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
Geometry-Calibrated DRO: Combating Over-Pessimism with Free Energy Implications	Reject	Summary: The paper demonstrates a root cause of the over-pessimism issue of existing distributionally robust optimization (DRO) methods: excessive focus on noisy samples. To mitigate this issue, the authors proposed a novel DRO method called Geometrically-Calibrated DRO (GCDRO). They introduce the free energy implicati...	Rebuttal 1: Rebuttal: We sincerely appreciate your approval of the theoretical and empirical contributions of this work. Thank you very much for the advice to improve this paper. And we address your concerns as follows: ### 1. Paper organization. Thanks for your suggestions. We will highlight the key difference betwe...	Summary: In this work, the authors propose a novel approach called Geometry-Calibrated Distributionally Robust Optimization (GCDRO) to address the over-pessimism issue in traditional Distributionally Robust Optimization (DRO) methods. DRO aims to optimize worst-case risk within an uncertainty set to mitigate the effect...	Rebuttal 1: Rebuttal: Thanks for your time and efforts in the reviewing process and we appreciate your approval of our methodological novelty, theoretical soundness, and empirical significance. * For baselines, we have compared most of the typical DRO methods, including two $\phi$-divergence DROs: KL-DRO and $\chi^2$...	Summary: This work investigates the impact of noisy samples on Distributionally Robust Optimization (DRO) algorithms. Using a simple model and empirically, they show how DRO variants tend to give to much importance to those samples, which make them overly pessimistic in finding the worst case distribution shift. To sol...	Rebuttal 1: Rebuttal: We sincerely appreciate your approval of the motivation, novelty, and potential impacts of this work. Thank you very much for the advice to improve this paper. And we address your concerns as follows: ### 1. How $G_N$ is obtained: We construct a $k$-NN graph using the raw covariates for each dat...	Summary: This paper introduces a novel approach to tackle the issue of overconservativeness in conventional DRO models. The proposed method incorporates data geometry properties into the design of the objective function and ambiguity set. The authors leverage the discrete geometric Wasserstein distance, initially prese...	Rebuttal 1: Rebuttal: We sincerely appreciate your approval of the motivation and novelty of this work. Thank you very much for the advice to improve this paper. And we address your concerns as follows: ### 1. Comparison between WDRO and GCDRO * when WDRO is better: WDRO could extend the distribution support and...	null	NeurIPS_2023_submissions_huggingface	2,023	null	null	null	null	null	null	null	null
Scalable Primal-Dual Actor-Critic Method for Safe Multi-Agent RL with General Utilities	Accept (poster)	Summary: The work extends constrained single-agent general utilities RL (where the objective is a non-convex function of occupancy measure) to a constrained multi-agent setting. Naturally, similar to single-agent, the constraints are satisfied in expectations. They assume the combined (global) objective to be additive ...	Rebuttal 1: Rebuttal: We thank the reviewer for the insightful questions. Please find our responses below. We highly appreciate your re-evaluation of our work and a kind reconsideration of the score. ### Weaknesses > Objective: ... exploration objective. A: We admit that our current formulation still requires structur...	Summary: This paper studies the safe MARL with general utilities and focuses on the setting of distributed training without global observability. The authors propose a primal-dual actor-critic method with shadow reward and \kappa-hop policy to tackle the problems of scalability and incorporating general utilities. They...	Rebuttal 1: Rebuttal: We thank the reviewer for appreciating our work. We hope our responses below provide further clarity. ### Weakness > Lack of baseline comparison in the main context: I understand that the authors want to display more theoretical analysis so they may move the baseline comparison into the appendix. ...	Summary: This paper considers the safe multi-agent rl in the fully-decentralized setting. In the system, agents need to satisfy local safety constraints, and they also have to maximize the joint objectives. Moreover, the objectives are general utilities, which have broader applications than common discounted reward. Au...	Rebuttal 1: Rebuttal: We thank the reviewer for appreciating our work. We hope our responses below provide further clarity. ### Questions > The potential scope ... distributed way. A: We would like to provide more concrete motivations for studying decentralized MARL. * First, while global state/action spaces are deco...	Summary: This paper considers the problem of safe/constrained multi-agent reinforcement learning (MARL) with general utilities and constraints, which could be generic functions of agents' occupancy measures. A distributed method is proposed for solving the constrained MARL problem with local constraints, where each age...	Rebuttal 1: Rebuttal: We thank the reviewer for the positive feedback. We hope our responses address your concerns and provide further clarity. > (Weakness) The presentation ... in Section 5. > (Weakness) The paper mostly focuses ... continuous state spaces. > (Question) How does the proposed method extend to the cas...	Rebuttal 1: Rebuttal: ## General Response We would like to express our sincere gratitude to the reviewers for reading our paper and providing valuable feedback. Below, we answer some common questions of the reviewers, including the paper organization, the separability of the objective, the **extension to contin...	NeurIPS_2023_submissions_huggingface	2,023	null	null	null	null	null	null	null	null
Training biologically plausible recurrent neural networks on cognitive tasks with long-term dependencies	Accept (poster)	Summary: The paper proposes a method for extracting long-term dependencies in biologically plausible leaky RNNs. The motivation of this work is well-presented and has contribution to the neuroscience field. Strengths: Strengths: * The paper demonstrates significant improvements in terms of performance when compared t...	Rebuttal 1: Rebuttal: We thank the reviewer for the positive comments. > In terms of presentation, it would be helpful to list the properties of datasets used in the experimental part. All cognitive tasks used to train our networks are currently explained in Appendix C in the supplementary. We will elaborate further ...	Summary: The authors suggest a method for accelerating the training of continuous time vanilla recurrent neural networks. Specifically, they add temporal skip connections during training, and gradually remove them throughout the training process. For tasks with long term memory requirements, this can provide up to 3 ti...	Rebuttal 1: Rebuttal: We thank the reviewer for the extensive review. We hope we can address them within the character limit. > The main claim is not verified. Our claim was referring to the reduction in number of training iterations. We understand that this would not be sufficient without data on wall-clock training...	Summary: In this study, the authors propose a couple of architectural innovations to train biological RNNs in a stable manner on a suite of 16 cognitive tasks, some of which require the integration of long-range temporal dependencies. The authors motivate their work from the lack of cortical evidence for gating mechani...	Rebuttal 1: Rebuttal: > ... they mention that using biological RNNs without gates is important but this claim is not backed with any advantage of biological RNNs over ones with gates... We thank the reviewer for raising this important issue. LSTM and GRU architectures were introduced to overcome challenges pertaining ...	Summary: This paper describes and tests several methods for training a class of “biologically-plausible” recurrent neural networks (RNN), which are discrete approximations of continuous-time leaky RNNs. This type of gate-less RNN model has difficulty learning dependences over long time intervals, so tweaks are applied ...	Rebuttal 1: Rebuttal: We thank the author for the positive review and helpful comments. We also thank the reviewer for the suggested relevance to surrogate gradient approaches used in training spiking neural networks – we will mention this in a revised version if given the opportunity. > I found many points in this pa...	Rebuttal 1: Rebuttal: We thank all reviewers for their time and effort in reviewing our submission. We will address common issues here. Several reviewers have suggested moving information between the main text and supplementary. After also taking into account additional work requested by reviewers, we have arrived at...	NeurIPS_2023_submissions_huggingface	2,023	Summary: This paper proposes a way to speed up biological RNN training by up to 3x using well placed skip connections during training time. Strengths: This paper will be of interest to researchers training biological neural networks, especially for hard cases where a 3x time speed up can be achieved. Presentation is c...	Rebuttal 1: Rebuttal: > This paper proposes a way to speed up biological RNN training by up to 3x using well placed skip connections during training time. This paper will be of interest to researchers training biological neural networks, especially for hard cases where a 3x time speed up can be achieved. Presentation i...	null	null	null	null	null	null
S-CLIP: Semi-supervised Vision-Language Learning using Few Specialist Captions	Accept (poster)	Summary: The paper presents S-CLIP, a novel semi-supervised learning method for training contrastive language-image pre-training (CLIP) models in specialized domains where limited image-text pairs are available. Vision-language models like CLIP have achieved impressive results in natural image domains, but they struggl...	Rebuttal 1: Rebuttal: Dear reviewer Q2NM, Thank you for your valuable feedback and comments. We appreciate your remarks on the strengths of our paper, including the problem significance, effective method, extensive experiments, and clear presentation. We will address your concerns and questions in the response below. ...	Summary: The paper studies semi-supervised learning for contrastive vision-language pre-training, during which unlabelled/unpaired data is also accessible. The authors formulate the pseudo-labeling as an optimal transport (OT) problem and propose to use the Sinkhore algorithm to produce pseudo labels, which is validate...	Rebuttal 1: Rebuttal: Dear reviewer XrtM, Thank you for your valuable feedback and comments. We appreciate your remarks on the strengths of our paper, including the technical soundness, sufficient evaluation and ablation, and clear presentation. We will address your concerns and questions in the response below. --- *...	Summary: The paper introduces an additional training method for vision-language models like contrastive language-image pre-training (CLIP). The proposed method consists of semi-supervised learning with two different pseudo labeling including caption- and keyword-level pserudo-label. These two pseudo labelings are compl...	Rebuttal 1: Rebuttal: Dear reviewer 7J1f, Thank you for your valuable feedback and comments. We appreciate your remarks on the strengths of our paper, including the problem significance, simple yet effective method, reasonable experiments, and clear presentation. We will address your concerns and questions in the resp...	Summary: Fine-tuning CLIP in specialized domains suffer the problem of limited available image-text pairs. This paper propose to address this challenge via semi-supervised learning methods with the additional unpaired images. Two pseudo-labeling strategies including the caption-level pseudo labeling and keyword-level p...	Rebuttal 1: Rebuttal: Dear reviewer CjR3, Thank you for your valuable feedback and comments. We appreciate your remarks on the strengths of our paper, including the problem significance, sufficient experiments, and clear presentation. We will address your concerns and questions in the response below. --- **[W1] Defin...	Rebuttal 1: Rebuttal: To answer the Q3 of reviewer EbWV, we have attached a PDF containing qualitative examples of caption-level and keyword-level pseudo-labels. Pdf: /pdf/24445cc683a624331539c896c337e12b58fdd980.pdf	NeurIPS_2023_submissions_huggingface	2,023	Summary: This paper proposes a method S-CLIP, to enhance CLIP in a semi-supervised way by leveraging captions (via optimal transport) and keywords (via image-keyword similarity) from other unpaired images in the same batch. This is beneficial for specialized domains like remote sensing, where image-text data is usually...	Rebuttal 1: Rebuttal: Dear reviewer EbWV, Thank you for your valuable feedback and comments. We appreciate your remarks on the strengths of our paper, including the problem significance, novel and simple method, comprehensive experiments and ablations, and clear presentation. We will address your concerns and question...	null	null	null	null	null	null
A U-turn on Double Descent: Rethinking Parameter Counting in Statistical Learning	Accept (oral)	Summary: The paper aims to explain the well-known phenomenon of double-descent in the scope of several traditional ML methods (random forests, boosting and linear regression on random Fourier features). The main message of the paper (in the mentioned scope) is that what appears to be a double-descent on the first sight...	Rebuttal 1: Rebuttal: We thank the reviewer for the very interesting, in-depth and constructive review! We were delighted by the assessment that Part 1 of our paper was "novel, surprising, important , interesting, neat and probably impactful" and, limited by space constraints, hope to resolve more pressing doubts about...	Summary: This paper critically looks at the double descent phenomena, and it is shown that the number of parameters on the x-axis does not represent the complexity well for double descent curves, and in essence, causes the double descent. If a more appropriate complexity measure is adopted, in this paper, the effective...	Rebuttal 1: Rebuttal: We would like to thank the reviewer for the very interesting, in-depth and constructive review! We are delighted by the reviewer’s overwhelmingly positive review of our paper, in particular the assessment that “it will have a high impact and is relevant for many researchers”! We also put much effo...	Summary: This paper proposes a new perspective for understanding the double descent behavior in non deep learning models: Once the number of raw parameters goes pass the interpolation threshold, the effective parameter number actually stops increasing and if one keeps increasing the number of parameters, the underlying...	Rebuttal 1: Rebuttal: We thank the reviewer for the thoughtful review -- we are very delighted by the very positive assessment of our work! Below, we respond to the points raised in the review in turn. The need to evaluate effective parameters on test samples. We appreciate this reviewer raising this point as we b...	Summary: This paper studies the double descent phenomenon where it has been suggested that optimal generalization performance is achieved as model complexity gets extremely large and training data is perfectly overfit. However, this work suggests that true model complexity is not described simply by the number of param...	Rebuttal 1: Rebuttal: We thank the reviewer for the thoughtful review -- we are delighted by the positive assessment of our presentation, relevance and our insights into the linear regression case study! Connections to Neural Networks. We completely agree that – as discussed in l. 356ff. – understanding double des...	Rebuttal 1: Rebuttal: We would like to once more thank all reviewers for their time and effort put into the review process -- we thoroughly enjoyed reading the interesting and constructive reviews we received. We were delighted by the overwhelmingly positive response to our paper, and are thrilled that all 6 reviewers ...	NeurIPS_2023_submissions_huggingface	2,023	Summary: This paper studies the double descent phenomenon focusing on the complexity axis used to display the phenomenon. For trees and boosting, they show that double descent is a result of peculiar axes used by one previous work and that an alternative formulation of complexity removes the phenomenon. They then propo...	Rebuttal 1: Rebuttal: We thank the reviewer for the thoughtful review -- we are delighted by the positive assessment of our presentation, complexity reparametrization and effective parameter measure! We hope to resolve doubts about any weaknesses in our point-by-point response below. **Inclusion of boosting and trees....	Summary: The paper analyzes the seeming appearance of double descent in non-deep models, and finds various ways in which it disappears when thinking more carefully about how the parameters are scaled. Trees and boosted models see double descent because parameters start increasing along a different and more effective di...	Rebuttal 1: Rebuttal: We thank the reviewer for the in-depth and constructive review! We were delighted by the assessment that our paper is "interesting", "well-presented " and "likely to encourage follow-on work in the field", and, limited by space constraints, hope to resolve doubts about any weaknesses in our respon...	null	null	null	null
Evaluating the Moral Beliefs Encoded in LLMs	Accept (spotlight)	Summary: The paper studies the encoded moral norms of various language models. To this end, the authors first describe different measurements: semantic likelihood, consistency, and uncertainty. To construct the dataset to analyze LMs, the authors chose to use OpenAI’s text-davinci-003 and GPT-4 to create the content of...	Rebuttal 1: Rebuttal: We thank the reviewer for their thoughtful feedback and questions. We have actioned minor comments on typos and references and address the main concerns below. **(1) Missing details on how the introduced dataset is annotated. What is Surge AI? How is the number of annotators justified? What is th...	Summary: The paper studied the moral beliefs encoded in large language models. For this purpose, the authors designed a systematic survey-based framework consisting of three novel metrics, i.e., semantic likelihood, consistency and uncertainty, and the MoralChoice dataset with 680 ambiguous moral scenarios and 685 less...	Rebuttal 1: Rebuttal: We thank the reviewer for the thoughtful questions and constructive feedback. We have incorporated the references into the revised draft and will address the concerns point by point (1) How can we guarantee that these survey questions really reveal the intrinsic beliefs of LLMs? Please see `...	Summary: This paper presents an broad investigation of the morality of LLMs. Authors craft the MoralChoice dataset, consisting of ~1300 moral scenarios with 50:50 high and low ambiguity, and probe 24 different LLMs (commercial and open-source). To assess the moral decisions of LLMs, authors create two measures for *sem...	Rebuttal 1: Rebuttal: Thank you for your support! Your thoughtful questions and constructive feedback have greatly improved the paper. (1) How a model's moral leaning in these scenarios actually influences downstream tasks or user interactions, and thus how this could harm users. Excellent Point! We agree and h...	Summary: They assess moral beliefs in LLMs by analyzing a mix of commonsense reasoning and behavior of the model in scenarios without clear answers. They treat the LLMs themselves as "survey respondents" rather than humans in prior empirical survey studies but apply the same methodologies. In their inquiry they attemp...	Rebuttal 1: Rebuttal: Thank you for the thoughtful questions, constructive feedback, and your support for our work! We really appreciate it! We address the questions point by point. **I don't know if the claim that "people use LLMs to get advice about difficult moral situations" is true. Just because LLMs might encode...	Rebuttal 1: Rebuttal: We thank all reviewers for their thoughtful feedback. Your reviews have greatly improved the paper. We really appreciate it! We are happy to hear that all reviewers found our work well-written and addressing a timely and important topic. We are grateful for the appreciation of our analysis (`t4xk...	NeurIPS_2023_submissions_huggingface	2,023	Summary: The contributions of this work can be summarized as such: - New questionnaire dataset to evaluate moral beliefs encoded in LLMs, called MoralChoice, which is **a by-product of the survey designed for analysis, rather than a benchmark. The authors call this as conducting a survey on LLMs, but I find it confusi...	Rebuttal 1: Rebuttal: We thank the reviewer for the insightful questions and comments. We appreciate your support! **While the experiments are interesting enough, it occurs to me that these are based on simple prompting of LLMs and we cannot be sure that these so-called moral beliefs actually transfer to safe or ha...	Summary: This work aims at creating a dataset containing moral reasoning scenarios and surveyed a wide range of language models on their value and consistency shown on the dataset. Specifically, the introduced MoralChoice dataset includes low and high ambiguity scenarios. The investigated language models range across o...	Rebuttal 1: Rebuttal: Thank you for the thoughtful feedback and questions. Your comments have greatly improved the paper. We address the main concerns point-by-point. (1) More extensive human investigation into the created dataset in the first place may be necessary. Whose moral values does the dataset represent?...	null	null	null	null
Information Design in Multi-Agent Reinforcement Learning	Accept (poster)	Summary: This paper studied a Markov signaling game and proposed to use the signalling gradient and extended obedience constraint to build a learning system that can effectively train the sender in the game to reach desirable equilibriums. Experiments performed on two toy problems suggest that the newly developed SGOC ...	Rebuttal 1: Rebuttal: We thank the reviewer for their insightful review. We now provide our responses. We are glad that the reviewer is interested in our work and asked many questions. Despite this, there are factual misunderstandings in the review which might have affected the evaluation of our work. We wish that wi...	Summary: This paper proposes a new agent-interaction framework called Markov signaling games, and the corresponding algorithm for mixed-motive tasks. Various experiments and discussions are provided to show the efficacy of the proposed method. Strengths: - The authors provide a new formulation of agent-interaction fra...	Rebuttal 1: Rebuttal: We thank the reviewer for their insightful review. We now provide our responses. We wanted to point out that our work is a learning-based extension of information design in sequential decision-making. It focuses on mixed-motive settings which are less addressed in existing MARL communication works...	Summary: This paper addresses the problem of multi-agent reinforcement learning in environments with mixed motives, which lie on the spectrum from fully cooperative to fully competitive scenarios. Specifically, it studies the problem of information design in this setting: how a sender might craft messages so that the ...	Rebuttal 1: Rebuttal: We thank the reviewer for their insightful review. We now provide our responses. ### QS1: Is it reasonable for the sender to access the receiver's actor and observation during training? What is the practical use of this model? Yes, it is reasonable for the sender to access the receiver's actor a...	Summary: The article tackles the scenario where two agents, a sender $i$ and a receiver $j$, act in an environment. The sender $i$ can send a message to the receiver $j$ at each time step. The sender's goal is to learn what information to send to the receiver such that it influences the latter's policy to maximize its ...	Rebuttal 1: Rebuttal: We thank the reviewer for their insightful review. We now provide our responses. The suggestions by the reviewer have reminded us that we should provide more detailed explanations for some fundamental concepts, which could also be confusing for other reviewers. We have improved the explanations o...	Rebuttal 1: Rebuttal: We notice some common concerns expressed by the reviewers. One is on the sender's access to the receiver, and the other is on MSG for mixed-motive MARL communication. These "concerns" are not issues and we are happy to explain why. ### Sender and receiver are in interest conflict. Can the sender ...	NeurIPS_2023_submissions_huggingface	2,023	Summary: This paper studies how multi-agent reinforcement learning agents influence each other by hsaring different kinds of information. Additionally, the authors study how different information affects the transition of the agents trajectories and what information can be ignored. To ensure efficient learning, a suita...	Rebuttal 1: Rebuttal: We thank the reviewer for their insightful review. We now provide our responses. ### Emergent communication This reviewer shows particular interest in understanding how information is conveyed semantically, especially in the context of emergent communication and the emergence of language in ...	null	null	null	null	null	null
Efficient Policy Adaptation with Contrastive Prompt Ensemble for Embodied Agents	Accept (poster)	Summary: This paper proposes a novel framework with an ensemble of visual contrastive prompts for efficient adaptation of vision-based robot learning tasks, named CONPE. The visual prompts are firstly learned from a contrastive learning procedure with different domain factors, which are expected to be domain orthogonal...	Rebuttal 1: Rebuttal: ### Author response for Weakness 1. [physical property] We appreciate the reviewer for this observation. The physical properties that we focus on include camera fields of view, camera positions and action magnitudes which are frequently considered embodiment configurations in the embodied AI l...	Summary: This paper proposes a novel contrastive prompt ensemble (ConPE) framework for embodied reinforcement learning (RL) agents. ConPE utilizes a pretrained vision-language model and a set of visual prompts to construct robust state representations. Each prompt is contrastively learned in terms of an individual doma...	Rebuttal 1: Rebuttal: ### Author response for Weakness 1. [more challenging and realistic tasks] The simulators that we used (e.g., AI2THOR, CARLA) are widely used in the field of Embodied AI [1, 2, 3, 4, 5]. The observation state from these environments contains a wide variety of objects, and their image quality ...	Summary: The authors propose an approach for rapid policy adaptation to unseen visual observations for embodied agents. Their approach called CONPE (contrastive prompt ensemble) leverages contrastive learning to learn an ensemble of prompts for policy visual encoders that enables the policy to ultimately adjust to the ...	Rebuttal 1: Rebuttal: ### Author response for Weakness 1. and Question 1. [prompt and zero-shot] In Line 122 in Section 3.2, the term "prompt" is defined as a learnable vector appended to the input, with the intent of extracting domain-invariant knowledge from pre-trained visual encoders. This definition of prompt ...	Summary: The authors propose a method to address zero shot adaptation to novel visual domains. They do so by: i) Learning a set of visual prompts via contrastive learning that encode specific domain invariant knowledge. ii) Using a prompt ensemble based on embedding similarity scores (attention) Strengths: * The autho...	Rebuttal 1: Rebuttal: ### Author response for Weakness 1. In Figure 5(a)-1) of our manuscript, we visualize intra prompted embeddings from two distinct domains, both defined by different configurations of the same domain factor, to evaluate their alignment. The figure specifies that the prompted embeddings from the...	Rebuttal 1: Rebuttal: We attached global author rebuttal. Please refer to it. Pdf: /pdf/d81fdc9100607f9ff976675d113a6aeb014186f3.pdf	NeurIPS_2023_submissions_huggingface	2,023	Summary: This paper presents a method for improving zero-shot domain adaptation for Embodied AI agents by learning a task-specific state representation. Specifically, $n$ visual prompts are learned by fine-tuning the agent's visual encoder (i.e. CLIP) contrastively on observations. Each visual prompt is specific to a ...	Rebuttal 1: Rebuttal: ### Author response for Weakness 1. \|Method\|0.5 M\|0.6 M\|0.7 M\|0.8 M\|0.9 M\|1.0 M\| \|-\|-\|-\|-\|-\|-\|-\| \|EmbCLIP \|$54.38$\|$60.35$\|$60.00$\|$66.00$\|$66.38$\|$70.72$\| \|ATC \|$56.61$\|$54.30$\|$54.12$\|$52.54$\|$61.61$\|$60.45$\| \|ConPE \|$64.52$\|$68.16$\|$69.91$\|$71.58$\|$71.50$\|$75.53$\| We appreci...	Summary: This paper presents a novel framework for “efficient policy adaptation”, i.e. adapting learned policies to changes in the domain (e.g., lighting, camera height etc.), using a very small amount of in-domain data (few-shot and zero-shot). The proposed approach comprises learning “visual prompts” for modeling eac...	Rebuttal 1: Rebuttal: ### Weakness 1-1. In Sec. 3.2, when using the phrase "fine-tune $\mathcal{T}_{\phi}$", it denotes the tuning process of visual prompts, a process we term visual prompt-tuning. This does not encompass parameter updating $\phi$ in CLIP visual encoder $\mathcal{T}$. Via the visual prompt-tuning, ...	null	null	null	null
Sheaf Hypergraph Networks	Accept (poster)	Summary: \ The primary research focus of this paper is to investigate the potential of neural networks for processing hypergraph datasets. The paper utilises the mathematical concept of sheaf, which describes how locally defined data on a space can be consistently pieced together, to enhance hypergraphs by associating...	Rebuttal 1: Rebuttal: We thank the reviewer for the very detailed review and constructive feedback. We are delighted the reviewers agree that we propose a sound method, with a pivotal role for the hypergraph neural networks progress. In the following, we discuss the main remarks raised by the reviewer, and we will in...	Summary: This paper generalizes hyper-graph Laplacians to hyper-graphs attached with sheaves, called Hypergraph Sheaf Laplacian, and constructued hyper-graph NNs with proposed Hypergraph Sheaf Laplacian. Experiments demonstrate the effectiveness of SheafHyperGNNs over normal HyperGNNs. Strengths: I find this paper wel...	Rebuttal 1: Rebuttal: We are grateful that the reviewer appreciates our work and believes that the novel direction is promising and beneficial to the community. We thank the reviewer for the constructive feedback and for pointing out to us a relevant related domain. In the following, we discuss in more detail each po...	Summary: The paper proposes cellular sheaf to enhance the high-order relationships among hypergraphs. Specifically, it first designs two formulations, linear sheaf hypergraph Laplacian and non-linear sheaf hypergraph Laplacian to construct the hypergraph structure. Then the two sheaf hypergraph Laplacians are leveraged...	Rebuttal 1: Rebuttal: Thank you for your thorough review and constructive feedback on our paper. We appreciate the time you've taken to review our work, and we would like to address your concerns and questions. Motivation and challenges The motivation for our work comes from recognizing that current hypergraph mo...	Summary: The paper introduces the cellular sheaves to hypergraph and proposes two sheaf hypergraph laplacians. In addition, they also propose two models: sheaf HNN and Sheaf HCN. The experiments demonstrated the effectiveness of the proposed models. Strengths: - The paper is well-written and well-organized. - The aut...	Rebuttal 1: Rebuttal: We warmly appreciate your time and effort invested in reviewing our paper. Your perceptive comments and constructive criticism are invaluable, and we are grateful. We are sorry for the incorrect reference and we will make sure it will be fixed in the final version of the paper. **Regarding the no...	null	NeurIPS_2023_submissions_huggingface	2,023	null	null	null	null	null	null	null	null
Label-Only Model Inversion Attacks via Knowledge Transfer	Accept (poster)	Summary: The paper proposes a novel label-only model inversion attack (MIA). The underlying idea is to train a surrogate model and then perform existing white-box MIAs to reconstruct characteristic samples from the target model's training data. Instead of using the target model to label a public dataset, the paper prop...	Rebuttal 1: Rebuttal: >It would, therefore, be interesting to see how the attacks perform on images of higher resolution and other, more recent architectures, We appreciate the Reviewer’s suggestion. We report additional experiment results in 2 parts: (A) The MI attack results using recent architecture ConvNeXt [Liu ...	Summary: This paper proposes a label-only model inversion (MI) attack. It focuses on a more challenging scenario where the target network only outputs the label. It trains a class-conditional GAN based on some public data and the label information provided by the target network. A surrogate network can be trained at th...	Rebuttal 1: Rebuttal: > What are the target labels? Are there overlapped identities/labels in the private and public datasets? Pls see Ln 223: “To ensure a fair comparison, when evaluating our method, we use the exact same experimental setup as BREPMI [6].” Following [6], each dataset is divided into two distinct/...	Summary: The paper focuses on the setting where the adversary has only label access to the target model. The authors suggested to use transfer learning and generative models to improve the model inversion attacks. In particular the authors proposed T-ACGAN which is a variation of ACGAN that leverages the access to the ...	Rebuttal 1: Rebuttal: > In Fig 1.c The authors show that the adversary does not have access to the hard labels ? is that correct? Thank you for pointing out this to us. It is a typo, the adversary has access to the hard labels. $ $ > Looking at the membership inference attack literature, the better way of measu...	Summary: This paper proposes a two-step label-only model inversion attack using knowledge transfer. In the first step, a surrogate model is trained via knowledge transfer. In the second step, a white-box model inversion attack is performed on the surrogate model. The authors explore several approaches to train the surr...	Rebuttal 1: Rebuttal: > It will be great if the authors can elaborate more on the difference between T-ACGAN and GAMIN [10]. The authors briefly describe GAMIN as an "an adversarial approach for black-box MI". However, the current proposal shares several similarities with GAMIN, namely the use of GANs to learn a surrog...	Rebuttal 1: Rebuttal: We thank all the reviewers for their valuable time and effort to review our work. We appreciate the Reviewers' kind comments and recognition. We would like to highlight that our paper focuses on Model Inversion, not Membership Inference as one of the reviewers suggested. Therefore, **comments of ...	NeurIPS_2023_submissions_huggingface	2,023	null	null	null	null	null	null	null	null
Blocked Collaborative Bandits: Online Collaborative Filtering with Per-Item Budget Constraints	Accept (poster)	Summary: This paper considers the online recommendation for a hard-clustered user. Problem setup summary: This problem involves $M$ users and $N$ items. the reward matrix $P \in R^{M \times N}$ is unknown. There are $T$ rounds in total, and each user receives a recommendation $\rho_u(t)$ at each round. Each user belo...	Rebuttal 1: Rebuttal: We thank the reviewer for making some excellent points. Below, we discuss them in details: *1) Somewhat artificial budget constraints 2) Many parts of the algorithms are in the appendix (mainly, Alg 4,5) 3) Alg depends on B, C that are hard-to-know before (in particular B).** 1) We apologize...	Summary: In this paper, the authors tackle a collaborative-filtering type multi-armed bandit problem. There are many users, each facing a multi-armed bandit problem. However, there are only $C$ base multi-armed bandit instances, where each user comes from one of these instances. The authors provide a novel algorithm wh...	Rebuttal 1: Rebuttal: *Knowledge of C* Please note that C is a small positive constant integer that does not scale with M,N,T (users,items,rounds). In fact, it suffices if we know a loose upper bound for C. Also, if C is unknown, then at the beginning, we can use a small number of rounds to estimate C. For instan...	Summary: The paper considers the problem of blocked collaborative bandits with multiple users that each belong to a cluster with associated bandit instance. The goal is to minimize the regret of each user while satisfying the constraint that each arm can be played at most B times for each user. The paper proposes an al...	Rebuttal 1: Rebuttal: We thank the reviewer for making some excellent points. Below, we discuss them in details: *As the papers main contribution is providing regret bounds with the existence of the budget constraint B, I believe the authors should provide the dependency of the regret bound on B.* The regret scal...	Summary: This paper studies the blocked collaborative bandits where each user represent a single bandit problem. The latent groups are assummed to exist among users in which all the users in a group share the same expected reward. But the constrant is placed on the number of serving time for each user to solve the chal...	Rebuttal 1: Rebuttal: We thank the reviewer for asking some great questions. Below, we discuss them in details: *time cost of this algorithm* Computationally, the main bottleneck of our algorithm is the matrix completion function Estimate invoked in Line 13 of the Explore Component (Algorithm 2). All the remain...	Rebuttal 1: Rebuttal: Experiments on Movielens-10M dataset Reviewer tct3 requested for additional experiments on real-world datasets - in response, we show a set of experiments on the popular MovieLens dataset. Relevant figures can be found in the attached pdf. We take the Movielens 10M dataset (https://grouplens...	NeurIPS_2023_submissions_huggingface	2,023	Summary: This paper studies blocked collaborative bandits. Multiple users, clustered in a fixed number of clusters according to the rewards obtained from them are provided recommendations in each discrete time instance till T timesteps. The paper proposes algorithms that minimize the cumulative regret with the constrai...	Rebuttal 1: Rebuttal: We thank the reviewer for making some excellent points. Below, we discuss them in details: ***A crisp summary of how the problem is MAB problem is missing from the introduction/setting section. For instance, the following points are not clear in the first reading of the first 2 sections; what con...	null	null	null	null	null	null
Reverse Engineering Self-Supervised Learning	Accept (poster)	Summary: This work presents a study of the structure of SSL representations with respect to sample labels (image identity), to class labels and to super-class labels. To that end, an analysis of the clustering properties of the embedding space according to such labels is performed. Additionally, probing linear layers a...	Rebuttal 1: Rebuttal: Dear Reviewer b2bW, Thank you very much for your thorough review and feedback on our paper. We value your constructive comments and would like to address the weaknesses and questions you raised: # Weaknesses: ## Originality: We appreciate the criticism that has been raised while also maintain...	Summary: This paper provides an empirical study on self-supervised learning (SSL) in terms of clustering. Specifically, the authors demonstrate 1) clustering process of SSL methods, 2) semantic learning of SSL methods based on random labels generated by various models, and 3) class hierarchies and intermediate layers o...	Rebuttal 1: Rebuttal: Dear Reviewer MBnG, Thank you very much for your thorough review and positive feedback on our paper. We are pleased you found our work well-written, clear, and consistent. Your encouraging words on our study's presentation, soundness, and contribution are greatly appreciated. We value your const...	Summary: This paper focuses on an empirical analysis of how self-supervised learning (SSL) works. While training artificial neural networks for classification under supervision has been extensively explored, the authors propose to shed more light on SSL process using VicReg + ResNet-L-H and a series of experiments on t...	Rebuttal 1: Rebuttal: We thank the reviewer for the valuable feedback. We've taken your insights into account and have expanded our analyses and clarifications as detailed below. # Weaknesses ## W1: While this work is an initial study into several geometrical phenomena in SSL, we've attempted to incorporate many networ...	Summary: To reveal the underlying mechanism of self-supervised learning methods, this paper mainly investigate their clustering characteristics. Empirical analyses show that these methods cluster samples based on their semantic classes. Additionally, the use of regularization terms, introduced to prevent representation...	Rebuttal 1: Rebuttal: We thank the reviewer for the valuable feedback. We've taken your insights into account and have expanded our analyses and clarifications as detailed below. # Weaknesses: ## W1: While we agree that deriving a novel SSL method based on these findings could be a compelling avenue for future explo...	Rebuttal 1: Rebuttal: Dear Reviewers, Thank you for your thoughtful and detailed reviews of our work. We appreciate your time and the constructive feedback you have provided. We have carefully considered your comments and concerns and would like to address them in a unified response: # 1. **Diversity of Architecture...	NeurIPS_2023_submissions_huggingface	2,023	null	null	null	null	null	null	null	null
On the Complexity of Differentially Private Best-Arm Identification with Fixed Confidence	Accept (poster)	Summary: The paper considers the best arm identification (BAI) problem under differential privacy (DP) constraints, where the goal is to privately identify the arm with the highest expected reward with probability at least $1 - \delta$. The paper proposes a DP variant of the top two algorithm, which achieves near-optim...	Rebuttal 1: Rebuttal: We thank the reviewer for the careful reading and astute observations to improve the manuscript. Technical novelty. Please refer to the general comment. Non-asymptotic sample complexity of AdapTT. Please refer to the general comment. Minor Typos. Thanks for detecting the typos. We w...	Summary: This work provides some insight into using the Laplace mechanism on the algorithms concerning the best arm identification problem bandits with a constraint of fixed confidence. The analysis carried out in this work enumerates general properties of differentially private FC-BAI and then proposes an algorithm ba...	Rebuttal 1: Rebuttal: We thank the reviewer for the time spent reviewing and the positive view of this work. We hope that the rebuttals done for the other reviewers will answer your future questions. Also, let us know if you have any other concerns.	Summary: This paper studies the problem of Best Arm Identification (BAI) for bandit problems under the constraints of differential privacy. There are some $K$ distributions and the solver interacts with the arms over a sequence of rounds. In each round they choose a distribution and receive a reward according to that d...	Rebuttal 1: Rebuttal: We would like to thank you for your time, thorough feedback, and precise remarks. We first answer the points raised in the "weaknesses" section. 1. The paper is hard to read for non-experts. The paper studies private BAI and builds upon recent advances in both domains. As such, for non-exper...	Summary: This work provides novel theoretical guarantees to quantify the cost of privacy in DP-FC-BAI (and $\epsilon$ global DP). Specifically, the authors show that the complexity depends on an information theoretic total variation metric. Although, this holds in the high privacy regime, this is reasonable and possibl...	Rebuttal 1: Rebuttal: We would like to thank the reviewer for the time spent reviewing, careful reading, and kind words about the contributions' novelty and significance. It is really encouraging for us. Missing references. We thank the reviewer for pointing us to these references. We will add them to the Related ...	Rebuttal 1: Rebuttal: We would like to thank the reviewers for acknowledging the strengths and soundness of the contribution as well as for their thoughtful comments and efforts towards improving the manuscript. In the following, we address the general concerns of reviewers that were common. As pointed out in the rev...	NeurIPS_2023_submissions_huggingface	2,023	Summary: This paper studies the problem of BAI with fixed confidence under ϵ-global Differential Privacy (DP). The authors derive a lower bound on the sample complexity of any δ-correct BAI algorithm satisfying ϵ-global DP. In addition, the authors also design an ϵ-global DP variant of the Top Two algorithm, named AdaP...	Rebuttal 1: Rebuttal: We thank the reviewer for the constructive feedback and kind words concerning the importance of the derived lower bound. 1. Dense paper and difficult to follow. The paper investigates private BAI and draws on very recent developments in both fields (like AKR21, AB22, JDB+22, JD22 etc.). As a ...	Summary: This paper studies the problem of best arm identification with fixed confidence under ϵ-global Differential Privacy. First, they provide a lower bound of the BAI-DP problem showing that two different regimes depend on the privacy budget of $\epsilon$. They propose an ϵ-global DP variant of the Top Two algorith...	Rebuttal 1: Rebuttal: We would like to thank the reviewer for the constructive feedback and the interesting questions. Real application motivating privacy at the level of intermediate actions and the single-agent setting. We respectfully disagree with the assessment of the reviewer regarding the practical groundin...	null	null	null	null
A representation-learning game for classes of prediction tasks	Reject	Summary: This paper extends and formulates learning representation procedure as a three-player game when the downstream task is known. Unlike conventional unsupervised learning algorithms, the proposed framework helps the learning process in the feature space be directed. They set up a general and basic mathematic form...	Rebuttal 1: Rebuttal: # Response to the strengths To clarify Figure 1 for Example 5: It shows that mixed representation is more important whenever $\{s_{i}\sigma_{i}^{2}\}$ are more spread ($\alpha$ is closer to $0$). Both regrets are decreasing as a function of $\alpha$, as can also be seen from the analytical expres...	Summary: This work studies the problem of dimensionality reduction given prior information about downstream tasks via game theoretic optimization point of view. They define a three player game. One player chooses the parameters of the representation function that reduces the dimensionality of the data. One player choo...	Rebuttal 1: Rebuttal: # Motivation While the concept of "representation learning" is rich, the main motivation of our paper is unsupervised dimensionality reduction (compression), in which the representation is restricted in its dimensionality. Our approach is mainly adequate whenever the compressed representation can...	Summary: This paper studies the important problem of representation learning, where the goal is to learn a representation that works well for all problems in a known problem class. It presents theoretical results for the special case where both the representation function and prediction function class are linear, and a...	Rebuttal 1: Rebuttal: # Answers to questions 1. The viability of mixed representations is for the _unsupervised_ learning problem we consider in the paper, whose goal is dimensionality reduction (our setting is indeed an unsupervised problem because only unlabeled features are provided to the learner, even though the r...	Summary: This paper studies the representation learning from the multi-layer gaming perceptive. The authors propose an min-max optimization problem for learning a good representation with the prior knowledge of the latter prediction tasks. The authors provide some theoretical analyses and examples in the linear case. F...	Rebuttal 1: Rebuttal: # Response to the weaknesses (1) We believe that proving the closed-form solution for the min-max problem in the linear-MSE caseis a major contribution of the paper since the proof is neither trivial or simple: It spans 9 pages and requires a variety of techniques. We also mention that this solut...	Rebuttal 1: Rebuttal: We responded to all the comments made by the 5 reviewers, as summarized below: * __Reviewer fmep (score 6):__ The reviewer questioned the notion of a three-player game compared to a two-player game. We explain why this is just a semantic issue, and what actually matters is the order the players t...	NeurIPS_2023_submissions_huggingface	2,023	Summary: This paper presents a game theoretic framework for learning representations in a general setup where some knowledge/prior about the prediction task is available. The representation learning task is cast as a certain game --- which is worked out exactly in the fully linear setting under the squared loss (castin...	Rebuttal 1: Rebuttal: # Response to the weakness We highly appreciate this observation, and will add it to the conclusion of the revised paper (with a proper acknowledgment). To re-iterate this idea, our paper links between optimal unsupervised pre-training, to nonconvex/nonconcave min-max problems. This is an importa...	null	null	null	null	null	null
On the Convergence to a Global Solution of Shuffling-Type Gradient Algorithms	Accept (poster)	Summary: This paper provides the convergence to a global solution of `shuffling' SGD (where the index set is permuted and apply gradient updates according to the permutation order) for a class of non-convex functions, in the over-parameterized setting. Strengths: The paper provides a good comprehensive background of p...	Rebuttal 1: Rebuttal: **Thank you for your constructive comments on our paper. We hope that our responses below answer all of your questions. Should you need any other clarification, please leave a comment and we are happy to discuss with you. While we are not able to upload a revision for the rebuttal, we will add all...	Summary: In this paper, the authors demonstrate the convergence of the shuffling version of Stochastic Gradient Descent (SGD) to a global solution for a class of non-convex functions. Notably, the analysis of the algorithm relies on more relaxed non-convex assumptions compared to previous works, while maintaining the s...	Rebuttal 1: Rebuttal: **Thank you so much for your time and for your positive evaluations of our paper. We also thank you for your appreciation of our novel complexity bound. We hope that our responses below answer all of your questions. Should you need any other clarification, please leave a comment and we are happy t...	Summary: Shuffling-type SGD is widely used in practice, for which the prior works only provide gradient norm guarantees in nonconvex settings. This paper studies the convergence to a global solution for shuffling-type gradient algorithms under nonconvex settings, under a new analytical framework of more relaxed nonconv...	Rebuttal 1: Rebuttal: **Thank you so much for your time and your appreciation that the contribution and soundness of our paper is good. We hope that our responses below answer all of your questions. Should you need any other clarification, please leave a comment and we are happy to discuss with you. While we are not ab...	Summary: The paper analyzes convergence properties of shuffling-type methods (shuffle SGD and incremental gradient) to the global solution under relaxed assumptions for non-convex overparameterized functions. Particularly, the paper relaxes frequently used PL-condition to the averaged PL of local functions $f_i$, which...	Rebuttal 1: Rebuttal: **Thank you so much for your time and your positive evaluations of our paper. We also thank you for your constructive comments and suggestions. We hope that our responses below answer all of your questions. Should you need any other clarification, please leave a comment and we are happy to discuss...	Rebuttal 1: Rebuttal: General Response: First of all, we would like to thank the AC and all other reviewers for your hard work and for reviewing our paper. To all reviewers: Below we first highlight our contributions for you to have a better overview of our paper, after which we clarify current misunder...	NeurIPS_2023_submissions_huggingface	2,023	null	null	null	null	null	null	null	null
Distributionally Robust Skeleton Learning of Discrete Bayesian Networks	Accept (spotlight)	Summary: The authors give novel algorithms to learn the structure of a Bayes net from noisy data. Strengths: Originality: Work seems original. Quality: Quality is high, writing-wise and content-wise; see questions. Clarity: Writing is clear; see questions. Significance: Topic is significant. Weaknesses: No signifi...	Rebuttal 1: Rebuttal: Thanks for your positive comments. Our rebuttals can be found below. Q1: Line 21: What do you mean by "ordinal?" If there is an ordinal relationship among variables, the relation is important to learning and inference. For example, if we are to rate the performance of an employee out of very...	Summary: The authors propose methods for learning Bayesian network skeletons based on discrete data when the data is corrupted by arbitrary noise. The authors show guarantees on discovering the correct skeleton in terms of the required sample size. Some empirical experiments are also presented. Strengths: + rigorous ...	Rebuttal 1: Rebuttal: Thanks for your very nice and helpful comments. **Q1: In Sec. 4, explain what exactly you mean my the F1 score: the F1 score is used for binary classification, so I'm guessing you're perhaps looking at presence/absence of each of the edges in the DAG. Please clarify further which methods are prop...	Summary: In this paper, the authors revisit the problem of structure learning of Bayesian networks. Here, we are interested to learn the underlying DAG structure that encodes the conditional independence properties of a Bayes net from random samples. A number of algorithms have been proposed to solve this problem over ...	Rebuttal 1: Rebuttal: Thanks for your insightful comments. Q1: A major weakness of the paper is that a number of assumptions are made in order to derive the results, which looks ad-hoc and too restrictive in my opinion. See page 4 of the paper for these assumptions. See also line number 271. Thanks for the good q...	Summary: The paper presents a method for learning the structure of discrete Bayesian networks from potentially corrupted data. It utilizes distributionally robust optimization and regression, optimizing the worst-case risk over a set of distributions. The approach works with general categorical variables without assumi...	Rebuttal 1: Rebuttal: Thanks for your positive comments. Please see our point-by-point responses as follows. **Q1: The author's assertion that they had limited space to relate their work to the existing literature is understandable. However, it is crucial to provide a clear positioning of their own work within the bro...	Rebuttal 1: Rebuttal: Dear reviewers, Thanks for your helpful reviews that would certainly improve our manuscript. Please see our rebuttals after individual reviews. We are more than glad to have further discussions or make clarifications in the discussion period. We would like to restate that our main contribution i...	NeurIPS_2023_submissions_huggingface	2,023	Summary: The paper considers the problem of learning the exact skeleton of general discrete Bayesian networks from potentially corrupted data. The estimator optimizes the worst case risk over a family of distributions within bounded Wasserstein distance and KL divergence to the empirical distribution. Under mild condit...	Rebuttal 1: Rebuttal: Thanks for the positive feedback. Please find our rebuttals below. Q1: Is it possible to extend the DRO framework to learn the structure of continuous Bayesian networks from corrupted samples? The proposed DRO framework is based on a regression approach, which is naturally tailored to tackli...	null	null	null	null	null	null
IPMix: Label-Preserving Data Augmentation Method for Training Robust Classifiers	Accept (poster)	Summary: This paper proposes IPMix, a data augmentation strategy to improve model robustness (and other safety metrics like calibration). The idea builds upon PixMix, a recent data augmentation strategy generating diversity via synthetic pictures. IPMix performs this at multiple scales (pixel, patch, image levels) and ...	Rebuttal 1: Rebuttal: Thank you for your detailed concerns. W1: First and foremost, I'd like to emphasize that the improvements we achieved are not minor. Our approach consistently shows improvement across various safety metrics by comparison with other methods, which is more meaningful from a practical perspective. ...	Summary: This work proposes a data augmentation method, named IPMix, which is designed to improve robustness without hurting clean accuracy. Specifically, IPMix utilizes different levels of data augmentation (image-level, patch-level, and pixel-level) to increase the diversity of training data and introduces structural...	Rebuttal 1: Rebuttal: Thank you for your valuable time in reviewing our paper. W1: Thank you for your valuable suggestion. We agree that RandAugment is a popular image-level method and believe that comparing with it can strengthen our work. While space constraints in our main paper limited a direct comparison with R...	Summary: The authors propose IPMix -- a new data augmentation technique applied during training that improves model robustness across a variety of test-time tasks (e.g., distributional shift, adversarial data augmentation) without compromising accuracy on in distribution tasks (when compared to a model trained in a tra...	Rebuttal 1: Rebuttal: Thank you for your valuable time in reviewing our paper. W1-1: "The motivation of using synthetic images. " Due to the difficulty in acquiring labeled data, as well as human bias and data privacy issues, more and more research focuses on obtaining synthetic data that can be generated without lab...	Summary: This paper proposes a new data augmentation method, IPMix, for vision models that integrates three levels of augmentation techniques (image level, patch level, and pixel level). The approach is label preserving and enforces extra structural complexity to enhance the robustness and other safety measures. Experi...	Rebuttal 1: Rebuttal: Thank you for your valuable time and detailed reviews. Below we will address your concerns about novelty and what problem we solved. W1-1: "novelty and technical contributions." * Diversity & Complexity: While data diversity and structural complexity are keys to model robustness, most previ...	Rebuttal 1: Rebuttal: First, we thank all reviewers for your help and support. Due to the limitation of input characters on OpenReview, we have to put some questions that the reviewers are generally concerned about here and answer them together. 1. Experiments on Transformer: Given the limited time and computatio...	NeurIPS_2023_submissions_huggingface	2,023	Summary: This paper presents IPMix, a new data augmentation strategy that enhances robustness without compromising clean accuracy. The proposed approach diversifies the training data with limited computational overhead. IPMix introduces structural complexity at different levels to generate a wider variety of images and...	Rebuttal 1: Rebuttal: Thanks for your detailed reviews and insightful comments! W1: Your suggestion is incorporated in the new version of the paper. Thank you. W2-1: "which component is vital for a specific safety measure." Thank you for your concerns. We will detailed analyze the impact of each part on different s...	null	null	null	null	null	null
Human spatiotemporal pattern learning as probabilistic program synthesis	Accept (poster)	Summary: The paper presents a "Language of Thought" (LoT) model which is meant to represent how humans learn concepts from small amounts of data. To test whether the LoT model is a suitable model for human cognition the authors present a task of 2-dimensional pattern learning and collect data from human participants wh...	Rebuttal 1: Rebuttal: Weaknesses - As to the suitability of the task, we have expanded our introduction and discussion to more clearly describe the general context discussed in our meta-response, as well as the work of Amalric et al. (2017) who looked at how children learn simple patterns on an octagon, and that of...	Summary: This work investigates a visual sequence prediction task in humans where a dot would move on a screen based on some kind of underlying function and the participant must predict where the dot will go next. The goal of the study is to understand people's underlying inductive biases, so only people's first attemp...	Rebuttal 1: Rebuttal: Weaknesses - With regard to your comment about the sequences we chose, have added a paragraph to the text to better explain how we chose the stimuli. Though the exact stimuli we used were somewhat ad hoc, we chose them carefully: we wanted them to be learnable by people within 20 timepoints and...	Summary: The paper investigates whether humans learn visual patterns by programming, i.e., inducing probabilistic models to fit observed data. In a synthesized visual prediction task, they evaluate human performance against standard parametric and non-parametric time-series models and Bayesian program synthesis models...	Rebuttal 1: Rebuttal: Strengths - The paper conducts a very interesting investigation and provides valuable insights on human pattern learning. - Their experiment method is interesting and appropriate and carefully carried out. - The paper is well written and I enjoy reading it. Thanks! Questions *"The ...	Summary: The paper tests human online prediction and learning in a moving dot task against a variety of probabilistic models. The paper primarily reports its behavioral results from human subjects in Section 2, then describes probabilistic models based on ridge regression, compositional and non-compositional Gaussian p...	Rebuttal 1: Rebuttal: Weaknesses: “Only one 'language of thought' hypothesis is given, and its design is ad-hoc and lacking in clear neural or motoric motivation, despite the paper attempting to take inspiration from computational modeling of drawing motor skills.” This work primarily aims to characterize people...	Rebuttal 1: Rebuttal: We would like to thank the reviewers for their thoughtful comments. We appreciate that they generally found our task and modeling work interesting and methodologically sound, even if there were some areas they thought could be improved. We will first try to summarize areas the reviewers agreed upo...	NeurIPS_2023_submissions_huggingface	2,023	Summary: In this paper, the authors study human function learning as an algorithmically structured process, comparing compositional models to behavior. Their experimental protocol is a spatial variant of function learning from the psychological literature, and requires participants to learn a sequence or function by it...	Rebuttal 1: Rebuttal: Weaknesses - Regarding contextualization and background, we agree that the current paper will be improved by more clearly contextualizing its contributions and the ways it differs from prior work. Though limited by space constraints, we have restructured our introduction and discussion to more e...	null	null	null	null	null	null
Understanding Diffusion Objectives as the ELBO with Simple Data Augmentation	Accept (oral)	Summary: The paper studies different weighting schemes of the loss functions of generative diffusion models. In particular, the authors investigate the relationship between a weighted (with respect to noise levels) loss and the proper Evidence Lower Bound (ELBO), discussing how the former can be written as a weighted i...	Rebuttal 1: Rebuttal: Thank you for your detailed review and valuable comments. Please note our top-level comment with additional experimental and theoretical results. Below we address specific questions. Q: I couldn't find a limitation nor societal impact section, neither a short paragraph on these points. A: Pl...	Summary: The paper takes a systematic approach to address two important aspects of the implementation of diffusion-based generative models: the objective function and the noise schedule, both of which have been conventionally set in a relatively empirical fashion ($i.e.$, based on sample quality). The paper first fram...	Rebuttal 1: Rebuttal: Thank you for your detailed review and thoughtful comments. Please note our top-level comment with additional experimental and theoretical results. Below we address specific questions. **Q: The part of the paper that derives the main conclusion of the paper [...], went a little bit too fast-paced...	Summary: This paper presents new theoretical insights on denoising diffusion objective. Namely, training a diffusion model with a non-uniform weighting can be expressed as a weighted integral of ELBOs, and monotonic weighting results in training ELBO with additive Gaussian noise augmentation. The authors also propose a...	Rebuttal 1: Rebuttal: Thank you for your constructive comments. We strongly agree with you that understanding weighted diffusion training objectives as ELBO with data augmentation gives us a way to fairly compare diffusion models with other likelihood-based models, and indicates that ELBO objective is still compatible ...	Summary: This is a nice paper that provides a comprensive framework for diffusion objectives, giving the ELBO the driver's seat. The paper shows that all of the axes along which diffusion objectives differ (elbo vs Ho-perceptual weighting, noise schedules, different parameterizations like x vs epsilon vs v) can all be...	Rebuttal 1: Rebuttal: Thank you for your valuable comments and kind words to our work. Please note our top-level comment with additional experimental and theoretical results. Below we address specific questions. **Q: Cite and discuss relationship with “Maximum Likelihood Estimation for Diffusion ODEs” (https://arxiv.o...	Rebuttal 1: Rebuttal: Dear Reviewers, Please see the attached PDF for a one-page PDF with a summary of added experimental results. We would like to thank all reviewers for providing constructive feedback that helped us improved the paper. We are encouraged that reviews think our paper: - "provides a comprensive f...	NeurIPS_2023_submissions_huggingface	2,023	Summary: The paper tackles the loss function of diffusion models from a theoretical point of view. The paper studies the relationship between the weighted loss and the ELBO objective. The paper shows that the diffusion model objectives can be formulated as a special case of weighted loss. In addition to the theoretical...	Rebuttal 1: Rebuttal: Thank you for your thoughtful feedback. Please note our top-level comment with additional experimental and theoretical results. Below we address specific questions. Q: Adding more experimental results could have been helpful. A: Please see our top-level comment and the PDF attached to it. We...	null	null	null	null	null	null
Mix-of-Show: Decentralized Low-Rank Adaptation for Multi-Concept Customization of Diffusion Models	Accept (poster)	Summary: This paper proposes Mix-of-Show for decentralized multi-concept customization using diffusion models. To tackle the concept conflict and identity loss problem, the authors propose the embedding-decomposed LoRA and the gradient fusion. The regionally controllable sampling is proposed to address the attribute bi...	Rebuttal 1: Rebuttal: Thanks for acknowledging our work. The concerns are addressed as follows: Notation: W: Weakness --- W.1: Thanks for raising this concern. Both the Custom Diffusion [1] and Mix-of-Show formulate multi-concept fusion as an optimization problem, and thus both methods are indeed related to ...	Summary: The paper presents Mix-of-Show, a framework that tackles the challenges of decentralized multi-concept customization in public large-scale text-to-image diffusion models. It introduces an embedding-decomposed LoRA (ED-LoRA) for single-client tuning and gradient fusion for the center node, enabling limitless co...	Rebuttal 1: Rebuttal: Thank you for acknowledging our work. Part of the discussion on time complexity and limitations is included in Section 2.2 and Section 4.1 in the supplementary material, respectively. Here, we provide a more detailed discussion. Notations: W: Weakness, L: Limitation --- W.1: There a...	Summary: This paper focuses on the problem of multi-concept customization of diffusion model. Compared to the previous methods, this work focuses on fusing multiple customized objects into a single image with different extra conditions such as poses contours. The paper did an extensive study to analyze the roles of tun...	Rebuttal 1: Rebuttal: Thanks for acknowledging our work. Regarding the weakness concerning quantitative experiments, we have included extensive quantitative evaluation in Section 3.1 of the supplementary material and added quantitative analysis of our main components (i.e., ED-LoRA, gradient fusion) in Table 1(...	Summary: This paper addresses a problem of significant practical importance, which is how to employ a number of individually fine-tuned concept plug-ins to produce a model, which can then be used to generate an image that contains any combinations of the concepts. The key method is center node fusion, which trains a fu...	Rebuttal 1: Rebuttal: Thanks for acknowledging the contribution of our work. We address the concerns as follows: Notations: W: Weakness --- W.1: Sorry. We might not fully grasp the meaning behind this statement, "What the paper meant was that the inverted embedding and weights for concept n must be trained b...	Rebuttal 1: Rebuttal: Thank you to all the reviewers for providing valuable comments to help us improve our submission. We would like to clarify that, due to the page limit, the quantitative evaluation, dataset details, limitation discussion and potential negative society impact discussion are provided ...	NeurIPS_2023_submissions_huggingface	2,023	Summary: The paper proposes a framework called Mix-of-Show for decentralized multi-concept customization of diffusion models. The framework uses embedding-decomposed LoRA for single-client tuning and gradient fusion for center-node concept fusion to preserve the in-domain essence of single concepts and support theoreti...	Rebuttal 1: Rebuttal: Thank you for providing many detailed comments. We will address your concerns by referring to the relevant contents in the supplementary material and the attached PDF (in the global response). Notations: W: Weakness, Q: Question, S: Suggestion --- W.1: Thank you for pointing that out. I...	null	null	null	null	null	null
Policy Optimization in a Noisy Neighborhood: On Return Landscapes in Continuous Control	Accept (poster)	Summary: This paper explores the concept of return landscapes in deep reinforcement learning agents for continuous control. The authors demonstrate that the mapping between policy parameters and return can exhibit high-frequency discontinuities, leading to variations in agent performance. They introduce the concept of ...	Rebuttal 1: Rebuttal: Thank you for your feedback! > "while Figure 5 demonstrates that we can smoothly interpolate between levels of stability along a single gradient direction" We apologize for the confusion. This is a typo in the original version of our paper: the direction that is used to interpolate between poli...	Summary: The paper investigates the mapping from policy parameters $\theta$ to returns $R(\theta)$ and the sensitivity of $R(\theta)$ to minor perturbations to $\theta$. The authors show that popular RL algorithms tend to traverse "noisy neighborhoods" in policy parameter space where the return is highly varied even fo...	Rebuttal 1: Rebuttal: Thank you for your feedback! > I suspect there is some way to quantify the claims in this section across a larger population of policy pairs, which I think would significantly strengthen this section. We thank the reviewer for the suggestion. To quantify whether there is a statistically signific...	Summary: In this study, the authors explore the underlying causes of the instability observed in the performance of deep RL policies. Their investigation revolves around examining the relationship between a policy and its corresponding return, with a focus on analyzing how minor variations in policy parameters impact t...	Rebuttal 1: Rebuttal: Thank you for your feedback! > This paper would benefit from providing additional explanation and motivation regarding the significance of investigating the mapping between policy parameters and return. A more thorough exploration of why this particular landscape is crucial to the focus of the wo...	Summary: This paper looks at visualizing the return landscape in typical continuous control environments of deep RL. The authors propose a statistic over the returns profile, namely the left tail probability, and try to present this as a suitable measure of policy quality in a series of experiments and visualizations. ...	Rebuttal 1: Rebuttal: Thank you for your feedback! > The idea of looking at return surfaces is not original (e.g., see [1], [2], and [3] among many others). While the return landscape has been studied in other works, our distributional perspective on it has no precedent. It is also important to distinguish between wo...	Rebuttal 1: Rebuttal: We thank the reviewers for their helpful comments and feedback. Overall, the reviewers found our perspective to be novel (ZEGo, Voa2, Lizq), our paper to be well-written (ZEGo, QuAp, Voa2, Lizq) and the quality of our figures and experiments to be high (Voa2, Lizq). Multiple reviewers (ZEGo, ErL...	NeurIPS_2023_submissions_huggingface	2,023	Summary: This work proposes a new perspective for investigating the stability of RL algorithms -- the return distribution of post-update policies. A simple rejection technique is developed correspondingly to increase the stability in that sense. Strengths: The paper is very well written and easy to follow. This new pe...	Rebuttal 1: Rebuttal: Thank you for your feedback! > In my understanding the concept the authors propose apply to stochastic environments and discrete actions as well. We thank the reviewer for the suggestion. We have run two experiments (Brax with stochastic policies and four games from the ALE, see General Response...	null	null	null	null	null	null
Towards Generic Semi-Supervised Framework for Volumetric Medical Image Segmentation	Accept (poster)	Summary: This paper tackles semi-supervised medical image segmentation over multiple domains, generalizing unsupervised domain adaptation and semi-supervised learning. A multiple branch encoder-decoder architecture with a shared encoder is proposed. The first branch is a diffusion model with V-Net architecture trained ...	Rebuttal 1: Rebuttal: Sincere gratitude to reviewer GWMf for the recognition of our idea and suggestion for supplementing the related work. We hope our answers can address all your concerns. We are looking forward to having further discussions with you. Regarding the differences with ACT: (1) **The settings are diffe...	Summary: This paper addressed both the semi-supervised issue and the domain adaptation issue for medical image segmentation. The proposed method incorporated the diffusion model and used a consistency-based regularization for training. Performance on several public datasets is good for the segmentation tasks. Strength...	Rebuttal 1: Rebuttal: Sincere gratitude to reviewer HPgc. We hope our answers can address all your concerns. We are looking forward to having further discussions with you. Q1: There are some recent papers considering the two tasks as well ... We believe there might be some misunderstandings regarding the differen...	Summary: This work proposes a generic framework for semi-supervised learning (SSL) in medical image segmentation. They identify two obstacles that prevent the applicability of SSL in real-world scenarios. Then they propose Aggregating & Decoupling to alleviate the issues under several settings, including class imbalanc...	Rebuttal 1: Rebuttal: Sincere gratitude to reviewer GWTw for the recognition of our idea and suggestion for extending the application scenarios to natural images. We hope our answers can address all your concerns. We are looking forward to having further discussions with you. **Weaknesses: My biggest concern about thi...	Summary: This paper proposes an Aggregating & Decoupling framework, which unifies the SSL, Class Imbalanced SSL, UDA, and SemiDG for volumetric medical image segmentation with one generic framework. Strengths: This paper explores a generic framework for the SSL, Class Imbalanced SSL, UDA, and SemiDG of volumetric medi...	Rebuttal 1: Rebuttal: Sincere gratitude to reviewer yd6u for the recognition of our idea. We hope our answers can address all your concerns. We are looking forward to having further discussions with you. Q1: How are the running time and parameters of the proposed method compared to other methods? We compared ...	null	NeurIPS_2023_submissions_huggingface	2,023	Summary: The paper presents a novel framework for semi-supervised medical segmentation that seeks to address three different scenarios simultaneously: lack of labelled data (standard semi-supervised learning -- SSL), domain shifts (unsupervised domain adaptation -- UDA) and generalization to new, unseen domains (semi-s...	Rebuttal 1: Rebuttal: Sincere gratitude to reviewer suEd for the recognition of our idea. We hope our answers can address all your concerns. We are looking forward to having further discussions with you. Q1: p2: "Currently, methods for …" While current Semi-DG methods can be evaluated within SSL and UDA, they hav...	null	null	null	null	null	null
Lexinvariant Language Models	Accept (spotlight)	Summary: This paper introduces lexinvariant language models, where the probability of any sequence is equal up to permutation of the vocabulary (e.g. P(A A B C) = P(B B C A)). In practice, the token embeddings within a Transformer model are replaced by random Gaussian vectors, which are consistent across a single seque...	Rebuttal 1: Rebuttal: Thank you for the helpful and thoughtful review! > While the authors acknowledge the practical limitations of the approach, convergence can become quite slow as vocabulary size grows. While a purely lexinvariant model may have practical limitations, we explore and discuss the potential benefits ...	Summary: This paper presents “lexinvariant models” which eschew learning an embedding matrix for the symbols in the sequence and instead of modeling the identity of the symbols, they model the repetition and co-occurence patterns of the symbols in a sequence. Specifically, these symbols are interchangeable among all th...	Rebuttal 1: Rebuttal: Thank you for the supportive and useful feedback! We’ve addressed your points below: > The theoretical analysis seems to have fairly loose bounds. Especially, the distribution matching looks feasible only for impractical long context lengths and there might be a large gap for realistic context le...	Summary: This work explores neural language modelling without fixed token embeddings. Being invariant to symbols forces the model to rely on discovering patterns across sequences (e.g. co-occurrences, repeating symbols) that enable it to infer the underlying symbol sequence. For example, suppose the symbol-substituted ...	Rebuttal 1: Rebuttal: First, we would like to express our appreciation for the comprehensive and encouraging feedback. We hope our response addresses any points of contention and believe the suggestions have already meaningfully strengthened the work. Thank you! > The probing experiments can be improved by discussing ...	Summary: The paper investigates the problem in which language model's tokens can be arbitrarily permuted (e.g., randomly interchanging two words $x$ and $y$ and stipulate that the new $x$ takes $y$'s meaning and vice versa -- such a process can happen over the entire vocabulary). The authors offer theoretical analysis ...	Rebuttal 1: Rebuttal: This review seems to be based on a misunderstanding of the key concept behind our paper (Equation (1)): we are considering permuting the underlying alphabet/"lexical symbols" as in a "substitution cipher", NOT changing the order of the tokens. For example, for a lexinvariant LM, p(“a a b”) = p(“b...	Rebuttal 1: Rebuttal: We would like to thank the reviewers for their thoughtful and encouraging comments. In particular, we appreciate each reviewer’s comments on the novelty of the notion of lexinvariance and the usefulness of the experiments. Aside from one review which appears to be predicated on a misunderstandin...	NeurIPS_2023_submissions_huggingface	2,023	Summary: The paper proposes a new language modelling approach, called lexinvariance, where entries in vocabulary are permuted (with an unknown permutation). The paper proves that a lexinvariant language model can predict language well: it converges to a true language model with a polynomial rate in terms of context ...	Rebuttal 1: Rebuttal: Thank you for the encouraging and thoughtful review! We’ve cleared up the notation you highlighted: > The experimental setting for char-permutation maybe biased. We should also consider "space" as a char. Otherwise, the model would have some good hint for word tokenization. In the experiments, w...	null	null	null	null	null	null
Improving neural network representations using human similarity judgments	Accept (poster)	Summary: The paper proposes a new objective function on the representation level that enforces a global structure similar to that of humans, while maintaining as much as possible the original local structure. The goal is to maintain a separability of classes (ie keep local structure) while placing similar classes close...	Rebuttal 1: Rebuttal: Responses > Why is LCKA the chosen RSM? The motivation behind the use of this measure could be made more clear in the text The appropriate representational similarity measure here ultimately relates to how one defines “global representational structure” mathematically. If one defines global ...	Summary: This paper proposes a linear transformation named gLocal transform to align the global structure of the representation space with human similarity judgments, while at the same time maintains the local structure of the original representation space. The method is shown to be superior than a naive transform that...	Rebuttal 1: Rebuttal: Response > In regards to the experiments on few-shot learning, it would be beneficial to not only compare the results with CLIP experiments but also include experiments with general baselines. This will provide a comprehensive evaluation and determine if the proposed approach consistently im...	Summary: This paper proposes to learn a linear transformation on top of pre-trained CLIP representations to improve alignment with human odd-one-out judgments from the things dataset. To do this, a similarity matrix is constructed based on pairwise similarities between the three images, and the representation is optimi...	Rebuttal 1: Rebuttal: Responses > The results seem to vary by dataset and by model - the benefits are mostly on the CLIP model, and ImageNet backbones are less consistent in their benefits when using the gLocal objective. I would be curious if similar trends would hold for OpenCLIP or DINO backbones. We report r...	Summary: This paper proposes a new alignment strategy (post-hoc) that manages to "re-align"/transform the learned feature space of modern neural networks into a new feature space that is correlated with that of humans. Authors find a way to do this by getting actual human judgment behavioural data and incorporate this ...	Rebuttal 1: Rebuttal: Responses > There is a lack of qualitative assessment, aside from the RDM visualizastion. Alongside the RDM visualizations which demonstrate differences in global similarity structure between human similarity judgments and neural net representations (Figure 5 and Figure E.1), we did a thorou...	Rebuttal 1: Rebuttal: We thank the reviewers for the overall very positive feedback. Although the reviewers’ concerns do not overlap , we would like to provide a general response where we address a few points that could be of interest to all reviewers. Adversarial alignment Reviewer 2yQv suggested to run the gLoc...	NeurIPS_2023_submissions_huggingface	2,023	null	null	null	null	null	null	null	null
Bucks for Buckets (B4B): Active Defenses Against Stealing Encoders	Accept (poster)	Summary: This paper studies active defenses against model stealing attacks, with a specific focus on SSL encoders. The authors propose B4B, a framework including three modules: 1) a coverage estimation to distinguish legitimate users and adversaries; 2) a cost function that maps the coverage to a penalty to prevent ste...	Rebuttal 1: Rebuttal: We thank the reviewer for the insightful comments. >Comparison with previous work All the previous active defenses add the same amount of noise to each returned representation. All of the them can be integrated into our framework in Block 2, where we can replace the addition of random noise,...	Summary: The authors proposed B4B, which aims at actively preventing model stealing while preserving high-utility representations for 133 legitimate users. The paper is overall well-organized and easy to follow. The proposed three blocks are clear and effective, namely the embedding space estimation, cost increasing, a...	Rebuttal 1: Rebuttal: >Baseline in Table 1 We extend Table 1 to include the static addition of noise to all the returned representations (as proposed in [13 (Section A.4),26,32]) as another baseline. We enumerate the following observations for such a defense: 1. If the amount of noise is small ($\sigma$=0.1) the...	Summary: The paper propose a defense over model stealing on-the-fly by assessing dataset distributions, and directly train on it. The paper is well-written and easy to read, while proposing a novel type of active defense over model stealing, without harming the performance of benign users. Strengths: The authors prop...	Rebuttal 1: Rebuttal: We thank the reviewer for the insightful comments. We address individual points below one by one: >“For large tech firms, they might have requests from diverse users and query a specific API many times, which also results in diverse usage.” If we correctly understand the setup, there are use...	Summary: This paper proposes B4B as a defense against model stealing attacks for pretrained encoders. The main assumption of the defense is that attack queries generally have a much broader coverage of the embedding space than queries from normal users. Based on this assumption, B4B utilizes Local Sensitive Hash and se...	Rebuttal 1: Rebuttal: We appreciate the positive, encouraging, and constructive feedback. >Answer to Weakness 1 (W1) & Question 1 (Q1): Querying with a small subset of ImageNet B4B does not classify a user as adversarial or benign but only adaptively calibrates the defense strength based on the coverage of the r...	Rebuttal 1: Rebuttal: We would like to thank all the reviewers for their positive feedback and insightful comments. The paper improved as a result of your reviews. We tackle a current urgent problem of designing an active defense for encoder stealing attacks (reviewers FtSs, N8vt, pRsJ). The proposed approach is inte...	NeurIPS_2023_submissions_huggingface	2,023	null	null	null	null	null	null	null	null
GNeSF: Generalizable Neural Semantic Fields	Accept (poster)	Summary: The paper combines "generalisable" neural radiance fields and semantic radiance fields. It does so in a relatively straightforward manner, borrowing and combining ideas in these two areas, and contributing a few technical advances on top. It demonstrates good segmentation accuracy both for producing the 2D seg...	Rebuttal 1: Rebuttal: ## Respose to Reviewer jaaU (R#5) Thanks for your reviews. We are not sure of the meaning of “test-time optimization” mentioned repeatedly in the reviews. We reckon that this refers to “per-scene optimization” and provides our clarification based on this understanding. We would also like to cla...	Summary: This paper introduces a generalizable 3D scene segmentation framework based on neural implicit representation. The proposed framework takes multi-view image features and semantic maps as inputs to reconstruct 3D geometric and semantic information. The generalizability of the approach enables the synthesis of ...	Rebuttal 1: Rebuttal: ## Respose to Reviewer G1Te (R#4) Q1: The article lacks clear explanations for many details, such as the detailed network structure and training strategy. This leads to poor reproducibility. A: We will release our source code and trained models after paper acceptance. We have provided so...	Summary: This paper, titled GNeSF, focuses on tackling the challenge of generalizable semantic segmentation using NeRF representation for both novel segmentation synthesis and 3D segmentation. The authors introduce several key contributions in this regard. Firstly, they propose utilizing a cost volume to construct a co...	Rebuttal 1: Rebuttal: ## Respose to Reviewer 1Moh (R#3) Q1: Lack of references and provide explicit details regarding the construction process. A: Thanks for your suggestions. We will add all mentioned references into the Related Work Section of the final version. For example, we will insert the following se...	Summary: This paper proposed GNeSF, a generalizable neural semantic fields by incorporating contextual and semantic prior information via a U-shaped CNN and a pre-trained segmentor. Combing multi-view fused image-level features as well as soft weighting using multi-view softmax probabilities, the overall pipeline could...	Rebuttal 1: Rebuttal: ## Respose to Reviewer kUSs (R#2) Q1: What is the main difference with Semantic-Ray [a]. A: Although Semantic-Ray [a] and our method both attempt to learn a generalizable semantic NeRF model, the approach taken by the two methods to solve this problem differs significantly. Semantic-Ray ...	Rebuttal 1: Rebuttal: There are figures and tables in the PDF. In Fig. 1, we show some successful and failure cases and explanations can be found in R#1Q7. Tab. 1 shows the comparison with other generalizable NeSF and explanations can be found in R#2Q2. Tab. 2 shows the effect of the number of nearby views and explanat...	NeurIPS_2023_submissions_huggingface	2,023	Summary: This paper proposes a generalizable 3D segmentation framework. The novel components include (a) a soft voting mechanism that incorporates image features, view difference aggregate the 2D semantic information for 2D views to 3D, (b) a visibility module to avoid being affected by occluded views. With only 2D sup...	Rebuttal 1: Rebuttal: ## Respose to Reviewer xsYz (R#1) Q1: Not enough intuition is given to explain the qualitative results. Fig. 3 shows that NeuralRecon predicts worse than the proposed method on bed. However, the authors didn't explain why this can happen (which design choice has potentially helped with this)?...	null	null	null	null	null	null
Reward-agnostic Fine-tuning: Provable Statistical Benefits of Hybrid Reinforcement Learning	Accept (poster)	Summary: The authors prove that augmenting an offline dataset with online data collection leads to improved sample complexity compared to online and offline RL in the tabular setting, under the condition of single-policy partial concentrability. This is an extension of previous work on single-policy concentrability, re...	Rebuttal 1: Rebuttal: Thanks a lot for the helpful comments and valuable feedback. Please find below our point-to-point response, which will also be incorporated into the final paper. 1. In the original submission, the main algorithm and several associated discussions were included in the appendix due to the page lim...	Summary: The paper proposes a new three step algorithm in the hybrid RL setting for tabular MDPs that improves in terms of sample complexity on the pure online and as well as the pure offline setting. The notion of single-policy partial concentrability is introduced which offers an intuitive concept that connects the s...	Rebuttal 1: Rebuttal: Thanks a lot for the helpful comments and valuable feedback. Please find below our point-to-point response, which will also be incorporated into the final paper. 1. If the number of the newly explored samples is smaller than the number of offline data samples, then the offline dataset and the da...	Summary: This paper studies hybrid reinforcement learning under tabular MDPs, where the agent has access to both a given offline dataset and online interaction with environment. This paper first proposes an interesting notion of single-policy partial concentrability, which relaxs previous concentrability assumption and...	Rebuttal 1: Rebuttal: Thanks a lot for the helpful comments and valuable feedback. Please find below our point-to-point response, which will also be incorporated into the final paper. 1. Frank-Wolfe-type method is an efficient first-order iterative algorithm to solve constrained convex optimization problems. We will ...	Summary: This paper investigates tabular reinforcement learning (RL) in a hybrid setting, where both an offline dataset and online interactions with an unknown environment are available. The main question addressed is how to effectively use online data to enhance and complement the offline dataset for policy fine-tunin...	Rebuttal 1: Rebuttal: Thanks a lot for the helpful comments and valuable feedback. Please find below our point-to-point response, which will also be incorporated into the final paper. 1. This work is mainly our theoretical attempt towards understanding the provable benefits of hybrid setting. While the current algor...	null	NeurIPS_2023_submissions_huggingface	2,023	Summary: This paper studies a tabular reinforcement learning setting where the agent tries to learn the optimal policy by both offline data (obtained by exploiting the optimal policy) and online exploration. A three-stage algorithm is proposed to obtain such a policy, which has lower sample complexity compared to pure ...	Rebuttal 1: Rebuttal: Thanks a lot for the helpful comments and valuable feedback. Please find below our point-to-point response, which will also be incorporated into the final paper. 1. When we say the initial state distribution is ``unknown'', we mean that this distribution is not revealed to the learner {\em a pr...	Summary: This paper studies hybrid RL (online RL with offline data) in the specialized setting of tabular MDPs, while providing tighter bounds than prior works. Namely, they seem to propose a novel and weaker concentrability notion called single-policy partial concentrability. The algorithm takes three stages: 1. estim...	Rebuttal 1: Rebuttal: Thanks a lot for the helpful comments and valuable feedback. Please find below our point-to-point response, which will also be incorporated into the final paper. 1. In the original submission, the main algorithm was included in the appendix due to the page limit. In the final version, we believ...	null	null	null	null
Minimum Description Length and Generalization Guarantees for Representation Learning	Accept (poster)	Summary: This paper obtains new generalization bounds of the conditional mutual information-flavor (Steinke and Zakynthinou, 2020), where the bounds depend only on the predictions that the algorithm induces on the supersample. These bounds are phrased more generally in terms of a KL divergence with symmetric priors, an...	Rebuttal 1: Rebuttal: We thank the reviewer for their careful reading of the paper, their interest in our work, and their constructive comments. Regarding CMI: We thank the reviewer for bringing our attention to CMI-based works and in particular the f-CMI-based follow-ups. We missed this connection, as the origina...	Summary: The paper advances state-of-the-art on information theoretic generalization guarantees for a class of representational learning problems. The contribution is positioned in the framework of information bottleneck approach and builds on the classical work of Blum and Langford for PAC-MDL bounds. The main result ...	Rebuttal 1: Rebuttal: We thank the reviewer for their interest and their thoughtful comments that help to ameliorate the quality of the work. We will improve the readability of the paper by gently introducing the used information-theoretic concepts. Among others, by explaining how by adding the "block-coding" techniqu...	Summary: Post-rebuttal I thank the authors for their response and have raised my score to an accept. Understanding the generalization properties of modern machine learning algorithms is one of the key challenges in statistical learning theory. Focusing on the classification setup, this paper first studies the ge...	Rebuttal 1: Rebuttal: We thank the reviewer for their interest in our work and their constructive comments. We will apply them to improve the quality of our work. Regarding the presentation: Thanks for all suggestions related to the presentation of our work. We will definitely take them into account for the next v...	Summary: This submission considers supervised learning problems and derives two types of generalization gap bounds. The first type states that generalization gap can be upper bounded by approximately $\sqrt{\mathrm{KL}(P \|\| Q) / n}$, where $P$ is the average (wrt data and training stochasticity) of predicted label dist...	Rebuttal 1: Rebuttal: We thank the reviewer for their detailed investigation of the paper, their interest, and their constructive comments. Weakness #1: As mentioned by the reviewer, our work is the first theoretical work that characterizes $gen(S,W)$ in terms of the complexity of latent variables, achieved by con...	Rebuttal 1: Rebuttal: We sincerely thank all the reviewers for their valuable time and their pertinent comments that helped to improve our work. We are excited to see that all reviewers appreciated the value of our work. We carefully read the comments of all reviewers and addressed them separately. Please do let us kno...	NeurIPS_2023_submissions_huggingface	2,023	null	null	null	null	null	null	null	null
Robust Nonparametric Regression under Poisoning Attack	Reject	Summary: The paper studies robust nonparametric regression under poisoning attacks. The input is samples from some fixed distribution and the goal is to approximate some unknown function on the input space. It is assumed that the values observed in q of the samples are adversarial. In this setting, classical approaches...	Rebuttal 1: Rebuttal: Thanks for this positive review. 1.Nonparametric regression is a fundamental problem. The paper studies a natural adversarial setting where a small number of input samples are corrupted adversarially. This research direction can produce more robust useful data-analytic methods. Reply: Thank...	Summary: A robust version of the classic non-parametric problem is studied, when the training data is under an adversarial poisoning attack. The work is primarily theoretical, and makes assumptions on Lipschitzness and boundedness of the underlying function, boundedness of the data density, restrictions on sharpness of...	Rebuttal 1: Rebuttal: Thanks for the comments. We believe that for theoretical works, even if the method is proposed much earlier, new theoretical analysis is still a meaningful contribution, especially if the minimax lower bound is provided and the upper and lower bound matches. Examples include ref. [1-3] in this reb...	Summary: The authors propose two new nonparametric regression methods that and study their resilience under poisoning attacks. The baseline for comparing the performance of the proposed methods is kernel regression (aka Nadaraya-Watson estimator). The proposed methods are two, since the initial idea that the authors ...	Rebuttal 1: Rebuttal: Thanks the reviewer for valuable comments. The space is quite limited here. We reply to these comments briefly. We are willing to discuss more during the discussion period later. 1.Missing a section with preliminaries or background knowledge Thanks for this suggestion. Loss function, kernel ...	Summary: This paper studies nonparametric regression, where an adversary can corrupt $q$ samples from the training set. The paper proposes robust estimators for this problem. Strengths: The paper studied the Huber loss minimization approach under adversarial noise, giving theoretical upper bounds and lower bounds. ...	Rebuttal 1: Rebuttal: Thanks the reviewer for this positive review. 1.Comment: (from the strengths) The paper studied the Huber loss minimization approach under adversarial noise, giving theoretical upper bounds and lower bounds. Reply: Studying the Huber loss minimization under adversarial noise corresponds to the f...	Rebuttal 1: Rebuttal: Thanks for the effort of all reviewers. We have received many valuable comments. We have the following general responses: 1.Clarification of the novelty. Robust nonparametric regression has been studied much earlier, such as the book [1] mentioned by the reviewer i98Z. However, early works o...	NeurIPS_2023_submissions_huggingface	2,023	null	null	null	null	null	null	null	null
Provably Robust Temporal Difference Learning for Heavy-Tailed Rewards	Accept (poster)	Summary: This paper studies the problem of reinforcement learning under heavy-tailed reward. The paper present two algorithms: robust TD learning and robust NPG that achieves polynomial sample complexity in the generative setting. Strengths: The problem of heavy tail reward is important and understudied in RL. Weak...	Rebuttal 1: Rebuttal: We thank the reviewer for the feedback and questions. Please refer to our general response above for our contributions, and differences of our work with the existing papers. --- Rebuttal Comment 1.1: Title: Thank you for the elaboration of contribution Comment: Thank you to the authors for elabo...	Summary: This paper studies the TD and NAC with heavy tailed reward noise. The authors provide both theoretical results and the numerical verification to support their claims. Strengths: - This paper is well written and easy to follow. The numerical results firmly support their theoretical claims and show the advantag...	Rebuttal 1: Rebuttal: We thank the reviewer for the feedback and questions. Below, we provide our answers to the comments in the Weaknesses section. - W1: Contributions. Please refer to our general response for our contributions and the differences between our paper and existing works in the literature. - *W2: Know...	Summary: This work discusses the failure of classical TD learning under heavy-tailed reward. This phenomenon motivates the author to propose a new technique, robust TD learning, with a dynamic gradient clipping mechanism to overcome this issue. The theoretical convergence rates for both in-expectation and high-probabil...	Rebuttal 1: Rebuttal: We thank the reviewer for the feedback and questions. Below, we provide our answers. - Question 1. Regarding Assumption 1.2, please also see our general response above. As the reviewer pointed out, our Assumption 1.2 implies that the random reward has finite mean and potentially infinite varian...	Summary: The paper studies the value function estimation problem when the reward of the MDP is heavy tail. A robust version of TD learning is proposed which introduces a clipping step to the algorithm. The authors give both the theoretical analysis and numerical experiments for the proposed algorithm. And they further ...	Rebuttal 1: Rebuttal: We thank the reviewer for the comments and questions. - We refer to our general response above for our contributions, and for differences with [4] (i.e., Bhandari et al., 2018). - We thank the reviewer for pointing out the typo, we will remove the subscript $k$ in the updated version. Below, ...	Rebuttal 1: Rebuttal: We would like to thank the reviewers for their very valuable feedback. In the following, we provide a general response to the questions commonly raised by the reviewers. In order to address specific questions, we will post individual comments. **1. Realizability assumption and approximation error...	NeurIPS_2023_submissions_huggingface	2,023	Summary: The authors propose and study variants of TD(0) and Natural Actor-Critic that provably converge under (on-policy in the case of TD(0)) linear function approximation when the rewards may not have finite second moment or variance, but instead a finite (1+p) moment for $p \in (0,1]$; the authors use this as the d...	Rebuttal 1: Rebuttal: We would like to thank the reviewer for the comments and questions. Below, we provide answers to the questions. - Question 1. We have provided our comments regarding the realizability assumption in our general response above. - Question 2. We thank the reviewer for this important question. W...	Summary: This paper discusses robust Temporal Difference (TD) methods for RL with heavy-tailed rewards. Specifically, it proposes robust variants of TD(0) and Natural Actor-Critic (NAC) for policy evaluation and RL control, respectively, and then derives their average iterate's convergence rates. Strengths: 1. Converg...	Rebuttal 1: Rebuttal: We would like to thank the reviewer for the questions and comments. Below, we provide our responses to the reviewer's comments and questions in Weaknesses and Questions, respectively. - W1: Assumption 3. We thank the reviewer for pointing this out. Please note that we have also addressed th...	null	null	null	null
Geometric Neural Diffusion Processes	Accept (poster)	Summary: This paper proposes several extensions to functional diffusion models. In particular, the focus is on extending the process to cover various geometric inputs (such as manifold values and symmetries), although there are some other components (such as a different formulation of the diffusion that allows for more...	Rebuttal 1: Rebuttal: We appreciate your thoughtful feedback and the positive assessment of our manuscript. Experimental benefits over the previous formulation The goal of Table 1 (bottom part) is to highlight a key advantage of our geometric approach to neural diffusion processes (NDP): the ability to generalize ...	Summary: In this paper, the authors propose a generative model over vector fields. The proposed approach is a combination of ideas from neural diffusion processes [1], equivariant learning of stochastic fields [2] and equivariant scores [3]. The authors represent vector fields as vectors in finite sets of points and us...	Rebuttal 1: Rebuttal: We are grateful for your insightful comments and the favourable evaluation of our manuscript. In the following, we address your primary concerns, aiming to further enhance your appreciation of our work. 1. Scaling to Higher Dimensionality We agree that extending our experiments to higher dim...	Summary: This paper proposes GeomNDP, a continuous-time functional diffusion model which can incorporate symmetries of the problem of interest. In particular, this model is constructed via a collection of finite-dimensional Ornstein-Uhlenbeck processes (one for each finite-dimensional marginal), which target a given pr...	Rebuttal 1: Rebuttal: We thank you for your feedback and the positive review. Below, we respond to your main concerns. Minor comments: >The submission was not properly anonymized We apologise for mistakenly attaching an early version of the appendix with the main pdf. >Line 85: it was somewhat unclear what exact...	Summary: The present thesis deals with the problem of learning geometric-sensitive Denoising Diffusion Models (DDMs) for modeling equivariant conditional processes in (non)-Euclidean observation spaces. In doing so, the authors extend existing work that defines diffusions over function spaces with infinite dimensions u...	Rebuttal 1: Rebuttal: Thank you for taking the time to produce such a thorough review. In particular, thanks for your concrete suggestion on how to improve section 3.1 - we have updated our manuscript accordingly. Furthermore, due to strong space constraints, we had to limit the number of references and definitions. T...	Rebuttal 1: Rebuttal: Dear Reviewers, We thank the reviewers for taking the time and effort to provide feedback on our submitted manuscript. We are glad that reviewers found our proposed method sound, well-motivated, well-written and to deliver an interesting and valuable contribution to the community. In light of t...	NeurIPS_2023_submissions_huggingface	2,023	null	null	null	null	null	null	null	null
Disambiguated Attention Embedding for Multi-Instance Partial-Label Learning	Accept (poster)	Summary: This paper proposes a new algorithm for a weakly supervised learning paradigm called multi-instance partial label learning, where each sample is represented by a multi-instance bag and its label is provided as a set of candidate labels. This is a particularly challenging learning paradigm since learners do not...	Rebuttal 1: Rebuttal: We greatly appreciate your constructive feedback and comments, which have proven instrumental in enhancing the quality of our paper. Your recognition of the positive aspects of our work is sincerely valued. In the subsequent sections, we present your comments and our responses in a point-by-point ...	Summary: The paper proposes the first deep learning-based method for multi-instance partial-label learning (MIPL) problem. The proposed method employs a disambiguation attention mechanism to approximate the contribution of each instance to the bag-level feature and then aggregates a multi-instance bag into a single vec...	Rebuttal 1: Rebuttal: Thank you for dedicating your time to reading and reviewing our paper. Your comments have provided valuable insights that greatly contribute to the enhancement of our work. In the following sections, we present your comments and provide detailed responses to each of them. > A notation in Section ...	Summary: The article proposes a novel algorithm called DEMIPL for addressing the challenges in multi-instance partial-label learning (MIPL) tasks. The algorithm uses a disambiguation attention mechanism to embed multi-instance bags into a single vector representation and a momentum-based disambiguation strategy to iden...	Rebuttal 1: Rebuttal: Thank you for providing constructive feedback and insightful comments to enhance our paper. We greatly value your recognition of the positive aspects of our work. We will now address the specific concerns you raised. > Recently, using distribution of labels to generate partial labels have been co...	Summary: The paper introduces a novel algorithm named Disambiguated attention Embedding for Multi-Instance Partial-Label learning (DEMIPL) to address the problem of multi-instance partial-label learning (MIPL). The proposed algorithm embeds multi-instance bags into a single vector representation using a disambiguation ...	Rebuttal 1: Rebuttal: Thank you for your valuable feedback and insightful suggestions to enhance the quality of our manuscript. We have diligently integrated your input into the revised version. In the following sections, we present your comments, followed by our detailed responses. > When discussing the experimental ...	null	NeurIPS_2023_submissions_huggingface	2,023	Summary: In this paper the authors study multi-instance partial label learning, where a bag of instance is associated with several candidate labels of which it is assumed one is correct. The propose a deep learning approach for such a problem. The key contribution is a "disambiguation attention mechanism" where they fi...	Rebuttal 1: Rebuttal: Thanks for your valuable feedback and insightful comments, which have greatly contributed to the enhancement of our paper. We also appreciate your acknowledgment of the positive attributes of our work. We will now address the mentioned negative aspects. Below, we provide a summary of your comments...	null	null	null	null	null	null
Intra-Modal Proxy Learning for Zero-Shot Visual Categorization with CLIP	Accept (poster)	Summary: Authors provide a series of detailed theories to support the conclusion that the inter-modal distance of the CLIP model is large and the gap can be reduced by introducing proxy features in visual feature space. They also propose the intra-modal proxy learning method to improve the performance of CLIP model wit...	Rebuttal 1: Rebuttal: R-DJum-W1-A: Yes, they are the same. 0.01 was obtained by CLIP and we did not change it. We will make it clear in revision. R-DJum-W2-A: Please kindly note that the result is obtained on ImageNet, where achieving a gain of 0.25\% is not trivial. Although the parameter can be eliminated to reduce ...	Summary: This work theoretically proves that the gap cannot be reduced sufficiently by minimizing the contrastive loss in CLIP and the optimal proxy for vision tasks resides only in the vision space. Two strategies are developed to further refine the pseudo label obtained by the text proxy to facilitate intra-modal pro...	Rebuttal 1: Rebuttal: R-fSSP-W1-A: The main contribution of this work is to theoretically understand the modality gap in CLIP, which has also inspired our novel proposal of proxy learning. To the best of our knowledge, this is the first work to explore intra-modal proxy learning for CLIP. Meanwhile, to better recover t...	Summary: Based on the phenomenon of modality gap, this paper analyzes that the proxy of visual space is better than that of text space, and explains this theoretically. This paper designs a test-time adaptation method for unlabeled test data Strengths: 1. The influence of the modality gap on image classification is a...	Rebuttal 1: Rebuttal: R-dkv6-W-A: Thanks for this great suggestion. With the contrastive pre-training, the textual feature will be aligned with the vision feature to enable zero-shot learning. It means that the similarity between the image feature $x_i$ and its corresponding text $t_i$ as in Proposition~1 can be lower ...	Summary: The paper aims to improve the zero-shot performance of CLIP, i.e.: the setting in which given a test image, we compute its visual embedding, computes its similarity to a pre-computer set of text query embeddings where each text query corresponds to a different class and finally output the classification result...	Rebuttal 1: Rebuttal: R-eKBJ-L-A: Thank you for the suggestion. We will add the following discussion. While our analysis is for the modality gap in contrastive pre-training with the architecture of dual encoders, it cannot be applied directly for other pre-training paradigms with different architectures. For example,...	Rebuttal 1: Rebuttal: We would like to thank all reviewers for their valuable time and insightful comments. We will answer (A) the weaknesses (W), questions (Q) and limitations (L) for each reviewer. The experiments suggested by Reviewer DJum can be found in the attached PDF. Pdf: /pdf/19164e7fc0acb5bc98d75ffcec96d6fc6...	NeurIPS_2023_submissions_huggingface	2,023	null	null	null	null	null	null	null	null
Moment Matching Denoising Gibbs Sampling	Accept (poster)	Summary: This paper proposes a new denoising scoring matching (DSM) technique to train energy-based models. They address the well-known inconsistency problem of DSM objectives by training a "noisy" model and then sampling "clean" samples from the "noisy" model by learning a denoising distribution. The contributions of...	Rebuttal 1: Rebuttal: We're grateful for the constructive feedback. We're glad the clarity, references, and experiments in our paper resonated with you. Your feedback greatly enhances the paper's quality. ### Typos and Reference We want to thank the reviewer for pointing out the typo $s(x)\equiv \nabla_x\log p(x)$. T...	Summary: This paper proposed a sampling scheme called (pseudo)-Gibbs sampling with moment machine. It can be applicable to the sampling problem from from the underlying clean model when a noisy model hat has been well-trained via DSM. The proposed sampling technique is used to sample images for image generation and s...	Rebuttal 1: Rebuttal: We appreciate the reviewer for acknowledging the strengths of our paper. ### Contribution and Effectiveness We wish to emphasize that this paper's central contribution is to elucidate the theoretical optimal full-covariance Gaussian approximation for the denoising distribution. As delineated in T...	Summary: This paper addressed the inconsistency problem in training energy-based models (EBMs) using denoising score matching (DSM). An efficient sampling framework, called (pseudo)-Gibbs sampling with moment matching, was proposed to enable effective sampling from the underlying clean model when given a well-trained '...	Rebuttal 1: Rebuttal: We're grateful to the reviewer for their valuable feedback, which has undeniably enhanced the quality of our work. The commendation on our theoretical contribution and presentation is heartening and much appreciated. Below, we respond to the queries presented. ### Wider scenarios and applications...	Summary: This work proposes a sampling method from implicitly defined clean model given a noisy model without retraining and demonstrates its applicability to high-dimensional data / images. Strengths: - this work presents interesting theoretical results such as the uniqueness of underlying clean model given noisy da...	Rebuttal 1: Rebuttal: We express our gratitude to the reviewer for recognizing the depth and value of our research. It's gratifying to see the acknowledgement of our novel theoretical findings. Here are our responses to the questions posed. ### Effectiveness of the Proposed Method We wish to emphasize that the central...	Rebuttal 1: Rebuttal: We want to thank all the reviewers again for the time spent and their valuable feedback that improves the quality of our work. Their recognition of our theoretical contribution, which forms the cornerstone of our paper, is deeply encouraging. Here's a succinct overview of our primary contributions...	NeurIPS_2023_submissions_huggingface	2,023	Summary: This paper proposes a sampling framework, i.e., (pseudo)-Gibbs sampling with moment matching, which aims at sampling from the implicitly-defined clean model given the noisy energy-based model well-trained by the denoising score matching method. It first provides a theoretical justification of the existence of ...	Rebuttal 1: Rebuttal: We thank the reviewer for their time and valuable feedback that improves the quality of our work. We are encouraged by the positive comments regarding our theoretical contribution and presentation. ### Hessian computation We want to highlight that our method can directly obtain the optimal covar...	null	null	null	null	null	null
Prefix-Tree Decoding for Predicting Mass Spectra from Molecules	Accept (spotlight)	Summary: This paper proposes a novel method for predicting a tandem mass spectrum from a given molecular graph. The key idea is to subdivide the problem into predicting fragmentation of the graph and then separately predicting intensities associated with each fragment. Relative to other recent methods that have adopt...	Rebuttal 1: Rebuttal: Thank you for your very constructive comments and suggestions to help improve the work! W1: “Physically inspired” should not necessarily be a desiderata A: We agree that the desiderata are somewhat interrelated and your point is well taken (and thanks — we have removed the hyphen!). W...	Summary: The manuscript presents a method called SCARF for the prediction of tandem mass spectra from molecular structures. This allows for building up large in-silico libraries that, together with measured spectra, can be used in spectral library search workflows for metabolite identification. From a general machine l...	Rebuttal 1: Rebuttal: Thank you for your very fair comments and feedback. W1: Putting SCARF in context with other recent methods A: Thanks for your suggestion. We will add a section to the Appendix explaining how SCARF has more flexibility as it does not require structural fragment labeling, higher speed, ...	Summary: The paper presents a new method for predicting mass spectra from molecules called SCARF, which stands for Subformulae Classification for Autoregressively Reconstructing Fragmentations. Mass spectra are sets of peaks that represent the masses and intensities of fragments of molecules after they are ionized and ...	Rebuttal 1: Rebuttal: Thank you for your comments and positive reception of our work. We believe you have identified several key discussion points and highlighted limitations of the work that would be intriguing starting points for future work. We now plan to incorporate a new section into the Appendix that explicitly ...	Summary: The authors propose a model, SCARF, that predicts Tandem-MS spectra in a two-step process: 1) predicting atomic subsets, 2) predicting peaks given the subsets. The model's performance is evaluated on two datasets (NIST20, Natural Product Library) using on spectral similarity, coverage, validity, and retrieval,...	Rebuttal 1: Rebuttal: Thank you for your comments! Point-by-point response below. W1: It would be helpful to provide some more context for the contents of the two datasets, NIST20 and NPLIB… A: We agree and will add a section in the Appendix with training dataset details by molecular weight and also molecul...	Rebuttal 1: Rebuttal: We thank all the reviewers for their thoughtful feedback and comments. Where relevant, we have addressed stated weaknesses (W), questions (Q) and limitations (L) with our answers (A). We also include an additional single page PDF containing additional referenced figures and tables....	NeurIPS_2023_submissions_huggingface	2,023	null	null	null	null	null	null	null	null
SceneScape: Text-Driven Consistent Scene Generation	Accept (poster)	Summary: The paper present a method for text-driven perpetual view generation. It takes as input a text-prompt and carema trajectory as input, optimizing a unified mesh representation for rendering coherent video sequence accordingly. The key idea is straight-forward by connecting text-driven inpainting and depth warpi...	Rebuttal 1: Rebuttal: We appreciate the comments and feedback from the reviewer. ### Outdoor scenes Extending our method to handle outdoor scenes is an interesting research direction that indeed poses several challenges, including the accuracy of the estimated depth, as pointed out by the reviewer, as well as handlin...	Summary: This paper approaches the task of text-conditioned generation of a walkthrough video upon (mostly indoor) scenes. It proposes to leverage pretrained text-to-image and pretrained depth prediction models, to generate these videos without finetuning. Implementation generates videos in an online manner by repeated...	Rebuttal 1: Rebuttal: We appreciate the comments and feedback from the reviewer. ### Camera motion and scope Please refer to our global response regarding our choice of default camera trajectory. Please also refer to the attached PDF and the provided video for sample extended results, including large camera turns and...	Summary: 1. This paper proposes a method for test-driven consistent scene generation using a pretrained text-to-image model and depth prediction models. 2. The updating of mesh representation and fine-tuning of the decoder and depth prediction module enable continuous scene generation. Strengths: 1. The continuous sce...	Rebuttal 1: Rebuttal: We appreciate the comments and the feedback from the reviewer. ### Error accumulation analysis Following the reviewer’s comment, we measured our error accumulation for our video evaluation set. To do so, we measured: (i) visual quality by measuring the CLIP aesthetic score as a function of fram...	Summary: In this paper, the authors study text-driven perpetual view synthesis with the input of text prompts describing the scene and desirable camera poses. Instead of training models on a specific domain, they propose an online scene generation method with a pretrained text-to-image model and a pretrained monocular ...	Rebuttal 1: Rebuttal: We appreciate the comments and the feedback from the reviewer. ### Camera Motion Please refer to the global response section, where we elaborate more on our camera motion. Text2Room (concurrent work) focuses on generating a mesh rather than synthesizing a coherent video as we do. Thus, they tai...	Rebuttal 1: Rebuttal: We would like to thank the reviewers for their thorough and valuable feedback. Please note that a link to the video file has been sent to the AC with additional results. Please see the per-reviewer responses for details. ### Camera motion The reason we focused on backward rotational camera moti...	NeurIPS_2023_submissions_huggingface	2,023	Summary: This paper proposes a text-driven perpetual view generation method that synthesizes long-term changing views of a scene and its corresponding 3D mesh, given the free-vocabulary text and camera trajectories. The authors apply an off-the-shelf pre-trained text-to-image diffusion model and a pre-trained depth pre...	Rebuttal 1: Rebuttal: We appreciate the comments and the feedback from the reviewer. ### Dynamic content generation By design, our method supports only static scenes since we use a shared unified mesh representation for all the frames. Thus, when providing SceneScape with a prompt that generates an object with dynamic...	null	null	null	null	null	null
First- and Second-Order Bounds for Adversarial Linear Contextual Bandits	Accept (poster)	Summary: This paper studies the adversarial linear contextual bandit setting, an online learning problem with $K$ arms, stochastic contexts in $\mathbb{R}^d$, and adversarial rewards (which are assumed to be a linear function of the context for each arm). Previous work proves a worst-case regret guarantee that is $\sqr...	Rebuttal 1: Rebuttal: To emphasize the above, an application of an algorithm designed to solve the linear bandit problem cannot be applied directly to the linear contextual bandit, and there are no such examples for this setting. With regard to your questions: 1. The meaning of the expectation is with respect to the f...	Summary: The submission considers semi-adversarial contextual bandit setting --- where d-dimensional contexts x_t are drawn iid, where some opponent picks K loss functions that are linear in the contexts at the beginning of each round, before the actual context x_t is drawn, and where he learner picks one of the K loss...	Rebuttal 1: Rebuttal: Indeed, we did omit a discussion of the epsilon approximations in the regret bound. As you note, one would require $\epsilon = O(1/T)$ to ensure that the contribution of the additive approximation sums over rounds to a constant. It is also correct that the epsilon approximation should be recalcul...	Summary: The authors explore the adversarial linear contextual bandit problem, where the context is i.i.d. drawn from a known fixed log-concave distribution, and the linear loss functions are adaptive adversaries. They provide first adaptive bounds for this setting, including a first-order bound based on the cumulati...	Rebuttal 1: Rebuttal: We want to emphasize that algorithms designed for solving the linear bandit problem cannot be applied to the contextual bandit problem, and thus there is no previous method that could simply yield the result that we claim in this paper. For a comparison with Ito et al., please see our general repl...	Summary: Addressing to the K-armed linear contextual adaptive adversary bandit problem with d-dimensional contexts over T rounds, using a truncated version of the continuous exponential weights algorithm over the probability simplex, the expected regret about the cumulative second moments of the losses is obtained. and...	Rebuttal 1: Rebuttal: Please see our general reply for answers to some of the points you raised. Here we only respond to the remaining points. Regarding time-varying distributions: it is common in contextual bandits to assume that both the contexts and the losses for the online recommendation arms are i.i.d. It is an ...	Rebuttal 1: Rebuttal: To all reviewers: We thank all reviewers for their insightful comments. Comparison to Neu \& Olkhovskaya (2020): We note that our approach is not a straight-forward extension of Neu \& Olkhovskaya (2020), which would lead to discrete exponential weights over arms (i.e.\ the corners of the simpl...	NeurIPS_2023_submissions_huggingface	2,023	null	null	null	null	null	null	null	null
Geometric Analysis of Matrix Sensing over Graphs	Accept (poster)	Summary: This work is novel in a sense of studying matrix sensing over graphs - and provides the first theoretical results on the optimization landscape of the MSoG problem - and intruduce a condition to characterize the optimization complexity Strengths: Its a theoretical work and ideas are nice, and it generalize th...	Rebuttal 1: Rebuttal: Real life applications and more experiments: we would like to thank the reviewer for the comment. We discussed the power state estimation problem and the numerical results the "application of MSoG problem" and the "numerical results" sections in the global response, respectively. Since our respo...	Summary: In this work, the author introduced a problem which they call matrix sensing over graphs. It is sort of a generalized version of PSD matrix completion problem. To fit in the new problem formulation, the author introduced a RIP type condition and showed that it characterizes the problem complexity. The major co...	Rebuttal 1: Rebuttal: 1. This work lacks support of applications: please see the "Application of MSoG problem" section in the global response. 2. *Generalization to the case when $M^$ is not PSD:** we would like to thank the reviewer for the insightful comment. Please see our response point 1 to Reviewer do7X fo...	Summary: This paper introduces the matrix sensing over graphs (MSoG) problem, which encompasses matrix sensing and matrix completion problems. The paper proposes a non-convex regularized optimization method for this new problem and provides theoretical guarantees for its solvability using the gradient descent algorithm...	Rebuttal 1: Rebuttal: 1. The problem is restricted to cases where the true unknown matrix is positive semi-definite: we would like to thank the reviewer for pointing out this limitation. In existing literature, there are two different formulations of the matrix completion problem: the positive semi-definite (PSD) c...	Summary: This paper considers a novelly proposed problem of matrix sensing over graphs (MSoG), which is a general case of matrix completion and matrix sensing problems. It proposes a novel variant of RIP condition named Ω-RIP condition and analyze the optimization landscape of MSoG problem. With an improved regularizer...	Rebuttal 1: Rebuttal: 1. The writing of this paper can be more clear: we would like to thank the reviewer for the constructive comment. We will follow the reviewer's suggestion to revise the presentation of the paper. 2. Application of MSoG problem: please see the "application of MSoG problem" section in the ...	Rebuttal 1: Rebuttal: 1. Application of MSoG problem: we show that the power state estimation problem [24,39] can be formulated as the MSoG problem (12). This is the most important data-analysis problem for power systems and it is solved every 5-15 minutes in practice using heuristic methods with little theoretical...	NeurIPS_2023_submissions_huggingface	2,023	null	null	null	null	null	null	null	null
Fairly Recommending with Social Attributes: A Flexible and Controllable Optimization Approach	Accept (poster)	Summary: Item-side group fairness (IGF) is crucial in recommendation tasks. Existing fairness notions focus on direct utility of item exposures but overlook the social utility, such as the influence of users' social attributes on exposed items. This paper introduces social attribute-aware IGF metrics and proposes a new...	Rebuttal 1: Rebuttal: > W1: Matrix factorization is not the mainstream choice for recommendation in practice. Tree model and deep model are the hot choice. Choosing them as backbone may further improve the impacts of this work. Response to W1: Thanks for your helpful suggestion. **As suggested by the reviewer,...	Summary: This paper considered both direct and social utilities on the Item-side group fairness (IGF). The authors proposed a new flexible and controllable optimization approach to trade-off the utilities while ensuring accuracy. Firstly, they proposed two social attribute-aware IGF metrics, NSP and NEO by leveraging u...	Rebuttal 1: Rebuttal: Thank you very much for your thoughtful and comprehensive feedback on our paper. With your helpful comments, we have revised the Introduction and Experiments Sections with the aim of presenting the relevant work in a more systematic and clear manner, as well as adding missing experimental ...	Summary: This paper aims to solve the Item-side Group Fairness (IGF) issue in Recommender Systems (RSs). The paper claims that existing IGF notions only focus on the direct utility of item exposures while ignore the influence from the social attributes of users. To tackle this issue, the paper proposes social attribute...	Rebuttal 1: Rebuttal: We sincerely thank you for the helpful suggestions. As suggested by the reviewer, we have added extensive experiments to address all your concerns. Below, we hope to address your concerns and questions to improve the clarity and quality of our paper. ### Idea and technique > W1(a): Th...	Summary: This paper introduces a new problem of fairness in recommendation tasks, specifically considering the social attribute of users and its influence on item exposures. The paper proposes social attribute-aware fairness metrics and formulates a multi-objective optimization problem to train recommender models with ...	Rebuttal 1: Rebuttal: We sincerely thank you for the helpful suggestions. As suggested by the reviewer, we have added extensive discussions and experiments to address all your concerns. Below, we hope to address your concerns and questions to improve the clarity and quality of our paper. ### Methodology > **W...	Rebuttal 1: Rebuttal: We sincerely thank all the reviewers for valuable comments and suggestions! Below we have tried our best to address the concerns point-to-point for each reviewer, and we are looking forward for further discussion about our work. In addition, we have uploaded a page of supplementary material here, ...	NeurIPS_2023_submissions_huggingface	2,023	null	null	null	null	null	null	null	null
Efficient Low-rank Backpropagation for Vision Transformer Adaptation	Accept (poster)	Summary: This paper proposed an efficient training method, called Low-rank Back Propagation via Walsh-Hadamard Transformation (LBP-WHT). Specifically, unlike previous works that target low-rank adaptation for model parameters, LBP-WHT aims to reduce the computational cost during backpropagation by projecting the gradie...	Rebuttal 1: Rebuttal: 1. > Theoretical complexity reduction cannot reflect the real speedup for training. It would be better for the authors to report training time (GPU hours) or training throughput (images/s) in main experiments, especially for Table 2. As our primary contribution is on the algorithm side, we u...	Summary: This paper introduces a method to efficiently finetune Vision Transformers (ViTs) for specific tasks such as image classification and semantic segmentation. The method, Low-rank Backpropagation via Walsh-Hadamard Transformation (LBP-WHT), reduces the computational cost of backpropagation in ViTs by projecting ...	Rebuttal 1: Rebuttal: 1. > Why only ViTs? That too only smaller scale ViT models? Yes, LBP-WHT works well with other architectures. In this paper, both EfficientFormer and EfficientFormerV2 are hybrid models of CNN-ViT that incorporate both multi-head attention modules from ViT and convolution layers from CNN. To...	Summary: The paper presents to compute the low-rank approximation for the matrices used in back-propagation, and thus it can save the computation in multiplication. The proposed approach is more efficient than LoRA because no need to backpropagate to the original weights, and experiments also support the claims. Stren...	Rebuttal 1: Rebuttal: 1. > Although the paper claims that the approach is complementary to other efficient approaches (such as quantization), there is no experimental support for it. Experimental results when combining LBP-WHT with LoRA To show that our method is complementary to other existing methods, w...	Summary: The author discusses the issue of computational complexity during matrix multiplication in the backpropagation of linear layers in the vision transformer architecture. To address this issue, the author proposes a method called Low-rank BackPropagation via Walsh-Hadamard Transformation (LBP-WHT), which projects...	Rebuttal 1: Rebuttal: 1. > *The literature survey section in this paper is a bit weak. The author should compare related/similar transformer models, particularly QLoRA, and similar work on "Hadamard product for low-rank bilinear pooling." A more thorough literature review is necessary in this regard. It would be benefi...	Rebuttal 1: Rebuttal: We genuinely appreciate the reviewer's insightful comments and suggestions in reviewing our paper. Below we include additional references used in the rebuttal for all reviewers. ## Notation for Additional Experiments We follow the same notation as the manuscript. To be more specific, \| Notatio...	NeurIPS_2023_submissions_huggingface	2,023	Summary: The paper proposes a cost-efficient back-propagation method using a low-rank trick for vision transformers. When transferring a huge-size vision transformer to the downstream task, the training cost would be high. The paper tackles the problem using low-rank approximation with Walsh-Hadamard Transformation (WH...	Rebuttal 1: Rebuttal: 1. > *The targeting scenario of the paper seems somewhat narrow. The paper focuses on vision transformers and its adaptation. It would be nice to discuss the possibility of the proposed method for the efficient pre-training scenario. Also, a discussion about general transformer adaptation beyond t...	null	null	null	null	null	null
To Stay or Not to Stay in the Pre-train Basin: Insights on Ensembling in Transfer Learning	Accept (poster)	Summary: This paper studies an interesting dilemma -- literature reports that ensembles work better if you end up in different "basins", but pre-training often confines you to a single "basin". The question is therefore, what should you do? This paper answers that you should start out with a pre-trained model and _sta...	Rebuttal 1: Rebuttal: Thank you for the thoughtful review of our work! Please allow us to address your concerns and answer the questions. 1. Large-scale experiments. As you correctly pointed out, in the submission we focused on smaller problems to be able to accurately compare the results with both Local and Glob...	Summary: This paper studies ensembling in the transfer learning setting. This is interesting because pre-training is important for strong predictive performance, yet has been shown to pre-determine the final basin in the loss-landscape after fine-tuning, reducing diversity. The authors explore the behavior of snapshot ...	Rebuttal 1: Rebuttal: Thank you for the thoughtful review of our work! Please allow us to address your concerns and answer the questions. 1. Experiments on additional downstream tasks. We started our project using the same standard benchmark suite (Aircraft, Caltech101, Cars, CIFARs, etc.) used in the work you me...	Summary: This paper focuses on ensembling in transfer learning scenarios, i.e., constructing ensembles starting from only one pre-trained checkpoint. Empirical results employing Snapshot Ensembles (SSE; Huang et al., 2017) demonstrate that leaving a pre-trained basin can degrade the quality of the ensemble. Based on th...	Rebuttal 1: Rebuttal: Thank you for the thoughtful review of our work! Please allow us to address your concerns and answer the questions. W1. StarSSE name and connection to Local DE. Your assessment of the conceptual closeness of StarSSE and Local DE is absolutely correct, and we discuss it in the paper (lines 29...	Summary: The authors consider techniques to faciliatate more efficient ensembling when in the transfer learning context.They propose a novel cyclical learning rate method, StarSSE, and compare it to standard methods like Snapshot Ensembles (SSE). They demonstrate that leaving the pretraining basin leads to worse perfor...	Rebuttal 1: Rebuttal: Thank you for the thoughtful review of our work! Please allow us to address your concerns and answer the questions. W1. Unclear quantitative comparison of SSE and StarSSE ensembles. Thank you for pointing out that flaw in our presentation. In all the experiments StarSSE ensembles showed sig...	Rebuttal 1: Rebuttal: First of all, we want to thank all the reviewers for their constructive and valuable feedback! We have found all the suggestions very relevant and will definitely incorporate them into the paper. We are happy that our analysis of the effectiveness of local and semi-local ensemble methods in the t...	NeurIPS_2023_submissions_huggingface	2,023	null	null	null	null	null	null	null	null
Extracting Reward Functions from Diffusion Models	Accept (poster)	Summary: The study focuses on extracting a reward function by comparing low-reward and high-reward decision-making diffusion models, sharing the similar goal with inverse reinforcement learning. The researchers define a relative reward function of two diffusion models and establish its existence and uniqueness under ce...	Rebuttal 1: Rebuttal: Weaknesses/ comments: >IRL algorithms typically formulate the problem as a two-player minimax game, assuming bounded rationality of expert data through a maximum entropy framework. In stochastic environments, these methods often analyze causal entropy. However, it appears that such methods ha...	Summary: The authors propose a new problem setting, similar to IRL, of extracting reward functions from decision-making diffusion models: one expert and one base (non-expert), where the reward is defined as a function that can steer the base model to the expert through classifier guidance. Using this intuition, they d...	Rebuttal 1: Rebuttal: Weaknesses/ Notes > The experiments lack proper evaluation on classifier guidance except for the locomotion tasks. As stated in the problem setting, the objective is to extract the reward function that can steer the base model to the expert model through classifier guidance, so this evaluation...	Summary: This paper proposed a reward extraction method for reinforcement learning from diffusion models. Theoretical justification and experiments in diverse domains and settings were used to support the proposed method. I have no background in reinforcement learning and I think I'm not qualified to evaluate the cont...	Rebuttal 1: Rebuttal: Weaknesses/ Questions >What's the benefit of using diffusion models rather than other probabilistic generative models, e.g. VAE, FLOW and GAN, for reward function extraction? R1: Diffusion models produce samples by gradual denoising, as opposed to VAEs and GANs. Thus, they can be seen (infor...	Summary: This paper proposed a method for extracting reward functions from diffusion models, which can be used for sequential decision-making. The authors defined the notion of a relative reward function of two diffusion models and show conditions under which it exists and is unique. They also propose a practical learn...	Rebuttal 1: Rebuttal: Notes/ Weaknesses mentioned: > The paper assumes that the two diffusion models have the same architecture and are trained on the same data distribution. This may limit the applicability of the method to scenarios where the diffusion models have different architectures or are trained on diffe...	Rebuttal 1: Rebuttal: We thank all reviewers for their time and valuable feedback, and appreciate that reviewers found our work novel and relevant. We have addressed each reviewer's questions and concerns individually below. Please let us know if there are any further questions. Best wishes, The Authors	NeurIPS_2023_submissions_huggingface	2,023	Summary: The paper introduces the following problem: given two diffusion models, a base model and an expert model, extract a guidance function such that the output distribution of the base model steered with the guidance function would resemble the output distribution of the expert model. The authors first formalize th...	Rebuttal 1: Rebuttal: Notes / Weaknesses mentioned: > .. explanation or justification to a relative reward function .. R1: The motivation behind the setting is that the base model represents general, unguided exploratory behavior (which could intuitively be thought of as a high-entropy exploration policy in th...	Summary: The authors explore the applications for diffusion models in IRL, and state that they provided the first method for extracting relative reward functions (RRF) from a pair of diffusion models, where the models differ in quality/performance. They propose a simple but intuitive idea: the difference in score func...	Rebuttal 1: Rebuttal: Notes / Weaknesses mentioned: > How useful is “extracting relative reward functions from two diffusion models”? Applications here don’t seem quite compelling. R1: Two potential use cases our work attempts to demonstrate are: (1) Extracting reward functions from decision-making diffus...	null	null	null	null
Policy Gradient for Rectangular Robust Markov Decision Processes	Accept (poster)	Summary: The paper investigates policy gradient methods in the context of unknown transition kernels in Markov Decision Processes (MDPs). The authors focus on a robust approach, where the uncertain transition kernel is assumed to lie within a specific uncertainty set, known as a rectangular uncertainty set. The authors...	Rebuttal 1: Rebuttal: >"In Section 3.2.1 it is written that the radius of the uncertainty set should be “small enough” such that transition kernels of the form are well defined. This, however, requires the nominal transitions kernel to be strictly positive everywhere...this modelling seems highly restrictive" We apo...	Summary: This paper proposes the robust policy gradient (RPG) algorithm as an efficient approach to solve rectangular RMDPs. The results are based on an explicit characterization of $\ell_p$-ball-constrained RMDPs, where the worst-case kernel can be viewed as a rank-one perturbation of the nominal kernel and the worst-...	Rebuttal 1: Rebuttal: > "numerical experiments seem relatively crude and incomplete" For each experiment, we fixed a state-space size and an action space size. An environment was generated by readily sampling a transition kernel and a reward matrix with proper dimensions. This was taken as an input to RPG or LP-RPG...	Summary: The paper presents a policy gradient method for solving rectangular-robust MDPs. They propose closed forms for the worst-case reward and transition models. They also present robust Q-value forms that can be efficiently updated by a proper Bellman operator. Furthermore, they show an expression for the robust po...	Rebuttal 1: Rebuttal: We thank the reviewer for the positive feedback. > "slightly increase external references and elaborate more to help (an otherwise unfamiliar) reader follow through the content" Thank you for the suggestion. We will go over the text thoroughly to elaborate more on the wording and the mathemat...	Summary: This paper develops policy gradients for robust MDPs, and proves several interesting properties of the worst-case kernel and occupancy measure along the way. The paper also shows that the robust Bellman operator is just the vanilla Bellman operator, with regularization of the vector's generalized variance func...	Rebuttal 1: Rebuttal: We first thank the reviewer for the constructive remarks. > "Notation is a bit hard to follow and is defined all over the place throughout section 3. It would be helpful to have a centralized place where all notation is defined." Thank you for the suggestion. Although we tried to centralize i...	null	NeurIPS_2023_submissions_huggingface	2,023	Summary: This present studied robust policy gradient to solve rectangular robust Markov decision process, where the authors provide a closed-form expression for the worst case of reward model and transition kernel. The provided explicit form of the robust policy gradient can be estimated from data with the same time co...	Rebuttal 1: Rebuttal: > "Leveraging the method to derive a computationally efficient and practical algorithm in continuous policy optimization" We thank the reviewer for the constructive remark. In fact, one of the goals in this work is to develop building blocks for robust continuous control. The derivation of an...	null	null	null	null	null	null
Understanding the Limitations of Deep Models for Molecular property prediction: Insights and Solutions	Accept (poster)	Summary: The authors systematically study the performance of deep vs. non-deep machine-learned models in the setting of molecular property prediction, assessing the impact of input representations and models on 15 representative datasets taken from MoleculeNet, a standard benchmark in molecular property prediction. T...	Rebuttal 1: Title: Response to Reviewer SHsm (1/2) Comment: > There is no discussion of works actually related to the topic of the paper -- that is, previous work that attempts to benchmark a variety of methods. Thanks for your valuable feedback! We have discussed previous works that benchmark various models in Line...	Summary: - The paper evaluates various (non-)deep models for molecular property prediction. Tree models using molecular fingerprints as inputs are shown to perform better than other competitors. - The paper claims also to contain an extensive empirical investigation into the unique patterns of molecule data and inducti...	Rebuttal 1: Title: Response to Reviewer rZmN (1/3) Comment: > *I do not fully understand why this paper was not submitted to the datasets & benchmark track, the research part seems minor (large parts are in the appendix), l67ff.: We provide the most comprehensive benchmark on MPP tasks to date and expose the limitation...	Summary: This paper investigates why classical ML models (i.e. anything except neural networks) often outperform neural networks on molecular property prediction tasks. The paper first presents a large study evaluating many deep and non-deep methods on a variety of datasets (seems to be mostly MoleculeNet but I was un...	Rebuttal 1: Title: Response to Reviewer ah6B (1/3) Comment: > In section 3, "irregular data patterns" is not really defined: what makes a molecule pattern "irregular"? The other 2 conclusions from this section are supported but just confirm what is in numerous previous works. In our study, “irregular data patterns”...	Summary: The authors demonstrate, through a large number of experiments across many datasets and tasks, that tree-based models on molecular fingerprints consistently outperform deep learning models in molecular property prediction, and attribute this to 1) the "non-smooth" nature of molecular labels (e.g. the existence...	Rebuttal 1: Title: Response to Reviewer 1yt1 (1/2) Comment: > Some tables and figures could be significantly improved to make them more easily comprehended. Thank you for your valuable advice, which will undoubtedly improve the quality and clarity of our manuscript! We promise to refine the tables and figures follow...	null	NeurIPS_2023_submissions_huggingface	2,023	Summary: The authors compared various deep learning and traditional non-deep machine learning model performance on multiple molecular property prediction tasks. Their findings indicate that deep models tend to underperform non-deep ones due to the irregular patterns of molecular data and relatively small data size. To ...	Rebuttal 1: Title: Response to Reviewer zBbT (1/3) Comment: > Insufficient novelty. The proposed methodology is well-known (e.g., NeRF, Fourier features, etc.). Thanks for your helpful reviews! We have addressed the novelty of our proposed methodology, IFM, by highlighting the key differences between our method and ...	null	null	null	null	null	null
Latent Diffusion for Language Generation	Accept (poster)	Summary: This work proposes a new type of diffusion language model based on latent representations. The proposed approach leverages the continuous representation space produced by pre-trained encoder-decoder language models (BART and FLAN-T5). To embed the variable-length encoder features, the author(s) propose to adop...	Rebuttal 1: Rebuttal: Re: Long-term Impact We appreciate the reviewer raising the important concern about the long-term impact of our approach. The reviewer is correct that autoregressive transformers have continued to advance rapidly, demonstrating exceptional capabilities in a variety of language tasks. Howeve...	Summary: The paper proposes a latent diffusion model for text generation, by stacking pre-trained powerful language models and fine-tuned projection layers and diffusion networks together. The input is first encoded into hiddens, then the diffusion model generates the hidden used as the condition of the language decode...	Rebuttal 1: Rebuttal: We discuss our contribution in our overall response. We would like to emphasize that at inference-time, there is no encoder output. Therefore the diffusion model is not providing a projection of the encoder output, but sampling novel latent codes. The diffusion network iteratively refines a laten...	Summary: The paper demonstrates that encoder-decoder language models can be used to efficiently learn high-quality language autoencoders and to learn continuous diffusion models in the latent space of language autoencoders so that continuous latent representations of natural language can be sampled using pre-trained de...	Rebuttal 1: Rebuttal: Motivation: As noted in our overall response, a key motivation of our work is exploring how diffusion models can provide complementary strengths to autoregressive generation for text. Our approach aims to combine the benefits of both methods by learning the diffusion model in the continuous l...	Summary: This paper introduces a novel approach to learning diffusion models in the latent space of an encoder-decoder language autoencoder. The proposed latent diffusion model utilizes continuous latent codes to generate text through sampling, leveraging a pre-trained decoder. The effectiveness of this approach is dem...	Rebuttal 1: Rebuttal: Novelty While we agree latent diffusion is commonly used for images, the adaptation to language is non-trivial and presents unique challenges given the discreteness of the domain. Latent diffusion models are reliant on the strength of the underlying autoencoder, and we introduce a method for ...	Rebuttal 1: Rebuttal: We would like to thank the reviewers for their thoughtful feedback on our paper. In particular, we appreciate Reviewer SqZp highlighting that the “utilization of pretrained encoder-decoder models to establish a feature space is a clever and efficient approach”. We also thank Reviewer 5zaG for not...	NeurIPS_2023_submissions_huggingface	2,023	Summary: This paper introduces a latent diffusion model for natural language generation, utilizing a continuous diffusion framework. The authors leverage pretrained encoder-decoder models, such as BART and Flan-T5, to establish a continuous feature space. They propose a Perceiver Resampler as a compression module to ha...	Rebuttal 1: Rebuttal: Multilingual Evaluation: We thank the reviewer for their thoughtful feedback on our paper. We agree that evaluating our approach on a multilingual dataset would further demonstrate the versatility of our approach. The extension of our method to such settings is straightforward given the avai...	null	null	null	null	null	null
Robust low-rank training via approximate orthonormal constraints	Accept (poster)	Summary: This work demonstrates that conventional low-rank training methods suffer from poor adversarial robustness, which is hypothesized to stem from ill conditioning of weight matrices during training. The authors address this by constraining low-rank weight matrices to be approximately orthogonal. Strengths: I fo...	Rebuttal 1: Rebuttal: Thank you for the encouraging feedback. We have now implemented new experimental evaluation with PGD attack and on one additional network (wideresnet) on both cifar10 and cifar100. All new results are available in the pdf attached to the 'general rebuttal comment'. Re the specific questions: > T...	Summary: This paper proposes a training approach that can improve the robustness of low-rank DNN models from scratch. This method enforces the low-rank components onto the Stifel manifold and ensures well-conditioning to achieve robustness. Experimental results show that the proposed training method demonstrates robust...	Rebuttal 1: Rebuttal: The reviewer claims that the paper is incremental because: (a) condition number as a measure of the robustness of compressed models is used already in [R1], and (b) the robustness of low-rank NNs is investigated in [R2]. We thank the reviewer for pointing these two papers out which we had missed....	Summary: In this paper, the authors propose a low-rank training algorithm that incorporates approximate orthonormal constraints. Through a comparative analysis with previous methods for training orthonormal neural networks, the experimental results demonstrate the superiority of the proposed algorithm. Strengths: 1. A...	Rebuttal 1: Rebuttal: Thank you for your feedback, we provide below our response addressing the raised questions > Some terminologies are not pre-defined. For instance, concepts like the Stiefel manifold and tangent space may be unfamiliar to readers who have not previously explored low-dimensional spaces. Moreover, i...	Summary: Some work has used low-rank matrix factorizations in order to reduce training and inference costs. However, the authors demonstrate that these low-rank models are actually more brittle, i.e., susceptible to adversarial perturbations -- they tend to have large/exploding singular values. Consequently, the author...	Rebuttal 1: Rebuttal: Thank you very much for your encouraging feedback. Below we provide our response to your weak points and questions. > Would be good to substantiate "improving the network robustness without affecting training nor inference costs" -- it's not obvious this is true to me since the method involves,...	Rebuttal 1: Rebuttal: We thank all the reviewers for their valuable feedback and useful comments and criticism. We provide a detailed response to all weak points and questions in the individual rebuttal blocks below. We use this general rebuttal block to report additional evaluation in response to a number of reviewers...	NeurIPS_2023_submissions_huggingface	2,023	Summary: A brief summary of the paper is as follows: - Motivation: Make deep neural networks more efficient while maintaining robustness to adversarial attacks. - The authors propose a combined low-rank and approximate orthonormalisation constraint to help simultaneous model conditioning and compression during training...	Rebuttal 1: Rebuttal: Thank you very much for your useful feedback. Below we provide a detailed response to all the raised criticism. We notice that all raised weaknesses, except for one point concerning a missing comparison/reference with paper [1], are entirely related to the missing concluding section and some mis...	null	null	null	null	null	null
Variational Gaussian processes for linear inverse problems	Accept (poster)	Summary: This paper analyzes posterior contraction rates for the inducing variables variational Bayes approximation for Gaussian process. In particular, mildly and ill-posed inverse problem settings, which are characterized by profiles of eigenvalues of measurement operators, are considered. For the both cases, the pos...	Rebuttal 1: Rebuttal: We thank the reviewer for the summary and the question on adaptivity, which is of high importance in practice. As it is common in the literature, our first article focuses on the non-adaptive case when the function space is known and derived guarantees in this setting. The next, natural, follow ...	Summary: This paper presents a theoretical analysis of the posterior contraction rates for the variational posterior in the context of linear inverse problems. Specifically, the variational posterior is obtained using the sparse Gaussian processes approach, with inducing variables learned through KL divergence minimiza...	Rebuttal 1: Rebuttal: We thank the referee for his/her interest in the paper and the constructive comments and suggestions. ## Theoretical literature on frequentist-Bayes analysis The monograph from Richard Nickl is multiscale analysis applied to non-linear inverse problems. This is a completely differnt techniques b...	Summary: This paper analyses posterior contraction rates in Variational GP for mildly and severely ill-posed linear inverse problems. The authors prove that under certain conditions variational posterior achieves same contraction rates as the true posterior. Moreover, they show that given enough inducing points (in spe...	Rebuttal 1: Rebuttal: Thank you for your interest in our paper. We acknowledge the brievity of the experimental part in the main paper due to the space limitations. Please note that our work besides proposing a new methodological approach also provides theoretical guarantees in general settings and considers three sp...	Summary: This theoretical paper addresses the problem of inferring a function $f$ when $y(x_i) = \[Af\](x_i) + \epsilon_i$ with $i=1, ...,n $ and $A$ is suitably defined linear operator with $\epsilon_i$ being i.i.d Gaussian noise. In words this is a linear inverse problem. A (standard) Bayesian approach is adopted ...	Rebuttal 1: Rebuttal: We would like to thank the reviewer for carefully reading our paper and for the helpful comments. ## Assumptions on the GP kernel We thank the referee for this excellent comment. It is indeed standard in the literature to model covariance functions in accordance to the linear operator $\mathcal...	Rebuttal 1: Rebuttal: # General rebuttal: We would like to thank the reviewers for carefully reading our manuscript, for acknowledging the importance of the topic ("important and challenging topic" - TgVX) and that the examples are " interesting and informative" (TgVX). We are happy to hear that they found the paper ...	NeurIPS_2023_submissions_huggingface	2,023	Summary: The authors propose an inducing-point variational Bayes approximation for linear inverse problems. They study the frequentist properties of the arising variational posterior approximation. Specifically, they prove contraction rates for the variational posterior, both in the mildly and severely ill posed case. ...	Rebuttal 1: Rebuttal: We would like to thank the reviewer for his/her interest in our work and the suggestions. ## Literature review While the three papers on variational Bayes published in AoS in 2020 are indeed all cited in our article [2,54,56], they consider direct problems, while our focus is on ill-posed inve...	null	null	null	null	null	null
Boundary Guided Learning-Free Semantic Control with Diffusion Models	Accept (poster)	Summary: This paper introduces the BoundaryDiffusion method for semantic image editing with frozen pre-trained Denoising Diffusion Models. The authors provide an analysis showing that the inverted latent variables have a different distribution from the prior Gaussian. From this observation, this paper suggests a method...	Rebuttal 1: Rebuttal: We deeply appreciate the feedback regarding the theoretical part of the distance effect. --- Comment 1: BoundaryDiffusion does not require training a neural network but requires the linear SVM classifier optimized from the annotated images. Response to Comment 1:     W...	Summary: The authors argue they can disentangle image attributes in the e and h-level latent spaces at the discovered mixing step $t_m$. The mixing step can be determined by "the largest radius shift (e.g., 4)" based on "the mean of the squared distance of a latent sample from the origin (and following the generati...	Rebuttal 1: Rebuttal: We have submitted our anonymized code to AC to eliminate the reproducibility concerns of the reviewer, and ensure that the reported results and performance are easily reproducible. We are also thankful for the careful reading and helpful suggestions on the writing, such as indicating the speci...	Summary: This paper proposes a method for semantic control in image editing using pre-trained denoising diffusion models (DDMs) without the need for additional network modules. The method involves optimizing the denoising trajectory solely via frozen DDMs and exploring the critical step in the denoising trajectory that...	Rebuttal 1: Rebuttal: We deeply appreciate the acknowledgement and the insightful questions, which we would like to include the clarifications in our final paper. --- - Comment 1: In this paper, certain assumptions are made about the structure of the latent space which are difficult to verify or justify. Hierarchi...	Summary: This paper proposes Boundary Diffusion, a method for controlling the output of frozen pre-trained diffusion models. Boundary Diffusion contains two parts: (1) an analysis from the characteristics of Gaussian distribution to determine the key mixing step (2) a boundary guidance method to control the output of t...	Rebuttal 1: Rebuttal: We are very grateful for the valuable comments, especially with the suggestion to include more experimental results to further enhance our paper. --- - Comment 1: Baseline comparison with the classifier guidance. - Response to Comment 1: We appreciate the comment. As suggested, we i...	Rebuttal 1: Rebuttal: ### Overall responses to all reviewers and ACs We thank all reviewers and ACs for the valuable feedback and for initiating insightful discussions. We also appreciate the acknowledgment of our work from both theoretical and experimental aspects, with the *theoretical analysis to better understand ...	NeurIPS_2023_submissions_huggingface	2,023	null	null	null	null	null	null	null	null
Zero-sum Polymatrix Markov Games: Equilibrium Collapse and Efficient Computation of Nash Equilibria	Accept (poster)	Summary: Cai (2016) proved that in normal-form polymatrix games the set of CCEs corresponds with the set of NEs. This paper provides a similar result for polymatrix, switching controller, Markov games. After introducing the formalism used throughout the paper, the main result is presented for both fixed horizon and in...	Rebuttal 1: Rebuttal: We thank the reviewer for their comments and positive evaluation of our work. We commit to improving our text using the reviewer's insightful and helpful suggestions. We will try to answer the reviewer's concerns one by one. Please also see the global rebuttal response for further discussion of th...	Summary: This paper defines a class of mulit-agent Markov games called zero-sum polymatrix Markov games, which is a generalization of the zero-sum polymatrix normal-form games. Specifically, it defines a class of Markov games where each state is a zero-sum polymatrix game. The main results of the paper shows that: in b...	Rebuttal 1: Rebuttal: We would like to thank the reviewer for the time and effort they put into reviewing our work. Further, we would like to thank them for positively assessing our work. We believe that their suggestions and comments are important in improving the readability and quality of our text. >* The swtiching...	Summary: This paper considers the problem of computing approximate Nash equilibria in Markov (aka stochastic) games. It is known that such equilibria can be computed efficiently in the non-stochastic setting when the game is zero-sum polymatrix. The authors show that approximate equilibria can also be computed efficien...	Rebuttal 1: Rebuttal: We thank the reviewer for their positive evaluation of our work, their thorough read,ing and very diligent editing comments. Your comments are very valuable in improving the quality of our text. We will most certainly address those in our revised text. Let us try to answer your questions. > Ques...	Summary: This paper explores a specific class of Markov games called zero-sum polymatrix Markov games, a stochastic generalization of one-shot polymatrix Markov games, and shows how an ϵ-approximate Nash equilibrium can be efficiently computed. The authors show equality between the set of coarse-correlated equilibria a...	Rebuttal 1: Rebuttal: We thank the reviewer for their challenging questions, constructive comments, and of course their time. Due to space constraints, we cannot answer as detailed as we would like to all of your questions. Please, see the global rebuttal for some further elaboration on our answers. > Solving nonzero-...	Rebuttal 1: Rebuttal: The authors would like to thank the reviewers for their constructive comments, their positive assessment, and their valuable time. Surely, the reviews will significantly improve the readability of our work. We apologize for the typos and, at times, parts of our writing might have been confusing. ...	NeurIPS_2023_submissions_huggingface	2,023	null	null	null	null	null	null	null	null
Monte Carlo Tree Search with Boltzmann Exploration	Accept (poster)	Summary: This paper introduces two variants of MCTS designed to perform more exploration during search than the commonly used UCT search policy, while resolving an issue with a previously introduced MCTS variant, Maximum ENtropy Tree Search (MENTS), which was designed to explore more than UCT, but which does not conver...	Rebuttal 1: Rebuttal: Thank you for taking the time to review and providing us with helpful comments. > The motivation for increasing exploration during planning/search could be better explained. Thank you for the suggestions on clarifying the motivations for using simple regret and increasing exploration. We will ma...	Summary: This paper proposes two new algorithms, Boltzmann Tree Search (BTS) and Decaying Entropy Tree Search (DENTS), based on the Maximum Entropy Tree Search (MENTS) method. The authors prove that in certain situations, MENTS fails to converge to the optimal policy. They analyze BTS and DENTS using simple regret and ...	Rebuttal 1: Rebuttal: Firstly, thank you for your review and comments to help us improve our paper. > The Komi in 9x9 Go is usually set to 7 (a balanced setting for both players). Black will have an advantage when Komi is set to 6.5. We chose a komi of 6.5 because it is commonly used by humans in 9x9 Go and a half-in...	Summary: << I have read the authors' rebuttal and have raised my score based on the discussion >> The paper addresses challenges in online tree-based search algorithms, such as UCT, which have shortcomings such as being prone to local optima and a slower convergence rate to the optimal action. An alternative approach,...	Rebuttal 1: Rebuttal: Firstly, thank you for reviewing our work and giving us valuable feedback. > Q3: To better understand the influence of hyperparameters on BTS and DENTS and gauge their robustness, would it be possible for the authors to compare their performance for varying values of the alpha parameter? We vari...	Summary: The paper presents two MCTS based algorithms called Boltzmann Tree Search (BTS) and Decaying ENtropy Tree-Search (DENTS). These algorithms overcome the limitation of two state-of-the-art algorithms pUCT and MENTS of under and over exploration respectively. Further, the paper provides simple “simple regret” an...	Rebuttal 1: Rebuttal: Firstly thank you for taking the time to review our work and provide helpful feedback. > The motivation for analyzing simple regret instead of cumulative regret is not clear. This limits the contribution of the proposed approach to some extent. Cumulative regret is more appropriate in situations...	Rebuttal 1: Rebuttal: We have attached a pdf with updated results following comments from the reviewers. In Figure 1 of the attached pdf we have updated results for DENTS running under the $\beta(m)=\alpha$ configuration in both the D-chain and Frozen Lake environments. These show empirically that DENTS can still mat...	NeurIPS_2023_submissions_huggingface	2,023	null	null	null	null	null	null	null	null
Frequency Domain-Based Dataset Distillation	Accept (poster)	Summary: This paper proposes a frequency-domain dataset distillation approach that can help to select the most important information in images to help distillation. To do this, the approach starts with a DCT to the input $x$ and then performs loss matching using existing dataset distillation approaches. Experiments dem...	Rebuttal 1: Rebuttal: We really appreciate your constructive review on our manuscript. $\textbf{[W1. A request for experiments on a large volume dataset]}$ We appreciate the valuable comments of the reviewer. The volume of the dataset could increase by two factors: either 1) the dimension of each instance in the data...	Summary: The paper proposes a novel way (code named FreD) to perform parameterization on dataset distillation tasks. Different from other methods which all operate on the spatial domain, this paper proposes to convert the spatial domain into frequency domain which are more concentrated and then perform distillation on ...	Rebuttal 1: Rebuttal: We really appreciate your constructive review on our manuscript. $\textbf{[W1. Differences in IPC settings for each table]}$ We apologize for any confusion caused to the reviewer by utilizing different IPC values across each table. In the case of HaBa, which is FreD's main baseline, conducting a...	Summary: This work introduces Frequency-based Dataset Distillation as a new means of parameterizing the distilled dataset that requires a much smaller memory footprint. By learning the distilled dataset in the frequency space (defined by the discrete cosine transform by default), a binary mask based on the class-wise E...	Rebuttal 1: Rebuttal: We really appreciate your constructive review on our manuscript. $\textbf{[W1. Specifications of the distillation loss utilized in the experiment and figures]}$ Thanks for pointing out our lack of explanation. Figure 2 in the main paper utilized DM as a distillation loss to compute the gradient....	Summary: This paper introduces a new parameterization method for dataset distillation, called FreD, that utilizes the frequency domain to distill a compact, synthetic dataset from a large-sized original one. This approach differs from conventional methods that focus on the spatial domain. It employs frequency-based tra...	Rebuttal 1: Rebuttal: We really appreciate your constructive review on our manuscript. $\textbf{[W1. Comparison with [1]]}$ Thank you for your valuable suggestion. [1] introduces a new parameterization method called memory addressing (MA). The key concept is to create a common representation by encapsulating the feat...	Rebuttal 1: Rebuttal: $\textbf{[Q1. The impact of utilizing different frequency transforms on FreD]}$ We utilized three frequency transforms as the reviewers noted: DCT, DFT, and DWT. We especially want to highlight the energy compaction property of DCT, where most of the signal information tends to be concentrated in...	NeurIPS_2023_submissions_huggingface	2,023	null	null	null	null	null	null	null	null
Alternation makes the adversary weaker in two-player games	Accept (spotlight)	Summary: The authors discuss a model of no-regret learning that simulates the use of alternating regret updates in a two-player game. In this model, the authors show that the classical lower bound of $\Theta(\sqrt{T})$ regret does not hold, and in fact, it is possible to achieve $\tilde O(T^{1/3})$ regret. Strengths...	Rebuttal 1: Rebuttal: Thank you very much for the work and the valuable comments. We will incorporate your comments/suggestions in the revised version of our work. 1. Thanks for spotting the typos! -Yes Remark 2.5 refers to Algorithm 4 - $\gamma$ is not missing from line 6 of Algorithm 6 (since we have $R(x)/\gamma$)....	Summary: This paper consider an online linear optimization (OLO) problem motivated by alternating game play in two-player games. Specifically, after the learner chooses $x^t$ and the adversary selects loss vector $c^t$, the learner suffers loss $(c^t + c^{t-1})x^t$ in the alternating OLO instead of $c^t x^t$ in the sta...	Rebuttal 1: Rebuttal: Thank you very much for your effort and your positive feedback. We will incorporate the minor typos that you pointed to in the revised version of our paper. Questions The results of the paper...in two-player games.: We believe that once both agents use the same algorithm (e.g. Algorithm~2) b...	Summary: The paper considers online linear optimization where losses are given by $x_t^{\top} (c_t + c_{t-1})$, where $c_t$ is the adversary’s choice at each round. The authors show that the standard $\Omega(\sqrt{T})$ lower bound no long holds, and give a $O(T^{1/3})$ regret algorithm over the simplex as well as a $O(...	Rebuttal 1: Rebuttal: Thank you very much for your work and valuable comments. Up next we answer the main points raised in the weakness part. We hope that the following discussion (that we will incorporate in our revision) will help reconsidering your suggestion. 1. To start...turn it is?: We remark that alternating...	Summary: The paper introduces a variant of the usual online learning model with linear losses, whereby the environment is constrained to produce losses of a specific form. Such a form appears naturally when studying alternating no-regret dynamics in games. The authors establish that this restriction weakens the environ...	Rebuttal 1: Rebuttal: Thank you very much for your work and your positive feedback. Thank you also for spotting the typos. Up next, we answer your main points that we will also incorporate in our revision. 1. My only reservation...used in the paper \ \ Alternating self-play has proven a valuable optimization tool i...	null	NeurIPS_2023_submissions_huggingface	2,023	null	null	null	null	null	null	null	null
SLaM: Student-Label Mixing for Distillation with Unlabeled Examples	Accept (poster)	Summary: This paper aims to systematically utilize the noise of teacher predictions in a knowledge distillation setting comprised of both labeled and unlabeled samples. The proposed method, SLaM linearly mixes the student predictions with the noise of the teacher predictions, thereby hoping that the student will learn ...	Rebuttal 1: Rebuttal: >All empirical results presented are based on "[...] the best test accuracy over all epochs... I would strongly suggest reporting something like the test performance at the last epoch to avoid overfitting the test set. ... completely independent test sets would be included. We agree with the re...	Summary: The paper introduces SLaM, a method for knowledge distillation using unlabeled examples. Its main contribution lies in the development of an efficient and data-agnostic approach that enhances student model performance in scenarios with limited labeled data but abundant unlabeled data. SLaM accomplishes this by...	Rebuttal 1: Rebuttal: >... but it also seems hard to guarantee the proposed method can get the pseudo-labels without noisy. Can the proposed method ensure that the pseudo-labels generated by SLaM are noise-free, thereby addressing the limitation discussed in the paper? How and why?* SLaM effectively improves the qua...	Summary: In their paper, the authors propose a knowledge distillation procedure that augments pseudo-labels predicted by a teacher models by its uncertainty, leading to a uncertainty-aware student learning routine. This is realized by a novel loss formulation that mixes student predictions with quantified uncertainty a...	Rebuttal 1: Rebuttal: >.. a more rigorous ablation study of the influence of the uncertainty quantification would be insightful. ... obtaining $\alpha(x)$ and $k(x)$ from an oracle with “perfect” estimations, and gradually contaminating these estimates. We agree with the reviewer that a "controlled" experiment by ad...	Summary: The paper attempts to resolve the problem of noisy teacher labels for unlabeled examples in knowledge distillation. Proposed Student-Label Mixing (SLaM) to introduce noise into student prediction when matching to teacher’s pseudo-label on unlabeled data. Experiments on standard benchmarks were used to demonst...	Rebuttal 1: Rebuttal: > The paper lacks the explanation on the intuition of the mixing solution on formula (1), w ... What is the intuition of the second term on formula (1)? Short Answer* * The high-level intuition of equation (1) appears in Lines 212-226 of the manuscript. * In Appendix C of the supplementary m...	Rebuttal 1: Rebuttal: ## General Response We want to thank all reviewers for taking the time to read our manuscript carefully and for providing constructive and insightful feedback. We are very encouraged by the positive comments of the reviewers on tackling a well-motivated and important problem (Reviewers E5Wk,...	NeurIPS_2023_submissions_huggingface	2,023	null	null	null	null	null	null	null	null
Online Learning under Adversarial Nonlinear Constraints	Accept (poster)	Summary: This paper studies the online convex optimization subject to certain constrains and assumes that such constrains are slowly changing. The main contribution is a $O(\sqrt{T})$ regret bound under such constrains. Strengths: This paper appears to prove some interesting technical results. However, due to its uncl...	Rebuttal 1: Rebuttal: Thank you for your valuable suggestions and important remarks. We answer the raised questions below. Weaknesses 1. "The purpose of introducing the constraints $\mathcal{C}_t$ for $t<T$ is unclear. ... Are you assuming that $x_t$ must be selected from $\mathcal{C}_t$? ..." CVV-Pro optimizes ...	Summary: The paper studies an online learning problem in which the learner aims at maximizing their reward subject to adversarial time-varying nonlinear constraints. The proposed algorithm guarantees a regret upper bound of order $T^{1/2}$ and converges to the feasible set at a rate of $T^{-1/2}$. Intrestingly, the fea...	Rebuttal 1: Rebuttal: Thank you for your valuable suggestions and constructive feedback. We answer the raised questions below. Weaknesses 1. "I would add to the related works, when discussing performance metrics, also some pointers to work on online learning with long-term constraints stemming from the literature...	Summary: This paper considers an online constrained learning scenario, where partial information about the objective and the constraints, which could potentially be time-varying, is available at each time slot. The proposed algorithm, referred to as "Constraint Violation Velocity Projection," constructs a velocity poly...	Rebuttal 1: Rebuttal: Thank you for your valuable suggestions and important remarks. We answer the raised questions below. Weaknesses "This is mostly a theoretical piece of work, but it is slightly unclear to me what the broader implications of the proposed algorithm would be. The two-player game experimental res...	Summary: The paper studies a constrained online learning problem. The considered problem has a notion of constraint violation which is different from previous online learning problems. Also, the environment chooses not only a cost function but also a constraint function at each round. The algorithm can observe the cost...	Rebuttal 1: Rebuttal: Thank you for your valuable suggestions and important remarks. We answer the raised questions below. Weaknesses 1. "The time-varying constraints are also considered in [1,2,3]. Thus, it is not appropriate to claim that the setup differs from traditional OCO in terms of the time-varying const...	Rebuttal 1: Rebuttal: We would like to thank the reviewers for the time they spent reviewing our paper and for their insightful comments. Several reviewers recognized the novelty of the setting studied and the theoretical guarantees provided, as well as the applicability of our algorithm to related problems. More spec...	NeurIPS_2023_submissions_huggingface	2,023	Summary: CVV-Pro is an online learning framework for non-stationary data streams, with support for handling constraints that may be unknown, adversarially time-varying, and nonlinear. It doesn't assume that the feasible set is known, but proves that the algorithm still converges to it a rate of 1/sqrt(T). The method i...	Rebuttal 1: Rebuttal: Thank you for your valuable suggestions and constructive feedback. We answer the raised questions below. Weaknesses 1.1 "The paper is not easy to read ... can only guess that it is some sort of feasibility function." Following the reviewer's comments, we have reorganize and improved the pres...	null	null	null	null	null	null
Label Robust and Differentially Private Linear Regression: Computational and Statistical Efficiency	Accept (poster)	Summary: This work studies the problem of $(\varepsilon, \delta)$-differently private linear regression. The covariates are drawn from a Gaussian distribution $N(0,\Sigma)$ with unknown covariance $\Sigma$ and the random noise in the model is either Gaussian or heavier-tailed. Further, a small fraction of the labels ca...	Rebuttal 1: Rebuttal: We thank the reviewer for the constructive feedback and address each comment as follows. 1. Polynomial dependence on $\kappa$: We kindly direct reviewer to our general response for further details.  \   2. The GD-type of methods has inherent $\kappa$ dependence: Our work and the prior ...	Summary: This paper introduces the first computationally and statistically efficient algorithm for differentially private and label-corruption robust linear regression. In particular, they improve over the best sample complexity achieved by non-robust algorithms in terms of the condition number of the covariance matrix...	Rebuttal 1: Rebuttal: We thank the reviewer for the constructive feedback and address each comment as follows. 1. Polynomial dependence on $\kappa$: We kindly direct reviewer to our general response for further details.  \   2. Proposition 3.9: Compared to [11], our proof is a simpler version that only cove...	Summary: This paper focuses on differentially private linear regression with potential label corruption. The proposed algorithm is based on DP-SGD and introduces new components such as full batch-size GD and adaptive norm clipping. The paper achieves the best sample complexity under the assumption of sub-Gaussian data ...	Rebuttal 1: Rebuttal: We thank the reviewer for the constructive feedback and address each comment as follows. 1. Distributional requirement: We thank the reviewer for pointing this out, since this was an important contribution of this paper that was not clearly written in the paper. We focused on the subGaussian case ...	Summary: This paper presents a variant of the Differentially Private Stochastic Gradient Descent (DP-SGD) algorithm aimed at achieving efficient linear regression under differential privacy, even in scenarios where a portion of response variables is adversarially corrupted. The introduced methodology involves a full-ba...	Rebuttal 1: Rebuttal: We thank the reviewer for the constructive feedback and address each comment as follows. 1. Experiments: A primary motivation for this project is to understand theoretically the fundamental tradeoffs involved in privacy, accuracy, and robustness in solving linear regression. Our algorithmic contr...	Rebuttal 1: Rebuttal: We address common concerns regarding the significance of our $\kappa$ improvement. For linear regression without adversarial corruption or privacy, the dependence on $\kappa$ can be removed by simply applying an OLS solution. For linear regression with robustness and privacy simultaneously, there ...	NeurIPS_2023_submissions_huggingface	2,023	null	null	null	null	null	null	null	null
Understanding and Improving Ensemble Adversarial Defense	Accept (poster)	Summary: This paper first propose to understand ensemble adversarial defense from a theoretical perspective. Then the authors develop iGAT, including a distributing strategy is to encourage each base classifier to keep improving over regions that they are relatively good at classifying. This leads to a hard distributin...	Rebuttal 1: Rebuttal: We thank you for the insightful comments, and address them below. # Weaknesses * On Weakness 1. Both our ranking and regularization designs are motivated by Assumption 4.4 (``Acceptable Classifier'') which identifies sufficient conditions that should be satisfied by a base classifier in orde...	Summary: This paper derives provable error reduction when shifting from employing one neural network to an ensemble of two neural networks, which provides new insights into ensemble adversarial defenses. Based on the theoretical findings, the authors propose iGAT to narrow the subspace where all base classifiers miscla...	Rebuttal 1: Rebuttal: We thank you for the positive comments and the accurate summary of our work and address the comments by grouping them into three categories: weaknesses, questions, and limitations. # Weaknesses * On Weakness 1. Encouraged by your comment, we have conducted a new experiment to examine performa...	Summary: 1. This work aims at creating a theoretical framework to explain why an ensemble of MLPs can be more robust than the single components. 2. The assumptions under which the ensemble of MLPs have more robustness are given. The first set of conditions are on each individual model itself. The second set of assumpti...	Rebuttal 1: Rebuttal: We thank you for the insightful comments and questions and address them below. # Weaknesses * On Weakness 1. We appreciate the reviewer for pointing this out. By the time we submitted this paper, we were only able to develop the error theory for two base MLPs under those reported assumptions....	Summary: This paper investigates enhancing the performance of ensemble-based adversarial defense methods against adversarial attacks. Considering the effectiveness of ensemble-based adversarial defense methods have been proven empirically while the understanding of their mechanism is rather limited, authors first propo...	Rebuttal 1: Rebuttal: We thank you for the very positive comments and the accurate summary of our work. We address the comments by grouping them into three categories on questions, weaknesses, and limitations. # Answers * Q1. DVERGE performs proficiently on both CIFAR10 and CIFAR100. This can be credited to its ap...	null	NeurIPS_2023_submissions_huggingface	2,023	Summary: The ensemble strategy has gained popularity in the field of adversarial defense, wherein multiple base classifiers are trained cooperatively to protect against adversarial attacks. While empirical evidence supports its effectiveness, the underlying theoretical explanations for why an ensemble of adversarially ...	Rebuttal 1: Rebuttal: We thank you for the insightful comments and address them regarding the following three aspects. # Answers * Q1. On ImageNet Results: We thank the reviewer for the suggestion. If allowed by time and the computing resource available to us, we would very much like to report the ImageNet result...	null	null	null	null	null	null
PackQViT: Faster Sub-8-bit Vision Transformers via Full and Packed Quantization on the Mobile	Accept (poster)	Summary: The paper proposes a 4-bit vision Transformer and claims its efficient deployment on mobile CPUs. The attention module is quantized with mixed quantization schemes, i.e., inputs are quantized using different methods, logarithmic and uniform linear. To run 4-bit inference on CPUs, the paper also presents a manu...	Rebuttal 1: Rebuttal: ReW1.: (1) For efficient implementation of GeLU, we revisit I-GeLU [5] segmentation function with a=-0.2888, b=-1.769, x$\in$Z as Eq. 1 in our appended PDF. Within the (-2,2) data range, the value of I-GeLU is {-0.0257, -0.1152, 0, 0.5919, 1.3885} under the integer domain. So we directly use...	Summary: This paper is about developing a quantization method for ViT, including DeiT and Swin Transformer. The developed algorithm can be run on mobile phones. Experiments were conducted on several datasets, with comparisons to other quantization methods. Strengths: a new strategy for quantization of ViT, mainly for...	Rebuttal 1: Rebuttal: ReW1: Thank you so much for improving our paper presentation. We will rewrite the final paragraph of our introduction as follows: Our contribution can be categorized into four aspects, including motivations, optimization of task accuracy, optimization of hardware efficiency, and evaluation re...	Summary: This paper introduces a quantisation framework for end-to-end ViT inference on CPU devices. The main idea is to exploit the byte-level data parallelism present in the SIMD units of CPUs, and present a comprehensive set of techniques to enable low-latency ViT inference. The proposed framework includes two parts...	Rebuttal 1: Rebuttal: ReW1: Thanks for your questions helping us clarify the paper. (1) We have reported our results with QAT in the lower part of Table 1 in our paper. We also provide our analysis of each method in the ablation study, as shown in Table 3 of our paper. (2) Our approach is compatible with PTQ. We re...	Summary: This paper introduces an innovative framework to optimize the implementation of Vision Transformers (ViTs) on mobile devices using sub-8-bit quantization. This framework identifies and addresses key challenges such as long-tailed distribution and systematic channel-wise outliers within ViT dataflow. Through th...	Rebuttal 1: Rebuttal: ReW1: For more results on another task, e.g., object detection, please refer to Reviewer yEDd: ReW3 which can validate that our proposed method can be generalized to other image domains. 2. Our 4-bit packed multiplier can be executed on mainstream processors on edge platforms (e.g., mobile pho...	Rebuttal 1: Rebuttal: We first sincerely thank every reviewer for your insightful and constructive feedback. Then, we will answer the specific questions from each reviewer. We upload a pdf file with figures (Figure A,B,C) and equations (Equation 1,2) that we will present in our rebuttal. Pdf: /pdf/98832492b0ccb0ff467...	NeurIPS_2023_submissions_huggingface	2,023	Summary: This work started from analyzing the distribution of weights and activations of ViT models, identified several characteristics of critical layers, and applied Integer and hardware friendly techniques to address the issues. Furthermore, the author demonstrated a fully quantized ViT (including LayerNorm, Softmax...	Rebuttal 1: Rebuttal: ReW1: Our difference between Ref [31] can be categorized into four aspects, including motivations, optimization of task accuracy, optimization of hardware efficiency, and evaluation results. Motivation: We analyze data distribution within ViTs (long-tail distribution and outlier analysis), wh...	Summary: In this paper, the authors proposed PackQViT, a fast and accurate sub-8-bit quantization for running ViT on mobile devices. To maintain accuracy, the authors proposed different quantization schemes for activation and weights, and an outlier-aware training stage. For efficiency, they proposed fully integer soft...	Rebuttal 1: Rebuttal: ReW1: We appreciate your valuable comments since this allows us to clarify this phenomenon. The accuracy of the full precision model in Table 1 of our paper is sourced from the original paper of DeiT and Swin, both of which do not perform the DGD distillation techniques. Q-ViT proposes DGD dis...	null	null	null	null
Neural Multi-Objective Combinatorial Optimization with Diversity Enhancement	Accept (poster)	Summary: The paper presents a novel neural heuristic approach, called NHDE, for Multi-Objective Combinatorial Optimization (MOCO) problems. The proposed approach enhances the diversity of Pareto solutions by introducing a multiple Pareto optima (MPO) strategy and an indicator-enhanced Deep Reinforcement Learning (DRL) ...	Rebuttal 1: Rebuttal: Thanks for your valuable comments. Firstly, we clarify the scope of this paper, since there seems to be a misunderstanding by the reviewer. We focus on Multi-Objective CO (MOCO), a domain different from the Single-Objective CO (SOCO), even though the latter can employ diversity enhancement. MO...	Summary: This paper proposes a number of techniques to improve the diversity of solutions in neural multi-objective combinatorial optimization (MOCO): 1) solving preference-conditioned subproblems dependently by taking a subset of the current Pareto front as input to a graph attention-based policy network; 2) using the...	Rebuttal 1: Rebuttal: Thanks for the valuable comments. W1: Reduced duplicated solutions. We implicitly show the reduced duplicated solutions by the metric \|NDS\|, i.e., the number of non-dominated solutions. Empirically, more non-dominated solutions mean fewer duplicated solutions and fewer dominated s...	Summary: This work proposes NHDE, a novel neural heuristic with diversity enhancement method, to solve multi-objective combinatorial optimization (MOCO) problems. The main contribution of NHDE is to overcome the crucial duplicated solution issue for the widely-used decomposition-based method, and generate a set of dive...	Rebuttal 1: Rebuttal: Thanks for the valuable comments. W1: Solution Distribution. We agree with the reviewer that the scalarization may affect the solution distribution. Our NHDE can better alleviate this compared to pure decomposition-based methods for two reasons: * NHDE can generate more diverse solutions as v...	Summary: This paper focus on solving multi-objective combinatorial optimization by neural methods. The authors observe that decomposition-based neural MOCO methods may acquire similar solutions for neighboring subproblems. Therefore, to enhance the diversity of the non-dominated solution set, they introduce the hypervo...	Rebuttal 1: Rebuttal: Thanks for the valuable comments. To Question 1: The reasons that HGA can guide the diverse solution construction without explicitly considering the structure of the existing and current solutions are as follows. 1. Intuition of implicitly considering the solution structure. Actually, the pr...	Rebuttal 1: Rebuttal: Many thanks for all reviewers' constructive and valuable comments. Following their suggestions, we have made the following main revisions: 1. Experiments: We have provided more experimental results and analyses, including the solution distribution, runtime analysis, the evidence results about...	NeurIPS_2023_submissions_huggingface	2,023	null	null	null	null	null	null	null	null
Proportional Response: Contextual Bandits for Simple and Cumulative Regret Minimization	Accept (poster)	Summary: The paper proposes a novel contextual bandit algorithm that balances cumulative and simple regret guarantees using a surrogate objective and optimal cover. By constructing the conformal arms sets (CASs) which improve the coverage guarantees for the optimal policy, the algorithm efficiently finds the policy tha...	Rebuttal 1: Rebuttal: Q1: The results in [1] do not contradict Theorem 3. We will make this clear in our discussion. The log(T) cumulative regret guarantee in [1] relies on the linear realizability assumption, while our framework here addresses more general model classes where the instance-dependent cumulative regret b...	Summary: This paper studies contextual bandits and develops a flexible algorithm that can be adapted to both cumulative regret minimization and simple regret minimization. New techniques are developed to analyze the proposed algorithm, e.g., constructing the conformal arm sets. The authors also prove a negative result ...	Rebuttal 1: Rebuttal: Regarding claims of computational efficiency: We thank the reviewer for raising this point and apologize for the confusion. The term “computationally efficient” is indeed ambiguous. In the revised version, we will be more precise in our explanation by emphasizing that we only make $O(1)$ oracl...	Summary: This paper focus on the continues objective between simple regret minmization (SRM) and cumulative regret minimization (CRM) in contextual bandits, and propose an algorithm that can be either optimal in SRM, or optimal in CRM. Meanwhile, a new uncertainty quantification is propsed. Numerical simulations valida...	Rebuttal 1: Rebuttal: We thank the reviewer for their encouraging remarks. Regarding doubling epochs: We thank the reviewer for raising this point and will discuss doubling epochs as a limitation of our algorithm. To clarify, our algorithm does not assume a fixed given $T$. If we pause exploration in the middle of...	Summary: This paper provides a computational efficient algorithm for contextual bandit that is capable of trading off between near-optimal minimax cumulative regret guarantees and instance-dependent simple regret guarantees. Additionally, It handles model misspecification, works in the continuous arm setting, and provi...	Rebuttal 1: Rebuttal: Regarding theoretical guarantees and optimality of the achieved trade-off: The algorithm RAPR comes with a tuning parameter $\omega\in[1,K]$. This tuning parameter helps us trade-off guarantees on simple and cumulative regret. We discuss our theoretical results and recap the optimality of our ...	Rebuttal 1: Rebuttal: We thank all the reviewers for their valuable feedback. Regarding feedback on simulations: We ran an additional simulation to compare our algorithm (RAPR) against other baselines and to better understand the balance between the two objectives we study. On a randomly generated linearly realiza...	NeurIPS_2023_submissions_huggingface	2,023	null	null	null	null	null	null	null	null
Active Learning-Based Species Range Estimation	Accept (poster)	Summary: This paper tackles the problem of species distribution modeling from limited observation data. An active learning approach is proposed that makes the assumption that an unmapped species can be represented by a weighted combination of candidate species models (previously trained offline on a disjoint dataset). ...	Rebuttal 1: Rebuttal: [RtuW-1] The proposed approach is a relatively straightforward application of active learning beyond the introduction of the candidate models. Our approach consists of a novel exploration of active learning to the problem of efficiently estimating spatially varying quantities when there are ...	Summary: Paper Summary This paper tackles the data-driven estimation of geographical habitat area for a given species; that is, given a certain geographic location (i.e., a "point on the map"), the goal is to predict how likely a certain species is present in this location. Prior efforts train models in an "offlin...	Rebuttal 1: Rebuttal: [yX8t-1] Is the proposed approach only applicable to new species that are sufficiently similar to species for which data is readily available? If the candidate hypothesis set does not sufficiently cover the space of possible ranges for a held out test species of interest this will indeed imp...	Summary: The paper explores a method for geographical range estimation of a species. One of the primary challenges in range estimation is the scarcity and costliness of data collection, particularly for rare species. To address this issue, the paper introduces an active learning framework as a solution. The proposed m...	Rebuttal 1: Rebuttal: [B3F8-1] The paper does not explicitly highlight the unique machine learning challenges of this problem or how they connect to the technical contribution of the paper. As noted on line 118, when presence and absence data is available, the range estimation problem for a single species can be...	Summary: This paper proposes an active learning framework for estimating the geographical range of species from sparse observations. The core idea is to exploit a large set of pretrained range estimation models from many different species. This paper also provides extensive experiments including detailed ablation studi...	Rebuttal 1: Rebuttal: [yFB1-1] The technical novelty is limited. It is reasonable, yet straightforward, to exploit a set of pretrained range estimation models for training a model for a target species. Respectfully, we disagree. To the best of our knowledge, we are the first work to show that it is possible to m...	Rebuttal 1: Rebuttal: We thank the reviewers for their constructive comments and insightful questions. We appreciate the detailed reviews and words of support regarding the overall quality and clarity of the paper provided by each of the five reviewers. As there are no significant common concerns from the reviewers, we...	NeurIPS_2023_submissions_huggingface	2,023	Summary: This paper proposes to: 1) Use crowdsourced data to obtain species range maps for a large set of data-rich species, and then 2) Learn a linear combination of these maps to represent the range of other, data-poor, species, while minimizing the number of required presence/absence queries via active learning. St...	Rebuttal 1: Rebuttal: [WCfR-1] What other application areas could your method be applied to and is this a good fit for NeurIPS? This is an interesting question. Our approach is applicable to any spatially varying quantity where it is possible to obtain a hypothesis set over possible related predictions. Potentia...	null	null	null	null	null	null
Theoretical Analysis of the Inductive Biases in Deep Convolutional Networks	Accept (poster)	Summary: The paper theoretically analyses the sample complexity of CNNs as compared to fully connected networks. The authors find that while fully connected networks require quadratic samples with input dimension, the local connections of CNNs reduce the sample complexity to linear, while weight sharing further reduces...	Rebuttal 1: Rebuttal: We thank the reviewer for the encouraging and constructive comments. In the following, we address the specific questions. 1. Experimental evidences [W1, Q2] First, we have conducted numerical experiments and the results align well with our theoretical predictions. We refer to the **Global R...	Summary: The authors study inductive biases in deep CNNs. They analyse approximation abilities of these networks and compare the learning performances of CNNs for regression with those of locally-connected networks (LCNs) and fully-connected networks. They show that CNNs of depth O(\log d) is sufficient for achieving u...	Rebuttal 1: Rebuttal: We thank the reviewer for the encouraging and constructive comments. In the following, we address the specific questions. 1. Related Works. [W1] We thank you for pointing out these references. [Mao et al., 2023] considered various kinds of spatially sparse functions. [Chui et al., 2019] ...	Summary: This paper provides a fundamental analysis of the superior performance of CNNs, specifically focusing on their downsampling and multichanneling capabilities. Moreover, it establishes evidence with theoretical guarantees that CNNs exhibit efficient learning of sparse functions compared to FCNs, attributing this...	Rebuttal 1: Rebuttal: We thank the reviewer for the comments and questions. The following is our response. 1. Weakness. - Lack generalization analysis. We already provided generalization analysis in Section 4 for learning sparse functions and Section 5 for disentangling the inductive biases of weight shar...	Summary: This paper studies the inductive bias of deep convolutional neural networks (CNNs) in three distinct ways. With $d$ the input dimensions, it is shown that (i) depth $O(\log{d})$ is sufficient to achieve universality, i.e. the capacity to approximate any function of $d$ variables, when there is downsampling bet...	Rebuttal 1: Rebuttal: We thank the reviewer for the valuable comments and questions and we will revise our manuscript accordingly. In the following, we response to the specific weakness and questions. (1) CNN with non-overlapping pathces and the universality [Weakness 1; Question 6,7] First of all, we agree tha...	Rebuttal 1: Rebuttal: Global Response to All Reviewers We thank all reviewers for their constructive feedback. In our forthcoming revision, we will integrate these comments, paying special attention to an enriched discussion on the related works suggsted by the reviewers, in particular [Poggio et al., 2017]. In ...	NeurIPS_2023_submissions_huggingface	2,023	null	null	null	null	null	null	null	null
A Holistic Approach to Unifying Automatic Concept Extraction and Concept Importance Estimation	Accept (spotlight)	Summary: This paper proposes a unifying framework to study common concept-based explainability methods. In particular, the authors observe that all the methods boil down to dictionary learning in the latent spaces of a neural network. Based on this observation, they (1) compare the different concept extraction/discover...	Rebuttal 1: Rebuttal: We thank the reviewer for carefully reading and providing a constructive review with relevant remarks and suggestions that will enhance the quality of the paper. Please find below a point-by-point answer: * About the lack of analysis of the Strategic Clustering graph: The main goal of this ar...	Summary: This paper presents a novel framework for unifying existing post-hoc concept-based explainability methods in a cohesive, formalized form. Such a framework enables a principled study of key properties of state-of-the-art concept-based explainability methods, such as ACE, T-CAV, CRAFT, and ICE, and allows the au...	Rebuttal 1: Rebuttal: We are delighted to learn that the reviewer found the paper enjoyable and useful for the community. We extend our sincere gratitude to the reviewer for investing time and effort in providing a thorough review! Please find below a point-by-point answer: * **Reviewer: Some key and more recent meth...	Summary: In this paper, the authors propose a unified perspective on post-hoc concept-based explanation methods. The main intuition underlying the work revolves around the fact that typical concept-based explanations can be considered as a two-stage process whereby, initially, a concept vocabulary is learned and, lastl...	Rebuttal 1: Rebuttal: We thank the reviewer for the very nice review. Once again, we extend our appreciation for the time and the very comprehensive review. Please find below a point-by-point answer: * **Reviewer : 'Learning a dictionary of concepts in an unsupervised manner does not give any guarantee that the extra...	Summary: The paper provides a unifying perspective for the concept extraction / importance estimation method via the dictionary learning framework. Namely, depending on the constraints on $U$ and $V$, the framework includes ACE, PCA, CRAFT and ICE as special cases. Then, the paper compares three difference approaches, ...	Rebuttal 1: Rebuttal: We thank the reviewer for the time and the suggestions in the reviews. Please find below a point-by-point answer: * **Reviewer : 'I think it would be beneficial to show the failure case of the proposed method. Figure 1/2 show compelling results for three classes (Espresso, Zucchini, and Lemon) b...	Rebuttal 1: Rebuttal: The reviewers found the paper "extremely well-written" "thoroughly enjoyable to read" and appreciating the "non-standard approach" taken. Regarding the theoretical work, you found it “interesting and extensively curated”, with reviewers finding the use of matroids in proofs "*very interes...	NeurIPS_2023_submissions_huggingface	2,023	null	null	null	null	null	null	null	null
Mitigating Test-Time Bias for Fair Image Retrieval	Accept (poster)	Summary: The paper considers fair text-based image retrieval, in particular when the target retrieval set is biased. The chosen fairness metric is _equal representation_, e.g., the number of retrieved male and female images should be identical. The authors conjecture that if the target retrieval set is biased, debiasin...	Rebuttal 1: Rebuttal: ### Response to Reviewer PMeK Thanks for your detailed review and constructive feedback. Before reading our dedicated response to your comments, we kindly suggest you start with our global response, which offers a broad overview and addresses some of the general concerns raised. Below is our ...	Summary: The paper tackles the problem of bias in text-image retrieval. The goal is to generate fair and unbiased image retrieval results given neutral textual queries. As opposed to previous approaches that aim to separate the learned representations of images and text from characteristics such as gender and race, the...	Rebuttal 1: Rebuttal: ### Response to Reviewer VHZT Thanks for your detailed review and constructive feedback. Before reading our dedicated response to your comments, we kindly suggest you start with our global response, which offers a broad overview and addresses some of the general concerns raised. Below is our ...	Summary: The paper proposes a method for mitigating test-time bias in image retrieval of vision-language models. More concretely, the authors demonstrate how prior work debiasing methods fail to ensure equal representation of different sensitive groups in the retrieval bag at test-time when the test set is imbalanced a...	Rebuttal 1: Rebuttal: ### Response to Reviewer TTMH Thanks for your detailed review and constructive feedback. Before reading our dedicated response to your comments, we kindly suggest you start with our global response, which offers a broad overview and addresses some of the general concerns raised. Below is our ...	Summary: This paper proposes a post-hoc fairness method for text-to-image retrieval tasks. Instead of pre-processing the dataset or in-processing via fairness-derived regularization, this paper focuses on the post-processing of the retrieved items. The proposed post-hoc bias mitigation (PBM) utilizes a pre-trained bias...	Rebuttal 1: Rebuttal: ### Response to Reviewer KCNe Thanks for your detailed review and constructive feedback. Before reading our dedicated response to your comments, we kindly suggest you start with our global response, which offers a broad overview and addresses some of the general concerns raised. Below is our ...	Rebuttal 1: Rebuttal: We sincerely appreciate your valuable feedback. We are grateful for the constructive comments and suggestions, which provided us with insights to improve the quality and clarity of our work. Below, we address the common concerns raised by the reviewers. ### **Concerns on the performance of group ...	NeurIPS_2023_submissions_huggingface	2,023	null	null	null	null	null	null	null	null
State Sequences Prediction via Fourier Transform for Representation Learning	Accept (spotlight)	Summary: The authors in this paper investigated the structural information in state sequences for reinforcement learning, and proposed the prediction of Fourier transformation for the input sequences, which is further used an auxiliary task for learning policies. Some theoritical results are provided to show the existe...	Rebuttal 1: Rebuttal: We thank the reviewer for the positive and insightful comments. We respond to each comment as follows and sincerely hope that our rebuttal could properly address your concerns. If so, we would deeply appreciate it if you could raise your score. If not, please let us know your further concerns, and...	Summary: Data-efficient RL exploits representation learning which learns representations by predicting long-term future states. However, inherent structural information in sequential state signals is ignored in existing methods. The authors propose SPF which exploits the frequency domain of state sequences. The propose...	Rebuttal 1: Rebuttal: We thank the reviewer for the insightful and valuable comments. We respond to your comments as follows and sincerely hope that our rebuttal could properly address your concerns. If so, we would deeply appreciate it if you could raise your score（"3: Reject"). If not, please let us know your further...	Summary: Unlike many prior model-based RL methods that predict one future state at a time, the authors present a Fourier-transform based approach that predicts future states over a fixed interval. They prove that a large difference in policy performance will cause a significant difference between the experienced states...	Rebuttal 1: Rebuttal: We thank the reviewer for the positive and insightful comments. We respond to each comment as follows and sincerely hope that our rebuttal could properly address your concerns. If so, we would deeply appreciate it if you could raise your score. If not, please let us know your further concerns, and...	Summary: The paper suggests a representation learning method that leverages the existing periodic structures in the state sequences of a policy. Specifically, they propose using the features obtained by the Fourier transform of the state sequence. To practically extract such features, they employ a Bellman equation-lik...	Rebuttal 1: Rebuttal: We thank the reviewer for the positive and insightful comments. We respond to each comment as follows and sincerely hope that our rebuttal could properly address your concerns. If so, we would deeply appreciate it if you could raise your score. If not, please let us know your further concerns, and...	Rebuttal 1: Rebuttal: We thank all the reviewers for the insightful comments and constructive suggestions, which are very helpful for us to strengthen this submission. We have submitted detailed responses to address the concerns raised by all the reviewers. Besides answering the technical questions, 1. We provide a s...	NeurIPS_2023_submissions_huggingface	2,023	null	null	null	null	null	null	null	null
Mitigating the Popularity Bias of Graph Collaborative Filtering: A Dimensional Collapse Perspective	Accept (spotlight)	Summary: The paper discusses the issue of popular bias in Graph Collaborative Filtering (GCF) and demonstrates its resulting Dimensionality Collapse (DC) in the embedding space. The paper points out that LightGCN, widely used in GCF, is prone to DC. The authors introduce a decorrelated penalty based on LogDet divergenc...	Rebuttal 1: Rebuttal: # Response to Reviewer 5 (VmvT) *We thank the reviewer for the constructive review and valuable questions that have helped us improve our work.* ## 1. Title emphasizes alleviating popularity bias but the analysis focus on the dimensionality collapse. - **Dimensionality collapse (DC) and Po...	Summary: The paper proposes a novel graph collaborative filtering (CF) approach which tackles the issue of item popularity bias. Starting from the (theoretically-proven) assumption that state-of-the-art graph CF architectures (such as LightGCN) may enhance the so-called dimensional collapse (DC) problem, the authors de...	Rebuttal 1: Rebuttal: # Response to Reviewer 4 (bxqe) *We thank the reviewer for the constructive review and valuable questions that have helped us improve our work.* To further convince reviewer of the value of our work, we have added Figure 1 in our rebuttal PDF. The occurrence of dimensional collapse is cl...	Summary: In this paper, the authors study the dimension collapse problem in graph-based recommenders. The authors provide both sufficient experimental and theoretical analysis on the occurrence of dimension collapse to popular items and propose a new loss function to encourage feature diversity. Experiments are also su...	Rebuttal 1: Rebuttal: # Response to Reviewer 3 (Qs55) *We thank the reviewer for the constructive review and valuable questions that have helped us improve our work.* ## 1. Does the dimension collapse exist only in graph-based collaborative filtering? How about matrix factorization? - The phenomenon of dimension...	Summary: This paper studies popularity bias in graph collaboration filtering through dimensional collapse analysis of the embeddings. The authors proposed the LogDet divergence based de-correlation penalty for GCF to improve the performance. Strengths: This work has thorough analysis of dimensional collapse in GCF an...	Rebuttal 1: Rebuttal: # Response to Reviewer 2 (CD6b) *We thank the reviewer for the constructive review and valuable questions that have helped us improve our work.* ## 1. The model names seem inconsistent it the experimental study (Fig. 4 and Tables 2-3). We apologize. We have now revised our manuscript accord...	Rebuttal 1: Rebuttal: Kindly check the rebuttal PDF for Figures (bottom of this panel) requested by reviewers. Individual responses across rebuttal refer to them. Pdf: /pdf/95cfb08d119db7a05b654c91ab7c7ecc28e55353.pdf	NeurIPS_2023_submissions_huggingface	2,023	Summary: This paper aims to study the problem that features tend to occupy the embedding space inefficiently. The authors proposes a decorrelation-enhanced GCF objective that promotes feature diversity. LogDet divergence is leveraged to form a new constraint loss. Strengths: 1. The studied problem of embedding space ...	Rebuttal 1: Rebuttal: # Response to Reviewer 1 (pKCn) *We thank the reviewer for the constructive review and valuable questions that have helped us improve our work.* ## 1. The dimensional collapse issue (also known as overcorrelation) has been studied in [1] but not cited. - Thank you for pointing out this relat...	null	null	null	null	null	null
Near-Linear Time Algorithm for the Chamfer Distance	Accept (poster)	Summary: The paper proposes a fast estimation of Chamfer distance which is a famous distance between two sets. The key idea is to use random sampling on the first set and then find the closest point on the other set. To enhance further the approximation performance and quality of samples, importance sampling is utilize...	Rebuttal 1: Rebuttal: We thank the reviewer for their detailed and thorough comments. See our answers to specific comments below. On the set of ground metrics which the algorithm works for. We emphasize that our methods work for any metrics which support a locality sensitive hash function (or such metrics squared...	Summary: This paper provides the first near-linear time approximation scheme to compute the Chamfer distance $\texttt{CH}(A,B) = \sum_{a \in A} \min_{b \in B} d_X(a,b)$ for any two point sets $A,B \in \mathbb{R}^d$ and $d_X$ is the distance measure under consideration. In particular, for $\|A\|, \|B\| \leq n$, and for any ...	Rebuttal 1: Rebuttal: We thank the reviewer for their detailed and thorough comments. See our answers to specific comments below. Applications of Chamfer Distance. In addition to being a proxy for earth-mover distance, Chamfer distance is widely and independently used in point-cloud reconstruction and completion ...	Summary: The paper gives an algorithm to approximate Chamfer distance between two sets within $(1 \pm \epsilon) $ of the original value. The brute force algorithm to calculate the distance exactly between two sets of size $n$ and with points in $d$-dimension is $O(n^2d)$ which is prohibitive for large values of $n$. Th...	Rebuttal 1: Rebuttal: We thank the reviewer for their detailed and thorough comments. See our answers to specific comments below. Experiments are performed under a variety of settings and code is provided. Thank you for the comment. We intend to make the code public if the paper is accepted. **Definition of Loca...	Summary: This paper studies the approximation algorithm for computing the Chamfer distance between two datasets. The Chamfer distance is a very useful distance metric in machine learning, and the naive approach for computing it takes $O(n^2d)$-time for $n$ points in $d$ dimensions. This paper proposes an approximation ...	Rebuttal 1: Rebuttal: On notation. We apologize for the lack of precision in notation in the main text. We will fix these issues. On metrics that admit a locality-sensitive hash. We focused on L2/L1 as, to our knowledge, these are the most typical cases. However, our methods work for any metrics which support ...	Rebuttal 1: Rebuttal: We thank all the reviewers for useful comments and feedback. We will fix the typos and presentation issues in the final version of the paper. In what follows in the individual rebuttals, we address the issues identified by the reviewers as weaknesses and/or listed as questions.	NeurIPS_2023_submissions_huggingface	2,023	Summary: This paper considers the problem of computing the Chamfer distance between two point sets. Informally, the Chamfer distance is simply the total cost of mapping points of A to their closest point in B. Practically, Chamfer distance seems to be used as a proxy for the earth mover’s distance. One can always comp...	Rebuttal 1: Rebuttal: We thank the reviewer for their detailed and thorough comments. We would first like to remark that, while it is true that Chamfer is usually used as a proxy for EMD, there are scenarios where Chamfer distance is actually the correct choice instead of EMD. This includes settings where duplication o...	null	null	null	null	null	null
An Efficient Dataset Condensation Plugin and Its Application to Continual Learning	Accept (poster)	Summary: The present paper studies the task of dataset condensation (DC) with a specific focus on its computational and condensation efficiency. The main idea is to condense the dataset in a low-rank representation space instead of in the high-dimensional pixel space. To do so, the author decomposes an image representa...	Rebuttal 1: Rebuttal: - Advantages in high-resolution images: Thanks for your suggestion, in fact our method works better when the original image resolution is higher. As shown in $\underline{\text{Fig.2 in the response}}$, we sampled a subset of ImageNet and tested the performance of DM and our LoDM at different...	Summary: The authors propose an efficient dataset condensation (DC) plugin that utilizes a low-dimensional manifold, namely two low-rank matrices, to match the large raw and smaller synthetic condensed dataset. This aims to reduce inefficient condensation when matching in image pixel space. They show that their method ...	Rebuttal 1: Rebuttal: - The impact of synthetic data: Indeed, in $\underline{\text{Fig.3(d) in the original manuscript}}$, it seems intuitively that the generated image is no longer similar to the real image, but from the perspective of model learning, the impact of this difference may be weak. On the one hand, as...	Summary: This paper introduces a novel dataset distillation technique, referred to as LoDM. Instead of storing data in pixel space, LoDM propose to store each condensed image in a low dimensional space which is represented by multiplication of two matrixes. By decreasing the memory budget allocated for each image, LoD...	Rebuttal 1: Rebuttal: - Relationship to works [1,2,3]: Thanks for your suggestions for related works. Although our work and [1,2,3] seem to be data parameterization methods, the motivation of our work is fundamentally different from theirs: IDC[1] constructs multi-formation synthetic images by upsampling, using ne...	Summary: This paper proposed a low-rank technique to reduce the storage of the compressed dataset. It successfully applied the LoRA technique in DC and DD problems, and also apply it to the CL problem. After applying the LoRA technique, the method can still achieve similar performance compared to the original DM. Str...	Rebuttal 1: Rebuttal: - The significance of our low-rank method: As described in our related work, low-rank is a long-standing and fundamental research point in the field of machine learning. LoRA is not the first work to study low-rank. The contribution of LoRA is to consider the characteristics of low intrinsic d...	Rebuttal 1: Rebuttal: Dear reviewers, the figures in the responses are provided in this supplementary material, which contains 3 figures in total: - Fig.1 is the principal component analysis of real and synthetic images of the dataset distillation method, which shows that both real and synthetic images are low-rank and...	NeurIPS_2023_submissions_huggingface	2,023	Summary: This paper proposes a simple yet effective plugin for data condensation/distillation methods. Specifically, utilizing the low-dimensional nature of image manifolds, it proposes to decompose each synthetic image into two low-rank matrices and optimize them instead. Experimental results demonstrate the effective...	Rebuttal 1: Rebuttal: - Our motivation and method matching: As shown in $\underline{\text{Fig.1(a,b) in the response}}$, we perform principal component analysis on real images and synthetic images of DM on the CIFAR10 dataset. It can be observed that both are dominated by a small number of top eigenvalues, so both ...	null	null	null	null	null	null
Convolutional Neural Operators for robust and accurate learning of PDEs	Accept (poster)	Summary: The authors propose a UNet-like architecture for learning solution operators of PDEs. The architecture is assembled from building blocks analogous to the classical UNet architecture, but individual components respect a continuous-discrete equivalence. This makes the proposed architecture a representation equiv...	Rebuttal 1: Rebuttal: We start by thanking the reviewer for your appreciation of the merits of our paper and your welcome suggestions to improve it. We address your detailed concerns below. 1. Regarding the reviewer's point about novelty vis-a-vis Karras et al 2021, while acknowledging that we have used building blocks...	Summary: In this paper, the authors propose a new operator learning framework called Convolution Neural Operator which although works on discrete space, satisfies the property of Continuous-Discrete Equivalence (CDE) property. They define CNO over a UNet architecture and provide some universal approximation proofs. Fur...	Rebuttal 1: Rebuttal: We start by thanking the reviewer for your appreciation of the merits of our paper and your welcome suggestions to improve it. We address your detailed concerns below. 1. We apologize about the possible lack of clarity in notation for $r$ on line 82 -- it refers to the smoothness index $r$ of the...	Summary: The paper under review proposes a CNN architecture as a neural operator for solving PDEs via neural networks. The goal is to preserve the underlying continuous structure while being implemented with discrete convolutional architecture, based on U-Net. The idea is to consider the space bandlimited functions i...	Rebuttal 1: Rebuttal: We start by thanking the reviewer for your appreciation of the merits of our paper and your welcome suggestions to improve it. We address your detailed concerns below. 1. Our focus in the empirical part of the paper was on accuracy and during model selection, we chose the hyperparameters for the m...	Summary: The paper re-introduces convolution neural networks to the operator learning setting. It uses interpolation including up-sampling and downsampling layer to make sure the convolution layer is well-defined in the function space. The paper also discuss the space of band-limited functions and the trade off between...	Rebuttal 1: Rebuttal: We start by thanking the reviewer for your comments which we address below. 1. Regarding the reviewer's question about "trade-off between $L^p-H^p$ functions and bandlimited functions" and concern that "CNO only learns bandlimited functions", we start by pointing out that in Theorem 3.1 of our pa...	Rebuttal 1: Rebuttal: At the outset, we would like to thank all four reviewers for their thorough and patient reading of our article. Their criticism and constructive suggestions will enable us to improve the quality of our article. If our paper is accepted, we will incorporate all the changes that we outline below in ...	NeurIPS_2023_submissions_huggingface	2,023	null	null	null	null	null	null	null	null
Task-Robust Pre-Training for Worst-Case Downstream Adaptation	Accept (poster)	Summary: The authors proposed a concept called task-robust pre-training, which means pre-training a model that guarantees a uniformly good performance over the downstream tasks. A minimax loss and a corresponding algorithm to solve it is provided with convergence analyses in the convex setting. Experiments on large-sca...	Rebuttal 1: Rebuttal: Thank you very much for your detailed feedback. We are particularly grateful for your insightful comments which have helped us identify key areas for enhancement in our manuscript. We would like to address the specific points you raised: 1. **Convergence Rate of the Minimax Optimization Algorithm...	Summary: The paper tackles the paper of downstream adaptation from a large pretrained "foundation" model with a focus on robutness: Some downstream tasks may be harder to transfer to than others. To tackle this, the paper proposes a pre-training strategy to encourage uniformly good performance across all downstream tas...	Rebuttal 1: Rebuttal: Thank you for the detailed feedback on our submission. We highly appreciate your insights and have carefully considered each point you raised. Let us clarify and address your concerns: Weeknesses 1. Weak Comparison to multi-task learning optimization methods: We acknowledge that incl...	Summary: In order to improve the robustness of the pre-trained model on downstream tasks, the authors propose a simple optimization algorithm, softmax weighted gradient descent, to minimize the worst-case expected risk of upstream tasks. Strengths: The idea of the article is simple and easy to understand, the proof is...	Rebuttal 1: Rebuttal: Firstly, thank you for your thoughtful feedback on our paper. I would like to delve deeply into the central theme of our work: the critical importance of robustness. Let us clarify and address your concerns: Consistent Improvement： 1. Achieving consistent improvements in downstream tasks sol...	Summary: This work introduces the concept of worst-case adaptation, as each pre-training task may transfer differently to various downstream tasks. In order to address this issue, the authors propose a novel approach using a minmax loss and weighted gradient descent to optimize for the worst-case loss among all pre-tra...	Rebuttal 1: Rebuttal: Thank you for your detailed feedback on our work. We truly value your constructive comments and have taken steps to address your concerns. 1. Discussion with some multi-task learning methods: We acknowledge the importance of including a discussion on multi-task learning methods. In the update...	Rebuttal 1: Rebuttal: We attach PDF files here. Pdf: /pdf/5822b26234d45dc93630b81b3df638154b540c28.pdf	NeurIPS_2023_submissions_huggingface	2,023	null	null	null	null	null	null	null	null
Trade-off Between Efficiency and Consistency for Removal-based Explanations	Accept (poster)	Summary: This paper studies the consistency and efficiency of removal-based explanation methods. By establishing the impossible trinity theorem, the authors show that consistency and efficiency cannot be achieved simutaneously. They further propose the interpretation error as a metric to gauge the trade-off between con...	Rebuttal 1: Rebuttal: We express our gratitude to the reviewer for dedicating their valuable time and effort to evaluate our manuscript. In the following sections, we provide detailed responses to the reviewer's comments: Reply to Weaknesses 1. Acknowledging that the presentation may be complex for some readers, ...	Summary: This paper addresses the problem of tradeoffs between interpretability, consistency, and efficiency of AI models. The authors present “the Impossible Trinity Theorem”, which posits that interpretability, efficiency and consistency cannot hold simultaneously. To quantify the tradeoff, the authors define the uti...	Rebuttal 1: Rebuttal: We express our gratitude to the reviewer for dedicating their valuable time and effort to evaluate our manuscript. In the following sections, we provide detailed responses to the reviewer's comments: Reply to Weaknesses 1. It is acknowledged that the main text may seem to lack sufficient res...	Summary: This paper explores the interplay between two properties of removal based explanations: efficiency and consistency. It established that these two properties cannot coexist when the set of predictive models (requiring explanation), and the set of interpreting models (doing the explaining) are disjoint. It propo...	Rebuttal 1: Rebuttal: We express our gratitude to the reviewer for dedicating their valuable time and effort to evaluate our manuscript. In the following sections, we provide detailed responses to the reviewer's comments: Reply to Weaknesses: 1. Definition of Consistency: We acknowledge that the term "consisten...	Summary: The paper analyses the trade-off between efficiency and consistency in the context of removal-based post-hoc explanations. The authors demonstrate that having an interpretability method that is both efficient and consistent is impossible. Further, on the problem of finding consistent explanations, the authors ...	Rebuttal 1: Rebuttal: We express our gratitude to the reviewer for dedicating their valuable time and effort to evaluate our manuscript. In the following sections, we provide detailed responses to the reviewer's comments: Reply to Weaknesses 1. The prioritization of consistency over efficiency in our paper stems ...	Rebuttal 1: Rebuttal: # General Response We are deeply grateful for the efforts and valuable feedback of the reviewers and area chairs in reviewing our manuscript. We have made a detailed response to each reviewer's concern, including some clarifications and additional experimental results, and adopted some of the rev...	NeurIPS_2023_submissions_huggingface	2,023	Summary: This paper proposes using interpretation error as a metric to gauge inconsistencies and inefficiencies. The authors proposed three algorithms founded on the standard polynomial basis to minimize the interpretation errors. Authors perform experiments using both image and text classification to demonstrate the s...	Rebuttal 1: Rebuttal: We express our gratitude to the reviewer for dedicating their valuable time and effort to evaluate our manuscript. In the following sections, we provide detailed responses to the reviewer's comments: Reply to Weaknesses We acknowledge the valid concerns surrounding prediction accuracy and co...	null	null	null	null	null	null
Minimum norm interpolation by perceptra: Explicit regularization and implicit bias	Accept (poster)	Summary: This paper investigates how shallow ReLU networks interpolate between known regions as the number of data points and parameters tends to infinity. The authors prove that minimizers of regularized risk functionals converge to minimum norm interpolants of a given target function in an infinite parameter limit. T...	Rebuttal 1: Rebuttal: > The paper does not show the effectiveness for practical neural network optimization, such as CNN and Transformer. This is true, and it is not the article's goal. While more advanced network architectures are of great practical interest, the theoretical instruments to describe them are not suffi...	Summary: This paper investigates the behavior of shallow ReLU networks in terms of convergence to minimum norm interpolants of given data. The authors present provable results when explicit regularization is included and empirical evidence when it is not. The paper concludes with a summary of the authors' empirical ins...	Rebuttal 1: Rebuttal: We appreciate the reviewer's feedback and their forthcoming assessment of their level of expertise. In order to make the article accessible and interesting to a broad audience, we will include a more detailed discussion of the fit/regularization trade off and potential implications in the final wo...	Summary: Paper summary: * The paper studies the convergence properties of 1-layer Relu nets trained on a regression task with Tikhonov regularisation. The main contribution of the paper is Theorem 2.1, in which the authors prove that under some assumptions on the data distribution and on the optimisation trajectory (or...	Rebuttal 1: Rebuttal: > The paper applies to a very limited setting, namely 1 layer Relu nets trained on regression tasks. As we note in the final paragraph of Section 2 (Lines 115 - 117), our methods apply to (regression problems in) all function spaces which satisfy a Rademacher bound, a suitable compact embedding a...	Summary: This paper studies the optimization and generalization for two-layer ReLU neural network models. Specifically, * for the generalization part, this paper gives a Barron-norm based generalization bound for the minimize of the $\ell_2$-regularized squared loss function. The bound involves with three terms: sampl...	Rebuttal 1: Rebuttal: > The related work description is not comprehensive for me. Since I am not that familiar with the Barron-norm related fields, I don't know if there exists other works that study the generalization of neural networks in the Barron space. If there indeed exists, it would be better if the authors co...	null	NeurIPS_2023_submissions_huggingface	2,023	Summary: The authors study the generalisation behaviour of two-layer neural networks in the over-parameterised regime, where the network can perfectly interpolate its data in many different ways. What can we say about the functions that are learnt using SGD in these cases? The authors address this question using mathem...	Rebuttal 1: Rebuttal: We would like to thank the reviewer for their kind feedback and insightful comments. We will replot the Figures in the final version with larger labels, if accepted.	Summary: This work studies the global minimizer of the empirical risk with $\ell_2$ norm regularization of the parameters. It considers a one hidden ReLU layer architecture with square loss. The main result roughly states that for some asymptotic regime on $\lambda$ (reg parameter), $n$ (number of data points) and $m$...	Rebuttal 1: Rebuttal: We are grateful to the reviewer for their detailed feedback. We are unable to fully address all points within the character limit, but we will take their full feedback into account for a revision. > What is the main message of this work? The key messages of this work are: 1. Neural network mini...	null	null	null	null
Birder: Communication-Efficient 1-bit Adaptive Optimizer for Practical Distributed DNN Training	Accept (poster)	Summary: With the increasing size of the model, communication compression plays an important role in distributed machine learning because the model needs to transmit among the nodes. Toward this goal, this paper introduces a one-bit adaptive optimizer named BRAM and further extends the method to hierarchical all-reduce...	Rebuttal 1: Rebuttal: We express our sincere gratitude for your volunteering time and your valuable comments. In response to your feedback, we have organized our response into the following points: Q1. *The literature review is poor. The authors did not comprehensively survey the works of one-bit compression, e.g....	Summary: The authors propose a novel 1-bit adaptive optimizer by compressing the total update in each iteration. Besides, the author modifies the adam update with the absolute value of the gradient. No bias correction is needed and error feedback is applied in algorithm 1. Hierarchical-1-bit-All-Reduce is proposed, whe...	Rebuttal 1: Rebuttal: We express our sincere gratitude for your volunteering time and your valuable comments. In response to your feedback, we have organized our response into the following points: Q1. All-gather instead of All-Reduce is used in figure 1 R1. Thank you for raising this concern. We hold a r...	Summary: This paper proposes a compute-efficient gradient compressor named BRAM which randomly quantifies the update direction into 1-bit for communication. A theoretical convergence bound is also provided to prove BRAM’s convergence property. To improve communication efficiency, BRAM also integrates a novel Hierarchic...	Rebuttal 1: Rebuttal: We express our sincere gratitude for your volunteering time and your valuable comments. We organize our response into the following points: Q1. *The top-1 accuracy drops to 74.+\% with globe batch size of 8k for training ResNet-50, while methods in MLPerf can reserve 75.9\% with batch size ...	Summary: The paper introduces a 1-bit adaptive optimizer called BRAM for distributed training of deep neural network models. BRAM combines the merits of SignSGD and Adam optimizers to achieve fast convergence speed, high inference performance, and reduced communication overhead. The proposed optimizer employs extreme 1...	Rebuttal 1: Rebuttal: We express our sincere gratitude for your volunteering time and your valuable comments. In response to your feedback, we have organized our response into the following points: Q1. *The theoretical convergence analysis of BRAM relies on several assumptions, such as bounded infimum, Lipschitz ...	Rebuttal 1: Rebuttal: I greatly appreciate the efforts put forth by the chairs and reviewers to ensure high-quality reviews. We have studied the comments and suggestions seriously and carefully, which have definitely helped us improve the work. We now respond to the reviewers' comments as follows. The reviewer comment...	NeurIPS_2023_submissions_huggingface	2,023	Summary: The paper proposes BRAM, a new 1-bit optimizer that directly quantizes the entire adaptive update. It theoretically proves that BRAM promises the same convergence speed as the full-precision Adam even with extreme 1-bit quantization. Considering the difference between intra-node and inter-node bandwidth, the p...	Rebuttal 1: Rebuttal: We express our sincere gratitude for your volunteering time and your valuable comments. In response to your feedback, we have organized our response into the following points: Q1. *It is unclear whether using the same hyperparameters for both the exponential moving averages of gradients and ...	null	null	null	null	null	null
Diverse Conventions for Human-AI Collaboration	Accept (poster)	Summary: This work suggests employing cross-play rewards to improve the measure of diversity in human-agent collaboration tasks, and presents an innovative training paradigm, namely, mixed-play, to address the challenge during optimizing cross-play, referred to as the "handshake" by the authors. As a result, the develo...	Rebuttal 1: Rebuttal: Thank you for the thorough review of our paper! We hope that the responses below answer your questions about our technique. ## Explanation for unintelligent behavior in simple scenarios Please refer to the “Clarification on the convention-aware agent” section in the common response above, where...	Summary: The paper introduces a new approach for generating diverse conventions by using a training objective that balances increasing self-play performance and minimizing cross-play performance with previously discovered conventions. To avoid handshake mechanisms that can game the proposed objective, the technique als...	Rebuttal 1: Rebuttal: Thank you for the thorough review of our paper! We hope that the responses below answer your questions about our technique. ## Choosing Convention-Aware Agent We explain the generation of a convention-aware agent in Equation 2, within section 2.1 of the paper, but we will edit the paper to re-e...	Summary: The authors develop CoMeDi which optimizes a cross-play objective but avoids handshakes that can prevent cooperation using mixed play. The algorithm is evaluated in three environments of increasing complexity and on an Overcooked level with real human users. Strengths: The paper is written well and the ideas...	Rebuttal 1: Rebuttal: Thank you for the thorough review of our paper! We evaluated our algorithm across the full Overcooked suite with baseline agents, as you recommended. We believe that some of your additional questions have been answered in the appendix (within the supplementary materials as Diverse_Conventions_Appe...	Summary: This paper introduces a new approach to dealing with poor generalisation performance in human-AI settings when actors may have developed different coordination conventions. Strengths: In general, I think the work is original and very clear in its presentation. In addition, the paper attempts to challenge a v...	Rebuttal 1: Rebuttal: Thank you for the thorough review of our paper! We’re glad that you played with the demo and think that our approach works well, even under strange conventions. We hope that the responses below answer your questions about our technique. ## Diversity of Overcooked’s Human test set We agree that ...	Rebuttal 1: Rebuttal: We thank all the reviewers for their thoughtful feedback. We appreciate that all reviewers found the paper well-written while tackling key challenges in human-AI collaboration research. We are happy that reviewers found our algorithm “well-formulated” (9YDu) and “elegant” (ghEw) while “avoiding th...	NeurIPS_2023_submissions_huggingface	2,023	null	null	null	null	null	null	null	null
Delayed Algorithms for Distributed Stochastic Weakly Convex Optimization	Accept (poster)	Summary: This paper investigates delayed stochastic optimization problems with composite objective functions. The authors demonstrate that, under weakly convex conditions, the convergence rate of the delayed stochastic subgradient method can be improved from $\mathcal{O}(\tau_{\max}/\sqrt{K})$ to $\mathcal{O}(\bar{\tau...	Rebuttal 1: Rebuttal: Response to Reviewer 93YJ We thank the reviewer for the time spent on our paper and the constructive suggestions. Weaknesses 1. The research direction is narrow, only focusing on a specific setting We politely disagree with the reviewer on "narrow, only focusing on a specific settin...	Summary: This paper studies distributed nonsmooth optimization with delayed updates. The authors study two kinds of objective functions, the first class is the sum of a weakly-convex nonsmooth function and a computation-friendly nonsmooth regularizer, while the second class is the sum of a composite function with a reg...	Rebuttal 1: Rebuttal: Response to Reviewer 6vL5 We appreciate the efforts of the reviewer in reviewing our paper. Questions. 1. Is it possible to generalize the inner function of the composite function to a mapping Yes. Our analysis actually can be directly extended to the case, where function $c$ is rep...	Summary: This paper presents DSPL, which is an optimization for a certain structured weakly convex problems. This algorithm can work with arbitrarily delayed updates, and outperforms existing algorithms in both theoretical rates and numerical experiments. Strengths: 1. The algorithm is robust to arbitrary delays. The ...	Rebuttal 1: Rebuttal: Response to Reviewer SDCr We thank the reviewer for the efforts put into the review process. Question. Is this paper's setting inherently simpler than nonconvex optimization. This is a good question, and we would like to answer the question for SGD and SPL respectively. For SGD, it is...	Summary: This paper studies delayed stochastic algorithms for weakly convex optimization. For a general weakly convex problem, the authors provide a sharper bound that improves the max delay in the previous bounds to "average" delays. For a weakly convex problem with a composition structure, they propose a variant, DSP...	Rebuttal 1: Rebuttal: Response to Reviewer Ld2g We appreciate the efforts the reviewer puts into the review process and thank the reviewer for acknowledging the novelty of our work. Questions. 1. Add labels to the axes of figures and split the figures. We thank the reviewer for the suggestion and we are ...	Rebuttal 1: Rebuttal: Clarification of the assumptions First, we thank all the reviewers for their efforts in the review process, and for acknowledging the novelty of our work. Both reviewer bGwb and 93YJ share the concern that the assumptions are made on stochastic functions and not the expectation ones, which ...	NeurIPS_2023_submissions_huggingface	2,023	Summary: The authors study a class of weakly convex optimization over a distributed network under the presence of delayed derivatives. They first analyze an existing algorithm and show that they can improve its complexity bound in terms of the dependence on the properties of the delay. They further provide another algo...	Rebuttal 1: Rebuttal: Response to reviewer bGwb We thank the reviewer for the time spent reviewing our paper. Question. Assumptions B1 and C1 are much stronger than the ones in the existing works; The reasons to get a better dependency of the delay is to have direct assumption on the stochastic functions and...	null	null	null	null	null	null
Optimize Planning Heuristics to Rank, not to Estimate Cost-to-Goal	Accept (poster)	Summary: The paper highlights the limitations of the conventional method for learning heuristics in search algorithms, which involves imitating the optimal cost-to-go function. The authors present an alternative approach that focuses on optimizing the correct ranking of state pairs by the heuristic. Because of that cha...	Rebuttal 1: Rebuttal: We thank the reviewer for valuable commplents. Due to space limit of, we will address his comments in the same order using lines for perfect matching. * l.111,117,157 We seen the expression in lines L.111 and L.117 the same. Can you clarify us, where do you see the difference exactly? The expres...	Summary: Planning with (possibly learned) heuristic function is important for both planning and learning communities. The paper focuses on the forward search without expanding redundant search states in their forward process. With the development theorem, the authors demonstrated how the proposed concept work in variou...	Rebuttal 1: Rebuttal: Thank you for an interest in our work. Do you discuss why the learning-based heuristic is promising, even though the training resource is required? Let us first clarify the computational complexity. The paper reports the total time of all experiments presented in the paper. This means that 1772...	Summary: This paper investigate the computational advantage if a forward-search algorithm utilises ranking rather than optimising cost-to-goal. I think the contribution is good in general. Strengths: 1. The problem is formulated well. 2. The introduction of the contents is suitable. 3. The experiment section is good...	Rebuttal 1: Rebuttal: The definition of "optimal ranking" is so important to this paper that it needs to be presented before first used. The authors should not use "(definition is below)" for this term. This is a good suggestion. We will incorporate the definition to the motivation. *I suggest the authors simplify ...	Summary: The paper considers the problem of using imitation learning for obtaining planning heuristics suitable for using variations of best-first search with a weighted sum of h() and g(), covering A* and GBFS. The paper gives both theoretical and empirical arguments for predicting a ranking of actions instead of pred...	Rebuttal 1: Rebuttal: Thank you for an interest in our work and great questions. What happens if there are multiple solution path with the same length of the plan (number of actions) and cost?* The behavior of the resulting heuristic function depends on the composition of the training set. Let us illustrate the beha...	null	NeurIPS_2023_submissions_huggingface	2,023	null	null	null	null	null	null	null	null
Protein Design with Guided Discrete Diffusion	Accept (spotlight)	Summary: This paper is focused on the task of multi-objective protein sequence design using guidance in diffusion models. Since protein sequences are discrete objects, classifier guidance tools from continuous diffusion are not directly applicable. To bridge this gap, the paper proposes a new method NOS, where the clas...	Rebuttal 1: Rebuttal: Thank you for your review. We have added DiffAb and RFDiffusion comparisons to the antibody lead optimization task in the rebuttal PDF. There we show that our method generates samples with much higher values of the multi-objective acquisition function than these competing diffusion methods. This r...	Summary: This paper proposes a discrete diffusion model guided by protein properties. They add a property predictor to the diffusion model, sharing the hidden states. The method is largely inspired by previous work, while experimental results are positive. Strengths: - The paper is well written. - The experimental res...	Rebuttal 1: Rebuttal: Thank you for your review and suggestions for improved comparisons. ADDITIONAL EXPERIMENTS We have performed experiments with two of the papers you reference, and the results are quite positive in favor of our method. The detailed results are contained in the rebuttal PDF, but the findings can ...	Summary: This paper proposes new methods for discrete diffusion guidance, making it possible to optimize protein sequences for local and global properties while retaining high sequence likelihood. Strengths: 1. The paper addresses a problem closely related to the NeurIPS community. 2. The methods sound reasonable. 3. ...	Rebuttal 1: Rebuttal: Thank you for your detailed response and suggestions. DiGress Comparison While DiGress looks like a closely related and compelling method, its application appears to be limited to graph representations of small molecules, with a maximum of 9 nodes. As our proteins have 1000 nodes, adapting ...	Summary: The submission introduces a new diffusion for protein design. It focuses on overcoming the challenge of discrete nature of protein sequences. The model succeeded in local and global properties optimization of antibody sequences making them more applicable for therapeutic purposes. Strengths: Originality: Pro...	Rebuttal 1: Rebuttal: Thank you for your supportive review and for emphasizing the impact and novelty of our submission! We have included the requested comparison with RFDiffusion in our rebuttal PDF. We also include DiffAb and several other local sequence search methods (AdaLead, PEX) for the protein design task in ou...	Rebuttal 1: Rebuttal: We are glad to share many exciting new results inspired by reviewer comments in the attached PDF, including a new comparison with RFDiffusion and DiffAb on antibody lead optimization (Figure 1) and two additional genetic algorithm baselines mentioned by gzgm (Figure 2). In both of these experiment...	NeurIPS_2023_submissions_huggingface	2,023	Summary: The paper presents a novel approach to applying diffusion models to optimize protein sequences, specifically for antibody design. Diffusion models have been successful in handling continuous data like images, but their direct application to protein structures faces challenges due to limited high-quality struct...	Rebuttal 1: Rebuttal: Thank you for the detailed feedback. First of all, we would like to push back on the assertion that our paper has inadequate description of models or training procedures or lack of reproducibility. We encourage the reviewer to take a look at the paper appendices and follow the anonymous repo lin...	null	null	null	null	null	null
Optimization of Inter-group criteria for clustering with minimum size constraints	Accept (poster)	Summary: The authors of this paper consider metric clustering problems, where the goal is to maximize inter-clustering objectives. Such objectives measure how "distinct" different clusters are. The first objective they consider is the minimum spacing one (MS). According to this, the goal is to maximize the minimum dist...	Rebuttal 1: Rebuttal: We are quite happy that you appreciated our work! In our comment directed to all referees, we discuss the fit for Neurips and potential applications. Questions: 1. About the papers you mentioned Reply. Thanks for pointing out the papers! Indeed, they are related to ours and we will c...	Summary: This paper considers the problem where we want to cluster $n$ points into $k$ clusters with the clustering objective trying to maximize the minimum distances between clusters (either the actual minimum distance or the weight of an MST defined between clusters). The authors show that greedy agglomerative cluste...	Rebuttal 1: Rebuttal: Thanks again for your time and effort! We discuss applications and the fit for Neurips in our comments addressed to all reviewers. Questions: *The experimental results do not convey much about the performance of the algorithms except for their relative performance, which is not very dramatic. ...	Summary: The paper considers the following problem: given a metric space on a set of $n$ points and a parameter $k$, partition the points into $k$ clusters $C_1, \ldots, C_k$. The paper considers two objective functions: consider a metric with vertices $X$ as the set of clusters and distance between two clusters being...	Rebuttal 1: Rebuttal: We thank the reviewer! However, we are a bit surprised by his/her evaluation, as discussed below. Score It is not clear what are the soundness and presentation problems in our work that justify a score 2 for these dimensions; there is no comment about it. Weakness The weakness that th...	Summary: The paper presents a body of theoretical results addressing the problems of minimum-spacing and minimum spanning tree clustering; both based on the concept of finding clustering solutions which maximise the inter-cluster distance(s). The former maximises the distance between the pair of closest clusters while ...	Rebuttal 1: Rebuttal: We greatly appreciate your kind words! We will be sure to review our text for typos and to incorporate the papers you mentioned to our Related Work section.	Rebuttal 1: Rebuttal: To all Reviewers We thank all the referees for your time and effort. We are happy that some referees classified our work as solid and extremely thorough and that all referees found our problem interesting. We have specific comments for each referee, but in the current one, we discuss potenti...	NeurIPS_2023_submissions_huggingface	2,023	null	null	null	null	null	null	null	null
Energy Transformer	Accept (poster)	Summary: This paper presents a new transformer architecture that is based on minimizing energy with two networks viewed as Hopfield networks (one for the attention within the data point, and another for global features). The result energy-based model resembles (and perhaps motivates) empirical choices in the vanilla tr...	Rebuttal 1: Rebuttal: > Limited evaluation (only on graphs)... I am not sure why the authors didn't try NLP. I think trying NLP is a very crucial step for convincing the community that ET works. Thank you for your valuable comment. We have included a more thorough discussion on this in our general response and in our ...	Summary: This paper proposes a new Transformer-like architecture which is is inspired by modern Hopfield networks, and falls into the category of energy-based models. The authors argue that the existence of an energy function leads to several advantages such as more principled design and theoretical interpretation. The...	Rebuttal 1: Rebuttal: Thank you for your enthusiastic evaluation of our work! > Insufficient discussion of limitations. Thank you for pointing this out, we will add the following discussion to the revised manuscript. So far, the proposed model has only been empirically tested on the graph domain. It remains unclear w...	Summary: This paper proposes a new transformer architecture namely Energy Transformer, consists of layer norm, multi-head energy attention and Hopfield network, within a single recurrent ET block. The ET block models the dynamics of token updates by minimising the sum of energy from both the multi-head attention and th...	Rebuttal 1: Rebuttal: > Empirical experiments on image reconstruction is hard to quantify and are also not compared to any of the SOTA algorithms. The value of the reconstruction error shown in the Appendix is hard to understand/evaluate without comparing to any other algorithms. Thank you for your valuable comment. ...	Summary: The paper presents a new architecture for transformer (ET). The architecture is designed by first defining an energy function that forces the alignment of the query of a patch to the keys of a few other patches and to make each patch to look like a natural image patch. A dynamics (differential equation) is ind...	Rebuttal 1: Rebuttal: > Despite its well-motivated design, the proposed structure still lacks rigorous theoretical justification for its effectiveness. Thank you for your insightful comments! We are glad to hear that you find ET's design well-motivated. We discuss the general theory of ET in Section 2. To summarize, E...	Rebuttal 1: Rebuttal: We sincerely thank all reviewers for their thoughtful feedback. We are excited that they highlight the strengths of the Energy Transformer (ET): 1. A novel architecture (q5cg, bQx9, zsPk, qJXt) 2. with a well-motivated design (q5cg, qJXt) built on top of Associative Memory fundamentals 3...	NeurIPS_2023_submissions_huggingface	2,023	null	null	null	null	null	null	null	null
Approximate inference of marginals using the IBIA framework	Accept (poster)	Summary: The paper presents a new approximation scheme for computing posterior marginals in graphical models such as Bayesian networks and Markov networks. The proposed approach is based on a previously introduced scheme called IBIA (Incremental Build Infer Approximate) which was initially developed for approximating t...	Rebuttal 1: Rebuttal: Thank you for your comments and questions. Following are our responses to each question. > how many evidence variables in each benchmark class? how do you compute the score when exact results are not available? We used evidence sets given in the repository and an average can be added to Table...	Summary: The paper proposes a new approximate inference algorithm for probabilistic graphical models (PGMs). The approach uses a framework called IBIA to perform marginal inference in discrete PGMs. The main idea is to perform Belief Propagation on a sequence of clique trees. In each step the clique tree is approximate...	Rebuttal 1: Rebuttal: Thank you for your comments and questions. Following are our responses to each question. > What is the effect of longer vs shorter sequences, how will this affect the experiment results? Error is introduced both in the approximate step during sequence construction and also during heuristic b...	Summary: The paper proposes a new approximate inference method for marginals of probabilistic graphical models (PGMs) based on the incremental build-infer-approximate (IBIA) framework. IBIA was introduced by Bathla & Vasudevan in an arXiv preprint in April 2023 to estimate the partition function (normalizing constant)....	Rebuttal 1: Rebuttal: Thank you for your comments and questions. We would also like to thank you for pointing out the typos. We will fix them in the revised version. Following are our responses to each question. > My main question is about the UAI 2022 competition: it seems to me that other submissions there perform...	null	null	Rebuttal 1: Rebuttal: We thank all reviewers for their insightful comments and questions. We are encouraged they found our idea to be novel (vgXS, SJmy), useful (vgXS) and interesting (vgXS, SJmy). We are glad that the reviewers found our evaluation sound (QKYy) and that our method is competiti...	NeurIPS_2023_submissions_huggingface	2,023	null	null	null	null	null	null	null	null
Architecture Matters: Uncovering Implicit Mechanisms in Graph Contrastive Learning	Accept (poster)	Summary: In the paper "Architecture matters: uncovering implicit mechanisms in graph contrastive learning", the authors study common graph-level and node-level graph contrastive learning (GCL) methods based on graph convolutional networks (GCNs). The authors compare GCL with contrastive learning (CL) in computer vision...	Rebuttal 1: Rebuttal: We thank Reviewer sd4m for careful reading and appreciating the novelty and potential impacts of our work. Below, we summarize and address your main concerns. --- Q1. Comment on how realistic is the assumption of Thm 4.1. It is unclear whether it is similar to augmentations in GRACE/GCA/ProG...	Summary: The paper explores graph contrastive learning (GCL) methods and identifies distinct phenomena that differentiate them from traditional visual contrastive learning (VCL) methods. It highlights that positive samples are not mandatory in GCL, negative samples are not necessary for graph and node classification, a...	Rebuttal 1: Rebuttal: We thank Reviewer 4ia5 for careful reading and constructive comments. We summarize and address your main concerns in the following. --- Q1. The consistency in experimental setups when comparing visual CL and graph CL. A1. Across our experiments, we follow the standard setting and data ...	Summary: This manuscript demonstrates the correlation between the graph neural networks and the contrastive learning loss. Specifically, the authors illustrate that 1) because of the fusion of neighbor nodes caused by GNNs, neighbor nodes act as positive sample pairs (alignment); 2) incorporating an over-smoothing solu...	Rebuttal 1: Rebuttal: We thank Reviewer RuBb for your detailed reading and encouraging comments of our work on its novelty and insights. Below, we summarize and address your main concerns. --- Q1. Line 86: The comparison with SCE is unclear. More explanations like the backbone comparison and learning objectives a...	Summary: The current paper focuses on understanding the implication of design choices borrowed from visual contrastive learning (VCL) for graph contrastive learning (GCL). The authors argue that the need of positive samples is relatively weak in GCL; the architecture itself can implicitly act as a positive sample gener...	Rebuttal 1: Rebuttal: We thank Reviewer nRNc for carefully reading and appreciating the value of our work on the re-evaluation of designs in GCL. Below, we summarize and address your main concerns. --- Q1. Several prior works have shown the surprisingly good performance of random models and how performing GCL oft...	Rebuttal 1: Rebuttal: Here we provide supplementary materials for rebuttal. Pdf: /pdf/7ef3f3587612471f5de88e14cabb34a1f1e5df86.pdf	NeurIPS_2023_submissions_huggingface	2,023	null	null	null	null	null	null	null	null
Efficient Training of Energy-Based Models Using Jarzynski Equality	Accept (poster)	Summary: This paper uses Jarzinsky Equality to analyze the training of Energy Based Model (EBM) using cross entropy loss. The authors highlight the problem of directly approximating the gradient using the unweighted mean of nonequilibrium samples and propose a modified algorithm based on reweighting. Strengths: I like...	Rebuttal 1: Rebuttal: Sampling efficiency. As with our method, the sample size (i.e. the number of wakers) must be large enough to guarantee statistical accuracy. There are few theoretical guarantees about this issue, unfortunately, and it is known that the Jarzynski weight can indeed become large if the sample siz...	Summary: The paper under consideration proposes a modification of the standard EBM training procedure inspired by Jarzynski equality from thermodynamics. The idea is as follows. It is known that EBM training requires sampling from the distribution for which we know only its energy function given by, e.g., neural networ...	Rebuttal 1: Rebuttal: We would like to thank the reviewer for their careful reading, positive view on the manuscript, and constructive suggestions. Below, we address the main points raised. Using mini-batches. Algorithm 1 can be easily generalized to use mini-batches, see Algorithm 4 in the paper, and it does no...	Summary: The authors consider EBMs family of algorithms optimized with a cross-entropy objective. They proposed a modification of the unadjusted Langevin algorithm (ULA) to remove bias from the optimization procedure, basically introducing correction coefficients for UCL gradient estimation. Their method is based on Ja...	Rebuttal 1: Rebuttal: We would like to thank the reviewer for their careful reading, positive view on the manuscript, and constructive suggestions. Below, we address the main points raised. Scope of the numerical experiments. As stated in the common reply, we did not have the computing resources necessary to perfo...	Summary: The paper reports an enhcements in the eficiency of the training of EBMs using Jarzinsky equality. Strengths: The paper has the potential to help understand the training of EBM better through a thermodynamics based approach. It is well written and elaborated. Weaknesses: The paper shows a more efficient tr...	Rebuttal 1: Rebuttal: We would like to thank the reviewer for their careful reading, and their suggestion to better explain underpinning from the statistical physics of our approach based on the Jarzynski equality (JE). Thermodynamic interpretation. This interpretation is actually the same as for physical systems:...	Rebuttal 1: Rebuttal: We would like to thank all of the reviewers for their careful reading, constructive suggestions, and overall interest in our work. In this reply, we answer to the main questions common to all reviewers; below, we address their concerns individually and give more details on each point. ***Regardin...	NeurIPS_2023_submissions_huggingface	2,023	Summary: One of the main challenges in the training of Energy-Based Models (EBMs) is the design of a tractable objective that allows for low-variance gradient estimation. Indeed, the maximum likelihood objective requires sampling from the current density model, which usually is computationally expensive and biased due ...	Rebuttal 1: Rebuttal: We would like to thank the reviewer for their careful reading, positive view on the manuscript, and constructive suggestions. Below, we address the main points raised. Scope of the experimental results. Unfortunately we did not have the computing resources necessary to perform experiments on ...	Summary: This paper examines EBM learning using tools from the Jarzynski and Sequential Monte Carlo. The method is meant to address limitations of current EBM methods, which often rely on simulated Markov Chains to calculate approximate gradient of the log normalizer. Since the chains in practice do not converge to the...	Rebuttal 1: Rebuttal: We would like to thank the reviewer for their careful reading, positive view on the manuscript, and constructive suggestions. Below, we address the main points raised. Scope of the experimental results. As stated in the common reply, we did not have the computing resources necessary to perfor...	null	null	null	null
Explainable and Efficient Randomized Voting Rules	Accept (poster)	Summary: This paper is concerned with introducing a simple randomization "step" to deterministic voting rules so as to increase efficiency, as measured through the distortion framework (and also min-sw). The goal is to retain both the explainability of deterministic voting rules while increasing efficiency. Strengths:...	Rebuttal 1: Rebuttal: We are sincerely grateful for your time and energy reviewing the paper. To answer your questions: > The fundamental premise of the paper is that deterministic => explainable. Is this a well-established fact in the voting community (besides appealing to [16] as in L46)? If not, then the whole inve...	Summary: The paper studies two types of simple (or in their words explainable) randomized voting rules with respect to their (utilitarian) distortion. These are, (1) randomized positional scoring rules where every agent gives a certain number of points to their first ranked, second ranked, etc. alternative and a winn...	Rebuttal 1: Rebuttal: Thank you for your time and helpful comments. Regarding your questions: > Have you studied the natural combination of your two (family of) rules? This is a nice idea, which we did not consider. These rules are only slightly less explainable than UR$_k$ rules, so if they lead to significant effi...	Summary: This paper studies distortions of two family of "explainable" randomized voting rules -- randomized positional scoring rules and random committee member rules. In the literature of explainable voting rules, various deterministic voting rules have been widely studied and believed to be explainable to some exte...	Rebuttal 1: Rebuttal: Thank you for your time and effort in reviewing our paper. Here are our ansewers to your questions: > Explainable AI...Nevertheless, it is not clear to me why adding one more randomization step won't jeopardize the explainability....I encourage the authors to at least elaborate more and make more...	Summary: This paper provides a discussion of and bounds for distortion (distance) and minimum social welfare of randomized additions to several voting rules. The main body of the paper introduces two versions of randomization for positional scoring rules (either using scores as the distribution or selecting a top-k and...	Rebuttal 1: Rebuttal: Thank you for your review and the time you took to provide such helpful feedback. In response to your questions: > ? After Theorem 3 there is a discussion of non-worst case results, is there any hope for average case or distributional results that would improve the worst-case bounds for the rando...	Rebuttal 1: Rebuttal: We thank all the reviewers for their detailed and helpful reviews. In response to some of the comments, we ran further experiments and included the results in the PDF attached to this common response. We will be happy to incorporate these results as well as the suggestions of the reviewers when we...	NeurIPS_2023_submissions_huggingface	2,023	null	null	null	null	null	null	null	null
MCUFormer: Deploying Vision Tranformers on Microcontrollers with Limited Memory	Accept (poster)	Summary: This paper focuses on the efficient deployment of vision transformers on microcontroller platforms with limited resources. Vision transformers face significant challenges when deployed on resource constrained MCUs due to their large volume of data movement and intensive computation size. The paper explores alg...	Rebuttal 1: Rebuttal: Thank you for careful reading and valuable comments. We will check the paper carefully, and modify the presentation of the ambiguous parts in the final version. We provide answers to the questions as follows: Q1: Compare with existing work. A1: Due to lack of operator library for vision transfo...	Summary: This paper introduces MCUFormer, which aims to deploy vision transformers on micrcontrollers. MCUFormer utilizes two key techniques: 1. search space evolution to search vision transformer architectures within the memory limits; and 2. the construction of a compiling library with efficient memory scheduling for...	Rebuttal 1: Rebuttal: Thank you for careful reading and valuable comments. We will check the paper carefully, and modify the presentation of the ambiguous parts in the final version. We provide answers to the questions as follows: Q1: Difference between AutoFormerV2[1] and MCUFormer A1: Traditional vision transforme...	Summary: The paper introduces a hardware-algorithm co-design framework, MCUFormer, that enables the deployment of vision transformers on microcontrollers (MCUs). This framework jointly designs transformer architectures and constructs an inference compiler in order to accommodate the constraints of memory resources. MCU...	Rebuttal 1: Rebuttal: Thank you for careful reading and valuable comments. We will check the paper carefully, and modify the presentation of the ambiguous parts in the final version. We provide answers to the questions as follows: Q1: How to reduce the memory A1: Since the memory footprint of vision transformers ex...	Summary: This work proposed a hardware-algorithm co-design method called MCUFormer to enable vision transformer deployment on MCUs with limited memory, which includes both model design and compilation optimization to meet the memory constrain of MCUs. Specifically, the authors utilize one-shot NAS to search the model t...	Rebuttal 1: Rebuttal: Thank you for careful reading and valuable comments. We will check the paper carefully, and modify the presentation of the ambiguous parts in the final version. We provide answers to the questions as follows: Q1: The confusion of compiling library and memory scheduling and the definition of me...	null	NeurIPS_2023_submissions_huggingface	2,023	Summary: This paper proposes a co-design method (MCUFormer) that enables the deployment of vision transformers on microcontrollers with limited memory. The method enlarges the search space of vision transformers by considering low-rank decomposition dimensions and patch resolution for memory reduction. The inference li...	Rebuttal 1: Rebuttal: Thank you for careful reading and valuable comments. We will check the paper carefully, and modify the presentation of the ambiguous parts in the final version. We provide answers to the questions as follows: Q: The role of low-rank decomposition dimensions and patch resolution for memory reduc...	null	null	null	null	null	null
Restless Bandits with Average Reward: Breaking the Uniform Global Attractor Assumption	Accept (spotlight)	Summary: This paper studies the infinite-horizon restless bandits problem and proposed a simulation-based framework, i.e., Follow-the-Virtual-Advice, to leverage single armed policy to solve a multi-armed problem, which gets rid of the difficult-to-verify condition, i.e., the uniform global attractor property. Removing...	Rebuttal 1: Rebuttal: We appreciate the reviewer’s time and effort put into the review. The reviewer has two main concerns: correctness of our results (Comments 1 and 2), and relationship with prior work by Ghosh et al. (ArXiv 22, AAMAS 23) (Comment 3). Regarding the correctness of our results, we believe that our proo...	Summary: This paper presents an algorithm providing an asymptotically optimal policy to solve restless bandits with N arms when N goes to infinity without using classical assumptions on the expected deterministic dynamics (uniform global attractor). Strengths: 1. The idea to couple the actual trajectory under the r...	Rebuttal 1: Rebuttal: We thank the reviewer for recognizing the strengths and contribution of our paper. Below we provide our responses to the comments. > Discussion on the synchronization assumption and its link with UGAP In our rebuttal to Reviewer JXov, we give a detailed discussion on SA and UGAP. We repeat it he...	Summary: In the late 80s and early 90s, Whittle formalized the restless multi-armed bandit framework and Weber & Weiss proved that Whittle’s index policy, which solves a relaxed version of the optimization problem, is asymptotically optimal under conditions of indexability. Since then, RMAB problems have attracted much...	Rebuttal 1: Rebuttal: We thank the reviewer for recognizing our paper’s contribution and strengths, and for the constructive comments. We provide our detailed responses below. > Relation with Ghosh et al. (AAMAS 23). We thank the reviewer for pointing out this related work. We address this point in detail in our ...	Summary: This paper studied, for both the discrete-time and continuous-time settings, the average-reward restless bandit problem without the uniform global attractor property. They proposed the simulation-based policies Follow-the-Virtual-Advice (FTVA) and FTVA-CT, which converts the optimal single-armed policy into th...	Rebuttal 1: Rebuttal: We thank the reviewer for the constructive comments. We recap the comments and present our responses below. We will revise our paper accordingly. > Assumption of homogeneous arms. We address this comment in the global rebuttal, as this is an important point. In a nutshell, our approach and resul...	Rebuttal 1: Rebuttal: We thank all the reviewers for their time and effort. In this global rebuttal, we address two points each raised by multiple reviewers. Our responses to each reviewer are provided in the individual rebuttals. ## Comparison with prior work by Ghosh et al. (@Reviewers s6TA, S8QK, TTwr) We thank Rev...	NeurIPS_2023_submissions_huggingface	2,023	Summary: This work studies the restless bandit problem. In this setting the agent controls a set of $N$ arms modelled as Markov Decision Processes with two actions (active, inactive). The MDPs are coupled with a constraint that forces the agent to maintain a fixed arm activation rate. The authors propose a novel algori...	Rebuttal 1: Rebuttal: We thank the reviewer for appreciating our work, and for the suggestion on further clarifying the role and importance of SA. Your understanding is well aligned with the intended role of SA and with our proofs. We will integrate your suggestion into Section 3.3, specifically within lines 185-193 an...	Summary: This paper studies the problem of optimizing a policy for restless bandits in the infinite horizon case. A new algorithm is proposed that can convert a single-arm policy into an N-arm policy using virtual trajectories of the states and actions of the single-arm policy. The main idea is to generate virtual acti...	Rebuttal 1: Rebuttal: We thank the reviewer for the constructive comments. Below we quote the reviewer's key comments and present our responses. > Compare the synchronization assumption (SA) with conditions in other papers, and better explain the significance of SA. Before presenting the comparison, we reiterate...	null	null	null	null
Tanimoto Random Features for Scalable Molecular Machine Learning	Accept (poster)	Summary: Random feature map can overcome the limitation of kernel methods, and reduces its computational cost to linear in the number of training data points. Recent advances in random feature map broaden its application to not only to stationary kernels, but also to non-stationary kernels such as product kernels. The ...	Rebuttal 1: Rebuttal: Thank you for reviewing our paper. We will try to respond to your questions / concerns below. > Uncertainty on the correctness of the error bound” This reviewer finds this a serious point. What is the reason behind this ? Bound not tight enough, or bug in experiments ? Please describe what the au...	Summary: The paper investigates two low-rank approximations for the tanimoto similarity in the context of molecules. The two approximations are both random features newly proposed in the paper using random hashes and sketches. As theory, the paper contributes a proof of optimality for a certain way of constructing one ...	Rebuttal 1: Rebuttal: Thank you for your kind review and very positive rating. We will try to respond to your questions / concerns below. > How relevant are hand-crafted kernels? Although hand-crafted kernels are less relevant than a decade ago, they are still very relevant in applications with small datasets (where ...	Summary: This work develops random feature maps to approximate the Tanimoto coefficient among molecular datapoints. Approximating the Tanimoto coefficients allows this measure to be used as a kernel in classical machine learning approaches on larger datasets such as clustering and Bayesian optimization. Strengths: The...	Rebuttal 1: Rebuttal: Thank you for your kind review and very positive rating. We will try to respond to your questions / concerns below. > With the continued march of progress in deep learning algorithms---and their application to domains of molecular science---it is likely this approach will not see significant stan...	Summary: This paper presents two methods for approximating a Tanimoto kernel using random features, shows that the approximations are good, and validates the utility of the proposed approach by applying it to three different settings using molecular fingerprint data. The writing is clear, and the contributions are cle...	Rebuttal 1: Rebuttal: Thank you for your kind review and very positive rating. We will respond to the negative points and questions from your review below. > One of the main theoretical result is really just to show that Equation (6) is positive definite not just on binary inputs (which was shown in 2005) but for real...	Rebuttal 1: Rebuttal: We would like to thank all reviewers for their constructive comments. We have left a response to each reviewer answering their questions, but would like to provide a summary here. First, we appreciate that many reviewers complimented various aspects of our work including: - Theoretical analysis ...	NeurIPS_2023_submissions_huggingface	2,023	Summary: The paper derives two random feature maps for Tanimoto kernels of fingerprint data. The feature maps are applied in GP and BO settings to allow accelerated primal learning. Strengths: The idea of featurising Tanimoto is clever and novel. The two feature maps are seemingly well derived, and theoretical analysi...	Rebuttal 1: Rebuttal: Thank you for your thorough review and thoughtful comments. We appreciated that you found our work to be “clever and novel” and recognized that it “open[s] the door for feature-based learning in [the] GP setting, which is potentially significant.” Due to the 6000 character limit for the rebuttal w...	null	null	null	null	null	null
Improvements on Uncertainty Quantification for Node Classification via Distance Based Regularization	Accept (poster)	Summary: This paper focuses on uncertainty quantification in node classification. The authors examine the drawbacks of UCE. Inspired by our theoretical analysis, they introduced a distance-based regularization technique to obtain a better representation network for the uncertainty quantification. Experimental results s...	Rebuttal 1: Rebuttal: Questions: The authors should better discuss the limitation of their work. We discussed the limitations of Theorem 1 in lines 159 to 165 on Page 4 and the limitations of Theorem 8 in lines 226-234 on Page 6. In addition, our proposed regularization terms proposed in Section 4.2 are most effective...	Summary: This manuscript studies the uncertainty quantification for the non-independent node-level prediction problem. The proposed framework is built upon the graph posterior network (GPN) with minimizing uncertainty cross-entropy as the main loss function. A distance-based regularization technique is further proposed...	Rebuttal 1: Rebuttal: Weakness 1 (W1): How can Eq. (1) and (2) derive the aleatoric uncertainty and epistemic uncertainty? The authors do not mention the rationale behind this derivation, which makes it hard to understand the whole process. The aleatoric and epistemic uncertainty measures defined in Equation 2 are the...	Summary: The paper proposes a distance-based regularization to improve the OOD detection and misclassification detection tasks on the graph node classification. The regularization enforces the clustered OOD nodes to be close in the latent space. Strengths: - The paper includes comprehensive experiments on multiple co...	Rebuttal 1: Rebuttal: Weakness 1 (W1). The aleatoric and epistemic uncertainty measures defined in Equation 2 are the same measures used in the PN and GPN papers. Please refer to Section 2 of the PN paper and Section 3.3 in the recent TMLR’23 survey paper by Ulmer et al. for more detailed discussions on these Dirichl...	Summary: ## Post Rebuttal Update I have engaged with the authors for the rebuttal, and found their responses informative, prompting me to increase my score from a 7 to an 8. ## Original Review The paper proposes theoretical results for Graph Posterior Networks (GPNs) that use the uncertainty cross-entropy loss (UCE...	Rebuttal 1: Rebuttal: Weakness 1 (W1). We have two assumptions in Theorem 1: (1) the underlying distribution of feature vectors belonging to class $k$, denoted by $X_k$, is disjoint to each other; (2) the multi-layer perception (MLP) module, $f_\theta$, and the normalizing flow module, $h_\phi$, in the posterior ne...	Rebuttal 1: Rebuttal: We append the ROC-PR curves for our model related to Table 2 in the attached pdf. Pdf: /pdf/e19e448e6d6d80667a458a8749becb6cc88d9e1d.pdf	NeurIPS_2023_submissions_huggingface	2,023	null	null	null	null	null	null	null	null
HEDNet: A Hierarchical Encoder-Decoder Network for 3D Object Detection in Point Clouds	Accept (poster)	Summary: This paper focuses on 3D object detection using outdoor LiDAR data. The authors proposed a multi-scale network that is built upon several encoder-decoder blocks, e.g., SED and DED. The motivation behind it is to use sub-manifold sparse feature down-sampling and up-sampling steps to capture long-range voxel rel...	Rebuttal 1: Rebuttal: #### Q1. Might lack of insights or theoretical contributions? Not sure the findings/designs are generalizable to other 3D deep learning tasks? To the best of our knowledge, not all works published at NeurIPS prioritize theoretical contributions. In fact, NeurIPS has been a platform for the public...	Summary: The paper proposed a Hierarchical Encoder-Decoder Network, which faciliated sparse encoder-decoder block to enhance the network to model long-range dependencies among sparse voxel features while maintaining high sparsity. The method consistently improve the detection performance with a higher detection rate. ...	Rebuttal 1: Rebuttal: #### Q1. How the SED block can effectively reduce the computation and maintain the sparsity? 1. As you said, each SED block encompasses SS + RS_1(down) + SS + RS_2(down) + SS + UP_2 + UP_1. To distinguish different Down and Up layers, we add subscripts here. We adopt the sparse inverse conv implem...	Summary: This paper focuses on the long range problems in 3D object detection. Compared to the common idea of large kernel or transformers, this hierarchical design is more smart and efficient. It designs an encoder-decoder style network for 3D object detection. Comprehensive experiments on nuScenes and Ways show the e...	Rebuttal 1: Rebuttal: #### Q1. The visualization comparison on receptive fields should be provided. As this work claimed that it achieves similar or better long-range detection capability than large kernel and transformers. We have presented some visualization examples in the provided PDF file (refer to `Author Rebutt...	Summary: The proposed method considers a much more straightforward approach to incorporating larger context for 3D detection from point clouds. In contrast to existing work which consider large kernels or transformers to increase the receptive field of 3D backbones, this work focuses on developing an encoder-decoder bl...	Rebuttal 1: Rebuttal: #### Q1. A more fine-grained analysis on nuScenes? \|\|\|\|\|\|\|\|\|\|\|\|\| \|-\|-\|-\|-\|-\|-\|-\|-\|-\|-\|-\|-\| \|Model\|mAP\|Car\|Truck\|Cons.veh\|Bus\|Trailer\|Barrier\|Motor.\|Bicycle\|Ped.\|Tr.Cone\| \|TransFusion-L\|65.2\|86.5\|60.8\|25.7\|74.2\|41.6\|70.2\|71.6\|56.8\|87.1\|77.1\| \|HEDNet (Ours)\|66.9\|87.8\|61.4\|27.7\|77.6\|48.0\|69.9\|72.6\|5...	Rebuttal 1: Rebuttal: To All Reviewers: Thank you so much for your affirmations and insightful feedback on our work. We are very glad to receive your constructive suggestions. Here, we give more discussions and explanations on those concerns mentioned by each reviewer respectively. To Reviewer vKEU: The visualizatio...	NeurIPS_2023_submissions_huggingface	2,023	Summary: The paper addresses the task of 3D object detention on real-world 3D inputs that possess challenges such as scalability to long-range dependency. The main suggestion is to incorporate encoder-decoder type blocks to capture long-range dependency, in contrast to previous works that were limited to the sparsity p...	Rebuttal 1: Rebuttal: #### Q1. It is not clear what part of the design choices stemmed from the fact the architecture is aimed to tackle object detection? And the comment on the limitation section is a bit vague? Since we only conducted experiments on 3D object detection in the paper, we didn't know whether the propose...	null	null	null	null	null	null
On the Exploration of Local Significant Differences For Two-Sample Test	Accept (poster)	Summary: This paper proposes a bidirectional two-sample test that can address both positive and negative gaps from one sample to the other. The test is proposed based on Hotelling's statistic, and the direction information is encoded via the mean difference. By searching test locations and an adequate Mahalanobis kerne...	Rebuttal 1: Rebuttal: [Q1] Tradeoff between local difference and overall sample complexity … seems not to address the tradeoff between the decreasing problem's difficulty and the decreasing sample complexity used to compute the statistic. [A1] We will clarify that we take the number of Mahalanobis kernels and number o...	Summary: The paper introduces a two-sample test that explores local variations in distribution. The authors provide a motivation for this specific formulation by highlighting two practical scenarios in cosmology and biology. They proceed with a literature review and conclude that none of the existing approaches is full...	Rebuttal 1: Rebuttal: [Q1] It is unclear which elements of the design (bi-directional p-values, tree partition, kernel design) contribute to overall performance of the algorithm. [A1] We will clarify that the bi-directional p-values is proposed to improve test power by exploiting directional information between mean e...	Summary: In recent years, the focus on two-sample testing has significantly increased due to its wide range of practical applications. This study advances this exploration by highlighting local significant differences in two-sample testing. The authors introduce the ME, a robust test for two-sample testing that leverag...	Rebuttal 1: Rebuttal: [Q1] To show the difference between the deep-kernel-based method and your proposed method, I can understand the merit of the proposed method, but more discussions are needed in this paper. [A1] We will add more discussions that Liu et al. (ICML2020) presented a deep kernel to adapt for variations...	Summary: This paper investigates the exploration of local significant differences in the context of two-sample tests. The authors propose a new test called ME_{MaBiD}, which utilizes multiple Mahalanobis kernels and introduces a bi-directional hypothesis to exploit local information. The embedding space is partitioned ...	Rebuttal 1: Rebuttal: [Q1] I think the weakness lies mainly in clarity. It fails to give a formal definition of the input and output of the problem and defined further what is counted as local or not. [A1] We will present a formal definition of the problem, and clarify that the input includes two i.i.d. samples from t...	Rebuttal 1: Rebuttal: We want to express our gratitudes to reviewers for their insightful comments, and we will improve the paper accordingly. We present the detailed responses for each review (below), and upload a PDF file with two figures, as suggested by a reviewer. Pdf: /pdf/962c2de7732e131ac3b6e9cc557fd5015e20d6...	NeurIPS_2023_submissions_huggingface	2,023	null	null	null	null	null	null	null	null
Enhancing Adversarial Robustness via Score-Based Optimization	Accept (poster)	Summary: This paper proposes a dynamic defense named ScoreOpt, which maximizes the likelihood of input images by minimizing the score-matching loss. To regulate the optimization, the authors propose Score Regularization Loss which bound the distance between the denoised original images and denoised purified images. St...	Rebuttal 1: Rebuttal: Thank you for your helpful comments and constructive suggestions. Our responses are as follows point by point. > The baseline model, DiffPure with Wide-ResNet-28, achieves 81.40% robustness in your paper. However, according to [1], the original paper where this baseline was proposed, it only ach...	Summary: This paper introduces an adversarial defense scheme named ScoreOpt, which optimizes adversarial samples at test-time, towards original clean data in the direction guided by score-based priors. Comprehensive experiments are conducted on multiple datasets, including CIFAR10/100 and ImageNet, demonstrating that S...	Rebuttal 1: Rebuttal: Thank you for your helpful feedback and positive recommendation. Our responses are as follows point by point. > I cannot find the values of iteration steps $M$ and $N$ used for Ours(o) and Ours(n) in, e.g., Tables 1, 2 and 4. Do the Ours methods and other AT/AP baselines require comparable infere...	Summary: This paper proposes an adversarial defense scheme named ScoreOpt, which uses purification technology to eliminate adversarial perturbations. Specifically, ScoreOpt formulates the adversarial defense as an optimization problem, which optimizes adversarial examples based on pre-trained score-based priors. And ex...	Rebuttal 1: Rebuttal: Thank you for your helpful feedback and constructive suggestions. We will address your concerns below. > In this paper, the pre-designed noise plan is focused on lower noise levels. However, some clear explanations are necessary for this lower noise level. For example, how this lower noise level ...	Summary: The paper proposes a new approach for defending against adversarial attacks using adversarial purification, i.e., removing the adversarial noise from the input image at test time. To that end, the authors leverage diffusion models -- which have been used recently for adversarial purification -- and propose sev...	Rebuttal 1: Rebuttal: Thank you for your helpful feedback and positive recommendation. Our responses are as follows point by point. > The presentation of the ideas needs to be improved... (A1) Thanks for your useful suggestions. We will refine the methodology section in the revised version. > The authors mix notatio...	Rebuttal 1: Rebuttal: We sincerely appreciate the thoughtful comments from all the reviewers. In response to some general concerns raised by the reviewers, we provide a unified reply here. And we address specific questions and feedbacks individually. Derivation Details. Some of the reviewers mentioned that the no...	NeurIPS_2023_submissions_huggingface	2,023	null	null	null	null	null	null	null	null
Diff-Instruct: A Universal Approach for Transferring Knowledge From Pre-trained Diffusion Models	Accept (poster)	Summary: The paper discusses the distillation of pre-trained diffusion models. The paper introduces a framework called Diff-Instruct, which allows for the transfer of knowledge from pre-trained DMs to other generative models in a data-free manner. Diff-Instruct utilizes a mathematical foundation based on minimizing a d...	Rebuttal 1: Rebuttal: Thank you for your useful feedback. We will address your concerns one by one in the following paragraphs. Q1. it remains unclear whether the method will be effective with larger images A1. We agree that larger-scale experiments, such as the text-to-image generation experiments are import...	Summary: This work distills diffusion models to a GAN-style generator. The author provides mathmatic connection to other works, as well as experimental comparisons between the distilled GAN-style generator with other (mostly) diffusion models. Strengths: - This work introduce a way to distill diffusion model to GAN...	Rebuttal 1: Rebuttal: Thank you for your reviews. We will address your concerns one by one in the following paragraphs. Q1. Lack of baseline method. A1. We are sorry for the confusion. In Table 1 and Table 2 in the main text, the CD, PD, and Diff-Instruct use the same teacher model and the same student mode...	Summary: The paper introduces a method to transfer the knowledge of a pre-trained diffusion model to an implicit generator model, such that the distribution of the generated samples from the generator matches that of the pre-trained DM. The training is data free, and it can be divided into two alternating phases: train...	Rebuttal 1: Rebuttal: Thank you for your valuable suggestions, we will incorporate them in the revision. In the following paragraphs, we will address your concerns one by one. Q1. The presentation of the algorithm can be improved. I am not very clear about how to estimate the integral in the gradient formulation ...	Summary: The paper proposes a method for data free knowledge distillation from large scale diffusion models. It tries to distill the knowledge contained into a diffusion model, into another diffusion model or an explicit image generation model like a GAN. They propose a novel divergence measure obtained by integrating ...	Rebuttal 1: Rebuttal: Thank you for your reviews. We will address your concerns one by one in the following paragraphs. Before that, we first give a summary of the main contributions of our work. In this work, we introduce the Diff-Instruct, which can utilize a pre-trained teacher diffusion model to train a student o...	Rebuttal 1: Rebuttal: Thank all reviewers for your valuable feedback. In the rebuttal period, we run an additional experiment to compare the computational and memory costs of Diff-Instruct with our baseline method Consistency Distillation in Table 3. The test was run on 2 Nvidia V100 GPUs with 128 batch size and PyTorc...	NeurIPS_2023_submissions_huggingface	2,023	Summary: The paper proposes a general framework for distilling a pretrained diffusion model into an arbitrary one-shot latent-variable model. The objective is a reverse KL divergence integrated over time, and an extra score model is learned for the current model distribution. The generator model parameter and the score...	Rebuttal 1: Rebuttal: Thank you for your valuable suggestions, we will take them in the revision. In the following paragraphs, we will address your concerns one by one. Q1. If the implicit model is randomly initialized with an arbitrary architecture (e.g. a StyleGAN generator), would this approach work? A1. ...	null	null	null	null	null	null
Bridging the Domain Gap: Self-Supervised 3D Scene Understanding with Foundation Models	Accept (poster)	Summary: The paper proposes a pre-training strategy to distill the knowledge for large image foundation models. The method proposed is based on a mask auto-encoder transformer model and the setup contains several modules or losses. First, the masking strategy is guided by segmentation labels obtained using foundation m...	Rebuttal 1: Rebuttal: With gratitude for the reviewer's valuable insights, we will make appropriate revisions to our paper. In the following section, we address each of the reviewer's questions and comments in detail. W1: The novelty of this method **1, Difference to the Language-Grounded Indoor 3D Semantic Segme...	Summary: This paper aims to improve the pre-training of 3D point cloud models by leveraging various foundational models, including vision models such as SAM and Grounding Dino, language models like ChatGPT, and vision-language models. To achieve this objective, the paper proposes several techniques: 1. Generating insta...	Rebuttal 1: Rebuttal: With gratitude for the reviewer's valuable insights, we will make appropriate revisions to our paper. In the following section, we address each of the reviewer's questions and comments in detail. W1: Experimental results are not sufficiently convincing **1, Compare with the best performanc...	Summary: a novel masking and patch-dropping strategy based on foundation models to refine the focus of the network on foreground representation learning for 3D masked autoencoders. a novel scene-level and object-level multi-modality knowledge distillation method that pre-trains a 3D network via features, semantic mas...	Rebuttal 1: Rebuttal: With gratitude for the reviewer's valuable insights, we will make appropriate revisions to our paper. In the following section, we address each of the reviewer's questions and comments in detail. W1: Lack of visualization results Thank you for the valuable feedback regarding the absence o...	Summary: This paper proposes a combination of several techniques to pre-train 3D models by leveraging 2D foundation models. Specifically, the paper proposes: a) an MAE pre-training guided by foreground masks generated by foundation models to mask the foreground masks more compared to the background, to encourage the mo...	Rebuttal 1: Rebuttal: With gratitude for the reviewer's valuable insights, we will make appropriate revisions to our paper. In the following section, we address each of the reviewer's questions and comments in detail. W1: The complexity. 1, Complexity of the framework. The complexity observed in our approach ...	Rebuttal 1: Rebuttal: We appreciate the time and effort spent reviewing our submission. In this response, we address the major concerns raised by the reviewers and respond to other concerns in individual rebuttal responses. Reviewer XLcq, MtnQ: The limitations of the paper not discussed Thank you for your thought...	NeurIPS_2023_submissions_huggingface	2,023	Summary: This paper proposes a methodology named Bridge3D to pre-train 3D models using features, semantic masks, and captions sourced from foundation models. Experiments have been conducted on different 3D understanding frameworks. Strengths: The overall architecture seems to be reasonable; using the modern foundation...	Rebuttal 1: Rebuttal: With gratitude for the reviewer's valuable insights, we will make appropriate revisions to our paper. In the following section, we address each of the reviewer's questions and comments in detail. W1: The complexity of the framework. Concerns on complexity. The complexity observed in our ...	null	null	null	null	null	null
Unified Convergence Theory of Stochastic and Variance-Reduced Cubic Newton Methods	Reject	Summary: The authors propose a new way of solving minimization problems that are not necessarily convex. The main element of their approach is that they use some helper functions which give auxiliary information about the optimization problem at hand. That enables them to give improved optimization algorithms. Strengt...	Rebuttal 1: Rebuttal: Thank you very much for the invested time and expertise, and for the positive feedback as well. We hope that all raised issues are properly addressed in our rebuttal. - Line 28: A saddle point would not be desirable if our goal is to find a minimum, finding a saddle point means that we misse...	Summary: In this paper, the authors introduce a meta-algorithm for second-order methods which they refer to as the ``helper framework''. Their helper framework unifies stochastic and variance-reduced versions of Newton's method, and the framework could also be viewed as auxiliary tasks when applied to first-order metho...	Rebuttal 1: Rebuttal: Thank you very much for the invested time and expertise, and for the positive feedback as well. We hope that all raised issues are properly addressed in our rebuttal. 1. Helper framework. Using a unified framework is always beneficial because it helps us see different approaches as connected....	Summary: The paper provides a unified analysis for the Cubic regularized Newton Methods with variance-reduced estimators to solve stochastic nonconvex optimization problems. The results rely on the helper framework which generalizes the use of stochastic estimators for first-order and second-order oracles. The analysis...	Rebuttal 1: Rebuttal: Thank you very much for the invested time and expertise, and for the positive feedback as well. We hope that all raised issues are properly addressed in our rebuttal. Weaknesses: - Yes, for the update of the snapshot we use options (4) and (5), the main reason is that we needed these two opt...	Summary: The paper studies stochastic Cubic Newton methods for solving possibly nonconvex minimization problems. The authors propose a flexible “helper framework” that unifies stochastic and variance-reduced second-order algorithms, offering global complexity guarantees. This framework allows arbitrary batch sizes, noi...	Rebuttal 1: Rebuttal: Thank you very much for the invested time and expertise, and for the positive feedback as well. We hope that all raised issues are properly addressed in our rebuttal. Addressing weaknesses: 1. Related works. We did not discuss various variance-reduction first-order methods because the ma...	Rebuttal 1: Rebuttal: A table comparing the convergence rates of different algorithms discussed in the paper. Pdf: /pdf/42fc561f53d3395f4ed287e0cd1faa96bbccb401.pdf	NeurIPS_2023_submissions_huggingface	2,023	null	null	null	null	null	null	null	null
Sequential Preference Ranking for Efficient Reinforcement Learning from Human Feedback	Accept (poster)	Summary: The authors propose that a novel RLHF framework called SeqRank to improve the feedback efficiency via sequential preference ranking. They propose two methods: sequential pairwise comparison and root pairwise comparison. Strengths: The paper is clear and writing is easy to follow. The main idea is theoreticall...	Rebuttal 1: Rebuttal: We truly appreciate the positive and thoughtful review by Reviewer f3sA, _e.g., clear and easy writing, theoretically sound idea, and promising experiment results_. We will address all raised concerns by the reviewer and revise the paper accordingly. ## Weakness **Weak1: The proposed method s...	Summary: This paper introduces a novel approach, named SeqRank, to enhance the efficiency of reinforcement learning from human feedback. By incorporating sequential preference ranking, the proposed method significantly improves average feedback efficiency. Evaluations demonstrate its superiority over existing baseline...	Rebuttal 1: Rebuttal: We truly appreciate the positive and constructive review by Reviewer Mtqb, _e.g., strong mathematical justification, validation of the effectiveness of the proposed solution in realistic settings_. We will address all raised concerns by the reviewer and revise the paper accordingly. ## **Weakness...	Summary: This paper studies the idea of the feedback efficiency of RLHF, which the authors define as the ratio between the number of trajectory pairs divided by the number of feedbacks. To optimize such a metric, the authors propose new algorithms that decide which trajectories to compare. In particular, the authors pr...	Rebuttal 1: Rebuttal: We truly appreciate all feedback from Reviewer tocg and will incorporate the reviewer’s comments to revise our paper to be more sound and clear. Since the reviewer’s concerns are mainly grounded in the InstructGPT paper [1], we would like to mention three major differences between the robotics dom...	Summary: This paper presents SeqRank a framework for labelling user preferences in a ranked fashion taking advantage of the transitive property of human preferences. Authors mathematically prove that their two variants of SeqRank (sequential pairwise comparison, and root pairwise comparison), achieve better feedback e...	Rebuttal 1: Rebuttal: We truly appreciate the positive and thoughtful review by Reviewer sxrB, _e.g., robust mathematical analysis and compelling results in a wide variety of tasks_. We will address all raised concerns by the reviewer and revise the paper accordingly. ## Weaknesses **Weak1: “Authors only test SeqR...	Rebuttal 1: Rebuttal: ## Global Response to All Reviewers (Part I) We sincerely appreciate all the reviewers for their thoughtful and constructive comments. We are encouraged that they found our motivation and idea clear, intuitive, and novel (sxrB, Mtqb, f3sA) to solve an important and timely problem in RLHF (toc...	NeurIPS_2023_submissions_huggingface	2,023	null	null	null	null	null	null	null	null
Debiasing Conditional Stochastic Optimization	Accept (poster)	Summary: This paper considers CSO problems and its finite-sum variant FCCO, and proposed new algorithms incorporating extrapolation and variance reduction. The proposed algorithms are shown to enjoy a better sample size in the inner layer which induces an improved sample complexity compared to existing CSO literature. ...	Rebuttal 1: Rebuttal: We thank the reviewer for careful proofreading and suggestions. We will incorporate the suggestions about writing. Please find below our responses to the specific questions and weaknesses that you mention. >**The additional regularity assumption (compared to [15]) on the bounded higher-order gr...	Summary: This paper presents a stochastic extrapolation technique with variance reduction to obtain improved sample complexity bounds for conditional stochastic optimization problem. Numerical study also support the effectiveness of the proposed algorithm. Strengths: - The stochastic extrapolation is a useful tool for...	Rebuttal 1: Rebuttal: We thank the reviewer for careful proofreading and suggestions. We will add the iteration complexities in the revision. >I am also curious about how the authors come up with this approach? Do authors inspired from any existing literature? We independently came up with this stochastic extrapo...	Summary: The paper highlights that optimizing conditional stochastic optimization problems is challenging as the sample-averaged gradient is biased. The biased gradient increases the sample complexity needed to achieve convergence. To account for the bias, the paper proposes stochastic extrapolation. The method shows t...	Rebuttal 1: Rebuttal: Thank you for your positive feedback. We also thank the reviewer for the suggestions on the writing and readability. We will incorporate the suggestions accordingly. Please find below our responses to the specific questions and weaknesses that you mention. >**The paper has minimal numeric studie...	Summary: The authors study stochastic optimization problems with nested conditional expectations, with access to first-order samples from the involved functions. With such an oracle access, it is non-trivial to build unbiased estimates of the gradient, hence normal stochastic optimization methods don't work. They aut...	Rebuttal 1: Rebuttal: Thank you very much for your encouraging comments and valuable feedback. Please find below our responses to the specific weaknesses that you mention. >last term of Eq (2) The reviewer is correct, there is a typo in Eq. (2), $\phi_\delta$ should be inside the expectation, and we will fix...	Rebuttal 1: Rebuttal: We thank all the reviewers for constructive and actionable feedback on the paper. We have additional experimental results, in particular, addressing the question of what happens if the regularity assumption gets violated. Description of Fig. 5 in the attached pdf: We consider an experimenta...	NeurIPS_2023_submissions_huggingface	2,023	Summary: The paper addresses the conditional stochastic optimization (CSO) problem. A noted drawback in CSO is the bias in its sample-averaged gradient, an offshoot of its nested structure, which necessitates an extensive sample complexity to achieve convergence. The authors introduces a new stochastic extrapolation te...	Rebuttal 1: Rebuttal: Thank you for appreciating our work. Please find below our responses to the specific weaknesses that you mention. >While the new bounds show improved rates, the results are based on additional regularity assumptions ... We agree that the regularity assumptions are for technical convenience ...	null	null	null	null	null	null
Learning Efficient Surrogate Dynamic Models with Graph Spline Networks	Accept (poster)	Summary: The paper builds upon a recent advances in deep-learning based collocation methods for physical simulation to improve prediction accuracy in continuous space and time. The authors proposed a hybrid model based on graph neural network and differentiable orthogonal spline collocation method. Interpolation in tim...	Rebuttal 1: Rebuttal: Thank you for your review and for acknowledging the strengths of our work in soundness, presentation, and contribution! We address your concerns point by point in the following sections. ### Why is collocation in time important? Your observation regarding the limited sampling of collocation poin...	Summary: The authors introduce GraphSplineNets, providing a framework for combining MPNNs like MGNs with interpolation/collocation methods. Their contributions also include two methods to surpass the performance of a naive combination: end-to-end training and collocation points adaptation. Temporal and spatial downsamp...	Rebuttal 1: Rebuttal: Thank you for the constructive and detailed review and thoughtful feedback! We provide responses to your concerns and questions below. ### Why do GraphSplineNets work well? Thank you for your question! This is an important point for our paper and as such we would like to clarify it. As you cor...	Summary: The authors propose a method based on deep learning that will speed up the prediction of physical systems by reducing the size of the grid and the number of tests to be carried out. The method consists of combining a graph neural network to predict some values with two orthogonal spline collocation, one spatia...	Rebuttal 1: Rebuttal: Thank you for your review and for acknowledging the strengths of our work in soundness, presentation, and contribution! We address your concerns point by point in the following sections. ### Availability of data for physical systems Your question delves into an important consideration. The avail...	Summary: The presented manuscript proposes a novel architecture to predict the evolution of the dynamical system. The proposed method has two key ingredients: orthogonal spline collocation for both time and spatial dimensions and adaptive selection of the collocation points. Extensive experiments demonstrate the perfor...	Rebuttal 1: Rebuttal: Thank you for the constructive review and feedback. We provide responses to your concerns and questions below. ### Too much attention to standard methods Thank you for your suggestion. We agree with you that certain sections detailing the standard method for solving structured linear systems can...	Rebuttal 1: Rebuttal: We thank all reviewers for providing valuable and constructive feedback on our work; we sincerely appreciate the time and effort dedicated to reviewing our manuscript. We are glad to receive such an overall positive evaluation on all aspects of the submission, including soundness, presentation, an...	NeurIPS_2023_submissions_huggingface	2,023	null	null	null	null	null	null	null	null
Should Under-parameterized Student Networks Copy or Average Teacher Weights?	Accept (poster)	Summary: This paper is a theoretical study of whether an under-parameterized student network should copy or average teacher weights, there teacher is a one-hidden-layer network with k hidden units and student is the same but with n<k hidden units. The paper derived the optimal “copy-average” configuration under a set o...	Rebuttal 1: Rebuttal: We thank the reviewer for their comments. Below, we address each point. We hope that the reviewer will find the results more intuitive. From a bird-eye view, our main contribution is the copy-average phenomenon that kicks in for underparameterized networks. On practical relevance, the results sugg...	Summary: The paper studies a shallow teacher-student setting in which a student network being trained attempts to recover the orthogonal weights of a teacher network, assuming normally distributed data. Specifically, the setting being studied is where the number of student neurons $n$ is far smaller than that of the te...	Rebuttal 1: Rebuttal: We thank the reviewer for the actionable feedback. Thanks also for acknowledging the fundamental difficulty of the teacher-student problem. Below, we respond to each feedback. Some of the results cover the full under-parameterization regime (n<k) which is much more general than what is understood ...	Summary: his paper proposes a constrained optimization approach to learning student networks $g(x) = \sum_{j=1}^n a_j \sigma (v_j \cdot x)$ that have fewer neurons than a two-layer teacher network $f^* (x) = \sum_{i=k} b_i \sigma( w_i \cdot x )$. (with $ n \le k $ as the student network has fewer neurons than the teach...	Rebuttal 1: Rebuttal: We thank the reviewer for the detailed feedback. Thanks for liking the results too! Below, we aim to address each one of the concerns. We are confident that the writing can be improved with minor updates on the text, following the suggestions of the reviewers. This would be a minor change overall ...	Summary: The goal of the paper is to understand the following approximation problem: given a student network and teacher network, how should one construct the student network such that the approximation error is minimized. The paper considers the underparameterized regime where the student network has fewer parameters ...	Rebuttal 1: Rebuttal: We thank the reviewer for the feedback and appreciating the novelty of the copy-average phenomenon. Below, we answer every question raised by the reviewer. We appreciate follow-up questions, if anything is unclear in the responses or if the reviewer has new questions. **Which of the $n-1$ teacher...	Rebuttal 1: Rebuttal: Regarding Figure 1 (response to ```Cg9m``` and ``` 6p1W``` ): $u_{ij}$ is the vector of correlations between the student incoming vector $w_i$ and the teacher incoming vectors $v_j$, i.e. $u_i = (u_{i1}, …, u_{ik})$. This notation is introduced in line 138 and added in the new caption (see th...	NeurIPS_2023_submissions_huggingface	2,023	null	null	null	null	null	null	null	null
Random Cuts are Optimal for Explainable k-Medians	Accept (oral)	Summary: The paper studies the optimal upper bound of explainable k-median and shows a tight analysis for the random coordinate cut algorithm. The explainable k-clustering model was first introduced by Dasgupta et al. [ICML’20] and was followed by a flurry of work to understand the optimal approximation factor to achie...	Rebuttal 1: Rebuttal: We thank the reviewer for suggesting how we can improve the presentation. We will address all your comments in the final version of the paper. 1. For Theorem 2.2, we do not assume the reference centers are optimal centers for k-medians. We show that for every set of k reference centers, the Rando...	Summary: The paper provides a tight analysis of the random-coordinate-cut (RCC shortly) algorithm for explainable k-medians in the $ \ell_1 $ metric. To be precise, the RCC algorithm was independently proposed by several works, and for the explainable k-medians in $ \ell_1 $ , the tightest proved competitive ratio is ...	null	Summary: The paper considers the problem of explainable $k$-medians which was recently introduced by Dasgupta et al, and analyzes the RandomCoordinateCut algorithm. The paper proves an upper bound of $2\ln k + 2$ on the competitive ratio of RandomCoordinateCut. This matches a previously known lower bound of $\Omega(\lo...	null	Summary: This paper considers the Explainable k-Medians problem, which is a more interpretable model for k-clustering than traditional models such as k-median. An explainable clustering partitions space based on coordinate threshold cuts, so that instead of a Voronoi diagram as in k-median / k-means, we end up with rec...	Rebuttal 1: Rebuttal: Yes, our result can be used for explainable k-means. Makarychev and Shan (2021) provided a cut-preserving terminal embedding from $\ell_2^2$ to $\ell_1$ with distortion $O(k)$. Thus, we can first run the terminal embedding and then use the Random Coordinate Cut algorithm on the new instance. Our a...	Rebuttal 1: Rebuttal: First of all, we would like to thank all the reviewers for their valuable feedback. We will address all their comments and implement all their suggestions in the final version of the paper. Several reviewers asked us to compare our proof and the proofs obtained independently by Gupta, Pittu, Sve...	NeurIPS_2023_submissions_huggingface	2,023	null	null	null	null	null	null	null	null
Imbalanced Mixed Linear Regression	Accept (poster)	Summary: The paper proposes a novel method for estimating parameters in a mixture of linear regression (MLR) model. The authors argue that the maximum likelihood estimator (MLE) does not perform well under extreme imbalanced cases and propose a sequential estimation approach using robust regression methods. The paper p...	Rebuttal 1: Rebuttal: We thank the reviewer for their detailed and constructive feedback. Below we address each point separately. We will carefully revise our manuscript according to the points raised by the reviewer. Weaknesses: * Notation and Terminology: - By "significantly imbalanced" we mean that the mixtu...	Summary: This paper introduces a novel algorithm for mixed linear regression suitable for balanced and imbalanced mixture settings. It proves the proposed method holds for a sufficiently imbalanced mixture rather than a balanced one through a recovery guarantee on the theoretical front. Empirical results on synthetic d...	Rebuttal 1: Rebuttal: We thank the reviewer for the helpful comments. * We thank the reviewer for bringing this issue to our attention. Our goal was to emphasize the advantages of our method, but we understand that the reader might miss our main message. We will revise our abstract and introduction and focus on our me...	Summary: In this paper, the authors take into account the problem of recovering a mixture of $K$ unknown linear regressions with imbalanced proportions, i.e. the numbers of samples for components in the mixture are substantially different. Rather than try to recover $K$ components at the same time as in previous work, ...	Rebuttal 1: Rebuttal: We thank the reviewer for their constructive feedback. Weaknesses: 1. Unfortunately, we do not currently have a finite-sample analysis or a robustness to outliers guarantee for our algorithm. We agree that these two guarantees would greatly improve the theoretical grounding of our method, and...	Summary: In this paper the authors study the problem of mixed linear regression, where instead of a single linear regression model, the response variable comes from a finite mixture of them. The authors propose an algorithm for estimating the regression coefficients, especially in the imbalanced case where the mixing p...	Rebuttal 1: Rebuttal: We thank the reviewer for their insightful comments. Weaknesses: 1. We thank the reviewer for bringing this unclear point to our attention. It might not be so well-known, but in general, robust regression algorithms succeed even when the number of outliers largely exceeds the number of inlier...	Rebuttal 1: Rebuttal: We thank the reviewers for their time and effort. We appreciate the constructive feedback, which will definitely improve the manuscript. Please find attached a one-page PDF with new figures that help answer some of the reviewers' concerns. Following the reviewers' comments, we describe below few i...	NeurIPS_2023_submissions_huggingface	2,023	Summary: This paper proposes a new method, Mix-IRLS, for mixed linear regression, in which data are drawn from a mixure of linear models whose coefficients we wish to estimate. The method is especially motivated by the failure of existing methods in the case of imbalanced mixtures (i.e., when some components are much r...	Rebuttal 1: Rebuttal: We thank the reviewer for their thoughtful comments. Weaknesses: 1. Thanks for bringing this up. We performed an additional experiment based on a Kaggle CO2 emission from vehicles dataset to address the reviewer's concern. This dataset approximately follows an MLR model, where the samples can...	null	null	null	null	null	null
Nash Regret Guarantees for Linear Bandits	Accept (poster)	Summary: In this research, the authors investigate the concept of Nash regret in the context of stochastic linear bandits. Nash regret represents the disparity between the optimal reward and the geometric mean of expected rewards, reflecting fairness considerations across different time instances within the Nash social...	Rebuttal 1: Rebuttal: *The paper does not discuss the computational aspects of finding the G-optimal distribution and sampling via the John ellipsoid method.* Thank you for this relevant question. Our algorithm (LinNash) is in fact computationally efficient. A detailed runtime analysis is provided in the common re...	Summary: The authors consider stochastic linear bandits, and the goal is to minimize Nash regret, which is a stronger notion than the standard regret, by replacing the arithmetic mean reward of the learner over time with the geometric one in the definition. Strengths: The paper is well written. I didn't check the proo...	Rebuttal 1: Rebuttal: *no discussion about the Nash regret of multi-armed bandits, which is an important special case.* Perhaps the reviewer missed this, but Nash Regret in the context of stochastic multi-armed bandits (MAB) has already been addressed in [4]. The writeup extensively cites this prior work (see Lin...	Summary: The paper extends algorithmic techniques for the linear bandits problem to the Nash Social Welfare objective function which ensures a level of fairness rather than merely the utilitarian welfare. Their work provides essentially optimal regret guarantees as compared to the traditional linear bandits problem (as...	Rebuttal 1: Rebuttal: *The paper lacks a comprehensive discussion on the related work surrounding this problem.* As suggested in the review, we will add some discussion on online resource allocation (including online algorithms for NSW maximization). ***Can you provide a brief sketch of the final theorem (inf...	Summary: This paper considers the problem minimize 'Nash regret' in linear bandit, where Nash regret here is defined as the difference between the optimum and the geometric mean of expected rewards induced by the bandit algorithm. The stochastic reward in this paper is assumed sub-Poisson. The authors give algorithm to...	Rebuttal 1: Rebuttal: *The term "Nash regret" has been previously used in [1] and [2], but with a different definition from the one presented in this paper. The current name "Nash regret" in this paper does not capture the characteristic of the geometric mean.* The papers [1] and [2] consider Nash equilibrium and...	Rebuttal 1: Rebuttal: *Numerical Simulations* Some reviewers have asked about numerical simulations. In the limited rebuttal time period, we have compared our algorithm LinNash with Thomson Sampling on synthetic data. Note that we chose Thomson Sampling (Algorithm 1 in [1]) instead of UCB/OFUL because it is crucia...	NeurIPS_2023_submissions_huggingface	2,023	Summary: This paper proposes an approach based on the Nash regret to extend the stochastic linear bandits framework with fairness and welfare considerations. The Nash regret relies on the geometric mean of expected rewards induced by the linear bandit algorithm, which tends to incentivize uniformly high rewards as oppo...	Rebuttal 1: Rebuttal: *Although the paper is technically well-set, the contribution seems incremental.* The work brings forth (1) important conceptual ideas and (2) entails substantive technical steps. We do not find these results to be incremental and believe the current paper will inspire future research. 1) On...	null	null	null	null	null	null
Adversarial Attacks on Online Learning to Rank with Click Feedback	Accept (poster)	Summary: This paper presented a study of reward poisoning attacks on UCB bandit algorithms with binary feedback and applies the designed attack to the online-learning-to-rank (OLTR) problem. The binary reward structure poses unique challenges to reward poisoning attacks, since the attacker needs to ensure the perturbed...	Rebuttal 1: Rebuttal: Q1. Can the author make the assumption in Theorem 4 more prominent? A1. Thanks for the suggestion. We agree it is better to define the assumption that the ranking algorithm would choose suboptimal items sublinear times more explicitly. We will make it as a separate assumption in the upd...	Summary: The paper introduces adversarial attack on online learning to rank. The attacker aims to fool the online learning to rank algorithm to choose the target item. The paper studied three victim models: stochastic L-armed bandit, OLTR with position-based model and OLTR with cascade model. Authors proposed three att...	Rebuttal 1: Rebuttal: Q1. Trivial$\_{K-1}$ seems to keep the same cost scaling on time as the proposed attack strategies. A1. We admit that the attack cost of Trivial$\_{K-1}$ is in the same order of $O(\log T)$ as our proposed algorithms. However, our algorithms achieve problem-dependent attack costs adapt...	Summary: This paper studies adversarial attack to online learning to rank, where a successful attack to an OLTR algorithm is such that a suboptimal item is recommended for linear in T times while with log T cost. This paper approaches this attack problem by studying attack to discrete MAB as warmup and then moves on to...	Rebuttal 1: Rebuttal: Q1. Comparison with the other work [1]. A1. We thank the reviewer for pointing us to this related work. We kindly remind that the submission date of this arXiv paper is after the NeurIPS deadline. We would like to make some clarifications on the difference between our work and theirs. ...	Summary: This paper studies attack strategies against UCB-based OLTR algorithms for position-based click model and cascade click model. Compared with prior works in adversarial attacks, OLTR problem leads to the additional challenge that the rewards are discrete and bounded, which requires different attack method desig...	Rebuttal 1: Rebuttal: Q1. Intuitive explanation of conservative estimation. A1. The main idea of the attack algorithm in [9] is to calculate the attack value per round such that the post-attack empirical estimates of the non-target arms are always less than that of the target arm (i.e., Eq. (1)). However, wi...	null	NeurIPS_2023_submissions_huggingface	2,023	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.