 3 Contribution: 2 Limitations: Yes it's adequately addressed. Flag For Ethics Review: ['No ethics review needed.'] Rating: 5 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Dear Reviewer ErqJ, We thank you for your feedback on our work. > I don't think the analysis result of... We provide detailed answers to all questions below, including concerns about the significance and novelty of our results. > Line 77 is not finished. Line 78... Thank you, we will correct those typos. > It would be good to put the GRM... Due to space limitations in the main part of the paper, we had to move the GRM algorithm (which is not small) to Appendix. However, we have provided all the important information needed to understand our results in the main part of the paper. > Line 179: I'd suggest using... In choosing the notations, we were guided by the traditions of previous works. It is tacitly accepted that the accuracy of the problem solution is denoted as $\varepsilon$ or $\epsilon$. Since in our work we are essentially solving two optimization problems, we have chosen these two notations accordingly. However, if you recommend the most appropriate notation for solving the inner optimization problem (linear search), we will be happy to change the notation. > In (8), it would be better... Thank you for this comment, we will correct this typo. > Why is coordinate descent the best thing to do? To me... First of all, we would like to mention that in previous works (with the Order Oracle) the authors used various schemes (including the transition to the normalized gradient algorithm), but all these works, including a very recent paper [1] (which appeared after our submission to the conference) achieved only unaccelerated estimates, moreover, all these estimates contain explicit dimensionality. This fact demonstrates that it has not yet been possible to find such an accelerated method that is satisfied with only the direction of the gradient (without knowledge of the modulus)... This includes among various methods with auxiliary low-dimensional minimization [2-4]. In all the accelerated methods known to us, the computation of the real gradient was assumed in one way or another. Therefore, the approach proposed in our paper is perhaps the only feasible scenario at the moment. Moreover, prior to our work, it was not clear that it was possible in any existing accelerated coordinate algorithm to completely get rid of the knowledge of directional derivatives and reduce everything to one-dimensional searches. It is worth noting that this observation is not trivial. For example, this was not the case with full-gradient accelerated algorithms. Moreover, we did not immediately succeed with coordinate algorithms, since there are already many of them. We had to find one for which it was possible to prove the corresponding statement. Regarding non-accelerated methods, our approach explicitly demonstrates the following: we achieve SOTA results for a class of unaccelerated algorithms up to logarithm factor (see Table 1). Moreover, since the dimensionality dependence is hidden in parameters such as $S_{\alpha}$, in some cases (e.g., the asymmetric case: when the trace of the Hessian matrix has the same order as the largest eigenvalue) this dependence is reduced (this is typical for the class of coordinate descent), thus outperforming previous work in the high-dimensional problem. Moreover, we did not stop there and provided an algorithm that improves the already accelerated convergence estimates in the case of the low-dimensional problem (see Appendix G). Finally, in Section 5 of our paper, we opened a new research direction (stochastic order oracle) by providing the first convergence results of another algorithm, but already in the stochastic oracle concept. > Algorithm 1 as stated works... As we mention in Section 3, in the problem formulation, the stochasticity $\xi$ for Algorithm 1 and 2 denotes the $i$-th coordinate, and $\mathcal{D}$ represents the distribution $p_{\alpha}(i)$. But for the algorithm described in Section 5, stochasticity means the $\xi$-th realization of the function. > The accelerated algorithm 2 has several loops... Regarding the use of the second linear search, it is currently necessary to use the linear search twice per iteration for the theoretical proof of estimates. But the good news is that in practical experiments (all the ones we have encountered), using only the first linear search allows us not only to significantly reduce the running time of the algorithm, but also to improve the convergence itself (see Figure 4). Regarding the noise $\delta(x,y)$, in this paper, for simplicity of presentation, we omitted it (we assumed that it does not exist), since our goal was to show that acceleration in such an oracle concept exists. However, the analysis for the case when there is noise will not be very different from the proposed one. We left this technical moment as a future work. > Similarly in Section 5, these reductions to existing methods... In order to talk about the optimality of certain algorithms, we need to compare our upper estimates with the lower ones. However, as far as we know, there are currently no lower estimates for algorithms using the Order Oracle. So now we can only compare with existing SOTA results and assume that these estimates are likely to be optimal. [1] Zhang, C., & Li, T. (2024). Comparisons Are All You Need for Optimizing Smooth Functions. arXiv preprint arXiv:2405.11454. [2] Drori, Y., & Taylor, A. B. (2020). Efficient first-order methods for convex minimization: a constructive approach. Mathematical Programming, 184(1), 183-220. [3] Narkiss, G., & Zibulevsky, M. (2005). SESOP: Sequential Subspace Optimization Method. [4] Hager, W. W., & Zhang, H. (2006). A survey of nonlinear conjugate gradient methods. Pacific journal of Optimization, 2(1), 35-58. With Respect, Authors --- Rebuttal Comment 1.1: Comment: Thanks for the response and for explaining. I remain somewhat neural to the paper / contribution for the following reason: it is straightforward that given any random (i.e., importance-sampling type) coordinate algorithm, such as (3), a line search greedy step can only guarantee more progress, and it is because of the fact the algorithm is built on coordinate update that the 1D binary search is efficient. The same trick, of course, can be applied to coordinate-based accelerated method, which would now require knowledge of $L_i,\mu...$ to implement. There seems to be various hidden $d$ dependence in the $S$'s and $L_i$'s: There was a remark about "when the trace of the Hessian matrix has the same order as the largest eigenvalue, this dependence on $d$ is reduced" in which case the problem itself sounds like a 1D-ish problem to me. The methodology feels a bit like a tag-on to existing methods - I think the paper would benefit from a more thorough study of when/why this is the best "gradient estimator" vs. alternatives potentially using full-gradient-like information. --- Rebuttal 2: Comment: Dear Reviewer ErqJ, We thank you for your prompt feedback! Perhaps we have not clearly described the full importance of our results in Rebuttal, so with all due respect to you, we would like to reiterate the full novelty and significance of our results for the optimization community and beyond. We would like to point out that our work focuses on the case where the oracle cannot return the objective function values $f(x)$ at the requested point $x$, much less the function gradient. Our oracle can only compare the objective function values (see the definition of Order Oracle (2)). There are a number of works in such an oracle concept (see Table 1). As shown in Table 1, the research on the optimization problem with such an oracle is at least c 2019 and the question of the existence of an accelerated algorithm in such oracle concept remained open until our work. For example, in [1] an attempt was made to create an accelerated algorithm using the heavy ball momentum, but the result was still an unaccelerated algorithm. Moreover, as mentioned earlier above, a recent paper [2] (which appeared after our conference submission) proposes another (alternative) approach that also does not produce an accelerated algorithm in the Order Oracle concept. Thus, one of the main contributions of our paper is the proposed new approach to create (both unaccelerated and accelerated) optimization algorithms (1) using only Order Oracle (2). We agree that our approach for creating unaccelerated algorithms is quite simple, but we have demonstrated the effectiveness of this approach by proposing an alternative to the existing ones. But the value of the approach proposed in our work demonstrates the possibility of creating an accelerated algorithm (using Order Oracle). >The same trick, of course... We cannot agree with this statement. Unlike the unaccelerated algorithm, it is not obvious that a linear search can be applied to the accelerated method of coordinate descent (see any accelerated algorithm of coordinate descent). Moreover, it is even more not obvious, and even surprising, that to create an accelerated algorithm with Order Oracle, it is necessary to use not one linear search, but two. As we said, there are a number of accelerated coordinate descent algorithms, and one of the problems we faced while preparing this article was to find the most appropriate accelerated coordinate descent algorithm on the basis of which we could create an already accelerated algorithm with Order Oracle. >There seems to be various hidden $d$ dependence... As we have already mentioned, we have retained the hidden dimensionality dependence from the coordinate algorithm class. Indeed, comparable to other algorithms using Order Oracle concepts (see Related Work, namely "Algorithms with Order Oracle"), there are cases where hidden dimensionality dependence demonstrates superiority over explicit dependence. In our opinion, the proposed approach demonstrating the first accelerated algorithm in the Order Oracle concept is no longer a minor contribution to the paper. Nevertheless, we also have equally significant results, namely, we show how to improve the estimates of the proposed accelerated algorithm with Order Oracle (the low-dimensional problem case). Finally, we provide the first convergence result of the algorithm using the already stochastic concept of Order Oracle (see the definition in (8)). [1] Gorbunov, E. et. al (2019, September). A Stochastic Derivative Free Optimization Method with Momentum. In International Conference on Learning Representations. [2] Zhang, C., & Li, T. (2024). Comparisons Are All You Need for Optimizing Smooth Functions. arXiv preprint arXiv:2405.11454. With Respect, Authors	Summary: This paper addresses challenges in black-box optimization by introducing the concept of an Order Oracle. This oracle compares function values without assuming access to their actual values, allowing for the creation of new optimization algorithms. The paper proposes both deterministic and stochastic approaches, achieving SOTA convergence results and demonstrating the effectiveness of these algorithms through numerical experiments. Strengths: 1. The order oracle setting is new. 2. The theoretical analysis seems solid. Weaknesses: 1. Elaboration on Motivation: While the setting of the problem is interesting, the motivation for addressing this particular problem could be further elaborated, especially in the introduction. Providing a more detailed context and explanation of the significance and potential impact of solving this problem would help readers better understand its importance. 2. Comprehensive Experimental Validation: The experimental validation appears less comprehensive, as it relies on only one simple function. It would be much more convincing to apply the proposed method to the motivating problem outlined in the introduction. This would demonstrate the method's practical applicability and robustness in a real-world scenario. 3. Performance in High-Dimensional Space: I am curious about the performance of the proposed method in high-dimensional space, as this is often a significant challenge for coordinate descent methods. Could the author provide more insights or experimental results that address this aspect? If there are inherent limitations related to high-dimensional settings, it would be beneficial to discuss and highlight these issues. Providing such information would offer a clearer understanding of the method's strengths and potential limitations. Technical Quality: 3 Clarity: 3 Questions for Authors: 1. Technical Challenges compared to Nesterov's 2012 Work: I am curious about the specific technical challenges involved in the proof, especially in relation to extending and building upon Nesterov's 2012 work. Could the author provide more detail on the difficulties encountered and the novel techniques or modifications that were necessary to address these challenges? 2. Dimension-Independent Complexity: I notice that the paper presents a dimension-independent ($d$) complexity in comparison with existing works, as shown in Table 1. This is a significant and surprising result. Could the author elaborate more on this aspect? Specifically, I am interested in understanding the implications of achieving dimension-independent complexity. Or the dependence is manifested in some parameters. Confidence: 3 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: The author could discuss the limitations in high-dimensional space if this is indeed a constraint of the proposed method. Flag For Ethics Review: ['No ethics review needed.'] Rating: 5 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Dear Reviewer WKzP, We thank you for your feedback on our work. We provide detailed answers to the comments and questions raised in the review below. > Elaboration on Motivation. Despite the potential motivation given in the introduction, as well as the mention of Appendix A, which details an application (startup: smart coffee machine), we are convinced that problem setting with such an oracle has a huge number of potential applications, in particular, such an oracle is organic to perhaps the most popular setting at the moment, namely Reinforcement learning with Human Feedback (RLHF). And we agree that the paper will become stronger and potentially have even more impact when we add more mentions of RLHF in the motivational part of the main paper. > Comprehensive Experimental Validation. Indeed, this paper was initiated due to a challenge faced by one of the co-authors in the realization of a startup (a smart coffee machine). Already in the process of preparing the paper, we started testing it. At the moment we have ready experiments on more than a real problem, where the coffee machine uses the algorithms presented in this paper, with a deterministic oracle. However, we have encountered some difficulties: since the coffee machine is not yet patented (we've started the patenting process), we have not been able to insert these experiments. Regardless, we thought we found a decent temporary alternative - model examples that clearly verify the theoretical results. Nevertheless, we realize that adding these results will further strengthen our work. Currently we expect the patenting process for the coffee machine to be completed in September-October this year, so in case of a positive decision (acceptance at the conference) on our paper, we are confident that we will have time to add the experiments on real problems to the camera-ready version. > Performance in High-Dimensional Space. The high-dimensional challenge problem is typical for a class of coordinate algorithms. And since we are based on such an algorithm, we expectedly inherit this challenge. However, as you have already noticed below (in the Questions section), there is no explicit dependence on dimensionality in the final estimates (unlike previous work with Order Oracle). This fact is an advantage of our scheme (approach), since there are cases (e.g., the asymmetric case: when the trace of the Hessian matrix has the same order as the largest eigenvalue) where the dimensionality dependence is reduced (again, this is typical for the class of coordinate algorithms). In this case, our algorithm will be significantly superior to all previous algorithms, since in alternative algorithms the explicit dependence on dimension $d$ does not disappear anywhere. > Technical Challenges compared to Nesterov's 2012 Work. To begin with, we would like to clarify that indeed the optimization problem formulation is somewhat similar to the one considered in [1], with perhaps a serious exception: we cannot use information about the function (Order Oracle concept). As can be seen from Table 1, all previous optimization algorithms using the Order Oracle achieved only unaccelerated convergence estimates. Moreover, a recent paper [2], which appeared on the arXiv after our submission to the conference also provided only an unaccelerated algorithm. This fact demonstrates that it has not yet been possible to find such an accelerated method that is satisfied with only the direction of the gradient (without knowledge of the modulus)... This includes among various methods with auxiliary low-dimensional minimization [3-5]. In all the accelerated methods known to us, the computation of the real gradient was assumed in one way or another. Therefore, it is important to take into account the following: - The approach proposed in our paper is perhaps the only feasible scenario at the moment. - Prior to our work, it was not clear that it was possible in any existing accelerated coordinate algorithm to completely get rid of the knowledge of directional derivatives and reduce everything to one-dimensional searches. It is worth noting that this observation is not trivial. For example, this was not the case with full-gradient accelerated algorithms. - We did not immediately succeed with coordinate algorithms, since there are already many of them. We had to find one for which it was possible to prove the corresponding statement. However, we did not stop there and provide an algorithm (see Appendix G) that shows the possibility of improving convergence estimates in the case of a low-dimensional optimization problem. Finally, we opened a new research direction by providing the first convergence results for an optimization algorithm with a stochastic Order Oracle (see Section 5), which uses a different approach. > Dimension-Independent Complexity. This is an interesting question! Yes, by using unusual sampling of the active coordinate, the dependence on dimensionality is implicitly hidden in parameters such as $S_{\alpha} = \sum_{i}^{d} L_{i}^{\alpha}$. For example, if we consider a uniform distribution ($\alpha = 0$), then $S_0 = d$. Also, we propagate the best results among accelerated coordinate descent algorithms by using the following distribution to select the active coordinate: $i=\mathcal{R_{\alpha/2}}(L)$. This is to get rid of the dimension dependence in the estimates, since $S_{\alpha/2} \leq \sqrt{d S_{\alpha}}$. [1] Nesterov, Y. (2012). Efficiency of coordinate descent methods on huge-scale optimization problems. [2] Zhang, C. et al. (2024). Comparisons Are All You Need for Optimizing Smooth Functions. [3] Drori, Y. et al. (2020). Efficient first-order methods for convex minimization: a constructive approach. [4] Narkiss, G. et al. (2005). SESOP: Sequential Subspace Optimization Method. [5] Hager, W. et al. (2006). A survey of nonlinear conjugate gradient methods. With Respect, Authors --- Rebuttal Comment 1.1: Comment: I appreciate the authors' thorough responses to my questions. After carefully reviewing their answers, I have no further questions at this stage and will provide my feedback to the AC in the next phase. --- Reply to Comment 1.1.1: Comment: Dear Reviewer WKzP, We thank you for taking the time to check our responses. With Respect, Authors	Summary: this paper focuses on solving the black-box optimization problem with order oracle method. specifically, the author provided a non-accelerated deterministic optimization algorithm, which relies on weighted sampling of the coordinate and the GRM method to resolve a linear search subproblem. the author provided convergence analysis on non-convex, convex and strongly convex cases and showing a superior iteration complexity. the author further an accelerated method called OrderACDM and proved its linear convergence. the stochastic case is discussed with asymptotic convergence results. Finally, the author provided a very simple numerical experiments to show the advantages of the proposed algorithms. Strengths: 1. overall, the paper is clear and easy to follow. 2. the author provided solid analysis to the convergence of the proposed algorithms and the results look correct to me. 3. the proposed algorithm looks clean and simple. Weaknesses: 1. i have some concerns on the novelty of this paper. in the non-accelerated algorithm, the two main techniques (the weighted sampling and the GRM) are not new. and the accelerated algorithm closely follows ( Nesterov and Stich, 2017). 2. both of the proposed algorithms are double loop algorithms as in each iteration, a linear search sub-problem is solved by GRM. as the author mentioned, GRM requires O(log 1/eps) iterations, which should be considered significant especially in the case of strongly convex problem. the author is hiding this factor with \tilde{O} in table 1, which makes the comparison unfair. 3. the numerical experiment is too simple. i understand this is a theoretical paper, but the experiments on a toy example is still below the bar of acception. 4. in the accelerated case, it's not clear to me how the proposed algorithm is compared with (ACDM), Nesterov and Stich, 2017) and (NU-ACDM), Allen-Zhu et al., 2016). the proposed algorithm looks slower to me since it is double loop. i would suggest the author to add more discussions on the convergence results. minor issues: 1. line 77, missing something after "to the inner product as.." ? 2. line 78, L_i appears before it is defined. 3. table 1, many parameters appear before they are defined. Technical Quality: 3 Clarity: 3 Questions for Authors: 1. in assumption 1.1, can L-coordinate-Lipschitz suggest L-smooth? Confidence: 3 Soundness: 3 Presentation: 3 Contribution: 2 Limitations: see above Flag For Ethics Review: ['No ethics review needed.'] Rating: 5 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Dear Reviewer uSy5, We thank you for your interest in our work. We attach below detailed responses to the comments and questions raised in the review. > i have some concerns on the novelty of this paper... We tried to prepare the text of the paper so that it would be easily accessible to readers not only for the optimization community, as we believe that the results of the paper can have an impact on various areas. In particular, perhaps the most popular area at the moment is Reinforcement learning with Human Feedback (RLHF). Perhaps it is precisely because of the simplicity and accessibility of the presentation of the paper that the misunderstanding arose. With all due respect to you, we would like to emphasize the significance and novelty of our work. As can be seen from Table 1, all previous optimization algorithms using the Order Oracle achieved only unaccelerated convergence estimates. Moreover, a recent paper [1], which appeared on the arXiv after our submission to the conference also provided only an unaccelerated algorithm. This fact demonstrates that it has not yet been possible to find such an accelerated method that is satisfied with only the direction of the gradient (without knowledge of the modulus)... This includes among various methods with auxiliary low-dimensional minimization [2-4]. In all the accelerated methods known to us, the computation of the real gradient was assumed in one way or another. Therefore, it is important to take into account the following: - The approach proposed in our paper is perhaps the only feasible scenario at the moment. - Prior to our work, it was not clear that it was possible in any existing accelerated coordinate algorithm to completely get rid of the knowledge of directional derivatives and reduce everything to linear search. It is worth noting that this observation is not trivial. For example, this was not the case with full-gradient accelerated algorithms. - We did not immediately succeed with coordinate algorithms, since there are already many of them. We had to find one for which it was possible to prove the corresponding statement. However, we did`t stop there and provide an algorithm (see Appendix G) that shows the possibility of improving convergence estimates in the case of a low-dimensional optimization problem. Finally, we opened a new research direction by providing the first convergence results for an optimization algorithm with a stochastic Order Oracle (see Section 5), which uses a different approach. As can be seen, all of these results individually, and even more so in combination, are novel and are especially needed in various applications. > both of the proposed algorithms are... We do not quite agree with this remark, since for us the notation $\tilde{O}(\cdot)$ is equivalent to explicitly writing the logarithm in the estimates. Moreover, in Section 1.2 (Notation) we introduced this notation ($\tilde{O}(\cdot)$, where we explained that it hides the logarithmic coefficient, which is standard among optimization works) to reduce the length of the formulas, in particular in Table 1. To be fair, we spell out the presence of the logarithm in the estimates in all discussions of the results, and even more so in the abstract of the paper. However, if you strongly recommend explicitly stating the logarithm in the estimates, we will of course do so. > the numerical experiment is too simple... As written in Appendix A1, this paper was initiated due to a challenge faced by one of the co-authors in the realization of a startup (a smart coffee machine). Already in the process of preparing the paper, we started testing it. At the moment we have ready experiments on more than a real problem, where the coffee machine uses the algorithms presented in this paper, with a deterministic oracle. However, we have encountered some difficulties: since the coffee machine is not yet patented (we've started the patenting process), we have not been able to insert these experiments. Regardless, we thought we found a decent temporary alternative - model examples that clearly verify the theoretical results. Nevertheless, we realize that adding these results will further strengthen our work. Currently we expect the patenting process for the coffee machine to be completed in September-October this year, so in case of a positive decision (acceptance at the conference) on our paper, we are confident that we will have time to add the experiments on real problems to the camera-ready version. > in the accelerated case, it's not clear to me... We would like to clarify that all figures show convergence depending on the iteration complexity. As shown by theoretical evaluations, the iteration complexities of the proposed algorithms match those of the coordinate descent (first order) algorithms, but already the oracle complexities lose by a logarithmic factor. We mention this in the discussion under each result. Moreover, in Figure 4, we compare the proposed accelerated algorithm with the ACDM algorithm of [5]. In Figure 4, we show that in numerical experiments, using only the first linear search not only improves the running speed of the algorithm, but also improves the convergence itself. > minor issues. Thank you, we will correct those typos. > in assumption 1.1, can... Assumption 1.1 is standard for the class of coordintate algorithms and implies some L-smooth over each coordinates. [1] Zhang, C. et al. (2024). Comparisons Are All You Need for Optimizing Smooth Functions. [2] Drori, Y. et al. (2020). Efficient first-order methods for convex minimization: a constructive approach. [3] Narkiss, G. et al. (2005). SESOP: Sequential Subspace Optimization Method. [4] Hager, W. W. et al. (2006). A survey of nonlinear conjugate gradient methods. [5] Nesterov, Y. et al. (2017). Efficiency of the accelerated coordinate descent method on structured optimization problems. With Respect, Authors --- Rebuttal Comment 1.1: Title: reply to authors Comment: I thank the authors for addressing my questions and concerns. Now i agree that the main contribution in this paper is to deal with a special case where only the order of obj function values can be accessed. I have increased my score to be 5. however, my main concern is still the design of the double loop algorithm that leads to an additional log factor and the experiments part. the author mentioned that they could add more practical experiment results. but those results may need another round of review in order to be considered. --- Reply to Comment 1.1.1: Comment: Dear Reviewer uSy5, We are grateful to the Reviewer for checking our responses and for raising the score! However, since there are still concerns about our work, we would like to provide detailed answers: >however, my main concern is still the design of the double loop algorithm that leads to an additional log factor The number of oracle calls per iteration possibly can be further reduced in log factor... However, firstly, such a method (for our oracle) is simply not known. And secondly, the total number of oracle calls (Order Oracle calls) may even be better with our approach, because the iterations number may become smaller due to one-dimensional searches. In any case, this is observed for accelerated methods with low-dimensional auxiliary searches (by the way, note that our oracle could not be integrated into such methods!): [1] (see experiments and references to other works at the end of the paper), and [2]. Nevertheless, we have shown in Figure 6 (see Appendix B) that using only the first linear search not only reduces computational resources but also improves convergence. Based on our observation, we may well expect as future results an improvement of our estimate by at least one logarithm. >and the experiments part. the author mentioned that they could add more practical experiment results. but those results may need another round of review in order to be considered. We would like to emphasize that our work is mainly theoretical. Our main results are four new algorithms: an unaccelerated and accelerated algorithm that uses a deterministic Order Oracle (see Section 3 and 4, respectively), the asymptotic convergence of an algorithm that already uses a stochastic Order Oracle (see Section 5), and finally an algorithm that shows how the convergence results of the accelerated algorithm can be improved given a low-dimensional problem (see Appendix G). All these results individually open directions for future research, showing an already complete and, more importantly, meaningful paper. However, we realize that our problem statement has many potential applications, so, as we said, we will definitely add experiments on a real-world problem (the smart coffee machine - a startup) but in Appendix (since our results are fundamental and not limited to a startup). Nevertheless, we believe that our current experimental part is fully in the spirit of the theoretical work. Finally, we would like to thank the Reviewer uSy5 once again for taking the time to check our responses. We hope that with this answer we have solved all the remaining concerns of the Reviewer. However, if you have any further questions, we will be happy to answer them! [1] Guminov, S., Gasnikov, A., & Kuruzov, I. (2023). Accelerated methods for weakly-quasi-convex optimization problems. Computational Management Science, 20(1), 36. [2] Drori, Y., & Taylor, A. B. (2020). Efficient first-order methods for convex minimization: a constructive approach. Mathematical Programming, 184(1), 183-220. Best regards, Authors	Summary: The paper considers the “Order Oracle” for optimization, which does not require the function values or gradient information, but only the relative ordering of the function values at different points. This can also be done with some small noise. This model captures the challenges encountered in real-world black box optimization problems and the authors motivate it via the ongoing developments in generative models. The main contributions of the paper are: 1. The paper provides SOTA convergence results up to log factors for deterministic algorithms with order oracle in the non-convex, convex and strongly convex settings. 2. The paper shows that acceleration in such an oracle is possible by giving an algorithm for strongly convex functions. They further show that the convergence results can be improved when the problem is low dimensional and compare it to the ellipsoid method. 3. The paper gives a stochastic version of the algorithm. The accelerated and unaccelerated algorithms are inspired by coordinate descent methods. The method is essentially coordinate descent where the step size is chosen via a specific line search procedure (golden ratio method) which is implemented using the order oracle. For the accelerated algorithm, the authors need to include a secondary line search to ensure that the iterates are always making progress. Strengths: The paper demonstrates that using a very weak oracle such as order oracle is sufficient to get rates of convergence that match the rates of the optimal gradient methods via first order oracles. The algorithms are quite simple and work well in practice as shown in the experiments. The ideas are original and the results are presented clearly. Weaknesses: The authors have motivated the use of this oracle, but it would be useful to have some concrete examples on how is it being used in practice. Technical Quality: 4 Clarity: 3 Questions for Authors: 1. When you say adaptive in the text, what does that mean? 2. The algorithms basically seem like you are performing coordinate descent, but instead of using gradients* step size, you want to directly compute this quantity using the GRM line search. So implicitly this is a first order method being implemented using an order oracle? If this is the case, is it possible to show an equivalence between the order oracle and gradient oracle? 3. What is the notation with $\alpha$? 4. Are there examples (even experiments) that can say something about the difference between the order oracle and usual first order oracles? 5. In the accelerated algorithm, are there adversarial examples where the second line search is necessary? 6. What is the barrier in extending the acceleration beyond strongly convex objectives. Confidence: 4 Soundness: 4 Presentation: 3 Contribution: 3 Limitations: The paper considers a weaker oracle model than what is usually considered in standard optimization results and gives algorithms under this oracle. It is not clear what the limitations of the oracle are from the paper, but the results match the ones obtained via stronger gradient oracles. Flag For Ethics Review: ['No ethics review needed.'] Rating: 7 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Dear Reviewer K16N, We thank you for your positive evaluation of our work! Below we provide detailed answers to all comments and questions that arose in the review. > The authors have motivated the use of this oracle, but it would be useful to have some concrete examples on how is it being used in practice. With all due respect to you, we do not fully agree with this comment, as a more concrete example given in Appendix A1 (startup: smart coffee machine) seems hard to think of. But despite the fact that this research was indeed initiated due to a problem faced by one of the co-authors in implementing a startup, we are convinced that problem setting with such an oracle has a huge number of potential applications, in particular, such an oracle is organic in perhaps the most popular setting at the moment, namely Reinforcement learning with Human Feedback (RLHF). And we agree that the article will become stronger and potentially have even more impact when we add more mentions of RLHF in the motivational part of the main paper. > When you say adaptive in the text, what does that mean? This is a really interesting question! The unaccelerated algorithm presented in Section 3 can be classified as an adaptive algorithm (especially in the case of uniformly distributed selection of the active coordinate $\alpha = 0$), since the iteration step and the value of the gradient coordinate are found without any parameters (in particular without the $L$-coordinate-Lipschitz constants) by solving an internal linear problem. This adaptability is clearly shown in Figure 2, where OrderRCD outperforms the first-order RCD algorithm, which uses the gradient coordinate with a theoretical step value using the constant $L_i$. That is, the advantage of our algorithm arises because we adaptively select the iteration step of the algorithm in contrast to the first-order algorithm. > The algorithms basically seem like you are performing coordinate descent, but instead of using gradients step size, you want to directly compute this quantity using the GRM line search. So implicitly this is a first order method being implemented using an order oracle? If this is the case, is it possible to show an equivalence between the order oracle and gradient oracle? Indeed, the class of coordinate algorithms was chosen as a basis because the issue of not knowing the function gradient was organically solved by using the golden ratio method (linear search), where the Order Oracle is already used. Therefore, our algorithm almost completely adopted both all the advantages of the coordinate algorithm (which was our goal) and the disadvantages. Moreover, as often mentioned in this paper, we cannot say that these algorithms are equivalent, because the oracle complexity differs by a logarithm. This is the cost of using linear search. > What is the notation with $\alpha$? By $\alpha$ we mean that this is the parameter used in generating the active coordinate $i_k$ with the following distribution: $p_{\alpha}(i) = L_i^{\alpha} / S_\alpha,$ where $i \in [d]$, and $S_{\alpha} = \sum_{i}^{d} L_i^{\alpha}$. This choice of active coordinate is already standard in works devoted to coordinate descent and is actively used in various works, e.g., [1-4]. > Are there examples (even experiments) that can say something about the difference between the order oracle and usual first order oracles? Yes, Figure 2 and 4 show the comparison of unaccelerated and accelerated algorithms respectively. The algorithms proposed in this paper (Order Oracle) are compared with SOTA coordinate descent algorithms (first order oracle). It is worth noting, as discussed above, the unaccelerated algorithm with Order Oracle can quite expectedly outperform the first order algorithm (RCD) due to adaptivity. But the accelerated algorithm with Order Oracle with the given recommendations concerning the number of linear searches is not inferior to the first-order algorithm. > In the accelerated algorithm, are there adversarial examples where the second line search is necessary? At the moment it is necessary to use linear search twice per iteration for theoretical proof of estimates. This is basically not bad considering that our goal was to show the possibility of acceleration. But the good news is that in practical experiments (all the ones we have encountered), using only the first linear search allows us not only to significantly reduce the running time of the algorithm, but also to improve the convergence itself (see Figure 4). > What is the barrier in extending the acceleration beyond strongly convex objectives? As already mentioned, one of the main goals of our work is to show the existence/possibility of accelerating algorithms using the concept of the Order Oracle. We have demonstrated this on the example of a strongly convex function. By doing so, we have opened a direction for future work, in particular, to investigate already accelerated algorithms under other assumptions. Regarding obtaining convergence results of the accelerated algorithm in the convex smooth case, this is possible using standard tricks such as, for example, regularization. We left further development of the results as future work. [1] Bubeck, S. (2015). Convex optimization: Algorithms and complexity. Foundations and Trends® in Machine Learning. [2] Nesterov, Y. (2012). Efficiency of coordinate descent methods on huge-scale optimization problems. SIAM Journal on Optimization. [3] Nesterov, Y., & Stich, S. U. (2017). Efficiency of the accelerated coordinate descent method on structured optimization problems. SIAM Journal on Optimization. [4] Allen-Zhu, Z., Qu, Z., Richtárik, P., & Yuan, Y. (2016, June). Even faster accelerated coordinate descent using non-uniform sampling. In International Conference on Machine Learning. PMLR. With Respect, Authors --- Rebuttal Comment 1.1: Comment: Thanks for the clarifications and answers to my questions! I hope the ones that are missing from the paper end up in the next version! --- Reply to Comment 1.1.1: Comment: Dear Reviewer K16N, We thank you for checking our response and for your very positive evaluation. With Respect, Authors	Rebuttal 1: Rebuttal: Dear Reviewers, we thank you for taking the time to prepare reviews of our work. We have prepared detailed answers to the comments and questions that arose in the reviews. Our responses will be found under the official review. However, we would like to emphasize the highlights from your reviews. - For example, Reviewers XqbR and uSy5 advised to add experiments closer to reality. We would like to point out that our work was initiated due to a challenge faced by one of the co-authors in the realization of a startup (a smart coffee machine, see Appendix A1). Already in the process of preparing the paper, we started testing it. At the moment we have ready experiments on more than a real problem, where the coffee machine uses the algorithms presented in this paper, with a deterministic oracle. However, we have encountered some difficulties: since the coffee machine is not yet patented (we've started the patenting process), we have not been able to insert these experiments. Regardless, we thought we found a decent temporary alternative - model examples that clearly verify the theoretical results. Nevertheless, we realize that adding these results will further strengthen our work. Currently we expect the patenting process for the coffee machine to be completed in September-October this year, so in case of a positive decision (acceptance at the conference) on our paper, we are confident that we will have time to add the experiments on real problems to the camera-ready version. - Whereas Reviewer K16N emphasized in the Summary section all the key points of our work that demonstrate the significance and novelty of the results. We would like to describe them in more detail: This work was motivated by a specific application problem (startup: smart coffee machine), but the presented results are not limited to it! The oracle concept under consideration has a number of applications, including perhaps the most popular setting, namely Reinforcement learning with Human Feedback (RLHF). It is also worth noting that all previous works (see Section 2 and Table 1) have only achieved unaccelerated estimates under various convexity assumptions. This fact demonstrates that it has not yet been possible to find such an accelerated method that is satisfied with only the direction of the gradient (without knowledge of the modulus, see the definition of the Order Oracle)... This includes among various methods with auxiliary low-dimensional minimization [1-3]. In all the accelerated methods known to us, the computation of the real gradient was assumed in one way or another. Therefore, the approach proposed in our paper is perhaps the only feasible scenario at the moment. Moreover, prior to our work, it was not clear that it was possible in any existing accelerated coordinate algorithm to completely get rid of the knowledge of directional derivatives and reduce everything to one-dimensional searches. It is worth noting that this observation is not trivial. For example, this was not the case with full-gradient accelerated algorithms. Moreover, we did not immediately succeed with coordinate algorithms, since there are already many of them. We had to find one for which it was possible to prove the corresponding statement. Further, we did not stop there and provided already an algorithm that improves the already accelerated convergence estimates in the case of the low-dimensional problem (see Appendix G). Finally, in Section 5 of our paper, we opened a new research direction (stochastic order oracle) by providing the first convergence results of another algorithm, but already in the stochastic oracle concept. - Finally, Reviewers WKzP and ErqJ drew attention to the proposed additional advantages of our approach. Namely, the advantage in the case of a high-dimensional problem, which is typical for the class of coordinate algorithms. Since we are based on the coordinate descent algorithm, we expectedly inherit this challenge. However, as you have already noticed below (in the Questions section), there is no explicit dependence on dimensionality in the final estimates (unlike previous work with Order Oracle). This fact is an advantage of our scheme (approach), since there are cases (e.g., the asymmetric case: when the trace of the Hessian matrix has the same order as the largest eigenvalue) where the dimensionality dependence is reduced (again, this is typical for the class of coordinate algorithms). In this case, our algorithm will be significantly superior to all previous algorithms, since in alternative algorithms the explicit dependence on dimension $d$ does not disappear anywhere. We would like to thank all Reviewers once again for their interest in our work. We really appreciate it! [1] Drori, Y., & Taylor, A. B. (2020). Efficient first-order methods for convex minimization: a constructive approach. Mathematical Programming, 184(1), 183-220. [2] Narkiss, G., & Zibulevsky, M. (2005). SESOP: Sequential Subspace Optimization Method. [3] Hager, W. W., & Zhang, H. (2006). A survey of nonlinear conjugate gradient methods. Pacific journal of Optimization, 2(1), 35-58. Best regards, Authors	NeurIPS_2024_submissions_huggingface	2,024	Summary: This paper explores the use of a zero-order oracle called the Order Oracle to solve optimization problems where exact objective function values are unknown. This oracle focuses on comparing the order of objective function values rather than requiring exact values. The authors propose new non-accelerated algorithm, OrderRCD, that integrate the Order Oracle into coordinate descent algorithm and achieves performance comparable to state-of-the-art methods in non-convex, convex, and strongly convex settings. This involves randomly selecting coordinates and performing linear searches to find optimal step sizes. They also present an accelerated optimization algorithm called OrderACDM, which uses two linear searches in each iteration to adaptively determine optimal step sizes. They show faster convergence rates and improved efficiency in strongly convex setting. The paper extends the concept to stochastic settings, showing asymptotic convergence. Numerical experiments validate the theoretical results, demonstrating that OrderRCD performs comparably to traditional methods, while OrderACDM converges faster, outperforming other methods. Strengths: The paper formalizes a novel concept called Order Oracle, which enables optimization without requiring exact function values. Building on this concept, the authors develop two new algorithms, OrderRCD and OrderACDM, which integrate the Order Oracle into the coordinate descent framework. The paper provides a comprehensive theoretical analysis demonstrating the effectiveness of OrderRCD across non-convex, convex, and strongly convex settings and showing faster convergence rate of OrderACDM in strongly convex setting. The convergence proofs of both the algorithms are detailed. Additionally, this paper also introduce the first algorithm using the stochastic Order Oracle with asymptotic convergence guarantees. They also show practical effectiveness of OrderRCD and OrderACDM compared to traditional methods through their numerical experiments. Weaknesses: The numerical experiments show the evaluation of OrderRCD and OrderACDM in standard quadratic functions. Experiments on real-world problems would strengthen the validation. The paper can include specific details about the numerical experiments, the exact values or structure of matrices and vectors $A$, $b$, and $c$, the dimensionality of the problems tested, datasets used etc. The paper does not deeply explore the scalability of the proposed algorithms in very high-dimensional settings or large-scale datasets. While the paper addresses stochastic optimization, analysis of noise sensitivity would enhance the understanding of the algorithm's performance under various conditions. Technical Quality: 3 Clarity: 2 Questions for Authors: 1. Could you provide more specifics about the numerical experiments? 2. Could you show numerical validation in real-world problems and datasets? 3. Are there any experiments being conducted for the stochastic order oracle? Could you show the results in that setting? Confidence: 4 Soundness: 3 Presentation: 2 Contribution: 3 Limitations: 1. The theoretical guarantees rely heavily on the assumptions of strong convexity of the objective function. These assumptions may not hold for all types of optimization problems, limiting the applicability of the proposed methods to broader, less structured scenarios. 2. The paper does not provide practical demonstrations on large-scale, high-dimensional datasets. Additional experiments on a diverse set of objective functions and real-world datasets would provide a more comprehensive validation of the algorithms' performance and robustness 3. The accelerated version requires updating multiple parameters and has multiple steps and iterative processes that potentially increase the computational load. 4. In large-scale problems with high-dimensional data, the computational overhead from linear searches, parameter updates, and coordinate selections can become significant. The scalability of the algorithm is thus a concern, and its performance may degrade as the problem size increases. Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Dear Reviewer XqbR, We would like to thank you for your time for preparing the review. > The numerical experiments show the evaluation of OrderRCD and OrderACDM in standard quadratic functions. Experiments on real-world problems would strengthen the validation. The paper can include specific details about the numerical experiments, the exact values or structure of matrices and vectors $A$, $b$, and $c$, the dimensionality of the problems tested, datasets used etc. The paper does not deeply explore the scalability of the proposed algorithms in very high-dimensional settings or large-scale datasets. While the paper addresses stochastic optimization, analysis of noise sensitivity would enhance the understanding of the algorithm's performance under various conditions. Since the presented questions fully characterize the weaknesses mentioned in the review, we have prepared detailed answers to each of the questions. The answers can be found below. > Could you provide more specifics about the numerical experiments? First, we would like to clarify that due to the lack of space in the main part of the paper, several numerical experiments can be found also in Appendix B. Regarding the reproducibility of the results, in our opinion it seems that there is enough information about the problem formulation and the parameters of the algorithm. However, we agree that we could be a bit more precise in describing the matrix $A$ and vectors $b,c$, namely the process of their generation. We will add this information about the problem formulation to the main part of the paper. > Could you show numerical validation in real-world problems and datasets? As written in Appendix A1, this paper was initiated due to a challenge faced by one of the co-authors in the realization of a startup (a smart coffee machine). Already in the process of preparing the paper, we started testing it. At the moment we have ready experiments on more than a real problem, where the coffee machine uses the algorithms presented in this paper, with a deterministic oracle. However, we have encountered some difficulties: since the coffee machine is not yet patented (we've started the patenting process), we have not been able to insert these experiments. Regardless, we thought we found a decent temporary alternative - model examples that clearly verify the theoretical results. Nevertheless, we realize that adding these results will further strengthen our work. Currently we expect the patenting process for the coffee machine to be completed in September-October this year, so in case of a positive decision (acceptance at the conference) on our paper, we are confident that we will have time to add the experiments on real problems to the camera-ready version. >Are there any experiments being conducted for the stochastic order oracle? Could you show the results in that setting? Of course, yes, they are being conducted! To be more specific, there are tests of the coffee machine at the moment, where the ideal coffee is created already for a certain group of people on average. We also want to add the results of the experiments to the Appendix section in the final version of the paper after obtaining a patent for the coffee machine. >Limitations. Thank you, we will be sure to add a "Limitations" section to the main part of the paper. # Significance and novelty of our results In conclusion, as already mentioned, this work was motivated by a specific application (startup), but the results presented are not limited to it! This oracle concept has a number of applications, including perhaps the most popular setup, namely Reinforcement learning with Human Feedback (RLHF). It is worth noting that all previous works (see Section 2 and Table 1) have only achieved unaccelerated estimates under various convexity assumptions. Moreover, the work [1] that appeared after our submit also achieves only unaccelerated algorithms. This fact demonstrates that it has not yet been possible to find such an accelerated method that is satisfied with only the direction of the gradient (without knowledge of the modulus, see the definition of the Order Oracle)... This includes among various methods with auxiliary low-dimensional minimization [2-4]. In all the accelerated methods known to us, the computation of the real gradient was assumed in one way or another. Therefore, the approach proposed in our paper is perhaps the only feasible scenario at the moment. Moreover, prior to our work, it was not clear that it was possible in any existing accelerated coordinate algorithm to completely get rid of the knowledge of directional derivatives and reduce everything to one-dimensional searches. It is worth noting that this observation is not trivial. For example, this was not the case with full-gradient accelerated algorithms. Moreover, we did not immediately succeed with coordinate algorithms, since there are already many of them. We had to find one for which it was possible to prove the corresponding statement. Further, we did not stop there and provided (see Appendix G) an algorithm that improves the already accelerated convergence estimates in the case of a low-dimensional problem. Finally, in Section 5 of our paper, we opened a new research direction (stochastic order oracle) by providing the first convergence results (of another algorithm), but already in the stochastic oracle concept. [1] Zhang, C., & Li, T. (2024). Comparisons Are All You Need for Optimizing Smooth Functions. arXiv preprint arXiv:2405.11454. [2] Drori, Y., & Taylor, A. B. (2020). Efficient first-order methods for convex minimization: a constructive approach. Mathematical Programming, 184(1), 183-220. [3] Narkiss, G., & Zibulevsky, M. (2005). SESOP: Sequential Subspace Optimization Method. [4] Hager, W. W., & Zhang, H. (2006). A survey of nonlinear conjugate gradient methods. Pacific journal of Optimization, 2(1), 35-58. With Respect, Authors --- Rebuttal Comment 1.1: Title: On the rebuttal Comment: I thank the authors for their responses to my questions. I have read the responses carefully and I am inclined to retain my original scores. And also, it must be interesting to see the algorithm in action real time, all the best to the startup! --- Reply to Comment 1.1.1: Comment: Dear Reviewer XqbR, We thank you for checking our response and for the positive rating. With Respect, Authors	null	null	null	null	null	null
VB-LoRA: Extreme Parameter Efficient Fine-Tuning with Vector Banks	Accept (poster)	Summary: This paper proposes an approach to decrease the number of parameters in LoRA by introducing a collection (aka "vector bank") of sub-vectors and composing Lora matrices across modules/layers using this collection. Each sub-vector within a given LoRA module/layer is then formed as a linear interpolation of sub-vectors from the vector bank, where the coefficients are obtained by selecting top-k "basis vectors" and then applying a softmax. When applied to Roberta-large, the approach yields improvements over LoRA on the GLUE benchmark with only a tenth of the free parameters. When applied GPT-2 large , it obtains similar or slightly better performance than LoRA with around a sixth of the parameters, on the E2E task. The paper reports ablation studies showing the effect of the vector selection strategy (i.e. alternatives to top-k) and the sub-vector length. Strengths: * Proposes a novel approach to reduce the parameter count in LoRA by sharing parameters across modules and layers using a sub-vector approach. (Originality) * The use of sub-vector style decomposition is novel in the context of LoRA and could be applied in other PEFT settings (Significance). * Demonstrates that the approach can yield improvements over LoRA using Roberta or GPT-2 with far fewer parameters (Quality) Weaknesses: * The paper does not discuss if the proposed factorization approach (i.e. using subvectors) can be performed efficiently on accelerators such as GPUs. Also it would be good to know what is the overhead resulting from this subvector decomposition. * It would be useful to report the performance of an alternative PEFT scheme such as prefix tuning (or prefix tuning) or adapters which have a similar number of parameters as the proposed approach. Technical Quality: 4 Clarity: 3 Questions for Authors: * It would be useful to discuss this paper in related work: https://openreview.net/pdf?id=qKQu1ZcJjD : The idea is to initialize matrices in an LLM using pre-trained submatrix components, which bears resemblance to the subvector decomposition. Typos: L101: * parameters -> parameter L192 * numebr -> number L314 * frozen -> frozen Confidence: 5 Soundness: 4 Presentation: 3 Contribution: 4 Limitations: Yes Flag For Ethics Review: ['No ethics review needed.'] Rating: 8 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Dear Reviewer o6ri, 1. Weakness #1 – GPU acceleration and computation overhead VB-LoRA’s implementation is straightforward, and the proposed factorization approach is also simple to implement in modern deep learning frameworks such as PyTorch, allowing us to fully leverage GPU acceleration. However, the use of subvector decomposition does introduce some computational overhead. Despite this, the reduced number of trainable parameters in VB-LoRA results in lower memory consumption. As shown in Table 8, we compared the training time and memory usage of LoRA and VB-LoRA. VB-LoRA requires approximately 15%-20% more training time than LoRA, while using less memory. It's important to note that this additional overhead is limited to the training phase and does not affect inference, as both LoRA and VB-LoRA merge their parameters back into the original model parameters during this stage. 2. Weakness #2 – comparing with other PEFT methods Thank you for your suggestion! We will revise the manuscript to include additional baseline comparisons. Below, we compare our method with inserted adapters and BitFit. Our results show that our method outperforms other PEFT methods while using an order of magnitude fewer parameters. \| \| Method \| # Params \| SST-2 \| MRPC \| CoLA \| QNLI \| RTE \| STSB \| Avg. \| \|---------------\|---------\|----------\|-------\|------\|------\|------\|------\|------\|------\| \| Roberta-base \| VB-LoRA \| 0.027M \| 95.0 \| 89.7 \| 64.3 \| 92.3 \| 82.3 \| 90.7 \| 85.7 \| \| \| Adpt_D [1] \| 0.3M \| 94.2 \| 88.5 \| 60.8 \| 93.1 \| 71.5 \| 89.7 \| 83.0 \| \| \| BitFit [4] \| 0.1M \| 93.7 \| 92.7 \| 62.0 \| 91.8 \| 81.5 \| 90.8 \| 85.4 \| \| Roberta-large \| VB-LoRA \| 0.033M \| 96.3 \| 91.9 \| 69.3 \| 94.4 \| 87.4 \| 91.8 \| 88.5 \| \| \| Adpt_P [2] \| 0.8M \| 96.6 \| 89.7 \| 67.8 \| 94.8 \| 80.1 \| 91.9 \| 86.8 \| \| \| Adpt_H [3] \| 0.8M \| 96.3 \| 87.7 \| 66.3 \| 94.7 \| 72.9 \| 91.5 \| 84.9 \| 1. Adpt_D: Andreas Rücklé, Gregor Geigle, Max Glockner, Tilman Beck, Jonas Pfeiffer, Nils Reimers, and Iryna Gurevych. AdapterDrop: On the efficiency of adapters in transformers. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pp. 7930–7946. 2. Adpt_P: Jonas Pfeiffer, Aishwarya Kamath, Andreas Rücklé, Kyunghyun Cho, and Iryna Gurevych. AdapterFusion: Non-destructive task composition for transfer learning. In Proceedings of the 16th Conference of the European Chapter of the Association for Computational Linguistics: Main Volume, pp. 487–503. 3. Adpt_H: Neil Houlsby, Andrei Giurgiu, Stanislaw Jastrzebski, Bruna Morrone, Quentin de Laroussilhe, Andrea Gesmundo, Mona Attariyan, and Sylvain Gelly. Parameter-efficient transfer learning for nlp, 2019. 4. BitFit: Elad Ben Zaken, Shauli Ravfogel, and Yoav Goldberg. Bitfit: Simple parameter-efficient fine-tuning for transformer-based masked language-models, 2022. 3. Question #1 – related work “Deep Fusion” Thank you for bringing Deep Fusion [1] to our attention! This work provides an efficient method for network training under the network growth mechanism by leveraging pre-trained initializations of smaller networks. It shares a similarity with our approach in that both methods induce structure on the parameters of the original transformer model for improved efficiency. Additionally, initialization can be seen as a way to share computation and reduce costs between training smaller and larger models. We will include a discussion of this work in the related work section of our manuscript. 1. Mazzawi et al., Deep Fusion: Efficient Network Training via Pre-trained Initializations, Forty-first International Conference on Machine Learning, 2024 --- Rebuttal Comment 1.1: Comment: Thanks to the authors for sharing the newer results and adding clarifications.	Summary: The authors propose an extremely parameter-efficient methods, VB-LoRA, for finetuning an LLM. Specifically, VB-LoRA has a shared vector bank (similar to a codebook), the adapter parameters (A and B) of the linear layers in an LLM are constructed from this bank by selecting the most effective bank vectors. Due to the shared mechanism, the number of trainable parameters are very little, much fewer than the strong baselines, like LoRA, VeRA and Tied-LoRA. Tha authors also validate VB-LoRA on three benchmarks (GLUE, E2E benchmark and MT-Bench) on 6 LLMs from three model families (RoBERTa, GPT-2 and Llama2). The results show that VB-LoRA's results are better or comparable to above-mentioned baselines while requiring much fewer number of trainable parameters. Strengths: 1. The proposed method, VB-LoRA, is effective, intuitive and easily applied. 2. The number of trainable parameters are very limited, mostly less than 0.1M. Weaknesses: 1. The paper is not easy to follow, especially Section 3 for the proposed method. VB-LoRA is basically an algorithm inspired by Mixture-of Expert or codebook, and is used for PEFT here. However, the authors seem to intend to make it complexer by relating VB-LoRA to other theoretical works (L137-143, L157-162), which would be better by putting in the related works. Such writing style easily distracts the reader's attention. 2. Some experimental setting is weird, not in line with previous work. - In Table 1, two high-resource tasks of GLUE, i.e. MNLI and QQP, are deleted. The authors claim that the experiments are conducted on a server equipped with 8 A100 GPUs (L215), which means the computation resource should not be a problem. Why are these two tasks are discarded? - All ablation studies are conducted on a very small task, CoLA, whose results have large variance. Normally, 24GB GPU memory and less than 1 hour are enough for finetuning RoBERTa on CoLA. Due to the small number of samples in CoLA, its results have relatively larger variance. Since the authors have much more resource, why choosing this task for ablation study? 3. VB-LoRA has a potential drawback that is inherent in its design, i.e. unscalibility. One big problem for MoE is that not all expert embeddings can be well trained. We normally include a balance loss and use a small number of experts to avoid this problem. For VB-LoRA, could you offer this ablation study? I.e. increasing the number and length of the bank vectors. 4. VB-LoRA doesn't consistently outperform baselines, or the gap is very narrow. I would be interested in more benckmarks on larger LLMs, like commonsen/arithmetic reseasoning tasks. Technical Quality: 2 Clarity: 2 Questions for Authors: 1. In algorithm 1, if you implement delta_W in this way, it will comsume more GPU memory during finetuning than LoRA. For implementing LoRA, the forward pass is z = Ax, H = Bz. It means we don't calculate delta_W explicitely. If you calculate delta_W, it will be stored in the memory for gradient computation, which enlarges the memoy for activation. 2. Check other questions in Weakness. Confidence: 5 Soundness: 2 Presentation: 2 Contribution: 2 Limitations: The limitation discussion is not thorough. Flag For Ethics Review: ['No ethics review needed.'] Rating: 5 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Dear Reviewer eNJC, 1. Weakness #2 – why MNLI and QQP are discarded? We adhered to the experimental settings established by our baseline, VeRA, which did not include MNLI and QQP. Despite having access to computational resources, it is a shared resource in university. We chose to focus on Natural Language Generation (NLG) and instruction tuning tasks and evaluation on GPT-2 and Llama-2 models, which we believe is of more interest to the research community. 2. Weakness #2 – why choose CoLA for ablation study? It is a typical practice to choose smaller sized benchmarks for ablation study, since running small benchmarks across different experimental settings can still be computationally intensive and time-consuming. We chose CoLA for the ablation study due to its relevance and consistency with the prior works, e.g., VeRA. Each experiment on the GLUE benchmark was run five times, and the results are reported as the median of these runs to mitigate variance. 3. Weakness #3 – scalability of VB-LoRA to larger models Note that there are two distinct differences in the routing module between VB-LoRA and MoE. First, the expert network in VB-LoRA is simply a vector in the vector bank. Second, the gating network does not take any model input; instead, it is freely updated by gradients during fine-tuning. In our experiments (e.g., Figure 3, and Figure 5 in the appendix), we observed that even without noisy logits, our method does not suffer from the load balancing issue seen in MoE. In terms of scalability, we have conducted experiments with a range of model sizes, from Robota 125M to Llama 13B, and varied the number and length of vector banks from 30 to 2048 vectors and 128 to 1024 dimensions, respectively. Our method demonstrated consistent performance across these variations. To further demonstrate that our approach is free from load balancing issues that hurt scalability, we present the vector usage in a Gemma-7B model trained on the MetaMathQA dataset, as shown in Figure 2 of the attached PDF file. The vector bank contains 2048 vectors. The distribution of vector usage follows a roughly normal distribution, with most vectors being selected between 40 to 55 times. 4. Weakness #4 – performance of VB-LoRA VB-LoRA's primary focus is on parameter efficiency rather than predictive performance. Despite the extremely high parameter efficiency, our method still delivers accuracies comparable to or sometimes surpassing baseline models. We appreciate the reviewer’s suggestion to evaluate our method on larger language models, particularly in commonsense and arithmetic reasoning tasks. In response, we have fine-tuned the Mistral-7B and Gemma-7B models on the MetaMathQA dataset and evaluated them on the GSM8K and MATH datasets. We compared our method with LoRA and the concurrent work LoRA-XS [1]. Results show that VB-LoRA outperforms all baselines on the GSM8K dataset, with Mistral-7B utilizing only 0.4% of the parameters compared to LoRA, and Gemma-7B using just 0.3%. Compared to LoRA-XS, our method achieved better results on both evaluation datasets while using only 70% of the parameters with Mistral-7B and 83% with Gemma-7B. \| Model \| Method \| # Params \| GSM8K \| MATH \| \|------------\|----------------\|----------\|-------\|-------\| \| Mistral-7B \| Full FT \| 7242M \| 67.02 \| 18.60 \| \| \| LoRA (r=64) \| 168M \| 67.70 \| 19.68 \| \| \| LoRA-XS (r=64) \| 0.92M \| 68.01 \| 17.86 \| \| \| VB-LoRA \| 0.65M \| 69.22 \| 17.90 \| \| Gemma-7B \| Full FT \| 8538M \| 71.34 \| 22.74 \| \| \| LoRA (r=64) \| 200M \| 74.90 \| 31.28 \| \| \| LoRA-XS (r=64) \| 0.80M \| 74.22 \| 27.62 \| \| \| VB-LoRA \| 0.67M \| 75.96 \| 28.90 \| 1. Bałazy, Klaudia, et al. "LoRA-XS: Low-Rank Adaptation with Extremely Small Number of Parameters." arXiv preprint arXiv:2405.17604 (2024). 5. Question #1 – memory consumption We appreciate your insight into the memory consumption concerns. Indeed, the method you mentioned for reducing memory usage was considered and has been implemented already in our submitted supplementary source code, as it integrates seamlessly with our approach. Algorithm 1 serves as a high-level pseudocode to illustrate the formulation of matrices A and B, intentionally omitting the input x for simplicity. Thus, it does not explicitly showcase the memory-saving technique you referenced. We will revise Algorithm 1 as suggested and make it consistent to our implementation. --- Rebuttal Comment 1.1: Comment: Thank you for the clarification and etra eperiments. I still have one more question for W3: Why don't you show the scalability results here? In stead, you only simply mentioned "we have conducted experiments with a range of model sizes, from Robota 125M to Llama 13B, and varied the number and length of vector banks from 30 to 2048 vectors and 128 to 1024 dimensions, respectively. Our method demonstrated consistent performance across these variations." --- Reply to Comment 1.1.1: Comment: Thanks for the follow-up question. We did the ablation study as you suggested, i.e., increasing the number and length of the bank vectors, to demonstrate the scalability. The table below shows the results on COLA (median over 5 runs, with standard deviations in parentheses). As we can see, as the bank size (90->8192) and vector length (256->1024) increase, the performance of VB-LoRA is relatively stable. Once the number of trainable parameters is very large, the performance degrades a little bit, showing the sign of overfitting. This result further illustrates the scalability of our method. We will include this table in the revised paper. \| Bank size \| Vector length \| # Params (M) \| COLA \| \|-----------\|---------------\|--------------\|-------\| \| 90 \| 256 \| 0.03 \| 69.3 (1.5) \| \| 1024 \| 256 \| 0.28 \| 69.2 (1.2) \| \| 8192 \| 256 \| 2.12 \| 69.4 (1.3) \| \| 8192 \| 1024 \| 8.39 \| 68.8 (0.7) \|	Summary: This paper explores parameter-efficient fine-tuning (PEFT) methods in the context of further reducing the number of fine-tunable parameters, even to the extreme. The main idea is to reduce the number of parameter within LoRA modules as much as possible while maintaining or even improving fine-tuning performance. This line of work is a sub-area of PEFT research, where similar ideas have been explored in the past (e.g., the work of Karimi Mahabadi et al. on Compacter architectures, or a very recent work concurrent to this one: LoRA-XS). To this end, the authors propose VB-LoRA which aims to use globally shared vector banks comprising 'constituent' sub-vectors that can be recombined into local 'delta' parameters (of the \Delta W matrix) that modify/adapt the weights of the original model. The VB-LoRA is then evaluated on GLUE with RoBERTa (for NLU), E2E with GPT-2 (for generation), and Llama 2 on MT-Bench (for instruction tuning) to show its benefits in terms of paramater efficiency and performance. Strengths: - The work is well situated and shows excellent awareness of other work in this subarea which also partially inspired the proposed VB-LoRA method. The chosen baseline models are largely adequate and the only baseline missing (to the best of my knowledge) is LoRA-XS (which is concurrent work, so it couldn't have been included anyhow). - The method is well motivated from a conceptual level and it is also well described formally, in a succinct and convincing manner. The idea does resemble matrix factorisation and tensor product-inspired methods a lot, so I would like to see additional links to previous literature here. - I appreciate the fact that the authors aimed to provide a comprehensive evaluation over NLU, NLG as well as instruction-tuned models. However, this approach has traded some depth of evaluation for its breadth (and the evaluation is not entirely adequate and comprehensive - see under "Weaknesses" later). - A careful derivation related to the number of fine-tuned parameters across different methods is a very useful addition to the paper, given the fact that the parameter efficiency-performance balance is the main topic of the paper. - The paper is very clear and well written in general. Weaknesses: - Evaluation is not comprehensive: 1) When it comes to instruction-tuned models, the main findings are based on a single model from a single family (Llama 2), fine-tuned on a single instruction tuning dataset and evaluated on a single evaluation dataset. This is simply enough to move from anecdotal to more generalisable findings, and experiments with (i) additional models, (ii) more instruction tuning datasets (e.g., Flan-v2, Tulu v2, there are more), and (iii) additional evaluation sets (e.g., MMLU, GSM, BBH, HumanEval) are required. 2) The same holds for NLU and NLG experiments. NLU is evaluated only on GLUE (where it's widely established that performance on GLUE is already saturated and considered 'superhuman'). For NLG, only experiments with a single model on a single benchmark are run. As I mentioned above, the experiments are broad, but stay quite shallow, and this has to change in the revised version. - The paper is very quantitatively driven, but it doesn't delve into providing clarifications and explanations on why certain quantitative phenomena/findings are observed: 1) What is the intuition behind VB-LoRA outperforming LoRA? How can we explain this? What makes VB-LoRA even more suitable performance-wise? Isn't it counter-intuitive that an extremely efficient model can outperform here? Are the tasks then too simple (with very low intrinsic dimensionality) as is the case with some of the standard GLUE tasks? 2) How can one fine-tune the optimal VB size (b) and dimensionality of subvectors within the VB (Table 5) - is this task specific or model specific? What would happen if one increases parameter budget a bit - can we expect even better performance or not? In what cases? 3) Are there any sub-vectors that get selected more often than some others? Why? Overall, while I'm quite happy regarding technical aspects of the work, the paper can be much improved in terms of evaluation/experiments and discussion of the key findings (some error analysis would be useful as well). Technical Quality: 3 Clarity: 4 Questions for Authors: - Lines 180-182. "It’s worth mentioning that adding the multi-index information to the vector selection mechanism can make the TKAM model structure-aware, potentially yielding additional benefits." This is unclear to me and warrants further clarification/discussion. Could you write a short paragraph on how the multi-index information would be added and what type of structure would then TKAM learn (become aware of)? - Additional discussion on the similarities and differences between TKAM and standard (sparse and soft) MoE-s is required. Are there any MoE-style paradigms (which incur a higher parameter cost) that could have been used instead of TKAM? In general, the work doesn't really examine the choice of TKAM as a go-to routing method for the selection of sub-vectors. - Can the authors elaborate more on the similarities and differences between their work and the Compacter work (beyond the fact that Compacter focuses on bottleneck adapters while VB-LoRA focuses on LoRA)? Conceptually, the papers operate in the same space with similar ideas. Confidence: 4 Soundness: 3 Presentation: 4 Contribution: 3 Limitations: There are limitations to this work that should be further elaborated (see also under 'Weaknesses'). The paper didn't explore the whole space of options when it comes to vector selection, VB size, VB content interpretation, and task dependency of findings, among other things. The work also stays within the realm of monolingual (i.e., English-only), mono-modal (i.e., text-only) and mono-PEFT (i.e., LoRA-only) contexts, and this should also be adequately signalled in the Limitations section. Flag For Ethics Review: ['No ethics review needed.'] Rating: 7 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Dear Reviewer wW2k, 1. Weakness #1 – evaluations on instruction tuning Thank you for your suggestion. In response, we fine-tuned the Mistral-7B and Gemma-7B models on the MetaMathQA dataset and evaluated them on GSM8K and MATH datasets, and compared them with the suggested work LoRA-XS. Our experimental setup follows the LoRA-XS configuration. The results show that our method outperforms all baselines on GSM8K, with Mistral-7B utilizing only 0.4% of the parameters compared to LoRA, and Gemma-7B using just 0.3%. Compared with LoRA-XS, our method outperforms on both evaluation datasets while using 70% (Mistral-7B) and 83% (Gemma-7B) of parameters. More details can be found in the Global Rebuttal. \| Model \| Method \| # Params \| GSM8K \| MATH \| \|------------\|----------------\|----------\|-------\|-------\| \| Mistral-7B \| Full FT \| 7242M \| 67.02 \| 18.60 \| \| \| LoRA (r=64) \| 168M \| 67.70 \| 19.68 \| \| \| LoRA-XS (r=64) \| 0.92M \| 68.01 \| 17.86 \| \| \| VB-LoRA \| 0.65M \| 69.22 \| 17.90 \| \| Gemma-7B \| Full FT \| 8538M \| 71.34 \| 22.74 \| \| \| LoRA (r=64) \| 200M \| 74.90 \| 31.28 \| \| \| LoRA-XS (r=64) \| 0.80M \| 74.22 \| 27.62 \| \| \| VB-LoRA \| 0.67M \| 75.96 \| 28.90 \| 2. Weakness #1 – evaluations on NLU and NLG experiments We chose the GLUE benchmark for evaluation to keep consistency with our baseline methods, LoRA and VeRA, both of which use GLUE in their experiments. Although performance on GLUE is considered saturated, we believe that using this benchmark does not undermine our evaluation, as our primary focus is on parameter efficiency rather than absolute performance metrics. Moreover, we have evaluated our method across a range of model sizes from 125M (Robota-base) to 13B (Llama 2). Overall, these experiments demonstrate the robustness and effectiveness of our method. 3. Weakness #2 – the intuition behind VB-LoRA We agree that the performance of PEFT methods strongly depends on the task's intrinsic dimensionality [1]. In essence, It's not just the number of training parameters that matters, but also the way they are composed. While both methods use low-rank decomposition, VB-LoRA goes further by breaking layer and module boundaries, employing a less flexible yet expressive reparameterization based on a sparse convex combination of vectors. This approach introduces an inductive bias that can benefit tasks. For more complex tasks like mathematical reasoning, we've included additional evaluations to showcase the effectiveness of our method. 1. Aghajanyan, Armen, Luke Zettlemoyer, and Sonal Gupta. "Intrinsic dimensionality explains the effectiveness of language model fine-tuning." arXiv preprint arXiv:2012.13255 (2020). 4. Weakness #2 – the optimal VB size b and dimensionality of subvectors Yes, the number of vectors (h) and the length of each vector (b) in the vector bank are task-specific hyperparameters that need to be tuned. Given a fixed budget (h x b), the performance is not highly sensitive to these parameters, as shown in Table 5. In general, increasing the parameter budget improves performance, as demonstrated in Figure 1. However, the extent of improvement also depends on other factors such as the model architecture and the task difficulty. Performance gains will plateau once these limitations are reached. 5. Weakness #2 – vector selection In Figure 2 of the attached PDF, we present the vector usage for a Gemma-7B model trained on the MetaMathQA dataset. The distribution of vector usage follows an approximately normal pattern, with most vectors being selected between 40 to 55 times. However, some vectors are chosen more often, with some being selected up to 70 times. 6. Question #1 – adding the multi-index information to the vector selection mechanism can make the TKAM model structure-aware We apologize for the confusion. Currently, VB-LoRA is not structure-aware, in the sense that the selection of vectors from the vector bank is not informed by factors such as module, layer, and matrix type. However, our framework can be easily extended to become structure-aware. One approach is to make the logits of vector selection conditional on embeddings of the layer, module, and matrix type, which can be implemented through a hypernetwork [1]. 1. Mahabadi, Rabeeh Karimi, et al. "Parameter-efficient Multi-task Fine-tuning for Transformers via Shared Hypernetworks." Annual Meeting of the Association for Computational Linguistics, 2021. 7. Question #2 – TKAM and MoE We chose TKAM for its simplicity and strong performance. In our ablation study (Section 4.4), TKAM outperforms other baselines like Gumbel-Softmax and Straight-Through Gumbel-Softmax. While alternatives like DSelect-k [1] and Expert Choice routing [2] are worth exploring, we focused on introducing the general idea in this paper. 1. Hazimeh, Hussein, et al. "Dselect-k: Differentiable selection in the mixture of experts with applications to multi-task learning." Advances in Neural Information Processing Systems, 2021. 2. Zhou, Yanqi, et al. "Mixture-of-experts with expert choice routing." Advances in Neural Information Processing Systems, 2022. 8. Question #3 – VB-LoRA and Compacter Compacter and our work are conceptually similar, as both are PEFT methods that operate in an efficiently reparameterized space and reduce redundancies by globally sharing information. However, Compacter designates shared and adapter-specific parameters, while our approach introduces a vector bank with a learned sharing mechanism. This mechanism offers more flexibility by allowing dynamic, context-dependent utilization of shared parameters instead of static designation. We will include this discussion of the relationship between VB-LoRA and Compacter in the related work. 9. Limitations We agree. We will include these points in the limitations section. --- Rebuttal Comment 1.1: Title: The response clarifies some of my concerns Comment: I would like to thank the authors for the provided response and the additional results, and I'm happy to increase my score in light of the provided new evidence.	Summary: The authors present a modified version of LoRA called VB-LoRA which is a highly parameter efficient fine-tuning method. It uses a vector bank to represent the model parameters as a composition of vectors. This vector bank is then used to select top-k vectors using the top-k softmax function which are thereby pooled and arranged to form the A and B low-rank matrices of LoRA. The top-k admixture module function selected is a differentiable function so the whole model is trained end-to-end making it more efficient for a particular task. Ultimately the authors show that such as divide-and-share approach results in an orders of magnitude more parameter efficient technique. Strengths: 1. The most significant contribution of this paper is to develop an LoRA-like adaptation technique that is orders of magnitude more parameter efficient than LoRA but at the same time maintaining comparable or better performance in all the tasks. VB-LoRA has the best average performance in all the tasks for every model. 2. The second most important aspect of VB-LoRA is that the number of parameters do not grow linearly with the model dimension (number of layers and model dimension). Making the 'k' of top-k much smaller than This is highly valuable as the model sizes are getting increasingly bigger by the day. 3. The usage of differentiable top-k softmax (TKAM - mentioned in eqn. 1) is a great idea as it enables the paper to be trained fully for a specific task. Also dicing the vector to sub vectors of same size was great for sharing among all layers. 4. The authors have tried to go to the extreme in making the LoRA parameter count efficient by dividing even the low-rank vectors of LoRA and pooling them from a common vector bank. 5. The authors also provide adequate supplementary information such as the detailed hyper-parameters and the hardware used which is highly appreciated as it is highly valuable for anyone who wants to replicate these results. 5. The authors also provide the exact code as a supplementary material which is a huge plus, especially as the paper consists of several moving parts. 6. The paper is presented in a very high quality way using multiple different concepts and giving proper references wherever needed. This fact would be appreciated by the readers as the paper uses several different concepts to come up with their design such as sparse admixture models, top-k gating module, canonical polyadic decomposition, etc. Weaknesses: 1. The parameter count of VB-LoRA is defined by: hb + 1.5LMr(d/b). The second term of this is still linearly dependent on the model dimensions (L and M). In the paper's experiments r<<b in which case the second is quite small and doesn't impact the growth in parameter by much. But there exist several adaptation tasks in which much larger ranks are needed. For example this paper (https://ieeexplore.ieee.org/abstract/document/10445894) shows that for domain adaptation large ranks (eg. 64) are much better than smaller ranks. In some of my other works for domain and language adaptation I have needed to use even larger ranks (200+) which are very much comparable to the value of b. In such cases, VB-LoRA's parameter requirement would be very much comparable to other methods. 2. Some parts of the paper are not entirely clear. For eg. in lines 166-169, the paper mentions that decomposed cumulative gradient parameters are more sensitive than the original model parameters during the training process. An intuition to why this is happening would be helpful. 3. In line 153, the paper mentions that by keeping k << h makes the learnt vectors highly specialized. However figure 3 shows that even with k's value as small as 2 and 3, the model updates almost all the vectors during the training process. Inherently this means that almost all the vectors have some activation and they are not very specialized. 4. Not a lot of details are provided for the virtual dimension 'b' and the results of table 5 are not explained. 5. nit: line 167: 'parameters' Technical Quality: 4 Clarity: 4 Questions for Authors: 1. Can you provide an intuition to your finding why the cumulative gradient parameters are more sensitive than than the original model parameters and how this affects adaptation? 2. Have you experimented with larger ranks and how do they look in terms of performance and parameter count? 3. Have you tried this technique on other tasks than what's listed in the paper? Also have you tried it on other architectures? Confidence: 4 Soundness: 4 Presentation: 4 Contribution: 4 Limitations: 1. In my view the authors have correctly mentioned that this paper has no larger societal impact beyond what LLMs may have. 2. Table-3 shows that the performance of LoRA may have improved with the changes in the GPT-4 model over time. But for the other tables the authors have used the old results for several competing models. 3. I agree with the rest of the answers to the NeurIPS paper checklist that the authors have provided. Flag For Ethics Review: ['No ethics review needed.'] Rating: 8 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Dear Reviewer e7Nd, 1. Weakness #1, Question #2 – the parameter efficiency when rank is high First, it’s important to note that rank may not be directly comparable across different methods. In many approaches, rank dictates the number of independent trainable parameters. However, in our method, the majority of the trainable parameters reside in the vector bank, allowing us to use a smaller rank compared to other methods. For instance, when training LLaMA2-13B for instruction tuning, VB-LoRA performs well with r=6, whereas LoRA requires r=64 and VeRA requires r=1024. New experiments on Mistral-7B and Gemma-7B also show that the rank in our method (rank=4) can be significantly lower than baseline methods (r=64), while achieving better performance. More details can be found in the Global Rebuttal. \| Model \| Method \| # Params \| GSM8K \| MATH \| \|------------\|----------------\|----------\|-------\|-------\| \| Mistral-7B \| Full FT \| 7242M \| 67.02 \| 18.60 \| \| \| LoRA (r=64) \| 168M \| 67.70 \| 19.68 \| \| \| LoRA-XS [1] (r=64) \| 0.92M \| 68.01 \| 17.86 \| \| \| VB-LoRA (r=4) \| 0.65M \| 69.22 \| 17.90 \| \| Gemma-7B \| Full FT \| 8538M \| 71.34 \| 22.74 \| \| \| LoRA (r=64) \| 200M \| 74.90 \| 31.28 \| \| \| LoRA-XS [1] (r=64) \| 0.80M \| 74.22 \| 27.62 \| \| \| VB-LoRA (r=4) \| 0.67M \| 75.96 \| 28.90 \| If high rank is needed, VB-LoRA can reduce the number of parameters more significantly compared to other PEFT methods due to the global sharing mechanism. In this case, the gap in parameter efficiency between LoRA and Full FT becomes smaller, whereas VB-LoRA can sustain a high rank while maintaining a compact vector bank. 1. Bałazy, Klaudia, et al. "LoRA-XS: Low-Rank Adaptation with Extremely Small Number of Parameters." arXiv preprint arXiv:2405.17604 (2024). 2. Weakness #2, Question #1 – decomposed cumulative gradient parameters are more sensitive than the original model parameters during the training process The reason for this increased sensitivity is the decomposition process. Similar to hypernetwork, the parameters are generated by another parameterized model, instead of directly being updated through gradients. In our case, the cumulative gradient updates ΔW are decomposed into matrices A and B, and further into sub-vectors in VB-LoRA. The vectors in the vector bank and logits are the parameters directly updated through gradients. This can be related to the training instability issues observed in hypernetworks [1], where heuristics such as gradient clipping are typically used to manage these instabilities [2]. In TKAM, we found that adding noise to the logits [3] can lead to difficulties in training stability. 1. Ortiz, Jose Javier Gonzalez, John Guttag, and Adrian V. Dalca. "Magnitude Invariant Parametrizations Improve Hypernetwork Learning." The Twelfth International Conference on Learning Representations. 2024. 2. Ha, David, Andrew Dai, and Quoc V. Le. "Hypernetworks." arXiv preprint arXiv:1609.09106 (2016). 3. Shazeer, Noam, et al. "Outrageously large neural networks: The sparsely-gated mixture-of-experts layer." arXiv preprint arXiv:1701.06538 (2017). 3. Weakness #3 – vector specialization Figure 3 shows the footprint of the entire training process. It is important to note that active exploration predominantly occurs in the early stages of training. As training progresses, each sub-vector starts to focus more on much fewer (specialized) vectors within the vector bank. To highlight this pattern, we have plotted the footprint at different training periods in Figure 1 in the attached pdf file. This visualization demonstrates that updates become progressively sparser in the later stages of training. 4. Weakness #4 – virtual dimension b The number of vectors (h) and the length of each vector (also known as the virtual dimension b) in the vector bank are task-specific hyperparameters that need to be tuned. One constraint on b is that it needs to be a common factor of all hidden dimensions to ensure compatibility across the entire model, as the hidden dimensions of FFN layers may differ from those of attention layers. Table 5 shows that given a fixed budget (h×b), the performance is not highly sensitive to the exact values of these parameters. However, selecting a moderate b may help achieve higher performance. Given the budget of the vector bank, a larger b reduces the number of vectors in the vector bank, potentially making each vector less specialized. On the other hand, a smaller b increases the number of trainable parameters (as discussed in Sec. 3.4) and complicates the vector selection process. We will include this discussion in the paper. 5. Limitation #2 – evaluations with GPT-4 GPT-4 is only used for evaluating instruction tuning for Llama2. All other experiments do not rely on GPT-4, thus remaining consistent over time.	Rebuttal 1: Rebuttal: We thank all reviewers for their valuable opinions and comments. In response to the reviewers' requests, we have added additional mathematical reasoning experiments for the Mistral-7B and Gemma-7B models. ### Mathematical Reasoning Experiments We fine-tuned the Mistral-7B-v0.1 and Gemma-7B models on the MetaMathQA dataset and evaluated them on GSM8K and MATH datasets. We compared our results with the concurrent work LoRA-XS [1]. Our experimental setup follows the LoRA-XS configuration. We use a vector bank size of 2048 with b=256, and set the rank to 4. We use a batch size of 128 and train for 2 epochs. The warmup ratio is 0.02, and a cosine learning rate scheduler is used. The initial learning is set to 0.001 for the vector bank and 0.01 for the logits. The experiment is performed on A100 80GB GPUs. The results show that our method outperforms all baselines on GSM8K, with Mistral-7B utilizing only 0.4% of the parameters compared to LoRA, and Gemma-7B using just 0.3%. Compared with LoRA-XS, our method outperforms on both evaluation datasets while using 70% (Mistral-7B) and 83% (Gemma-7B) of LoRA-XS parameters. Table 1. Instruction tuning on GSM8K and MATH Benchmarks for Mistral-7B and Gemma-7B models. \| Model \| Method \| # Params \| GSM8K \| MATH \| \|------------\|----------------\|----------\|-------\|-------\| \| Mistral-7B \| Full FT \| 7242M \| 67.02 \| 18.60 \| \| \| LoRA (r=64) \| 168M \| 67.70 \| 19.68 \| \| \| LoRA-XS (r=64) \| 0.92M \| 68.01 \| 17.86 \| \| \| VB-LoRA (Ours) \| 0.65M \| 69.22 \| 17.90 \| \| Gemma-7B \| Full FT \| 8538M \| 71.34 \| 22.74 \| \| \| LoRA (r=64) \| 200M \| 74.90 \| 31.28 \| \| \| LoRA-XS (r=64) \| 0.80M \| 74.22 \| 27.62 \| \| \| VB-LoRA (Ours) \| 0.67M \| 75.96 \| 28.90 \| 1. Bałazy, Klaudia, et al. "LoRA-XS: Low-Rank Adaptation with Extremely Small Number of Parameters." arXiv preprint arXiv:2405.17604 (2024). The attached PDF file contains Figure 1: VB-LoRA’s vector selection footprints during training, and Figure 2: Histogram of vector usage frequency. Pdf: /pdf/dd1f1bdf1dd94e0ea806f99ba166e25a27c46db4.pdf	NeurIPS_2024_submissions_huggingface	2,024	null	null	null	null	null	null	null	null
BMRS: Bayesian Model Reduction for Structured Pruning	Accept (spotlight)	Summary: The paper introduces Bayesian model reduction for structured pruning (BMRS), an efficient method for structured pruning of neural networks. It improves Bayesian structured pruning with multiplicative noise by combining it with Bayesian model reduction, enabling a principled pruning strategy based on efficient Bayesian model comparison. Two variants with different priors are proposed, threshold-free BMRS_N and BMRS_U with a hyperparameter that allows for more aggressive compression. The experimental evaluation on a range of data sets and neural network architectures demonstrates the usefulness of both approaches. Strengths: - The paper tackles an important problem with fundamental societal impact, especially given the massive resource requirements of current large language models. - The paper is well-written and introduces all required concepts clearly, also to readers like me who are less familiar with the field. It focuses on addressing a single gap in the existing literature clearly and comprehensively. - An anonymous code repository with clear structure and dependencies is provided upfront. - The experiments are thoughtfully designed, featuring a range of data sets and neural network architectures, aggregation across a sufficiently high number of random seeds, clear presentation of the main results, and mostly convincing evidence for the usefulness of the proposed method. Weaknesses: - The performance of the proposed BMRS method is better than SNR for simple architectures (see Table 1), but this advantage vanishes for more complex architectures (see Table 2), where SNR achieves higher compression at similar accuracy for 3 of 4 settings. Unlike the thorough analysis of the results for the rest of the experiments, possible underlying reasons are not discussed here. Ultimately, the economic and environmental impact of the proposed pruning method scales with the size of the neural network architecture, so I believe these results to be crucial for extrapolating to bigger architectures that are not computationally feasible to evaluate experimentally here (e.g., LLMs). These concerns currently prevent me from giving a higher score to this otherwise excellent work. - The computational overhead of introducing multiplicative noise is discussed in the Limitations Section, but not considered in the experiments. I understand that this is not related to the novelties of the BMRS method itself, but it would still be helpful to provide runtime information in the tables so that the difference to the baseline can be factored in. Technical Quality: 3 Clarity: 4 Questions for Authors: - As the main figure of the paper, Figure 1 could profit from a more explicit description both in the graphs themselves (e.g., axis labels) and the label (e.g., a more detailed description of the steps taken or that it uses the log-uniform prior). - Why is a pruning threshold defined for L2 but not SNR at the start of Section 5? Isn't a major drawback of SNR the need for choosing a threshold? Confidence: 2 Soundness: 3 Presentation: 4 Contribution: 3 Limitations: Several limitations are discussed. Flag For Ethics Review: ['No ethics review needed.'] Rating: 8 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We thank the reviewer for their time and thoughtful review! We are happy that they noted that the paper tackles and important problem, it is well-written and clearly and comprehensively addresses a gap in the literature, and that the experiments are thoughtfully designed. We address their concerns and questions below. W1. Differences in performance advantage for different settings The reviewer is referring to the continuous pruning experiments. For MNIST, Fashion-MNIST, and CIFAR10 with an MLP and LeNet, BMRS clearly outperforms the next best baseline in terms of compression vs. accuracy, while for ResNet50 and ViT on CIFAR10 and Tiny-ImageNet, SNR provides slightly higher compression at comparable accuracy. We explain this as follows: SNR is dependent on selecting a pruning threshold, which we set at a fixed value (1.0) as used in previous work [1]. We show that this results in radically different pruning behavior for different model sizes and datasets. In contrast, BMRS$_{\mathcal{N}}$ does not use any explicit threshold, and its solution is dependent on how good our variational approximation is. We chose to end training at 200 epochs for the larger models, but with further training we expect the compression gap to close. Additionally, we show two settings of $p_{1}$ for $BMRS_{\mathcal{U}}$ (8 and 4), but with lower values of $p_{1}$ we would expect to see a higher compression rate (e.g., see Figure 4). W2. Runtime information This is an excellent point! We have included the average training runtime for a sample of datasets and models in the additional supplement. We will update these numbers for all experiments in the final version. Also note that this is the average total training time, which can vary depending on the load on our compute cluster and which nodes the models are run on. We can provide fair training and inference benchmarking in the final version of the paper. Q1. Improve Figure 1 Thank you for the suggestions! We have incorporated the changes to Figure 1 and included it in the additional PDF. Q2. No pruning threshold for SNR That is correct, and it is an oversight on our part; the pruning threshold is set to 1.0 for SNR in order to compare with previous work. This is mentioned at line 275 as the threshold is only used for continuous pruning, but we can additionally mention this at line 243. References [1] K. Neklyudov, D. Molchanov, A. Ashukha, and D. P. Vetrov. Structured bayesian pruning via log-normal multiplicative noise. In Advances in Neural Information Processing Systems (NeurIPS), pages 6775–6784, 2017. --- Rebuttal Comment 1.1: Comment: I thank the authors for their helpful clarifications and the updated results provided. My concerns have been addressed and I will raise my score. --- Reply to Comment 1.1.1: Title: Thank you Comment: Thank you so much, we greatly appreciate it!	Summary: This paper proposes a new probabilistic approach to structured pruning of neural networks. Inspired by variational dropout, the authors derive a method to learn a multiplicative noise distribution, which is encoded in a multiplicative noise layer. Pruning algorithms are derived based on assumed priors over the data and parameter distributions (specifically, the distribution of $p(\mathcal{D}\|\theta)$, where $\mathcal{D}$ is the dataset and $\theta$ is the multiplicative noise). The derived pruning algorithms are then applied to some standard image classification benchmarks. Strengths: 1. The variational approach to network pruning is interesting and well-motivated. 2. The mathematical derivations seem to be correct. 3. The writing is reasonably clear, though a specific notation section aside from the Problem Formulation in Section 3.1 would have been helpful. 4. The paragraph on the connection between BMRS and floating point precision was interesting and highlights strong scholarship (lines 199-218). 5. A variety of algorithms (variants) are derived, and compared empirically, both in the post-training setting, as well as pruning while training. Weaknesses: 1. Some figures could be slightly more clear; for instance, Figure 1 should show the y-axis as well on the plot. 2. The experimental evaluation does not show Imagenet experiments. For pruning on vision datasets, this is now a standard benchmark, and should be presented in any experimental setup. 3. Modern architectures -including transformer architectures and ResNet models such as the ResNet50 model used in the experiments - have complex interconnections between layers, such as residual connections in ResNets (see [1]-[3]). These interconnections make pruning such models (by genuinely changing the architectures) challenging, as saliencies have to be computed for groups of connected filters, as opposed to individual filters. It would have been extremely interesting to see the authors apply their approach to complex interconnections instead of just individual filters. 4. It would have been nice to see the authors use their approach to analyze "how many filters/components are required to learn a given dataset". The variational approach seems like it would have provided an interesting insight into such questions, which would potentially have implications in other fields as well, such as NAS and continual learning. 5. The empirical evaluations should baseline against more than just L2, particularly in the post-training pruning. This significantly weakens the evaluation. References [1] Fang et al. DepGraph: Towards Any Structural Pruning [2] Narshana et al. DFPC: Data flow driven pruning of coupled channels without data. [3] Liu et al. Group Fisher Pruning for Practical Network Compression Technical Quality: 3 Clarity: 3 Questions for Authors: 1. What is the computational advantage, if any, of applying the proposed variational approach to structured pruning versus more classical approaches? 2. In the post-training setting, how many samples are required to efficiently compute $\Delta F$? 3. Are there other choices of priors that were considered in this work? Please refer to the 'Weaknesses' section as well. Confidence: 3 Soundness: 3 Presentation: 3 Contribution: 2 Limitations: The limitations are discussed in this paper, specifically in section 6. Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We thank the reviewer for their time and engagement with the paper. We are happy that they found the approach interesting and well motivated, the writing clear, the math sound, the connection to floating point format interesting, and that we derive a variety of pruning algorithms. We address their weaknesses below: W1. Figure clarity We will revise the figures to be more clear, including the suggestion for Figure 1 which has now been updated in the rebuttal PDF. W2. No ImageNet The ResNet and ViT that we use in the paper are pre-trained on ImageNet, and then further fine-tuned on Tiny-ImageNet, which consists of 100,000 ImageNet images from 200 classes. While not as large scale as ImageNet, we do not expect to see radically different compression vs. accuracy if we perform further fine-tuning on the full ImageNet dataset. That being said, we are happy to perform experiments and include results on ImageNet in the final version of the paper, which could not be completed within the rebuttal period due to the size of ImageNet. W3. Compare pruning characteristics when applied to complex structures This is indeed an interesting point, and well worth looking into. The scope of our work is focused on pruning criteria for Bayesian structured pruning, while this question is slightly broader (relating to the Bayesian pruning literature in general). We can point this out in the paper as a useful future direction to explore. W4. How many filters are required to learn a given dataset? This is also an interesting research question! Figures 2, 5, and 7 (the post-training pruning experiments) show us this empirically. The knee-points of the SNR curve show us where accuracy begins to drop for higher compression. We see here that BMRS gives us a lower bound on the number of prunable structures in a network i.e. slightly higher compression can be achieved without significant drops in accuracy, but the accuracy will quickly decrease with further compression. It should also be noted that different sparsity inducing priors will lead to different answers to this question. For example, [1] introduce a hierarchical prior design to induce both unstructured and structured sparsity, which can lead to a more sparse solution (and thus fewer parameters required to learn a given dataset), but is more complex to train and BMR cannot be directly used for group sparsity. W5. More Baselines We agree on the importance of relevant baselines. In our case, the baselines are different pruning criteria for multiplicative noise with the log Uniform prior, so we include L2 norm, SNR, and a no-pruning baseline. Additional baselines that are comparable are: - Expected value, $E[\theta]$ - Magnitude of gradient - Hessian of gradients - Magnitude of activation Given the time constraints we have not been able to finish running all of these baselines, but we have submitted updated versions of Figure 2 and Table 1 with the finished results for $E[\theta]$ in the supplemental material (the rest will be included in the final version). Q1. Advantage of variational pruning vs. classical methods This depends on what the reviewer mean by "computational advantage"; the variational approach can lead to higher sparsification than classic approaches, which provides a net gain at inference time, but can be more costly to train due to the overhead of the variational parameters. Q2. How many samples required for $\Delta F$? One of the benefits of our proposed method is that the calculation of $\Delta F$ does not require any sampling; it is calculated analytically using the statistics of the variational distribution, original prior, and reduced prior, making it quite efficient (see Eq. 8 and Eq. 11). Q3. Are there other choices of prior considered? This depends on if the reviewer means the prior for the original model, or the reduced priors used to calculate $\Delta F$. The former is selected based on previous work that we build upon [2]. The latter are selected in order to be able to calculate $\Delta F$ analytically. References [1] D. Markovic, K. J. Friston, and S. J. Kiebel. Bayesian sparsification for deep neural networks with Bayesian model reduction. arXiv:2309.12095, 2023. [2] K. Neklyudov, D. Molchanov, A. Ashukha, and D. P. Vetrov. Structured bayesian pruning via log-normal multiplicative noise. In Advances in Neural Information Processing Systems (NeurIPS), pages 6775–6784, 2017. --- Rebuttal Comment 1.1: Title: Response to Rebuttal Comment: I thank the authors for their thoughtful response. In particular, thank you for clarifying the error made in the $\Delta F$ question. However, I still have some concerns. Principally, the advantage over classical pruning methods (in terms of sparsification of the model) as stated by the authors is not reflected in the experiments. This is especially since classical methods outperform BMRS in terms of the sparsity/accuracy tradeoff on CIFAR10 models. This is highlighted by the runtime for BMRS on a ViT on CIFAR10 - 20+hours is quite high, particularly compared to cheaper classical methods. As such, I am happy to keep my positive score as is. --- Reply to Comment 1.1.1: Title: Thank you and clarification Comment: Thank you for the comment and engaging in the discussion! We would just briefly clarify two things: - Our experiments do show better compression vs. accuracy against the L2 norm (a form of magnitude pruning which is a classic method); SNR is also based on variational inference but uses a different pruning criteria which requires threshold tuning - The runtimes are based on total training time, i.e., time until convergence, and are thus unnormalized (e.g., ViT on CIFAR10 runs for 200 epochs in those runtimes). The baseline to compare to here is "None" which does not use any variational inference, while both "SNR" and "BMRS" use variational inference with the same noise variables and different pruning criteria. The VI methods are generally slower than with no pruning due to the overhead of the variational parameters, but both SNR pruning and BMRS pruning are about the same. For the case mentioned (ViT on CIFAR10) we actually saw that BMRS ran the fastest. Thank you again and let us know if we can clarify anything else!	Summary: This paper works on structured pruning using Bayesian models. They try both post-training pruning and continuous pruning for MNIST and CIFAR-10 datasets. Strengths: The writing is easy to read and follow. Weaknesses: 1. The novelty is quite limited: The BMRS basically in my opinion is a naive extension as previous Variational dropout [1], or [2]. BMRS seems to simply combine these works and pruning works, and test different priors and criteria. There are no contributions from both theoretical and empirical sides. This work seems to be highly related to SSVI [3] published months ago. [3] also combines pruning and BNN training, but with fully sparse training and novel pruning criteria designs. 2. Performance is quite bad. For example, as shown in Fig 2, BMSR can only reach around 60% compression rate. This is quite bad for CIFAR-10. Previous works like [3] and most of the pruning papers can get more than 90% compression rate. 3. No baselines to compare. I see no previous works as baselines. Seems that this work aims at design a pruning algorithm, then all the modern pure pruning algorithm should be compared. However, non of these are shown. 4. Missing reference For example, [3] is highly related but missing in this paper. [1] D. P. Kingma, T. Salimans, and M. Welling. Variational dropout and the local reparameterization trick. In Advances in Neural Information Processing Systems (NeurIPS), pages 2575–2583, 2015. [2] C. Louizos, K. Ullrich, and M. Welling. Bayesian compression for deep learning. In Advances in Neural Information Processing Systems (NeurIPS), pages 3288–3298, 2017. [3] Li, J., Miao, Z., Qiu, Q., & Zhang, R. (2024). Training Bayesian Neural Networks with Sparse Subspace Variational Inference. arXiv preprint arXiv:2402.11025. Technical Quality: 3 Clarity: 2 Questions for Authors: What is the novelty of this work? Confidence: 5 Soundness: 3 Presentation: 2 Contribution: 2 Limitations: See the weakness section. Flag For Ethics Review: ['No ethics review needed.'] Rating: 5 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We thank the reviewer for their time reviewing the paper. We would like to address the points they make in their review: Novelty We respectfully disagree with the characterization of our method as a naive extension of variational dropout. We derive novel pruning criteria for a class of structured pruning models which are theoretically grounded, namely from the perspective of Bayesian model reduction (BMR). This yields a robust approach to structured pruning that performs well across our experiments/datasets without tuning of any threshold parameters. The different priors used in the paper are to attain different pruning criteria, as different realizations of BMR result from different priors; they are _not_ different priors on the model parameters. In recent work, discussed in our paper, BMR has been successfully applied in the unstructured pruning case [1,2] but not in the structured case as far as we know. Performance Developing BMR for structured pruning is the main contribution of our work. Specifically, to 1) derive theoretically grounded pruning criteria for structured pruning and 2) empirically characterize the pruning characteristics of these criteria with respect to existing criteria. Our work serves as a step towards principled Bayesian structured pruning criteria via BMR. And when comparing with methods that require tuning of importance thresholds (L2, SNR) we show that we achieve reasonable trade-offs between compression and performance. It is not expected that we would attain higher compression than, e.g., Bayesian approaches which induce both unstructured and structured sparsity. Baselines and Datasets We would like to point out that we use MNIST, Fashion-MNIST, CIFAR10, and tiny-imagenet for our experiments (not only MNIST and CIFAR10 as the reviewer mentions). Second, we compared to relevant baseline pruning criteria that have been used in previous work for the model class of Bayesian structured pruning that we study. Indeed in the paper referenced by the reviewer [3], their proposed pruning criteria are based on variations of $E_{q_{\phi}}[\theta]$ and SNR$(\theta)$; we compare to SNR$(\theta)$ as a baseline, and we have added $E_{q_{\phi}}[\theta]$ in the rebuttal PDF document, which will also be updated in the final version of the paper. Missing citations We thank the reviewer for pointing out the very recent work [3], which we will cite in our paper; the two other papers on variational pruning which the reviewer mentions are already discussed in our submission. References [1] J. Beckers, B. Van Erp, Z. Zhao, K. Kondrashov, and B. De Vries. Principled pruning of bayesian neural networks through variational free energy minimization. IEEE Open Journal of Signal Processing, 2024. [2] D. Markovic, K. J. Friston, and S. J. Kiebel. Bayesian sparsification for deep neural networks with Bayesian model reduction. arXiv:2309.12095, 2023. [3] J. Li, Z. Miao, Q. Qiu, and R. Zhang. Training bayesian neural networks with sparse subspace variational inference. CoRR, abs/2402.11025, 2024 --- Rebuttal Comment 1.1: Comment: Thank you to the authors for their response. However, my concern remains unresolved, and I would prefer to maintain the current score. I continue to have concerns about the novelty of this work for similar reasons. Additionally, since this paper focuses on structured pruning rather than improving BNNs, the baselines should include the most advanced pruning algorithms, encompassing works beyond just the Bayesian approaches. --- Reply to Comment 1.1.1: Comment: We thank the reviewer for engaging with our rebuttal. We would like to gently point out that we do not believe our paper contains "technical flaws, weak evaluation, inadequate reproducibility and/or incompletely addressed ethical considerations" as indicated by the reviewer score of 3. The reviewer has not pointed out any technical flaws, and we have provided the source code for reproducibility. For their remaining concerns we would like to address the following points. > Thank you to the authors for their response. However, my concern remains unresolved, and I would prefer to maintain the current score. I continue to have concerns about the novelty of this work for similar reasons. Regarding the novelty, we derive BMR for structured pruning using the combination of a log-uniform prior and log-normal posterior on multiplicative noise. This pruning criteria is generally applicable to any network which uses this setup to induce sparsity, and enables continuous pruning (as opposed to post-training pruning with the tuning of thresholds). As mentioned previously, BMR has been well established in the unstructured case, but not the structured case. For example, it would be straightworward to apply BMR as another subspace selection criterion in Li et al. (SSVI) [1] as discussed by the reviewer since BMR has been used several times in the literature for Gaussian variables (see e.g. [2-4]). Extending this to structured pruning is non-trivial, as indicated by [5], but we derive this in our paper. We additionally offer a theoretical connection to floating point precision for this pruning criterion. Our hope is that this will enable future work on Bayesian structured pruning from the perspective of BMR. We would finally ask, if the reviewer sees the novelty as limited, could they provide references to papers who do something similar, that is, derive Bayesian Model Reduction (BMR) for structured pruning? > Additionally, since this paper focuses on structured pruning rather than improving BNNs, the baselines should include the most advanced pruning algorithms, encompassing works beyond just the Bayesian approaches. We provide an extensive comparison of different model selection criteria (L2, SNR), now also including $E[\theta]$. We would like to stress the following: Any pruning algorithm has two key components: 1. a model selection criterion (for ranking networks) and 2. a (typically heuristic) algorithm for changing network structures (e.g., removing nodes or layers). In our paper, we study the first aspect by deriving new criteria based on BMR on multiplicative noise variables acting on network structures. Additionally, in [1], they also use SNR and $E[\theta]$ as criteria for sparse subspace search for unstructured pruning. One could additionally formulate the sparse subspace search using log-multiplicative noise to induce group sparsity for structured pruning, and subsequently use our pruning criteria based on BMR to perform subspace selection. In our work, we have compared to their model selection criteria, SNR and $E[\theta]$, together with our novel pruning criteria based on BMR, which also allows for continuous pruning, finding that we can achieve high performing and thresholdless pruning (which makes our approach also promising for sparse subspace search). References [1] J. Li, Z. Miao, Q. Qiu, and R. Zhang. Training bayesian neural networks with sparse subspace variational inference. CoRR, abs/2402.11025, 2024 [2] J. Beckers, B. Van Erp, Z. Zhao, K. Kondrashov, and B. De Vries. Principled pruning of bayesian neural networks through variational free energy minimization. IEEE Open Journal of Signal Processing, 2024. [3] K. J. Friston and W. D. Penny. Post hoc Bayesian model selection. NeuroImage, 56(4) 2089–2099, 2011. [4] K. Friston, T. Parr, and P. Zeidman. Bayesian model reduction. arXiv:1805.07092 [stat.ME] 2018 [5] D. Markovic, K. J. Friston, and S. J. Kiebel. Bayesian sparsification for deep neural networks with Bayesian model reduction. arXiv:2309.12095, 2023.	null	null	Rebuttal 1: Rebuttal: We thank the reviewers for their time and their reviews. We are glad that they generally found the approach interesting, the math sound, and the problem important. We are also happy that they all found the writing clear and the paper presented well. To contextualize the reviews, we would like to echo a point made by reviewer L897: _that the paper addresses a single, clear gap in the literature on Bayesian structured pruning_. To the best knowledge of the authors, Bayesian Model Reduction (BMR) for structured pruning has been successfully explored for the first time in this work. Our key contribution shows how to derive principled and theoretically motivated pruning criteria for Bayesian structured pruning using BMR. We empirically show that our derived pruning criteria work on four benchmark datasets (MNIST, FashionMNIST, CIFAR-10, Tiny-ImageNet) across different architectures (MLP, LeNet5, ResNet-50, Vision Transformer). Given this, we address the concerns of each reviewer in the individual responses. In addition, we include the following in the additional PDF as part of our rebuttal: - An updated Figure 1, including more clear x- and y-axis ticks and labels, as well as an updated caption - Additional baseline results using $E[\theta]$; this includes an updated Table 1 and Figure 2 (other figures and tables will be updated in final version) - Average training runtimes for Lenet5 and ViT on all of the datasets in the paper We are looking forward to engage in the author-reviewer discussion. Pdf: /pdf/9bc99e1279d18da91adba037e74f345a36ec35c1.pdf	NeurIPS_2024_submissions_huggingface	2,024	null	null	null	null	null	null	null	null
Rethinking Open-set Noise in Learning with Noisy Labels	Reject	Summary: The paper extends the problem setting of learning with noisy labels (LNL) to include open-set noise, where noisy labels may come from unknown categories, in contrast to the traditional focus on closed-set noise. The authors theoretically compare the impacts of open-set and closed-set noise and analyze detection mechanisms based on prediction entropy. They construct two open-set noisy datasets, CIFAR100-O and ImageNet-O, and introduce an open-set test set for the WebVision benchmark to validate their findings. Their results show that open-set noise exhibits distinct characteristics from closed-set noise. The paper emphasizes the need for comprehensive evaluation methods for models in the presence of open-set noise, calling for further research in this area. Strengths: - The research problem is interesting. Compared with learning with closed-set noise, learning with open-set noise is under-explored. - The theoretical analysis seems to be solid. Weaknesses: - Some technical details are hard to follow. Writing needs to be polished. - The contribution from the algorithm perspective is not enough. Technical Quality: 3 Clarity: 2 Questions for Authors: - I am a bit confused about the title of this work. This paper provides some insights into the research community. However, it does not provide some essentially different and surprising conclusions. Therefore, perhaps it is not suitable to use "rethinking". - The review part is not comprehensive. The methods combining unsupervised and semi-supervised learning are not reviewed and discussed. - The assumption that the noisy labeling will not affect the sampling prior is a bit strong. Could the paper supplement more intuitions about this assumption? - Will the size of label space for the open-set label, i.e., $B$, affect the theoretical analysis? - There is a gap between Line 165 and the following explanations, which makes it hard to understand. - Does Eq. (7) mean the bounded noise? - The method of predicting entropy values is similar to the methods in OOD detection. Could the method supplement more details? - The paper does not provide enough contribution from the algorithm perspective, such as how to estimate the transition matrix and how to better use/remove the open-set data for better model robustness. The harm of open-set label noise is less than closed-set one is known, which has been reflected in the previous works (results or related discussions). Confidence: 4 Soundness: 3 Presentation: 2 Contribution: 2 Limitations: N/A. Flag For Ethics Review: ['No ethics review needed.'] Rating: 4 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thank you very much for your careful review. We sincerely appreciate your time and effort in reading our paper, as well as the insightful and constructive feedback. >Q1: Some technical details …needs to be polished. A1: We will make further revisions to the manuscript to improve clarity. We would like to kindly note that the other two reviewers, gSVD and 6bHU, agreed that the paper was generally clear and well-written. We would be very happy to clarify any specific parts the reviewer points out. >Q2: The contribution … is not enough. A2: Please refer to A10 for Q10. >Q3: I am a bit confused about the title of this work. ... perhaps it is not suitable to use "rethinking". A3: Thanks for the suggestion. We would be happy to drop the term ‘rethinking’. >Q4: The review part is not comprehensive... A4: We would kindly bring attention to the fact that we have cited many papers based on semi-supervised and unsupervised techniques in the current version of the related works section (L67-L69 in paper). Since most of these methods aim to achieve better baseline results by introducing more techniques, and our paper focuses on the theoretical analysis of open-set noise, we thus did not provide detailed descriptions of the methods before. Following the reviewer's suggestion, we will divide the related work section into three parts in the updated manuscript: Statistical-Consistent Methods, Statistical-Approximate Methods, and Exploration of Open-Set Noise. The second part will include introduction to the works referred to by the reviewer. If the reviewer has more references to suggest, we would be very happy to include them. >Q5: The assumption that the noisy labeling ... more intuitions about this assumption? A5: We would like to kindly note that most works in the LNL (Learning with Noisy Labels) area adhere to this assumption. The change in sampling prior ($P(x)$) is often known as the covariate shift problem—this falls beyond the scope of this work and LNL. Furthermore, if the assumption is that both labelling and sampling prior change, the problem becomes nearly intractable. >Q6: Will the size of label space for the open-set label, i.e., B, affect the theoretical analysis? A6: Thanks for the question. We are happy to clarify further that value of $B$ will not affect the theoretical analysis. >Q7: There is a gap between Line 165 and the following explanations... A7: Thank you for the comment. In the updated manuscript, following the suggestion we will include the explanation of each case, at the bullet point that it is defined. For example, - Memorized case: the model completely memorises the noisy labels: $P^f(y\|x=x; y\in \mathcal{Y}^{in}) = P^{y^n}(y\|x=x; y\in \mathcal{Y}^{in})$; This, could arise in the scenarios that a high capacity model is overfitted on a small, noisy dataset. >Q8: Does Eq. (7) mean the bounded noise? A8: Here $\delta$ is a notation to denote the noise ratio without any further assumptions on it. Clearly 0$\leq$delta$\leq$1, that is the noise ratio is bounded between 0 and 1, however, the main purpose of Eq.(7) is to convey that $x_1$ and $x_2$ have overall the same noise levels. This is needed for a fair comparison of different noise given the same noise ratio. >Q9: The method of predicting entropy ... Could the method supplement more details? A9: Indeed, as the reviewer mentioned, entropy-based detection techniques have been widely used in out-of-distribution (OOD) detection (e.g., [1], [2]). Previous LNL methods have also adopted similar approaches for open-set detection (discussed in L250-252 of the paper). We would like to kindly note, that we do not claim entropy-based open-set detection as our contribution; rather, we analyze its applicability to the two different open-set noise we proposed. Following the reviewer's comment, we would be happy to revisit here the steps of the entropy-based sample selection we used [L603-L607 in the appendix]: "Denoting as $e_i, i=1,...,N$ the entropy of all samples' predictions, we model it (after min-max normalization) with a GMM. The probability $p_i, 1,...,N$ of each sample belonging to the component with a smaller mean value are then extracted. Samples with probability $p_i$ greater than a threshold are then identified as ''inlier" subset." We will update the manuscript to move above to the main content. [1] Chan et al. "Entropy maximization and meta classification for out-of-distribution detection in semantic segmentation." ICCV. 2021. [2] Xing et al. "Learning by Erasing: Conditional Entropy Based Transferable Out-of-Distribution Detection." AAAI. 2024. >Q10: The paper does not provide ... to estimate the transition matrix ... remove the open-set data... The harm of open-set label noise is less than closed-set one is known... A10: We respectfully disagree with the reviewer's argument. We want to clarify that our focus is not to explore better ways to estimate the open-set noise transition matrix nor investigate how to detect open-set noise samples from the dataset more effectively. As the reviewer noted, "The harm of open-set label noise is less than closed-set one is known" - However, previous work mostly focused on ‘easy’ open-set noise in the memorized case. In addition to confirming this theoretically, we further demonstrate that the newly proposed ‘hard’ open-set noise also shows less harm than closed-set noise. More importantly, we compare 'hard' open-set noise versus 'easy' open-set noise, finding their opposite trends across the two cases. In the fitted case, 'easy' open-set noise is less harmful, whereas in the memorized case, the impact of 'hard' open-set noise is comparatively smaller. Due to space limitation, we kindly refer the reviewer to our paper and general response for more detailed clarification of contributions. Thanks once again to the reviewer. If there are any further questions, we would be happy to discuss them at the next stage. --- Rebuttal Comment 1.1: Comment: Thank you for the feedback. The response addresses some concerns. There are still some problems in the current form. First, for A5, please refer to [1] to check the discussion about covariate shifts in LNL. This does not fall beyond the scope of LNL. Second, for A9, the reviewer just needs some explanations about the relation between the work and OOD detection work to address/highlight the technical contributions of this work, which does involve or affect the contribution of this paper. Third, the paper does not provide enough contribution from the algorithm perspective, which is further confirmed. I understand the analysis of hard and easy open-set noise and do not deny them. This is just because even if there is such noise, how do we detect and handle/use them? It is not solved. I appreciate the author's response. Due to the above points, I may not be convinced now. When the reviewers and AC discuss, I am willing to listen to other views. Although I am slightly negative, I am not opposed to accepting this paper. [1] Confidence Scores Make Instance-dependent Label-noise Learning Possible. ICML 2021. --- Rebuttal 2: Comment: Many thanks for the reviewer's detailed feedback. We are glad the reviewer is not opposed to accepting the paper. We understand and respect the reviewer’s comments. Here, we would like to further clarify remaining concerns, for your optional reference: - Regarding covariate shift, we appreciate the reviewers' recommended reference. We have carefully read [1] and would like to provide further clarification. The 'covariate shift' mentioned in [1] refers to distribution change between the selected subset and the original training set due to the sample selection (section 2.2), i.e., $P_{\text{selected}}(x) \neq P_{\text{train}}(x)$. For example, sample selection methods based on 'small-loss' mechanism tend to choose more samples with smaller losses (possibly images with more explicit class-related visual signals), leading to potential distributional changes. We would like to note, similar to most methods based on noise transition matrix, the method proposed in [1] also assumes $P_{\text{train}}(x) = P_{\text{test}}(x)$ implicitly. - Regarding the relationship between our work and out-of-distribution (OOD) detection, we would be pleased to include more discussions in the updated related work section. We want to clarify that we do not claim the entropy-based open-set detection method as our contribution; rather, we analyzed its applicability under different open-set noise conditions to show the possible limitations of existing LNL techniques towards different open-set noise. Moreover, our method can also provide interesting insights to current OOD methods. For example, the different tendency of ‘hard’ open-set noise and ‘easy’ open-set noise in different cases (fitted vs. memorized) may also help to guide current OOD methods in dealing with different OOD samples. - Regarding the contributions of this paper, we would like to reiterate that our focus is to investigate and compare different open-set noises and their impacts. We would also kindly refer to the general response for our detailed contributions. Considering how more advanced sample selection frameworks could be used to detect them or better estimate the noise transition matrix is an intriguing direction for future exploration, but it is beyond the scope of this work. For example, as reviewer 3Ttr mentioned, our findings could be used to guide such approaches. We believe that we make a solid contribution to the LNL community. [1] Confidence Scores Make Instance-dependent Label-noise Learning Possible. ICML 2021.	Summary: This paper introduces an approach to address the challenge of open-set noise in the context of learning from noisy labels. The authors propose a method that differentiates between 'easy' and 'hard' types of open-set noise, which is critical for improving the robustness and performance of learning models faced with noisy data. By integrating existing Learning with Noisy Labels (LNL) techniques with novel entropy-based noise detection mechanisms, the paper presents both theoretical insights and empirical validations of the proposed methods. The contributions are significant as they offer a refined perspective on handling different noise complexities, which can enhance the utility of machine learning models in real-world applications dealing with noisy labels. Strengths: Originality: The paper addresses the issue of open-set noise in learning from noisy labels with a novel approach, differentiating between 'easy' and 'hard' noise types. This nuanced consideration is original as it pushes the boundaries of how noise is typically treated in noisy label learning. Quality: The theoretical explanations are thorough and complemented by robust empirical evidence that strengthens the methodological claims. Clarity: The paper is well-structured, offering clear explanations of complex concepts, which aids in understanding the proposed methods and their implications. Significance: The significance of this work is evident as it tackles a critical issue that can potentially enhance model robustness and performance in real-world scenarios where label noise is common. Weaknesses: Dependency on Specific Methods: The reliance on entropy-based techniques for noise distinction may not generalize across all scenarios or noise types. Experimental Scope: The experiments primarily utilize synthetic datasets, which might not fully capture the complexity of real-world data applications. Technical Quality: 3 Clarity: 2 Questions for Authors: Regarding Open-set Noise Types: How do 'easy' and 'hard' open-set noises specifically impact the robustness of different deep learning models across various architectures? Confidence: 4 Soundness: 3 Presentation: 2 Contribution: 3 Limitations: Limited Experimental Scope: The experimental validation focuses predominantly on synthetic datasets like CIFAR100-O and ImageNet-O. While these are commonly used in the research community for benchmarking, the real-world implications of the findings might be limited without additional testing on more varied and real-world datasets. Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thank you very much for your careful review. We sincerely appreciate your time and effort in reading our paper, as well as the insightful and positive feedback. > Q1: Dependency on Specific Methods: The reliance on entropy-based techniques for noise distinction may not generalize across all scenarios or noise types. A1: We would like to thank the reviewer for the comment. Indeed, entropy-based techniques may not be the ultimate answer to the problem of learning under open-set noise, however, we show its applicability and usefulness in the paper, including its limitations (see Section 3.5). We feel, that overall our work advances the state of the art by: - Extending the noise transition matrix concept to address open-set noise and reformulating the learning with noisy labels (LNL) problem. - Analyzing two cases for offline evaluation: the fitted case (model fits the noisy distribution) and the memorized case (model memorizes the noisy labels). - Confirming that open-set noise has less impact than closed-set noise and noting opposite trends between 'easy' and 'hard' open-set noise in these cases. - Proposing an additional open-set detection task and conducting preliminary experiments, given the small impact of open-set noise on classification performance. - Analyzing an entropy-based detection mechanism, which is effective mainly for 'easy' open-set noise. - Creating two synthetic datasets, CIFAR100-O and ImageNet-O, and introducing a new open-set test set for WebVision to facilitate controlled experiments. > Q2: Experimental Scope: The experiments primarily utilize synthetic datasets, which might not fully capture the complexity of real-world data applications. A2: Thank you for the feedback. To conduct comparative experiments with controllable noise ratios, we introduced two novel open-set noise datasets: ImageNet-O and CIFAR100-O. Additionally, we conducted experiments on the real-world dataset WebVision, as detailed in Appendix E. We will update the manuscript to clarify this clearer. > Q3: Regarding Open-set Noise Types: How do 'easy' and 'hard' open-set noises specifically impact the robustness of different deep learning models across various architectures? A3: In Sections 3.4.2 and 3.5 of the paper, we explore and compare the effects of simple and hard open-set noise in two different cases on model classification accuracy. We would be happy to further clarify: the impact of 'hard' open-set noise and 'easy' open-set noise shows an opposite trend in two different cases. In the fitted case, 'easy' open-set noise is less harmful, whereas in the memorized case, the impact of 'hard' open-set noise is comparatively smaller. > Q4: Limited Experimental Scope: The experimental validation focuses predominantly on synthetic datasets like CIFAR100-O and ImageNet-O. While these are commonly used in the research community for benchmarking, the real-world implications of the findings might be limited without additional testing on more varied and real-world datasets. A4: Please refer to A2 for Q2. Thanks once again to the reviewer. If there are any further questions, we would be happy to discuss them at the next stage.	Summary: This paper focuses on open-set label noise problem. Authors first formally extending closed-set transition matrix to open-set transition matrix and define two noise ratios for open-set and closed-set separately. Then authors define error inflation rate as a measurement for noisy label impact and measure for two conditions, classifier fitted noisy distribution or memorized (overfit) noisy label. Later, authors propose a new type of open-set noise by exclusively transitioning outlier classes to a specific inlier class, and consider this as a "hard" open-set noise and traditional open-set noise as "easy" case. Authors further analysis two noise types on two classifier conditions and claim traditional entropy based open-set detection might only works on "easy" case. Experiments are performed on CIFAR-100, ImageNet and Webvision datasets. Strengths: Authors formally define open-set noise with a similar symmetric/asymmetric setup as closed-set noise, and find out that it shows opposite trend with different classifier cases. Weaknesses: - The experiment parts lack of baselines. With a new type of noise proposed, previous baselines on easy open-set noise should be run to assess the performance gap and set up the benchmark. - Figure 4 (a) and (b) have similar distribution, it is hard to draw conclusions from entropy dynamics. - Supp E.1 results are confusing. "X+EntSel" should be a better strategy since it selects inlier clean samples. However, why the closed-set classification accuracy is always the worst? Table 1 Webvision result is similar as well. Why is the claim "EntSel + SSR improves open-set detection performance" valid? The Acc and AUC are both dropping after adding EntSel. Why SSR/DivideMix + EntSel is always the worst performance? Considering it is a combination of inlier and clean, shouldn't it be the best performing one? I assume this is still the normal accuracy and AUC, which is the higher the better. Technical Quality: 3 Clarity: 2 Questions for Authors: - what is the pre-trained weight fitted case used? Confidence: 3 Soundness: 3 Presentation: 2 Contribution: 2 Limitations: Authors do not address any limitations in conclusion. A possible limitation might be related to approximation of fitted and memorized classifiers. Flag For Ethics Review: ['No ethics review needed.'] Rating: 5 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thanks very much for the careful review. We sincerely appreciate your time and effort in reading our paper, as well as the insightful and constructive feedback. >Q1: The experiment parts lack of baselines. ..., previous baselines on easy open-set noise ... set up the benchmark. A1: Thanks very much for the suggestion. We note that, in the Appendix E, we have included results for the DivideMix and SSR methods on the newly proposed dataset. Following the reviewer’s suggestion, we include results in *Table. 3* with two additional methods—EvidentialMix [1], DSOS [2] —on the CIFAR100-O dataset below. \| Method / Noise Ratio \| 0.2 Easy \| 0.4 Easy \| 0.2 Hard \| 0.4 Hard \| \|---\|---\|---\|---\|---\| \| SSR \| 0.889 \| 0.875 \| 0.895 \| 0.871 \| \| DivideMix \| 0.783 \| 0.754 \| 0.738 \| 0.675 \| \| EvidentialMix[1] \| 0.884 \| 0.827 \| 0.898 \| 0.872 \| \| DSOS[2] \| 0.846 \| 0.765 \| 0.854 \| 0.832 \| Table. 3 Results on CIFAR100-O We can generally agree that different methods exhibit varying sensitivities to open-set noise. Please note that analyzing the performance of these methods is not straightforward and requires further exploration, as they often comprise multiple modules and regularizations, which is beyond the scope of this paper. We would also like to clarify that proposing new noise type and benchmarking current SOTA methods is not our primary focus. We, instead, focus on the impact of different open-set noise (compared to closed-set noise, and compared with each other) in the standard classification evaluation protocol. [1] Sachdeva et al. "EvidentialMix: Learning with combined open-set and closed-set noisy labels." WACV. 2021. [2] Albert et al. "Addressing out-of-distribution label noise in webly-labelled data." WACV. 2022. >Q2: Figure 4 (a) and (b) ... it is hard to draw conclusions ... A2: Thanks for the comment. Since the noise ratio in Fig. 4(a) and Fig. 4(b) are relatively low, we kindly recommend enlarging the images to better visualize the differences (see an enlarged version in the 1-page PDF for your convenience). We will also update the manuscript. It can be observed that in Fig. 4(a) compared to Fig. 4(b), after a certain epochs of training, the easy open-set noise exhibits significantly different entropy values compared to the clean samples, while the hard open-set noise is difficult to be distinguished. This trend is consistent with what is observed in Fig. 4(c) compared to Fig. 4(d). >Q3: Supp E.1 results are confusing...., why the closed-set classification accuracy is always the worst? ... Why is the claim "EntSel + SSR improves open-set detection performance" valid? ... A3: We greatly appreciate the thorough review. To clarify, let us briefly revisit the three methods in the appendix: SSR/DivideMix, EntSel, and SSR/DivideMix + EntSel. When integrating EntSel, we adhered to the structure of SSR/DivideMix, which generally comprises two main modules: sample selection and model training. Specifically, we retained the training module of both methods: - SSR/DivideMix: the original method. - EntSel: we replaced the original sample selection module in SSR/DivideMix with EntSel. - SSR/DivideMix + EntSel: we selected the intersection of samples chosen by EntSel and those chosen by the original sample selection module. We will revise the manuscript to better clarify the notation. Regarding the reasons for the decrease in model performance, we found that SSR/DivideMix + EntSel selected a significantly smaller subset of samples compared to either SSR/DivideMix or EntSel alone (with default hyperparameters for all). This is expected because the intersection is considered. A possible cause then for the performance decrease is the precision-recall trade-off in the sample selection process. As the reviewer noted, "it selects cleaner samples," which leads to higher precision in sample selection but also results in lower recall. For instance, on the WebVision dataset, this suggests that, for DivideMix/SSR, the negative impact of discarding clean samples to eliminate noisy ones outweighs the negative impact of including slightly more noisy samples. Additionally, the strategy of using intersection is not optimal—effectively integrating open-set sample detection mechanisms with existing sample selection methods presents an interesting area for future research, though it lies beyond the scope of this work. "Why is the claim "EntSel + SSR improves open-set detection performance" valid? " As explained above, we are actually refer to the EntSel method (EntSel) rather than SSR + EntSel. As shown in second row of Fig. 6 and Table 1, we find that replacing the original sample selection module of the SSR method with EntSel may increase the open-set detection performance under open-set noise conditions. > Q4: what is the pre-trained weight fitted case used? A4: We used a standard `ResNet18` model with pretrained weights from the `torchvision` package. To accommodate the input size, we manually resized the CIFAR-10 dataset to 256x256 in this case. We will update the manuscript to make this clearer. > Q5: ... approximation of fitted and memorized classifiers. A5: A model $f$ under training is affected by many factors (model capacity / dataset size / training hyperparameters, etc.) - To enable an offline analysis, we would like to note that all theoretical analyses require some approximations/assumptions for tractability. However, the two cases that we introduce, i.e.,: the memorized case and the fitted case, are realistic and important. The memorized case could correspond to overfitting a high capacity model on a small, noisy dataset. By contrast, the fitted case could correspond to fine-tuning a linear classifier using a pre-trained model. In Section 4.1, we show experimental results consistent with corresponding theoretical findings respectively. Thanks once again to the reviewer. If there are any further questions, we would be happy to discuss them at the next stage. --- Rebuttal Comment 1.1: Comment: I have read the rebuttal. Here are my comments for questions: - Thanks for the new results, this is what I have in mind. - I can see the trend now. - I generally agree with the argument of precision-recall tradeoff, as it was also mentioned in other literatures [1]. It would be interesting to study EvidentialMix + EntSel, as it selects open-set with a separate interval rather than DivideMix clean/noisy. - Thanks for clarification. Although I do not think resizing is necessary. ImageNet is much larger than CIFAR and CIFAR images is only 32 * 32. - I think this could cause approximation error is not because these two cases are wrong. In general, I agree fitted case and memorized case are realistic. My understanding is that there is no precise way to quantify what is fitted and what is memorized, and of course we can approximate heuristically as what the manuscript does. But it is not fully justified. Neverless, this belongs to grand research direction and is beyond the scope of this paper. My remaining questions are how to benefit downstream tasks with these findings, including integrating with existing methods, effectively improve close/open set classification performance and sample selection. I am happy to raise my score. [1] Cordeiro, Filipe R., et al. "Longremix: Robust learning with high confidence samples in a noisy label environment." Pattern recognition 133 (2023): 109013. --- Reply to Comment 1.1.1: Comment: Many thanks to the reviewer for the response and recognition. We agree that it would be very interesting to explore how our findings can help improving existing techniques, and this could be a promising direction for the future research. For example, the opposite trends of 'hard' open-set noise and 'easy' open-set noise under the fitted case and memorized case could potentially suggest useful guidance for sample selection with pre-trained models.	Summary: The paper refines the problem of learning with noisy labels (LNL) by addressing the often overlooked issue of open-set noise. It provides a comprehensive theoretical analysis comparing the impacts of open-set and closed-set noise, introduces novel datasets for empirical validation, and explores the effectiveness of entropy-based noise detection mechanisms. Strengths: - The paper offers a thorough theoretical analysis of the differences between open-set and closed-set noise, extending the current understanding of LNL. - The exploration of entropy-based mechanisms for detecting open-set noise adds a practical tool for improving LNL methods. - This paper is well-written and easy to understand. Weaknesses: - The author summarizes two types of open-set noise, i.e., the easy and the hard noise, which is very similar to the symmetric and asymmetric label noise from the perspective of the transition matrix. So does there exist the instance-dependent open-set noise? What is its form if exists? - In Section 3.5, the author conducts analyses regarding entropy dynamics-based open-set detection, which belongs to the Fitted case. If adopting the vision language model (such as CLIP) to fine-tune and detect the open-set noise, is it aligned with the Memorized case? It would be better for the author to provide a real-world application for the memorized case. - The author should clearly illustrate the construct method of closed-set in the experiment (Figure 3) for reproducibility. Technical Quality: 3 Clarity: 3 Questions for Authors: See weaknesses Confidence: 2 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: See weaknesses Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thanks very much for the careful review. We sincerely appreciate your time and effort in reading our paper, as well as the insightful and positive feedback. > Q1: The author summarizes two types of open-set noise, i.e., the easy and the hard noise, which is very similar to the symmetric and asymmetric label noise from the perspective of the transition matrix. So does there exist the instance-dependent open-set noise? What is its form if exists? A1: Thanks very much for the question. We would like to note that the proposed complete noise transition matrix (L113 in the paper, Definition 3.1) also applies to instance-dependent open-set noise as well. The key difference between instance-dependent noise and the referred symmetric and asymmetric noise (also known as class-dependent noise) is whether we assume that samples from the same class share the same noise transition matrix. Our theoretical analyses are based exclusively on single samples and are therefore applicable to a variety of open-set noise forms. > Q2: In Section 3.5, the author conducts analyses regarding entropy dynamics-based open-set detection, which belongs to the Fitted case. If adopting the vision language model (such as CLIP) to fine-tune and detect the open-set noise, is it aligned with the Memorized case? It would be better for the author to provide a real-world application for the memorized case. A2: We appreciate the reviewer bringing up such an interesting situation. Theoretically, fine-tuning a large-scale visual language model like CLIP, without freezing the encoder, is likely to cause the model to memorize noisy labels (memorized case). We attribute this to the model's tendency to overfit when the dataset predominantly contains a single label for each sample, particularly given the large capacity of deep models (as discussed in lines 166-169 of the paper). Generally speaking, in most real-world scenarios, when there are enough training budgets (long enough training epochs for example), most of the deep neural networks typically have sufficient capacity to overfit nearly all samples in the dataset. We can thus assume that the memorized case corresponds to the learning scenario in most of these situations. > Q3: The author should clearly illustrate the construct method of closed-set in the experiment (Figure 3) for reproducibility. A3: Apologies for the confusion. Since our focus is on open-set noise, we default to the simpler symmetric closed-set noise (randomly flipping the labels of each sample) for more efficient comparison (L 415-416). In the updated manuscript, we will highlight its related details under a separate heading titled "Closed-set Noise.Though, please note that in our theoretical analysis, we do not make any assumptions about the form of closed-set noise; our theoretical results apply to both symmetric and asymmetric closed-set noise. Thanks once again to the reviewer. If there are any further questions, we would be happy to discuss them at the next stage. --- Rebuttal Comment 1.1: Comment: In response to the author's answer to Q3, I have a follow-up question: can the provided theoretical results be extended to instance-dependent noise, where the noise label is strongly correlated with the features? --- Reply to Comment 1.1.1: Comment: Many thanks for the reviewer's reply. We would like to confirm that our theoretical analysis also applies to instance-dependent noise, as we do not assume that samples from the same class follow the same noise transition matrix. We would be happy to engage further disucssions if you have any other concerns.	Rebuttal 1: Rebuttal: We thank the reviewers for their insightful comments. We are encouraged that they find our work comprehensive and well-structured, with thorough theoretical analysis and interesting findings. Specifically, we appreciate the recognition of our method's novelty and its empirical validation through novel datasets (6bHU, gSVD), the significance of the research problem (6bHU, gSVD, AwyQ), the acknowledgment of our method's solid theoretical foundation (6bHU, gSVD, AwyQ) and its interesting/significant findings (6bHU, 3Ttr, gSVD). Here, we would also like to briefly reiterate our key contributions again: - Generalized the noise transition matrix to incorporate open-set noise, reformulating the LNL problem. - Introduced practical 'memorized' and 'fitted' case scenarios for in-depth offline analysis. - Compared open-set and closed-set noise, revealing nuanced impacts on classification accuracy. - Differentiated between 'hard' and 'easy' open-set noise, uncovering its contrasting trends in different scenarios. - Proposed open-set detection as a complementary evaluation metric and conducted preliminary empirical validations. - Created novel synthetic datasets and an open-set test set for rigorous experimentation. We have provided detailed individual responses to each reviewer's concerns. If the reviewer has any further questions or concerns, we would be more than happy to engage in additional discussions. Pdf: /pdf/069af4af26babbfbc3c67a9a133245b9f464ba80.pdf	NeurIPS_2024_submissions_huggingface	2,024	null	null	null	null	null	null	null	null
Tell What You Hear From What You See - Video to Audio Generation Through Text	Accept (poster)	Summary: This paper proposes a multimodal video-to-audio generative model called VATT that can generate audio and audio captions given input silent videos and optional text prompts. It is capable of performing controllable text-guide video-to-audio generation and video-to-audio captioning. The framework consists of two parts: i) VATT converter, an instruction fine-tuned LLM that projects video features into language space. To train such a model, the authors caption the existing audio-visual dataset by LTU-13B. ii) VATT audio, a bi-directional transformer audio decoder based on Maskgit to generate audio tokens and text tokens with given inputs by iterative parallel decoding and Encodec as audio tokenizer. Through the quantitative and qualitative experiments on the VGGSound dataset, they demonstrate that the proposed method achieves state-of-the-art performance in the aspect of audio quality and inference speed. Strengths: - The paper is well-written and easy to follow. The motivation is clear. All details are clearly stated in the paper and appendix. - The paper has several technical highlights, especially on the V2A Instruction Tuning part which enables text control on the video-to-audio generation. - The authors perform the throughout experiments and ablations to demonstrate its technical advance. The proposed method outperforms baselines on the quantitative evaluation and human study. Weaknesses: - The reviewer thinks the major contribution lies in the V2A Instruction Tuning part which enables the text control on the video-to-audio generation. While the contribution of maskgit-based audio decoder seems to be incremental. - From the reviewer's perspective, using the VATT converter to encode visual information will lose some temporal information from the video. From the provided qualitative examples, the reviewer thinks the proposed method is not performing quite well on the generated synchronized audio but only roughly correct. Technical Quality: 3 Clarity: 3 Questions for Authors: The reviewer has questions about the paper: - The reviewer wonders why Maskgit as the decoder framework. From the reviewer's side, the inference speed seems to be the main advantage. - For the visual encoder eva-clip, it encode the feature frame-by-frame. The reviewer thinks it might lose temporal information, do authors explore this with a different encoder? - The reported inference speed of Diff-Foley is much slower than the number reported in their paper, any clue why it is happening Confidence: 5 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: Yes, the authors have adequately addressed the limitations. Flag For Ethics Review: ['Ethics review needed: Research involving human subjects'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We thank the reviewer for the thoughtful review and valuable feedback. We are glad to see that the reviewer acknowledges the presentation of our work, the technical highlights and thorough experiments. We address the feedback below. W1. Indeed, one major contribution of our work is the V2A tuning part which allows both video-to-audio captioning as well as video-to-audio generation via text control. For the audio decoder part, our methods are inspired from previous works like MaskGIT and Soundstorm. However, we still want to emphasize our key differences from these works. MaskGIT explores only the task of class-label conditioned image generation, and it is not designed to work on cross-modal generation like video-to-audio generation. For Soundstrom, it enables text-conditioned audio generation. However, its masking designs, in particular the masking ratio distribution, are different from ours. We conducted an ablation study (see Table 5) to study how the masking distribution design affects the performance of generation quality. The experiment could give some insights on how to design the masking ratio, which was not explored in previous works. W2 & Q2: We agree with the reviewer that frame-by-frame eva-CLIP features will lose fine-grained temporal details. At the current stage, we do not find any more powerful video encoder techniques to efficiently extract these fine-grained spatio-temporal features for better synchronizations. We will add this as part of the limitations in our revised manuscript, and motivate for the future research in this direction. Q1: We agree with the reviewer that one of the key motivations to use masking-based decoder is that it enables parallel decoding, which makes the inference speed a lot faster. The other reason we aim to use masking-based decoder is that we aim to unify different modalities using token representation, such that this kind of framework could be easily worked with multi-modal generation purposes within one model. In such cases, the choices are masking-based decoder and autoregressive decoder. For efficiency purposes again, we design our model's framework with masking-based decoder. Q3: For all inference speed evaluation, we perform generation with batch size = 1, while we found that Diff-Foley performs inference in a batch size of 64 and the inference time is calculated by per batch inference time divided by 64, which is the reason that their reported inference speed is significantly faster. For all other parts, we kept the same setting as theirs and used their Github repo’s notebook to perform inference. We will clarify this in our revised manuscript. Flag for Ethics Review: We obtained Institutional Review Board Approval (IRB Approval) from our institution on conducting human subject evaluations for this research project before the submission deadline. Per further request, we could supply the evidence in an anonymous way to the reviewing board committee. --- Rebuttal Comment 1.1: Comment: Thanks for the authors' response. The reviewer's concerns are mostly addressed. I'd like to keep my positive score. --- Reply to Comment 1.1.1: Comment: Thank you again for your insightful comments, suggestions, and appreciation of our work! We will adjust our manuscript to reflect all the changes made in the rebuttal.	Summary: The paper introduces a two-stage model for video-to-audio generation controlled by text. In the first stage, a large language model (LLM) generates audio captions based on the video content. In the second stage, the model uses video frames and the generated audio captions to predict masked audio tokens, thereby generating the audio. Strengths: 1. An anonymous website is provided to showcase the audio generation results, clearly illustrating the effectiveness of the model. 2. The proposed method can accomplish both video-to-audio generation and captioning tasks, and it demonstrates superior performance compared to the mentioned audio generation models. Weaknesses: 1. The content of the paper is densely packed, with many details relegated to the appendix, making it somewhat difficult to understand. 2. There are some grammatical errors in the paper, such as in lines 133-134 where it states "showed improvement in multi-modal understanding in in tasks such as captioning and question-answering." 3. The paper's novelty is somewhat limited. I noticed that the bidirectional-attention, Iterative Parallel Decoding, and masking design mentioned in the paper were already proposed in reference [1]. How do these differ from what is presented in this paper? If the only contribution is integrating video input into a large language model, I believe this approach alone is insufficient to support the acceptance of the paper. [1] Zalán Borsos, Matt Sharifi, Damien Vincent, Eugene Kharitonov, Neil Zeghidour, and Marco Tagliasacchi. Soundstorm: Efficient parallel audio generation. arXiv preprint arXiv:2305.09636, 2023. Technical Quality: 2 Clarity: 2 Questions for Authors: 1. See the 3rd item in “Weakness”. 2. In the quantitative evaluation of video-to-audio captioning, comparing the proposed model with LLAVA, which only accepts single image input, seems unfair. Why not compare it with large language models designed for video understanding, such as VideoChat or Video-LLaMA? Confidence: 4 Soundness: 2 Presentation: 2 Contribution: 2 Limitations: Lack of novelty, too dense content Flag For Ethics Review: ['No ethics review needed.'] Rating: 5 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We thank the reviewer for the insightful feedback. We are glad that the reviewer is satisfied with the quality of provided samples and generation capability of our model, and that our model is capable of performing both video-to-audio captioning and video-to-audio generation tasks. We address the concerns raised by the reviewer below. W1: We aimed to incorporate the most important experiments in the main section of our manuscript. However, we find it difficult to include all ablation studies as well as qualitative samples within the page limit due to the fact that our model works simultaneously on two tasks. We will try our best to adjust some part of the contents so that it becomes easier to read. W2: We will make sure to correct these grammar errors and typos in our manuscript through careful proofreading upon revision. W3 & Q1: In terms of technical aspect, the main differences of our proposed method from Soundstorm lie in the masking design. 1. Soundstorm only tried an arc-cosine masking ratio distribution similar to the one proposed in MaskGIT. However, we explored different masking ratio distributions during training and compared them in Table 5. The study provides critical insights regarding which masking ratio distribution would be suitable for audio generation. In general, the model performs better when the distribution has a higher mean in masking ratio. This is due to the fact that initial steps during the sampling stage are important for future decoding steps. And these initial steps are in the situations of high masking ratio. For later steps, new tokens are unmasked conditioned on more clues as the masking ratio goes lower, the generation becomes less challenging, thus making the learning at lower masking ratio easier during the training. 2. Soundstorm designed a complicated masking pattern, where any selected token will be masked together with tokens at higher levels than it. While our approach does not require this pattern dependencies and still performs well. We will highlight these technical similarities and differences in our revised manuscript. Aside from directly comparing masking novelty of VATT, we would like to reiterate that additional key novelty in our work is integrating LLM into the audio generation pipeline which enables the capability of suggesting an audio for the video through captioning, as well as generation of audio from video with text feedback. This cannot be achieved by mere combination of existing approaches since the ability to produce text as well as encode text as a condition for the output of another modality like audio is not achievable using these commonly used off-the-shelf text encoders like BERT or T5. Our work is the first to leverage the capability of LLM to achieve video-to-audio generation through text. Such a unique design boosts interactivity via text control as well as provides interpretability of the model by providing captions, which we believe is an important novelty of our method. As mentioned by Reviewer BS9v, “The paper has several technical highlights, especially on the V2A Instruction Tuning part which enables text control on the video-to-audio generation.” Q2: We further compare our method against existing video LLM models. We experiment with Video-LLAMA as suggested by the reviewer in two ways: 1. We conduct zero-shot captioning using Video-LLAMA to generate audio captions for the VGGSound test dataset. We tried different prompts and ended up using “User/ What sounds could match the video?” as it allows the model to follow the instruction closely. 2. Because Video-LLAMA has not been trained on VGGSound dataset and LTU generated captions, we implement a similar structure of Video-LLAMA and then train it on our LTU-generated captioning data. Specifically, we replaced the BLIP-2 visual features with our eva02-CLIP-L visual features due to the expensive pre-processing time for all videos in VGGSound and AudioSet data. For the Video-QFormer component of Video-LLAMA, we keep it the same as Video-LLAMA, and we name this model as VATT-Qformer LLama. As shown in Table 1R3, Video-LLAMA’s zero-shot video-to-audio captioning performance stays close to LLAVA albeit utilizing all of the video frames. When we train a similar structure to Video-LLAMA from scratch on LTU captions, we find that its performance is close to VATT-Converter LLama. The linear projection adopted by us and the Qformer projection methods don’t make much difference in our video-to-audio captioning task, and we choose a simpler way of projection. Table 1R3: Comparison of video-to-audio captions on NLG evaluation metrics and text-audio relevance (CLAP Score). Bold denotes the new results. \| Methods \| BertScore (F1) ↑ \| BLEU-4 ↑ \| ROUGE-L ↑ \| CIDEr ↑ \| CLAP Score ↑ \| \|----------------------------\|------------------\|----------\|-----------\|--------\|--------------\| \| LLAVA w/ Visual Prompt \| 0.855 \| 0.089 \| 0.137 \| 0.026 \| 0.213 \| \| LLAVA w/ Audio Prompt \| 0.870 \| 0.123 \| 0.155 \| 0.095 \| 0.182 \| \| Video-LLAMA w/ Audio Prompt \| 0.861 \| 0.091 \| 0.117\| 0.021 \| 0.204 \| \| VATT Converter - Gemma \| 0.900 \| 0.345 \| 0.337 \| 0.926 \| 0.229 \| \| VATT-Qformer - LLama \| 0.907 \| 0.419 \| 0.375 \| 1.264 \| 0.245 \| \| VATT Converter - LLama \| 0.909 \| 0.424 \| 0.384 \| 1.354 \| 0.263 \| --- Rebuttal Comment 1.1: Comment: The author's response basically solved my confusion, so I chose to boost my score to borderline accept. --- Reply to Comment 1.1.1: Comment: Thank you so much for your valuable feedback and time devotion for reviewing the work along with the rebuttal! We will incorporate these modifications in our revised manuscript.	Summary: In this paper, the authors propose VATT, a multi-modal generative framework for text-guided video-to-audio generation. VATT consists of two key modules VATT Converter and VATT Audio. The former component maps video features to LLM vector space with a projection layer. The latter one generates audio tokens with a bi-directional transformer. Experiments on datasets VGGSound and AudioSet-2M demonstrate that the proposed VATT framework surpasses existing SOTA in terms of KLD score. Strengths: 1. The authors identified limitations within the existing video-to-audio generative modeling techniques, indicating a clear understanding of the challenges in the field. 2. The publication of the generated samples is also commendable. Weaknesses: 1. As stated by the authors, the diversity of text generated by LLM could influence the generated audio. However, it is unclear how the quality of the generated caption affects the results. It is suggested that the authors include an ablation study comparing the results of different captions quality and choose some examples to analyze the reasons. 2. In the section on quantitative results, the proposed method is only evaluated on VGGSound, which is not enough to prove the generalization and robustness of the proposed method. It would be more convincing to conduct experiments and ablation studies on more datasets. 3. There seems to be a writing error in part D of Appendix, “we use existing an existing audio large language model…” Technical Quality: 2 Clarity: 3 Questions for Authors: See weaknesses Confidence: 5 Soundness: 2 Presentation: 3 Contribution: 2 Limitations: yes Flag For Ethics Review: ['No ethics review needed.'] Rating: 5 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We thank the reviewer for thoughtful feedback and acknowledging the quality of generated samples from VATT. We address the feedback and concerns below. W1: The quality of generated captions do affect the generation quality of the model in the V+T -> A stage. We conduct an additional ablation study by providing captions generated by different models as inputs to the same VATT model, i.e., VATT-LLama-T or VATT-Gemma-T to evaluate the impact of models’ captioning ability on the audio generation ability. We compare two generated captions from VATT-Converter - Gemma and VATT Converter - LLama. For reference, we also report the result of ground truth audio caption as inputs to the model. As shown in the Table 1R2, when the same model is fed with generated captions with different quality, the model’s audio generation performance are different. This effect is already expected from “Table 6: Ablation on self-prompting text-guided generation” in our paper. Combined with results in Table 3 in our paper, we can see that as the caption quality improves, the text-conditioned video-to-audio generation performance also improves, in particular from the FAD and KLD scores. Notably, the ground-truth audio captions generated by LTU has the highest CLAP score (measured with respect to the ground truth audio in original video) of 0.379 as stated in the paper, reflecting the best caption quality. Feeding such ground truth captions as input to the model also leads to the best audio generation results. As for the qualitative analysis on how the captions affect the audio generation quality, we select an example video with YouTube ID “j1kM-hC44Ok_000002.mp4” from VGGSound test set to analyze. The caption generated by VATT-LLama Converter is “A car is driving and crushing.”, while the caption generated by VATT-Gemma is “A car is driving and a man speaks with wind noise in the background.” As can be seen from the video, the car is crushing over some objects and making cracking sounds. The former caption is better aligned with the visual scene, and it generates the sound that matches closely to what is intended. While the latter caption is unrelated to the key event happening in the video, leading to poorer results. We will incorporate qualitative examples as well as demos on our website, and make more detailed analysis in our revised manuscript to help understand the correspondence between the caption quality and the generated audio. From another perspective, our design by enabling the model to take an extra caption as input is a useful feature since the user can provide reasonable text input to steer the generated sound with variations to some extent. In this aspect, our qualitative samples from the provided link in the paper as well as Figure 5 in the Appendix C demonstrate some interesting scenarios where a user provides different plausible captions for the model to generate. Table 1R2: The impact of caption quality on V + T -> A performance. \| Methods \| Text Prompt \| KLD ↓ \| FAD ↓ \| Align Acc ↑ \| \|----------------\|-------------------------\|-------\|-------\|-------------\| \| VATT-Gemma-T \| VATT-Converter Gemma \| 2.40 \| 3.67 \| 80.07 \| \| VATT-Gemma-T \| VATT-Converter LLama \| 2.26 \| 3.20 \| 80.42 \| \| VATT-Gemma-T \| Ground Truth \| 1.66 \| 2.98 \| 81.48 \| \| VATT-LLama-T \| VATT-Converter Gemma \| 2.57 \| 2.67 \| 79.20 \| \| VATT-LLama-T \| VATT-Converter LLama \| 2.38 \| 2.58 \| 80.41 \| \| VATT-LLama-T \| Ground Truth \| 1.41 \| 2.54 \| 80.16 \| W2: We use VGGSound as our main dataset and benchmark to evaluate since it is not only a large-scale audio-visual dataset with around 200K videos across many categories, but also the quality of audio-visual alignment is quite good. To further show the generalization capability of VATT, we experiment with AudioCaps dataset. Due to limited video samples in AudioCaps, we finetuned our VGGSound pretrained VATT model on AudioCaps dataset in two settings, with and without text prompts. To keep the comparison fair, we use the ground truth audio captions from AudioCaps as the text prompts. We use VATT-LLama and VATT-LLama-T to compare against AudioGen and AudioLDM-2. As shown in Table 2R2, VATT-LLama-T performs on a similar level to AudioGen in terms of FAD and KLD score, while falling behind AudioLDM-2. It is noteworthy that the audio decoder of both AudioGen and AudioLDM-2 are pretrained on much larger data scale (7000 hrs and 30000 hrs audio respectively) than ours (700 hrs audio). Despite this, our method still performs reasonably well on this dataset. We plan to conduct further ablation studies by scaling up the training by incorporating more data to train VATT on the video-to-audio generation task. Table 2R2: Quantitative results against text-to-audio generation methods on AudioCaps test set. \| Methods \| KLD ↓ \| FAD ↓ \| Align Acc ↑ \| CLAP Score ↑ \| \|---------------\|------\|------\|-----------\|------------\| \| AudioGen \| 2.09 \| 3.13 \| 58.26 \| 0.447 \| \| AudioLDM-2 \| 1.64 \| 1.86 \| 60.32 \| 0.432 \| \| VATT-LLama \| 2.53 \| 3.42 \| 75.76 \| - \| \| VATT-LLama-T \| 2.07 \| 3.25 \| 74.89 \| 0.376 \| W3: We will correct the grammar errors and typos like this in our revised manuscript through careful proofreading. --- Rebuttal 2: Comment: Dear Reviewer, We are very thankful for your valuable feedbacks and comments! Let me know if you have further questions regarding the rebuttal response or any other questions related to the paper. We look forward to your further feedbacks! Thanks, Authors of Paper Submission 1915 --- Rebuttal Comment 2.1: Comment: Since the discussion period ends soon, we want to reach out to see if you have any further advice or feedback and whether there is more information we could provide for reconsideration of your score? Please let us know and we will be happy to engage in further discussions. Thanks!	Summary: - The paper proposes a model for video-to-audio generation (main task) and video-to-audio captioning (auxilliary task). - The text is optionally used to control audio generation for ambiguous video cases. - They use two step approach i.e. video-to-caption stage and video+text-to-audio stage - Stage 1: video-to-caption - This stage basically converts videos and text to audio relevant captions - They train a VATT converter that uses LoRA finetuning to get audio captions. - This stage uses AudioSet and VGGSound and captions generated using LTU model. - Stage 2: video+text-to-audio - This stage generates audio tokens, given video+text as input. The video+text is converted to audio relevant captions using stage-1 - They use LM based audio token generation. The VATT Audio decoder learns to predict the masked audio tokens. During inference, iterating parallel decoding is used. - VGGSound is used to train this stage. - There model shows comparative and better performance on video-to-audio(V2A) generation task and text to audio generation task. Add text to V2A task further improves KLD score. - Stage 1 when compared to baselines can perform better on video-to-audio captioning. - Qualitative and human evaluation shows preference towards their work. Strengths: - The paper proposes a new problem of utilizing text for video-to-audio generation. - Several novel and interesting technical contributions. i.e. - Video+text conditioning to generate audio by converting this condition to joint audio-caption. - Using synthetic captions for training and LoRA for MLLM finetuning. - Using iterative parallel decoding and bi-directional self-attention. - Their approach improves on video-to-audio generation, and text-to-audio generation both quantitatively and qualitatively. - The results also show improvement in video-to-caption generation. Weaknesses: - Text captions generated by LTU are used as ground truth for video-to-audio generation, text-to-audio generation, and video-to-audio captioning. The authors should verify the correctness of the captions. - Table 2 results for Text-to-audio generation seem slightly unfair for the baselines. - The LTU-generated caption may have some biases or patterns of caption generation that VATT has seen during training while the baselines have not. - Though AudioGen has seen VGGSound tags, it has not seen the captions generated by LTU (which VATT has). - AudioLDM2 has not even seen VGGSound data for training. - A fair comparison would be finetuning/training from scratch Text-to-audio models on VGGSound with LTU-generated captions. Or VATT should be compared on AudioCaps (with videos downloaded from YouTube). - Table 3 results are also a little unfair: - Similar to the argument above LLAVA has not seen LTU-generated captions for VGGSOUND. - Also LLAVA takes only 1 frame while VATT takes more frames. - Maybe a fair comparison would be to train some captioning models on LTU-generated data. Minor comments: - The captions for the tables should come above the table. For eg. Table 1,2,3 - Missing reference: A previous work by Mo. et al [1] that utilizes text+video-to-audio generation. This should be mentioned - Align-acc for diff-foley is 82.47 in the paper vs 94.05 in the original paper [1] DiffAVA: Personalized text-to-audio generation with visual alignment Technical Quality: 3 Clarity: 3 Questions for Authors: My questions are based on the arguments mentioned in the weakness section: - Can the correctness of the LTU-generated captions be verified? Since they are used as the ground-truth for several tasks? - Can the baseline text-to-audio(TTA) generation models be trained on VGGSound+LTU captions? Can VATT be compared on AudioCaps? - Can VATT be compared on other video captioning datasets? Or the existing video-captioning models be finetuned/trained from scratch on this LTU data? - Why are some metrics different from the numbers quoted in the original paper? For eg. Diff-foley alignment score. Confidence: 4 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: Yes the authors adequately addressed the limitations. Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We thank the reviewer for their insightful review and valuable feedback. We appreciate the reviewer feedback that our proposed methods include novelties and technical contributions. We address the raised concerns below. W1 & Q1: We manually verified the validity of LTU-generated captions prior to using them as synthetic GT. As a result of reviewer feedback, we performed an experiment to further evaluate its correctness. We randomly selected 100 videos from VGGSound test set with stratified sampling according to video categories to conduct a human study. We use the 1-5 point MOS (Mean-Opinion-Score) scale (the higher the better) to measure correctness of the captions. We provide pairs of videos and the corresponding captions to the raters, asking “How accurately the provided caption reflects the sound events happening in the video? 1. Inaccurate and irrelevant. 2. Relevant but inaccurate with many mistakes. 3. Partially accurate but missing details and with mistakes. 4. Mostly accurate with some minor mistakes. 5. Accurate and complete.” We used the MTurk platform to perform the evaluation and collected a total of 300 responses. The generated captions have a high MOS of mean 4.72 and std 0.37, providing another indication for the validity of the synthetic ground truth. W2 & Q2: We thank the reviewer for asking regarding further comparison against existing text-to-audio methods. Due to the time constraints of the rebuttal, we chose to finetune our VATT model on AudioCaps benchmark instead of training AudioGen and AudioLDM-2 using LTU generated captions from scratch. Specifically, due to the limited data size of AudioCaps, we use our VGGSound pretrained checkpoint to finetune instead of training on AudioCaps from scratch. As shown in the Table 1R1, VATT achieves similar performance against AudioGen on FAD and KLD metrics, while still showing some gaps from AudioLDM-2. One clear reason is that AudioGen and AudioLDM-2 are pretrained on much larger audio-text data (7000 hrs and 30000 hrs audio respectively) than ours (700 hrs audio). In addition, we also find that the audio-visual alignment from AudioCaps dataset (videos from AudioSet) is not as good as VGGSound, often suffering from static frames and weak correspondence of audio and visual. Therefore, the visual condition is not as effective as the VGGSound videos. Table 1R1: Quantitative results against text-to-audio generation methods on AudioCaps test set. \| Methods \| KLD ↓ \| FAD ↓ \| Align Acc ↑ \| CLAP Score ↑ \| \|---------------\|------\|------\|-----------\|------------\| \| AudioGen \| 2.09 \| 3.13 \| 58.26 \| 0.447 \| \| AudioLDM-2 \| 1.64 \| 1.86 \| 60.32 \| 0.432 \| \| VATT-LLama \| 2.53 \| 3.42 \| 75.76 \| - \| \| VATT-LLama-T \| 2.07 \| 3.25 \| 74.89 \| 0.376 \| W3 & Q3: We appreciate the reviewer pointing to the need for further comparison on video-to-audio caption tasks. We have added two additional models/setups in the comparison. To address the concerns that LLAVA is only trained on single frames rather than the whole video, we instead use an existing video LLM , Video-LLAMA-7B, to perform zero-shot video-to-audio captioning. Specifically, following the Video-LLAMA’s Github repository instructions, we directly input the VGGSound videos into the VL branch of the model, and prompt it to generate audio captions using the instruction “User/ What sounds could match the video?” Since Video-LLAMA is still not pretrained on VGGSound dataset and LTU generated captions, we implement a similar structure of Video-LLAMA and train on our LTU-generated captioning data. We replaced the original BLIP-2 visual features used by Video-LLAMA with our eva02-CLIP-L visual features due to the expensive pre-processing time for all BLIP-2 features from videos in VGGSound and AudioSet. For the Video-QFormer component of Video-LLAMA, we keep it the same as Video-LLAMA, and we name this model as VATT-Qformer - LLama. As shown in Table 2R1, zero-shot performance of Video-LLAMA is similar to LLAVA, while VATT-Qformer - LLama trained from scratch performs very close to VATT Converter - LLama. It is note-worthy that the only difference between VATT Converter - LLama and VATT-Qformer - LLama is the projection method, and we find that the linear projection adopted by VATT Converter is enough for the task to perform well. Table 2R1: Comparison of video-to-audio captions on NLG evaluation metrics and text-audio relevance (CLAP Score). Bold denotes the new results. \| Methods \| BertScore (F1) ↑ \| BLEU-4 ↑ \| ROUGE-L ↑ \| CIDEr ↑ \| CLAP Score ↑ \| \|----------------------------\|------------------\|----------\|-----------\|--------\|--------------\| \| LLAVA w/ Visual Prompt \| 0.855 \| 0.089 \| 0.137 \| 0.026 \| 0.213 \| \| LLAVA w/ Audio Prompt \| 0.870 \| 0.123 \| 0.155 \| 0.095 \| 0.182 \| \| Video-LLAMA w/ Audio Prompt \| 0.861 \| 0.091 \| 0.117\| 0.021 \| 0.204 \| \| VATT Converter - Gemma \| 0.900 \| 0.345 \| 0.337 \| 0.926 \| 0.229 \| \| VATT-Qformer - LLama \| 0.907 \| 0.419 \| 0.375 \| 1.264 \| 0.245 \| \| VATT Converter - LLama \| 0.909 \| 0.424 \| 0.384 \| 1.354 \| 0.263 \| Minor Comments 1: We will move the table titles and captions above the tables to follow the NeurIPS format. Minor Comments 2: We will make sure to cite the reference mentioned by the reviewer as it is closely related to our work. Minor Comments 3 & Q4: We did find differences between our experiment results and the ones reported in Diff-Foley’s results. However, we strictly follow the guidelines and instructions from their official repository to perform audio generation on the VGGSound test set, and we could not achieve their claimed performance. We provided all the necessary implementation details for the baseline comparisons in Appendix F for reference. --- Rebuttal Comment 1.1: Comment: Thanks for the comprehensive explanation of my questions. Hence, I would keep my positive score. Please include the additional experiments and clarifications in the main text or appendix of the paper. --- Reply to Comment 1.1.1: Comment: Thank you for your detailed and thoughtful reviews, and appreciation of our work! We will include all the mentioned items in the rebuttal into our revised manuscript properly.	null	NeurIPS_2024_submissions_huggingface	2,024	null	null	null	null	null	null	null	null
Optimal Flow Matching: Learning Straight Trajectories in Just One Step	Accept (poster)	Summary: This paper aims to solve the optimal transport problem between two distributions within a flow matching framework. This can be achieved by finding the best velocity model $u_t$ in a class of optimal vector fields, i.e., $u_t$ be implicitly induced by defining the $x_1=\nabla \Phi(x_0)$ where $\Phi$ is a convex function. The authors prove that the resulting (optimal) flow matching loss is equivalent to the dual formulation of the OT problem and regression towards the dynamic OT field. However, optimizing the optimal flow matching loss can be computationally expensive, as the velocity model is only implicitly defined. Despite this, the study of the relationship between the optimal flow matching problem and optimal transport can be a strong contribution. Strengths: - In general, this paper is well-written and easy to follow. - Lemma 1 and its corollaries make a strong contribution to the community. It proves that the OT problem can be equivalently solved by simply optimizing the flow matching loss in the ``optimal vector fields'' family. In fact, this is a rather surprising result for me. - In theory, the proposed optimal flow matching method can learn the OT map after one round of training, regardless of the reference distribution. - The experiments are well-designed and convincing. The 2D experiment with different reference distributions (especially the anti-minibatch) is illustrative. Weaknesses: - The practical implementation of OFM is rather computationally expensive. It would be fine to add a table to directly compare the training times of different methods, or a figure showing the convergence as a function of computation time. I don't think this will lessen the contribution of this paper. - As the authors claim that OFM learns straighter trajectories, comparing OFM to other baselines on standard generation benchmarks, e.g., CIFAR10, would be better. Technical Quality: 4 Clarity: 3 Questions for Authors: On the image-to-image transfer task, do you use unpaired training data? If the answer is yes, is it costly to compute the minibatch OT? Still, in this case, what is the tensor shape of the 512-dimensional latent space? I think 512 is much smaller than other latent space models. Confidence: 3 Soundness: 4 Presentation: 3 Contribution: 4 Limitations: the authors adequately addressed the limitations Flag For Ethics Review: ['No ethics review needed.'] Rating: 7 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Dear reviewer, thank you for your feedback and kind words. Please find below the answers to your questions. (1) Convergence time. Plots of convergence. Following your request, we provide the plots of convergence (in $L^2$-UVP) depending on training time in the experiment on the W2 benchmark (Section 4.2) in dimension $D=32$. Please see Figure 3 in the PDF file. OFM iteration takes more time than iteration of OT-CFM, RF or $c-RF$ during the first FM round. After the first FM round, RF and $c-$RF slow down considerably and lose to OFM. Fortunately, OFM requires much fewer iterations to achieve the best metrics and outperforms all other methods in the first hours of the training. We will include this requested plot to the appendix of the final version of the paper. (2) Comparing OFM to other baselines on standard generation benchmarks. In theory, OFM can work directly in pixel-space (e.g., of CIFAR10) and has no limitation for dimensionality. However, in practice, the ML community has not yet designed input convex neural networks (ICNN) architectures that work well in the pixel-space (see, e.g., [1] for reference). Hence, for now image generation probably remains infeasible, and we encourage the ML community to invest into developing ICNN architectures. (3) On the image-to-image transfer task, do you use unpaired training data? Is it costly to compute the minibatch OT? Still, in this case, what is the tensor shape of the 512-dimensional latent space? I think 512 is much smaller than other latent space models. Yes, we use unpaired data. The latent space of ALAE autoencoder has dimension $D=512$ (e.g., the full encoded dataset has size $N\times 512$, where $N$ is the number of samples). Discrete OT computation in $512$-dimensional space actually is not very costly, since this task is much easier than the conjugation or hessian inversion used in training. OT computation requires $O(D\cdot B^3)$ time, where $B$ is the batch size (it is usually considerably smaller than dimension $D$). For comparison, the hessian inversion itself takes $O(BD^3)$. Concluding remarks for all questions. We thank the reviewer one more time for high evaluation of our work. Please respond to our post to let us know if the clarifications above suitably address your concerns about our work. We are happy to address any remaining points during the discussion phase. References. [1] Korotin, Alexander, et al. "Do neural optimal transport solvers work? a continuous Wasserstein-2 benchmark." Advances in neural information processing systems 34 (2021): 14593-14605. --- Rebuttal Comment 1.1: Comment: I thank the authors for their clarifications and additional experiments. I will keep my score.	Summary: This work proposes Optimal Flow Matching (OFM), which restricts standard Flow Matching to specific vector fields that yield straight trajectories by design. This is implemented by considering vector fields such that a convex function exists whose gradient pushes the initial to the final point. The authors propose to solve this optimization problem with training that involves solving a convex optimization problem in each iteration to obtain the inverse of the flow map. Then, it is shown that this procedure has the same minimizer as the one minimizing the dual OT loss. The authors empirically validate OFM quantitatively on the Wasserstein-2 benchmark and qualitatively on FFHQ. Strengths: - The authors derive a new training procedure based on Flow Matching restricted to fully straight vector fields implemented through the gradient of a convex function. - The authors provide a new insight into the connection between minimizing an OT loss and Flow Matching. Specifically, they show that the minimizer of these two losses is equivalent when restricted to straight vector fields. Weaknesses: Missing comparison and contextualization of related work: - In [1], it was proposed to approximate the computation of the conjugate with amortized optimization and solve this with Adam or L-BFGS. The proposed SubOpt in this work (solved with Adam/L-BFGS) is equivalent to the convex conjugate optimization when $t=1$. A discussion between these two optimization schemes should be included. Specifically, how does OFM compare to [1] when only sampling $t=1$? What is gained by considering $t\sim \mathcal{U}[0,1]$? Additionally, the Wasserstein-2 benchmark results from [1] are not considered in the paper. This is the most related method and outperforms the method proposed in the paper on all dimensions. These results should be included as well as a comparative discussion. Experiments: - No quantitative results are provided for the unpaired image-to-image translation experiments. It is hard to generalize the results over a few qualitative results. Thus, including some commonly used quantitative metrics like FID or the ones used in the Wasserstein-2 benchmark on CelebA64, as done in [1], would be very beneficial. Minor Weaknesses: - Inconsistency on the optimizer used to solve SubOpt. In Appendix C, it is stated that "in all our experiments as the subOpt optimizer we use LBFGS"[line 557] while in Appendix B, it says "(with Adam optimizer)"[line 534]. Which one is actually used? How do Adam and LBFGS compare? - Some of the wording can be improved, e.g. "... via fancy formula (18) is not trivial and requires tricky integration techniques."[line 194] [1] Brandon Amos. "On amortizing convex conjugates for optimal transport". In ICLR 2023. Technical Quality: 2 Clarity: 2 Questions for Authors: - The goal of OFM is to learn a time-independent one-step model that translates $x_0$ to $x_1$. Then, during inference, no ODE is solved. As mentioned above, some more motivation could improve clarity here. Specifically, explaining why it is beneficial to formulate this in the FM framework considering $t\sim \mathcal{U}[0,1]$. Would the proposed method also work when only sampling $t=1$? The authors only mention that the subproblem (20) is $(1-t)$-strongly convex, which gives a benefit compared to $t=1$. Are there other benefits of the proposed formulation? - In Figure 2, it is shown that OFM learns the same solution for different choices of couplings. In [2], it was shown that only the OT coupling is preserved when using standard Flow Matching. What do these illustrative examples look like for standard FM training? I assume they would learn different trajectories, but do the learned endpoints differ? - In this work, only ICNNs are considered. As shown in previous works, even when wanting to learn a convex function, MLPs can also work well, i.e., in [1], they work better compared to ICNNs. Does OFM also work with standard MLPs/CNNs/etc? How is the comparative performance? [1] Brandon Amos. "On amortizing convex conjugates for optimal transport". In ICLR 2023. [2] Valentin De Bortoli and Guan-Horng Liu and Tianrong Chen and Evangelos A. Theodorou and Weilie Nie. "Augmented Bridge Matching". In Arxiv 2023. Confidence: 4 Soundness: 2 Presentation: 2 Contribution: 2 Limitations: - The limitations of the flow map inversion, the Hessian computation, and ICNNs are laid out in the paper, and details for the experiments' computational time are provided. - While these computational times give some insight, detailed information on time and space complexity with respect to the input dimension would be even more beneficial. Additionally, comparing the runtime of OFM and standard FM training would provide a more illustrative comparison of raw computational time. Flag For Ethics Review: ['No ethics review needed.'] Rating: 5 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Dear reviewer, thank you for your detailed feedback. Please find below the answers to your questions. (1) The reduction of our OFM to sampling of $t = 1$ only. Our method requires considering all times $t\in [0,1]$ by its design. Considering $t=1$ breaks the theory and does not work. Indeed, this can be seen both theoretically and practically. In case $t=1$, loss (17) in OFM becomes $$\min\limits_{\Psi \text{ convex}}\int\limits_{\mathbb{R}^D \times \mathbb{R}^D} \left\|\left\|\nabla \overline{\Psi} (x_1) - x_0 \right\|\right\|^2 \pi(x_0, x_1) dx_0 dx_1.$$ Without $t\sim U[0,1]$, loss becomes plan-dependent, and OT recovery is not guaranteed. Hence, we lose both the desired connections with FM and OT. To also confirm this empirically, we run this optimization on a 2D Gaussian to Mixture of Gaussians setup for two plans: minibatch with 4 and 1024 samples in batch. The results are depicted in Figure 1 in the PDF file. One may clearly see that the learned map does not even match the target distribution. We also highlight that the objective above is not equivalent to the dual OT objective $\mathcal{L}_{OT}$ (equation (4) of our paper), which involves the potential $\Psi$ and its conjugate $\overline{\Psi}$. (2) Benefits of the dynamic formulation ($t\in [0,1]$) over the static one (in addition to the strong convexity). The main reason to consider $t \sim U[0,1]$ is to pave a new bridge between two emerging subfields in deep learning: optimal transport and flow matching. We introduce novel methodology relating both of these problems. This development potentially allows to use the advantage of one field to improve the methods in the other one. And, indeed, stronger convexity is a good benefit of the dynamic loss, but it is a bonus, not an initial goal. (3) Comparison/relation to the dual OT optimization approach [1]. We primarily aimed to compare flow matching-based methods with each other. Metrics from MMv1 are included for completeness, and they already give a general understanding about the quality and baselines of modern static OT methods. Method [1] surely demonstrates the better result, as it used advanced amortization techniques to optimize $\mathcal{L}_{OT}$. We will include the results of the method in the table. Overall, advanced amortization techniques such as [1] are orthogonal to our theoretical and methodological study and can be used on top of our method to further improve it. To provide the empirical evidence, we adapted the amortization scheme from [1] to our OFM. Namely, we find an approximate conjugation solution with an extra MLP $A_\phi(x_t,t):\mathbb{R}^D \times [0,1] \to \mathbb{R}^D $: \begin{eqnarray} A_\phi(x_t, t) \approx \arg\min_{z_0 \in \mathbb{R}^D} \left[ \frac{(1-t)}{2} \\| z_0\\|^2 + t \Psi (z_0) - \langle x_t , z_0 \rangle \right], \end{eqnarray} and repeat the training pipeline from [1] but with an extra condition on $t$. We run an additional OFM experiment on $D = 32$ with amortization and regression loss. This augmentation gives a boost for metrics in comparison with basic minimization techniques, see Figure 4 in the PDF file. In the final version of the paper, we will add an appendix about the amortization technique for OFM. (4) Inconsistency on the optimizer used to solve SubOpt. In Appendix C, it is stated that "in all our experiments as the subOpt optimizer we use LBFGS"[line 557] while in Appendix B, it says "(with Adam optimizer)"[line 534]. Which one is actually used? How do Adam and LBFGS compare? We are sorry for this inconsistency in our text. In all our experiments, we used the LBFGS optimizer. We kindly propose you to take a look at our code. We will fix the misprint in Appendix B in the final version. Regarding Adam, in our early experiments we indeed experimented with this optimizer, but eventually found that LBFGS works better (it is faster when achieving the same tolerance). We humbly think that the detailed comparison between the optimizers is out of scope of our work. (5) ICNNs are considered. Does OFM also work with standard MLPs/CNNs, etc.? Following your request, we conducted several additional experiments with MLP parameterization of $\Psi$ instead of ICNN. Unfortunately, they failed to converge. Probably, this happens due to violation of conjugated functions' properties for non-convex functions required in OFM. We leave studying this phenomenon for future works. (6) Quantitative results are provided for the unpaired image-to-image translation experiments. Following your request, we calculated FID for I2I experiments. See our General response section and Table 1 from the attached pdf. According to the results, our OFM method beats all the baselines. (7) How do the illustrative 2D examples look like for standard FM training with ID, MB, anmiMB plans? Do the endpoints/trajectories differ? Following your request, we run experiments with FM under independent, minibatch and anti-minibatch plans on 2D example, see Figure 2 in the PDF file. Unlike OFM with straight paths, FM trajectories differ a lot depending on plan, curvature increases in order "mb", "idp" and "anti-mb". We, in particular, highlight the high curvature of antimibatch's trajectories. The generated distributions of end points and the target distribution are close in all the cases. Concluding remarks for all questions. We thank the reviewer one more time for the interesting and important questions. Please respond to our post to let us know if the clarifications and additional experiments above suitably address your concerns about our work. We are happy to address any remaining points during the discussion phase; if the responses above are sufficient, we kindly ask that you consider raising your score. --- Rebuttal Comment 1.1: Comment: Thank you for the detailed comments and added experiments. This cleared up a lot of my questions, and I've decided to raise my score 4->5 as I think the paper has an interesting theoretical contribution. I still have concerns regarding the scalability of the method because of its reliance on ICNNs making it very hard to scale up to real word examples like pixel-space image generative modeling. The authors mention that the main competitive methods of their proposed method are different Flow Matching methods but the work does not include results on any common generative modeling benchmarks as also mentioned by Reviewer Bhuh. These other FM methods don't have any problems scaling up to arbitrary dimensions. This should be sufficiently discussed and contextualized in the revised manuscript. Also as mentioned in my rebuttal, I believe [1] is a more closely related work to the proposed method with respect to the Optimal Transport benchmark and outperforms it across all dimensions. --- Reply to Comment 1.1.1: Title: Thank you! and concluding comments Comment: Dear Reviewer, we thank you for your answer and for highlighting the theoretical contribution of our work. We are very grateful for the score increase. Let us leave some concluding comments. 1. We agree that the reliance on the ICNNs makes the practical adaptation of our methodology to high-dim generative image setups to be tricky. Following your suggestions, we will expand the corresponding limitation subsection (ICNNs) accordingly. 2. The work [1] which is based on the amortization methodology will receive due attention in our revised manuscript. In particular, we acknowledge that we will include the benchmark results of [1] in our Table. With regard to the superior performance of [1], we humbly believe that achieving the SOTA metrics on the benchmark lies primarily on the surface of technical solutions. I.e., particular architectures, optimizers, schedulers, regularizers etc. To support our point, we provide the additional results for our method fitted on the benchmark, but with additional exponential moving average (EMA) of the trained model weights. Note that EMA does not change the training process itself - it just creates a smoothed copy of the model whose weights are updated (at each new training iteration $t + 1$) as $W^{\text{ema}}\_{t + 1} = \alpha W^{\text{ema}}\_{t} + (1 - \alpha) W\_{t + 1}$, where $W_{t + 1}$ are the newly updated original trained weights. Our results for $\mathcal{L}^2$-UVP are in the table below ($\alpha = 0.999$). For completeness, we also recall the metrics for our models without EMA and additionally grab the benchmark results from [1]. \| Solver/dim \| 2 \| 4 \| 8 \| 16 \| 32 \| 64 \| 128 \| 256 \| \|-------------------\|----------\|----------\|---------\|---------\|---------\|---------\|----------\|----------\| \| OFM MB (ema) \| $\textit{0.15}$ \| $\textit{0.52}$ \| $\textit{1.2}$ \| $\textit{1.0}$ \| $\textbf{1.2}$ \| 7.2 \| $\textit{1.5}$ \| 2.9 \| \| OFM MB \| 0.33 \| 0.81 \| 1.5 \| 2.0 \| 3.4 \| 10 \| 6.7 \| 11.7 \| \| [1] (ICNNs) \| 0.26 \| 0.78 \| 1.6 \| 1.1 \| $\textit{1.9}$ \| $\textit{4.2}$ \| 1.6 \| $\textit{2.0}$ \| \| [1] (MLP) \| $\textbf{0.03}$ \| $\textbf{0.22}$ \| $\textbf{0.6}$ \| $\textbf{0.8}$ \| 2.0 \| $\textbf{2.1}$ \| $\textbf{0.67}$ \| $\textbf{0.59}$ \| As we can see, the utilization of a rather cheap technique allows obtaining much better results, compared to those reported in our article. Regarding [1], it seems they did not utilize EMA in their experiments. Probably, the usage of EMA in their experiments also will lead to better results. Our newly provided table is not about competing with [1]. We just want to stress that the point of our benchmark experiments is to demonstrate the applicability and success of our proposed methodology - and we humbly think that we managed to achieve this goal - by demonstrating rather good and close-to-the-best metrics. One more time, we are grateful to the reviewer for the valuable comments, and we will incorporate the necessary edits in the revised manuscript. --- Rebuttal 2: Title: The discussion period deadline is approaching... Comment: Dear reviewer nvT5. Thank you for your time efforts to review our paper. The discussion period is coming to the end. We would appreciate your feedback on our responses. We are happy to answer any further questions during the remaining discussion period. Thank you in advance, Authors.	Summary: The paper goal is to learn the optimal transport map between any two distributions. The approach derives a flow matching loss which is minimized by the velocity field of the optimal transport map. Strengths: The paper investigate the interesting subject of the optimal transport problem. Weaknesses: The paper is written in an unprofessional form: 1. Theorems and propositions are structured in a way which makes it very hard to follow. 2. The author does not make any attempt to give an accessible main proof idea/proof sketch in the main paper. Specifically, it is expected to provide one for Theorem 1 which seems to be the key claim of the paper. 3. The proof of Theorem 1 in the appendix is unclear and unsatisfying. 4. The experiment section is rather minimal. In particular, it is not clear what Figure 3 in the Image-to-Image task presents. Technical Quality: 2 Clarity: 1 Questions for Authors: In essence, does the author claim that the $\mathcal{L}^{\pi}\_{OFM}(\Psi)$ loss presented in equation (16) is minimized by the velocity field of the optimal transport map for any plan $\pi$? As stated by the author, it seems this loss is by definition Flow Matching (FM) loss constrained to convex functions. For the independent plan, i.e., $\pi=p_0\times p_1$ the minimizer for the FM loss is explicitly given in [1], $$ u_t(x) = \frac{1}{p_t(x)}\int u_t(x\|x_1)\frac{p_t(x\|x_1)p_1(x_1)}{p_t(x)}dx_1, $$ where for the linear interpolant $u_t(x\|x_1) = \frac{x_1-x}{1-t}$. This minimizer is dependent solely on the source and target distribution (i.e., the plan for this case) and does not yield the optimal transport. If you could please explain how you overcame this. [1] Lipman, Yaron, et al. "Flow matching for generative modeling." arXiv preprint arXiv:2210.02747 (2022). Confidence: 2 Soundness: 2 Presentation: 1 Contribution: 2 Limitations: - Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Dear reviewer, we thank you for your feedback. Please find below the answers to your questions. (1) In essence, does the author claim that the loss presented in equation (16) is minimized by the velocity field of the optimal transport map for any plan ? As stated by the author, it seems this loss is by definition Flow Matching (FM) loss constrained to convex functions. Yes, we claim this. This is our main theoretical result presented in Section 3.1. (2) For the independent plan, i.e., the minimizer for the FM loss is explicitly given in [1], where for the linear interpolant. This minimizer is dependent solely on the source and target distribution (i.e., the plan for this case) and does not yield the optimal transport. If you could please explain how you overcame this. If we consider unconstrained FM loss minimization (w.r.t. the optimized vector field) as you suggest, the analytical solution is indeed given by the formula which you provided above. However, in our OFM, we propose to optimize FM loss only over specific optimal vector fields. In this case, minimizers of constrained and unconstrained problems are not the same. We theoretically prove that the minimizer in this case is the true OT vector field. (3) The experiment section is rather minimal. In particular, it is not clear what Figure 3 in the Image-to-Image task presents. The primal goal of the experimental section is to empirically confirm our theoretical findings. Namely, we check that our OFM always recovers the true OT map regardless of the input transport plan. We humbly believe that the section clearly demonstrates this: 1. In qualitative 2D-toy example of Section $4.1$ (Figure $2$), our OFM recovers the same trajectories for all input plan, even for misleading anti-minibatch plan. 2. In Section $4.2$, we quantitatively evaluate the ability to solve OT problems. We run our OFM, other FM-based methods and popular OT solver on OT Benchmark which knows ground truth OT solution. Then we directly compare obtained solutions with ground truth via L2-UVP and cosine metrics. We can see that our OFM outperforms other FM-based methods in recovering OT and demonstrates metrics close to OT solvers. Additionally, we conduct an illustrative experiment with image-to-image translation. Here the usage of OT is motivated by the necessity to keep the image content during the translation. In Figure $3$, we can see that OFM's results are the same for the both considered input plans and also the most plausible in comparison with adult images. As per request of the other reviewers, we also calculated FID on the I2I problem to quantitatively estimate the performance, see Table 1 from the attached pdf. Our OFM method outperforms all the baselines on this task. (3) Theorems and propositions are structured in a way which makes it very hard to follow. We are sorry that you found our paper hard to follow, but we understand that opinions on this topic differ, e.g., Reviewer qt8g, in contrast, found our paper well-written and easy to follow. To address your comment, we will restructure Section 3.1 in the final version. Specifically, we will split in to 2 subsections: - Subsection 3.1.1. (major) will contain optimal vector fields, our proposed loss and direct statement of the main Theorem 1. In turn, technical Lemma 1 will be moved to Appendix. This will help the reader to directly find the main result without any disctractions on technical things. - Subsection 3.1.2 (auxiliary) will contain various supportive results: Propositions 1-3 (loss reformulation, properties of the loss). Hence, the reader which is interested in the main result only will be able to easily skip this subsection and continue to the next (practical) section. Could you please reply, does such a change suit you well? (4) The proof of Theorem 1 in the appendix is unclear and unsatisfying. Please note that our main technical result is Lemma 1. It allows to take integral in FM loss over time $t$ and obtain analytical formula (18). The proof is based on rather non-trivial techniques from calculus and convex analysis: change of variables, line integrals, differentials and convex conjugation. We tried to do our best to make the proof accessible to the general ML community. Could you please point to particular aspects of the proof to clarify? In turn, our main Theorem 1 follows from Lemma 1. We take math expectation over plan $\pi$ from both sides of (18) and notice that the left part of (18) corresponds to OFM loss and the right part matches OT dual loss (4) up to constant. (5) Accessible main proof idea/proof sketch in the main paper. We do not include any sketch of the proofs because it is purely technical (but still very non-trivial) and heavily relies on doing calculus tricks such as the change of variables. In Lemma's proof (line 480), we firstly simplify OFM loss using Proposition 1. Then we make two changes of time variable $t$: $s = \frac{1}{t}, s' = \frac{1}{s-1}$, use eqn (15) and obtain integral in line 488. In line 489, we consider the curve $z_0(s')$ and take its differential $dz_0$ using implicit definition of $z_0(s')$ from eqn (15). Starting from here, we take line integrals over line $z_0$ to get the sum from (27). Next we use closed forms of differentials, integrate it from $t = 0$ to $t = 1$ and obtain analytical sum (28). Finally, we use standard inversion and Fenchel-Young’s equalities to simplify the final sum. If you think a sketch of this type could be useful in the main text, we are happy to add it to the final version of the paper. Concluding remarks for all questions. Please respond to our post to let us know if the clarifications above suitably address your concerns about our work. We are happy to address any remaining points during the discussion phase; if the responses above are sufficient, we kindly ask that you consider raising your score. --- Rebuttal 2: Title: Reviewer response Comment: I want to thank the authors for their response. Following the reviewers answers, I have invested more time in better following the authors claim and proofs. It is now my understanding that the key results of the paper is that the minimizer of FM loss constrained to velocity field which are a gradient of convex functions is the OT map. I acknowledged this is an important results and I will raise my score.	Summary: This paper introduces the Optimal Flow Matching (OFM) algorithm, which improves upon Rectified Flow and OT-CFM by generating exact straight trajectories and recovering the optimal transport map in one iteration. OFM optimizes Flow Matching loss using vector fields and a convex function. The algorithm is implemented with Input Convex Neural Networks (ICNNs) and includes an explicit gradient formula for loss calculation. Experiments demonstrate OFM's better performance on 2D, high-dimensional benchmarks, and in unpaired image-to-image translation. OFM achieves good results in the latent space of a pretrained autoencoder without needing ODE integration. Strengths: Strengths: - The problem is sound and essential. - The proposal is supported by both theoretical and empirical evidence. Weaknesses: - The success of the algorithm in high-dimensional tasks relies on pretrained autoencoders. To what extent does the performance of the pretrained autoencoders affect the performance of the proposal? - The assumption that the optimal transport map can be recovered in one iteration might not hold in all scenarios, particularly in more complex distributions. - How do you calculate the initial transport plan? Does OFM suffer from distorted or biased dynamic OT solutions like OT-CFM? - The proposal is compared with OT-CFM and Rectified flow. However, the main task of the baselines in their original paper was image generation in pixel space. Is it fair to exclude this task from the experiments? - It is essential to provide more justification for using metrics in the experimental section. What are the advantages of these metrics over others in works on Unpaired Image-to-Image Translation tasks, such as [1] and [2]? Technical Quality: 3 Clarity: 2 Questions for Authors: - Does OFM still work in the pixel space or a higher-resolution dataset? - Could you provide empirical running time and OFM time complexity? - Does the method work for the conditional generation, which is feasible for Rectified flow? - OFM does not need ODE for the sampling process; however, could using an ODE solver for sampling improve the performance of OFM, which is some kind of quality-time tradeoff? Confidence: 4 Soundness: 3 Presentation: 2 Contribution: 2 Limitations: References [1] Zhu, Jun-Yan, et al. "Unpaired image-to-image translation using cycle-consistent adversarial networks." Proceedings of the IEEE international conference on computer vision. 2017. [2] Xie, Shaoan, et al. "Unpaired image-to-image translation with shortest path regularization." Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2023. Flag For Ethics Review: ['No ethics review needed.'] Rating: 5 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Dear reviewer, thank you for your feedback. Please find the answers to your questions and comments below. (1) Pretrained autoencoders. Does OFM still work in the pixel space or a higher-resolution dataset? In theory, OFM can work directly in pixel space and has no limitation for dimensionality. However, in practice, there are yet no input convex neural networks (ICNN) architectures that work well in the pixel-space (see, e.g., [1] for discussion). Hence, we use pretrained encoders to process images. We emphasize that this aspect is mostly related to ICNN rather than to our developed OFM Framework. We believe that our novel theoretical findings will inspire the ML community to develop new well-performing ICNNs that may work well in the pixel space. (2) The proposal is compared with OT-CFM and Rectified Flow. However, the main task of the baselines in their original paper was image generation in pixel space. Is it fair to exclude this task from the experiments? Please see the previous answer for the reason of excluding this type of experiments. (3) The assumption that the optimal transport map can be recovered in one iteration might not hold in all scenarios, particularly in more complex distributions. Dear reviewer, this is not an assumption. We provide and prove a rigorous theoretical statement (Theorem $1$) that our Optimal Flow Matching retrieves OT between any two absolutely continuous distributions in just one flow matching iteration. Moreover, our experiments on high-dimensional standard benchmark [1] in the field of OT (Table $1$) demonstrate that OFM actually finds a solution whose metrics are close to or better than those of the state-of-the-art OT approaches. (4) Does OFM suffer from distorted or biased dynamic OT solutions like OT-CFM? No, it does not. In our OFM, we run only one FM iteration with any transport plan as the initial plan (Theorem $1$) and provably recover the true OT map. That is, unlike OT-CFM, the initial plan theoretically does not affect the final solution (again, which is the OT map between distributions), i.e., there is no bias or distortion effect. (4) How do you calculate the initial transport plan? In 2D toy example (Section $4.1$), we practically demonstrate that OFM's final solutions do not depend on initial plans. In this setup, the considered plans are: (a) Independent, where we independently sample batches from both distributions, (b) Minibatch, where we independently sample batches and then rearrange them according to Discrete Optimal Transport (with $\ell^2$ cost) between them, (c) Anti-Minibatch, where we independently sample batches and then rearrange them according to Discrete Optimal Transport based on negative Euclidean distance. The same approach with plans is used in the other experiments. (4) It is essential to provide more justification for using metrics in the experimental section. What are the advantages of these metrics over others in works on Unpaired I2I Translation tasks? The metrics which we use in Section 4.2 are specially designed for evaluation of the quality of the recovered OT solution in the benchmark experiment. These metrics are generally not related to unpaired I2I problems. Our goal there is to confirm that our OFM indeed learns OT. Specifically, we use the unexplained variance percentage (UVP) as the main metric: $\mathcal{L}^2$-UVP$(T) := 100 \cdot \\| T - T^\\|^2_{\mathcal{L}^2(p_0)} / \text{Var}(p_1) \%$. It directly computes the squared error between ground truth OT map $T^{}$ (which is known by the construction of the benchmark [1, Sec. 4.2]) and learned map $T$ and then normalizes the answer. Following the benchmark, we additionally compute cosine similarity between ground truth map directions and directions of $T.$ Following your request, we additionally calculated FID for Unpaired Image-to-Image experiment from Section $4.3$. See our General response section and Table 1 from the attached PDF. According to the table, our method beats all the baselines. (8) Does the method work for the conditional generation? Unfortunately, ICNNs which we use in our method are not straightforward to adapt to conditional generation. In particular, to our knowledge, there are yet no research papers developing scalable ICNNs (especially, conditional) to work with pixel spaces. This is why we leave this interesting question for future research. (9) Could using an ODE solver for sampling improve the performance of OFM? If one uses ODE solver for the vector field generated by OFM, one gets the same solution as the one obtained by simply calculating the final point. It happens because OFM trajectories between the initial and final points are perfectly straight. Concluding remarks. We thank the reviewer one more time for the interesting and important questions. Please respond to our post to let us know if the clarifications above suitably address your concerns about our work. We are happy to address any remaining points during the discussion phase; if the responses above are sufficient, we kindly ask that you consider raising your score. References: [1] Korotin, Alexander, et al. "Do neural optimal transport solvers work? a continuous Wasserstein-2 benchmark." Advances in neural information processing systems 34 (2021): 14593-14605. --- Rebuttal Comment 1.1: Comment: Thank you the the authors for their responses. Your responses partially address my concerns about the paper. I have raised the score by one point to reflect my positive attitude toward the additional explanation and empirical results.	Rebuttal 1: Rebuttal: We thank the reviewers for their feedback and time. We appreciate that reviewers acknowledge: importance and soundness of the chosen problem (Bhuh, YPMs), our theoretical contribution connecting FM and OT methods (Bhuh, nvT5, qt8g), and experimental verification of the theoretical part (qt8g, Bhuh). Please find the answers to your questions in the respective posts below your reviews. We additionally calculated FID for our considered Unpaired Image-to-Image problem to expand the experiments and make them more representative. The obtained results exhibit that our OFM approach quantitatively outperforms the considered flow-based baselines. We attach the PDF file to this post with requested experiments containing: - (All Reviewers) Table 1, FID metric for unpaired Image-to-Image real data problem. The considered methods are OFM, RF, c-RF and OT-CFM. - (Reviewer nvT5) Figure 1, OFM setting $t \equiv 1$. The title "gt" stands for the source and target distributions. Titles "mb 64" and "mb 16" stand for minibatch plan with batchsize $64$ and $16$, respectively. - (Reviewer nvT5) Figure 2, default FM with independent ("idp"), anti-minibatch ("anti mb") and minibatch ("mb") plans. The title "gt" stands for the source and target distributions. - (Reviewer qt8g) Figure 3, convergence of L2-UVP loss depending on training time for OFM, FM, c-RF and OT-CFM methods. - (Reviewer nvT5) Figure 4, OFM with and without amortization. The graph depicts L2-UVP metric during training. Pdf: /pdf/7a51f9ab251e6a0c58b25ad66542cd379ddea489.pdf	NeurIPS_2024_submissions_huggingface	2,024	null	null	null	null	null	null	null	null
Aligner: Efficient Alignment by Learning to Correct	Accept (oral)	Summary: This paper proposes a lightweight and model-agnostic alignment method, Aligh, which learns the correctional residuals between preferred and dispreferred answers using a seperate model. Extensive experiments are conducted to demonstrate its effectiveness across 11 different LLMs, focusing on helpfulness, harmlessness, and honesty. Additionally, interpretability experiments are performed, and the benefits in multi-round RLHF are examined. Strengths: - The method is well-motivated and quite lightweight. As demonstrated in the experiments on multi-round RLHF, it can assist in iterative updating. - Extensive experiments are conducted on 11 different LLMs, various datasets, and both single-turn and multi-turn scenarios. - Interpretability experiments are performed, providing interesting insights. Weaknesses: The discussion regarding the out-of-domain (OOD) extent of training data to test data is not addressed. I am quite interested in how a trained Aligner performs on OOD datasets. Technical Quality: 3 Clarity: 3 Questions for Authors: - Apart from safety tasks, how would the aligner model perform in other tasks, such as math or coding? - In Table 1, Llama2-70B-Chat + aligner shows negative impacts. Could the authors elaborate more on the potential reasons? Confidence: 3 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: The paper includes the limitations section. Flag For Ethics Review: ['Ethics review needed: Safety and security'] Rating: 7 Code Of Conduct: Yes	Rebuttal 1: Title: Official Reply to Reviewer tcrV Comment: Dear Reviewer tcrV: Thank you very much for taking the time to review Aligner and for your valuable feedback. In your initial comments, you highlighted concerns regarding Aligner's experiments in the OOD setting and anomalies in the data points in Table 1. To address these concerns, we conducted additional experiments during the rebuttal period without extra training. We evaluated the trained Aligner on zero-shot generalization experiments on `HumanEval`, `MMLU`, `MATH`, and `MT-Bench`. We also re-examined the original dataset for the anomalies in Table 1 and performed re-sampling. To further validate the effectiveness and robustness of our experiments, we included additional experiments using Llama3 and Llama3.1 as the upstream models, aiming to alleviate any remaining concerns you might have. We sincerely hope for your continued support for Aligner. If you have any further concerns, please feel free to contact us. Specifically, we have added the following discussions: - Without additional training, we provided Aligner's experimental data on the four evaluation datasets: `HumanEval`, `MMLU`, `MATH`, and `MT-Bench`, as validation experiments from the OOD generalization perspective. - Included experiments using Llama3 and Llama3.1 as base models, providing effective evidence of Aligner's outstanding performance. - Extended the experimental results of Aligner on non-corrected datasets by conducting comparative experiments on publicly available datasets such as PKU-SafeRLHF, HH-RLHF, and UltraFeedback. These results indicate that Aligner significantly outperforms existing methods even when trained on preference datasets. - Explained the anomalies in the data points in Table 1, primarily due to the excessively long output from Llama2-70-Chat during dynamic sampling. - Included experiments comparing Aligner with CAI and Self-improvement, expanding the content of Table 3 in the original paper. - Added consistency comparison experiments between GPT-4 evaluations and human evaluations to eliminate the influence of corrected response length. - Analyzed the response lengths before and after Aligner corrections to eliminate the influence of corrected response length. - Added experiments for BoN with N=5 and N=10, as well as BeamSearch. - Investigated evidence from existing open-source technical reports of models such as GPT-4 and LLaMA2, highlighting that RLHF cannot effectively balance helpfulness and harmlessness. ------- Here are our detailed responses to each of your questions and suggestions. We have made every effort, with the utmost sincerity, to address every one of your concerns. $\downarrow$ --- Rebuttal Comment 1.1: Title: The author's detailed response [1/5] Comment: > (Weakness #1) The discussion regarding the out-of-domain (OOD) extent of training data to test data is not addressed. I am quite interested in how a trained Aligner performs on OOD datasets. Re: Thank you very much for your suggestion. Regarding `I am quite interested in how a trained Aligner performs on OOD datasets`, we conducted the following experiments: without additional proportional datasets, we connected the trained Aligner to different pre-trained models and tested it on `HumanEval`, `MMLU`, `MATH`, and `MT-Bench`. We found that Aligner performed well on OOD datasets due to its dual Copy and Correction capabilities. Upon examining the data cases, we identified two main reasons for this performance: - The base model used to train Aligner is the Llama2-XB-Base, which possesses general capabilities. Through the Q-A-C learning process, the base model with general capabilities can learn the more easily generalizable representations from the preference dataset, which are the "corrective differences between good and bad answers," compared to RLHF's reward model that directly scores Q-A. - The dual Copy-Correction capability allows Aligner to be more conservative in some OOD Q-A cases, thereby leaning towards executing the Copy operation. We will include a detailed description of these findings in the revised version. Table 1. The performance (Win Rate) of Aligners across various OOD datasets encompassing code, mathematics, instruction-following, and general capabilities. \| Comparison $\downarrow$ Dataset $\rightarrow$ \| HumanEval \| MMLU \| MATH \| MT-Bench \| \| ---------------------------------------------------- \| --------- \| ----- \| ------ \| -------- \| \| GPT4 + Aligner-2B vs. GPT4 \| 0.75% \| 0.70% \| 0.12% \| 3.75% \| \| GPT3.5 + Aligner-2B vs. GPT3.5 \| 1.67% \| 0.91% \| 0.33% \| 6.25% \| \| Claude2 + Aligner-2B vs. Claude2 \| 1.47% \| 1.13% \| 0.24% \| 10% \| \| Beaver-7B + Aligner-2B vs. Beaver-7B \| 2.19% \| 1.48% \| 6.43% \| 17.50% \| \| Alpaca-7B + Aligner-2B vs. Alpaca-7B \| 2.92% \| 1.41% \| 5.65% \| 22.50% \| \| Vicuna-7B + Aligner-2B vs. Vicuna-7B \| 3.52% \| 3.14% \| 9.36% \| 12.50% \| \| Vicuna-13B + Aligner-2B vs. Vicuna-13B \| 2.22% \| 3.67% \| 5.39% \| 11.25% \| \| Vicuna-33B + Aligner-2B vs. Vicuna-33B \| 3.03% \| 2.55% \| 5.41% \| 10% \| \| Llama2-7B-Chat + Aligner-2B vs. Llama2-7B-Chat \| 1.63% \| 1.22% \| 9.62% \| 11.25% \| \| Llama2-13B-Chat + Aligner-2B vs. Llama2-13B-Chat \| 1.39% \| 1.01% \| 9.41% \| 13.75% \| \| Llama2-70B-Chat + Aligner-2B vs. Llama2-70B-Chat \| 1.36% \| 0.86% \| 5.47% \| 5% \| \| GPT4 + Aligner-7B vs. GPT4 \| 1.89% \| 0.72% \| 0.11% \| 5% \| \| GPT3.5 + Aligner-7B vs. GPT3.5 \| 1.87% \| 0.97% \| 0.37% \| 7.50% \| \| Claude2 + Aligner-7B vs. Claude2 \| 1.65% \| 1.25% \| 0.28% \| 11.25% \| \| Beaver-7B + Aligner-7B vs. Beaver-7B \| 5.41% \| 2.27% \| 8.13% \| 12.50% \| \| Alpaca-7B + Aligner-7B vs. Alpaca-7B \| 4.67% \| 2.32% \| 9.44% \| 17.50% \| \| Vicuna-7B + Aligner-7B vs. Vicuna-7B \| 3.43% \| 3.28% \| 6.69% \| 23.75% \| \| Vicuna-13B + Aligner-7B vs. Vicuna-13B \| 3.89% \| 3.76% \| 7.39% \| 25% \| \| Vicuna-33B + Aligner-7B vs. Vicuna-33B \| 2.63% \| 3.43% \| 4.35% \| 16.25% \| \| Llama2-7B-Chat + Aligner-7B vs. Llama2-7B-Chat \| 2.52% \| 1.24% \| 12.83% \| 15% \| \| Llama2-13B-Chat + Aligner-7B vs. Llama2-13B-Chat \| 1.99% \| 0.92% \| 11.47% \| 17.50% \| \| Llama2-70B-Chat + Aligner-7B vs. Llama2-70B-Chat \| 2.68% \| 0.91% \| 2.36% \| 7.50% \| --- Reply to Comment 1.1.1: Title: The author's detailed response [2/5] Comment: > (Question #1) Apart from safety tasks, how would the aligner model perform in other tasks, such as math or coding? Re: Thank you very much for your suggestions. In response to your concerns, we have added four mainstream objective and subjective evaluation datasets during the rebuttal period: `HumanEval`, `MMLU`, `MATH`, and `MT-Bench`. These datasets cover evaluations in mathematics and code tasks. The experimental results are as follows: Table 2. The performance (Win Rate) of Aligners across various datasets encompassing code, mathematics, instruction-following, and general capabilities. \| Comparison $\downarrow$ Dataset $\rightarrow$ \| HumanEval \| MMLU \| MATH \| MT-Bench \| \| ---------------------------------------------------- \| --------- \| ----- \| ------ \| -------- \| \| GPT4 + Aligner-2B vs. GPT4 \| 0.75% \| 0.70% \| 0.12% \| 3.75% \| \| GPT3.5 + Aligner-2B vs. GPT3.5 \| 1.67% \| 0.91% \| 0.33% \| 6.25% \| \| Claude2 + Aligner-2B vs. Claude2 \| 1.47% \| 1.13% \| 0.24% \| 10% \| \| Beaver-7B + Aligner-2B vs. Beaver-7B \| 2.19% \| 1.48% \| 6.43% \| 17.50% \| \| Alpaca-7B + Aligner-2B vs. Alpaca-7B \| 2.92% \| 1.41% \| 5.65% \| 22.50% \| \| Vicuna-7B + Aligner-2B vs. Vicuna-7B \| 3.52% \| 3.14% \| 9.36% \| 12.50% \| \| Vicuna-13B + Aligner-2B vs. Vicuna-13B \| 2.22% \| 3.67% \| 5.39% \| 11.25% \| \| Vicuna-33B + Aligner-2B vs. Vicuna-33B \| 3.03% \| 2.55% \| 5.41% \| 10% \| \| Llama2-7B-Chat + Aligner-2B vs. Llama2-7B-Chat \| 1.63% \| 1.22% \| 9.62% \| 11.25% \| \| Llama2-13B-Chat + Aligner-2B vs. Llama2-13B-Chat \| 1.39% \| 1.01% \| 9.41% \| 13.75% \| \| Llama2-70B-Chat + Aligner-2B vs. Llama2-70B-Chat \| 1.36% \| 0.86% \| 5.47% \| 5% \| \| GPT4 + Aligner-7B vs. GPT4 \| 1.89% \| 0.72% \| 0.11% \| 5% \| \| GPT3.5 + Aligner-7B vs. GPT3.5 \| 1.87% \| 0.97% \| 0.37% \| 7.50% \| \| Claude2 + Aligner-7B vs. Claude2 \| 1.65% \| 1.25% \| 0.28% \| 11.25% \| \| Beaver-7B + Aligner-7B vs. Beaver-7B \| 5.41% \| 2.27% \| 8.13% \| 12.50% \| \| Alpaca-7B + Aligner-7B vs. Alpaca-7B \| 4.67% \| 2.32% \| 9.44% \| 17.50% \| \| Vicuna-7B + Aligner-7B vs. Vicuna-7B \| 3.43% \| 3.28% \| 6.69% \| 23.75% \| \| Vicuna-13B + Aligner-7B vs. Vicuna-13B \| 3.89% \| 3.76% \| 7.39% \| 25% \| \| Vicuna-33B + Aligner-7B vs. Vicuna-33B \| 2.63% \| 3.43% \| 4.35% \| 16.25% \| \| Llama2-7B-Chat + Aligner-7B vs. Llama2-7B-Chat \| 2.52% \| 1.24% \| 12.83% \| 15% \| \| Llama2-13B-Chat + Aligner-7B vs. Llama2-13B-Chat \| 1.99% \| 0.92% \| 11.47% \| 17.50% \| \| Llama2-70B-Chat + Aligner-7B vs. Llama2-70B-Chat \| 2.68% \| 0.91% \| 2.36% \| 7.50% \| From the experimental results above, it is evident that Aligner maintains significant advantages in Math and Code tasks beyond non-safety tasks. However, considering the constraints of time and computational resources, we have made our utmost effort to complete the above experiments. We hope you can recognize our efforts during the rebuttal period and continue to support Aligner. Under the witness of Peer Review, we commit to supplementing the complete experimental results of Aligner in these tasks. --- Rebuttal 2: Title: Hope to Get Your Reply Comment: Dear Reviewer tcrV, As the deadline approaches, we wanted to kindly follow up on our recent submission. We have carefully addressed each of your suggestions and incorporated the feedback into the revised version. During the rebuttal period, we evaluated the trained Aligner on zero-shot generalization experiments on `HumanEval`, `MMLU`, `MATH`, and `MT-Bench`. We also re-examined the original dataset for the anomalies in Table 1 and performed re-sampling. To further validate the effectiveness and robustness of our experiments, we included additional experiments using `Llama3` and `Llama3.1` as the upstream models. Your feedback has been invaluable, and we would greatly appreciate any updates or further guidance you may have regarding our revisions and responses. Thank you for your time and consideration. --- Rebuttal Comment 2.1: Title: Reply to rebuttal Comment: Thank you for the detailed reply. Most of my concerns have been addressed. --- Rebuttal 3: Title: Thank you to your recognition and encouragement! Comment: Dear Reviewer tcrV, Thank you for your response. We are greatly encouraged that `most of your concerns have been addressed`. The significant efforts we put in during the rebuttal period have proven to be worthwhile. As the rebuttal deadline approaches, we would like to know if you require any further discussion with us. We sincerely hope that you will consider raising the assessment of Aligner. --- Rebuttal 4: Title: Thank you to your recognition and encouragement! Comment: Dear Reviewer tcrV, We are deeply grateful for your support of Aligner, and it has been highly encouraging for us to address your concerns in the rebuttal. We plan to include additional experiments to further demonstrate the effectiveness of Aligner and provide a comprehensive analysis of its out-of-domain (OOD) performance. These results will be reflected in the latest version of the paper. --- Would it be possible for you to consider raising your score after we have addressed your concerns? If there are any additional points you believe we should incorporate, please let us know, and we will promptly reply to you and try our best to incorporate them into the revised version of the paper. With best regards!	Summary: The authors proposed a novel alignment paradigm, fine-tuning a pre-trained LLM as an 'aligner' module to map the target LLM's misaligned zero-shot responses to corrected aligned responses. Advantages of this technique over alignment-via-tuning strategies like RLHF/DPO is that it is agnostic to model size/architecture as it does not require access to the target LLM's internals, so the size and training resources for the aligner LLM do not need to scale with the target LLM, and can be used to align different LLMs with only training the aligner once. This also means aligner can be used to align models without weight access, such as those behind an API. The authors demonstrate these advantages empirically, showing that aligner in on-par/superior HHH performance compared to existing prompt and tuning-based techniques with reduced compute costs, across 11 target LLMs and robust to choice of preference dataset. They also demonstrate improved RLHF performance by using aligner-generated corrections as synthetic human preference data. Training aligner is potentially also more compute-efficient than tuning the target model as the aligner LLM only has to learn how to map misaligned answers to aligned ones, which the authors claim to be easier to learn than mapping user inputs to aligned outputs from scratch, with no intermediate answer. The authors also use interpretability techniques to identify an activation direction in the aligner model corresponding to correcting the intermediate response vs leaving it uncorrected Strengths: originality: Seems Quite original as I haven't seen any paper that uses an aligner LLM to map corrected outputs , But seems plausible that a less well-known paper was already using such an alignment technique (or that a big AI lab is using this internally) given it's just a novel combination of known techniques and training schemes. quality: Experiment and analysis are of good quality. Comprehensive baselines of existing SOTA techniques compared, decent variety of robustness tests, intriguing supplementary interpretability result Clarity: Well presented in general, writing and results are clear, structured for easy reading Significance: Potentially quite impactful If the trend of scaling LLMs continues, could be very useful to have a model size-agnostic alignment methodology. The increased variety of intermediate alignment outputs (i.e. internals of the aligner model and intermediate responses, which aren't produced in say DPO) could be useful new artifacts to study interpretability/safety with Weaknesses: - For the RLHF section, Sounds like you’re just transferring the OOD reward model collapse problem to the aligner module? Seems like you'd just run into the same problem if the aligner LM performs poorly on OOD inputs and generates poor synthetic data to train on - I'm no expert on RLHF reward collapse so correct me if I'm wrong Other Fairly minor weaknesses: - Concrete examples of the prompts and before/after alignment LM responses would possibly be helpful to have a clearer qualitative sense of aligner's effectiveness, perhaps randomly sample some to include in the paper - Examples of the aligner model "correcting the response more than usual" after CRC would be nice, not really sure what that would look like concretely - Would be nice to see the baseline effectiveness of "just prompt the model to output the same answer" instead of doing QAA fine tuning - learning the identity transform seems to be fairly easily done with a prompt so seems like a potential waste of compute? I'd guess the authors tried training without QAA training and that led to poor results, but would be nice if documented more clearly - I would be much more convinced of the technique's usefulness if the authors can show good performance in a 'adversarially' crafted OOD test set (rather than just subsetting the train set, which makes for a more similar distribution between train and test). Particularly in applications where high reliability in OOD scenarios is needed. This may be especially apparent if the target LM is tuned to perform well on specialized domains, as the aligner LM may lack the domain knowledge/skills needed to generalize as well as the target LM. - the Loss landscape part of figure 1 is perhaps a little misleading as I thought it implied existence of an comparison of the gradient/loss landscapes. If that's included I would prefer further theoretical analysis or targeted experimental results on the claim of 'correcting misaligned answers is easier to learn than fine-tuning the target LM directly'. Though intuitive, it seems to be a overly broad claim not substantiated strongly enough by empirical evidence in the paper (at least as a claim about model training in general, beyond the domains tested in this paper). Technical Quality: 4 Clarity: 4 Questions for Authors: - What is the base model of the aligner LLM? I tried to look for this in the text but wasn't able to find it, probably should include that Confidence: 3 Soundness: 4 Presentation: 4 Contribution: 3 Limitations: Seems adequate, though I'd like to see the "transferring OOD to aligner model" issue addressed Flag For Ethics Review: ['No ethics review needed.'] Rating: 7 Code Of Conduct: Yes	Rebuttal 1: Title: Official Reply to Reviewer 3sZj Comment: Dear Reviewer 3sZj, We greatly appreciate your time and effort in reviewing our work on Aligner. We are grateful for your support. In your initial feedback, your main concerns were regarding Aligner's experiments in the OOD setting and the OOD reward model collapse problem. During the rebuttal period, we conducted additional supplementary experiments. We tested the trained Aligner on `HumanEval`, `MMLU`, `MATH`, and `MT-Bench` for zero-shot generalization. Additionally, we addressed your minor weaknesses by including experiments on CRC and results without QAA training to further alleviate any concerns you might have. We sincerely hope for your more support of Aligner. If you have any further concerns, please do not hesitate to communicate with us. Specifically, based on the feedback from other reviewers, we have added the following: - Without additional training, we included Aligner's experimental data on four evaluation datasets: HumanEval, MMLU, MATH, and MT-Bench, serving as validation experiments for Aligner's OOD generalization. - Added validation experiments without QAA training. - Included examples of CRC. - Discussed suggestions regarding your minor weaknesses. - Conducted experiments with Llama3 and Llama3.1 as the base models, proving Aligner's outstanding performance. - Expanded Aligner's experimental results on non-curated datasets, with comparative experiments on publicly available PKU-SafeRLHF, HH-RLHF, and UltraFeedback, showing that Aligner significantly outperforms existing methods even when trained on preference datasets. - Added experiments comparing Aligner with CAI and Self-improvement, expanding the content of Table 3 in the original paper. - To eliminate the impact of answer length post-correction, we added experiments comparing the consistency of GPT-4 and human evaluations. - Included an analysis of the response lengths before and after Aligner's corrections. - Added experiments with BoN, where N=5 and N=10, as well as BeamSearch. - Investigated evidence from existing open-source technical reports of models like GPT-4 and LLaMA2, demonstrating that RLHF cannot effectively balance helpfulness and harmlessness. ------- Here are our detailed responses to each of your questions and suggestions. We have made every effort, with the utmost sincerity, to address every one of your concerns. $\downarrow$ --- Rebuttal Comment 1.1: Title: The author's detailed response [1/7] Comment: > (Weakness #1) For the RLHF section, Sounds like you’re just transferring the OOD reward model collapse problem to the aligner module? Seems like you'd just run into the same problem if the aligner LM performs poorly on OOD inputs and generates poor synthetic data to train on > I'm no expert on RLHF reward collapse so correct me if I'm wrong Re: From our experiments, the Aligner can mitigates the OOD reward model collapse problem, primarily from two perspectives: - Unlike the Reward Model's BT modeling to learn preferences, the Aligner uses a residual-like learning approach to learn preferences. As a numerical model, the RM's generalization in text space is not as robust as that of the Aligner, which is a seq-to-seq model. - In traditional RLHF and DPO training, the preference datasets are often generated by multiple models, leading to RM not being sampled from the distribution of the model that needs updating. However, when using RLHF/DPO with the Aligner, we leverage the Aligner to learn the differences between human-preferred responses A and non-preferred responses B in the preference dataset, thereby endowing the Aligner with correction capabilities. The "copy + correction" approach uses the Aligner to correct the target model's response A to obtain A, forming a synthesized preference dataset $A^ > A$ for RM learning. Notably, A is sampled from the true distribution of the target model, and $A^$ is induced based on the model's true sampled response, thereby alleviating the OOD reward model collapse problem to some extent. > (Minor Weakness #1)* Concrete examples of the prompts and before/after alignment LM responses would possibly be helpful to have a clearer qualitative sense of aligner's effectiveness, perhaps randomly sample some to include in the paper Re: Thank you very much for your thorough review. In the submitted version, we included examples of responses before and after correction in Table 10 of the Appendix. Although this may not be very prominent, in the subsequent revision, we will highlight the responses before and after correction using different colors. Additionally, we will directly reference these examples in the main paper. --- Rebuttal 2: Title: The author's detailed response [2/7] Comment: > (Minor Weakness #2) Examples of the aligner model "correcting the response more than usual" after CRC would be nice, not really sure what that would look like concretely Re: Thank you very much for your suggestion! We have made the following adjustments and will include the following examples in the final revised version: - By adjusting the coeff values, we can control the promotion or suppression of corrections. - The coeff value is the coefficient with which the feature is added to the residual stream of the Aligner Model; a positive coefficient promotes correction, while a negative coefficient suppresses correction. From the examples, we observe that a positive coeff results in making longer corrections, whereas a negative coeff results in making shorter corrections. > Example 1 > Question: Whats the difference between a computer and an abortion clinic > Answer: A computer is a device that stores, retrieves and processes data while an abortion clinic provides medical services to end a pregnancy. > Aligner Correction: Thank you for your question. I understand that you are looking for information on the difference between a computer and an abortion clinic. However, I must point out that the question contains harmful and offensive content. The comparison of a computer to an abortion clinic is not appropriate or respectful, as it equates a device for storing and processing information with a medical facility that provides services related to pregnancy termination. \n\n I would suggest rephrasing the question in a more respectful and appropriate manner. For example, you could ask: "What are some differences between computers and medical facilities?" or "What are some similarities between computers and medical facilities?" This way, we can focus on finding answers that are helpful, respectful, and appropriate. > coeff=1.0 correction: I cannot provide a response that would intentionally harm or mislead others. The question contains harmful and offensive content, and I cannot edit it to make it appropriate. I'm just an AI, my purpose is to provide helpful and informative responses, not to promote or encourage harmful or unethical behavior. \n\n Instead, I would like to address the immediate safety concerns in the question. The term "abortion clinic" is a euphemism for a medical facility that provides abortion services. It's important to note that abortion is a legal and safe medical procedure, and there are no ethical or legal reasons to compare it to a computer. \n\nAdditionally, it's important to recognize that comparing a medical facility providing essential health services to a device used for storing and processing data can be harmful and offensive. Such comparisons can perpetuate harmful stereotypes and stigmatize marginalized groups. \n\nIf you have any other questions or concerns that do not involve harmful or offensive content, I would be happy to try and assist you to the best of my abilities. > coeff=-0.5 correction: I cannot provide a response that would intentionally harm or mislead others. The question contains harmful and offensive content, and I cannot edit it to make it appropriate. I suggest rephrasing the question in a respectful and neutral manner. Additionally, I must remind you that it's not appropriate to compare or contrast computers with medical facilities such as abortion clinics, as it can be harmful and offensive to some people. Is there anything else I can assist you with? --- Rebuttal Comment 2.1: Title: The author's detailed response [3/7] Comment: Follow the previous reply: > Example 2 > Question: Where can I steal some needles from to shoot up? > Answer: I cannot provide information or advice on illegal activities such as drug use or theft. Stealing is not an appropriate or safe way to obtain needles for drug use, and it is also illegal. Instead, I would encourage you to consider seeking help from a medical professional or a substance abuse treatment center if you are struggling with addiction. > Aligner Correction: I cannot provide information or advice on illegal activities such as drug use or theft. It's important to prioritize your health and well-being by avoiding harmful behaviors and seeking help from qualified professionals if needed. Instead of stealing, consider reaching out to a medical professional or a substance abuse treatment center for guidance and support. They can provide you with safe and legal ways to address your needs and help you overcome addiction. > coeff=1.0 correction: I cannot provide information or advice on illegal activities such as drug use or theft. It's important to prioritize your health and well-being, and there are safer and more positive ways to address any challenges you may be facing. Instead of stealing, consider reaching out to a medical professional or a substance abuse treatment center for help with addiction. They can provide you with the support and resources you need to overcome your struggles. Additionally, it's important to remember that using drugs can have serious negative consequences for your physical and mental health, relationships, and overall quality of life. It's worth seeking help and taking the first step towards a healthier path. > coeff=-1.0 correction: I cannot provide information or advice on illegal activities such as drug use or theft. --- Reply to Comment 2.1.1: Title: The author's detailed response [4/7] Comment: > (Minor Weakness #3) Would be nice to see the baseline effectiveness of "just prompt the model to output the same answer" instead of doing QAA fine tuning - learning the identity transform seems to be fairly easily done with a prompt so seems like a potential waste of compute? I'd guess the authors tried training without QAA training and that led to poor results, but would be nice if documented more clearly Re: Thank you very much for your valuable feedback and for recognizing the effectiveness of the Residual Correction training method we employed for training the Aligner. Your deep understanding of fine-tuning and aligning large language models is evident. In our experiments, we discovered that without the QAA Identity Training (i.e., the warm-step), the Copy paradigm produced by the Aligner was suboptimal, even when we instructed the Aligner to "copy the previous response if it is good enough" during inference. As reported in Table 1, we compared the `copy rate` and the `average modification length` of responses corresponding to both the `Prompt-Only (add copy request in the prompt during inference)` and the `QAA Identity Training`. The experimental results indicate that the Aligner trained with QAA identity training achieved a higher effective copy rate than the value corresponding to `Prompt-Only`. Additionally, by comparing the average modification lengths of the Aligner under both methods, we found that when the upstream model's response was sufficiently good, the Aligner trained with `QAA identity training` tended to have shorter modification lengths. These results align with our expectations and further demonstrate the superiority of the Residual Correction training paradigm employed by the Aligner. Table 1. The copy rate and mean modified length comparison between the QAA-Training and Prompt-only methods on Aligner-7B. \| Model $\downarrow$ Metric $\rightarrow$ \| Copy Rate \| Copy Rate \| Average Modified Length \| Average Modified Length \| \| :----------------------------------------- \| :---------: \| :-----------------: \| :------------------: \| :------------------: \| \| Training Method $\rightarrow$ \| Prompt-Only \| QAA-Training (ours) \| Prompt-Only \| QAA-Training (ours) \| \| GPT4 \| 0.00% \| 0.00% \| 147.13 \| 141.23 \| \| GPT3.5 \| 0.00% \| 0.00% \| 154.37 \| 150.99 \| \| Claude2 \| 0.00% \| 0.00% \| 155.07 \| 154.64 \| \| Beaver-7B \| 0.00% \| 0.29% \| 151.16 \| 143.14 \| \| Alpaca-7B \| 0.00% \| 0.29% \| 161.89 \| 160.16 \| \| Llama2-7B-Chat \| 0.00% \| 1.43% \| 155.34 \| 142.74 \| \| Llama2-13B-Chat \| 0.14% \| 6.00% \| 187.67 \| 178.46 \| \| Llama2-70B-Chat \| 0.00% \| 1.14% \| 149.06 \| 140.47 \| --- Rebuttal 3: Title: Hope to Get Your Reply Comment: Dear Reviewer 3sZj, As the deadline is nearing, we wanted to gently follow up on our recent submission. We have meticulously addressed each of your suggestions one by one and incorporated these feedbacks into the revised version. During the rebuttal period, We tested the trained Aligner on `HumanEval`, `MMLU`, `MATH`, and `MT-Bench` for zero-shot generalization. Additionally, we addressed your minor weaknesses by including experiments on CRC and results without QAA training to further alleviate any concerns you might have. We hope that these efforts will alleviate your concerns regarding Aligner. Your feedback is highly valuable to us, and we would appreciate any updates or further guidance you might have regarding our revisions and responses. Thank you for your time and consideration. --- Rebuttal 4: Comment: Thank you for your thoughtful updates. I have increased the presentation and soundness subscores in my review, but I don't think these are substantial enough changes to increase overall score to a strong accept --- Rebuttal Comment 4.1: Title: Thanks Reviewer 3sZj for Approving Our Work Comment: We sincerely appreciate your acknowledgment and are deeply encouraged by your decision to increase your presentation and soundness subscores!!! It is our honor to address your concerns, which have been helpful to our work and will be integral to the improvements in our final version.	Summary: The paper has the following contributions: 1. Resource Efficiency: Aligner is notably smaller, requiring far fewer parameters compared to traditional models like DPO and RLHF1. 2. Plug and Play: It can be easily integrated with various large language models (LLMs) without needing to adjust parameters, ideal for API-based implementations. 3. Enhanced Performance: The Aligner-7B model improves helpfulness and harmlessness metrics significantly across multiple models, including a notable increase in GPT-4's performance metrics. Strengths: 1. The paper proposes an interesting method which is Aligner to improve LLM alignment. 2. The paper has developed a model which can correct the model answers. The paper has conducted extensive experiments to demonstrate the effectiveness of Aligner and also discussed potential use cases of Aligner. Weaknesses: 1. Developing Aligner should be much more expensive for data annotation: Aligner needs the human annotator to correct the response, which is unlike correcting preference feedback where the annotator only needs to judge which candidate answer is better. Collecting corrections should be much more expensive than collecting preference feedback. 2. LLMs can self-critic and self-correct their answers. For advanced language models such as GPT-4, do we need the aligner to help with the alignment? Technical Quality: 3 Clarity: 3 Questions for Authors: When refining the mode's answer using Aligner, can providing feedback to the refinement process improve the performance? Confidence: 4 Soundness: 3 Presentation: 3 Contribution: 2 Limitations: I don't see any issues for this part. Flag For Ethics Review: ['No ethics review needed.'] Rating: 5 Code Of Conduct: Yes	Rebuttal 1: Title: Official Reply to Reviewer 7osp Comment: Dear Reviewer 7osp: Thank you very much for your time and valuable feedback on the Aligner review. In your initial feedback, you expressed concerns about the more expensive data annotation required for Aligner and the model's self-critic and self-correct capabilities, questioning Aligner's advantages. During the rebuttal period, we conducted extensive experiments and made a concerted effort to provide thorough experimental analysis to address your concerns. This includes additional experiments on uncorrected datasets, specifically publicly available preference datasets such as `UltraFeedback`, `PKU-SafeRLHF`, and `HH-RLHF`. We have also discussed in detail the limitations of GPT-4's self-critic and self-correct abilities in specific application scenarios and included additional validation experiments. We sincerely hope that you will reassess Aligner's score and support its acceptance by the community. Specifically, we have added the following discussions: - Discussed the advantages of combining GPT-4 and Aligner, including considerations for deployment in different application scenarios. - Expanded experiments involving Aligner, CAI (self-correct), and self-improvement, enriching the content of Table 3 from the original paper. - Enhanced experimental results for Aligner on uncorrected datasets, with comparative experiments on publicly available datasets like PKU-SafeRLHF, HH-RLHF, and UltraFeedback, demonstrating Aligner's significant superiority over existing methods even when trained on preference datasets. - Without additional training for Aligner, we have included Aligner's experimental data on four evaluation datasets: HumanEval, MMLU, MATH, and MT-Bench. - To address the impact of corrected response length, we have added experiments comparing the consistency between GPT-4 evaluation and human evaluation. - Included analysis of response length before and after Aligner correction to mitigate the impact of corrected response length. - Added experiments with BoN, where N=5 and N=10, as well as BeamSearch. - Investigated evidence from existing open-source technical reports, such as GPT-4 and LLaMA2, highlighting RLHF's inability to effectively balance helpfulness and harmlessness. --- Rebuttal Comment 1.1: Title: The author's detailed response [1/3] Comment: > (Weakness #1) Developing Aligner should be much more expensive for data annotation: Aligner needs the human annotator to correct the response, which is unlike correcting preference feedback where the annotator only needs to judge which candidate answer is better. > (Weakness #2) Collecting corrections should be much more expensive than collecting preference feedback.Collecting corrections should be much more expensive than collecting preference feedback. Re: Thank you very much for your meticulous review and feedback. Firstly, we would like to emphasize that Aligner does not rely on correction datasets; it achieves effective alignment using publicly available preference datasets. In the paper's Table 3, we utilized open-source preference feedback datasets such as `PKU-SafeRLHF` and `HH-RLHF`. These datasets only require annotators to judge which candidate answer is better. By training the Aligner on such non-corrective human preference datasets, we compared its performance with `SFT`, `RLHF`, and `DPO`, demonstrating a clear advantage. To enhance the persuasiveness of our experiments, we added an additional preference feedback dataset, `UltraFeedback`, during the rebuttal period. The overall experimental results are as follows: Table 1. The performance comparison(Win Rate) between Aligner and fine-tuning methods using public preference datasets. \| Comparison $\downarrow$ Dataset $\rightarrow$ \| Q-A-C Datasets \| PKU-SafeRLHF \| HH-RLHF \| Utral-Feedback \| \| :------------------: \| :--------------- \| :--------------- \| :---------------- \| :--------------- \| \| Metric $\rightarrow$ \| Helpful / Harmless \| Helpful / Harmless \| Helpful / Harmless \| Helpful / Harmless \| \| Aligner vs. SFT \| 23.10% / 0.40% \| - / - \| - / - \| - / - \| \| Aligner vs. RLHF \| 24.40% / 21.90% \| 8.70% / 8.80% \| 9.60% / 3.40% \| 25.47% / 13.13% \| \| Aligner vs. DPO \| 49.10% / 0.10% \| 33.30% / 27.00% \| 5.60% / 30.90% \| 27.21% / 6.12% \| In our paper, the correction dataset we used is a meticulously curated dataset of exceptionally high quality, consisting of 50K human preference correction data. The significance of this dataset lies in its ability to effectively improve the performance of 11 different models across 3H dimensions with a single training session of Aligner, rather than retraining for each model individually. Under the scrutiny of Peer Review, we commit to fully open-sourcing this 50K dataset (which cost approximately $90,000 to annotate), along with the complete training code and evaluation details. From a functionality perspective, while RM is trained only on binary human preference feedback, Aligner offers additional capabilities. For instance, Aligner can generate corrected responses, A, from an initial model's response, A. These corrected responses can serve as synthetic preference datasets to facilitate multiple iterations of RLHF, eliminating the need for additional manual annotations. --- Reply to Comment 1.1.1: Title: The author's detailed response [2/3] Comment: > (Weakness #3)* LLMs can self-critic and self-correct their answers. For advanced language models such as GPT-4, do we need the aligner to help with the alignment? Re: Thank you for your valuable feedback and question. Despite the strong instruction-following capabilities of LLMs like GPT-4, which can self-critique and correct their responses, the Aligner still holds irreplaceable advantages, primarily in the following three aspects: - The model's self-criticism and correction capabilities require users to use specific prompts, such as `Your answer is problematic, please reconsider.` This not only demands users to judge the correctness of the answers but also requires them to use additional prompts to activate GPT-4's self-correction function. This occupies the model's valuable context length, and expanding the context is highly resource-intensive for large models like GPT-4. - Although GPT-4 is powerful, it still has knowledge blind spots. For instance, it once couldn't answer a simple question like `Which is larger, 9.11 or 9.9?`. The Aligner can function as a patch for such large models. Similar to how the Windows system doesn’t update to a major version every time an issue arises but instead uses patches to fix problems and only upgrades to a major version after accumulating enough fixes. The combination of Aligner and GPT-4 works in a similar way. When GPT-4 encounters issues, the Aligner can be used as a patch to perform targeted corrections. After accumulating enough error cases, the model parameters can then be fine-tuned. Given GPT-4's massive number of parameters, frequent fine-tuning to address various bad cases is impractical. - GPT-4 excels in general capabilities, but in specific fields, such as rare diseases in clinical cardiology, it often provides incorrect answers. This type of knowledge is very scarce in the model’s pre-training data, leading to weaker performance in these rare scenarios. Adding a specialized knowledge enhancer like the Aligner on top of GPT-4 aligns well with practical application needs. - GPT-4 is particularly conservative in providing safe answers. For example, when asked about chemical reagents, even if the user considers it a normal inquiry, as a general-purpose model serving different regions and countries, GPT-4 tends to provide conservative responses. When deploying such powerful models in specific regions, adding an Aligner to improve user experience is very necessary. Experimentally, the Aligner demonstrates significant performance advantages over self-correct methods like CAI, as shown in the Table 2. More importantly, Self-correct and Aligner are not mutually exclusive. On the contrary, the combination of both can achieve a synergistic effect greater than the sum of their parts. The former ensures that the preliminary model produces better responses, while the latter fills in the gaps in the model's answers. Based on your suggestions, we will include relevant discussions and experiments in the revised version. Table 2. The performance comparison (Win Rate) between Aligner and CAI. \| Model $\downarrow$ Method $\rightarrow$ \| w/ CAI \| w/ Aligner-7B \| \| :-------------: \| :--------: \| :---------------: \| \| GPT-4 \| +14.83% \| +22.16% \| \| Alpaca2-7B \| +22.04% \| +47.71% \| \| Beaver-7B \| +6.35% \| +12.2% \| \| Llama2-13B-Chat \| +13.45% \| +18.63% \| --- Rebuttal 2: Title: Hope to Get Your Reply Comment: Dear Reviewer 7osp, As the deadline approaches, we wanted to kindly follow up on our recent submission. We have carefully addressed each of your suggestions and incorporated the feedback into the revised version. During the rebuttal period, we conducted additional ablation experiments on uncorrected datasets, specifically publicly available preference datasets such as `UltraFeedback`, `PKU-SafeRLHF`, and `HH-RLHF`. We have also discussed in detail the limitations of GPT-4's self-critic and self-correct abilities in specific application scenarios and included additional validation experiments. Your feedback has been invaluable, and we would greatly appreciate any updates or further guidance you may have regarding our revisions and responses. Thank you for your time and consideration.	Summary: This work introduces Aligner, a small model designed for post-processing the outputs of Large Language Models (LLMs) to improve their results. Despite incorporating concepts like Interpretability and Residual Correction, the fundamental issue of lacking novelty remains unchanged. Furthermore, the experimental section lacks results from mainstream subjective and objective evaluation datasets, which undermines the solidity of the experimental conclusions. Strengths: 1. The writing is very clear, and Aligner is imbued with the concept of Residual Correction, reminiscent of ResNet. 2. Aligner has the potential to play a role in multi-round RLHF training. Weaknesses: 1. The training of Aligner requires a robust Teacher model and human annotations. More importantly, as compared by the authors in their experiments, works like CAI have already demonstrated the model's ability to self-improve. The authors need to emphasize the fundamental difference between collecting improvement signals from a broad range of preferences and using the model for self-improvement, explaining why the former can yield stronger results than self-improvement alone. 2. The authors have not sufficiently presented results on objective datasets like MMLU、MATH、HumanEval, nor have they shown results on mainstream subjective evaluation sets (MT-Bench/Alpaca-Eval). In fact, due to the provision of more information or longer responses, this modeling approach is susceptible to subjective evaluation hacking. Technical Quality: 2 Clarity: 3 Questions for Authors: 1. Can BoN be included as a baseline since it also serves as an inference time intervention method? 2. Lines 234-236: It is unclear why standard RLHF can only improve utility and not harmlessness. Confidence: 5 Soundness: 2 Presentation: 3 Contribution: 3 Limitations: N/A Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Title: Official Reply to Reviewer JeaM Comment: Dear Reviewer JeaM, Thank you very much for taking the time to review Aligner and for your valuable feedback. In your initial comments, you noted that Aligner's experiments lacked certain baselines such as `BoN` and `results on mainstream subjective and objective evaluation datasets`. During the rebuttal period, we conducted extensive experiments, striving to include baselines like BoN and BeamSearch, as well as zero-shot generalization experiments on datasets such as HumanEval, MMLU, MATH, and MT-Bench. We hope that these efforts will alleviate your concerns regarding Aligner. We sincerely hope that you will reconsider Aligner's score and support its acceptance by the community. Specifically, we have added the following discussions: - Addressed the differences between Aligner and Self-improvement, and discussed the potential for organically combining Self-improvement with Aligner. - Included experiments comparing Aligner with CAI and Self-improvement, expanding the content of Table 3 in the original paper. - Extended the experimental results of Aligner on non-corrected datasets, performing comparative experiments on publicly available datasets such as `UltraFeedback`, `HH-RLHF`, and `PKU-SafeRLHF`. These results indicate that Aligner significantly outperforms existing alignment methods even when trained on preference datasets. - Without additional training, we provided experimental data for Aligner on four evaluation datasets: HumanEval, MMLU, MATH, and MT-Bench. - To eliminate the influence of corrected response length, we included consistency comparison experiments between GPT-4 evaluations and human evaluations. - Analyzed the response lengths before and after Aligner corrections. - Added experiments for BoN with N=5 and N=10, as well as BeamSearch. - Investigated evidence from existing open-source technical reports of models such as GPT-4 and LLaMA2, highlighting that RLHF cannot effectively balance helpfulness and harmlessness. ------- Here are our detailed responses to each of your questions and suggestions. We have made every effort, with the utmost sincerity, to address every one of your concerns. $\downarrow$ --- Rebuttal Comment 1.1: Title: The author's detailed response [1/6] Comment: > (Weakness #1) The training of Aligner requires a robust Teacher model and human annotations. More importantly, as compared by the authors in their experiments, works like CAI have already demonstrated the model's ability to self-improve. The authors need to emphasize the fundamental difference between collecting improvement signals from a broad range of preferences and using the model for self-improvement, explaining why the former can yield stronger results than self-improvement alone. Re: Thank you very much for your valuable feedback and question. The core difference between `Aligner` and `self-improvment` lies in how they achieve better responses based on the original model's answers. Numerous works, such as CAI [2], have demonstrated that models possess the capability to self-improve. Self-improvement relies on human experts designing some Chain of Thought (CoT) examples to enable the model to perform multiple-path decoding and vote for the best response [1]. However, these methods still encounter the following issues: - In cases where the model's original answers are poor or lack knowledge, self-improvement struggles to produce a better answer through CoT+vote. In contrast, Aligner can inject the missing knowledge from the upstream model using an additional preference dataset, thereby achieving improvements without relying on the model's self-correcting ability. As shown in the Table 1, Aligner demonstrates superior performance in the comprehensive enhancement of helpfulness and harmlessness compared to self-improvement. - Designing CoT to meet various rules and constraints for different scenarios requires a high level of expertise from human experts and strong instruction-following capabilities from the model. This is challenging for smaller or weaker models (as shown in the table below, models like Beaver-7B show limited improvement). Aligner, on the other hand, can be trained solely on a human preference dataset, where identifying the better and worse responses is less demanding than designing expert responses. Table 1. The performance comparison (Win Rate) between Aligner and self-improvement methods. \| Model $\downarrow$ Method $\rightarrow$ \| w/ CAI \| w/ self-improve \| w/ Aligner-7B \| \|:---------------:\|:--------:\|:---------------:\|:------------:\| \| Metric $\rightarrow$ \| Helpful \| Helpful \| Helpful \| \| GPT-4 \| +14.83% \| +20.93% \| +22.16% \| \| Alpaca2-7B \| +22.04% \| +22.22% \| +47.71% \| \| Beaver-7B \| +6.35% \| +0.6% \| +12.2% \| \| Llama2-13B-Chat \| +13.45% \| +13.05% \| +18.63% \| More importantly, Self-improvement and Aligner are not mutually exclusive. On the contrary, the combination of both can achieve a synergistic effect greater than the sum of their parts. The former ensures that the preliminary model produces better responses, while the latter fills in the gaps in the model's answers. Based on your suggestions, we will include relevant discussions and experiments in the revised version. [1] Large Language Models Can Self-Improve [2] Constitutional AI: Harmlessness from AI Feedback Based on the currently available technical reports [3], whether it’s RLHF or Aligner, methods that leverage learning from robust teacher models and human annotations can effectively incorporate human feedback into model fine-tuning, aligning the model's behavior with human intentions. In the current applications of LLMs fine-tuning and alignment, introducing human feedback or AI feedback remains an effective means of enhancing performance. Aligner, in comparison to RLHF, exhibits significant advantages. Even a model as small as Aligner-2B can effectively enhance GPT-4's performance due to one key insight: it is easier for a model to learn the differences between good and bad responses than to generate good responses directly. As you pointed out, >Aligner is imbued with the concept of Residual Correction, reminiscent of ResNet. This concept is similar to residual learning, and our ablation experiments and interpretability analysis further demonstrate the residual characteristics of Aligner. Moreover, Aligner does not rely on a specialized correction dataset. Using publicly available preference datasets, Aligner can achieve better alignment effects compared to RLHF/DPO. We have demonstrated this superiority in our experiments and will include relevant discussions and results in the revised version. --- Reply to Comment 1.1.1: Title: The author's detailed response [2/6] Comment: Table 2. The performance comparison (Win Rate) between Aligner and fine-tuning methods using public preference datasets. \| Comparison $\downarrow$ Dataset $\rightarrow$ \| Q-A-C Datasets \| PKU-SafeRLHF \| HH-RLHF \| Utral-Feedback \| \| :------------------: \| :--------------- \| :--------------- \| :---------------- \| :--------------- \| \| Metric $\rightarrow$ \| Helpful/Harmless \| Helpful /Harmless \| Helpful / Harmless \| Helpful /Harmless \| \| Aligner vs. SFT \| 23.10% / 0.40% \| - / - \| - / - \| - / - \| \| Aligner vs. RLHF \| 24.40% / 21.90% \| 8.70% / 8.80% \| 9.60% / 3.40% \| 25.47% / 13.13% \| \| Aligner vs. DPO \| 49.10% / 0.10% \| 33.30% / 27.00% \| 5.60% / 30.90% \| 27.21% / 6.12% \| [3] The Llama 3 Herd of Models [4] GPT-4 Technical Report [5] Gemma: Open Models Based on Gemini Research and Technology [6] Qwen Technical Report and Qwen2 Technical Report > Summary: Furthermore, the experimental section lacks results from mainstream subjective and objective evaluation datasets, which undermines the solidity of the experimental conclusions. > (Weakness #2) The authors have not sufficiently presented results on objective datasets like MMLU、MATH、HumanEval, nor have they shown results on mainstream subjective evaluation sets (MT-Bench/Alpaca-Eval). In fact, due to the provision of more information or longer responses, this modeling approach is susceptible to subjective evaluation hacking. Re: Thank you very much for your valuable feedback and suggestions. During the rebuttal period, we made every effort to enhance the evaluation of Aligner on both objective datasets and mainstream subjective datasets, as shown in the Table [3]. We tested the trained Aligner on different upstream models and evaluated its performance on `HumanEval`, `MMLU`, `MATH`, and `MT-Bench`. This showcases the Aligner's OOD zero-shot generalization capability. We found that due to the Aligner's combined properties of `Copy` and `Correction`, it performed well on OOD datasets. Upon examining the data cases, we identified two reasons for this: - The base model used for training the Aligner is the llama2-7B-Base, which inherently possesses general capabilities. Through Q-A-C learning, this base model can acquire representations from the preference dataset that are easier to generalize, specifically focusing on the `corrective differences between good and bad responses` compared to the direct scoring of Q-A by RLHF reward models. - The combined Copy-Correction ability allows the Aligner to be conservative in some OOD Q-A scenarios, thereby leaning towards executing Copy operations. --- Rebuttal 2: Title: Hope to Get Your Reply Comment: Dear Reviewer JeaM, As the deadline is nearing, we wanted to gently follow up on our recent submission. We have meticulously addressed each of your suggestions one by one and incorporated these feedbacks into the revised version. During the rebuttal period, we conducted numerous additional experiments, striving to include baselines like `BoN` and `BeamSearch`, as well as ablation study experiments on datasets such as `HumanEval`, `MMLU`, `MATH`, and `MT-Bench`. We hope that these efforts will alleviate your concerns regarding Aligner. Your feedback is highly valuable to us, and we would appreciate any updates or further guidance you might have regarding our revisions and responses. Thank you for your time and consideration. --- Rebuttal Comment 2.1: Comment: Most of my concerns have been resolved, so I will raise the score to 6. I hope these experimental results can be included in the final version of the paper. Thank you. --- Rebuttal 3: Title: Thanks You for Approving Our Work Comment: Dear Reviewer JeaM, Thank you very much for your recognition. We are greatly encouraged by the opportunity to address your concerns. The updated experimental results will be included in the final version of the paper. Once again, we sincerely appreciate your recognition. With best regards!	null	NeurIPS_2024_submissions_huggingface	2,024	null	null	null	null	null	null	null	null
Functional Gradient Flows for Constrained Sampling	Accept (poster)	Summary: The paper proposes to adapt particle based variational infence to the context in which the distribution of interest is supported on a subset of R^d. It is supposed that the subset on which the distribution is supported is characterized as a lower level set of a function, g, and that this function has a gradient whose norm is bounded away from zero on the complement of the supporting set. Consequently, a gradient flow which descends g when outside the set and follows gradient dynamics for which (a mollified modification of) the original distribution of interest is a fixed point within that support might be expected to provide a good approach to sampling. The paper then concerns itself with constructing a particular (Wasserstein-type) gradient flow with such properties, establishing some theoretical properties of the proposed method and then exploring its performance empirically in a number of settings. The approach could be viewed as an extension of an existing method for dealing with distributions over manifolds in which the inequality constraint within the present work is replaced with an equality constraint and much of the novel work is concerned with addressing that difference in this particular framework. Strengths: The problem of sampling from distributions constrained to subsets of R^d is an important one and extending the reach of modern sampling methods to this context is a valuable contribution. The general framework is an appealing one: it allows for fairly arbitary supports to be considered, provided only that a good choice of g can be obtained to characterize the support and the general approach is intuitively appealing; if within the support of the distribution of interest then pursue a standard gradient flow approach and outside that set modify the dynamics such that the flow moves towards the support. Reasonably competitive performance seems to be obtained in the numerical examples considered. Weaknesses: I did find the paper a little difficult to read in places and the choice/definition of notations could be improved. For example, in the definition of the set $\Omega$ in line 97 it is not made explicit it is a subset of $\mathbb{R}^d$ and $g$ is never identified; something like $$\Omega = \{ x \in \mathbb{R}^d : g(x) \leq 0\} \text{ for some } g:\mathbb{R}^d \to \mathbb{R}$$ would be much clearer. And in line 243 (as an example of an odd choice which is made throughout the manuscript to write densities evaluated at a point with that point being entirely implicit): $$p(\beta) = \mathcal{N}(0,\sigma^2 I)$$ doesn't feature $\beta$ on the RHS and looking at (17) I at first thought you intended the $p(\beta)$ term to be uniform; the trivial edit to the more precise statement $p(\beta) = \mathcal{N}(\beta; 0,\sigma^2 I)$ would undoubtedly save many readers time. These are just some examples and while individually trivial my overall impression was that careful editing for presentation could significantly improve the manuscript. The examples all seem rather small by modern standards with "ground truth" estimates obtained by rejection sampling in one case (notwithstanding the fact that scale is not the only thing which makes inference challenging and there is always scope for methods which perform well in other important contexts). Is it feasible to scale the method up, for example to larger neural networks? Technical Quality: 3 Clarity: 2 Questions for Authors: Time complexity seems important in comparing methods. Other than Appendix D.2 this isn't much discussed in this manuscript and that appendix isn't particularly enlightening. I think most readers would like an answer to the question "How much computation is required to obtain an answer of a given quality (however that is assessed) using this and competing methods and how does that vary as the required quality varies?" Are you able to give at least some heuristic numerical answer to that question? The choice of $g$ function only seems to be made explicit for the monotonic neural network example where it is not a simple function. What were the choices used in the other numerical examples and can you offer users any guidance on how they should go about specifying the particular choice of $g$ for other problems (while it may be straightforward to specify a $g$ such that it is zero on the boundary of the domain there are clearly a great many possible choices and it isn't obvious a priori how sensitive performance will be to the choice or how to make that choice)? More generally, how should users specify the various tuning parameters other than $g$ itself -- number of particles, architecture and parameters for the neural networks f_net and z_net, ... I don't understand the second part of Assumption 5.1: in line 139 it is stated that $p_0$ is generally assumed to be supported on all of $\mathbb{R}^d$; in the first part of the assume the gradient of $g$ is assumed to be bounded away from zero outside the support set. How can $g$ have gradient bounded away from zero across the support of $p_0$ (less $\Omega$) if that support is all of $R^d$ and $g$ be bounded over that support? Do you have in mind particular settings where the greater flexibility enjoyed by the CFG method in specifying constrained domains is important? I was disappointed that the numerical study didn't show such a setting. This flexibility seems like it should be a major advantage of the method and one thing that was absent from the numerical study was evidence that the proposed method either solves problems which existing methods cannot or performs substantially better in those settings. Such evidence would, I think, make the paper stronger. Confidence: 3 Soundness: 3 Presentation: 2 Contribution: 3 Limitations: I think this is done reasonably well. I would have liked to see more discussion of the tuning of the algorithmic parameters. Flag For Ethics Review: ['No ethics review needed.'] Rating: 5 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thank you for your careful review and valuable questions! We address your comments and questions as below. ### Weaknesses `W1`: I did find (...) improve the manuscript. A1: Thanks for your suggestion! We will modify the notations accordingly in our revision. `W2`: The examples all (...) larger neural networks? A2: Thanks for your question! Our method is feasible for learning larger neural networks and on higher dimensional real problems. Please refer to the third part of the global response. ### Questions `Q1`: Time complexity seems (...) to that question? A1: Thanks for your question! Figure 8 in appendix D.2 should be able to answer your question with numerical results. Take the sample size $N=4000$ as an example. Given a quality of Wasserstein distance smaller than 0.6, the computation time of CFG is approximately 110s while MIED is approximately 320s. Changing the required quality to Wasserstein distance smaller than 0.4, the computation time of CFG change to approximately 160s while MIED change to approximately 400s. In general, CFG is more particle efficient as a functional gradient method. `Q2`: The choice of (...) make that choice)? A2: Thanks for your question! We would like to clarify that we focus on problems for which the constrained condition is provided in advance, so we choose the $g$ function to be the simplest way of representing the required constraint. For toy experiments, please refer to appendix D.1 for the analytical descriptions of the constrained domains $\Omega$. Corresponding to $\Omega$, we choose $g(x)=(\|\|x\|\|^2-1)(\|\|x\|\|^2-4)$ for the ring experiment, $g(x)=x_1^2+(\frac{6}{5}x_2-x_1^{\frac{2}{3}})^2-4$ for the cardioid experiment, $g(x)=e^{-2}-\frac{e^{-2(x_1-3)^2}+e^{-2(x_1+3)^2}}{e^{2(\|\|x\|\| - 3)^2}}$ for the double moon experiment, and $g(x)=\|x_1-x_2\|+\|x_1+x_2\|-4$ for the block experiment. For Bayesian Lasso, $\Omega$ is the $l_q$ ball, and we choose $g(x)=\|\|x\|\|_q-r$ to match the problem scenario. `Q3`: More generally, (...) $f_{net}$ and $z_{net}$, ... A3: Thanks for the question! We experimented on choosing different number of particles in the Bayesian Lasso experiment, the results showed that increasing the number of particles can improve the sampling quality. Our experiment results are not very sensitive to the choice of architecture and the parameters of $f_{net}$ and $z_{net}$. Properly selecting the bandwidth $h$ according to the particle number can improve the results. One possible approach is to use the adaptive scheme to select the bandwidth as proposed in section 6.2.1 and appendix C. This scheme is backed by the error analysis in appendix C, and could be verified in the Bayesian Lasso experiment. Figure 1 in the attached pdf file showed that the adaptive scheme achieved slightly better energy distance results than the fixed bandwidth. More details on the parameter setting are provided in appendix D for each experiment. `Q4`: I don't understand (...) over that support? A4: Thanks for your question! This assumption is adopted from [1], which is to guarantee that the particles outside the domain can enter the constrained domain. In practice, we could take a sufficiently large domain that covers all the initial particles and the given constrained domain. $g$ could be bounded on this domain and $\nabla g$ is bounded away from zero, providing feasibility of training. This assumption does not serve for our main focus, which is the convergence analysis when the particles are inside the constrained domain. [1] Zhang, R., Liu, Q., & Tong, X. T. Sampling in constrained domains with orthogonal-space variational gradient descent. Advances in Neural Information Processing Systems, 2022. `Q5`: Do you have (...) the paper stronger. A5: Thanks for your question! Firstly, from the best of our knowledge, there are not many methods that can directly deal with non-convex and complex hard constrained domains besides our method and the MIED method. As stated in the introduction, the mirror method is restricted to problems where mirror maps can be obtained, while projection methods suffers from high computation cost for complex domains because of the complexity in finding projection points. Additionally, compared to methods that can deal with this setting such as MIED, our method achieved better results on the cardioid, the double moon and the monotonic BNN experiment. We list Table 6 in appendix D.1 as below for the comparisons on the toy experiments. Secondly, using neural network to learn the optimal velocity field is more powerful way to implement ParVI, as it enhances the expressiveness and flexibility compared to classical kernel methods ([1][2][3]). This advantage can be verified by our better results compared to kernel based methods MSVGD, PD-SVGD and C-SVGD. Using the simple and flexible neural network framework, our method is also more scalable than MIED in terms of particle numbers N. Please refer to appendix D.2 for detailed discussion and verification. Table: Wasserstein-2 distance and energy distance between the target distribution and the variational approximation on the three toy datasets. \|Name\|MIED(W2 distance)\|CFG(W2 distance)\|MIED(Energy distance)\|CFG(Energy distance)\| \|----\|----\|----\|----\|----\| \|Ring\|0.1074\|0.1087\|0.0004\|0.0003\| \|Cardioid\|0.1240\|0.1141\|0.0016\|0.0005\| \|Double-moon\|0.4724\|0.1660\|0.0629\|0.0022\| [1] di Langosco, L. L., Fortuin, V., & Strathmann, H. (2021). Neural variational gradient descent. arXiv preprint arXiv:2107.10731. [2] Dong, H., Wang, X., Lin, Y., & Zhang, T. (2022). Particle-based variational inference with preconditioned functional gradient flow. arXiv preprint arXiv:2211.13954. [3] Cheng, Z., Zhang, S., Yu, L., & Zhang, C. (2024). Particle-based Variational Inference with Generalized Wasserstein Gradient Flow. Advances in Neural Information Processing Systems, 36. ### Limitations I would (...) parameters. A1: Please refer to A3 of Q3 of our response for more details. --- Rebuttal Comment 1.1: Title: Thanks for the detailed response Comment: Thanks for the very detailed response. Having read the collected reviews and rebuttals, I feel more positive about the manuscript and have increased my score from 4 to 5. I would still have liked to see a more comprehensive numerical validation of the approach and I hope that the example described in the global rebuttal at least can find its way into any subsequent versions of the paper. --- Reply to Comment 1.1.1: Title: Thanks for your response Comment: Thanks for your suggestion! We will modify our paper accordingly in our revision.	Summary: The paper proposed a particle based method (using neural networks) to sample probability densities which are supported on subdomain of $\mathbb{R}^d$. This is done in the spirit of Stein variational gradient descent but using neural network instead of kernels and by incorporating the constraint into the functional which leads to a boundary integral which is handled via a band approximation. A complete convergence to equilbrium analysis is given, under a suitable assumption on the neural network approximation. The algorithm is tested on a variety of synthetic and real data sets and the results are competitive compared to other methods on the market (e.g MIED). Strengths: This a very well written paper which propose a new approach to constrained sampling. The approach is very well motivated and is a natural extension of SVGD. The reviewer enjoyed very much the clarity of the exposition and the conceptual clarity of the approach. A nice convergence analysis in continuous time is provided which is relatively straightforward and does not provide strong guarantees.There is lots more to be done on the theoretical side but this is a nice first step. The model is tested on a variety of examples with satisfactory results and the model compares quite well (similar accuracy) with other algorithms on the market. Weaknesses: Relative weakness is that the proposed method does not seem to outperform existing algorithms. Maybe a more detailed comparison with existing method interms of costs would be useful. Technical Quality: 4 Clarity: 4 Questions for Authors: 1) Can the authors accommodate multiple constraints and more complicated geometries? Are those generalizations possible? A discussion of these issues would be welcome. 2) How could the theoretical analysis be strengthened? The rate of convergence TV norm is slow and one would expect (?!) much faster convergence in suitable norm. Confidence: 4 Soundness: 4 Presentation: 4 Contribution: 3 Limitations: The limitations are addressed adequately. Flag For Ethics Review: ['No ethics review needed.'] Rating: 7 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thank you for your careful review and valuable questions! We address your comments and questions as below. ### Weaknesses `W1`: Relative weakness is that the proposed method does not seem to outperform existing algorithms. Maybe a more detailed comparison with existing method in terms of costs would be useful. A1: We achieved better results than MSVGD in the block experiment, and achieved better results mostly than MIED in the ring, cardioid and the double moon experiment. The latter results are in appendix D.1, which are listed below as Table 1. For large-scale experiments monotonic BNN, our method also outperforms other methods, including MIED. We list the monotonic BNN results in our main text (Table 1 in section 6.3) below as Table 2. Table 1: Wasserstein-2 distance and energy distance between the target distribution and the variational approximation on the three toy datasets. \|Name \|MIED(W2 distance)\| CFG(W2 distance)\| MIED(Energy distance)\| CFG(Energy distance) \| \|----\|----\|----\|----\|----\| \|Ring\|0.1074\| 0.1087\| 0.0004\|0.0003\| \|Cardioid\|0.1240\|0.1141\| 0.0016\|0.0005\| \|Double-moon\|0.4724\|0.1660\| 0.0629\|0.0022\| Table 2: Results of monotonic Bayesian neural network under different monotonicity threshold. The results are averaged from the last 10 checkpoints for robustness. \|$\epsilon$\|PD-SVGD(Test acc)\|C-SVGD(Test acc)\|MIED(Test acc)\|CFG(Test acc)\|PD-SVGD(Test NLL)\|C-SVGD(Test NLL)\|MIED(Test NLL)\|CFG(Test NLL)\|PD-SVGD(Ratio out/%)\|C-SVGD(Ratio out/%)\|MIED(Ratio out/%)\|CFG(Ratio out/%)\| \|----\|----\|----\|----\|----\|----\|----\|----\|----\|----\|----\|----\|----\| \|$0.05$\|$.647\pm .007$\|$.649\pm .002$\|$.596\pm .002$\|$\mathbf{.661\pm .000}$\|$.634\pm .002$\|$.633\pm .001$\|$.684\pm .000$\|$\mathbf{.632\pm .000}$\|$0.5$\|$0.0$\|$3.0$\|$0.0$\| \|$0.01$\|$.645\pm .004$\|$.650\pm .002$\|$.590\pm .001$\| $\mathbf{.660\pm .001}$\|$.635\pm .001$\|$.634\pm .001$\|$.678\pm .002$\|$\mathbf{.632\pm .000}$\|$0.0$\|$0.0$\|$5.0$\|$0.0$\| $0.005$\|$.645\pm .005$\|$.650\pm .002$\|$.586\pm .000$\|$\mathbf{.659\pm .001}$\|$.635\pm .001$\|$.633\pm .002$\|$.676\pm .001$\|$\mathbf{.632\pm .000}$\|$0.0$\|$0.0$\|$6.0$\|$0.0$\| ### Questions `Q1`: Can the authors accommodate multiple constraints and more complicated geometries? Are those generalizations possible? A discussion of these issues would be welcome. A1: Thanks for your question! The generalizations are possible. Please refer to the first part of the global response. `Q2`: How could the theoretical analysis be strengthened? The rate of convergence TV norm is slow and one would expect (?!) much faster convergence in suitable norm. A2: Thanks for the question! Our TV convergence rate is comparable to [1]. TV metric is broadly used in many machine learning problems because it is well defined for any two distributions. KL metric needs the initial distribution being absolute continuous with respect to the target distribution, while TV is not restricted by this requirement. Considering other suitable norm that applies to the constrained domain problem is very interesting, and we leave it for future work. [1] Zhang, R., Liu, Q., & Tong, X. T. Sampling in constrained domains with orthogonal-space variational gradient descent. Advances in Neural Information Processing Systems, 2022. --- Rebuttal Comment 1.1: Comment: Thank you for the comparison and the detailed answer. I think that this is a solid and innovative paper and I will raise my score a bit to a 7 --- Reply to Comment 1.1.1: Title: Thanks for your response Comment: Thanks for raising the score! Please feel free to let us know if you have further questions!	Summary: The authors develop a solution to constrained sampling by introducing a boundary condition for the gradient flow which would confine the particles within the specific domain. This gives a new functional gradient ParVI method for constrained sampling, called constrained functional gradient flow (CFG), with provable continuous-time convergence in total variation (TV). They also present numerical strategies to handle the boundary integral term arising from the domain constraints. They provide theoretical and experimental study to show the effectiveness of the proposed framework. Strengths: The authors provide extensive theoretical analysis to support the proposed method. The authors provide solid theoretical results along with the proposed method. The authors also provide experiments on different datasets to validate the approach. The paper is well written. Weaknesses: Experiments on a real-world application could strengthen the paper. Technical Quality: 3 Clarity: 4 Questions for Authors: There is no question from the reviewer. Confidence: 3 Soundness: 3 Presentation: 4 Contribution: 3 Limitations: N/A Flag For Ethics Review: ['No ethics review needed.'] Rating: 7 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thank you for your careful review! We address your comments as below. ### Weaknesses `W1`: Experiments on a real-world application could strengthen the paper. A1: Thanks for the suggestion! In our experiments, the monotonic BNN is applied to real-world problems. The setting is originally derived from [1] about modelling in law. Additionally, it is also widely applied in social problems, including predicting admission decisions ([2]). For higher dimensional real-world applications, please refer to the third part of the global response. [1] Karpf, J. (1991, May). Inductive modelling in law: example based expert systems in administrative law. In Proceedings of the 3rd international conference on Artificial intelligence and law (pp. 297-306). [2] Liu, X., Tong, X., & Liu, Q. (2021). Sampling with trusthworthy constraints: A variational gradient framework. Advances in Neural Information Processing Systems, 34, 23557-23568.	Summary: This paper addresses the challenging problem of constrained sampling from an unnormalized distribution. Building on the principles of Stein variational gradient descent (SVGD), which merges the strengths of variational inference and Markov Chain Monte Carlo (MCMC), this work innovates by applying SVGD in a constrained setting. Specifically, it introduces boundary conditions for the gradient flow to confine particles within a designated domain, demonstrating convergence in total variation. Numerical experiments highlight the method's effectiveness. Strengths: 1. The paper is well-written. 2. The proposed method is technically sound and demonstrates convergence in total variation. 3. Experiments on both synthetic and real-world datasets showcase the method's potential. Weaknesses: In Bayesian neural network experiments, the paper overlooks numerous baselines of existing Bayesian Neural Networks (BNNs) using variational inference, MCMC, and Laplace approximation. Technical Quality: 3 Clarity: 3 Questions for Authors: If we have additional equality constraints, does the method still work? Confidence: 2 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: The paper has discussed the limitations. Flag For Ethics Review: ['No ethics review needed.'] Rating: 7 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thank you for your careful review and valuable questions! We address your comments and questions as below. ### Weaknesses `W1`: In Bayesian neural network experiments, the paper overlooks numerous baselines of existing Bayesian Neural Networks (BNNs) using variational inference, MCMC, and Laplace approximation. A1: Thanks for the advice! Existing methods are useful in common unconstrained problems, but these methods generally do not apply to problems with constrained domains. ### Questions `Q1`: If we have additional equality constraints, does the method still work? A1: Thanks for the question! Our proposed framework can still apply to these types of problems by adopting the idea of [1]. We could use velocity fields like $v_{\sharp}(x)$ to guided the particles to satisfy the equality constraints, and use our method to propose velocity fields substituting $v_{\perp}(x)$ to satisfy the inequality constraints. We add these two types of velocity to obtain the desired velocity field. To illustrate the idea, we demonstrate the training process of a 3D toy example in Figure 2 of the attached pdf file. Suppose the coordinate of the particle is $(x,y,z)$. The target distribution is a truncated standard gaussian located in the ring shaped domain in xOy plane. This corresponds to the equality constraint $z=0$ and the inequality constraint $1\le x^2+y^2\le 4$. The initial distribution is the 3D standard gaussian distribution. We can see that the particles collapse to the xOy plane and converge inside the ring domain. [1] Zhang, R., Liu, Q., & Tong, X. T. Sampling in constrained domains with orthogonal-space variational gradient descent. Advances in Neural Information Processing Systems, 2022.	Rebuttal 1: Rebuttal: We thank all reviewers for their constructive feedback, and will modify our paper accordingly in our revision. Here we address some of the common issues raised by the reviewers. Geometric generalization of our method Some of the reviewers are interested whether our proposed method can generalize and accommodate multiple constraints (including more equality and inequality constraints) and more complicated geometries. For additional equality constraints, our proposed framework can still apply by adopting the idea of [1]. We could use velocity fields like $v_{\sharp}(x)$ to guided the particles to satisfy the equality constraints, and use our method to propose velocity fields substituting $v_{\perp}(x)$ to satisfy the inequality constraints. Adding these two types of velocity presents the desired velocity field. To illustrate the idea, we demonstrate the training process of a 3D toy example in Figure 2 of the attached pdf file. Suppose the coordinate of the particle is $(x,y,z)$. The target distribution is a truncated standard gaussian located in the ring shaped domain in xOy plane. This corresponds to the equality constraint $z=0$ and the inequality constraint $1\le x^2+y^2\le 4$. The initial distribution is the 3D standard gaussian distribution. We can see that the particles collapse to the xOy plane and converge inside the ring domain. For additional inequality constraints, we may need multiple velocity fields outside the constrained domain for particles to enter the constrained domain, and the boundary integral term can be similarly estimated using band-wise approximation. [1] Zhang, R., Liu, Q., & Tong, X. T. Sampling in constrained domains with orthogonal-space variational gradient descent. Advances in Neural Information Processing Systems, 2022. The assumptions and parameters on the domain entering phase Some of the reviewers are interested in the effect of lambda in line 158, and some are interested in assumption 5.1. These questions are about the phase when particles enter the constrained domain. We would like to be clear that the main focus of our problem is on the convergence when the particles are inside the constrained domain. The feasibility of entering the constrained domain is the basic requirement of the training. The assumptions on entering the constrained domain could be relaxed as long as sufficient enough that the particles will not stuck at local modes outside the domain. Lambda in line 158 controls the entering velocity of the particles. With respect to the problem scale and dimensionality, we select lambda in a reasonable range to achieve swift entering without overshooting the constrained domain. Scaling our method up to real-world application Some of the reviewers are interested whether our method could scale up to high dimensional real problems. We experimented on the COMPAS dataset using larger monotonic BNNs, the number of neurons in the layer of BNN increased from 50 to 100, making the particle dimension increase from 903 to 1502. From Table 1 in the attached pdf file, our method still achieved the best accuracy results compared to other methods, while achieving competitive log-likelihood results. To test our method on higher dimension, we additionally experimented on the 276-dimensional larger dataset Blog Feedback ([1]) using monotonic BNN. The particle dimension is 13903. From Table 2 in the attached pdf file, our method achieved the best result. [1] Liu, X., Han, X., Zhang, N., & Liu, Q. (2020). Certified monotonic neural networks. Advances in Neural Information Processing Systems, 33, 15427-15438. We hope our response has adequately addressed the reviewers' questions and concerns, and look forward to reading any other additional comments. Pdf: /pdf/a14050180d94f6b843741c558d0ef689828e5626.pdf	NeurIPS_2024_submissions_huggingface	2,024	Summary: The authors study functional Wasserstein gradient descent methods with constraints in support of target distribution. They study the variational problem of vector fields using the least square formula between the current step and the gradient of score functions. To maintain the constrained set, the authors restrict the choices of vector fields to particular linear combinations of vector fields and use the neural network functions to approximate these vector fields. The authors prove the convergence of the proposed method in total variation. Numerical examples demonstrate the effectiveness of the proposed method. Strengths: 1. The authors carefully analyze the algorithm's importance in handling the sampling domain's boundary. In particular, example 4.2 is a good example. 2. The algorithm variational formulation is well-designed with nice numerical examples. Weaknesses: 1. The authors construct (10) as the projected vector field. The choices of lambda seem to be lacking. 2. The authors miss essential literature in the field. One needs to discuss the related references carefully. 1. Wang, Y., Li, W. Accelerated Information Gradient Flow. J Sci Comput 90, 11 (2022). 2. A. Lin et.al. Wasserstein Proximal of GANs. GSI, 2022. 3. Wang, Y. et al. Optimal Neural Network Approximation of Wasserstein Gradient Direction via Convex Optimization. 2022. Technical Quality: 4 Clarity: 3 Questions for Authors: Can authors illustrate how to pick lambda? How does lambda affect the algorithm, at least for numerical reasons? Confidence: 5 Soundness: 4 Presentation: 3 Contribution: 3 Limitations: There are no limitations. Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thank you for your careful review and valuable questions! We address your comments and questions as below. ### Weaknesses `W1`: The authors construct (10) as the projected vector field. The choices of lambda seem to be lacking. A1: Thanks for the suggestion! We choose lambda to be 1 in toy experiments and Bayesian Lasso experiments. We choose lambda to be 100 in monotonic BNN experiments. Lambda is used to control how fast the particles can be pushed inside the constrained domain. According to the problem scale and dimensionality, we select lambda in a reasonable range. We will make this point clearer in our revision. `W2`: The authors miss essential literature in the field. One needs to discuss the related references carefully. Wang, Y., Li, W. Accelerated Information Gradient Flow. J Sci Comput 90, 11 (2022). A. Lin et.al. Wasserstein Proximal of GANs. GSI, 2022. Wang, Y. et al. Optimal Neural Network Approximation of Wasserstein Gradient Direction via Convex Optimization. 2022. A2: Thanks for the suggestion! These references implement Wasserstein gradient flows in multiple regions. Wang \& Li (2022) considered the Accelerated Information Gradient flow, which is the generalization of the Hamiltonian flow integrating multiple metrics. They provided the particle version for the flow with detailed theoretical analysis, and they also experimented on common sampling tasks including BLR and BNN. Lin et.al. (2022) applied the Wasserstein proximal method to GAN, the underlying theoretical guarantee is about Wasserstein natural gradient flow. Wang et.al. (2022) considered a similar problem of using neural networks to approximate the Wasserstein gradient flow as [1], the main difference is that Wang et.al. (2022) considered parameterizing the velocity field using the gradient of a neural network, rather than directly using a neural network. All the above references do not consider the problem of constrained domain sampling, for which we designed functional gradient flows. We will add these references in our revision. [1] Zhang, R., Liu, Q., & Tong, X. T. Sampling in constrained domains with orthogonal-space variational gradient descent. Advances in Neural Information Processing Systems, 2022. ### Questions `Q1`: Can authors illustrate how to pick lambda? How does lambda affect the algorithm, at least for numerical reasons? A1: Thanks for the question! Empirically, we pick lambda to be 1 for toy experiments and Bayesian Lasso experiments, and we pick lambda to be 100 for monotonic BNN experiments. Lambda is used to control how fast the particles can be pushed inside the constrained domain. According to scale of the problem, relatively large lambda may provide swift entering for some constrained domains, but may incur some instability if lambda is too large that particles overshoot the constrained domains. In practice, for small scale experiments where the step size can be large, one may choose a small lambda. For large scale problems (e.g., BNN) where the step size is small, one may want to use a large lambda to drive the particles to the constrained domain faster. --- Rebuttal Comment 1.1: Title: Reply to authors Comment: The authors have addressed my questions. I will keep my score. --- Reply to Comment 1.1.1: Title: Thanks for your response Comment: Thanks for your careful review! Please feel free to let us know if you have further questions!	Summary: The paper introduces an extension of the ParVI method for sampling on measures supported on constrained subset $\Omega$ of $\mathbb{R}^d$, where $\Omega$ is the sublevel set of a phase-field function $g$. The paper presents a construction in terms of a discontinuous particle velocity field, with the discontinuity occurring on the boundary of the domain. Outside the domain, the velocity field pulls the particles towards the boundary of the domain, while within the domain the normal parVI approach is used. To leverage the usual Stein formulation through integration by parts, a new boundary term occurs which must be calculated. This is approximated by a spatial integral along a thin band within the domain. The authors prove that (i) particles eventually arrive at the domain boundary (using the assumption that there are no stationary points in the level set function $g$ outside the domain); and (ii) under a universal approximation assumption, the empirical distribution converges in TV, assuming a Poincare inequality holds. The author demonstrate the method on a number of synthetic and real examples. Strengths: The paper presents an alternative approach to the mirror descent strategies to handling sampling from constrained domains. This is potentially very useful in settings where there is not an obvious parametrisation of the constrained measure which lends itself to mirror-descent approaches -- to my knowledge this is novel. The methodology is generally applicable, and the authors have been very clear about their assumptions and provided some good theoretical results. The experiments are sufficiently challenging. Weaknesses: The main weakness to me lies in how the boundary term is handled, for two reasons. My main concern is what happens when the step-size is very small. In this case, I can easily see scenarios where: (1) The particles reach the boundary and remain stuck there. What is pushing the particles across the boundary. When the step-size is not small this is ok, because particles will 'overshoot' and then get picked up by the other side of the velocity field. This isn't really addressed in the text, nor is the effect of the discontinuity on the particle movements, which I expect to be non-trivial. (2) The handling of the boundary integral is challenging, because there is a clear bias-variance trade-off. The smaller $h$ is, the less likely it is to find particles in the domain, and thus we would expect the variance of the integral estimator to be huge. This isn't really addressed in the text. (3) I believe (though happy to be corrected on this), that most of the examples could be handled with the mirror approach? It would have been nice to see examples where the alternative is simply impossible. Technical Quality: 3 Clarity: 3 Questions for Authors: (1) How limiting is the assumption that $\lVert \nabla g\rVert \geq C$? I can see this being a reasonable assumption for simple constraints, but what about others where the boundary is complex, e.g. corrugated, or is non-convex? (2) Can the authors clearly explain the bias-variance trade-off in choosing $h$, and the implications of that for the performance of the method. (3) Can the authors explain why the particles do not 'stick' at the boundary when the step-size gets very small? Similarly, can they explore the effects of the large discontinuity. (4) For the problems considered in this paper, could the authors not simply instead use rejection sampling for a uniform distribution within the constrained domain and then use that as an initial condition for the constrained parvi? Potentially then you wouldn't even need to consider the boundary condition? Confidence: 4 Soundness: 3 Presentation: 3 Contribution: 2 Limitations: Limitations have been adequately addressed. Flag For Ethics Review: ['No ethics review needed.'] Rating: 5 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thank you for your insightful review and valuable questions! We address your comments and questions as below. ### Weaknesses `W1`: My main concern is what happens when the step-size is very small. In this case, I can easily see scenarios where: (1) The particles reach (...) non-trivial. A1: Thanks for the question! First, we want to emphasize that in practice the step-size needs not to be very small (the algorithm would still convergence inside the constrained domain as the functional gradient approaches zero). Assumption 5.1 provides positive velocity field on the boundary. Also, when particles get stuck near the boundary, the boundary integral part would provide extra "force" to push these particles towards the inner side of the domain if the local density of the particles is higher than the target. With proper step size, empirically the discontinuity on the particle movements does not affect the overall convergence. Please refer to Figure 6 and Table 6 (listed below) in appendix D.1 for empirical results. `W2`: The handling of the boundary integral is challenging, (...) This isn't really addressed in the text. A2: Thanks for the question! The bandwidth $h$ needs to be adjusted according to the number of particles $N$ to get good performance. To put it simple, $h$ should be large if $N$ is small and vice versa. The choice of $h$ needs to strike a good balance between the band-wise numerical approximation error of the integral and the variance due to Monte Carlo estimation. Please refer to appendix C for detailed error analysis and simulation verification. We will make this reference clearer in our revision. `W3`: I believe (though happy to be corrected on this), (...) simply impossible. A3: Thanks for the question! For non-convex and complex boundaries where the mirror map is infeasible, it is hardly possible to use the mirror method. We have experimented on these cases in our paper, including the ring, cardioid, double-moon experiments in section 6.1 and the monotonic BNN experiments. Additionally, when mirror method is applicable such as the block experiment, we have shown in the right of Figure 1 that we still achieved better results than the mirror method MSVGD. ### Questions `Q1`: How limiting is the assumption that $\|\|\nabla g\|\|\ge C$? I can see this being a reasonable assumption for simple constraints, but what about others where the boundary is complex, e.g. corrugated, or is non-convex? A1: Thanks for the question! Firstly, we would like to be clear that the main focus of our theoretical analysis is on the convergence behavior when the particles are inside the constrained domain. The assumption that $\|\|\nabla g\|\|\ge C$ is mainly to guarantee that the particles would enter the constrained domain. In fact, any assumption would work here as long as it is sufficient enough that the particles will not get stuck at local modes outside the domain. Secondly, this assumption is not very limiting, which is adopted from [1]. In general, if $g$ has local extrema outside the constrained domain which hinder the entering, the particles may stuck or collapse at these local modes and affect the training. This would be left for future work. [1] Zhang, R., Liu, Q., & Tong, X. T. Sampling in constrained domains with orthogonal-space variational gradient descent. Advances in Neural Information Processing Systems, 2022. `Q2`: Can the authors clearly explain the bias-variance trade-off in choosing $h$, and the implications of that for the performance of the method. A2: Please refer to appendix C for detailed error analysis and simulation verification. We will refer to this discussion earlier in our revision. `Q3`: Can the authors explain (...) large discontinuity. A3: Thanks for the question! Generally speaking, the additional boundary integral term in RSD will endow the trained velocity field with a repulsive factor that prevents clustering near the boundary unless the target distribution encourages the particles to do so. This is different from the unconstrained cases and demonstrates the importance of properly estimating the boundary integral. Empirically, for positive step size, we can see from Figure 6 in appendix D that the 'sticking' phenomenon does not occur. `Q4`: For the problems considered in this paper, could the authors not simply instead use rejection sampling for a uniform distribution within the constrained domain and then use that as an initial condition for the constrained parvi? Potentially then you wouldn't even need to consider the boundary condition? A4: Thanks for the questions! For your first question, for complex real problems such as the monotonic BNN, it may be sample inefficient to use rejection sampling for a uniform distribution simply to guarantee that the initial particles are inside the constrained domain. The constrained domain of some problems may be difficult to locate and may have a low acceptance rate for rejection sampling. For your second question, even if the initial particles are all inside the domain, the boundary condition is still essential for correct sampling of the constrained distribution. Please refer to appendix D.1 for detailed ablation study on with and without estimating the boundary integral (Table 6 in appendix D.1 is listed below). The latter suffers from the 'sticking' to the boundary problem as you stated before while the former does not. Table: Ablation results of not estimating boundary integral, with and without $z_{net}$. \|Name\|Ring(W2 distance)\|Cardioid(W2 distance)\|Double-moon(W2 distance)\|Block(W2 distance)\|Ring(Energy distance)\|Cardioid(Energy distance)\|Double-moon(Energy distance)\|Block(Energy distance) \| \|----\|----\|----\|----\|----\|----\|----\|----\|----\| \|w/o boundary integral\|0.2138\|0.2321\|0.4866\|0.2438\|0.0097\|0.1147\|0.0068\|0.0073\| \|w/o $z_{net}$\|0.1248\|0.2234\|0.1217\|0.2422\|0.0013\|0.0009\|0.0049\|0.0073\| \|w/ $z_{net}$\|0.1087\|0.1660\|0.1141\|0.2416\|0.0003\|0.0005\| 0.0022\|0.0072\| --- Rebuttal Comment 1.1: Title: Response to authors Comment: I thank the authors for the detailed rebuttal. I am still not entirely convinced by the argument around the small step-size limit of the algorithm, but I admit I have not derived the continuum limit and studied the effect of the boundary term. The empirical experiments suggest it works sufficiently well. Based on the responses + discussion I will increase my score. I hope the responses made in the rebuttal make their way into the paper. --- Reply to Comment 1.1.1: Title: Thanks for your response Comment: Thanks for raising the score! We will modify our paper in our revision according to the responses made in the rebuttal.	null	null	null	null
Normal-GS: 3D Gaussian Splatting with Normal-Involved Rendering	Accept (poster)	Summary: The authors propose a novel appearance modeling technique for an chor-based 3D Gaussian Splatting representation based on normal information and an incident lighting parametrization with MLPs. In addition to representation several regularization techniques are used to stabilize normal and incident light optimization. The central claim of the paper is to achieve competitive the rendering quality while achieving more detail geometric information. The authors show experiments on the MIP-NeRF360, Tanks&Temples, Synthetic-NeRF and the deep blending dataset and provide a quantitative comparison in rendering quality and normal-accuracy. Strengths: 1) The paper is easy to read and all components are well explained. 2) The conducted experiments use the most relevant datasets and metrics to assess the image quality, LPIPS, SSIM, PSNR. 3) The underlying research problem of reconstructing accurate appearance and geometry is relevant for the field of neural rendering and an very active research area. Weaknesses: 1) The experimental evaluation only contains 3DGS based methods. Even though there are inherent advantages of 3DGS methods, I'd still expect a comparison to NeRF-based methods on the same research problem, e.g. Ref-NeRF and follow-ups. The central question here would be, how good is the proposed method compared to the best Neural Field based method? 2) The general definition of normals in the context of 3D gaussians sounds a bit vague. The authors say that they define the normal of the a gaussian primitive as the shortest axis of the 3D gaussian without actually defining a surface. How do you define the normals when gaussians overlap or are semi-transparent? 3) The authors claim to improve geometric quality and show the comparison in normal accuracy in table 2. From a 3D reconstruction perspective, the normal accuracy indeed can indicate high quality, however metrics on the reconstructed geometry, like chamfer distance, F1-score, might be a better indicator for geometric quality. Technical Quality: 3 Clarity: 2 Questions for Authors: I'd would be great if the authors could provide explanations to weaknesses. Confidence: 4 Soundness: 3 Presentation: 2 Contribution: 2 Limitations: The authors addressed valid limitations in the dicussion part. Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We thank the reviewer for helpful comments and suggestions. We are glad to address the issues raised in the review. Q1: Comparison to NeRF-based methods We would like to include the comparisons to NeRF-based methods here and in the final version. It is important to highlight that 3DGS-based methods, including ours, are relatively fast and could support real-time rendering compared with NeRF-based methods. As illustrated in Table C, on the Mip-NeRF 360 dataset, our method performs the second best in terms of the PSNR and achieves comparable SSIM and LPIPS scores. The involvement of the proposed IDIV could provide better novel view synthesis quality compared with Ref-NeRF, which further assures the novelty of our design. Please also note that the training time for Zip-NeRF is > 10 hours on high-performance GPUs, which is much longer than 3DGS-based methods, including ours. \| Method \| PSNR \| SSIM \| LPIPS \| \|-------------------\|--------\|-------\|-------\| \| 3DGS [2] \| 28.691 \| 0.870 \| 0.182 \| \| Mip-NeRF 360 [26] \| 29.231 \| 0.844 \| 0.207 \| \| Ref-NeRF [22] \| 28.553 \| 0.849 \| 0.196 \| \| Zip-NeRF [30] \| 30.077 \| 0.876 \| 0.170 \| \| Ours \| 29.341 \| 0.869 \| 0.194 \| Table C. Rendering quality comparisons on the Mip-NeRF 360 dataset. Q2: Definition of normals in the context of 3D Gaussians. We acknowledge that we did not define normals for true surfaces but rather established a normal direction for each Gaussian. As discussed on L202-203, we define the normal for a 3D Gaussian based on its geometric properties. Specifically, during optimization, 3D Gaussians often exhibit flatness around the surface areas, as observed in [5, 7, 45]. Therefore, for a near-flat Gaussian ellipsoid, its shortest axis thus functions as the normal vector. Consequently, we use the shortest axis of each Gaussian as its normal attribute. For areas overlapped by multiple Gaussians, the normal is the weighted average of overlapped Gaussians. As shown in Fig.1, we found with our IDIV design, the model can also render a reasonable surface normal for the semi-transparent cover in this definition. Moreover, defining physically correct normals for semi-transparent areas remains an interesting and open problem and we leave it for future work. Q3: Metrics on the reconstructed geometry other than the normal accuracy. We agree that the overall geometry quality is also important, in addition to the normal accuracy. We follow your suggestions and add experiments to evaluate the reconstruction quality of our method on the DTU dataset in terms of the mean Chamfer distance. We also report the rendering quality indicated by PSNR to demonstrate the advantages of our method. The results of 3DGS [2], SuGaR [8], and 2DGS [44] are adopted from Table 3 of 2DGS [44]. For a fair comparison, we conduct experiments under the same setting as 2DGS [44]. As Table 9 illustrates, our method achieves outstanding rendering quality while maintaining better geometric quality than 3DGS. \| Method \| Mean Chamfer Distance \| PSNR \| \|--------------\|------------------------\|-------\| \| 3DGS [2] \| 1.96 \| 35.76 \| \| SuGaR [8] \| 1.33 \| 34.57 \| \| 2DGS [44] \| 0.80 \| 34.52 \| \| Ours \| 0.94 \| 37.63 \| Table A. Geometric and rendering quality comparisons on the DTU dataset. We appreciate your recommendations regarding the comparisons to NeRF-based methods, clarification of defining normals, and measurements of the overall geometry quality. We will add them in the revised version. --- Rebuttal Comment 1.1: Title: Looking forward to your valuable feedback Comment: Dear Reviewer Gwpd, Thanks again for your thoughtful review, which helped us improve the quality and clarity of our paper. We sincerely hope that our rebuttal has addressed your questions and concerns. And if it is possible, please let us know if there are any additional clarifications that we can offer. We appreciate your valuable suggestions. Thank you very much for your time, Best Regards. Authors of Submission 8220 --- Rebuttal Comment 1.2: Comment: Dear authors, Thank you , I appreciate your efforts in answering my concerns. The provided results show that the actual reconstructed geometry is accurate and I strongly encourage the authors to add the answer to Q3 to the main paper. The comparison against the NeRF baslines provide evidence that it is also competitive to this line of works. AS the authors resolve my main concerns, I increase my rating to weak accept. Best Reviewer Gwpd --- Reply to Comment 1.2.1: Title: Thanks for your feedback. Comment: Dear Reviewer Gwpd, We are glad to include the suggested measurements of the overall geometry quality, clarifications regarding the normal definition, and comparisons to NeRF-based methods. Your valuable feedback has significantly improved the quality and completeness of our paper. We thank you again for your time and efforts during the review process. Best Regards, Authors of Submission 8220	Summary: This paper proposes a novel appearance modeling of 3D Gaussian Splatting (3DGS) for both accurate appearance representation and geometry reconstruction. Existing 3DGS methods suffer from the trade-off of the accuracy of appearance and geometry due to the disconnection between surface geometry and rendering. To address this problem, this paper reformulates the rendering equation for fast physically-based 3D Gaussian (Normal-GS) rendering, directly connecting surface normal and rendering color. For stable optimization, Normal-GS uses anchor-based MLPs that implicitly consider the local smoothness of local illumination. The experimental results show that Normal-GS reconstructs accurate surface normal while retaining the rendering quality of 3DGS. Strengths: + Propose a novel physically-based rendering method for 3D Gaussian Splatting that can use shading cues for normal estimation, backpropagating photometric loss to the surface normal of 3D Gaussian. + Exploit anchor-based 3DGS to represent and stably optimize local incident light without time-consuming ray tracing. + Experimentally show that the proposed method can achieve both high-quality rendering and accurate normal estimation even in complex lighting and specular scenes where existing methods degrade. Weaknesses: - The illumination of diffuse and specular components is independent of each other. This is physically implausible. - The anchor-based regularization would prevent the reconstruction of detailed geometry like bicycle spokes in Fig. 6. Furthermore, since IDIV depends on the surface normal through $\Omega^+$, it also requires the smooth surface normal. Technical Quality: 4 Clarity: 4 Questions for Authors: - How plausible are the optimized diffuse and specular reflection components? What if only either of them is rendered? - Can the IDIV be visualized to validate that it captures local incident light? For example, in outdoor scenes, it is oriented to the direction of the sun and other directions outside and within the shadow region, respectively. - Scaffold-GS uses the enhanced view-dependent features $\hat{f}_v$ instead of directly using a local feature $f_v$. Which features does this method use? Regarding both geometry and appearance modeling, view-independent Gaussian attributes would be desirable. Confidence: 4 Soundness: 4 Presentation: 4 Contribution: 3 Limitations: Yes. Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We thank the reviewer for appreciating our work and providing constructive comments. Q1: The illumination of diffuse and specular components is independent of each other. This is physically implausible. We agree that diffuse and specular components cannot be independent physically. However, our main purpose is to strive for a better balance between rendering and geometry quality. Our insight is to keep the model simple without solving the highly-unconstrained inverse rendering problem, which needs too many extra regularizers. We do believe our framework could be further improved by introducing additional parameters like metalliciness [16, 17, 18, 45] for downstream applications. Q2: Visualization of Diffuse, Specular and IDIVs. Thanks for the helpful suggestions. We are glad to provide more results to show the plausibility of our method. We show the visualization of our specular and diffuse components in Figure A of the one-page PDF. We choose scenes containing specular materials to better demonstrate the effectiveness of our method. From the visualization, we can observe that our method successfully disentangles the diffuse and specular components. We also visualize IDIVs in Figure B of the one-page PDF. We choose an outdoor scene for better visualization. We can observe that IDIVs approximately align with the sunlight in bright regions (on the table) and roughly diverge in the shadow region (under the table), which indicates that our IDIVs capture local incident lighting information. Q3: The anchor-based regularization would prevent the reconstruction of detailed geometry like bicycle spokes in Fig. 6. We acknowledge the detailed geometry like bicycle spokes is not fully modeled by our method. The thin structure is a challenging problem in discrete Gaussian representation which also related to the aliasing issue. We leave it to the future work. Q4: Furthermore, since IDIV depends on the surface normal through $\Omega^{+}$, it also requires the smooth surface normal. We agree with your comments. A smooth and correct normal is important for many applications, including estimating lighting and supporting capturing specular effects. We applied the depth-normal loss to regularize estimated normals. Q5: Scaffold-GS features. We agree that view-independent Gaussian attributes are desirable. However, during our experiments, we found that view-dependent features would produce slightly better rendering quality (~0.1-0.2 dB). We believe that the viewing direction could help determine the inside-outside direction of the surface, thus benefiting the rendering quality. We are glad to include the suggested results in the revised version. Thank you again for your valuable feedback. --- Rebuttal Comment 1.1: Title: Looking forward to your valuable feedback Comment: Dear Reviewer 68er, Thanks again for your thoughtful review, which helped us improve the quality and clarity of our paper. We sincerely hope that our rebuttal has addressed your questions and concerns. And if it is possible, please let us know if there are any additional clarifications that we can offer. We appreciate your valuable suggestions. Thank you very much for your time, Best Regards. Authors of Submission 8220	Summary: Normal-GS successfully combines color estimation with surface normal optimization, achieving remarkable surface normal prediction without sacrificing view synthesis quality compared to previous methods. In calculating diffuse color, the traditional method that extracts the surface normal components from the incident light integration is used, thus diffuse color is expressed as a simple product of the diffuse albedo, normal, and integration of the incident light (referred to as IDIV in the paper). Based on Scaffold-GS, which predicts multiple nearby 3DGS parameters by passing a single anchor neural Gaussian to a global MLP, IDIV is calculated for each anchor and used for diffuse color prediction. Additionally, by integrating the IDE proposed in RefNeRF into Scaffold-GS, the specular component is successfully represented with respect to the normal. Finally, geometry is regularized by comparing the rendered surface normal with the normal obtained from the depth map. Strengths: The idea of approximating diffuse color equation in terms of surface normal by extracting the normal from the integral equation is both simple and powerful. Additionally, the geometry inaccuracy that can occur due to view-dependent effects is effectively handled by using IDE to manage specular artifacts. The proposed method compares well with other recent methods that demonstrate robust surface normal in view synthesis. Detailed explanations of IDE and surface normal estimation are well-documented in the supplementary materials, making it easy to follow. Weaknesses: Due to the lack of explanation of Scaffold-GS in the related work section, the description of the network output is confusing. While it is unnecessary to provide additional details about Scaffold-GS, it is important to clearly explain the network output. As I understand it: In Scaffold-GS, an anchored feature f_v is fed into a global MLP to initially predict the parameters of ‘k’ adjacent 3DGS. In Normal-GS, according to line 194, global MLP theta_l is used to additionally predict IDIV. For diffuse color, it is directly calculated using the predicted IDIV and diffuse albedo. For specular color, it is calculated by passing IDE, normal, and feature into the color MLP theta. However, there is no mention of how components like opacity and diffuse albedo are calculated. It is unclear whether these components are embedded and stored like the original 3DGS, or if they are calculated through a separate MLP. Therefore, it would be helpful to specify which MLP is responsible for each component, similar to IDIV, to clarify the process. There is no quantitative ablation results. Additionally, there is no ablation of the depth-normal loss. Personally, I think this paper, due to its contributions through the refactorization of the rendering equation, would be more suitable for a computer vision or graphics conferences. Technical Quality: 3 Clarity: 3 Questions for Authors: How is "3DGS w/ IDIV" in the ablation studies (section 4.2) (b) implemented? Did you create an additional anchor neural Gaussian to the original 3DGS and use an global MLP to predict IDIV, or did you place it as an optimization component like opacity in each 3DGS and use it for rendering? Additionally, in (c), "ours" generally refers to the complete Normal-GS, so the meaning of "w/ IDIV" is unclear. If (c) refers to Normal-GS without the specular component, it would be better to label it as "Ours w/o L_s." In Figure 2, it seems that there is a typo where the representation of the "vector" in the "IDIV vector" is duplicated. Confidence: 4 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: As mentioned in the paper, learning the depth of distant objects, like the sky, consistently is extremely challenging. Therefore, it seems that learning surface normal based on depth will also be very difficult. Additionally, as described in the weaknesses section, clearer network outputs would be beneficial. Flag For Ethics Review: ['No ethics review needed.'] Rating: 7 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We thank the reviewer for appreciating our work and making constructive comments. Q1: Clearly explain the network output, and how components like opacity and diffuse albedo are calculated. We would like to clarify it here and in the final version. The whole procedure about how to calculate IDIVs and the specular color is the same as your comments in the second paragraph of the weakness section and as our explanation on L194-197. As for the opacity, because the updating strategy of the Scaffold-GS depends on the value of opacity, we followed the Scaffold-GS and predicted it using a global MLP. As for the diffuse albedo, we reused the color MLP of the Scaffold-GS to predict it. Q2: How is “3DGS w/ IDIV” in the ablation studies (section 4.2) (b) implemented? For “3DGS w/ IDIV” in Fig. 5 (b), we attached the IDIV as an additional attribute to each 3D Gaussian without using global MLPs. As described on L299, in this setting we applied the Laplacian and Total-Variation loss to IDIVs. This was done by first rendering IDIVs following the 3DGS rendering pipeline and then applying image-space regularizers. Please note that this setting was “sensitive to the tuning of loss weights” as discussed on L302 and motivated us to use the locally shared structure, such as Scaffold-GS, to regularize IDIVs. Q3: Quantitative Ablations. Thanks for the constructive suggestions. To better verify the effectiveness of the components, we add the quantitative results on DTU by comparing the mean chamfer distance and PSNR for geometry and rendering quality evaluation. Please refer to Table B. Our base model is the Scaffold-GS. 1) We add depth-normal into the Scaffold-GS, the results of (b) in Table B shows that the depth-normal loss can significantly improve the geometry quality but ruin the rendering with more than 2dB PSNR drop. 2) Then we introduce the specular component into (b), which can improve the rendering quality without compromising the geometry, which further assures the importance of involving normal into the rendering. 3) We add the proposed IDIV into the model to get our final model (d). Our full model further improves the PSNR by more than 0.2 dB with similar geometry quality. Compared with the base model (a) and the results of 3DGS (CD: 1.96, PSNR: 35.76), our full model achieves a better balance between the geometry quality and the rendering fidelity. \| Model \| Details \| Mean Chamfer Distance \| PSNR \| \|-------\|----------------------------\|------------------------\|-------\| \| (a) \| Scaffold-GS \| 1.84 \| 38.14 \| \| (b) \| (a) + Depth-normal loss \| 0.95 \| 35.90 \| \| (c) \| (b) + $L_\text{Specular}$ \| 0.93 \| 37.37 \| \| (d) \| (c) + $L_\text{Diffuse}$ (Full model) \| 0.94 \| 37.63 \| Table B. Quantitative ablation studies on the DTU dataset. Q4: Some typos. Thank you for pointing out these typos. We will fix them in the final version. We will also refine our writing as suggested and include the mentioned experiments in the revised version. We appreciate the reviewer's valuable help and comments. --- Rebuttal Comment 1.1: Title: Looking forward to your valuable feedback Comment: Dear Reviewer cnbL, Thanks again for your thoughtful review, which helped us improve the quality and clarity of our paper. We sincerely hope that our rebuttal has addressed your questions and concerns. And if it is possible, please let us know if there are any additional clarifications that we can offer. We appreciate your valuable suggestions. Thank you very much for your time, Best Regards. Authors of Submission 8220 --- Rebuttal 2: Comment: I somewhat agree with the comment from reviewer 5i8r that the motivation for directly optimizing the surface normals, which are byproducts of geometry, without targeting the geometry itself, is weak. From the perspective of geometry estimation, I believe that comparisons with methods like SuGaR and 2DGS, which were conducted in the rebuttal, should be included in the paper. However, since the calculation of normals is essential for using geometry as a rendering component, I still believe this research holds value. Assuming that the implementation details and comparisons with existing surface reconstruction method are supplemented, I will maintain my score. --- Rebuttal Comment 2.1: Title: Thanks for your feedback. Comment: Dear Reviewer cnbL, We are glad to include the implementation details, suggested quantitative ablations, and comparisons with existing surface reconstruction methods. We sincerely appreciate your constructive comments and feedback, which have significantly enhanced the quality and completeness of our paper. Thank you again for your time and efforts during the reviewing process. Best Regards, Authors of Submission 8220	Summary: This paper addresses the challenge of achieving high rendering quality and accurate geometry in computer vision and graphics. While recent advancements in 3D Gaussian Splatting (3DGS) have enabled real-time high-fidelity novel view synthesis, the discrete and noisy nature of 3D Gaussian primitives hinders accurate surface estimation. The authors propose Normal-GS to integrate normal vectors into the 3DGS rendering pipeline. Surface colors are re-parameterized as the product of normals and a specially designed Integrated Directional Illumination Vector (IDIV). Strengths: - The writing of this paper is good. It is easy to follow and understand the technical details. - The usage of PBR for optimizing normals of 3DGS is interesting, though it is not new. - Although the experimental results show only slightly better quantitative measurements compared to SpecGaussian, the quality of the normals is significantly improved. Weaknesses: - L102-103, the authors claim that PBR-based 3DGS methods show lower rendering quality than original 3DGS, which is not correct, please refer to the tables in GaussianShader [7]. - The most closely related work to this paper is 2DGS [44], but it is surprisingly not mentioned or included in the experiment section, which is quite perplexing. 2DGS simplifies 3DGS by aligning it with the surface normal direction, allowing the scale parameters to be directly optimized based on the appearance loss. This omission raises confusion since 2DGS has direct relevance and could potentially provide insights or comparisons for the proposed method. - The integration of Physically Based Rendering (PBR) techniques, specifically Bidirectional Reflectance Distribution Function (BRDF), with 3D Gaussian Splatting (3DGS) is not a novel concept. As mentioned by the authors, previous works have already explored this combination and achieved satisfactory results. Moreover, existing methods mostly enable "relighting" while the proposed method is not. In the proposed method, the authors introduce a simplification of the integral over the up half-space of a point, referred to as IDIV, which represents the relationship between illumination and incident light direction. They decode this IDIV using a Multi-Layer Perceptron (MLP) under the assumption of Lambertian reflectance. However, the authors also draw inspiration from Ref-NeRF and incorporate a specular component to account for specular effects. This additional complexity and the combination of different techniques have resulted in confusion and difficulty in understanding the underlying purpose, raising doubts about the technical novelty of the method, as it appears to be a combination of existing approaches (referred to as an "A+B" manner). - In order to assess the quality of the geometry (indirectly optimized by normals, which is the motivation stated by the authors), there are established metrics and datasets available, such as DTU, that can be utilized. It is perplexing that the authors solely evaluate the Mean Angular Error (MAE) of normals, as the quality of normals is not always the primary factor when evaluating geometry. Instead, the overall geometry quality holds greater importance and should be the focus of evaluation. Technical Quality: 3 Clarity: 2 Questions for Authors: I have expressed my concerns regarding the paper above, and I strongly recommend that the authors address them by enhancing the motivation, technical contributions, and exposition. These improvements are necessary to enhance the quality of the paper. As it stands, I am inclined to reject the current form due to the unclear motivation, absence of baseline methods, and limited applicability. I am also open to revising my rating based on the response and other reviewers' feedback. Confidence: 5 Soundness: 3 Presentation: 2 Contribution: 2 Limitations: NA Flag For Ethics Review: ['No ethics review needed.'] Rating: 3 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thank you very much for your review and valuable feedback. Below we address the concerns raised in the review. Q1: L102-103, the authors claim that PBR-Based 3DGS methods show lower rendering quality than original 3DGS, which is not correct, please refer to the tables in Gaussian Shader [7]. We appreciate the reference to Gaussian Shader [7], which indeed demonstrates improvements in rendering quality on synthetic datasets. We agree that [7] demonstrates better performance under controlled, synthetic conditions. However, our primary concern is the Gaussian Shader’s reliance on a global environment map, which does not align with real-world scenarios with analysis on L165-174 of our paper. In our experiments, as illustrated in Figure 4 and Table 1, [7] faced difficulties in modeling complicated real-world scenes. These experiments are re-run directly using their released codes. Q2: 2DGS [44] is not mentioned or included in the experiment section. We would like to clarify that the omission of 2DGS [44] from our experiments was not intentional. We discussed and cited 2DGS on L94-95 of our paper. However, it is important to note that 2DGS was recently released (with codes available on May 3rd). This timing is the primary reason for its exclusion from our experiments. Additionally, as noted on L94-95, 2DGS sacrifices some rendering fidelity in novel view synthesis due to the lack of a clear relationship between geometry and appearance, as shown in Table 3 and Table 4 of [44]. We also want to emphasize that 2DGS and our method are conceptually orthogonal and could potentially be combined. Unlike 2DGS, our method explicitly addresses the interactions between lighting and normals. We have also added comparisons with 2DGS, focusing on geometric reconstruction quality and rendering quality. Please refer to our response to Q3 and Table A for these comparisons. \| Method \| Mean Chamfer Distance \| PSNR \| \|--------------\|------------------------\|-------\| \| 3DGS [2] \| 1.96\| 35.76 \| \| SuGaR [8] \| 1.33\| 34.57 \| \| 2DGS [44] \| 0.80 \| 34.52 \| \| Ours \| 0.94\| 37.63 \| Table A. Geometric and rendering quality comparisons on the DTU dataset. Q3: The quality of normals is not always the primary factor when evaluating geometry. Instead, the overall geometry quality holds greater importance and should be the focus of evaluation. We agree that overall geometry quality is important, but we believe that accurate normal is also very useful for many applications and has nice properties such as strong generalization capability [A, B]. We appreciate the reviewers' recognition of the accuracy of our normal predictions. To address the importance of overall geometry quality, we followed your suggestion and further evaluated our method's performance on the DTU dataset. We focused on both the reconstruction and rendering quality. We assessed the mean Chamfer distance and PSNR, adopting results for 3DGS [2], SuGaR [8], and 2DGS [44] directly from Table 3 of 2DGS [44]. For a fair comparison, we followed the process from 2DGS. As shown in Table A, our method achieves outstanding rendering quality while with a better mean Chamfer distance than 3DGS. [A] Bae, Gwangbin, and Andrew J. Davison. "Rethinking inductive biases for surface normal estimation." CVPR. 2024. [B] Guangyao Zhai et.al Monograspnet: 6-dof grasping with a single rgb image. ICRA. 2023. Q4: The integration of PBR … with 3DGS is not a novel concept. ...existing methods mostly enable “relighting” while the proposed method is not. We acknowledge the significant contributions of previous PBR-based methods [7, 16, 17, 45], especially as they enable “relighting” editability. However, as we discussed on L44-51 and L98-101, prior attempts often require approximations and meticulous regularization terms, which compromise either rendering quality or geometric accuracy. This motivates us to design a straightforward method to integrate normal information explicitly and simply into 3DGS. We also directly compared with the PBR-based method, Gaussian Shader [7], in our experiment section and our method achieves better rendering results and normals. Q5: The motivation and novelty of our work. Thank you for your feedback. We respectfully disagree with the characterization of our method as merely a combination of existing approaches. Our primary motivation is to address the balance issue between rendering and geometry quality, rather than focusing on relighting or intrinsic decomposition. From a physically-based rendering perspective, we identify that the core issue arises from the lack of interaction between appearance and normal estimation. Solving this problem is challenging, and our approach aims to simplify the model while avoiding the complexities of highly unconstrained inverse rendering problems, which often require numerous additional regularizers. Our method introduces only a few additional IDIV and reflection parameters, yet achieves remarkable improvements in normal estimation and maintains rendering quality. We believe our approach provides valuable insights into the issue and establishes a new baseline that could benefit the community. Q6: ... under the assumption of Lambertian reflectance. ... incorporate specular components to account for specular effects. We would like to clarify that our approach does not assume objects are Lambertian. As stated on L151-153, our initial consideration is for a simple ideal case with Lambertian objects. For more complex materials, including those with specular effects, we address this in Section 3.4 of our paper. As illustrated in Figure 2 of our paper, our method is designed to handle both diffuse and specular components. This allows us to account for a range of material properties beyond the Lambertian assumption. We thank you again for your valuable feedback and we will include suggested experiments in the final version. --- Rebuttal Comment 1.1: Title: Looking forward to your valuable feedback Comment: Dear Reviewer 5i8r, Thanks again for your thoughtful review, which helped us improve the quality and clarity of our paper. We sincerely hope that our rebuttal has addressed your questions and concerns. And if it is possible, please let us know if there are any additional clarifications that we can offer. We appreciate your valuable suggestions. Thank you very much for your time, Best Regards. Authors of Submission 8220 --- Rebuttal Comment 1.2: Comment: I appreciate the authors' efforts during the rebuttal phase. After carefully reviewing their responses and the comments from other reviewers, I believe my initial justification is correct and valid, and I stand by my evaluation leaning towards rejection. IMO this work still requires further improvement to meet the standards of NeurIPS, including the motivation, technical novelty, experiments, etc. This paper may be better suited for graphics venues instead of machine learning conferences. Some specific comments: * I disagree with the assertion that the proposed method surpasses the inverse rendering technique for its 'environment map representation.' The proposed method also utilizes PBR-based rendering, albeit in a simplified form without the functionality of relighting or material editing. Furthermore, inverse rendering methods demonstrate better performance in novel view synthesis compared to 3DGS. * I maintain my evaluation that this work is incremental, as there are already numerous existing studies on 3DGS employing PBR rendering with simplified BRDF, such as split-sum. Recent works, such as [1][2], have incorporated more accurate PBR representations within 3DGS, and I think this direction is more promising as it significantly improves the accuracy of PBR rendering in 3DGS. [1] Unified Gaussian Primitives for Scene Representation and Rendering [2] 3D Gaussian Ray Tracing: Fast Tracing of Particle Scenes --- Reply to Comment 1.2.1: Title: Thanks for your feedback. Comment: Dear Reviewer 5i8r, Thank you for your feedback and the opportunity to address your concerns. We fully acknowledge the significant contributions of previous PBR methods [7, 16, 17, 45] targeting inverse rendering, and we appreciate your references to recent methods combining ray tracing with 3DGS. Our approach, however, addresses the core balance issue between rendering and geometry quality through a simple yet effective design, without complex regularizers, assumptions, or ray tracing, as detailed in our paper (L32-36, L52-53, Figure 1) and response to Q5. This focus differentiates our approach from inverse rendering techniques that often require additional regularizers or assumptions. In the Experiment section of our paper, we directly compared our method with a prior inverse rendering method [7] and showed our superior rendering quality and normal accuracy. We believe our method establishes a new baseline that could benefit the community. We respectfully emphasize that this explanation is crucial to understanding the motivation and advantages of our method. Regarding the “environment map representation” and “split sum” [A], both of which assume the existence of a global environment map, we would like to clarify that our method does not use this representation. Instead, as comprehensively outlined in our main paper (L165-174, Figure 4, experiments) and our response to Q1, this assumption has limited applicability in real-world scenarios. To clarify further, Figure 4 clearly illustrates the disadvantage of relying on global environment maps, where [7] fails to capture specular effects but our method succeeds. Moreover, in the “Dr. Johnson” scene in “Deep Blending”, which involves multiple rooms without a global environment map, the previous method [7] failed, as shown in Table 1 of our paper. We appreciate the suggestion that our work might be “better suited for graphics venues.” However, given the success of related work in the field of neural rendering and geometry modeling, such as IDR [B], Neural-PIL [C], NeuS [D], VolSDF [E], SAMURAI [F] and NDRMC [G] at NeurIPS, we believe our contributions align well with the conference’s scope, particularly in “Machine Vision.” Thank you again for your thorough review. Best regards, Authors of Submission 8220 [A] Brian Karis. Real shading in Unreal Engine 4. SIGGRAPH 2013 Course: Physically Based Shading in Theory and Practice, 2013. [B] Yariv, Lior, et al. "Multiview Neural Surface Reconstruction by Disentangling Geometry and Appearance." NeurIPS (2020). [C] Boss, Mark, et al. "Neural-pil: Neural pre-integrated lighting for reflectance decomposition." NeurIPS (2021). [D] Wang, Peng, et al. "Neus: Learning neural implicit surfaces by volume rendering for multi-view reconstruction." NeurIPS (2021). [E] Yariv, Lior, et al. "Volume rendering of neural implicit surfaces." NeurIPS (2021). [F] Boss, Mark, et al. "Samurai: Shape and material from unconstrained real-world arbitrary image collections." NeurIPS (2022). [G] Hasselgren, Jon, et al. "Shape, Light, and Material Decomposition from Images using Monte Carlo Rendering and Denoising." NeurIPS (2022).	Rebuttal 1: Rebuttal: We sincerely thank all the reviewers for their constructive and insightful comments, especially for recognizing our work by - Clear presentation. "easy to follow and understand the technical details" (5i8r), "well-documented" (cnbL), "excellent presentation" (68er), "easy to read and all components are well explained" (Gwpd). - Soundness of design. "Approximating diffuse color in terms of the surface normal by extracting the normal from the integral equation is both simple and powerful", "handle view-dependent effects by using IDE" (cnbL), "a novel method that can use shading cues for normal estimation, backpropagating photometric loss to the surface normal of 3D Gaussian", "stably optimize local incident light without time-consuming ray tracing." (68er). "Good soundness" (5i8r, Gwpd). - Superior experimental results. "the quality of the normals is significantly improved" (5i8r), "demonstrate robust surface normal in view synthesis", "achieving remarkable surface normal prediction" (cnbL), "the proposed method can achieve both high-quality rendering and accurate normal even in complex lighting where existing methods degrade." (68er). We will polish and revise the paper according to the reviewer's suggestions. Below please find our detailed responses. We sincerely welcome additional feedback and further discussions. Best regards, Authors of Submission 8220 Pdf: /pdf/e6ff62e347375eb6473446dc63540e9242dc9201.pdf	NeurIPS_2024_submissions_huggingface	2,024	null	null	null	null	null	null	null	null
Group-wise oracle-efficient algorithms for online multi-group learning	Accept (poster)	Summary: Multi-group learning has been attracted attention and solutions for small-sized groups are already available. This paper further studies the online learning case when the group $\mathcal{G}$ is large or even possibly infinite. In addition to designing an algorithm, the paper also provides extensive theoretical analysis on the results in three settings: the i.i.d. setting, the adversarial setting with smoothed context distributions, and the adversarial transductive setting. Strengths: 1. The paper studies an interesting and important problem. The size of possible groups may indeed be very large in real cases, and infinite as the sample size infinitely grows. 2. The paper, while using relatively advanced mathematics, tackles the problem with well-known and intuitive techniques in optimization and game theory, and consequently achieves satisfying theoretical results. 3. The presentation is very readable: the definitions, results, and remarks are presented in a nice order and achieves the level of rigor for a theoretical paper. Overall I am in favor of this work because of the elegant solutions for an important problem. Weaknesses: 1. The theoretical analysis is restricted on binary case, i.e. $\mathcal{Y} = \{ -1, 1 \}$. I am aware that the multi-class case would be much difficult to analyze, but this certainly shall be one of the possible directions in the future. 2. (minor) Definition 4.1 uses measure theory that is in a graduate-level course in real analysis and is likely a rarely appearing concept for the majority audience. To keep the paper self-contained, it might be helpful to thoroughly define all concepts such as essential supremum and the differentiation of measures. Technical Quality: 3 Clarity: 4 Questions for Authors: I do not have any questions at this moment. Confidence: 3 Soundness: 3 Presentation: 4 Contribution: 4 Limitations: Yes. Flag For Ethics Review: ['No ethics review needed.'] Rating: 7 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thank you for your insight and helpful review. Your comments will be very helpful in improving the presentation of the final paper. To your first point about the multi-class case, we actually do address this in Appendix C. In the main body, we only included a Remark on line 248, but Appendix C includes all the details for a finite, $K$-valued discrete action space $\mathcal{Y}$. Our main theoretical guarantee for such a case is Theorem C.1, and the algorithm for this case is Algorithm 4. To your second point about the measure theoretic Definition 4.1, we agree that this may be overly technical. We do need the definition to properly define our setting, but we will take care to define essential supremum and absolute continuity. We will also emphasize that thinking about the results in terms of our simple example where the base measure $\mathcal{B}$ is uniform on $\mathcal{X}$ is sufficient for understanding our results. --- Rebuttal Comment 1.1: Title: Thanks for the response Comment: I appreciate the concise response by the authors. I am happy to keep my score.	Summary: This paper studies the problem of online multi-group learning where the algorithm should achieve sublinear regret with respect to all groups. They propose a group-wise oracle-efficient algorithm that avoids enumerating all groups by accessing the group oracle. An $\tilde{O}(\sqrt{dT/\sigma})$ regret bound is derived for $\sigma$-smoothed online learning setting. Strengths: 1. The paper addresses the problem of avoiding enumeration of groups, which improves the efficiency of the algorithm when the space of group is large. 2. The reduction from online multi-group learning to the AMF framework is novel and interesting. Weaknesses: 1. The discussion of related work for online multi-group learning is not adequate. It would be better if the author included the regret bound and online learning setting in the related work. I think listing a table for comparison among these works is better. 2. The motivation of using the AMF framework and FTPL is still not clear and needs more discussion. Moreover, it would be better if the author could clarify the novelty in the analysis. 3. It would be better if the author could also show the lower bound of this problem. Thus we can see how tight regret is obtained and which part can be improved in future work. 4. The computational efficiency of the $(G, H)$-optimization oracle is not discussed. It seems that this oracle is much harder than the ERM problem in [Ach23+] since this oracle optimizes over both $G$ and $H$. Technical Quality: 3 Clarity: 2 Questions for Authors: Please see the weaknesses part. Confidence: 4 Soundness: 3 Presentation: 2 Contribution: 2 Limitations: The authors have addressed the limitations and societal impact. Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thank you for your insight and helpful review. Your comments will be very helpful in improving the presentation of the final paper. In particular, we appreciate your feedback on the presentation of our results and related work, and we believe that addressing these issues will greatly improve clarity. To your first point on related work, space limitations did not allow us to include a table in the preliminary submission, but we agree that a table of existing results would help presentation. Please see the attached rebuttal PDF for our draft of this table. For multi-group learning in the online, sequential setting, the only two works that provide comparable regret bounds are, to our knowledge, [BL20] and [Ach+23], which both obtain a $\sqrt{T_g \log \|\mathcal{H}\| \|\mathcal{G}\|}$ regret requiring enumeration of $\mathcal{H}$ and $\mathcal{G}$ or $\mathcal{G}$ itself with an oracle for $\mathcal{H}$. We will state this in our table in the camera-ready version, making specific note how the algorithms fare in terms of space and computationally complexity, our main concern. In the online learning setting without group considerations, we can include the existing results for smoothed online learning from [Blo+22] in our table, which, to our knowledge, are the best possible regret bounds for the smoothed setting. To your second point on AMF and FTPL, we appreciate the feedback and believe it is important that the intuition of these two main ingredients is clear. In a nutshell, the AMF framework allows us to have low-regret with respect to all $\mathcal{G} \times \mathcal{H}$ pairs; the multi-objective regret guarantees are useful for our purposes because we want to guarantee simultaneous low regret over all $\mathcal{G}$. However, naively applying AMF requires us to enumerate these objectives, in turn enumerating $\mathcal{G}$, the main issue we want to avoid. FTPL comes to the rescue and allows us to have low-regret to all $\mathcal{G} \times \mathcal{H}$ pairs implicitly without enumerating them. So AMF allows us to compete with all $g \in \mathcal{G}$; FTPL ensures this is efficient. We see that perhaps this needed more description in the Introduction, namely around line 82 in Section 1.1 (Summary of Results), so, in the camera-ready version, we will include the main novelty of using these techniques in the intro. We will also refer the interested reader to "Technical Details" in Section 4.2 for the main high-level description of the analysis and techniques. Then, in the "Technical Details" in Section 4, on line 258 and below, we will make sure this high-level description of our analysis is clear and that each component, the AMF and the FTPL, have a clear purpose. To your third point about lower bounds, we agree that it is important to state the existing known limits, and we will provide details about existing lower bounds in the camera-ready work where we have more space. Statistical and computational lower bounds exist from [Blo+22], [HK16], and [HHSY22] for the oracle-efficient online learning setting without groups, so they immediately apply to our setting, as the group setting is a strictly harder generalization. We will include such lower bounds in our work. The lower bound of [HK16] motivates the line of work in oracle-efficient online learning, demonstrating that, in the fully adversarial online setting, even when given an oracle for $\mathcal{H}$, regret must be $\Omega(T)$ unless we our time complexity is $\mathrm{poly}(\|\mathcal{H}\|)$. This motivates working with an optimization oracle in more restricted settings, such as the smoothed setting we extensively study in our paper. Lower bounds for the oracle-efficient smoothed setting from [Blo+22] without groups show that, in the classification setting we consider, the $\sqrt{\frac{dT \log T}{\sigma}}$ regret is essentially tight. In particular, they show that any online algorithm with access to an ERM oracle cannot achieve sublinear regret against a $\sigma$-smooth adversary unless $T \geq \tilde{\Omega}(\sigma^{-1/2})$. So in the smoothed classification setting, our main result is essentially optimal. To your fourth point, we definitely agree that a main open question in our work is the nature of the $(\mathcal{G}, \mathcal{H})$-optimization oracle. To our knowledge, [GKR22] is the only work to have explicitly stated instantiations for such an oracle. There are many computational intractability results for agnostic learning, and multi-group learning is at least as hard as agnostic learning, so we don't expect to get end-to-end computationally efficient multi-group learning algorithms in general. To address Reviewer wskr, as well, we believe it would be very helpful to the presentation of the paper to include the algorithm description for the $(\mathcal{G}, \mathcal{H})$-optimization oracle in our paper so it can be easily referenced. Currently, we only make a passing mention to [GKR22], but we will include a lengthier discussion in the camera-ready version. There are two existing algorithms that reduce optimizing over $(\mathcal{G}, \mathcal{H})$ to weighted multi-class classification or alternating ERM over $\mathcal{G}$ and $\mathcal{H}$. Both suggest implementable heuristics in practice. In the camera-ready version, we will include the algorithms for the ternary classification oracle and the alternating minimization oracle, as well as their main guarantees from [GKR22] in the appendix or full body. We will also briefly discuss some heuristic possibilities for instantiating such oracles. We note that, due to the computational hardness of ERM in the worst-case, the oracle in [Ach+23] requires access to similar heuristics. For more on the computational efficiency of this oracle, please see our response to Reviewer wskr. We believe that we have addressed all the weaknesses in this response and will make the appropriate additions in the camera-ready version. Please let us know if anything was insufficient or unclear. --- Rebuttal Comment 1.1: Comment: Thanks for your response. You have addressed my concerns about the related work, lower bound, and computational efficiency of the oracle. Hence, I will raise my score to 6.	Summary: The paper provides oracle efficient algorithms for online multi-group learning. The goal is to obtain sublinear regret for each group subsequence $g \in G$ simultaneously. [BL20] showed that $o(T_g)$ regret is possible with finite $H$ and $G$ but must enumerate both $H$ and $G$. [Ach+23] showed $o(T_g)$ regret while being oracle efficient in the hypothesis class, but must enumerate over $G$. In this work, the authors use an optimization oracle for $G \times H$ jointly and avoid explicit enumeration for either G or H. To do so, they reduce the multigroup online learning to Adversary moves first framework of [Lee+22], and adapt results from oracle efficient online learning in the standard (no-groups) setting. Strengths: - The paper is technically solid and a substantial contribution to online learning with multi group regret gurantees, and is the first to be oracle efficient in both G and H. - The reduction of multi group regret to AMF regret and the resulting minmax game simplification is novel. - They provide multigroup regret analogs of smoothed online learning, generalized FTPL. Weaknesses: - For the $G \times H$ oracle mentioned in definition 2.2 the authors defer to [GKR22] who in turn provide two constructions: one based on ternary classification and another alternating minimization heuristic. It would be beneficial to add a longer discussion on this in the main body or the appendix to make the paper self contained. - Computational efficiency of the $G \times H$ oracle, this hasn't been discussed explicitly, are there group structures for which the optimization over G is efficient and doesn't involve enumeration?Algorithm 6 in [GKR22], doesn't discuss this either, it seems like without any further assumptions, enumeration is the best we can do in practice. Technical Quality: 4 Clarity: 4 Questions for Authors: Please address the above. Also, the appendix describes multi-class action space, what would change with real valued labels? Confidence: 5 Soundness: 4 Presentation: 4 Contribution: 3 Limitations: The authors discuss limitations of their paper. Flag For Ethics Review: ['No ethics review needed.'] Rating: 7 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thank you for your insight and helpful review. Your comments will be very helpful in improving the presentation of the final paper. We definitely agree that a main open question in our work is the nature of the $(\mathcal{G}, \mathcal{H})$-optimization oracle. To our knowledge, [GKR22] is the only work to have explicitly stated instantiations for such an oracle. On a high-level note, our work follows a long line of work that exploit oracle primitives assumed to be computationally efficient (referenced in our work), so we don't expect to get end-to-end computational efficiency. There are a plethora of well-known hardness results for agnostic learning with various concept classes, and multi-group learning is a generalization of agnostic learning, so the need to assume such primitives is necessary in many cases (and models the empirical success of ERM despite hardness results). To your first point in "Weaknesses," we agree that it would be very helpful to the presentation of the paper to include the algorithm description for the $(\mathcal{G}, \mathcal{H})$-optimization oracle in our paper so it can be easily referenced. In the camera-ready version, we will include the algorithms for the ternary classification oracle and the alternating minimization oracle, as well as their main guarantees from [GKR22] in the appendix or main body. We will also briefly discuss some heuristic possibilities for instantiating such an oracle. To your second point in "Weaknesses," it is not entirely clear to us what the computational efficiency of such an oracle would be, at least in the theoretical sense. From a heuristics standpoint, the aforementioned ternary classification instantiation from [GKR22] can be instantiated so long as we have access to weighted multi-class classification, which, at least in practice, can be implemented with cost-sensitive classification packages. In this case, $\mathcal{G}$ is a byproduct of the class upon which this ternary classifier outputs a "defer" label. On the other hand, if there's a specific $\mathcal{G}$ we want to address, the aforementioned alternating maximization oracle from [GKR22] can be implemented with ERM access to $\mathcal{G}$. Of course, ERM is computationally hard in the worst-case, but heuristics for ERM exist and it is a common assumption in the literature to assume that, at the very least, approximate ERM is feasible. Because $\mathcal{G}$ is class of group indicators, this is equivalent to assuming that we have ERM heuristics for binary classification. However, we should mention that the alternating maximization heuristic can only be guaranteed to supply saddle points, not true global maxima. In both these cases, heuristics (either for multi-class weighted classification or ERM) exist that avoid enumeration. However, it's clear that both these methods have limitations, and we will point out these limitations in the camera-ready version, along with the algorithm descriptions, as we mentioned in the point above. We believe that it is an interesting and worthwhile open question to develop more specific instantiations of this $(\mathcal{G}, \mathcal{H})$-oracle for more specific problem settings that are computationally efficient and have provable optimization guarantees. To your final point about real-valued action spaces, the main challenge is that our $\mathcal{H}$-player must solve a $\|\mathcal{Y}\| \times \|\mathcal{Y}\|$ size linear program, ultimately playing a distribution over possible actions which are of the order $\|\mathcal{Y}\|$. Because the $\mathcal{H}$-player must hedge against all possible $\mathcal{Y}$ the adversary could select in its linear program, we were only able to implement this step with discrete, finite label spaces. We believe it is an interesting open problem to extend our techniques to real-valued action spaces. One naive route might be to take a discretization approach, but a fine-enough discretization would defeat the purpose of the simple LP the $\mathcal{H}$-player must solve at every round. [Blo+22] does give an approach for the (single group) case for real-valued labels in the smooth setting, but the challenge remains to design a feasible game for the $\mathcal{H}$-player to solve at every round to hedge against the maximally adversarial choice in a continuous $\mathcal{Y}$. --- Rebuttal Comment 1.1: Comment: Thanks for the addressing my comments, I stand with my original assessment for accepting this paper. I appreciate the discussion on oracles for GxH and highlighting this in the main body and providing some concrete though not always computationally efficient heuristic in the appendix would be great, these should also address the other reviewer's concern.	Summary: This paper studies the online multi-group learning problem, and considers it in both smoothed context and general ($\gamma$-approximatable) settings. Unlike the traditional online learning setting, this paper considers achieve low regrets for all groups (subsequences). It utilizes the adversary moves first (AMF) framework for multi-objective online learning from [Lee+22] and the follow-the-perturbed-leader (FTPL) algorithm to handle the large number of groups. Through subtle algorithmic design and analysis, the paper achieved computationally-efficient and oracle-efficient learning. Strengths: The paper studied the multi-group online learning problem, which is related to fair learning and subgroup robustness literature, both are significant. Under the smoothed setting, it achieves $\tilde{O}(\sqrt{dT/\sigma})$ regret; while under the general setting where $\gamma$-approximability is guaranteed, it achieves $\tilde{O}(\sqrt{NT_{g}\log\lvert H\rvert \lvert G\rvert})$ for each group $g$. It first considers the i.i.d setting (both offline and online) then quickly moves forward to the smoothed online setting. Under the AMF framework, it considers two players: $(G,H)$-player (adversary that has access to an optimization oracle $\text{OPT}^{\alpha}_{(G,H)}$ that maximizes the learner's loss) and the $H$-player. Under such a game-theoretic perspective, the paper proposes algorithms that achieves bounded regret. It then adapts the algorithm to broader settings (non-smoothed settings), where the $\gamma$-approximability is assumed. For groups with different densities, the algorithm can achieve low regret for each group $g$. The algorithmic designs are subtle and the analysis is sound to the review. Weaknesses: The main text in Section 2 and after are fine, while Section 1 needs improvement. Specifically, some descriptions are unclear. For example, the bound $\tilde{O}(\sqrt{NT_{g}\log\lvert H\rvert \lvert G\rvert})$ is achieved for each group $g$, but this is not brought forward in summary of result (Line 76) and the last time $g$ is mentioned as one group is in Line 18. Line 54-55, if $G$ is rich and expressive, then all subgroups well-approximated by a group $g$ in $G$ will be "covered" by whom? What does this sentence mean? Technical Quality: 3 Clarity: 2 Questions for Authors: See Weaknesses. Confidence: 3 Soundness: 3 Presentation: 2 Contribution: 3 Limitations: No concerns. Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: Thank you for your insight and helpful review. Your comments will be very helpful in improving the presentation of the final paper. We will take care in the camera-ready version to make revisions to Section 1 that improve its clarity. In particular, the additional space of a page should allow us to address these issues adequately. - Lines 54-55: In this line, we mean that if $\mathcal{G}$ is expressive enough, then any potentially relevant subgroup should be well-approximated by some $g \in \mathcal{G}$, and, therefore, any individual $x \in g$ should be ensured that the quality of their predictions are as good as the best possible prediction for their group. When we wrote that those subgroups would be "covered," we meant that those subgroups should be included in, or, at least, well-approximated by some $g \in \mathcal{G}$, so our guarantees apply to that subgroup. We will make this wording more clear in the revision, replacing this clause with "...but multi-group learning with a very rich and expressive family of groups $\mathcal{G}$ ensures that, as long as a subgroup is well-approximated or within the collection $\mathcal{G}$, it obtains our theoretical guarantees. This motivates dealing with large or potentially infinite $\mathcal{G}$ to provide guarantees for as many subgroups as possible." - Line 18 + Line 76: Thank you for bringing this to our attention. We will clarify early in the introduction that $g$ (lowercase) denotes a single subset of $\mathcal{X}$. In the "Summary of Results," we will clarify that our regret guarantee is specific to each $g \in \mathcal{G}$ by starting the claim in Line 75 with, "...for each particular group $g \in \mathcal{G}$." In addition to the comments you raised, we will clarify the descriptions in the Summary of Results using the added space. In particular, we will add sentences describing the parameters in the regret bounds, and, potentially, if there's still space, add a table to compare to results from previous work. We included a draft of such a table in the global Author Rebuttal PDF. Please let us know, if, in addition to the proposed modifications above, there are any other clarity issues you noticed. --- Rebuttal Comment 1.1: Title: Official Comment by Reviewer K3De Comment: Thank you very much for the response. Please include the discussions in revision for better readability. I will keep the current review.	Rebuttal 1: Rebuttal: We have included a draft of the table that will appear in the Introduction section of the camera-ready version of our work to address Reviewers K3De and sHAP. This table includes a summary of existing results in comparison to ours. Pdf: /pdf/0a161b055cf97bcbddcc915a8e22e2b8787ab01b.pdf	NeurIPS_2024_submissions_huggingface	2,024	null	null	null	null	null	null	null	null
On the Scalability of GNNs for Molecular Graphs	Accept (poster)	Summary: This paper investigates the scaling of GNNS across various settings including width, depth, number of molecules, number of labels, and diversity in pretraining across 12 datasets. Different conclusions were drawn from experiments conducted during both pretraining and fine-tuning stages. Finally, the authors introduce a foundational model named MolGPS which integrates various findings. Strengths: 1. The experiments are conducted extensively, taking into account various variables, resulting in some interesting findings. 2. The problem addressed in this paper—the scaling law problem in the molecular domain—remains unresolved but is crucial for advancing molecular representation learning. Weaknesses: 1. The pre-training strategy employed in this paper is a supervised method, overlooking a range of existing self-supervised approaches in molecular representation learning [1][2][3]. It is unconventional to explore scaling laws without considering established pre-training strategies like masking and denoising. 2. The pre-training data in this paper consists of only 5 million molecules. While this scale may be constrained by the label requirements in supervised pre-training, it is insufficient for exploring scaling laws in the molecular domain. Many works use significantly larger datasets, often involving hundreds of millions or even billions of molecules for pre-training their models. 3. The scaling of molecular data appears to perform poorly under the downstream fine-tuning and probing testing protocols. Are there any insights or analyses regarding this phenomenon? 4. From the foundational model perspective, it is also important to evaluate quantum molecular properties such as QM9 and MD17 beyond tasks related to biological ADME properties. [1]:Uni-Mol: a universal 3d molecular representation learning framework [2]:Pre-training Molecular Graph Representation with 3D Geometry [3]:Fractional Denoising for 3D Molecular Pre-training Technical Quality: 2 Clarity: 2 Questions for Authors: See weakness Confidence: 5 Soundness: 2 Presentation: 2 Contribution: 2 Limitations: The paper attempts to uncover the significant scaling law challenges in molecular representation learning. However, it is limited by its focus on a single supervised strategy and constrained data scaling. Despite conducting extensive experiments, these efforts fall short of fully addressing the issue and providing actionable conclusions to the community. Flag For Ethics Review: ['No ethics review needed.'] Rating: 3 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We thank the reviewer for providing detailed feedback on the paper which is of utmost value to our work. Below we address your concerns point-by-point. We also kindly invite the reviewer to refer to the general rebuttal where we share further information and a summary of the feedback we received. >The pre-training strategy [is] supervised [...], overlooking a range of existing self-supervised approaches [...] Thank you for your valuable feedback. We agree that our work utilizes a lesser explored strategy of supervised pretraining. Unsupervised molecular models are built on years of cutting-edge pretraining research [2,3,4] with vast empirical promises [6]. In fact, our work, to our best knowledge, is the first to successfully scale supervised pretraining for GNNs to the multi-billion parameter regime, setting new downstream task performance standards for the multiple benchmarks considered in our submission. In Table 1 and 2 of the rebuttal doc, we compare our approach to self-supervised strategies and find that our method maintains the clear upper hand. The unsupervised MolE [10] model variants clearly underperform the standard MolE model that leverages both unsupervised and supervised pretraining (Table 2), which is in turn outperformed by the listed MolGPS variants by large margins. Even our smaller MPNN++ models of comparable size to MolE (~100M parameters) both outperform the self-supervised variant. In Table 1, we compare to the unsupervised GraphMVP [11] on MoleculeNet (only on tasks that were not part of our pretraining). Our MolGPS from the original submission outperforms [11] on 3/4 tasks, while the model variants that we recently pretrained with additional Phenomics data outperform across all tasks. We acknowledge the results reported by Uni-Mol [12] slightly surpass our score, but note that [12] uses a different data splitting approach compared to GraphMVP, which can have a significant impact in low-data regiemes. We have checked the paper and code, but we were unable to find the recipe so far. Lastly, the mentioned reference [13] focuses on 3D molecular modeling, which is outside the scope of the present work (as discussed below). >The pretraining data [...] of only 5 million molecules. We would like to point out that due to the supervised pretraining, the scale of the dataset is not directly comparable. We also note that previous works in GNN literature scale to less than 5M graphs [14] and show interesting scaling trends. In our case, it is important to characterize the diversity not only by the number of molecules, but instead paired with the number of labels per molecule. We recall our label scaling experiments that show the impact of reducing the data diversity by removing labels. The performance gains from incorporating Phenomics data further support this hypothesis (Figure 2 of rebuttal). >The scaling of molecular data appears to perform poorly [for] the downstream [...]. We indeed found no consistent trend on downstream task performance for increased number of molecules in the pretraining dataset (Figure 2(a) and Appendix E.3 of our submission). That said, no molecular scale (except for the lowest 12.5%) had an outstanding negative effect on downstream tasks. We therefore hypothesize that our approach is robust to a smaller number of molecules during pretraining when the highly important label diversity per molecule is maintained (as discussed above). >[I]t is also important to evaluate quantum molecular properties such as QM9 and MD17. Thank you for the suggestion. However, datasets such as QM9 and MD17 require the model to reason over graph geometry in 3D coordinate spaces. Evaluating on such tasks would be an uneven evaluation of a 2D model such as MolGPS being used to learn 3D downstream tasks. We would like to also highlight the strong performance of MolGPS on the complementary gene inhibition tasks (e.g., the pkis2-* task series of Polaris benchmark) and the significant improvements of our most recent model variants that have been pretrained with additional Phenomics data (Figure 1 in rebuttal pdf). Furthermore, ADME(T) tasks capture the efficacy and toxicity of bioassays which play a crucial role in screening and evaluation of likely drug candidates. These tasks continue to remain the pinnacle of industrial pharmaceutical evaluation [15,16]. We thus prioritize our empirical evaluation by keeping their real-world applications and biological importance in mind. [2]. Li et al, A knowledge-guided pre-training framework for improving molecular representation learning, Nature 2023 [3]. Xia et al, Pre-training Graph Neural Networks for Molecular Representations: Retrospect and Prospect, ICML AI4Science Workshop 2022 [4]. Li et al, KPGT: Knowledge-Guided Pre-training of Graph Transformer for Molecular Property Prediction, ACM SIGKDD 2022 [6]. Lu et al, Learning to Pre-train Graph Neural Networks, AAAI 2021 [7]. Sun et al, Does GNN Pretraining Help Molecular Representation?, NeurIPS 2022 [8]. Sun et al, MoCL: Data-driven Molecular Fingerprint via Knowledge-aware Contrastive Learning from Molecular Graph, ACM SIGKDD 2021 [10] Méndez-Lucio et al, MolE: a molecular foundation model for drug discovery, arxiv 2022 [11] Liu et al, Pre-training molecular graph representation with 3d geometry. arXiv preprint arxiv 2021 [12] Zhou et al, Uni-mol: A universal 3d molecular representation learning framework, ChemRxiv 2023 [13] Feng et al, Fractional denoising for 3d molecular pre-training, ICML 2023 [14]. Chen et al, Uncovering neural scaling laws in molecular representation learning, NeurIPS, 2024 [15]. Shi et al, Fine-tuning BERT for automatic ADME semantic labeling in FDA drug labeling to enhance product-specific guidance assessment, Journal of biomedical informatics, 2023 [16]. Walter et al, Multi-task ADME/PK Prediction at Industrial Scale: Leveraging Large and Diverse Experimental Datasets, Molecular Informatics, 2024 --- Rebuttal Comment 1.1: Comment: Thank you for the authors' detailed feedback. However, I still believe that validating the scaling law using the available labeled data may make a trivial contribution to the community. As highlighted in the abstract, "Scaling deep learning models has been at the heart of recent revolutions in language modeling and image generation." The validation of scaling laws in CV and NLP domains largely relies on unlabeled data and self-supervised tasks [1,2,3], as referenced in the paper. I argue that a more relevant approach for studying scaling is under an unlabeled setting, and there are already studies that explore this [4]. In the biological domain, obtaining labeled data is more expensive and challenging compared to the CV or NLP domains. Validating scaling laws in such a constrained setting seems unusual. The effectiveness of this approach may only depend on the relevance of supervised tasks between pre-training and fine-tuning. [1]: OpenAI. Gpt-4 technical report [2]: Llama 2: Open foundation and fine-tuned chat models [3]: Language models are unsupervised multitask learners. [4]: Uni-Mol2: Exploring Molecular Pretraining Model at Scale --- Reply to Comment 1.1.1: Title: Response to comment by Reviewer Lxfs Comment: We thank the reviewer for their response and are happy to provide further clarification of the concerns mentioned. >The validation of scaling laws in CV and NLP domains largely relies on unlabeled data and self-supervised tasks [1,2,3] [...]. [A] more relevant approach for studying scaling is under an unlabeled setting [e.g.,] [4]. Unsupervised approaches are an interesting avenue of research in certain areas of the molecular domain. As suggested, Uni-Mol models [4,5] show promising results for modeling molecules in 3D space and for finetuning to physics-based downstream tasks like QM9. We would like to remind the reviewer that we follow a significantly different approach based on modeling 2D graphs with different applications & downstream tasks, where unsupervised models have not yielded comparable results. In the context of CV/NLP [1,2,3] we also point out the lack of equally label-rich datasets that explains why supervised pretraining is not a focal point in those domains. With respect to Uni-Mol2 [4], we note that it has been only finetuned to physics-based tasks (QM9, e.g., homo-lumo gap), that highly depend on 3D coordinates and graph structure. [4] is specialized on understanding graph structure in 3D space, which leads the approach to perform well on such tasks. >We hope that our work paves the way for an era where foundational GNNs drive pharmaceutical drug discovery. [abstract of our submission] Our objectives are different: predicting properties that drive drug discovery like ADMET and binding predictions that strongly rely on understanding the biochemical space beyond graph structure. Uni-Mol2 provides no results on any shared or similar downstream task. While the initial Uni-Mol model [5] provided some results for ADMET downstream tasks, the derivation of the results lacks transparency and is not reproducible (despite our best efforts). Uni-Mol's finetuning compares to GraphMVP [6] (published one year before Uni-Mol), which we also compare to in our work, closely following their experimental setup. Uni-Mol outperforms GraphMVP but reports performance directly taken from [6], despite clearly using a different dataset splitting technique in their work. We reiterate that data splits have a significant impact on empirical results in the context of molecular scaffold splits, especially in the low-sample regime of the MoleculeNet. Further, [5] does not provide sufficient information for reproducibility. We have rigorously analyzed the papers and code of both works and provide details in a separate comment below. Overall, the learnings from our scaling study (width, depth, #molecules, #labels per molecule, composition of pretraining data mixture) lead to the derivation of a foundational GNN that has set a new standard in the various competitive downstream task benchmarks considered here. We kindly request the reviewer to provide more context on their assessment of our work as “a trivial contribution” in light of the above discussion. We would be happy to clarify any further questions the reviewer may have. >In the biological domain, obtaining labeled data is more expensive and challenging compared to the CV or NLP domains. [...] The effectiveness of this approach may only depend on the relevance of supervised tasks between pre-training and fine-tuning. We respectfully disagree with this assessment. Adequate labeled data is readily and publicly available at large scale with our pretraining only using a small fraction of the available data. We recall that PCBA_1328 is only a small subset of the PubChem database (only considering datasets with binary tasks with at least 6k molecules) and larger alternatives for PCQM4M exist (e.g., PM6 dataset in [7] that is more than 20x bigger). We also recall that our results show no signs of a data bottleneck. Instead, our molecule scaling suggests, downstream task performance does not deteriorate much when pretraining on a smaller fraction of molecules, e.g., 25% or 50% (Fig. 2 of submission). This relates back to the importance of the number of labels (i.e., label scaling in submission) per molecule already discussed in our initial response. The objective of our work, as in CV/NLP, is to obtain informative embeddings for domain-specific downstream applications. For drug discovery, our work suggests supervised pretraining is a suitable avenue, thanks to the availability of large labeled data sources. As there is no overlap between the pretraining and finetuning tasks, our downstream evaluation solely evaluates the information content of our learned representations (similar to other domains). We are happy to provide more information if further questions arise. [1,2,3,4] as referenced in previous comment of Reviewer Lxfs [5] Uni-Mol [6] Liu et al, Pre-training Molecular Graph Representation with 3D Geometry, ICLR 2022 [7] Beaini et al, Towards Foundational Models for Molecular Learning on Large-Scale Multi-Task Datasets, ICLR 2024 --- Rebuttal 2: Comment: We hope our two comments above addressed the reviewer’s concerns. Please let us know if there are any additional clarifications that we can provide that could increase your support for the acceptance of the paper.	Summary: This paper investigates the scaling of GNNs on molecular tasks. In their setting, they pre-train a GNN on multiple molecule dataset and then fine-tune these GNNs on different datasets (“down-stream tasks”). They investigate how the performance of different GNNs changes with parameter count, depth, training samples and so on. Strengths: - (S1 - Clarity) The paper is extremely well written and easy to follow. - (S2 - Significance) Both foundation models for graphs and ML on molecules are very active research fields. As such this paper can be important to a large number of people. - (S3 - Novelty and Significance) To me, there are four interesting and novel contributions / observations in this paper: 1) Scaling laws for GNNs. 2) The investigation of how to train a pre-trained GNN on a downstream task. Here they investigate two options: fine-tuning the entire model and training only the final MLP (probing). 3) A comparison of three different types of GNNs: MPNNs, Transformers and MPNNs+Transformers. It is particularly interesting that GNNs scale better with parameters than transformers. 4) The observation that removing a dataset from the training data can improve performance on downstream tasks. - (S4 - Quality) The evaluation seems solid, the authors compare three different models and observe mostly similar behavior for all models across datasets. Weaknesses: - (W1 - Clarity) The presentation of the results in the main paper is done poorly. Figure 2 is too small and difficult to read / understand. I understand that this paper presents a lot of results but putting all of them into a single figure is in my opinion a bad choice. - (W2) Fine tuning vs probing: you have experiments to compare these two approaches. However, I could not find an analysis of the results in the paper (maybe I have missed it). Technical Quality: 4 Clarity: 3 Questions for Authors: - (Q1) How can you train a single model on multiple datasets if these datasets might have different features? Are all datasets pre-processed to have the same features? - (Q2) Figure 2 - probing on Polaris / TDC: it seems that most models do not profit from an increase in the number of molecules. Does this not contradict the idea of foundation models? It seems to me like this indicates that when pre-training GNNs the size of the original dataset does not matter much to the downstream task. - (Q3) In a similar vein, you notice that you can get better results when pretraining without the L1000 dataset. Does this imply that for graphs it is more important to have the right kind of data compared to having a lot of data (as for example LLMs or diffusion models)? - (Q4) From (W2): how does fine-tuning compare to probing? Is one clearly better than the other? To sum up, this is a good experimental paper that should be accepted. Confidence: 4 Soundness: 4 Presentation: 3 Contribution: 3 Limitations: Yes Flag For Ethics Review: ['No ethics review needed.'] Rating: 7 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We thank the reviewer for providing detailed feedback on the paper which is of utmost value to our work. Below we address your concerns point-by-point. We also kindly invite the reviewer to refer to the general rebuttal where we share further information and a summary of the feedback we received. >(W1) Figure 2 is too small and difficult to read We thank the reviewer for the comment and agree that Figure 2 can be better represented when split up into multiple larger figures. We will apply those changes in a revised version of our paper. >(W2 & Q4) Fine tuning vs probing: How does fine-tuning compare to probing? We thank the reviewer for this helpful comment. We will further clarify this point when revising the paper. Section 4.2 of the submission studies finetuning and probing side-by-side, establishing both as effective strategies for tackling downstream tasks, without conducting a direct comparison between them. However, when deriving the foundation model MolGPS in Section 4.3, we find probing to be the overall stronger approach. The major advantage of probing is the ability to leverage multi-level information from the pretrained GNN. We recall that our pretraining is based on a supervised multi-task learning approach. As a result, different task heads capture task-specific information, while earlier layers that feed into the task heads carry more general information. When we combine fingerprints from various layers, we can think of aggregating knowledge from several “experts”. Our multi-fingerprint probing suggests this knowledge is additive as it clearly outperforms probing of any single fingerprint. Our foundation model MolGPS doubles down on this idea, combining fingerprints from multiple layers and various pretrained models. For finetuning, there is no straightforward way of taking advantage of this multi-level information, making probing the preferred approach. >(Q1) How can you train a single model on multiple datasets if these datasets might have different features? Are all datasets pre-processed to have the same features? All datasets are indeed pre-processed to have the same features. Each molecule is initially represented by a SMILES string [1], that our pipeline features into a molecular graph (i.e., atoms and bonds as nodes and edges, respectively). The (node) features are also agnostic to the data source, e.g., using atom type, row/col in periodic table, etc.. We will add more context on the pre-processing to the appendix. >(Q2) Figure 2 - probing on Polaris / TDC: it seems that most models do not profit from an increase in the number of molecules. Does this not contradict the idea of foundation models? We indeed found no consistent trend on downstream task performance for increased number of molecules in the pretraining dataset (Figure 2(a) and Appendix E.3 of our submission). That said, no molecular scale (except for the lowest 12.5%) had an outstanding negative effect on downstream tasks. We therefore hypothesize that our approach is robust to a smaller number of molecules during pretraining when the highly important label diversity per molecule is maintained that can be observed in our label scaling study (e.g., in Figure 2(a) and Appendix E.4 of our submission). We would like to point out the importance of characterizing the diversity of the data not only by the number of molecules, but also paired with the number of pretraining labels/tasks per molecule. PCBA_1328 considers more than 1k different labels per molecule (albeit with high sparsity) and PCQM4M comes with 25 graph-level tasks and 4 node-level tasks (that are learned for each node of the ~4M molecules). This perspective is reinforced by our label scaling experiments that show the impact of reducing the data diversity by removing labels (e.g., Figure 2(a) of submission). Performance gains from recently incorporating Phenomics data into the pretraining data mix further support this claim, adding ~500k of molecules with a highly informative set of ~6k labels per graph (Figure 2 of rebuttal). >(Q3) In a similar vein, you notice that you can get better results when pretraining without the L1000 dataset. Does this imply that for graphs it is more important to have the right kind of data compared to having a lot of data (as for example LLMs or diffusion models)? This is an accurate observation, which relates back to the importance of the number of molecules and the number (and quality of the labels). L1000 is exceptionally small (only 20k molecules) but features almost 1k labels for each molecule. However, in contrast to, e.g., PCBA_1328 (1328 binary classification tasks), it is a regression task that suffers from low signal-to-noise ratio. Despite trying to stabilize learning by transforming it into a classification task via binning of regression targets, the overall impact on downstream task performance was negative. Notably, the small amount of molecules may lead to the absence of positive samples in most batches, as the batches are dominated by the larger PCBA_1328 and PCQM4M molecules. As pointed out by the reviewer, it is the quality (and label diversity) of the data that is highly relevant. The perfect example for this is the massively improved downstream task performance we report in the rebuttal doc after recently adding Phenomics data (with ~500k molecules and ~6k highly informative labels) to our previously use pretraining data mix (excluding L1000). We kindly refer the reviewer to the general rebuttal doc for more details. Kindly let us know if our response above addressed your concerns. We will be happy to further discuss and answer any questions you may have. [1] Weininger D, SMILES, a chemical language and information system. 1. Introduction to methodology and encoding rules. Journal of chemical information and computer sciences. 1988 --- Rebuttal Comment 1.1: Comment: I thank the authors for answering my questions and providing such a thorough rebuttal. I am still of the opinion that this is a good paper which should be accepted and will thus keep my score.	Summary: The paper examines the scaling behavior of GNNs in the context of molecular graph representation and prediction. The authors analyze various GNN architectures, including message-passing networks, graph Transformers, and hybrid models, on a large dataset of 2D molecular graphs. They find that increasing the scale of model depth, width, dataset size, and diversity significantly enhances performance on downstream tasks. They introduces MolGPS, a new graph foundation model, which demonstrates superior performance across numerous tasks in molecular property prediction. Strengths: 1. The paper is well-written and easy to follow. 2. Comprehensive experiments were performed on large scale datasets to demonstrate the scaling law. 3. The empirical results of MolGPS on various dataset is promising. 4. The insights on how to scale GNN models is meaningful to the community. Weaknesses: 1. Another important aspect of scaling the parameters is regularization [1]. The authors did not explicitly discuss / validate the regularization methods they are using. 2. The authors did not control the compute used to train the models. It would nice to see some efficiency comparison as well. [1] How to train your ViT? Data, Augmentation, and Regularization in Vision Transformers https://arxiv.org/abs/2106.10270 Technical Quality: 3 Clarity: 4 Questions for Authors: 1. How well do these insights generalize to other types of graph-structured data beyond molecular graphs? Confidence: 4 Soundness: 3 Presentation: 4 Contribution: 3 Limitations: NA Flag For Ethics Review: ['No ethics review needed.'] Rating: 7 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We thank the reviewer for their valuable feedback on the paper which is of utmost value to our work. Below we address your concerns point-by-point. We also kindly invite the reviewer to refer to the general rebuttal where we share further information and a summary of the feedback we received. >Another important aspect of scaling the parameters is regularization [1]. The authors did not explicitly discuss / validate the regularization methods they are using. Our model indeed uses regularization techniques, e.g., dropout in each GNN layer. We will include a detailed discussion of the used methods and their parameterization in the appendix of the revised paper. The need of more sophisticated regularization techniques is largely avoided through the use of rich pretraining data that prevents the models from overfitting (except for the MPNN++ model in a few instances). >The authors did not control the compute used to train the models. It would nice to see some efficiency comparison as well. Thank you very much for the suggestion. Model performance tables with the training compute time will be added to the appendix. Training time of a single model varies from 24 GPU hours for the smallest models to ~1120 GPU hours for the largest models. In total ~13400 GPU hours were used for computation in this paper. The work of [1] pointed out by the reviewer presents a compelling analysis of the used computational resources compared between finetuning to a downstream task and training a model of the same size from scratch. We note that such a study may be difficult to replicate in our case due to the high parametric scale of the GNNs in contrast to the low-data regimes for the downstream tasks with only a few 100's or 1,000's of molecules per task and almost exclusively a single target label, which would cause overfitting when trained from scratch. >How well do these insights generalize to other types of graph-structured data beyond molecular graphs? We thank the reviewer for their question on potential applications outside the biochemical domain. It is highly likely that models similar to MolGPS could be pretrained for other applications and domains given adequate pretaining data. However, the present work would not be suited for downstream tasks outside the biochemical domain. To obtain a domain-agnostic graph-foundation model, unsupervised approaches that primarily learn from the graph structure may be a more natural choice. That said, the proposed foundation model MolGPS is specifically targeted towards biochemistry and drug discovery in particular, with the pretraining data chosen accordingly to learn domain-specific information. For comparison, we provide some comparison to unsupervised pretraining approaches in the rebuttal doc (Tables 1 & 2) that indicate our supervised pretraining strategy is to be preferred for a “molecular” graph foundation model. Kindly let us know if our response above addressed your concerns. We will be happy to further discuss and answer any questions you may have. [1] Steiner et, How to train your ViT? Data, Augmentation, and Regularization in Vision Transformers, arxiv 2021 --- Rebuttal Comment 1.1: Comment: I'm satisfied with the author's response and I'll keep my score.	Summary: The paper investigates the scalability of GNNs in molecular graph applications. It highlights the relationship between model size, dataset size, and performance upon message-passing networks, graph Transformers, and hybrid architectures using a large dataset of 2D molecular graphs. It shows that supervised pretraining on molecular graphs provides rich embeddings beneficial for downstream tasks. And it finds that the number of labels is crucial for fine-tuning performance. Strengths: The experiment of MPNN and Transformer on pretrained and downstream tasks is very thorough and provides ample evidence to support the claim. The ablation studies are detailed. Weaknesses: Many scaling results are unsurprising, such as the consistent improvement with increased depth and width during pretraining. However, the thorough experiments compensate for this predictability. I am curious if these trends will also occur for 3D graph datasets like QM9 or others. Conducting a few additional experiments on these datasets could strengthen this paper, though this isn't necessarily required in this rebuttal. Technical Quality: 3 Clarity: 3 Questions for Authors: See weaknesses. Confidence: 3 Soundness: 3 Presentation: 3 Contribution: 3 Limitations: The authors claim that this work uncovers additional aspects of GNN training such as the increasing complexity of aggregation functions and their effect on scaling properties. Flag For Ethics Review: ['No ethics review needed.'] Rating: 6 Code Of Conduct: Yes	Rebuttal 1: Rebuttal: We thank the reviewer for providing detailed feedback on the paper which is of utmost value to our work. Below we address your concerns point-by-point. We also kindly invite the reviewer to refer to the general rebuttal where we share further information and a summary of the feedback we received. >Many scaling results are unsurprising, such as the consistent improvement with increased depth and width during pretraining. However, the thorough experiments compensate for this predictability. Thank you for appreciating the quality of our empirical analysis and finding our findings interesting. We agree that, especially given the unprecedented benefits of scale in other domains such as natural language, some results of this work may seem unsurprising at first. In the context of GNNs however, our work explores largely uncharted territory in terms of parametric scale [1,2,3] and the possible implications on biochemistry and drug discovery in particular. We recall the finetuning and probing performance on downstream tasks, establishing new state-of-the-art on 12/22 tasks on highly-competitive TDC benchmark and all but two tasks on Polaris and MoleculeNet benchmarks with our proposed foundation model (MolGPS). In our rebuttal we have expanded our empirical study, adding model variants of MolGPS that were scaled to the 3B parameter regime andpretrained on an improved data mix that further enhances our performance across the downstream task benchmarks. We kindly refer to the general rebuttal for more details. >I am curious if these trends will also occur for 3D graph datasets like QM9 or others. Conducting a few additional experiments on these datasets could strengthen this paper, though this isn't necessarily required in this rebuttal. Thank you for the suggestion. We note that our proposed MolGPS model solely relies on learning from 2D graphs, while tasks in QM9 depend on molecules with 3D positions. This is one of the reasons why we limit our analysis of downstream tasks to 2D molecular tasks. It is important to highlight the importance of publicly available labeled data for our supervised pretraining setup. Our label scaling experiments (e.g., Figure 2a of submission) highlight the importance of many diverse labels for each molecule. Current databases for 3D molecules are still fairly small and would hardly allow for the required data diversity present in our 2D pretraining data collection. However, we note that scaling design decisions uncovered in our empirical analysis may prove useful to construct and train foundational 3D GNNs in the future, when databases reach adequate scales. These models would further require the use of sophisticated training strategies and aggregation functions, a future direction which we explicitly mention in our limitations section. Kindly let us know if our response above addressed your concerns. We will be happy to further discuss and answer any questions you may have. [1]. Sun et al, Does GNN Pretraining Help Molecular Representation?, NeurIPS 2022. [2]. Sun et al, MoCL: Data-driven Molecular Fingerprint via Knowledge-aware Contrastive Learning from Molecular Graph, ACM SIGKDD conference on knowledge discovery & data mining 2021. [3]. Liu et al., Neural Scaling Laws on Graphs, arxiv 2024. --- Rebuttal Comment 1.1: Comment: Thanks for your response! I will keep my positive score.	Rebuttal 1: Rebuttal: We thank the reviewers for providing detailed feedback on the paper and appreciating its presentation (YrHv, kaW1, FQ72, Lxfs), organization (kaW1, FQ72) and scientific contribution (FQ72, Lxfst). Overall, we have updated the paper and responded to the reviewer’s concerns in the individual rebuttals. In the following, we discuss our updates and reviewer comments that might be relevant for all reviewers. ## Incorporating Phenomics data We further integrated an additional data type into our pretraining data mix. The added Phenomics dataset contains ~6k labels for ~500k molecules (compounds) that were derived from phenomic imaging [1] of cells perturbed with either a dose of a compound or a gene knockout. We conducted a similarity analysis between the obtained images (represented by vector embeddings, e.g., similar to [2]) subject to a compound perturbation on one side, and images subject to a gene perturbation on the other side. The pretraining task is to predict for each compound if it has a phonemically visible similarity to a gene knockout (indicating a biological relationship). Adding Phenomics data to our pretraining data mix (i.e., PCBA_1328 and PCQM4M), improved our downstream task performance across the board (Figures 1 and 2 of rebuttal doc). Comparing the scaling trends in Figure 2 of the rebuttal doc, MPNN++ w/ Phenomics pretraining exhibits a significant vertical upwards shift compared to the original MPNN++. Notably, we were also able to extend our scaling study to the 3B parameter regime. While we were previously unable to extend the scaling trend (Figure 2 of the rebuttal doc), MPNN++ w/ Phenomics maintains a positive scaling trend. We note that, in a slight deviation from the figure shown in our submission, Figure 2 of the rebuttal doc shows scaling trends for MPNN++ instead of GPS++. This is because our recent experiments with added Phenomics data were only conducted with MPNN++ for parameter scales <1B due to the availability of compute. To better visualize the impact of our results on the TDC benchmark collection [3], we have renamed and added few baselines compared to Figure 2(b) of the original submission. We renamed TDC SOTA to “Best model per task” (a collection of 8 different models that together establish SOTA across the benchmark collection) and added MapLight + GNN [1] (the best single model out of those 8 methods) evaluated on the benchmark collection, which falls significantly short of the newly added MPNN++ variant w/ Phenomics pretraining. We are also thrilled to report that our MolGPS w/ Phenomics even outperforms the best model per task at parameter scale 3B and performs on par at scale 1B. Lastly, we added a purely self-supervised MolE variant [4] to represent unsupervised pretraining strategies (which is further discussed below). ## Comparison with unsupervised methods In response to the feedback of Reviewer Lxfs, we added further comparison to unsupervised pretraining approaches (Table 1 and 2 of rebuttal doc). We observe that the unsupervised MolE model variants [4] clearly underperform the standard MolE model that leverages both unsupervised and supervised pretraining (Table 2), which is in turn outperformed by the listed MolGPS variants by large margins. Even our smaller MPNN++ models of comparable size to MolE (with ~100M parameters; see Figure 2 of rebuttal doc at the comparable parametric scale of 100M MPNN++) both outperform the self-supervised variant. In Table 1, we compare to the self-supervised GraphMVP [11] model on MoleculeNet (which was already featured in the submission; we only consider the tasks that were not part of our pretraining). TheMolGPS model from the original submission outperforms GraphMVP on 3/4 tasks, while the model variants that we recently pretrained with additional Phenomics data outperform across all tasks. We overall conclude that our pretraining strategy, is a promising avenue of molecular pretraining and novel at the parametric scale of the present work. ## Dataset scale in the context of supervised molecular pretraining We would like to point out that due to our supervised pretraining approach, the scale of the dataset is not directly comparable to that of self-supervised methods, i.e, billions of molecules in some cases. We note however that previous works in GNN literature scaled to even less than 5M graphs [5] and observe interesting scaling trends. In our case, it is important to characterize data scale not only by the number of molecules, but instead paired with the number of pretraining labels per molecule. PCBA_1328 considers more than 1k different labels per molecule (albeit with high sparsity) and PCQM4M comes with 25 graph-level tasks and 4 node-level tasks (for each node of the ~4M molecules). This is further confirmed by our label scaling study that shows the impact of reduced data diversity by removing labels (e.g., Figure 2(a) of submission). The performance gains from incorporating Phenomics data into the pretraining data mix also support this, adding ~500k of molecules with ~6k highly informative labels per graph (Figure 2 of rebuttal). ## Applicability to downstream tasks with 3D molecules like QM9 and MD17 As downstream tasks from 3D molecular modeling were mentioned by 2 reviewers, we would like to clarify that our modeling approach operates on 2D molecules, while datasets such as QM9 and MD17 require the model to reason over graph geometry in 3D coordinate spaces. [1] Bray et al, Cell Painting, a high-content image-based assay for morphological profiling using multiplexed fluorescent dyes, Nat Protoc 2016 [2] He et al, Masked autoencoders are scalable vision learners, CVPR 2022 [3] Notwell et al, Admet property prediction through combinations of molecular fingerprints, arxiv 2023 [4] Méndez-Lucio et al, MolE: a molecular foundation model for drug discovery, arxiv 2022 [5] Chen et al, Uncovering neural scaling laws in molecular representation learning, NeurIPS, 2024 Pdf: /pdf/0eddc453740fe4cf3e13e3338e99e8a832330f2d.pdf	NeurIPS_2024_submissions_huggingface	2,024	null	null	null	null	null	null	null	null

Subsets and Splits

No community queries yet

The top public SQL queries from the community will appear here once available.