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 |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
Adversarial Moment-Matching Distillation of Large Language Models | Accept (poster) | Summary: To improve knowledge distillation for large language models, the authors first motivate an RL-based formulation that aims to minimize the imitation gap while matching on and off-policy moment bounds, and then introducing an adversarial training algorithm that achieves this by posing it as a two-player minimax ... | Rebuttal 1:
Rebuttal: We sincerely appreciate your detailed and constructive comments, as well as your support of our work. We hope our responses adequately address your concerns.
**Q1. Is the comparison between SFT and the other methods fair in terms of compute budget? It might be interesting to see if increasing the... | Summary: This paper applies a reinforcement learning (RL) framework to the problem of auto-regressive text generation, framing knowledge distillation as a task of minimizing the imitation gap between teacher and student policies. The authors provide a theoretical analysis demonstrating that the proposed momentum-matchi... | Rebuttal 1:
Rebuttal: We sincerely appreciate your detailed and constructive comments, as well as your support of our work. We hope our responses adequately address your concerns.
**Q1. a. Could you provide a detailed comparison of resource consumption and memory costs relative to the baseline methods? b. Additionally... | Summary: The paper introduces a novel approach to knowledge distillation for Large Language Models (LLMs) using an adversarial training method that incorporates both on and off-policy distillation. The method jointly learns a critic that estimates Q-values while updating both the Q-function and the student model to mor... | Rebuttal 1:
Rebuttal: Thank you for taking the time to read our paper and for providing constructive comments. We hope our detailed responses will adequately address your concerns.
**Q1. How exactly is the Q-value function parameterized? Is it an extra head on the model, or a new model entirely?**
> Please refer to th... | Summary: This paper proposed an adversarial moment-matching approach for knowledge distillation of LLM. The idea is to reformulate the knowledge distillation from an imitation learning perspective and derive both on-policy and off-policy bounds for the imitation gap between the teacher and student models. The authors p... | Rebuttal 1:
Rebuttal: Thank you for taking the time to read our paper and for providing constructive comments. We hope our detailed responses will adequately address your concerns.
**W1. Ablation studies and analysis of the impact of on-policy and off-policy objective. And analysis how each of them effect the overall ... | Rebuttal 1:
Rebuttal: ## Global Response to Common Questions
**Q1. The parameterized estimation of Q-value function.**
> The Q-value function $f_{\phi}(y_{<t}, y)$, where $y_{<t}$ denotes the state of the $0:t-1$ tokens and $y \in \mathcal{V}$ denotes an action of the $t$-th token, was estimated with a neural network, ... | NeurIPS_2024_submissions_huggingface | 2,024 | null | null | null | null | null | null | null | null |
Queueing Matching Bandits with Preference Feedback | Accept (poster) | Summary: This paper proposes an algorithm that learn optimal allocation in a multi-arm bandit problem involving queues.
The algorithm ensures the system stability while having a sub-linear regret.
Strengths: The algorithm leverages on both max-weight and UCB (or TS) to provide stability and no-regret in a multi-arm b... | Rebuttal 1:
Rebuttal: We appreciate your time to review our paper and comments. Below, we address each comment.
**There are notation confusions:**
We appreciate your helpful comments in clarifying our paper. We will include a notation table in our final (Please refer to Table 1 in the attached PDF). As mentioned in l... | Summary: This study examines multi-class multi-server asymmetric queueing systems, where jobs arrive randomly, and unknown job-server service rates are modeled by a feature-based Multinomial Logit (MNL) function. The proposed UCB and Thompson Sampling algorithms aim to stabilize the queues while learning service rates,... | Rebuttal 1:
Rebuttal: We appreciate your feedback and positive evaluation of our work. Below, we address each of your comments.
**The quality of the figures can be improved:**
We will make sure to improve the quality of the figures in our final version.
**How about the performance and calculation efficiency when N... | Summary: The work introduces a new bandits framework that handles the problem of matching queuing jobs (agents) with preferential servers (arms). The work extends beyond the match making problem with the objective of stabilizing the queue by learning the preferential nature of agents to arms. The authors propose two ne... | Rebuttal 1:
Rebuttal: We appreciate your feedback and positive evaluation of our work. Below, we address each of your comments.
**The previous algorithm seems to work for the proposed algorithm:**
Our problem includes novel factors, which are inspired by the real-world such as online job markets or ride-hailing platf... | Summary: This paper studies queueing matching bandits. It proposes a framework that involves multiple queues and multiple servers: in each round, jobs arrive randomly at each queue; the learner assigns jobs to servers; and each server picks and serves its preferred job according to a feature-based linear model. The goa... | Rebuttal 1:
Rebuttal: We appreciate your feedback and positive evaluation of our work. Below, we address each comment.
**Require solving an NP-hard optimization:**
NP-hardness is common in almost all assortment (and many other combinatorial) optimization problems. Hence, we do not think that this is particularly a we... | Rebuttal 1:
Rebuttal: Thank you for taking the time to review our paper. We are encouraged by your feedback, noting that our problem is both novel and interesting, inspired by real-world challenges. We have attached a PDF that includes additional experimental results and a notation table, addressing the reviewers' comm... | NeurIPS_2024_submissions_huggingface | 2,024 | null | null | null | null | null | null | null | null |
IntraMix: Intra-Class Mixup Generation for Accurate Labels and Neighbors | Accept (poster) | Summary: The paper introduces IntraMix, a novel method for augmenting graph data to improve the performance of Graph Neural Networks. IntraMix addresses two major issues in graph datasets: the lack of high-quality labeled data and inadequate neighborhood information. The proposed method leverages Intra-Class Mixup to g... | Rebuttal 1:
Rebuttal: Thank you for your constructive comments and suggestions. Next is our response.
**Q1.The method involves several steps, which might complicate the implementation and increase the computational overhead.**
A1. Concerns about complexity are reasonable but unnecessary for IntraMix.
First, we anal... | Summary: This paper aims to improve the performance of graph neural networks (GNNs) for node classification problem by generating high-quality labeled nodes and enriching node neighbors. It first uses pseudo-labeling to transform the unlabeled nodes into low-quality labeled nodes and performs mixup to generate high-qua... | Rebuttal 1:
Rebuttal: Thank you for your constructive comments and suggestions. Next is our response.
**Q1. In Table 1, the improvements of the proposed method is not significant, which should be discussed more clearly.**
A1. Thank you for raising this question. We will discuss it further.
**Significance:** We belie... | Summary: This paper proposes an intra-class mixup generation method to generate high-quality labeled data to improve the performance in the node classification task.
Strengths: 1. The extensive experimental results show that the proposed IntraMixup outperforms most of the baseline methods across several datasets.
2. T... | Rebuttal 1:
Rebuttal: Thank you for your constructive comments and suggestions. Next is our response.
**Q1.Why does the proposed method achieves better performance than NodeMixup?**
A1. This is a great question. We discuss our advantages over NodeMixup in Sec 1, and Appendix E.2. We analyze the reasons in terms of a... | Summary: This paper propose IntraMix, a data augmentation approach for node classification with graph neural networks. IntraMix effectively mixes node features of nodes in the same class based on pseudo labels to generate new nodes, and then link the generated nodes to selected nodes in the graph. The authors conduct s... | Rebuttal 1:
Rebuttal: Thank you for your constructive comments and suggestions. Next is our response.
**Q1. The authors should clarify the distribution assumption on noises, since the constraint cannot be satisfied if $\epsilon_1,...\epsilon_{|C|}$ are i.i.d. Gaussian random variables.**
A1. We apologize if our expre... | Rebuttal 1:
Rebuttal: We thank all reviewers for their diligent efforts in evaluating our submission. We will revise the paper to address the comments raised by each reviewer. We are glad the reviewers find that
- Our method is elegant and interesting, efficiently addressing two major challenges in graph data augmenta... | NeurIPS_2024_submissions_huggingface | 2,024 | null | null | null | null | null | null | null | null |
Not All Tokens Are What You Need for Pretraining | Accept (oral) | Summary: The paper analyzes token-level training dynamics in continued pretraining, identifying four loss patterns: persistent low loss, persistent high loss, increasing loss, and decreasing loss. Motivated by these patterns, the paper proposes a modification to language modeling called Selective Language Modeling (SLM... | Rebuttal 1:
Rebuttal: Dear Reviewer pLtQ,
Thank you for your detailed review and thoughtful feedback, especially for finding our token-level loss research novel, our SLM idea clever and effective, and our analysis comprehensive!
### **W1. Data efficiency vs. Training time**
Thank you for your suggestion! We believe ... | Summary: The authors explore how loss for specific tokens changes in continued pre-training and note that they fall into four categories (high->high, high->low, low->high, low->low) with each category having at least 10%. They run continued pre-training on tokens that are learnable and domain-useful (judged by referenc... | Rebuttal 1:
Rebuttal: Dear Reviewer 3eTg,
Thank you for your thoughtful review and for recognizing the novelty, effectiveness, and significance of our work!
### **Definition of “number of tokens”**
> The greatest weakness (in my opinion) is that "tokens" in many cases could refer to "number of tokens after % filteri... | Summary: The authors propose a method to train LLMs on the most influential tokens selectively. They suggest training a reference model on a small high-quality corpus using the standard CLM loss. They then compute the excess loss of each token in the training corpus as a difference in losses of the reference model and ... | Rebuttal 1:
Rebuttal: Dear reviewer QeQV,
Thank you for your comprehensive review and positive remarks!
### **Pre-training from scratch**
> The experiments in the paper are performed in the continued pre-training setting and the impact of the original pertaining performance is not discussed in the paper. It is poss... | null | null | null | NeurIPS_2024_submissions_huggingface | 2,024 | null | null | null | null | null | null | null | null |
GFT: Graph Foundation Model with Transferable Tree Vocabulary | Accept (poster) | Summary: The paper proposes GFT, a graph foundation model based on computation tree. Extensive experiments and theoretical analyses are conducted to show the effectiveness of GFT across diverse tasks.
Strengths: 1. Addresses a significant challenge (identifying transferrable patterns) in graph foundation models.
2. In... | Rebuttal 1:
Rebuttal: We appreciate the reviewer’s recognition of our work and thank the reviewer for detailed feedback. We address your concerns as follows.
> W1: The experiments are unconvincing and most likely unfair.
[W1.1] *Outdated supervised baselines.* The baselines are not outdated, as our primary comparison... | Summary: This paper explores the concept of the transferable token in graph foundation models. Specifically, the paper proposes to use the computation tree as the transferable token for graph learning and prove its efficiency from both theoretical and empirical perspectives. Then, the paper proposes GFT model, GFT firs... | Rebuttal 1:
Rebuttal: We greatly appreciate Reviewer M9ZQ for the insightful feedback and are deeply inspired by your recognition of our contribution as “the first to demonstrate the potential existence of transferable tokens in text-attributed graphs.” We address your concerns as follows.
> W1: How to define the 𝑗th... | Summary: This paper proposed a novel computation tree method to improve the transferability between the pre-train model and downstream tasks. This paper rethinks the transferable pattern in graphs as computation trees and validate their transferability both empirically and theoretically. The proposed GFT leverages comp... | Rebuttal 1:
Rebuttal: Dear Reviewer U8xo,
We appreciate the reviewer’s recognition of originality, quality, clarity, and significance of our work, and thank the reviewer for detailed feedback. We address your concerns as follows.
> W1: Differences between computation trees and subgraphs.
Please refer to the [W1] in... | Summary: The paper proposes a new graph foundation model based on the tree structure, which is called GFT. GFT leverages computation trees to define tokens within the transferable vocabulary, which improves model generalization and reduce the risk of negative transfers. Comprehensive experiments and theoretical analys... | Rebuttal 1:
Rebuttal: Dear Reviewer H6gM,
We greatly appreciate your acknowledge of our contribution and insightful feedback to help us refine our work. We address your concerns as follows:
> W1: Is there any other structures transferable? How to demonstrate what tokens learn exactly?
>
Yes, there are other transf... | Rebuttal 1:
Rebuttal: Dear ACs and reviewers,
We thank the reviewers for their feedback and constructive suggestions. The positive feedback has truly inspired us, highlighting the importance of exploring transferable patterns on graphs (H6gM, M9ZQ, hEws), the novelty of our GFT (H6gM, U8xo, M9ZQ), the meaningfulness o... | NeurIPS_2024_submissions_huggingface | 2,024 | null | null | null | null | null | null | null | null |
Imprecise Label Learning: A Unified Framework for Learning with Various Imprecise Label Configurations | Accept (poster) | Summary: This paper introduces imprecise label learning (ILL), a framework for the unification of learning with various imprecise label configurations, such as partial label learning, semi-supervised learning, noisy label learning, and a mixture of these settings. They propose an EM based method with closed-form learni... | Rebuttal 1:
Rebuttal: Thanks for the reviewer's efforts reviewing this paper and acknowledge of our work. We address the concerns as follows.
> Given the extensive content, the main text of the paper omits certain details, which may not be immediately straightforward to follow for some parts, e.g., section 3.2.
Thank... | Summary: This paper introduces a framework that unifies various imprecise label configurations, with an EM modeling for imprecise label information. The framework is demonstrated can be adapted to partial label learning, semi-supervised learning, and noisy label learning, and the combinations of all above. The experime... | Rebuttal 1:
Rebuttal: Thanks for the reviewer's efforts reviewing this paper and acknowledge of our work. We address the concerns as follows.
---
> The implementation of EM over all possible labelings may increase the computation time.
For all the weak supervision settings studied in this paper, our method introduce... | Summary: The article addresses the challenge of learning with imprecise labels in machine learning tasks, such as noisy or partial labels. Traditional methods often struggle with multiple forms of label imprecision. The authors introduce a novel framework named Imprecise Label Learning (ILL) that serves as a unified ap... | Rebuttal 1:
Rebuttal: Thanks for the reviewer's time reviewing this paper and suggestions on the relevant works. We address the concerns raised as follows.
---
> The article's innovation is limited...which suggests that the presented techniques may not be as novel as claimed.
Thanks for mentioning these relevant wor... | Summary: The paper provides a unified view on various imprecise-label learning frameworks, such as semi-supervised-, partial-label-, or noise-label learning, through the lens of the expectation-maximization algorithm. In addition to unifying these existing setups, EM naturally allows treating combinations of the above ... | Rebuttal 1:
Rebuttal: We thank the reviewer for the detailed and informative review.
We recognize several of the concerns expressed by the reviewer and agree that addressing these will indeed enhance the paper. We have attempted to do so below. We hope this is satisfactory.
---
> Neither abstract...
The setting stud... | null | NeurIPS_2024_submissions_huggingface | 2,024 | null | null | null | null | null | null | null | null |
Local Superior Soups: A Catalyst for Model Merging in Cross-Silo Federated Learning | Accept (poster) | Summary: In this paper Local Superior Soups (LSS) is proposed to minimize communication rounds in federated learning (FL) using pre-trained models, specifically tackling data heterogeneity challenges. LSS achieves this by employing sequential model interpolation, maintaining connectivity, and integrating diversity and ... | Rebuttal 1:
Rebuttal: Thank you for your constructive feedback on our manuscript. We appreciate your positive comments on the effectiveness and soundness of our proposed method, as well as the clarity of our writing.
**Expermental Scope Clarification.**
We would like to clarify that the current set of experiments is s... | Summary: This paper proposes a method called Local Superior Soups (LSS), a novel technique for model merging in cross-silo federated learning aimed at reducing communication rounds while enhancing model performance. This paper introduces random interpolation, diversity term, and affinity term to alleviate the need for ... | Rebuttal 1:
Rebuttal: We appreciate your thorough review and valuable feedback on our paper. Thank you for your positive comments on the comprehensiveness of our experiments and for recognizing the key contribution of our proposed two metrics (i.e., affinity and diversity) in quantifying model quality for model selecti... | Summary: This paper proposes LSS, a model interpolation-based local training technique to reduce the number of communication rounds required. The intuition is to regularize local models to connected low-loss valleys, so the aggregated model may have lower loss. LSS is empirically evaluated on a variety of datasets and ... | Rebuttal 1:
Rebuttal: Thank you for your thoughtful and constructive feedback. We appreciate that you found our figure illustrations clear and our experimental scope comprehensive, covering diverse datasets and types of distribution shifts. We also value your suggestions for improving our manuscript.
We have addressed... | null | null | Rebuttal 1:
Rebuttal: We sincerely thank the AC and reviewers for their valuable time. We are delighted to see that the reviewers highlight our **"paper is well written"**, recognizing the importance of our idea of **"bridging two low-loss valley to reduce communication rounds"**, and our proposed method **"effectively... | NeurIPS_2024_submissions_huggingface | 2,024 | null | null | null | null | null | null | null | null |
Reliable Learning of Halfspaces under Gaussian Marginals | Accept (spotlight) | Summary: This paper studies the problem of agnostic reliable learning with Gaussian margin. It gives a novel algorithm with improved running time and sample complexity bound and this suggests that agnostic reliable learning is easier than agnostic learning. It also gives a Statistical Query lower bound matching some te... | Rebuttal 1:
Rebuttal: We thank the reviewer for the time and effort in reviewing our paper and for the positive feedback.
> I do find the SQ lower bound a bit weak. It doesn't have a dependence on $\epsilon$, is there any known lower bounds on $\epsilon$?
*Response:* The point of our SQ lower bound is that the joint... | Summary: This paper considers learning halfspaces with Gaussian marginals in a reliable learning setting, where the learner has to guarantee that the error of the output classifier is less than $\epsilon$ (and we assume such a classifier exists). It is known that the reliable learning problem can be efficiently reduced... | Rebuttal 1:
Rebuttal: We thank the reviewer for the time and effort in reviewing our paper and for the positive feedback. We respond to the point raised by the reviewer regarding the significance of our result.
Our result is the first to establish a computational separation between reliable and agnostic learning in t... | Summary: This work studies agnostic learning of halfspaces in the reliable learning model, which guarantees a halfspace with nearly no false positives and a nearly optimal false negative rate, where the optimal false negative rate is defined relative to a class of halfspaces with no false positives. The authors prove s... | Rebuttal 1:
Rebuttal: We thank the reviewer for the time and effort in reviewing our paper and for the positive feedback. Below, we provide specific responses to the points and questions raised by the reviewer.
> In line 222, I’m confused about the signs. Doesn’t the interval $[t^*,\infty]$ correspond to positive labe... | Summary: This paper studies reliable learning of halfspaces in $d$ dimensions. Reliable learning is a framework in learning theory in which the learning algorithm is required to output a classifier $f$ satisfying:
- the probability that f makes a false-positive error is at most $\epsilon$
- The probability that f makes... | Rebuttal 1:
Rebuttal: We thank the reviewer for the time and effort in reviewing our paper and for the positive feedback. Below, we provide specific responses to the points and questions raised by the reviewer.
> Is it possible that for every small constant $c$, if $\alpha$ is promised to be at least a constant, then ... | Rebuttal 1:
Rebuttal: We thank all reviewers for taking the time to read our manuscript carefully and for providing constructive and insightful feedback.
We provide detailed responses to each reviewer separately. We look forward to engaging in further discussion with the reviewers, answering questions, and discussing... | NeurIPS_2024_submissions_huggingface | 2,024 | null | null | null | null | null | null | null | null |
KVQuant: Towards 10 Million Context Length LLM Inference with KV Cache Quantization | Accept (poster) | Summary: KVQuant presents a method for applying low-bit activation quantization to the Key-Value Cache, a major bottleneck in long context LLM's generation inference. The authors propose strategies (per-channel, per-token, pre-RoPE) tailored to the distribution characteristics of keys and values, as well as the distrib... | Rebuttal 1:
Rebuttal: Thank you for taking the time to review our paper!
> As a minor weakness, due to the extensive experimental results and content, there are many appendices linked in the main text and frequent parenthetical explanations, which can make the paper somewhat challenging to follow. There is room to im... | Summary: This paper proposes KVQuant, which consists of 4 techniques for improving the performance of low-bit KV cache quantization. By observing the distribution of Key and Value cache, it proposes to use channel-wise quantization for Key cache before RoPE and token-wise quantization for Value cache. It also adopts no... | Rebuttal 1:
Rebuttal: Thank you for the detailed feedback on our work!
>The paper seems to be a combination of several existing quantization methods), e.g. choosing to use per-channel quantization for Key cache, using non-uniform quantization, and dense-and-sparse quantization to improve performance.
With respect to ... | Summary: This paper proposes KVQuant, a quantization framework for enabling long context window inference through compressing KV cache activations. Specifically, the KVQuant framework incorporates several techniques including per-channel key quantification, per-RoPE key quantification, non-uniform KV cache quantificati... | Rebuttal 1:
Rebuttal: Thank you for taking the time to review our paper.
> In the per-vector dense-and-sparse quantization step, the numerical outliers are stored as high precision in a separate sparse matrix. It is not quite clear how to achieve this in practice in a hardware-friendly way without affecting inferenc... | Summary: The paper presents KVQuant, a low precision quantization method for KV cache activations in LLMs to reduce memory consumption during inference with large context windows. KVQuant applies per-channel pre-RoPE quantization on Key cache, exploits non-uniform datatype for quantization, and keeps the per-vector out... | Rebuttal 1:
Rebuttal: Thank you for taking the time to review our paper.
> The paper does not sufficiently clarify the baseline methods and implementations used for comparison.
The baseline methods that we used to compare with KVQuant are (i) uniform quantization with and without grouping, (ii) non-uniform quantiza... | Rebuttal 1:
Rebuttal: Thank you to all of the reviewers for taking the time to review our paper and to provide us with valuable feedback. We included responses for each of the questions that the reviewers have posed.
Pdf: /pdf/95a510379ed52c79cff66d022ba5659aad988ba7.pdf | NeurIPS_2024_submissions_huggingface | 2,024 | null | null | null | null | null | null | null | null |
Latent Plan Transformer for Trajectory Abstraction: Planning as Latent Space Inference | Accept (poster) | Summary: Authors propose a novel method Latent Plan Transformer (LPT). The key idea is to modify Decision Transformer (DT) approach by adding latent variable conditioning instead of return-to-go. This latent variable is assumed to represent a "plan" that the agent will follow. The motivation of replacing return-to-go i... | Rebuttal 1:
Rebuttal: Thank you for your valuable feedback! We greatly appreciate the opportunity to clarify and expand our work.
> W1: Lack of intuition: plugging in a sampled fixed random variable conditioned on the final return performs better than DT with step-wise RTG
Thank you for this insightful observation. In... | Summary: The paper introduces the Latent Plan Transformer (LPT), a novel framework for trajectory generative modeling in the absence of step-wise reward. This framework employs a top-down latent variable model, using a temporally-extended latent variable z to represent a plan for decision-making. The framework comprise... | Rebuttal 1:
Rebuttal: Thank you for your thorough and constructive review of our paper! We shall address your questions point by point.
> W1: Potential training and inference inefficiency because of MCMC sampling in more complex scenarios
Thank you for your insightful question! We would like to clarify that with caref... | Summary: Building on the idea of decision transformer, the paper introduces a new generative model based decision-making agent called Latent Plan Transformer (LPT). Instead of directly generating the trajectories and returns as in the prior work, LPT would first generate a latent vector, and then generate the trajector... | Rebuttal 1:
Rebuttal: Thank you for recognizing the importance of our approach to modeling plans in decision-making processes! We greatly appreciate your insightful feedback and the opportunity to clarify and expand upon our work.
> W1: What are the connections between LPT and some theoretical analysis in MDP represen... | null | null | Rebuttal 1:
Rebuttal: ## Global Response
We would like to thank all reviewers for their helpful feedback! Here, we would like to add some clarifications based on points raised by multiple reviewers.
**[Latent variable $z$ formulation]**
We first higlight that the latent variable of our model is $z$, whose prior dist... | NeurIPS_2024_submissions_huggingface | 2,024 | null | null | null | null | null | null | null | null |
CoFie: Learning Compact Neural Surface Representations with Coordinate Fields | Accept (poster) | Summary: This paper proposes a new architecture for SDF auto-decoding task. It divides the whole SDF points into voxels and builds local coordinates for each surface patch included in valid voxels. Respective shape coding is learned separately and a generalizable MLP is used for SDF decoding. For MLP, a quadratic layer... | Rebuttal 1:
Rebuttal: Thank you for the detailed feedback, your acknowledgment of our theoretical proofs of fitting quadratic local patches, and the potential of our method. We will address your concerns as follows.
---
**[W1, Baseline]** Thank you for pointing us to NKSR. However, **NKSR is not a directly comparable... | Summary: ### Motivation
- In the realm of implicit 3D shape representations, local-based solutions [4, 19, 32] (which decomposes the target shape into a set of local surfaces to model) provide higher accuracy but at the cost of a higher number of parameters to optimize.
- The authors argues that this number of paramet... | Rebuttal 1:
Rebuttal: Thank you for your detailed and positive comments. Thank you for acknowledging the novelty, the theoretical foundation, and the good performance of our work. We will address your comments as follows.
---
**[W1, Baselines]** Although we understand your concern, we cannot find more directly compar... | Summary: This paper introduces CoFie, a novel neural surface representation method designed to efficiently learn and represent complex 3D shapes. CoFie addresses the challenges of existing methods by introducing a coordinate field that optimizes the representation of local shapes, significantly reducing shape errors an... | Rebuttal 1:
Rebuttal: Thanks for your positive comments, and your acknowledgment that our theoretical proof is sufficient and the articulation is clear. We will address your comments as follows.
---
**[W1, grid size]** Our method has demonstrated good performance in modeling detailed geometry as shown in Figure 3, Fi... | Summary: Paper proposes CoFie — 3D shape representation as a set of latents arranged on a regular voxel grid. Each latent encodes an oriented local quadratic patch. This local oriented patch defines local SDF in a local coordinate frame which is decoded via conditional MLP for which the last layer is quadratic: it defi... | Rebuttal 1:
Rebuttal: Thanks for your detailed feedback, and your acknowledgement of the efficiency of our model, the clarity of our writing and figure, and our clear ablation study. We will address your comments and some **factual errors** in your comments as follows.
---
**[W1, AutoSDF and SDFusion]** The mentioned... | Rebuttal 1:
Rebuttal: We would like to thank the reviewers for their detailed feedback. We received positive comments from the reviewers who appreciated our clear ablation and convincing evaluation (LpYZ, Kq6m), the efficiency of our method (V5fn, 4hsy), good performance (V5fn, 4hsy), clear theoretical proof (4hsy, Kq6... | NeurIPS_2024_submissions_huggingface | 2,024 | Summary: The paper proposes a local prior based method where the model is trained on a dataset and learns a prior over local patches of the shapes, and then at test time reconstructs patches by optimising to minimise the reconstruction error w.r.t. the trained model's prior. While this has already been done in the lite... | Rebuttal 1:
Rebuttal: We thank you for the positive comments and your acknowledgment of the thoroughness of our ablation study. We will address your comments as follows.
---
**[W2, Benefit of Coordinate Field]** We kindly note that the coordinate field is a part of our representation, as the coordinate field is invol... | null | null | null | null | null | null |
When to Sense and Control? A Time-adaptive Approach for Continuous-Time RL | Accept (poster) | Summary: The paper focuses on continuous-time reinforcement learning. Typically the control and the measurements of continuous-time systems occur at discrete time points. These interventions generally come with a cost. The paper proposes a time-adaptive approach to reduce the number of these interventions in an optimal... | Rebuttal 1:
Rebuttal: Thank you for your comments and valuable feedback!
# Weaknesses
1. *“The references in the paper lack consistency in formatting. Conference names are sometimes written in full, sometimes abbreviated, and occasionally omitted altogether.”:* Thanks for spotting that, we updated the references in t... | Summary: The paper introduces a framework for reinforcement learning named Time-adaptive Control & Sensing (TACOS). The TACOS framework reformulates the problem of continuous-time RL into an equivalent discrete-time Markov decision process (MDP) that standard RL algorithms can solve. Additionally, a model-based version... | Rebuttal 1:
Rebuttal: First, thank you a lot for your positive and valuable feedback! We will indeed incorporate the proposed ideas in the updated version of the paper.
# Weaknesses
1. *“Although the problem posed by the paper is novel and interesting, the solution simply involves incorporating time into the action s... | Summary: This paper proposes a novel time-adaptive RL method framework (TaCoS) for continuous-time systems with continuous state and action spaces. The framework shows that the settings of interactions having costs or a budget of interactions, can be formulated as extended MDPs, that can be solved with standard RL algo... | Rebuttal 1:
Rebuttal: Thanks a lot for the positive and valuable feedback!
# Weaknesses
1. *Verification of “Intuitively, the more stochastic the environment, the more interactions we would require to stabilize the system.”:* We verify this empirically. In Figure 3 in the paper, bottom row, we analyze the setting wh... | Summary: Reinforcement learning (RL) is effective for discrete-time Markov decision processes (MDPs), but many systems operate continuously in time, making discrete-time MDPs less accurate. In applications like greenhouse control or medical treatments, each interaction is costly due to manual intervention. To address t... | Rebuttal 1:
Rebuttal: Thank you for your comments and valuable feedback!
# Questions
1. *Real-world applications and global frequencies*: As you correctly described, the main difference between standard discrete-time control and our setting is that in the standard discrete-time setting, the interaction (measuring th... | Rebuttal 1:
Rebuttal: We thank all the reviewers for their valuable and useful feedback. We have attached a pdf with 3 additional experiments that study the following setting:
1. *Figure 6:* We further show the effect of environment stochasticity on the frequency of interactions. We empirically show that with larger st... | NeurIPS_2024_submissions_huggingface | 2,024 | null | null | null | null | null | null | null | null |
A General Protocol to Probe Large Vision Models for 3D Physical Understanding | Accept (poster) | Summary: This paper aims to evaluate how well large-scale vision models encode 3D properties of the scenes depicted in images. The paper proposes a general and lightweight protocol that involves training discriminative classifiers on the features of pre-trained models to assess their encoding of several physical proper... | Rebuttal 1:
Rebuttal: Thanks for your valuable comments. We have provided responses below:
**W1: Whether the proposed probes can reflect the 3D understanding: The material, support relation and shadow properties can be well identified with 2D clues (appearance and 2D spatial location). For occlusion and depth, using l... | Summary: This paper investigates several mainstream large vision models for their 3D physical understanding. The authors curated a binary classification benchmark covering a set of important 3D properties based on publicly available 2D image datasets and linearly probed different layers and different time steps of the ... | Rebuttal 1:
Rebuttal: Thanks for your valuable comments. We have provided responses below:
**W1: In-depth analysis and understanding: What’s the speculations and thoughts on the cause of varied performance? Does difference in data and training objectives play a part?**
Both data and the training objective may play a ... | Summary: In order to efficiently examine whether large vision models have explicit feature representations for different properties of the 3D physical scene, the paper proposes to linearly probe the features of different layers (and different time steps) from LVMs on specially designed binary classification problems. E... | Rebuttal 1:
Rebuttal: Thanks for your valuable comments. We have provided responses below:
**W1: Downstream application of pre-trained LVMs**
Concerning downstream applications, in the Appendix Section F, we showed that the features selected by the linear probe could be used for the task of normal prediction.
We hav... | null | null | Rebuttal 1:
Rebuttal: We thank the reviewers for their appreciation of the importance of the problem and our approach, and for their comments.
Reviewer Lz9q:"The first to investigate 3D knowledge in LVMs in a lightweight manner with linear probing; extensive experiments; paper well-written; the problem is crucial to i... | NeurIPS_2024_submissions_huggingface | 2,024 | null | null | null | null | null | null | null | null |
Grounding Multimodal Large Language Models in Actions | Accept (poster) | Summary: This paper studies how to "ground" multimodal LLMs (MLLMs) to the action spaces of agents with various embodiments. The authors examine "Action Space Adapters" (parameterization strategies; ASAs) for various embodiments and MLLMs and identify principles for constructing ASAs based on the target action space. T... | Rebuttal 1:
Rebuttal: We thank the reviewer for appreciating the technical motivation, importance of the problem, and the exhaustive experimentation. The reviewer's clarification questions are addressed below and will be added to the final manuscript, which we believe will greatly improve it.
**1. No specific example ... | Summary: In this paper the authors present a way to adapt a Vision and Language model to perform action execution tasks in embodied environments. Specifically, systematically evaluate different ways of predicting actions both on task having both discrete and continuous action spaces. Thanks to this evaluation, it is po... | Rebuttal 1:
Rebuttal: We thank the reviewer for the comments and suggestions. We address the reviewer’s points below.
**1. Details on VQ-VAE training.**
Thank you for the suggestion; we will add these details to the paper. We train the VQ-VAE learned tokenizers to reconstruct actions from the same dataset used for su... | Summary: The paper empirically studies how to properly ground MLLMs into embodiments, with a particular focus on the action representations, including the continuous and discrete solutions. The authors conduct a thorough study on 7 methods, encompassing over 114 tasks. The research indicates that for continuous actions... | Rebuttal 1:
Rebuttal: We thank the reviewer for the comments and suggestions. We address the reviewer’s points below.
**1. Conclusions may not be applicable to other MLLM architectures or scales.**
In the main rebuttal PDF, we show results with a different MLLM (PaliGemma [1]) and a different finetuning method (full ... | null | null | Rebuttal 1:
Rebuttal: We thank the reviewers for their useful and insightful feedback. Reviewers highlighted that our study is one of the first on how to properly ground MLLMs in embodied action (f4rS, YrxP) and provides concrete and actionable takeaways that will be useful for future research (YrxP, eGot). We highligh... | NeurIPS_2024_submissions_huggingface | 2,024 | null | null | null | null | null | null | null | null |
Sketching for Distributed Deep Learning: A Sharper Analysis | Accept (poster) | Summary: This paper considers the sketch-DL framework for distributed / federated learning studied by prior works such as Rothchild et al., Song et al. The authors identify that for these works, their convergence either has a dependence on the dimension (under rather minimal assumptions) which is unfavorable, or assume... | Rebuttal 1:
Rebuttal: We thank the reviewer for the suggestion. We will include a table comparing our results with [Song et al. 2023] in the final paper. Here, we provide brief summary:
**Q1. Comparison with [Song et al. 2023]:**
Before comparing, we briefly summarise the difference in our setup, assumptions and resul... | Summary: This paper provided ambient dimension-independent convergence and communication complexity analysis for sketching-based gradient descent algorithms using the second-order geometry of loss Hession.
Strengths: * This paper proves that the dimensional dependence comes from the global smoothness assumption and pr... | Rebuttal 1:
Rebuttal: We thank the reviewer for carefully reading our paper and providing constructive feedback. Below we provide answers to each of the concerns:
**Q1: SGD v/s GD**:
Firstly we point out our theoretical result in Theorem 1 must be compared to Theorem E.1 in [Song et al. 2023] which also studies GD. F... | Summary: This paper presents a novel analysis of sketching-based distributed learning algorithms for deep neural networks. The authors provide a tighter, dimension-independent convergence analysis by leveraging the restricted strong smoothness (RSS) property of deep learning loss functions. This work addresses a gap be... | Rebuttal 1:
Rebuttal: We are thankful to the reviewer for the suggestions and comments. Below we provide answers to each of the concerns:
**Q1. Robustness to violation of assumption on eigenvalues:**
Under the assumption on eigenvalues, i.e. $\kappa = o(1)$, we provide a dimension-independent convergence rate. Howeve... | null | null | Rebuttal 1:
Rebuttal: **General Response**:
We thank the reviewers for their positive comments about our work and their efforts in providing meaningful feedback. In particular, we thank Reviewer `LRSh` for stating that our results are **quite important** ,**resolve big issues from previous work** under **reasonable as... | NeurIPS_2024_submissions_huggingface | 2,024 | null | null | null | null | null | null | null | null |
Test-time Adaptation in Non-stationary Environments via Adaptive Representation Alignment | Accept (poster) | Summary: This paper proposes a test-time adaptation (TTA) method focusing on adapting to a non-stationary environment.
The proposed method DART aims at minimizing the cumulative loss over time steps, which is equivalent to minimize the distribution gap between the source and current features.
To capture unknown changin... | Rebuttal 1:
Rebuttal: Thank you for your detailed review and thoughtful feedback. We will revise the manuscript accordingly. Specific questions are answered below.
---
**Question in W2:** Providing more detailed explanations of theoretical analyses in Sec. 3.4 would be beneficial. For example, where do $\mathcal{O}(T... | Summary: This paper proposes a novel algorithm called Ada-ReAlign for sequentially adapting a model to non-stationary environments. The proposed algorithm uses a set of base learners, each equipped with different learning window sizes, with a meta learner that combines their outputs. This online ensemble allows an adap... | Rebuttal 1:
Rebuttal: Thank you for your detailed review and thoughtful feedback. We will revise the manuscript accordingly. Specific questions are answered below.
---
**Q1:** Should the number of base learners used depend on the downstream problem? How is the number of base models determined and the size of the slid... | Summary: The paper investigates test-time adaptation in a non-stationary environment where unlabeled data batches arrive sequentially with changing data distributions. The authors propose non-stationary representation learning to project the changing distribution back to the original source data distribution and update... | Rebuttal 1:
Rebuttal: Thank you for your detailed review and thoughtful feedback. We will revise the manuscript accordingly. Specific questions are answered below.
---
**Q1:** How does Theorem 1 apply to deep models, considering that the theory was originally formulated for linear models?
**A1:** Theorem 1 shows tha... | null | null | Rebuttal 1:
Rebuttal: We are very grateful to all reviewers for your valuable feedback, which has definitely helped in improving our paper. We respond to the questions raised respectively and report all the tables and figures in the supplementary PDF.
Pdf: /pdf/bf940babfb9349b0f776dedbcd488d1dd6f40d81.pdf | NeurIPS_2024_submissions_huggingface | 2,024 | null | null | null | null | null | null | null | null |
Quantum algorithm for large-scale market equilibrium computation | Accept (poster) | Summary: This paper proposes quantum faulty proportional response (FPR) dynamics, a quantum algorithm that mimics classical proportional response (PR) dynamic with quantum speed-up, to compute the "part of" market equilibrium. Specifically, a market equilibrium specifies the allocation of each good to each buyer, and t... | Rebuttal 1:
Rebuttal: We thank the reviewer for the comments. We address the reviewer's concerns and questions as follows.
### Allocation computation for all buyers and goods.
> The goal of FPR is to compute the allocation of some pre-specified good to some pre-specified buyer, which seems fairly limited, since classi... | Summary: The paper studies the problem of computing an equilibrium for the Fisher market. Roughly speaking, this is a fractional matching problem with a particular objective function, namely a sum of logarithms of utilities. In the classical world, this problem is usually approximated with a local search algorithm, but... | Rebuttal 1:
Rebuttal: We thank the reviewer for their comments and their appreciation of the discussion section. We address the reviewer's concerns and questions as follows.
### Scope and venue.
In regards to the scope of the paper and compatibility with an artificial intelligence/machine learning conference, we beli... | Summary: This paper proposes a quantum-assisted algorithm for market equilibrium computation by first introducing the faulty proportional response dynamics and then constructing its quantum implementation. This algorithm has provable quadratic speedups. Simulation studies show its effectiveness.
Strengths: 1. The pape... | Rebuttal 1:
Rebuttal: We thank the reviewer for their comments and concerns. We address the concerns and questions as follows:
### Regarding robustness and additional error terms.
Our analysis of FPR dynamics is not constrained to error terms stemming from quantum amplitude estimation only but applies to errors from o... | Summary: This paper focuses on market equilibrium computation with linear utility. Traditional algorithms like proportional response (PR) dynamics, despite having nearly linear runtimes relative to the number of buyers and goods, struggle with the massive scale of modern internet-based markets.
The authors introduce a... | Rebuttal 1:
Rebuttal: We thank the reviewer for the comments.
### Access to accurate valuation.
> I am wondering about the practicality of the algorithm, since we need an accurate valuation profile of all buyers and goods, and the current limitations of quantum hardware pose significant challenges.
Our assumption of... | Rebuttal 1:
Rebuttal: We sincerely thank all the reviewers for their detailed comments and suggestions, as well as for appreciating our contributions.
- Reviewer B4u4 remarked that our results are "valuable as a theoretical study" and our results provide "significant improvements in efficiency".
- Reviewer vuEH ackn... | NeurIPS_2024_submissions_huggingface | 2,024 | null | null | null | null | null | null | null | null |
Promoting Fairness Among Dynamic Agents in Online-Matching Markets under Known Stationary Arrival Distributions | Accept (poster) | Summary: This paper studies an online bipartite matching problem where each offline type has a capacity and the nodes of each online type arrive according to a Poisson process. Each offline type can serve a certain subset of online types, and each online node needs to be served or discarded immediately upon its arrival... | Rebuttal 1:
Rebuttal: Q1. When we state that all hardness results provided in the paper are independent of any benchmarks, we mean that all competitiveness results are computed directly against the performance of a clairvoyant optimal policy (OPT), rather than any upper bound on OPT (e.g., the optimal value of a bench... | Summary: This paper considers the online matching problem under known stationary arrival distributions. Each offline agent of type i has a matching capacity b_i, and each online agent of type j arrives according to an independent Poisson process of rate \lambda_j. The objective is to maximize the minimum matching rate ... | Rebuttal 1:
Rebuttal: Thanks for your comments. As for the technical aspects, we want to highlight our parameter-dependent competitive analysis for sampling-based policies. Specifically, we incorporate two parameters—the optimal LP value and the minimum offline serving capacity—into the analysis and the final competiti... | Summary: This paper investigates a variant of the online matching problem, focusing on maintaining long-term fairness at both individual and group levels. A key finding is that the optimal competitive ratio achievable without rejecting any items is capped at 1/2. By implementing a novel sampling algorithm, SAMP, the au... | Rebuttal 1:
Rebuttal: Q1 ``For the First Come, First Serve (FCFS) for Online Matching with Long-run Fairness (OM-LF) we get a maximal matching
thus it achieves a 1/2 approximation ratio...''
We politely disagree, and explain why in the following two points.
First, the objective of (individual) long-run fairness is d... | Summary: This work focuses on a fair online bipartite matching problem under poisson arrivals, where the objective is to maximize the minimum number of matches across groups of the online nodes (a max min objective). It is showed that if rejecting online nodes is not allowed, then any algorithm will have at best a $1/2... | Rebuttal 1:
Rebuttal: Q1 ``It is mentionned in the related work that Huang et al showed that similar results can be obtained in the KIID and
Poisson arrival models...''
Yes. As stated in the paper by Huang et al., for any algorithm, its competitiveness can be translated between the two models up to a multiplicative fa... | null | NeurIPS_2024_submissions_huggingface | 2,024 | null | null | null | null | null | null | null | null |
Rethinking Human Evaluation Protocol for Text-to-Video Models: Enhancing Reliability, Reproducibility, and Practicality | Accept (poster) | Summary: This paper introduces the Text-to-Video Human Evaluation (T2VHE) protocol, a standardized approach for assessing the quality of videos generated from text.
Besides, this paper comprehensively evaluates some famous text-to-video models with their proposed reliable human assessment, which is very exciting and v... | Rebuttal 1:
Rebuttal: We sincerely thank reviewer LQLU for the valuable comments and give detailed responses to all questions. Due to space constraints, we submit part of the replies in "Official Comment", for convenience, we have summarized conclusions for each reply. Looking forward to further discussion with you.
*... | Summary: This manuscript presents a novel, standardized human evaluation protocol specifically tailored for Text-to-Video (T2V) models. The protocol encompasses a meticulously designed suite of evaluation metrics supplemented with robust annotator training resources. The effective deployment of Latent Action Ratings (L... | Rebuttal 1:
Rebuttal: We sincerely thank reviewer RMCR for the valuable comments and give detailed responses to all questions. Due to space constraints, we submit part of the replies in "Official Comment", for convenience, we have summarized conclusions for each reply. Looking forward to further discussion with you.
*... | Summary: This paper introduces the Text-to-Video Human Evaluation (T2VHE) protocol, a standardized approach for evaluating text-to-video (T2V) models, addressing the challenges posed by recent advancements in T2V technology. The T2VHE protocol supposedly offers a solution through well-defined metrics, annotator trainin... | Rebuttal 1:
Rebuttal: We sincerely thank reviewer AEKZ for the valuable comments and give detailed responses to all questions. Due to space constraints, we submit part of the replies in "Official Comment", for convenience, we have summarized conclusions for each reply. Looking forward to further discussion with you.
*... | Summary: This paper designs an evaluation protocol and surveys a number of video generation papers published over the last few years. The authors point out shortcomings of evaluation protocols in these priors works (e.g. they often limit studies to video quality comparisons with no clear definitions of video quality, ... | Rebuttal 1:
Rebuttal: We sincerely thank reviewer 2pJH for the valuable comments and give detailed responses to all questions. Due to space constraints, we submit part of the replies in "Official Comment", for convenience, we have summarized conclusions for each reply. Looking forward to further discussion with you.
*... | null | NeurIPS_2024_submissions_huggingface | 2,024 | null | null | null | null | null | null | null | null |
A Simple Framework for Generalization in Visual RL under Dynamic Scene Perturbations | Accept (poster) | Summary: This paper introduces SimGRL for vision-based DRL tasks. It tackles two challenges: the imbalanced saliency, where an agent repeatedly favors saliency maps from the recent one in stacked frames, and the issue of overfitting to particular background areas rather than concentrating on elements crucial to the tas... | Rebuttal 1:
Rebuttal: We thank the reviewer for your thoughtful comments. We address individual comments in the following.
**[Weakness]**
>**[W1] Need for an explicit articulation of its motivation and contributions**
**(A)** We will add a paragraph to clarify the motivation and contribution of our paper.
***
>**... | Summary: This paper proposes two simple, yet crucial modifications to the Generalization in the Visual RL pipeline where during the test time, the visual observation consists of various degrees of dynamically varying backgrounds.
Concretely, the authors identify two issues that inhibit RL agents from generalization: (... | Rebuttal 1:
Rebuttal: We thank the reviewer for your thoughtful comments. We address individual comments in the following. Additionally, we sincerely appreciate your eagerness to engage in further discussions regarding our work!
**[Weakness]**
>**[W1] Clarity w.r.t the attribution masking argument**
**(A)** *To avoi... | Summary: The paper first identify two pitfalls with the visual RL via gradient-based attribution mask, i.e., imbalanced saliency and observational overfitting. To address these two pitfalls, the paper proposes two novel modifications. One is for the encoder where encode each frame and stack encoded representations inst... | Rebuttal 1:
Rebuttal: We thank the reviewer for your thoughtful comments. We address individual comments in the following.
**[Weakness]**
>**[W1] Comparison with robust RL algorithms**
**(A)** We clarify that our work aims to address the generalization issue in a purely **model-free vision-based RL** setting, which r... | Summary: The paper introduces SimGRL, a framework aimed at enhancing generalization in vision-based deep reinforcement learning (RL) under dynamic scene perturbations. It addresses two critical issues in existing visual RL methods: imbalanced saliency and observational overfitting. SimGRL employs architectural modifica... | Rebuttal 1:
Rebuttal: We thank the reviewer for your thoughtful comments. We address individual comments in the following.
**[Weakness]**
***
>**[W1] Dependency on augmentation quality**
**(A)** The effectiveness of the shifted random overlay augmentation may depend on the quality and diversity of the natural imag... | Rebuttal 1:
Rebuttal: Dear reviewers, we appreciate all your valuable comments and the time you've dedicated to reviewing our paper.
Here, we briefly summarize the answers to some commonly asked questions and upload a PDF file with some additional figures to aid in understanding.
+ The attribution masks were used sol... | NeurIPS_2024_submissions_huggingface | 2,024 | null | null | null | null | null | null | null | null |
RefDrop: Controllable Consistency in Image or Video Generation via Reference Feature Guidance | Accept (poster) | Summary: This paper targets the task of subject consistency in image and video generation. Previous consistency generation methods [53,68] are based on concatenated attention. The authors reformulate concatenated attention in a manner similar to classifier-free guidance, simplifying the constant C matrix to a constant ... | Rebuttal 1:
Rebuttal: Thank you for your thorough comments! They are quite insightful!
> The generated images appear to have limited diversity. In Figure 13, many examples of humans have very similar poses and styles.
>
Please see the global response for this matter. In short, we propose to exclude the first upsampl... | Summary: The authors analyzed the current methods of consistent content generation and revealed that the concatenation of reference features in the self-attention block can be reformulated as a linear interpolation of image self-attention and cross-attention between synthesized content and reference features with a con... | Rebuttal 1:
Rebuttal: Thank you for your review and detailed questions!
> If a subsection is added to show the flexibility of the method across different models and potential limitations/considerations when applying it, that would be great.
>
We try our method on the most recent capable open-source model, FLUX-dev, ... | Summary: This paper focuses on the challenge of ensuring consistency in the generation of images and videos. Deep learning and artificial intelligence techniques are utilized in image and video generation for generating new images or generate video frames based on given inputs, such as text prompts or reference images.... | Rebuttal 1:
Rebuttal: Thank you for your review!
> The lack of feature-based comparisons using metrics like Kullback-Leibler (KL) divergence in the experiments.
>
KL divergence measures the divergence between two distributions, which would require at least hundreds of images in our case. However, we aim to measure t... | Summary: This paper presents RefDrop, a method that allows users to control the influence of reference context in a direct and precise manner. More specifically, the proposed method is training-free, which means it can be used plug-and-play without the need to train a separate image encoder for feature injection from r... | Rebuttal 1:
Rebuttal: Thank you for your review!
> It does not dis-entangle spatial control with appearance control.
>
Thank you for raising this issue. Please see the global response and *Figures* **1** and **2** in the rebuttal PDF for a detailed explanation. In short, we propose excluding the first upsampling blo... | Rebuttal 1:
Rebuttal: We thank the reviewers for their valuable comments. We are pleased that the reviewers find our method intuitive (Reviewer 5p1r) and plug-and-play (Reviewer NJob, c2sE), and that they consider our paper well-written (Reviewers 5p1r, c2sE). We appreciate the acknowledgment from all reviewers that ou... | NeurIPS_2024_submissions_huggingface | 2,024 | null | null | null | null | null | null | null | null |
Fast Iterative Hard Thresholding Methods with Pruning Gradient Computations | Accept (poster) | Summary: Iterative hard thresholding (IHT) is used to select the k most important features in an ordinary least squares (OLS) linear regression model, that is, the model parameter vector is constrained to have only k non-zero entries. It seems that most practical IHT methods to solve the constrained problem are based o... | Rebuttal 1:
Rebuttal: We sincerely appreciate the reviewer’s positive feedback on our paper. As mentioned in the reviewer's comment, previous methods for accelerating IHT have focused on reducing the number of iterations. In contrast, our approach accelerates IHT by reducing the computation cost per iteration. Accelera... | Summary: This paper studies iterative hard thresholding (IHT) as a canonical method for sparse linear regression. With precomputed X^TX and X^Ty, the computational cost of the algorithm is dominated by the gradient updates. To reduce the computational cost, this work proposed a pruning procedure at each step of IHT to ... | Rebuttal 1:
Rebuttal: We sincerely appreciate the reviewer's thoughtful review of our paper. Below are our answers to the questions and weaknesses raised by the reviewer.
> I find the terminology "pruning" a bit confusing. My understanding is that it refers to setting a particular entry $z_{j}^{t}$ to zero depending o... | Summary: The paper proposes to pruning the computation of marginal gradients in the IHT algorithm to accelerate the updating steps. For that, the upper bound $\overline{z}_j^{t}$ of the component $z_j^t$ in the gradient step is proposed in Definition 1 and unnecessary elements in $\mathbf{z}^t$ that must be thresholde... | Rebuttal 1:
Rebuttal: We appreciate the reviewer’s constructive comments on our paper. To begin, we address the questions raised by the reviewer below.
> 1. Line 42-45 is duplicated with the abstract.
We thank the reviewer for this comment. We will revise those sentences accordingly.
> 2. Definition 2 is less inform... | Summary: The authors accelerate the iterative hard thresholding (IHT) method, whose purpose is to find the k most important elements from a linear regression model. Specifically, they safely prune unnecessary elements with upper bounds on the element values. The experiment shows significant speedup for the proposed met... | Rebuttal 1:
Rebuttal: We extend our sincere gratitude to the reviewer for working on our paper. We address the weaknesses and questions raised by the reviewer below.
> The importance of the work seems to be not clearly conveyed.
A1. As mentioned in lines 13—15, IHT is widely used in various fields, such as feature se... | Rebuttal 1:
Rebuttal: We would like to thank the reviewers for working on our paper. We have received many constructive comments, which we intend to incorporate to improve the quality of our work. We believe we have addressed most of the comments and would be grateful if the reviewers could respond to our response.
In... | NeurIPS_2024_submissions_huggingface | 2,024 | null | null | null | null | null | null | null | null |
FSEO: A Few-Shot Evolutionary Optimization Framework for Expensive Multi-Objective Optimization and Constrained Optimization | Reject | Summary: This paper introduce a meta-learning framework into few-shot optimization to assist the surrogate modelling in expensive evaluation setting. The authors parameterize a mapping function to get the hidden feature of the solution space and then integrate such mapping into a gaussian kernel function as a deep kern... | Rebuttal 1:
Rebuttal: Weakness 1: Given that the likelihhod-based loss function (Eq. 4) should be maximized to fit the samples from all of the related tasks, why its update should follow a gradient descent rather a gradient ascent? Correct me if I was wrong.
Response: Thanks for your comment.
In practice, the maximi... | Summary: This paper proposes Meta Deep Kernel Learning (MDKL), a new surrogate for SAEAs. MDKL consists of a deep kernel with meta-learning. Empirical studies demonstrate its effectiveness in expensive multi-objective optimization and constrained optimization.
Strengths: 1. This paper is well-written and easy to follo... | Rebuttal 1:
Rebuttal: Weakness:
Thanks for your concerns on our contributions. We would like to clarify why our contributions are significant.
1. Contributions on new model:
1.1. Although previous meta-learned deep kernels are able to adapt themselves with the data from the new task, the parameters of these model... | Summary: The authors developed a few-shot evolutionary optimization framework to effectively solve the multi-objective EOPs and constrained EOPs.
Strengths: The proposed method can solve the multi-objective EOPs and constrained EOPs with little data, especially for the engineering problems.
Weaknesses: The learning r... | Rebuttal 1:
Rebuttal: Weakness: The learning results may rely on the relation degree of different tasks.
Response: Thanks for your comment.
We conducted an experiment in Appendix F to show that very similar related tasks were not necessary for the meta-learning of our algorithm.
In fact, task similarity is a convent... | Summary: This paper proposes a new surrogate-assistant evolutionary algorithm that utilizes a Gaussian process with Deep Kernel Learning as the surrogate model. The method employs few-shot meta-learning to learn from multiple tasks to construct the surrogate. It is then integrated with the existing MOEA/D-EGO algorithm... | Rebuttal 1:
Rebuttal: Weakness1: How parameters from different source tasks collectively form the experience, how parameters from both source and target tasks jointly create this experience, are not clearly addressed.
Response: Thanks for your comment.
1). We would like to clarify that all related tasks provide only... | null | NeurIPS_2024_submissions_huggingface | 2,024 | null | null | null | null | null | null | null | null |
Learning Versatile Skills with Curriculum Masking | Accept (poster) | Summary: This paper presents CurrMask, a novel masked prediction approach for unsupervised RL pretraining, which learns skills of different complexity through block-wise masking and adaptively adjusts masking schemes in a curriculum for training efficiency. In contrast to previous methods that perform random masking at... | Rebuttal 1:
Rebuttal: Dear Reviewer Wq3n,
We sincerely appreciate your time to review our paper and your valuable feedback. We will address your concerns in detail below.
>W1 Non-stationary reward distribution may increase instability of the training process
Regarding the concern about non-stationary reward distribu... | Summary: The paper presents a method that learns skills through curriculum masking. Specifically, the approach CurrMask can automatically arranges the order of different masking schemes for training. The algorithm is tested on Deepmind Control Suite tasks, and show positive results in representation learning, zero-shot... | Rebuttal 1:
Rebuttal: Dear Reviewer CzXA,
We sincerely appreciate your time to review our paper and your valuable feedback. We will address your concerns in detail below.
>Comparison to existing skill learning methods with offline data
Firstly, we would like to clarify that our paper focuses on masked prediction. Th... | Summary: This work proposes a curriculum learning approach to reinforcement learning skills from masked trajectory sequences. Given a set of pre-collected environment samples (here from a TD3 agent in 9 different mujoco domains), the proposed MaskCurr curriculum treats the agent's progress (target-loss decrease) like a... | Rebuttal 1:
Rebuttal: Dear Reviewer znVZ,
Thank you for your insightful review of our work! We appreciate your valuable suggestions and will respond to your questions in detail.
>W1 Explanations on "versatile" or "diverse" skills
In our study, we use the terms "versatile skills" or "diverse skills" to refer to the m... | Summary: This paper proposes a curriculum masking pretraining paradigm for RL training, which is based on the block-wise masking schemes and is able to decide the block size and mask ratio automatically. Specifically, the authors design a masking pool with different masking scheme of different block size and mask ratio... | Rebuttal 1:
Rebuttal: Dear Reviewer f2Vb,
Thank you for your valuable feedback! We appreciate your positive remarks about the soundness, presentation, and contribution of our work. Below, we address the specific points you raised:
> Comparing token-wise and block-wise masking
We have conducted experiments on token-wi... | Rebuttal 1:
Rebuttal: We give our sincere thanks to all the reviewers for their insightful suggestions and positive feedback of our work. For all the figures that we add for further discussion, we submit a PDF file in this global rebuttal.
Pdf: /pdf/a285a3d8fc58b4034aaf6491cfcd10f16ea5b1d8.pdf | NeurIPS_2024_submissions_huggingface | 2,024 | null | null | null | null | null | null | null | null |
Director3D: Real-world Camera Trajectory and 3D Scene Generation from Text | Accept (poster) | Summary: This paper presents Director3D, a text-to-3D generation framework that creates realistic 3D scenes with adaptive camera trajectories. It includes a Cinematographer (Traj-DiT) for generating camera trajectories, a Decorator (GM-LDM) for initial scene generation, and a Detailer (SDS++ loss) for refinement. Using... | Rebuttal 1:
Rebuttal: We thank the reviewer for your appreciation of the **high spatial consistency**, the **interesting task**, the **good idea**, and the **clear writing** of Director3D. We address your questions point-by-point below.
---
***Q1: The fairness of the quantitative comparison and more scene-level basel... | Summary: The paper proposes a scene-generation method from text input. The framework utilizes three models that first generate a trajectory, then produce 3D Gaussians, and finally refine through SDS loss.
Strengths: - The design of using a trajectory generator and 3DGS diffusion is novel and impressive.
- The propose... | Rebuttal 1:
Rebuttal: We thank the reviewer for your appreciation of the **novel and impressive design** and the **brilliant visual results** of Director3D. We address your questions point-by-point below.
---
***Q1: More scene-level baselines.***
Thanks for your helpful advice. For a more comprehensive evaluation, w... | Summary: This paper presents Director3D, a novel text-to-3D generation framework designed to generate both real-world 3D scenes and adaptive camera trajectories. Specifically, the authors propose the Traj-DiT to generate adaptive camera trajectories, which treats camera parameters as temporal tokens and performs condit... | Rebuttal 1:
Rebuttal: We thank the reviewer for your appreciation of the **clearly written paper**, the **novel and effective idea**, the **realistic and consistent scene synthesis**, and the **impressive results** of Director3D. We address your questions point-by-point below.
---
***Q1: More scene-level baselines.**... | Summary: This paper proposes a framework for simultaneous text-to-3D scene and camera trajectory generation. The authors propose a 3-stage pipeline to (1) generate a dense camera trajectory from input text, (2) use multi-view latent diffusion from a sparse subset of the generated trajectory to generate the 3D scene rep... | Rebuttal 1:
Rebuttal: We thank the reviewer for your appreciation of the **practical problem**, the **potential applications**, the **clear presentation**, the **easy-to-follow method**, and the **well-written paper** of Director3D. We address your questions point-by-point below.
---
***Q1: The meaning of “real-world... | Rebuttal 1:
Rebuttal: ---
## ***Global Rebuttal***
We express our gratitude to all reviewers for their recognition of the **potential applications** (Reviewer YUdV), the **novel and interesting idea** (Reviewers Lv1o & sdLT & yFX5), the **brilliant results** (Reviewers Lv1o & sdLT & yFX5), and the **well-written pape... | NeurIPS_2024_submissions_huggingface | 2,024 | null | null | null | null | null | null | null | null |
Improved Analysis for Bandit Learning in Matching Markets | Accept (poster) | Summary: This paper considers two-sided matching bandit problems, where the goal is to minimize regret against a player's optimal stable matching. The state-of-the-art algorithms achieve a regret bound of $KlogT/\Delta^2$. The authors suggest an algorithm using an adaptive online Gale-Shapley to achieve a regret boun... | Rebuttal 1:
Rebuttal: We thank the reviewer Pjvv for the valuable comments. Please find our response below.
- $N^2$ regret for the general market
In the matching market problem, each player needs to navigate both his individual explorations and game interactions with others. Considerable work has been done to invest... | Summary: The paper studies the bandit learning problem in two-sided matching markets and provide improved regret analysis that nearly match the lower bound (although with slightly different definition of instance-dependent gaps). The bound is particularly useful when the number of players is much smaller than the numbe... | Rebuttal 1:
Rebuttal: We thank the reviewer Y8Bw for the valuable comments. Please find our response below.
- Gap definition
We agree that investigating the optimal dependence on the preference gap is an interesting future direction. Among all of the existing works, there are three types of gap definitions due to d... | Summary: The paper proposed a new algorithm for bandit learning in matching markets and showed new results on regret bounds. However, the combination of bandits with matching markets is wired. Is there any evidence that practitioners would like to use bandits for learning in markets? The novelty of the theory is also u... | Rebuttal 1:
Rebuttal: We thank the reviewer SWbh for the valuable comments. Please find our response below.
- Technical novelty
Dealing with individual players' exploration-exploitation trade-offs has been extensively studied in the literature. However, in matching markets, researchers have not swiftly comprehended ... | Summary: This work studies the problem of bandit learning in two-sided matching markets, where the number of players $N$ is smaller than the number of arms $K$. The players' preferences $\mu\_{i,j}$ are unknown but the arms' utility preferences $\pi\_{i,j}$ are known. Two algorithms with improved regret bounds are prop... | Rebuttal 1:
Rebuttal: We thank the reviewer unKj for the valuable comments. Please find our response below.
- New gap definition in Theorem 4.1, lower bound for the gap
Recall that the regret dependence on the gap is $1/\Delta^2$, which means the larger the gap, the better the regret. Though we define a new gap in T... | null | NeurIPS_2024_submissions_huggingface | 2,024 | null | null | null | null | null | null | null | null |
SyncTweedies: A General Generative Framework Based on Synchronized Diffusions | Accept (poster) | Summary: SyncTweedie attempts to elucidate the design space for synchronized diffusion. Synchronized diffusion means to optimize some representation (e.g., 3D mesh) by jointly diffusing its lower-dimensional projections (e.g., 2D image). The paper analyzes existing synchronized diffusion approaches on different tasks u... | Rebuttal 1:
Rebuttal: Thank you for recognizing our work "enjoyable to read, and the analysis is very clear, elucidating the design space." We will correct the typos accordingly.
Blurry outputs of depth-to-360-panorama.
---
Depth-to-360-panorama involves Equirectangular projection which introduces more distortions th... | Summary: This paper introduces SyncTweedies, a novel framework to generate diverse visual content such as ambiguous images, panoramas, mesh textures, and Gaussian splat textures. The method uses a synchronization process that averages outputs of Tweedie's formula across multiple instance spaces, eliminating the need fo... | Rebuttal 1:
Rebuttal: Thank you for your recognition of our work such as "the framework's versatility is impressive", and "the methodological innovation could inspire new directions in the field of generative models." Below, we answer your insightful comments.
How does $\texttt{SyncTweedies}$ handle scenarios where ma... | Summary: This paper investigates content generation within a target space using pretrained diffusion models operating in projected subspaces. The authors analyze five variants of the DDIM procedure, performed separately in each subspace and aggregated in the target space using known projection and unprojection operator... | Rebuttal 1:
Rebuttal: Thank you for your in-depth reviews and feedback. We find that some of the reviews have been addressed in the appendix but are missing in the main paper. Please understand that these details were moved to the appendix due to page limits. However, we agree that providing more details in the main pa... | null | null | Rebuttal 1:
Rebuttal: We appreciate all reviewers for their constructive and insightful comments. Some comments are addressed in the global rebuttal due to length constraints.
***Please refer to the attached PDF file for qualitative and quantitative results.***
[KJxR] Do cases mathematically converge to the same meth... | NeurIPS_2024_submissions_huggingface | 2,024 | null | null | null | null | null | null | null | null |
Mimicking To Dominate: Imitation Learning Strategies for Success in Multiagent Games | Accept (poster) | Summary: This paper addresses the issue of training instability and slow convergence in MARL caused by the changing strategies of other agents. It proposes reducing the uncertainty faced during training by imitating the opponents' strategies. To address the challenge that opponents' actions are usually unobservable, th... | Rebuttal 1:
Rebuttal: We thank the reviewer for reading our paper and providing positive feedback.
> The performance of many methods in the experiments fluctuate. How many seeds were used in the experiments? Previous experiments on SMACv2 usually uses more than 5 random seeds.
The main reason the win-rate curves are... | Summary: This paper presents a new framework of multi-agent reinforcement learning (MARL) by modeling opponents’ behaviors through imitation learning.
Strengths: The motivation and the method is well described and the performance is tested in extensive experiments with challenging tasks against SOTA methods.
Weakness... | Rebuttal 1:
Rebuttal: We thank the reviewer for reading our paper and for the positive feedback.
> If I understand correctly, an important feature of the proposed framework is that all ally agents jointly learns a single joint model of the enemies. On the other hand, the SupMAPPO agents learn the enemies’ next states... | Summary: The paper studies cooperative-competive MARL. It utilizes imitation learning to comprehend and anticipate the next actions of the opponent agents (enemies), aiming to mitigate uncertainties of the controlled agents (allies) with respect to the game dynamics.
Strengths: - The paper studies a very interesting p... | Rebuttal 1:
Rebuttal: We thank the reviewer for carefully reading our paper and providing us with constructive feedback.
> Related work needs improvement. In opponent modelling, the authors claim that: "All the aforementioned related works require having access to opponent’ observations and actions during training and/... | null | null | Rebuttal 1:
Rebuttal: We thank the reviewers for carefully reading our paper and providing constructive feedback, which we have made efforts to address. Please find a summary of our responses below.
**Reviewer bLoo** mentioned some related work in opponent/agent modeling. In response, we have explicitly provided a dis... | NeurIPS_2024_submissions_huggingface | 2,024 | null | null | null | null | null | null | null | null |
MemoryFormer : Minimize Transformer Computation by Removing Fully-Connected Layers | Accept (poster) | Summary: In this paper, the authors present a new transformer model named MemoryFormer, which utilizes locality-sensitive hashing to replace the matrix multiplication operation of the fully-connected layer. This paper can minimize the computational complexity of transformer by removing all the computations except for s... | Rebuttal 1:
Rebuttal: ### **1. Data regarding the number of parameters.**
We report the number of parameters for each model in the row "Model Size" in table below.
||Pythia-70M|MF-tiny|Pythia-160M|MF-small|Pythia-410M|MF-base|pythia-1B|pythia-2.8B|pythia-6.9B|
|-|-|-|-|-|-|-|-|-|-|
|Layers/Hidden_dim|6/512|6/512|12/768... | Summary: A FLOPs-reduction strategy designed for transformer is proposed by this paper. The author introduces MemoryFormer, a transformer model that is built using the Memory Layer instead of fully-connected layer. According to the paper, the author claims that the MemoryFormer has the minimum computation because the M... | Rebuttal 1:
Rebuttal: ### **1. Combining MemoryFormer with other efficient attention method.**
We incorporate the efficient attention modules proposed by Linformer and cosFormer with MemoryFormer-tiny, and train each model for 8000 steps on PILE dataset.
|Attn. Type|baseline MHA|Linformer|cosFormer|
|-|-|-|-|
|**Val pp... | Summary: This paper introduces a novel neural network call MemoryFormer. This is a modified version for transformer that eliminates the dense layer (FC layer). The proposed MemoryFormer tries to address the problem of high computational complexity of decoder-style generative models. Concretely, the author proposes the ... | Rebuttal 1:
Rebuttal: ### **1. Data regarding the inference latency.**
We measure the inference latency for different models on Intel Xeon 8377C CPU and A100 GPU using custom kernel, and show the results in the table below.
||Pythia-70M|MF-tiny|Pythia-160M|MF-small|Pythia-410M|MF-base|
|-|-|-|-|-|-|-|
|Layers|6|6|12|1... | Summary: This work proposes to replace most linear layers of transformers by trainable hash-tables. The new modules---called memory layers---rely on locality sensitive hashing to obtain relevant indices within several hash tables, and returns a linear combination of the associated vectors. To overcome the non-different... | Rebuttal 1:
Rebuttal: ### **1.1. What is the experimental setup used to trained Pythia models? How many steps are used during training?**
All models are trained for 143000 steps. The total batch size is 1024 samples and the sequence length of each sample is 2048. We use exactly the same experimental setups, such as Ad... | Rebuttal 1:
Rebuttal: We provide the training loss curves for Pythia-410M and MemoryFormer-base in the rebuttal PDF file.
Pdf: /pdf/f17948070a768cbd0c69a780ac06a7e1df1eb344.pdf | NeurIPS_2024_submissions_huggingface | 2,024 | null | null | null | null | null | null | null | null |
Achieving Constant Regret in Linear Markov Decision Processes | Accept (poster) | Summary: The paper studies constant regret guarantees for linear MDPs. The main result is the first algorithm with this guarantee. The algorithm is a modification of LSVI-UCB with a careful quantization technique and some form of action elimination. This improves upon the regret of previous works by a factor of $\log K... | Rebuttal 1:
Rebuttal: Thank you for your detailed feedback, and we address your questions as follows:
**Q1. Why is it significant to improve the logarithmic dependence on $K$?**
**A1.** We would like to first highlight that achieving constant regret is actually an important topic in bandits (Abbasi et al., Papini et ... | Summary: This paper proposed a constant regret learning algorithm for linear MDPs with approximation error (misspecification). Specifically, it proved an instance dependent regret that is independent of the number of episodes. In addition, the algorithm does not require prior knowledge of misspecification level or subo... | Rebuttal 1:
Rebuttal: Thank you for your positive feedback. We address your questions as follows:
**Q1. How do the layer-dependent quantization and certified estimator contribute to getting rid of the $\log K$ dependence?**
**A1.** We would like to highlight the contribution of the layer-dependent quantization and ce... | Summary: This paper gives a constant regret (total regret independent of the number of episodes $K$) algorithm for online reinforcement learning of linear MDPs. The environment is considered to be a $\zeta$-approxinate linear MDP as in (Jin et al 2019), with an assumption that the misspecification level ($\zeta$) is no... | Rebuttal 1:
Rebuttal: Thank you for the positive feedback on our work! We will address your questions as follows:
**Q1. Adding some experimental results in addition to the theoretical result**
A1. Thank you for the suggestion. We have included some experimental results on a synthetic dataset during the rebuttal phase... | Summary: The paper presents improved regret bounds for Linear MDPs when the suboptimality gap \Delta is known. Most crucially (a bit surprisingly), the regret bounds are constant in a number of episodes K
Strengths: The contributions are broadly theoretical, and the removal of dependence on the number of episodes K se... | Rebuttal 1:
Rebuttal: Thank you for your positive feedback. We address your questions as follows.
**Q1. What's the meaning of notation $\lVert wb_h^\pi\rVert_2^2$ in Proposition 3.2?**
**A1.** Sorry for the typo caused by the LaTeX macros. The $\lVert wb_h^\pi\rVert_2^2$ should be $\lVert \mathbf{w}_h^\pi\rVert_2^2$... | Rebuttal 1:
Rebuttal: We first thank all reviewers for their valuable feedback.
In response to Reviewers 8jf2, 33Lb, and YEKy, we added experiments on synthetic datasets to verify the performance of the algorithm and the contribution of each component. Specifically, we consider a linear MDP with $S = 4$, $A = 5$, $H ... | NeurIPS_2024_submissions_huggingface | 2,024 | null | null | null | null | null | null | null | null |
Voxel Proposal Network via Multi-Frame Knowledge Distillation for Semantic Scene Completion | Accept (poster) | Summary: This paper introduces Voxel Proposal Network which achieves semantic scene completion from both voxel and BEV perspectives. Beginning with confident voxel proposals, the information is propagated to other voxels, aiming to handle dynamic aspects. Multi-Frame Knowledge Distillation is also incorporated to accur... | Rebuttal 1:
Rebuttal: We sincerely thank you for your valuable comments on our paper's motivation, writing, and experiment. Below, we respond to your questions point-to-point.
**Q1: At the beginning of the abstract, this paper mentioned that most previous methods use 3D/2D convolutions or attention mechanisms, having ... | Summary: To directly construct scene geometry and accurately propagate features from relted voxels, the paper proposes VPNet with a voxel proposal mechanism to identify confident voxels for completion and a multi-frame knowledge distillation scheme to fuse information from multi-sweep LiDAR. VPNet achieves better perfo... | Rebuttal 1:
Rebuttal: We sincerely thank you for your valuable comments on our paper's symbols and motivation. Below, we respond to your questions point-to-point.
**Q1: For clarity, the paper uses massive symbols to elaborate technical details which is hard to follow. I think there could be some abstraction.**
**Answ... | Summary: The author introduced the Voxel Proposal Network (VPNet), a dual-branch semantic scene completion method with two key innovations.
First, the Confident Voxel Proposal (CVP) module, which includes offset learning and voxel proposal, generates a confident feature map based on the semantics-embedded feature map, ... | Rebuttal 1:
Rebuttal: We sincerely thank you for your valuable comments on our paper's experiments and core contribution. Below, we respond to your questions point-to-point.
**Q1: Lack of Comparison with State-of-the-Art Methods: The performance of VPNet is not compared with current state-of-the-art methods such as SC... | Summary: This paper focuses on semantic scene completion. The authors propose a novel voxel proposal network and combine multi-frame knowledge distillation technique to reconstructs the scene geometry and implicitly models the uncertainty of voxel-wise semantic labels by presenting multiple possibilities for voxels. Th... | Rebuttal 1:
Rebuttal: We sincerely thank you for your valuable comments on the motivation and figure. Below, we respond to your questions point-to-point.
**Q1: My biggest concern is that the author's motivation for using MFKD is not clearly explained. Why use it to predict voxel-wise labels?**
**Answer:** Here, we in... | Rebuttal 1:
Rebuttal: We sincerely thank all reviewers for their constructive comments. We provide a rebuttal PDF containing extra figures and tables to answer the common questions raised by reviewers. **The detailed explanations can be found in the point-to-point response to every reviewer**.
**Q1: Too complicated fi... | NeurIPS_2024_submissions_huggingface | 2,024 | null | null | null | null | null | null | null | null |
PointAD: Comprehending 3D Anomalies from Points and Pixels for Zero-shot 3D Anomaly Detection | Accept (poster) | Summary: The paper titled "PointAD: Comprehending 3D Anomalies from Points and Pixels for Zero-shot 3D Anomaly Detection" introduces PointAD, a novel framework designed for zero-shot 3D anomaly detection by leveraging both point clouds and RGB images. The approach builds on the strong generalization capabilities of CLI... | Rebuttal 1:
Rebuttal: # Response to Reviewer dthJ
We appreciate your recognition of our work. Your insightful suggestions help make our sensitivity analyses more comprehensive and convincing. In addition to the response to your valuable feedback, we conducted further explorations on rendering quality and input resolut... | Summary: The paper introduces PointAD, the first approach to explore the domain of zero-shot 3D anomaly detection, leveraging CLIP's strong generalization to identify anomalies in previously unseen objects. It offers a unified framework that integrates 3D point clouds with 2D renderings, employing hybrid representation... | Rebuttal 1:
Rebuttal: # Response to Reviewer GgEU
Thanks for acknowledging our work. Your insightful comments help highlight our technological contributions in comparison to concurrent works and further promote comprehensive evaluations.
Due to the character limit, please refer to the bottom of **General Response** ... | Summary: The paper propose a unified framework, PointAD,to detect 3D anomalies in a ZS manner. Hybrid representation learning is proposed to incorporate the generic normality and abnormality semantics into PointAD. PointAD can incorporate 2D RGB information in a plug-and-play manner for testing, which can perform ZS M3... | Rebuttal 1:
Rebuttal: # Response to Reviewer 1QVN
Thank you very much for your acknowledgment and insightful comments. Your comments inspire us to explore the robustness of PointAD from the rendering process and input under different conditions.
**Q1: If the renderings are of poor quality, could the model's performan... | null | null | Rebuttal 1:
Rebuttal: # General Response
Dear Reviewers and ACs,
We very much appreciate the insightful and detailed review. We are delighted to hear the encouraging comments from all of the reviewers, including "`novel approach`" (Reviewer **1QVN** and **dthJ**), "`methodological soundness`" (Reviewer **GgEU**), "`st... | NeurIPS_2024_submissions_huggingface | 2,024 | null | null | null | null | null | null | null | null |
FreqMark: Invisible Image Watermarking via Frequency Based Optimization in Latent Space | Accept (poster) | Summary: This paper considers the problem of watermarking of images. In particular, it introduces a method called FreqMark where watermark is embedded in the latent frequency space obtained after variation auto encoder (VAE) encoding. Numerical results are carried out on test datasets containing 500 randomly selected i... | Rebuttal 1:
Rebuttal: Dear Reviewer TJux,
We sincerely appreciate your thoughtful review and constructive suggestions on our paper. We have carefully considered your comments and have taken them into account.
> W1: One of the main weaknesses of the work is that there is no convincing proof or theoretical justificati... | Summary: This paper introduces FreqMark, a novel invisible watermarking method that enhances digital content protection through optimization in the image's latent frequency space. Experiments have been conducted to demonstrate the robustness against regeneration attacks like VAE and diffusion model.
Strengths: 1. The ... | Rebuttal 1:
Rebuttal: Dear Reviewer kwvu,
Thank you very much for your careful reading and recognition of our work, we will address each one sequentially below:
> W1: Lacks some theoretical analysis. In the experiments, do you need to use the same VAE model for the attack as you have used during watermarking training... | Summary: This paper proposes a method called FreqMark that is able to prevent the invisible watermarks from the regeneration attack. By using the unconstrained optimization of the image latent frequency space obtained after VAE encoding, the proposed FreqMark achieves better robustness against the regeneration attacks ... | Rebuttal 1:
Rebuttal: Dear Reviewer BgZJ,
Thank you for your insightful feedback and questions. We address the concerns as follows:
> W1: The major contribution of this paper is to introduce a kind of image watermark robust to regeneration attacks. However, this part is not highlighted in its title. The authors are s... | Summary: The authors propose a new post-processing watermark for imagery, FreqMark. The FreqMark embeds a binary message into the frequency domain of a VAE-encoded image via a small perturbation, and then following IFFT + decompression, utilizes a pre-trained image encoder to extract the message. A PSNR + LPIPS metric ... | Rebuttal 1:
Rebuttal: Dear Reviewer BZ7h,
Thank you for your suggestions. We address the concerns as follows:
> W1: In my opinion, the FreqMark is an incremental variant of the StegaStamp [1]. ... The resemblance of equation (10) in this manuscript to the loss function equation (2) in [1] begs the question of novelty.... | null | NeurIPS_2024_submissions_huggingface | 2,024 | null | null | null | null | null | null | null | null |
Bridge the Modality and Capability Gaps in Vision-Language Model Selection | Accept (poster) | Summary: This paper considers a zero-shot image classification strategy by selecting the most appropriate Pre-Trained VLM from the VLM Zoo, relying solely on the text data of the target dataset without access to the dataset’s images. Two challenges, i.e., the “Modality Gap” across two different modalities, and the "Cap... | Rebuttal 1:
Rebuttal: We appreciate the reviewer's comments. Below are our responses:
* Q1：Use open-source datasets may violate zero-shot data restrictions.
* A1:
1. Firstly, we want to emphasize that our method focuses on how to use open-source datasets to help us select the appropriate models from the VLM Zoo ... | Summary: With the popularity of Vision Language Model (VLM) research in recent years, many versions have emerged, forming the VLM Zoo. This paper aims to select the most appropriate pre-trained VLM from the VLM Zoo, relying solely on texts of the target dataset without access to images. Two challenges are analyzed, nam... | Rebuttal 1:
Rebuttal: We appreciate the reviewer's comments. Below are our responses:
* Q1: Why not use combination of images and texts (or even text-image pairs) for VLM selection？
* A1:
1. Firstly, we want to emphasize that in our paper, "Language-Only" means that for the target new task's dataset, we can only ... | Summary: The paper proposes SWAB the modificaiton of LOVM to mitigate the negative impact of two gaps of LOVM: the modality gap and capability gap. The resutls show the effectiveness of SWAB.
Strengths: 1. The motivation of this paper is clear and structure is easy to follow.
Weaknesses: 1. For the modality gap, the ... | Rebuttal 1:
Rebuttal: We appreciate the reviewers' comments. Below are our responses:
* Q1: How many image samples were used in the experiment.
* A1: We followed LOVM's approach [1] to use 23 datasets. Each of the 23 datasets was used as a test dataset in turn, with the remaining 22 datasets serving as open-source dat... | Summary: This article focuses on selecting the best model from a visual-language multimodal model zoo for specific downstream tasks without having images of those tasks. It provides a detailed analysis of two challenges faced in this problem — Modality Gap and Capacity Gap. This paper proposes a method called SWAB, whi... | Rebuttal 1:
Rebuttal: We greatly appreciate the reviewers' recognition of our work. Below are our responses to the relevant questions:
* Q1: The performance of ModelGPT is slightly inconsistent with the results in its paper.
* A1: This is because we conducted multiple repeated experiments using different random seeds. ... | Rebuttal 1:
Rebuttal: We would like to express our deepest gratitude to the reviewers for the meticulous examination of the paper and their insightful and valuable comments. We acknowledge that most reviewers observed the shining point, saying the proposed method is novel (qHTS, mpun), and interesting (qHTS, mpun), wel... | NeurIPS_2024_submissions_huggingface | 2,024 | null | null | null | null | null | null | null | null |
Testing Calibration in Nearly-Linear Time | Accept (poster) | Summary: This paper considers the property testing of calibration, under the lower distance to calibration metric. By property testing, the authors adapt the standard definition in the TCS literature, where if the distance is at least $\epsilon$ then the algorithm will reject, and accept if the distance is 0. The main ... | Rebuttal 1:
Rebuttal: Thank you for your positive review and careful reading. We are also encouraged that we were able to develop a practical solver for our LP of interest by using its structure, which uses less sophisticated tools than the recent LP solvers from the TCS community.
Theorem 1’s runtime does not depend ... | Summary: This paper introduces a property testing formulation of verifying calibration, along with efficient algorithms for solving this problem as well as a relaxed/tolerant version. The empirical results support the theory and justify the efficacy of the proposed algorithms.
Strengths: **Originality:** The calibrati... | Rebuttal 1:
Rebuttal: Thank you for your encouraging feedback.
We agree with you that applications of our results to model selection in more concrete settings is an exciting potential implication of our paper, and believe that this is a natural further direction to take our techniques. Because the algorithmic theory ... | Summary: The paper studies the problem of *calibration testing*.
Here, we are given a distribution $D$ over outcomes and the goal
is to decide if the distribution is calibrated; specifically,
the property testing problem they formulate distinguishes
between perfectly calibrated distributions and those that
are $\epsilo... | Rebuttal 1:
Rebuttal: Thank you for your encouraging review. We are glad that you found the problem we study interesting, and that our paper makes important contributions to it.
We apologize for any confusion in framing. We agree that our Theorem 1 is really just a fast solver for the empirical smCE. However, we prese... | Summary: This paper studies testing the calibration of predictors through joint distributions and contributes efficient methods using appropriate measures.
Strengths: Originality:
There are new methods.
The work can be considered a novel combination of well-known techniques.
It is clarified how this work differs fro... | Rebuttal 1:
Rebuttal: We thank the reviewer for your careful reading – we are glad that you found our work to be complete, and that you found the results important.
We were confused by the reviewer’s comment about the lack of care or honesty in evaluating our work – we take such concerns very seriously, and would ver... | null | NeurIPS_2024_submissions_huggingface | 2,024 | null | null | null | null | null | null | null | null |
On the Noise Robustness of In-Context Learning for Text Generation | Accept (poster) | Summary: The paper introduces a new method to deal with noisy annotations for in-context learning. The authors suppose that the examples that cause higher perplexity than their neighbors are more likely corrupted than their neighbors. So, the authors suggest replacing the examples, causing suspiciously high perplexity ... | Rebuttal 1:
Rebuttal: Thanks for your recognition and the valuable suggestions. Please find our response below.
**1. Evaluation on the not-demonstration-selection-based baselines [W1, L1]**
Thank you for the great suggestion. Here, we add Zero-Shot baseline, as well as some CoT-related baselines, including Zero-Sho... | Summary: The paper "On the Noise Robustness of In-Context Learning for Text Generation" investigates how LLMs handle noisy annotations during in-context learning (ICL). The authors propose a method called Local Perplexity Ranking (LPR) that replaces noisy candidates with nearby neighbors that are less noisy. They also ... | Rebuttal 1:
Rebuttal: Thanks for your recognition and the valuable suggestions. Please find our response below.
**1. The implementation code is missing [W1]**
The code and data have been provided in the **Supplementary Material** of the original submission. Once published, we will upload the code and data to **GitHub... | Summary: This paper proposes Local Perplexity Ranking (LPR), a method to improve the robustness of in-context learning for text generation tasks when dealing with noisy annotations. The key contributions are:
- Empirically demonstrating that noisy annotations hurts performance of in-context learning for text generatio... | Rebuttal 1:
Rebuttal: Thank you for the constructive and elaborate feedback. Please find our response below.
**1. Clarification of motivation and evaluation [W1, Q1]**
There might be some misunderstandings, which are clarified in the following.
* **"The evaluated tasks are simple"**. In this work, we evaluate the ef... | null | null | Rebuttal 1:
Rebuttal: # **General Response**
We thank All the reviewers for their time and valuable comments. We are glad that reviewers find this work focuses on an **important** and **practical** problem (tEJo, DyDA) with **clear** analysis (Fz4p). We are also encouraged that reviewers find that the method is **nove... | NeurIPS_2024_submissions_huggingface | 2,024 | null | null | null | null | null | null | null | null |
What is Flagged in Uncertainty Quantification? Latent Density Models for Uncertainty Categorization | Accept (poster) | Summary: This paper proposes a model-agnostic framework named DAUC to categorize uncertain examples flagged by UQ methods, which introduces the confusion density matrix and provides post-hoc categorization for model uncertainty. Besides, this paper categorizes suspicious examples identified by a given uncertainty metho... | Rebuttal 1:
Rebuttal: We sincerely appreciate the reviewer's efforts and constructive comments in improving our paper. We will respond to each of the questions in turn:
---
### Q1: The categorization of uncertain examples.
A1: The categorization of uncertainty into three non-exclusive classes (OOD, Bnd, and IDM) is a ... | Summary: In this paper, the authors present a unique approach for classifying uncertainty into distinct categories, providing insights into the reasons behind labeling a particular sample as suspicious or highly uncertain. They develop a kernel density-based confusion density matrix for any neural network which separat... | Rebuttal 1:
Rebuttal: We deeply appreciate the reviewer's time and effort in evaluating our paper. We will now answer each of the questions:
---
### Q1: Choice of the latent space
A1: We compute kernel similarity scores within the latent space, as opposed to the input space or basic semantics space, to ensure that the... | Summary: I have read the other reviews and all rebuttals. The other reviewers are positive overall, and the authors provided very detailed and thorough rebuttals. I have increased my score from "5: Borderline accept" to "6: Weak Accept".
***
***
***
***
The authors propose an approach for categorizing examples which ... | Rebuttal 1:
Rebuttal: Thank you for your encouraging comments and appreciation of our paper's clarity, novelty, and presentation. We now address each specific question in turn:
---
### Q1: Working with more classes, computation, scalability, and additional experiments for Sec. 4.4.
A1: Technically, there's **no strict... | Summary: Paper introduces a framework to detect and categorize different model uncertainty types in classification setting. Proposer model-agnostic (with some assumptions of model structure) uncertainty quantification (UQ) relies on kernel density estimation on latent space representation by defining scores for OOD, Bn... | Rebuttal 1:
Rebuttal: We sincerely appreciate the reviewer's thoughtful insights and constructive feedback. We will respond to each of the questions in turn.
---
### Q1: Dataset and model architecture choices.
A1:
Our primary directive in selecting the Dirty-MNIST dataset was to anchor our work in a realm of reprodu... | Rebuttal 1:
Rebuttal: We extend our sincere gratitude to all reviewers for their insightful comments, valuable suggestions, time, and efforts in evaluating and improving our paper.
We thank all reviewers for their affirmation of our work’s **novelty** (reviewers: sVgx, hHZH, 7E4k), **presentation** (reviewers: sVgx, h... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Don’t blame Dataset Shift! Shortcut Learning due to Gradients and Cross Entropy | Accept (poster) | Summary: This paper theoretecally and empirically showed that the inductive bias of default- ERM maximizing the margin causes shortcut learning in a linear perception task.
It proposed uniform margins that leads to models that depend more on the stable than the shortcut feature and suggested loss functions encourage u... | Rebuttal 1:
Rebuttal: Thank you for your thoughtful feedback. We addressed your questions and concerns below. If any residual concerns remain, we would be glad to discuss further. If no concerns remain, we would appreciate it if you could raise your score.
**[The theory itself has limited applicability due to the line... | Summary: This paper provides an in-depth analysis of the phenomenon of "shortened learning" in machine learning models, especially in the context of perceptual tasks. The authors confirm that basic empirical risk minimization (ERM) methods tend to prefer models that depend on shortcut features, even when models can ach... | Rebuttal 1:
Rebuttal: Thank you for your thoughtful feedback. We addressed your questions and concerns below. If any residual concerns remain, we would be glad to discuss further. If no concerns remain, we would appreciate it if you could raise your score.
**[How does MARG-CTRL perform in scenarios where the shortcut ... | Summary: Having had my concerns addressed by the authors I have updated my score.
-----------------------------------------------------------------------------------------
* The paper proposes an explanation for why neural networks tend to learn spurious features over stable fratures which lead to a lower loss and b... | Rebuttal 1:
Rebuttal: Thank you for your thoughtful feedback. We addressed your questions and concerns below. If any residual concerns remain, we would be glad to discuss further. If no concerns remain, we would appreciate it if you could raise your score.
**[I find the scatter approach to testing many margin based lo... | Summary: This paper explores the phenomenon of models using easy-to-learn spurious features (aka, shortcuts) instead of reliable but harder-to-learn true features. They find in their theoretical that max-margin relies on the spurious feature while controlling for uniforming margin induces learning the true feature. Wit... | Rebuttal 1:
Rebuttal: Thank you for your thoughtful feedback. We addressed your questions and concerns below. If any residual concerns remain, we would be glad to discuss further. If no concerns remain, we would appreciate it if you could raise your score.
**[The informal statement of Theorem 1 is too vague. In partic... | Rebuttal 1:
Rebuttal:
# General response
We thank the reviewers for their thoughtful feedback. We are glad the reviewers find that
- Our paper is well-written
- “The comments and explanations are clear and concise, and follow a clear narrative” - Xk5J
- “The paper is well-written and easy to understand” - PgqC
... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Enhancing Adversarial Contrastive Learning via Adversarial Invariant Regularization | Accept (poster) | Summary: The authors propose AIR to regulate the the process of contrastive learning. They first analyze the causal graph of contrastive learning under the adversarial setting, and try to enforce $p\left(y^R | \tilde x\right)\cdot p\left(\tilde x | x\right)$ (Eq.4) to be invariant under different interventions. Intuit... | Rebuttal 1:
Rebuttal: Many thanks for your supportive comments! Please find our responses below.
> 1. [Reply to W1] Our proposed AIR is less sensitive to the hyper-parameter (HP) than SIR.
Automatic schedule of the HP could be a good extension to our work.
We provide an ablation study of HP in Table 7. In ***Point ... | Summary: This paper proposes a novel adversarial contrastive learning method, which introduces causal reasoning method to obtain robust feature representation and enforce independence from style factors. The idea is simple and effective. In addition, the experiments are sufficient to prove the effectiveness of the prop... | Rebuttal 1:
Rebuttal: Many thanks for your constructive comments! Please find our replies below.
> 1. [Reply to W1] In ***Table A*** of Rebuttal Highlights, we report the p-value obtained by conducting a Student's t-test to show that our improvement is **significant**.
***Table A*** shows that the p-value is consiste... | Summary: This paper proposes to tackle adversarial contrastive learning via causal reasoning. Specifically, the authors discuss the scenario where adversarial examples are involved in standard invariant regularization. The authors also provide analysis of proposed invariant regularization to justify the rationality. Th... | Rebuttal 1:
Rebuttal: Many thanks for your positive comments! Please find our replies below.
> 1. [Reply to Q1] We demonstrate the results of the baseline as follows. We will update them in revision.
The results show that our proposed IR consistently performs better than the baseline.
|Label ratio = 1\% | ACL | ACL... | Summary: This paper proposed a method to advance adversarial contrastive learning by utilizing a technique called causal reasoning. The adversarial invariant regularization (AIR) proposed in this paper demonstrated a style factor. Additionally, the effectiveness of the proposed method was empirically shown using CIFAR1... | Rebuttal 1:
Rebuttal: Many thanks for your constructive comments! Please find our replies below.
> 1. [Reply to W1] We argue that our theoretical analysis is **non-trivial**.
Directly applying adversarial data to paper [1] cannot obtain AIR. It is because SIR [1] in Eq. (8) aims to enforce $p(y^R|x)$ to be style-ind... | Rebuttal 1:
Rebuttal: [**Rebuttal Highlights**]
Many thanks for supportive comments from Reviewers **ZaSM** and **Ai1a** as well as constructive comments from Reviewers **fVht** and **paeS**!
Here, we would like to demonstrate extra empirical results to resolve your concerns.
> 1. [For Reviewers **fVht** and **pae... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Hybrid Distillation: Connecting Masked Autoencoders with Contrastive Learners | Reject | Summary: This paper introduces a new hybrid distillation method for Vision Transformer. It is hybrid in the sense that two teacher models pre-trained by two different approaches, namely Contrastive Learning and Masked Image Modeling, are adopted in the distillation. In this way, the student model inherit both the diver... | Rebuttal 1:
Rebuttal: **Q1: [What kind of distillation approach is used in Section 2.]:**
**Our distillation setup remains consistent with [1][2]**, where only the features in the last layer of the ViT teacher are utilized as distillation objectives. We introduce various decoders, including linear projection and trans... | Summary: This work presents Hybrid Distillation, which attempts to distill from both supervised/CL and MIM frameworks. The work begins by revealing certain observations regarding the interplay between self-supervised pre-training and the concepts of diversity and discrimination. Subsequently, the authors propose the Hy... | Rebuttal 1:
Rebuttal: **Q1: [The experimental setup in Section 2.]:**
Please refer to the response to Q1 of Reviewer yZYM.
**Q2: [The description of the metrics.]:**
Please refer to the response to Q2 of Reviewer yZYM.
**Q3: [The analysis of the figures lacks details.]:**
The visualization of average head distan... | Summary: This paper introduce a new distillation method that complimentary harmonizes two distillation method of different properties.
Strengths: * Hybrid Distillation obtained higher accuracies than DeiT, MAE, and CLIP using them.
* Explanation with analyses (NMI, AHD, and attention map visualization)
* The paper is ... | Rebuttal 1:
Rebuttal: **Q1: [The efficiency compared to methods without distillation.]:**
i) Our Hybrid Distill, similar to many other distillation methods [1][2][8][9][10], requires an additional distillation stage after obtaining the pretrained model. **The additional costs are inherent to distillation-based method... | Summary: The paper conducts sufficient experiments and theoretical analysis on diversity and discrimination.
Meanwhile, the authors propose a simple yet effective hybrid distillation that combines contrastive learning pre-train and MIM pre-train.
This hybrid distillation achieves significant improvement on downstream... | Rebuttal 1:
Rebuttal: **Q1: [Figure 4 seems to be wrong.]:**
Figure 4 visualizes the per-layer average attention distance. In Figure 4, we further average the per-head attention distance and obtain a single average attention distance for each layer to reflect the overall distance in each layer and facilitate compariso... | Rebuttal 1:
Rebuttal: We appreciate that the reviewers recognize the pros of our paper: the soundness (gLfn, 5Rvs, yZYM) and the contribution (gLfn, yZYM) of our paper, the experimental design to re-examine discrimination and diversity (gLfn, 5Rvs, CGRR, dWYQ), the simple but effective and intuitive idea (gLfn, yZYM), ... | NeurIPS_2023_submissions_huggingface | 2,023 | Summary: This paper explores the subject of representation learning, focusing on two aspects: discrimination and diversity. Contrastive learning (CL) exhibits superior discrimination capabilities but suffers from limited diversity. Conversely, masked image modeling (MIM) offers greater diversity but shows weaker discri... | Rebuttal 1:
Rebuttal: **Q1: [Consistent usage of the terms "asymmetric X" throughout the paper]:**
Thanks for the suggestions. The terms “asymmetric X” are consistent to indicate the asymmetric designs, which are our primary focus in Section 2. In detail, the terms '**asymmetric architecture**' and '**asymmetric des... | null | null | null | null | null | null |
An Optimization-based Approach To Node Role Discovery in Networks: Approximating Equitable Partitions | Accept (poster) | Summary: The paper offers a new perspective on the problem of role extraction while defining node roles based on the ideas of equitable partitions and and graph-isomorphism tests, the Weisfeiler-Leman algorithm. The paper studies two associated optimization problems (cost functions) inspired by graph isomorphism testin... | Rebuttal 1:
Rebuttal: Dear Reviewer rRxj,
Thank your for your review.
Regarding your questions:
1. Expressivity of GNNs: The cited references do, in fact, prove the statement. The contraposition of Lemma 2 in (Xu et. al [43]) yields "If the WL test decides G_1 and G_2 are isomorphic, then any GNN maps G_1 and G_2 to... | Summary: This paper presents a novel perspective on the problem of node role extraction in complex networks, highlighting its distinctions from community detection. The authors propose a definition of node roles and introduce two optimization problems based on graph-isomorphism tests, the Weisfeiler-Leman algorithm, an... | Rebuttal 1:
Rebuttal: Dear Reviewer WiNy,
Thank your for your review.
Regarding your questions:
1. MMSBM and RID$\epsilon$R. While the comparison with RIDR does seem interesting, we were unable to find an implementation for neither RIDR nor inference of the mixed membership SBM. If time permits, we would like to als... | Summary: The paper considers a relaxed definition of the coarsest equitable partition (CEP), which equals the final partition of Weisfeiler-Leman or the color refinement algorithm in the original version. The authors allow to specify the number of cells of the partition and derive a related optimization problem. From t... | Rebuttal 1:
Rebuttal: Dear Reviewer WiNy,
Thank your for your review.
Regarding your question, the results of Algorithm 1 heavily depend on the clustering algorithm used. If one was to use a clustering algorithm that does not assign two data points with the same value to different clusters, then Algorithm 1 would ou... | Summary: The authors propose the notion of equitable partitions from graph isomorphism literature in order to partition the nodes of a network according to their structural roles. They study two optimization problems for approximately recovering such equitable partitions. Analogous to the SBM model, the authors devise ... | Rebuttal 1:
Rebuttal: Dear Reviewer ZAin,
Thank you for your review.
Regarding your question, one has to keep in mind that we are concerned with finding "global" roles. That is, we consider the whole dataset as one graph with multiple connected components. On the proteins dataset, there are $1278$ classes of the *ex... | Rebuttal 1:
Rebuttal: Dear Reviewers,
Thank you all for your reviews.
A common theme in your reviews was a need to broaden the experimental part of the paper. We have thus extended our experiments by two experiments proposed by Reviewer nMtV and originally conducted by [1]. In particular, this includes the downstre... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Generator Identification for Linear SDEs with Additive and Multiplicative Noise | Accept (poster) | Summary: This paper presents conditions for identifying the generator of a linear stochastic differential equation (SDE) with additive and multiplicative noise. The authors derive sufficient conditions for identifying the generator of both types of SDEs and offer geometric interpretations.
Strengths: I think this pa... | Rebuttal 1:
Rebuttal: Thank you for your helpful comments. We address your comments point by point as follows. We have also updated our paper according to your comments, and we believe that the quality has been significantly improved thanks to your insightful comments.
**Answers to weaknesses:**
>1. **W:** practical... | Summary: This paper addresses the interventional identifiability of two classes of stochastic differential equations. Theoretical results are illustrated with simulation of identifiable and non-identifiable cases.
Strengths: The paper is carefully written and the theoretical results appear sound, although I could not... | Rebuttal 1:
Rebuttal: Thank you for your helpful comments. We address your comments point by point as follows. We have also updated our paper according to your comments, and we believe that the quality has been significantly improved thanks to your insightful comments.
**Answers to weaknesses:**
>1. **W:** target a... | Summary: This article derives relationships for identifying generators of linear SDEs in both the case of additive and multiplicative noise. It is written in a very classically statistical manor, with a great number of citations from the statistical research community.
Strengths: This article identifies two apparently... | Rebuttal 1:
Rebuttal: Thank you for your helpful comments. We address your comments point by point as follows. We have also updated our paper according to your comments, and we believe that the quality has been significantly improved thanks to your insightful comments.
**Answers to weaknesses:**
>1. **W:** Proofs c... | Summary: This paper derives some sufficient conditions for identifying linear stochastic differential equations (SDEs). The validity of the identifiability conditions is clarified on synthetic simulation data.
Strengths: - The identification problem of ordinary/stochastic differential equations has been a hot topic in... | Rebuttal 1:
Rebuttal: Thank you for your helpful comments. We address your comments point by point as follows. We have also updated our paper according to your comments, and we believe that the quality has been significantly improved thanks to your insightful comments.
**Answers to weaknesses:**
>1. **W:** The paper... | null | NeurIPS_2023_submissions_huggingface | 2,023 | Summary: The present work considers the problem of identifying the generators of autonomous linear
stochastic differential equations (SDEs) with an additive or multiplicative
noise term from the law of its solution process (e.g., in case of additive noise
this is a Gaussian process) with a given fixed initial state. Th... | Rebuttal 1:
Rebuttal: Thank you for your helpful comments. We address your comments point by point as follows. We have also updated our paper according to your comments, and we believe that the quality has been significantly improved thanks to your insightful comments.
**Answers to weaknesses:**
>**W1:** introduce g... | Summary: As I am not an expert in causal inference nor in SDEs, I will start with summarising the goal of the paper how I understood by reading it and some references therein. This summary will certainly reveal my ignorance regarding this research area, but I prefer to give this transparency and hope my review will be ... | Rebuttal 1:
Rebuttal: Thank you for your helpful comments. We address your comments point by point as follows. We have also updated our paper according to your comments, and we believe that the quality has been significantly improved thanks to your insightful comments.
**Answers to weaknesses:**
>1. **W:** The beginn... | null | null | null | null |
Algorithm Selection for Deep Active Learning with Imbalanced Datasets | Accept (poster) | Summary: This paper proposes a novel algorithm called TAILOR that adaptively chooses the active learning algorithm from the candidate set of multiple algorithms. Novel reward functions are proposed that encourage choosing the algorithm such that both class diversity, as well as informativeness, are maintained while s... | Rebuttal 1:
Rebuttal: We would like to thank the reviewer for the detailed review and we would like to address the reviewer's concerns below.
### Ablation Study on Rarest Class Performance
As shown in the PDF attached to our overall rebuttal (Figure 3), TAILOR significantly outperforms all other methods in improving ... | Summary: This paper proposes a new method for algorithm selection in active learning. The problem is treated as a multi-armed bandit problem and the proposed method is based on Thompson sampling. Extensive experiments are conducted on several datasets, including both multi-class and multi-label classification setting.
... | Rebuttal 1:
Rebuttal: We would like to thank the reviewer for the detailed review and we would like to address the reviewer's concerns below.
### Motivation Behind Algorithm Selection (Why not always BADGE?)
There are many cases BADGE is not the best. For example, under class imbalance, BADGE has been shown to underp... | Summary: The paper proposed TAILOR, a Thompson Sampling framework for active learning algorithm selection for unlabeled, possibly imbalanced datasets by framing it as a multi-arm bandit problem. The authors compared with random Meta and another meta-framework ALBL, along with other AL methods on 10 multi-class multi-la... | Rebuttal 1:
Rebuttal: We would like to thank the reviewer for the detailed review and we would like to address the reviewer's concerns below.
### Clarification on Reward Design and Choices:
Regarding the choice of reward function, we believe this largely depends on the goal of using AL in practice. If the practitione... | Summary: Selecting the most appropriate active learning algorithm for a given dataset poses a significant challenge when applying active learning in real-world scenarios. This challenge stems from the fact that the performance of different active learning strategies could vary significantly across various scenarios and... | Rebuttal 1:
Rebuttal: We would like to thank the reviewer for the detailed review and we would like to address the reviewer's concerns below.
### Clarification on Diversity based algorithms
As we have noted in Appendix A.1, all of the popular diversity-based algorithms rely on iterative procedure in choosing examples ... | Rebuttal 1:
Rebuttal: We would like to thank all of the reviewers for their insightful reviews. Attached PDF includes all of the figure plots mentioned in our rebuttals to each individual reviewer.
Pdf: /pdf/633dbe5a8b624387efeaf20f5cb3b028fc3177ff.pdf | NeurIPS_2023_submissions_huggingface | 2,023 | Summary: This paper proposes an algorithm, TAILOR, that iteratively and adaptively selects candidate active learning algorithm to gather class-balanced examples. Experimental results demonstrate that TAILOR achieves comparable or better performance than the best of the candidate algorithms.
Strengths: The paper prese... | Rebuttal 1:
Rebuttal: We would like to thank the reviewer for writing a detailed review, but the reviewer appears to have some fundamental misunderstanding of what our paper is studying.
### Research Problem and Related Work
We are not proposing another active learning algorithm. Our goal is to come up with an online a... | null | null | null | null | null | null |
Discover and Align Taxonomic Context Priors for Open-world Semi-Supervised Learning | Accept (poster) | Summary: This paper addresses the open-world semi-supervised learning (OSSL) problem and proposes taxonomic context priors discovering and aligning (TIDA), which considers the taxonomic hierarchy of classes.
Basically, the proposed method is based on [19], which assigns pseudo-labels to unlabeled samples by means of t... | Rebuttal 1:
Rebuttal: >**Q1:** Compare with DCCL[46].
>
> **A1:** **i).** Motivation: DCCL[46] aims to enhance representation learning by applying contrastive learning among dynamically estimated conceptions. In contrast, our TIDA aims to enhance pseudo label quality by hierarchically clustering samples and discovering... | Summary: Previous research has primarily focused on using pre-defined single-granularity labels as priors for recognizing novel classes. However, classes naturally adhere to a taxonomy, enabling classification at multiple levels of granularity and offering richer supervision through underlying relationships. To address... | Rebuttal 1:
Rebuttal: > **Q1&Q4:** Can pre-defined taxonomic priors be utilized?
>
> **A1:** We **CANNOT** directly construct taxonomic priors using language models, such as the mentioned WordNet[1]. This because we have no knowledge about novel classes in open-world SSL (neither semantic names nor attributes).
> Thu... | Summary: This paper tackles open-world semi-supervised learning and proposes to use multi-granularity labels as taxonomic context priors to leverage hierarchical supervision to enhance representation learning and improve the quality of pseudo labels. A taxonomic context discovery module is used to construct hierarchica... | Rebuttal 1:
Rebuttal: We sincerely thank you for your valuable comments! Please find our detailed response below.
>**Q1:** Compare with additional baselines.
>
>**A1:** Below, we compare our method with [1-5]. Our method achieves the best results on all datasets, as shown below. For a fair comparison, following [1-5],... | Summary: In this paper, a new Taxonomic context pIrors Discovering and Aligning (TIDA) which exploits the relationship of samples under various granularity is proposed. TIDA comprises two key components: i) A taxonomic context discovery module that constructs a set of hierarchical prototypes in the latent space to disc... | Rebuttal 1:
Rebuttal: We sincerely thank you for your valuable comments! Please find our detailed response below.
>**Q1:** The hyperparameters $\alpha$ and $\beta$ depend on the specific datasets, and experiments need to be conducted to determine their values for different datasets.
>
>**A1:** To avoid over-tuning par... | Rebuttal 1:
Rebuttal: We sincerely thank the ACs and reviewers for their great effort in handling our paper.
We have appropriately addressed all concerns raised by the reviewers. These include providing more comparisons with recent baselines/backbones (Reviewer #k9he, #Qvan, #uA1Ps), conducting more ablation studies o... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Learning Invariant Representations of Graph Neural Networks via Cluster Generalization | Accept (poster) | Summary: This paper aims to handle the structure shift problem in GNN. The paper proposed a novel approach -- Cluster Information Transfer (CIT) to address this challenge. CIT first computes cluster center representation during training. CIT then extract the cluster center from the node embedding to retrieve the cluste... | Rebuttal 1:
Rebuttal: 1. > W1, Q1：Learning process and theoretical explanation
First, we don't think that cluster information is harmful. The correlation between cluster and labels does help us make predictions. However, this assumption holds when the cluster information does not change, i.e., the structure shifts do... | Summary: The paper introduces a new mechanism called Cluster Information Transfer (CIT) to improve the generalization ability of GNNs to various and unknown test graphs with structure shift. The CIT mechanism enhances the diversity of nodes by combining different cluster information with the nodes, which helps GNNs lea... | Rebuttal 1:
Rebuttal: 1. >W1, Q1： Fair Novelty
We agree that the spectral clustering mechanism has been introduced previously, however, in this work, our main contribution resides in the development of the CIT mechanism and spectral clustering is just one component of it. The CIT mechanism is a flexible and succinct a... | Summary: This paper focuses on the invariant learning of graph neural networks and proposes a cluster information transfer mechanism with two statistics: the mean of cluster and the variance of cluster. The authors prove that with CIT mechanism, the model is able to capture the cluster independent information, so as to... | Rebuttal 1:
Rebuttal: 1. > Q1：The time complexity of the proposed model
Our CIT mechanism has two parts of computation: Clustering process and Cluster Information Transfer process. Let $N$ represent the number of nodes, and $K$ represent the number of clusters. The computational complexity of clustering process is $\... | Summary: The paper tackles an important question of learning invariant representations of GNNs. The authors show that once the test graph pattern shifts, the reliability of GNNs becomes compromised. Then they propose the cluster information transfer mechanism, which can be easily combined with current GNNs to improve t... | Rebuttal 1:
Rebuttal: 1. > W1： In section 2, GAT performs worse than GCN
We consider that the attention mechanism of GAT is a potential catalyst for the observed phenomenon. Existing research [1, 2] suggests that the attention mechanism is easily influenced by the distribution of neighbor characteristics. In Section 2... | Rebuttal 1:
Rebuttal: We sincerely thank all the reviewers for your careful reading and suggestions. Due to character constraints, we have uploaded the experimental results in this PDF.
Pdf: /pdf/107f70a3251daf942d5c1683475f184426e20cfe.pdf | NeurIPS_2023_submissions_huggingface | 2,023 | Summary: The paper introduces a novel Cluster Information Transfer (CIT) mechanism to enhance the generalization capability of Graph Neural Networks (GNNs) in the presence of structural shifts. The authors provide theoretical analysis, showing that the impact of cluster information on the classifier diminishes during t... | Rebuttal 1:
Rebuttal: 1. > Q1：More GNN results
Thanks for the suggestion. We evaluate another two popular GNNs (APPNP [2] and GCNII [3]), where the results are shown as follows.
| | 0.5 0.05 | 0.45 0.1 | 0.4 0.15 | 0.35 0.2 | 0.3 0.25 | 0.25 0.3 |
|-------|----------|----------|----------|----------|----------|... | null | null | null | null | null | null |
The expressive power of pooling in Graph Neural Networks | Accept (poster) | Summary: The paper analyzes the expressive power of pooling (not to be confused with readout) in message passing GNNs (MP-GNNs). The paper gives a condition (in Theorem 1) on the construction of the POOL function, under which there is a choice of the MP-GNN which can separate the same graphs that the 1-WL test can sepa... | Rebuttal 1:
Rebuttal: Thank you for the thoughtful comments and detailed review.
> Put the injectivity of pooling in the right context and setting. The construction does not show that POOL is injective for general message-passing functions with trainable parameters.
We completely agree that the objective of Theorem 1... | Summary: The authors present a comprehensive analysis of pooling operations in Graph Neural Networks (GNNs) from both theoretical and practical perspectives. Additionally, they introduce a refined dataset, EXPWL1, specifically designed to facilitate the analysis of the expressiveness of pooling operations. Remarkably, ... | Rebuttal 1:
Rebuttal: Thanks for the positive evaluation of the paper and for the useful suggestions.
> Discuss the limitations and whether the results apply to continuous features.
Thanks for the suggestion. In the conclusions, we now discuss the following limitations:
"*Firstly, the conditions of Th.1 are suffici... | Summary: This paper analyzes the expressive power of pooling operators in Graph Neural Networks (GNNs) and derives three conditions for a pooling operator to fully preserve the expressive power of the Message Passing (MP) layers preceding it. The derived conditions are as follows:
a) The pooling operator should extra... | Rebuttal 1:
Rebuttal: Thanks for your comments and suggestions.
> MUTAG, PTC-MR, IMDB-B, and IMDB-MULTI were left out.
We performed the experiments on these datasets. In addition, we also performed experiments on ENZYMES and REDDIT-5K (the latter is still running and will be completed in a few days). Table 5 with the... | Summary: This paper presents a study on the performance and expressiveness of various pooling operators in Graph Neural Networks (GNNs) in both theoretical and empirical ways. In detail, the authors identify the sufficient conditions that a pooling operator must satisfy to fully preserve the expressive power of the ori... | Rebuttal 1:
Rebuttal: Thanks for your valuable feedback. We believe to have addressed all the reviewer's concerns and requests, as discussed in the following.
> Provide examples of EXPWL1.
We inserted in the supplementary material figures showing pairs of WL-1 distinguishable graphs from the EXPWL1 dataset. A few exa... | Rebuttal 1:
Rebuttal: We thank the reviewers for recognizing the value of our work and for the useful feedback, which allowed to improve our paper.
Summary of changes:
- We addressed all the reviewers' requests by introducing explanations and several modifications in the text and the figures.
- We added an experiment... | NeurIPS_2023_submissions_huggingface | 2,023 | Summary: Studies on the expressive power of Graph Neural Networks (GNNs) have garnered extensive attention. However, these studies have been limited to flat GNNs. Some hierarchical pooling methods, such as diff-pool, have been proposed. Evaluating the ability of pooling operations directly is challenging, so the perfor... | Rebuttal 1:
Rebuttal: We thank the reviewer for the feedback. There are important misunderstandings that, hopefully, are clarified in the following.
> 1. The rationale of hierarchical pooling vs. simpler global pooling.
The rationale for using hierarchical pooling rather than global pooling is not to improve the expr... | null | null | null | null | null | null |
Shared Adversarial Unlearning: Backdoor Mitigation by Unlearning Shared Adversarial Examples | Accept (poster) | Summary: This paper proposes a Shared Adversarial Unlearning (SAU) method for mitigating backdoor attacks. The motivation behind this method is the recognition that not all adversarial samples are effective for backdoor defense. Therefore, it is important to identify the adversarial samples that truly contribute to bac... | Rebuttal 1:
Rebuttal: **Q1. For the combination of ABL**
**A1:** Thanks. We would like to point out that ABL is an in-processing method that can access the whole training dataset for poisoned sample detection and unlearns the suspicious backdoor samples, while our method SAU is a post-processing method and can only us... | Summary: This paper proposes a method to defend against backdoor attacks in deep neural networks through adversarial training techniques. By exploring the connection between adversarial examples and poisoned samples, the authors propose an upper bound for backdoor risk and a bi-level formulation for mitigating backdoor... | Rebuttal 1:
Rebuttal: **Q1.Suggestion of comparison to provable/certified backdoor defense methods [1,2,3].**
**R1:** We acknowledge the importance of certified robustness against backdoor attacks. In backdoor defense, methods can be categorized into two classes: **empirical methods which aim to develop effective and... | Summary: This paper analysed the relationship between adversarial examples and poisoned examples.
Then, this paper proposed a fine-tuning strategy to purify the poisoned model.
Strengths: 1 This paper is easy to follow.
2 This paper provides some experiments that support the proposed method.
3 This paper has som... | Rebuttal 1:
Rebuttal: **Q1. For the concern about terminology**
**R1.** We appreciate your interest in our work and the terminology we use. We would like to clarify the following points:
* As backdoor learning is an emerging area, we follow the typical paradigm of scientific research like [4,5] to first formulate the... | Summary: summary: The paper proposes a mitigation approach for backdoor attacks in ML (attacks where a model predicts target classes for poisoned samples when triggered by the adversary). The authors analyze the relationship between backdoor risk and adversarial risk [adversarial examples and poisoned samples] to creat... | Rebuttal 1:
Rebuttal: **Q1. Suggestion for paper layout**
**R1:** Thanks for your constructive suggestion. We will update the layout in the revised manuscript by moving the suggested contents and other important contents from the supplementary material to the main manuscript, to make it more self-contained and legible... | Rebuttal 1:
Rebuttal: # Common Response
**Q1. Concerns about the perturbation set $\mathcal{S}$ in Assumption 1 that restricts the trigger magnitude, including its practicality, and its generalization to backdoor attack with excessive-magnitude trigger.**
**A1:** Thanks for the insightful comments. The concerns are ... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
The Crucial Role of Normalization in Sharpness-Aware Minimization | Accept (poster) | Summary: This work investigates the role of the normalization factor $1/||\nabla L(x,w)||$ in the perturbation step of the Sharpness-Aware Minimization (SAM) algorithm. To this end, the authors compare the behavior of gradient descent, SAM and a version of SAM without the normalization (USAM) via experiments and extens... | Rebuttal 1:
Rebuttal: Thank you for your insightful comments! Here are our responses:
---
### Why Compare SAM to USAM?
We agree that USAM is by no means practical. However, the reason for comparison is solely to understand SAM better and doesn't have to do with USAM's impracticality. In particular, we are trying to *... | Summary: The paper investigates the role of the normalization term in the recently proposed optimization algorithm Sharpness-Aware Minimization (SAM). The authors theoretically, and empirically, study the differences among SAM, its un-normalized counterpart USAM, and gradient descent, under multiple settings (strongly... | Rebuttal 1:
Rebuttal: Thank you for your insightful comments! Here are our responses:
---
### Is the Main Takeaway "Using SAM instead of USAM"?
The takeaway is not this one if interpreted literally, but slightly more nuanced. Let us summarize the **two main takeaways** of our paper, one for the theoretical ML commun... | Summary: - The paper investigates the difference between SAM and Unclipped SAM (USAM) in terms of stabilizing and drifting along the manifold.
- The paper reveals a significant disadvantage in USAM compared to SAM, which was previously thought to be similar in various papers.
- The paper provides mathematical and exper... | Rebuttal 1:
Rebuttal: Thank you for your insightful comments! Here are our responses:
---
### Should the Choice of $\rho$ Be Different in SAM and USAM?
Yes, the optimal tuning of algorithmic parameters $(\eta,\rho)$ likely differs between SAM and USAM. However, we would like to emphasize that **we are not comparing ... | Summary: **I have read the rebuttal and raised the score to a 7.** This paper studies the importance of the normalization step in SAM (normalizing the gradient used to perturb the weights) for stability and effectiveness at finding a flat region in a manifold of minima (drifting along a continuum of minima). They provi... | Rebuttal 1:
Rebuttal: Thank you for your insightful comments! Here are our responses:
---
### USAM is Not Practical
We agree that USAM is by no means practical. However, the reason for comparison is solely to understand SAM better and doesn't have to do with USAM's impracticality. In particular, we are trying to **u... | Rebuttal 1:
Rebuttal: Dear Reviewers,
We thank all the reviewers for the valuable feedback. Specifically, thank you for appreciating that our motivation is "**clear and interesting**" (Reviewers pAM7, chBL, and a57m), our paper is "**well-written, easy-to-read, and thorough**" (Reviewers ymqN, pAM7, and a57m), the th... | NeurIPS_2023_submissions_huggingface | 2,023 | Summary: The authors argue that the normalization of the gradient in the SAM model during the ascent step is beneficial. In particular, classic SAM is shown theoretically and empirically in certain scenarios to not-diverge and to be able to drift along the minima manifold. This implies that SAM is more robust in the ch... | Rebuttal 1:
Rebuttal: Thank you for your insightful comments! Here are our responses:
---
### Comparison Between Our Paper and Wen et al. (2023); Compagnoni et al. (2023)
Thanks for mentioning these two seminal papers on SAM! We will definitely discuss more in our revision. Here is a draft of our discussion:
1. Wen... | null | null | null | null | null | null |
Towards Personalized Federated Learning via Heterogeneous Model Reassembly | Accept (poster) | Summary: The paper proposes to use the recently published model reassembly technique (NeurIPS 2022) to obtain personalized models through federated learning. At each round, the server collects the current models from the clients and uses reassembly to generate new candidate models, potentially training some stitching l... | Rebuttal 1:
Rebuttal: Thanks for the reviewer's valuable comments. Hope our responses below adequately address your concerns.
`>>> W1` ***Historical information maintaince***
`>>> WA1`
(1) Candidate models are reassembled using the layers from the models of active clients. Our designed optimization approach enfor... | Summary: The authors designed a method to train a model with FL when clients have heterogeneous model architectures. They designed a model reassembly technique that stitches together parts of DNNs. pFedHR also creates personalised models for each client without requiring server-side data or explicit human guidance.
S... | Rebuttal 1:
Rebuttal: We truly thank you for the insightful comments and suggestions. We hope our responses can address your concerns.
`>>> W1` ***Motivation of applying stitching technique***
`>>> WA1`
(1) Thanks for the suggestions. Our approach is motivated by the challenge that clients may have different model... | Summary: In this paper, the authors introduce a technique that tackles the challenge of enabling collaboration among client models with different network structures in federated learning. Unlike traditional knowledge distillation (KD)-based approaches, the proposed model involves dividing the heterogeneous models into ... | Rebuttal 1:
Rebuttal: Thanks for the reviewer's valuable comments.
`>>> W1` ***Utilization of other models***
`>>> WA1`
Thanks for your suggestion. We conduct experiments using MobileNetV1, MobileNetV2, and MobileNetV3 as our client models. For the skipping connection, we treat the block as a whole without assembl... | Summary: The authors design a novel approach to make it possible for clients equipped with different model structures to cooperate in the federated learning framework. Specifically, the server will reassemble models into different parts and assemble them together. After that, they propose a similarity-based approach to... | Rebuttal 1:
Rebuttal: Thanks for the reviewer's comments. The responses to the weaknesses are shown as below.
`>>> W1` ***Stitching layer training***
`>>> WA1`
I do appreciate your suggestion. After we stitch the models with stitching layers, we freeze the parameters of the selected layers but only train the param... | Rebuttal 1:
Rebuttal: We truly appreciate the insightful questions. We provide responses to Q1 from Reviewer xdZD. We hope it could help adequately address your concerns.
In the pdf file, there is a figure for the generated model under the homogeneous setting. Thanks again for your question.
Pdf: /pdf/b1a36c1e16dfde7d... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Active Learning for Semantic Segmentation with Multi-class Label Query | Accept (poster) | Summary: The paper studied active learning for semantic segmentation. Instead of using each pixel as a query for annotation, this paper used a super-pixel as a query. Since each super-pixel may contain multi-class labels, thus a multi-class labeling scheme is used instead of dominant class labeling. However, it introdu... | Rebuttal 1:
Rebuttal: We appreciate your insightful feedback and valuable suggestions that helped improve our paper substantially. All the suggestions and experiments will be incorporated into the paper. Please find our responses to the comments below.
___
**Q1. Effect of hyperparameters**
A1. Thank you for the detail... | Summary: This manuscript proposes a new issue on active segmentation from multi-class label query. Concretely, the authors are motivated from the shortcomings of existing active learning paradigm, i.e., [1][2] query the whole image pixel annotations, which costs annotation; [3] selects pixels for annotation, which perf... | Rebuttal 1:
Rebuttal: We appreciate your insightful feedback and constructive suggestions, which help improve our paper substantially. We will address all the comments and include additional experiments in the revision. Please find our responses to the comments below.
___
**Q1. Clarifying Line 28 in our manuscript**
A... | Summary: This work introduces a new active learning framework for semantic segmentation, involving a novel query design that uses a multi-hot vector for class representation with in a region, two novel loss functions for effective multi-class label supervision, and an acquisition function considering class uncertainty ... | Rebuttal 1:
Rebuttal: We sincerely appreciate your insightful feedback and suggestions that helped improve our paper substantially. All the suggestions will be incorporated into the main text and the appendix. Please find our detailed responses to the comments below.
___
**Q1. Clarification of our contributions from a ... | Summary: This paper proposes an active learning approach for semantic segmentation using multi-class label queries. Different from dominant class labeling methods, where an oracle is asked to select the most dominant class by a single click, this paper instead designs a multi-class labeling approach that asks the oracl... | Rebuttal 1:
Rebuttal: Thank you for your insightful feedback and constructive comments that help improve our paper substantially. We will incorporate all the suggestions into the paper. Please find our detailed responses to the comments below.
___
**Q1. Insights to determine the importance of each loss term**
A1. We a... | Rebuttal 1:
Rebuttal: We sincerely thank all the reviewers for their constructive comments. All reviewers appreciated the novelty of our multi-class labeling, comprehensive experiments, and the clarity of the paper. In addition, they recognized that our training method is both reasonable and sound (DZFu), well suited t... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Time-Reversed Dissipation Induces Duality Between Minimizing Gradient Norm and Function Value | Accept (poster) | Summary: This paper introduces novel notion of H-duality, which shows unexpected connection between time-reversed coefficients of methods and optimality w.r.t. gradient norm or function value.
Strengths: Very inspiring work, introduced notion is quite curious. I think that such an idea will be interesting to most of r... | Rebuttal 1:
Rebuttal: We're grateful for your thorough evaluation and the perceptive observations you've shared. These will certainly assist us in enhancing our paper even more. I'd like to address a few matters in response to your questions.
**Q1) Clarification on the phrase "efficiently reduce function values/gradi... | Summary: This work defines H-duality, which is a one-to-one correspondence between methods that minimize function values and methods that minimize gradient magnitudes, under the assumption that the objective function is convex and $L$-smooth. It is proved in both discrete and continuous time dynamics that, when one met... | Rebuttal 1:
Rebuttal: We appreciate the thoughtful comments and suggestions you've offered regarding our paper.
**W1) Assumption of $L$-smooth convex function**
Please refer to the response for W1) of reviewer 2PtR.
**W2, Q1) What about the necessity of $N$ in the H-duality?**
Please refer to the response for Q2) ... | Summary: This paper studies the duality for optimization algorithms. As claimed by the authors, the notion they present is distinct form any previously known duality or symmetry relations. In this work, the authors present a one-to-one correspondence between methods efficiently minimizing function values and methods ef... | Rebuttal 1:
Rebuttal: We value the considerate remarks and recommendations you've provided concerning our paper.
**W1) Regarding the relativity of NeurIPS and this topic**
Indeed, the primary value of our work lies in its theoretical novelty. In our view, H-duality is an unexpected discovery that may enrich our under... | Summary: This paper shows an intriguing symmetry in convex optimization between first-order methods that efficiently minimizes the function value and those that efficiently minimizes the squared gradient norm. In particular, they show that if a specific Lyapunov function upper bounds the function value suboptimality fo... | Rebuttal 1:
Rebuttal: Thank you for your detailed review and your insightful comments. They certainly will aid us in refining our paper further. I would like to clarify a few points based on your queries.
**W1) Why does the H-duality theorem intuitively hold?**
Let us explain the motivation of the main theorem and ... | Rebuttal 1:
Rebuttal:
# Common Response
We would like to express our sincere appreciation to the reviewers for their time and constructive criticism.
Several reviewers (Reviewers 8tCz, Fq7n, and x8JQ) strongly appreciated the theoretical novelty and value of our paper, acknowledging the innovative aspects of H-duali... | NeurIPS_2023_submissions_huggingface | 2,023 | Summary: This paper presents the concept of H-duality, which refers to the duality between optimization algorithms. The notion involves two sets of algorithms: 1) those that efficiently reduce the function value, and 2) those that efficiently reduce the gradient magnitude. The authors create the H matrix, which contain... | Rebuttal 1:
Rebuttal: We appreciate your thoughtful comments and suggestions regarding our paper.
First and foremost, we would like to emphasize that the primary focus of our paper lies in its theoretical contributions rather than constructing a faster method. However, we firmly believe that H-duality opens up the p... | null | null | null | null | null | null |
Optimizing over trained GNNs via symmetry breaking | Accept (poster) | Summary: To overcome the symmetry issue when solving inverse problems on trained GNNs, this paper proposes two types of symmetry-breaking constraints to break symmetry. The authors construct an indexing algorithm, and prove that the resulting graph indexing satisfies the proposed symmetry-breaking constraints. They al... | Rebuttal 1:
Rebuttal: Sincerely appreciate these comments and questions.
**Weakness 1(1) [Experimental verification of molecular constraints]**\
Section 3.1 is not a contribution of our paper but a description of how we followed the literature. This section just provides background for our numerical experiments.
We ... | Summary: This paper investigates the optimization of trained GNNs. This is a permutation-invariant problem and all points on an orbit of the symmetric group have the same performance. The authors proposed some symmetry-breaking approaches to have smaller search space and hence less redundancy and more efficient optimiz... | Rebuttal 1:
Rebuttal: Thank you for these considerations about the relationship and difference between our work and the literature.
**Weakness [Literature review]**\
Sincere thanks for this useful comment on the literature review: indeed we will add more details in the final paper. MIP solvers typically detect symmetr... | Summary: This paper suggests using a symmetry breaking indexing for nodes in a graph. The argument is that for inverse problems, such as molecule design, graph isomorphism results in finding many equivalent solutions. The symmetry breaking indices are supposed to partially alleviate this redundancy. They use pre-traine... | Rebuttal 1:
Rebuttal: Many thanks for these helpful comments and suggestions.
**Weakness [How many symmetric solutions are removed]**\
We acknowledge that the limited information in Section 3.2 and Table 1 is insufficient for our purposes. Table 1 is not used to show that MIP can find *many* feasible solutions under d... | Summary: This paper presents novel symmetry-breaking constraints for optimizing trained graph neural networks (GNNs) and addresses the graph isomorphism issue. The authors develop two mixed-integer optimization formulations and evaluate their methods in the context of molecular design.
Strengths: 1. **Relevance and A... | Rebuttal 1:
Rebuttal: Sincere thanks for these comments which are fair and help us clarify and improve our work.
**Weakness 1 \& Question 1 [Compatibility of S1 and S3]**\
Thanks for mentioning this. Our original paper ignored the compatibility of S1 and S3, but the reviewer is correct that we should add this to the p... | Rebuttal 1:
Rebuttal: To all reviewers,
Thank you very much for taking your precious time to review our paper and give great comments! Your feedback is really helpful for us to clarify and improve our work. More importantly, we are very lucky to have four reviewers focus on different aspects of our paper:
- Reviewer ... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
An Alternating Optimization Method for Bilevel Problems under the Polyak-Łojasiewicz Condition | Accept (poster) | Summary: This paper proposed a generalized alternating method for nonconvex bilevel optimization with PL condition in lower level. Meanwhile, it provided convergence analysis for the proposed method under a new stationary metric. Some experimental results demonstrate efficiency of the proposed method.
Strengths: This ... | Rebuttal 1:
Rebuttal: We thank the reviewer for the careful reviews. However, we feel the reviewer might have misunderstood some technical points in our paper which leads to the concern on the correctness of the conclusion in the paper. Our response to your comments follows.
**Q1: Does $\lambda_g>0$ implies positive ... | Summary: This work studies an important problem of bilevel optimization with non-convex lower level function. The authors reformulate the problem as a constraint optimization and prove a constraint qualification (CQ) for the reformulated problem under mild assumptions, and show that other stronger CQ do not hold under ... | Rebuttal 1:
Rebuttal: We thank the reviewer for appreciating our contribution and the constructive review. Our response to your comments follows.
**Q1: Rigorous definition of calmness.**
Thank you for your suggestions and we will formalize the definition of calmness. To answer your specific confusion, $h(x,y)$ can ... | Summary: The paper studies unconstrained bilevel optimization under the PL condition at the lower level. All previous work is limited to strongly-convex lower-level problems. The authors propose a new convergence criterion for this setting based on stationary-point seeking reformulation, supporting it with the calmness... | Rebuttal 1:
Rebuttal: We thank the reviewer for appreciating our presentation and recognizing our work as "a timely contribution to the literature." Our response to your comments follows.
**Q1: Concerns on the algorithmic idea.**
Thank you for recognizing the necessity to extend bilevel optimization (BLO) to the PL... | Summary: This paper addresses the bilevel optimization problem, which may feature a non-convex lower-level problem with non-unique optimal function values. The lower-level problem can be transformed into a constraint optimization problem using function value or gradient-based constraints. The paper establishes necessar... | Rebuttal 1:
Rebuttal: We thank the reviewer for appreciating our contribution and the constructive review. Our response to your comments follows.
**Q1: Possibility of extending the proposed stationary condition to the general non-PL nonconvex setting.**
Thanks for raising this intriguing point! Currently, generalizi... | Rebuttal 1:
Rebuttal: ## General Response ##
We sincerely thank the reviewers for their constructive comments. All reviewers have agreed that our work has substantially relaxed the strongly convex lower-level assumptions in bilevel optimization and proved the optimal complexity in terms of $\epsilon$. Comments from al... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Scalable Fair Influence Maximization | Accept (poster) | Summary: The paper proposes an algorithm for fair influence maximization. The objective is defined as a sum of powers of a utility function for each cluster, indicating that each cluster should receive some influence. Reducing the exponent will bring the objective to a fairer distribution of influence. An algorithm bas... | Rebuttal 1:
Rebuttal: ### **Response to Reviewer uZjd**
Thank you for your insightful comments and suggestions. We reply to all your comments below.
> **Weakness 1**: Baselines: While the theoretical guarantee is the key contribution, the experiment part of the paper can be strengthened. How much do we gain from this... | Summary: The authors apply fairness to the influence maximization problem (IM). On a social network, IM models which node subset should be used to trigger the spread of information to maximize its effects. A motivating application is the selection of leaders for natural disaster preparedness, where in existing methods,... | Rebuttal 1:
Rebuttal: ### **Response to Reviewer RYE8**
Thank you for your insightful comments and suggestions. We reply to all your comments below.
> **Weakness 1**: While I think the paper is reasonably clear, it can be quite dense at times with its heavy use of notation. I found it particularly hard to follow in S... | Summary: The authors study a variant of the popular influence maximization problem
that incorporates fairness constraints. Given a partition of the node set
of the graph into communities, the goal is to select seeds subject to
budget constraints that maximizes influence as well as reduces the
influence gap between comm... | Rebuttal 1:
Rebuttal: ### **Response to Reviewer ScbK**
Thank you for your thorough and insightful comments as well as acknowledging our theoretical analysis. We reply to all the points below.
>**Weakness 1**: The importance of/need for the unbiased estimator is not discussed. Firstly, in Section 3.1, it is pointed o... | Summary: This paper aims to study the problem of fairness-aware influence maximization over a community structure under a welfare fairness notion that balances fairness level and influence spread using an exponentially-weighted sum over communities objective with an fractional exponent parameter α expressing inequality... | Rebuttal 1:
Rebuttal: ### **Response to Reviewer tPHQ**
We thank the reviewer for the insightful comments and suggestions. We reply to all your comments below.
>**Weakness 1**: Limited novelty with respect to previous work in [7].
**A1**: Literature [7] proposes a novel notion of fairness for influence maximization ... | null | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Curriculum Learning With Infant Egocentric Videos | Accept (spotlight) | Summary: This paper studies the development of visual intelligence by providing a curriculum for self-supervised video representation learning. The curriculum is aligned with the age order of the infants whose egocentric video data are used as training samples. Experiment results demonstrate that training with data fro... | Rebuttal 1:
Rebuttal: We thank the Reviewer for detailed comments and for highlighting the values of this line of work. Below we addressed each weakness and question mentioned by the Reviewer.
**Reviewer** pointed out the lack of a quantitative measurement for “slowness” of the video, as well as corresponding experime... | Summary: The authors aim to explore whether the natural structures and regularities in infant visual experience acquired via video recording when infants wearing head-mounted cameras can facilitate pre-training self-supervised learning representation learners. The authors divide the infant data into 3 groups (according... | Rebuttal 1:
Rebuttal: We thank the Reviewer for detailed comments and for pointing out a number of strengths. Below we addressed each question mentioned by the Reviewer.
**Reviewer** asked whether the ratio of the two “number of components” could potentially be used to validate the effectiveness of the learning models... | Summary: The work is essentially about an experimental evaluation to show the benefit of curriculum learning using ego-centric videos acquired with from head-mounted cameras and very young children. The authors discuss the use of self-supervised learning on the data ordered according to developmental principles and emp... | Rebuttal 1:
Rebuttal: We thank the Reviewer for detailed comments and for the enthusiasm about our approach. Below are our responses to the Reviewer’s questions.
**Reviewer** asked for a more convincing justification on why we used large data sets and complex architectures.
**Our response:** While we did use advanc... | Summary: This paper aims to study how developmental changes impact visual learning. It used infant egocentric videos to pre-train video autoencoders with self-supervised learning loss. It separated the infant data by age group and evaluated the importance of training with a curriculum aligned with developmental order.
... | Rebuttal 1:
Rebuttal: We thank the Reviewer for the comments and appreciate the Reviewer's enthusiasm about the goals of the paper. Below we addressed each weakness mentioned by the Reviewer.
**Reviewer:** The contribution of this paper is limited.
**Our response:** We believe that the manuscript delivers several su... | Rebuttal 1:
Rebuttal: Our paper tackles a question at the heart of human and machine learning: How do we compare the learning abilities of human infants and machines? Our study takes an important step in this direction by (1) showing that generic learning algorithms (with no hardcoded knowledge about objects or space) ... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Comparing Apples to Oranges: Learning Similarity Functions for Data Produced by Different Distributions | Accept (poster) | Summary: This work considers PAC-learnable algorithms for estimating a intra-group
similarity function where:
1. intra-group similarity functions are metrics and given in advance,
2. inter-group similarities satisfy a notion of the triangle inequality.
It is assumed that points are always labeled by the group to which ... | Rebuttal 1:
Rebuttal: Thank you for your review. We will certainly take your feedback into consideration.
We first address the weaknesses that you mentioned:
1. Even though there is a description of the simple algorithm in lines 225-228, we will try to make it more explicit.
2. We will try to work on the overal... | Summary: The paper is about a mechanism to compare elements of a set that belong to two distinct distributions when two distinct sets of features for each distribution. This is a natural problem that arises in many practical settings. The paper presents the underlying problem and the formal framework with some guar... | Rebuttal 1:
Rebuttal: Thank you so much for your positive comments. We are glad to see our work getting appreciated.
Regarding the weakness you mentioned. This is indeed a great question and we can certainly investigated it even more. Our experimental results show that $\max (p_\ell(\epsilon, \delta), p_{\ell'}(\epsil... | Summary: This paper studies the problem of learning a similarity function of items generated from different distributions. Precisely, there are two groups of items, and the items in each group follow a specific distribution and the group has an intra-cluster metric. This paper wants to learn a cross-group metric from l... | Rebuttal 1:
Rebuttal: We sincerely thank the reviewer for their very positive feedback. We are glad to see that our work gets appreciated.
With regards to your question, we hypothesize that in practice, the vast majority of application the similarity scores will not be forming a metric. Take for instance our second mo... | Summary: This paper involves finding an inter-group similarity function with (limited) oracle advice when one has access to the intra-group metrics. The authors give theoretical guarantees about their algorithm and claim to have proven lower bounds on the performance of any algorithm.
Strengths: I like this problem. T... | Rebuttal 1:
Rebuttal: We would like to thank the reviewer for their thoughtful comments.
First we address the mentioned weaknesses of the paper.
1. Even though are algorithm are relatively simple, their analysis and the analysis of the lower bounds are highly non-trivial.
2. Here we address the concerns regarding th... | Rebuttal 1:
Rebuttal: We thank all of our reviewers for their time and their thoughtful comments. We reply to each one individual below. | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Temporal Causal Mediation through a Point Process: Direct and Indirect Effects of Healthcare Interventions | Accept (poster) | Summary: This paper studies how to estimate the direct and indirect effctes of healthcare interventions. The general idea of this paper is to model the mediation process and outcome process jointly. More specifically, it considers the mediation process as a temporal point process conditioned on the past mediation, outc... | Rebuttal 1:
Rebuttal: We are glad that you find our problem setting and approach both interesting and relevant. Thank you for highlighting the clarity of our motivation and the design of the causal model and analysis!
**Regarding treatment–outcome correlation and potential bias:**
In our real-world case-study, as disc... | Summary: The paper aims to estimate the direct and indirect treatment effects of healthcare interventions. The authors model the mediator as a point process and propose a non-parametric mediator–outcome model where the mediator is assumed to be a temporal point process that interacts with the outcome process. The autho... | Rebuttal 1:
Rebuttal: Thank you for your feedback and questions, which we hope to address in the following and with which we have improved the manuscript:
**1. Confounding bias of treatment $A$ on mediator $M$ and outcome $Y$:**
Please note that the estimation of $E[Y|A,M]$ is not confounded by $A$, because $A$ is inc... | Summary: The paper defines direct and indirect effects in complex healthcare time-series as dynamic stochastic processes and theoretically provides causal assumptions for identifiability. This model allows for an external intervention influencing both mediator and outcome sequences simultaneously and captures time-dela... | Rebuttal 1:
Rebuttal: Thank you for appreciating the articulation of our proposed method and for your valuable questions. We hope we can further clarify and address your concerns in the following:
**Regarding scalability:**
There appears to be a misunderstanding, as the works referred to [Chen+18, Luo+20, Zhang+21] ar... | Summary: The paper's outstanding qualities lie in its well-articulated presentation and its precise experimental design. It amalgamates the earlier research findings of [Zeng et al., 2021] and [Hızlı et al., 2022] with the innovative notions put forth by [Robins et al., 2022] on indirect effects. The authors tackle pra... | Rebuttal 1:
Rebuttal: Thank you for your thorough feedback! In the following we hope to address your concerns regarding originality and novelty:
**Dynamic causal mediation with point process mediator:**
We emphasize that the whole formulation of the problem with point process mediator and outcome, as well as defining ... | Rebuttal 1:
Rebuttal: We thank all reviewers for their time and effort in writing thoughtful and valuable reviews. Overall, the reviews appear positive about our submission. We are happy for the explicit commendations of our realistic and relevant problem setup, robust modeling approach, and clarity of our presentation... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Interpreting Unsupervised Anomaly Detection in Security via Rule Extraction | Accept (poster) | Summary: This paper proposes a rule-set extraction method for a black-box anomaly detection model that is trained with only normal or unlabeled data.
There are many methods to interpret the learned model, but, this paper claims that there are few methods designed for anomaly detection.
The proposed method first extract... | Rebuttal 1:
Rebuttal: We thank Reviewer jw73 for the detailed and constructive review. We will address all the expository comments in the final version of the paper.
**Major clarification.**
>Is it better to present it as a more generic method since there are other areas of industry where rule sets are required?
Th... | Summary: The paper deals with the problem of "explainable" unsupervised machine learning (ML) for anomaly detection (AD), with a focus of network intrusion detection (NID). The paper argues that while abundant effort focused on providing _supervised_ ML techniques that are explainable, this is not the case for _unsuper... | Rebuttal 1:
Rebuttal: We thank Reviewer Abmj for the detailed and constructive review. We are very glad to hear so much constructive advice from a security expert like you. Many of the references are very helpful for NIDS research, which we would like to cite in our paper.
>Given that CIC-IDS17 is flawed, how much ti... | Summary: The authors propose and Distribution Decomposition Rules and Boundary
Inference Rules to make black boxes more interpretable.
They use Interior Clustering Tree IC-tree to find distribution
decomposition rules. The IC-tree algorithm splits the data on a
feature value at each node recursively. The feature val... | Rebuttal 1:
Rebuttal: We thank Reviewer eS8h for the detailed and constructive review, especially thanks for your recognition of our work. We will address all the expository comments in the final version of the paper.
>line 220: What is the reasoning for the initial vs auxiliary explorers?
The initial explorers set t... | Summary: The paper presents a technique to build a decision tree (DT) that uses the predictions from any unsupervised anomaly detection algorithm to split at the nodes when constructing the DT. This is named as the Interior Clustering Tree. The DT will be used to extract interpretable rules.
Strengths: 1. The approach... | Rebuttal 1:
Rebuttal: We thank Reviewer mhUf for the detailed and constructive review. We will address all the expository comments in the final version of the paper.
> Very similar to [1].
Thanks for your comment. We have thoroughly read the provided paper, which proposes the one-class decision tree (OC-Tree). Though... | Rebuttal 1:
Rebuttal: We thank the reviewers for their careful reading and detailed and considerate feedback.
We are glad that reviewers agree that our paper tackles an important problem (“an important problem domain” by mhUf, “tackles an open research problem” by AbmJ, “problem tackled in this paper is important” by ... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Counting Distinct Elements Under Person-Level Differential Privacy | Accept (poster) | Summary: Suppose a data set consists of (user, item) pairs. The paper provides an estimate of the count of the number of distinct items that satisfies user level differential privacy. This differs from existing work in two ways. Existing work 1) considers item streams with item, not user, level differential privacy and... | Rebuttal 1:
Rebuttal: We would like to thank the reviewer for their thoughtful comments.
**W1.** We agree that the title could be misleading, we will rename the next revision to “Counting Unique Elements Under Person-level Differential Privacy”.
**W3 & W3.** Thank you very much for pointing out the issues in our re... | Summary: The authors study the fundamental problem of counting the number of distinct elements in a dataset
in a user-level DP setting, where a user can contribute an unbounded number of items. The main
contribution of the paper is an approach to obtain lower bounds through a bounded sensitivity
count and a bias-varian... | Rebuttal 1:
Rebuttal: We would like to thank you for reviewing the paper.
**Q1. Why $M_{\ell,\varepsilon}(D) - \frac{\ell}{\varepsilon}\log(1/2\beta)$?**
Under user-level DP, it is impossible to compute the exact number of distinct elements since a single user could contribute an arbitrary number of unique elements ... | Summary: Edit: Overall decision recommendation changed from borderline reject to borderline accept under the expectation that limitations and future work directions as discussed in the reviewer discussion are incorporated into the manuscript.
The manuscript studies the problem of approximating the number of unique ite... | Rebuttal 1:
Rebuttal: We would like to thank the reviewer for their thoughtful comments. We address the questions and weaknesses listed.
**Q1a. -- bias of estimator** Under the assumption that a bound $\ell$ on the number of distinct elements per person is known, it is possible to estimate $DC(D)$ in an unbiased manne... | Summary:
This paper proposes an algorithm to count the maximum number of unique items one can get from considering from every individual in a dataset at most l items, while preserving differential privacy.
Strengths:
The paper offers an (as far as I know) original contribution, which is simple and nice. The exp... | Rebuttal 1:
Rebuttal: We would like to thank you for reviewing the paper, we respond below.
**Applications:** We believe that applications of our algorithms are abundant. COUNT(DISTINCT …) is a very common SQL aggregation that is missing in all existing differentially private SQL solutions. [A concrete example](https:... | Rebuttal 1:
Rebuttal: Plots that we requested by reviewers are in the attached file.
Pdf: /pdf/bf196beedbc71e9f9993b5c3d9fc9c37cc4d5f20.pdf | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Optimistic Natural Policy Gradient: a Simple Efficient Policy Optimization Framework for Online RL | Accept (spotlight) | Summary: This paper proposes a simple and efficient policy optimization method with rigorous theoretical guarantee. The algorithm combines the natural gradient ascent and the optimistic off policy evaluation and is computationally efficient. They provide the condition under which the algorithm is guaranteed to be effic... | Rebuttal 1:
Rebuttal: We thank the reviewer for the valuable and positive feedback.
Please see our response below.
**Q1.** I understand that you only need to run the algorithm one for $\log(1/\delta)$ times, but how do you estimate the values of every output policy to accuracy $\epsilon$? Can you show exactly how you... | Summary: The paper presents $\texttt{OPTIMISTIC NPG}$, a new algorithm that combines the natural policy gradient with optimistic policy evaluation to encourage exploration in online RL. The algorithm demonstrates computational efficiency and achieves optimal sample complexity in linear MDPs, outperforming existing stat... | Rebuttal 1:
Rebuttal: We thank the reviewer for the valuable and positive feedback.
**Q1.** The technical novelty of the algorithm for a general function class appears unclear, and further explanation is needed on how to construct confidence sets when working with function classes beyond linear functions.\
**A1.**
... | Summary: This work proposes a simple-efficient efficient policy optimization algorithm for online RL. When specialized to linear MDPs, the algorithm improves the best known result by a factor of $d$. Moreover, this is the first efficient policy optimization algorithm under general function approximation.
Strengths: 1 ... | Rebuttal 1:
Rebuttal: We thank the reviewer for the valuable and positive feedback.
Please see our response below.
**Q1.** From what I know, [Ruosong Wang et al.] claims they developed a provably efficient (both computationally and statiscally) Q-learning algorithm. For policy optimization under general function app... | Summary: This paper proposes optimistic NPG for online RL, which is computationally efficient and enjoys the polynomial sample complexity while learning near-optimal policies. The paper is well-written and organization is clear.
Strengths: There are several strengths:
1. Sample complexity benefits: compared with other... | Rebuttal 1:
Rebuttal: We thank the reviewer for the valuable and positive feedback. | null | NeurIPS_2023_submissions_huggingface | 2,023 | Summary: This paper presents a model-free policy optimization algorithm Optimistic Natural Policy Gradient for online and episodic MDPs. The authors present sample complexity results which are better than existing results for linear MDPs. This makes it computationally efficient and optimal dimension dependence, first o... | Rebuttal 1:
Rebuttal: We thank the reviewer for the valuable and positive feedback. Please see our response below.
**Q.** Zanette et al., (2020) were able to improve the sample complexity of FQI-style algorithm by paying the price in computation. What is that the proposed algorithm compensates on, if anything, to ach... | null | null | null | null | null | null |
Statistical Limits of Adaptive Linear Models: Low-Dimensional Estimation and Inference | Accept (poster) | Summary: The paper studies the statistical limits of some adaptive linear models.
The paper defines the notion of $(k,d)$-adaptivity (Definition 2.1), and then it proves, under some conditions and for a $(k,d)$-adaptive model and failure probability $\delta$, that
- (Theorem 3.1) the estimation error of the adaptive c... | Rebuttal 1:
Rebuttal: We would like to thank the reviewer for taking the time to review our paper and provide helpful feedback! We truly appreciate your comments and suggestions, and believe they can make our work better.
In the following, we hope to address each of the points made in the review, in the order as they ... | Summary: In "Statistical Limits of Adaptive Linear Models: Low-Dimensional Estimation and Inference" the authors consider the problem of estimating a low-dimensional signal in a high-dimensional linear model where data collection is allowed to be adaptive. The notion of adaptivity employed in this paper restricts itsel... | Rebuttal 1:
Rebuttal: We would like to thank the reviewer for taking the time to review our paper and provide helpful feedback! We truly appreciate your comments and suggestions, and believe they can make our work better.
In the following we hope to address each of the points made in the review, in the order as they c... | Summary: This paper considers the issue of adaptive data collection in a linear regression model. To summarize the main idea, let us focus on the leading example in the paper (that is, Example 2.1, treatment assignment). In this example, a patient is treated based on effectiveness of the previous treatments as well as ... | Rebuttal 1:
Rebuttal: We would like to thank the reviewer for taking the time to review our paper and provide helpful feedback! We truly appreciate your comments and suggestions, and believe they can make our work better.
In the following we hope to address each of the points made in the review, in the order as they c... | Summary: The paper introduces a new data collection assumption that captures the partially adaptive data and then derives a bound for scaled MSE of order $k\log n$, where $k$ is the number of entries that are collected adaptively. Finally they also introduce a novel estimator for single coordinate inference which has ... | Rebuttal 1:
Rebuttal: We would like to thank the reviewer for taking the time to review our paper and provide helpful feedback! We truly appreciate your comments and suggestions and believe they can make our work better.
In the following, we hope to address each of the points made in the review, in the order they came... | Rebuttal 1:
Rebuttal: ## Additional simulation:
As suggested by some of the reviewers, in this global rebuttal we display the effect of degree of adaptivity $k$ on the single coordinate estimation. These simulations provide empirical validation to the theory developed in the paper (Theorem 3.1, 3.2). To better visuali... | NeurIPS_2023_submissions_huggingface | 2,023 | Summary: As I have communicated with the area chair, I will not be reviewing due to a conflict of interest. Submitting default ratings intended to be ignored below.
Strengths: NA
Weaknesses: NA
Technical Quality: 3 good
Clarity: 3 good
Questions for Authors: NA
Confidence: 1: Your assessment is an educated guess.... | null | Summary: This paper investigates degree of adaptivity in data impacts the performance of estimating a low-dimensional parameter component in high-dimensional linear models. The main result is giving an error bound of a low-dimensional component that does not have diension dependence. They propose an estimator TALE. Fo... | Rebuttal 1:
Rebuttal: We would like to thank the reviewer for taking the time to review our paper and provide helpful feedback! We truly appreciate your comments and suggestions and believe they can make our work better.
In the following we hope to address each of the points made in the review, in the order as they ca... | null | null | null | null |
Universality and Limitations of Prompt Tuning | Accept (poster) | Summary: This paper examines the theoretical capacity and limitations of prompt tuning. The authors have demonstrated the possibility of constructing a large transformer model that is sufficient for prompt-tuning to exhibit universal approximation over a Lipschitz function space. However, they have also shown the limit... | Rebuttal 1:
Rebuttal: We sincerely thank Reviewer AVty for the feedback and valuable suggestions. We have addressed each of the concerns below:
**Regarding additional experiments on more datasets**
Please see our general response for experiments on additional datasets
**Significance of memorization capacity of neura... | Summary: This paper wants to provide the insights on "when and how to perform prompt tuning to adapt a pretrained transformer to downstream tasks", by theoretically quantifying the universality (i.e., universal approximators) and limitations (i.e., representation capacity) of prompt tuning, in complementing a lot of wo... | Rebuttal 1:
Rebuttal: We sincerely thank Reviewer adez for the feedback and valuable suggestions. We have addressed each of the concerns below:
**Regarding explanation on L-Lipschitz function**
In this work, we consider a sequence-to-sequence function $f$ with input $X \in [0, 1]^{d \times m}$. $f$ is an L-Lipschitz ... | Summary: This paper focuses on understanding the role and theoretical underpinnings of soft-prompt tuning in transformer-based architectures, which is a technique used to adapt pretrained language models for new tasks. Despite the empirical effectiveness of this method, there's a lack of theoretical understanding of ho... | Rebuttal 1:
Rebuttal: We sincerely thank Reviewer E8G5 for the feedback and valuable suggestions. We have addressed each of the concerns below:
**Regarding the extension of Theorem 1 to decoder models**
Thank you for your valuable suggestion. The universality result in Theorem 1 can be extended to decoder models with... | Summary: This paper presents theoretical analysis of prompt tuning by 1) formally define the attention and prompting operations, 2) deriving the universality of prompt tuning with a strong transformer model, 3) further discussing the limitation of prompt tuning under the case of single and multi-layered transformer mod... | Rebuttal 1:
Rebuttal: We sincerely thank Reviewer VZsv for the feedback and valuable suggestions. We address each of the concerns below:
**Regarding additional experiments**
Please see the general response.
**Regarding discussions on potential directions to improve prompt-tuning**
From the analysis in Theorem 2 and... | Rebuttal 1:
Rebuttal: We thank all the reviewers for their feedback and valuable suggestions.
**General response on additional experiments**
As requested by most reviewers, we add additional experiments to further validate our theoretical results in the paper.
In Section 7, we designed two experiments to validate o... | NeurIPS_2023_submissions_huggingface | 2,023 | Summary: In this work, the authors embark on exploring the capabilities of prompt tuning in the continuous regime,
contrasting it with fine-tuning, as an initial endeavor toward theoretical comprehension. The authors prove
by construction that prompt tuning admits universal approximation within the space of Lipschitz
f... | Rebuttal 1:
Rebuttal: We sincerely thank Reviewer 1J6k for the positive feedback and valuable suggestions. For the corresponding additional experiments to validate our theorems, please see the general response. If the response satisfies your concerns, we hope that you will improve the score to reflect the same. We are ... | null | null | null | null | null | null |
Unpaired Multi-Domain Causal Representation Learning | Accept (spotlight) | Summary: The authors study the setting of unsupervised learning where observations belong to several domains, and we only observe the marginal distribution of each domain. A set of latent variables generates the observations, where a subset of latents are shared across domains. The authors provide the first identificat... | Rebuttal 1:
Rebuttal: We thank the reviewer for the supportive review and helpful comments.
*This paper makes two contributions. The first is to extend identifiable single-domain linear ICA to the multi-domain setting. The second is to extend single-domain graph identifiability to the multi-domain setting. In both cas... | Summary: In this paper, the authors address unpaired multi-domain causal representation. In detail, the authors learn the representations of the observed data from different domains that consist of causally. To achieve this, the authors consider the data generation process where the relationship between the latent vari... | Rebuttal 1:
Rebuttal: We thank the reviewer for their thorough review, which raises several helpful points. We hope that our response addresses their remaining concerns.
# Response to weaknesses
**1.)** We thank the reviewer for pointing out reference [1], which is indeed related to our work, and we will gladly add a... | Summary: - The paper considers causal representation learning from unpaired multi-domain data, with latent variables both shared and specific to domains.
- Its key contribution is a new identifiability result for linear causal models with non-Gaussian noise, linear mixing function, and a number of other assumptions.
- ... | Rebuttal 1:
Rebuttal: We thank the reviewer for their positive comments regarding the quality of our writing, the importance of the studied causal representation learning problem, and the novelty of our result. We hope that the following responses addresses their remaining questions.
# Response to weaknesses
*The con... | Summary: This work tackles the problem of learning the latent causal structure from multiple unpaired domains. Under a linear non-Gaussian condition, this work presents the identifiability guarantees for the joint distribution over the domains and the causal structure within the shared latent partition. Synthetic data ... | Rebuttal 1:
Rebuttal: We thank the reviewer for their positive review of our work and their supportive comments on the motivation of the problem and the presentation of the paper. We hope that the following responses addresses their remaining questions.
# Response to the weaknesses
### Linearity
We agree that the li... | null | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
$S^3$: Increasing GPU Utilization during Generative Inference for Higher Throughput | Accept (poster) | Summary: This work builds an LLM inference platform, called S^3, around a sequence length predictor.
The sequence length predictor is used to
1. batch LLM generations
2. pre-allocate kv cache (where all seq have similar predicted seq len)
S^3 also has a method for handling seq len prediction errors. They do pipelined g... | Rebuttal 1:
Rebuttal: - **How does $S^3$ perform against pre-allocating S tokens at a time?**
Table below shows the throughput compared to the reviewer’s recommended system with S=64. Compared to $S^3$, pre-allocation of S tokens allows larger batch size but at the cost of more frequent evictions. That is the schedule... | Summary: The paper proposes a scheme that increases the throughput during inference on Transformer large language models (LLM). Typically, LLMs require large amounts of memory, for model parameters and for the KV (key/value) cache. The KV cache size depends of the output sequence length, which is not known when inferen... | Rebuttal 1:
Rebuttal: - **Real traces and SLO**
We agree with the reviewer and believe that the paper could be strengthened by using real-world request traces and realistic latency SLOs. However, as mentioned both by the reviewer and in Section 6 Limitations, we acknowledge that the traces are not available to the res... | Summary: The paper tackles the problem of predicting the number of generated tokens for transformers in text generation tasks. This will help with better memory allocation and batch size management. The previous systems either used dynamic memory allocation (Hugging Face) which incurs inference overhead, or preallocate... | Rebuttal 1:
Rebuttal: - **Do different models generate sequences with different lengths?**
Different models showed similar output sequence length distributions that were skewed toward short sequences.
- **The rationale behind using sequences/s.**
We wanted to provide the intuition of throughput in terms of how many m... | Summary: This paper presents a new solution to the challenges of GPU underutilization and increasing batch size in the generation task of Large Language Models (LLMs), rooted in the memory-intensive requirement to retain the K and V values of prior tokens. To tackle these issues, the paper proposes an efficient framewo... | Rebuttal 1:
Rebuttal: - **The discrepancy between the batch sizes in $S^3$ and Oracle**
The predictive mechanism of $S^3$, elaborated upon in Section 3 under the Predictor paragraph, revolves around forecasting the length bucket. $i^{th}$ bucket contains sequences with lengths in [max_seq_len/num_buckets * i, max_seq_... | Rebuttal 1:
Rebuttal: We extend our sincere gratitude to the reviewers for their thoughtful and insightful feedback on our NeurIPS submission, titled “$S^3$: Increasing GPU Utilization during Generative Inference for Higher Throughput.” Their meticulous evaluation and constructive comments have undeniably enriched the ... | NeurIPS_2023_submissions_huggingface | 2,023 | Summary: The paper proposes a simple yet effective systematic solution to increase GPU usage and throughput during inference. The authors first make an interesting observation that the existing large language models are bounded by memory and the inefficiency is caused by the lack of awareness of the sequence length. Ac... | Rebuttal 1:
Rebuttal: - **More details of $S^3$ in Section 3 Subsection Length-aware sequence scheduler**
In addressing the bin-packing problem, we leveraged the decreasing first-fit algorithm, which orchestrates bin-packing by prioritizing the largest to smallest load sequence. Notably, $S^3$ orchestrates its packing... | null | null | null | null | null | null |
Beta Diffusion | Accept (poster) | Summary: In this submission, the authors introduce a diffusion model for range-bounded data.
To do so they introduce a noising process that is multiplicative (and not additive) that leads to conditionals that are beta distributed and converges to a beta distribution which is independent of the original datapoint.
They ... | Rebuttal 1:
Rebuttal: Thank you for your valuable questions. We will now provide detailed clarifications.
>*W1: Why Section 2.4 is needed?*
To implement Algorithm 1 for training beta diffusion, Section 2.4 is not necessary. However, it is needed to demonstrate how discretization functions under beta diffusion. Specif... | Summary: This work introduces a new type of diffusion model based on the beta distribution. The end of this is to defined a diffusion model on data that lives inside a specified range, i.e. on an interval [a,b].
This is achieved by first defining a forward noising process conditioned on the initial data sample $x_0$, ... | Rebuttal 1:
Rebuttal: Thank you for providing a comprehensive summary of our paper. We deeply appreciate your positive feedback on the technical details, overall quality, and recognition of the potential for future work. In what follows, we will address any concerns you may have, and we are confident that our response ... | Summary: This paper addresses a prevalent assumption found in deep generative diffusion models, which is the Gaussian assumption in both the forward and reverse processes. In this study, the authors explore the utilization of the beta distribution in these processes and establish several key properties. Firstly, they a... | Rebuttal 1:
Rebuttal: We appreciate your excellent summary of the key contributions of our paper and your recognition of the uniqueness of the proposed methodologies. We are delighted that you found our paper to be well-written and presented with clarity. Below please see our point-by-point response.
> *W1: Absense of... | Summary: The presented work proposes to design the forward and reverse process of diffusion models based on beta distributions (as opposed to Gaussian noise). It also showcases a new loss term based on KL-divergence upper bounds (KLUB) that can aid performance over the conventional ELBO-based diffusion model objective ... | Rebuttal 1:
Rebuttal: The reviewer found our paper to be well-written and the exploration of beta diffusion as an interesting and relevant direction, but was very concerned about the lack of discussion and comparison with related works. We aim to decisively alleviate these concerns with our response and revision.
> *W... | Rebuttal 1:
Rebuttal: # Response to All
We express our gratitude to all four reviewers for their valuable insights and suggestions. Their comments have been instrumental in identifying relevant recent works on alternative diffusion processes, refining the paper's positioning within the literature, elucidating the prop... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Exploiting Negative Samples: A Catalyst for Cohort Discovery in Healthcare Analytics | Reject | Summary: This paper presents a Shapley value based cohort discovery, by constructing "Negative Sample Shapley Field" that possesses isotropy property. By doing so, negative samples can be effectively clustered and separated with respect to the Shapley values.
Strengths: I think this paper points out many important pr... | Rebuttal 1:
Rebuttal: We would like to express our gratitude for your detailed comments regarding our paper and respond to them point by point in detail below.
**W1.**
In Eq.1, given a metric $M$ such as accuracy, two negative samples are assigned the same data Shapley value ($s_i = s_j$) if and only if these two sam... | Summary: The paper addresses the cohort discovery problem for supervised learning in the machine learning for healthcare domain. Positive examples of the cohort are easy to identify while it is not as straightforward to determine which negative examples should be admitted into a cohort. To deal with this problem, the ... | Rebuttal 1:
Rebuttal: We appreciate your detailed feedback on our paper. In response to the raised concerns about novelty, experimental evaluation, and clarity, we provide detailed explanations below.
**W1.**
We wish to highlight our main novelty, which is identifying the research gap arising from the asymmetry betwee... | Summary: This paper describes a method to understand the set of unlabelled / negative samples in a healthcare data-set. In this setting, one typically has a set of patients with a particular label, such as indicidence of a particular disease, and a large set of unlabelled samples. Training a classifier involves selecti... | Rebuttal 1:
Rebuttal: We thank the reviewer for the positive feedback and valuable advice, particularly the suggestion to benchmark our proposal against positive-unlabelled learning methods, which has greatly contributed to enhancing our paper. As for the several weaknesses and questions you mentioned, our responses ar... | Summary: In healthcare analytics, cohort constructions is one of the key steps that drives the analysis. For most problems, where the outcome of interest is a disease, the problem has asymmetrical formalism - while patients with disease are defined using string criterion and are homogenous w.r.t problem the negative se... | Rebuttal 1:
Rebuttal: We would like to express our gratitude for your constructive comments regarding our paper, especially the suggestion to compare our method with contrastive PCA. We have incorporated the recommended experiment and provided a comprehensive complexity analysis. Please find our detailed responses addr... | Rebuttal 1:
Rebuttal: We would like to extend our profound gratitude to all the reviewers for their insightful comments and constructive suggestions, which have played a pivotal role in elevating the quality of our paper. In this global response, we wish to first emphasize the primary novelty and contributions of our w... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Sounding Bodies: Modeling 3D Spatial Sound of Humans Using Body Pose and Audio | Accept (spotlight) | Summary: This paper deals with the problem of predicting sound fields around human bodies and proposes a method that exploits binaural audio signals and human body motions. The proposed method first encodes binaural audio signals and human poses and then decodes them into audio signals that are supposed to be captured ... | Rebuttal 1:
Rebuttal: 1. Comparison with previous methods for sound field prediction
We recognize that a comparison with other methods would be beneficial. However, as pointed out by the reviewer as well, the sound field reconstruction from egocentric data is a novel problem and no existing methods can be directly app... | Summary: The paper introduces a novel model for a novel task to render a 3D spatial sound from human body motion and audio collected by headset microphones. The authors also present a novel dataset containing human body motion and audio for this task. The model takes encoded audio, pose features, and target microphone ... | Rebuttal 1:
Rebuttal: Weaknesses:
1. Subjective evaluation
We agree that subjective evaluations are important. At the same time, the proposed work deals with sound spatialization, and it is not straightforward to perceptually evaluate spatialization quality. Study participants should be able to see the scene in 3D a... | Summary: This work introduces an approach for generating spatial audio from a 3D human pose and microphones placed close to the subject’s head, e.g. on a VR/AR headset, as is the case here. The authors first collect a multimodal dataset that contains 3D bodies and audio, recorded using multiple Kinects for body trackin... | Rebuttal 1:
Rebuttal: Weaknesses:
We thank the reviewer for suggestions that would help make the related works section more comprehensive. We will add a paragraph that discusses mentioned works. NAF and INRAS both leverage implicit representations for sound field modeling, but they focus on room impulse responses and ... | Summary: The paper tackles the task spatializing a mixture of speech and body sounds at different points in a sphere around a human body without explicitly recording or knowing the sound source location. Towards that goal, the paper captures a new dataset of humans speaking and making different body sounds in a very co... | Rebuttal 1:
Rebuttal: Questions/concerns:
1. a) & b) Generalization to other environments
We would like to clarify that we are considering a VR telepresence application where people interact in a VR space as full-body avatars. In this scenario we can distinguish between two environments: (1) the real physical environ... | Rebuttal 1:
Rebuttal: We would like to thank all the reviewers for taking the time to review the paper and their help in improving the quality of the manuscript with their constructive feedback. We replied to each reviewer’s comments and suggestions individually below. Attached to this rebuttal is also a PDF that inclu... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Optimized Covariance Design for AB Test on Social Network under Interference | Accept (poster) | Summary: This paper focused on designing a randomization scheme at the cluster level for the A/B test. It proposed and derived an upper bound for the MSE of the HT estimator, which was targeted to be minimized to optimize the experiment design. This article treated the covariance matrix of the treatment vector as the d... | Rebuttal 1:
Rebuttal: First of all, thanks for your time and efforts in the reviewing process, and we appreciate your approval on the problem formulation and content presentation of our work. We address your concerns as follows.
## Not comprehensive experiments
We conduct an array of simulations with a substantial po... | Summary: In this paper, authors present a novel experimental design to be used for randomized experiments under network interference.
Authors begin by considering a baseline adjusted Horvitz--Thompson estimator (which crucially requires knowledge of $Y_i(\mathbf{0})$) and a pre-specified clusterings.
Given these two th... | Rebuttal 1:
Rebuttal: Thanks a lot for such a detailed and informative review, and your feedback will undoubtedly contribute to the enhancement of our work. We address your concerns as follows.
## Confidence interval
We think it poses a significant challenge due to the estimation of nuisance parameters, such as $\gam... | Summary: In this paper, the authors propose a new algorithm for A/B test design under network effect with cluster level randomization. By derivation of an upper MSE upper bound with bias-variance trade-off, the authors reparameterize to directly optimize the covariance of the treatment vector. An efficient PGD algorith... | Rebuttal 1:
Rebuttal: First of all, thanks for your time and efforts in the reviewing process and we appreciate your approval of the theoretical and empirical contributions of this work. We provide some comments on your interesting questions as follows.
## Clustering issue
The role of clustering indeed introduces int... | Summary: This paper presents an optimized covariance design for A/B tests on social networks with interference. The authors address the challenge of accurately estimating the global average treatment effect (GATE) in the presence of network interference. They propose a method to balance bias and variance in experimenta... | Rebuttal 1:
Rebuttal: First of all, thanks for your time and efforts in the reviewing process, and we address your concerns as follows.
## Comparability assumption
The comparability assumption is predicated on the premise that the magnitude of interference remains akin to the direct effect at the cluster level. This ... | null | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Multi-scale Diffusion Denoised Smoothing | Accept (poster) | Summary: This paper proposed a new technique for certified adversarial robustness based on randomized smoothing. Two certification schemes are proposed in the paper: first is cascaded randomized smoothing where multiple smoothed classifiers of different smoothing factors are aggregated together in an efficient way; the... | Rebuttal 1:
Rebuttal: We sincerely appreciate your thoughtful comments, efforts, and time. We respond to each of your questions one-by-one in what follows. Please let us know if you have any comments/concerns that we have not addressed up to your satisfaction.
---
**Q1. Computational overhead**
Thank you for the sug... | Summary: This paper presents a practical method to address the issues of over-smoothing and over-confidence in randomized smoothing, thereby enhancing our understanding of the trade-off between accuracy and robustness. For over-smoothing, the author introduces a multi-stage approach for computing collective certified r... | Rebuttal 1:
Rebuttal: We sincerely appreciate your thoughtful comments, efforts, and time. We respond to each of your questions one-by-one in what follows. Please let us know if you have any comments/concerns that we have not addressed up to your satisfaction.
---
**Q1. “Cascaded smoothing cannot improve certified ro... | Summary: This paper identifies the accuracy-robustness trade-off in randomized smoothing and the over-smoothing and over-confidence issues in diffusion denoised smoothing. It then proposes a to finetune diffusion models to mitigate these issues by first passing a datapoint through a cascaded smoothing pipeline to obtai... | Rebuttal 1:
Rebuttal: We sincerely appreciate your thoughtful comments, efforts, and time. We respond to each of your questions one-by-one in what follows. Please let us know if you have any comments/concerns that we have not addressed up to your satisfaction.
---
**Q1. The higher $\sigma$, the longer runtime?**
Tha... | Summary: This paper builds on diffusion based denoised randomized smoothing. It analyzes two scenarios of model errors which are over-smoothing and over-confidence. To alleviate these issues, the authors proposed deploying cascaded randomized smoothing where samples are first smoothed with large variance that gradually... | Rebuttal 1:
Rebuttal: We sincerely appreciate your thoughtful comments, efforts, and time. We respond to each of your questions one-by-one in what follows. Please let us know if you have any comments/concerns that we have not addressed up to your satisfaction.
---
**Q1. Comparison with previous data-dependent approac... | Rebuttal 1:
Rebuttal: Dear reviewers,
We thank all the reviewers’ efforts to improve our manuscript. This common response addresses a concern raised by multiple reviewers, viz., on our diffusion fine-tuning scheme compared to classifier fine-tuning. We also kindly ask you to find the attached rebuttal PDF in this resp... | NeurIPS_2023_submissions_huggingface | 2,023 | Summary: This work aims to improve robustness-accuracy tradeoff in image classifiers, using the recent smoothing ideas. The authors propose two techniques for improving adversarial robustness without loss of accuracy. One is to fine-tune the prepended denoiser with a regularized loss, to reduce miss-classification of s... | Rebuttal 1:
Rebuttal: We sincerely appreciate your thoughtful comments, efforts, and time. We respond to each of your questions one-by-one in what follows. Please let us know if you have any comments/concerns that we have not addressed up to your satisfaction.
---
**Q1. Suggestions for better clarity - e.g., notation... | null | null | null | null | null | null |
Conformal PID Control for Time Series Prediction | Accept (poster) | Summary: The paper proposes an adaptation to existing adaptive conformal prediction in two novel ways: (i) by tracking the quantile via online regression over the running sum of the errors; and (ii) by incorporating a second model to anticipate the quantile in the next instant ("scorecasting"). The authors include theo... | Rebuttal 1:
Rebuttal: We thank the reviewer for the accurate summary, and the engagement (esp. on reproducibility!)
See the main response (in a separate comment) for a summary of the extensive improvements we have made to the paper. Below we reference the specific points related to your critique. (We have shortened/pa... | Summary: The paper tackles the problem of conformal UQ under distribution shifts. This is a very realistic setting where the upstream prediction model cannot be frequently updated, and we still want to certify some level of safety. The paper cleverly fames the problem as a PID control problem, and provides a practical ... | Rebuttal 1:
Rebuttal: > The paper tackles the problem of conformal UQ under distribution shifts. This is a very realistic setting where the upstream prediction model cannot be frequently updated, and we still want to certify some level of safety. The paper cleverly frames the problem as a PID control problem, and provi... | Summary: This paper examines the issue of how to parameterize conformal prediction in the time series setting. It argues that standard conformal inference methods would not provide valid inferences in the sequential setting, which lacks exchangeability. To resolve this, the paper suggests using online quantile tracking... | Rebuttal 1:
Rebuttal: > This paper examines the issue of how to parameterize conformal prediction in the time series setting. It argues that standard conformal inference methods would not provide valid inferences in the sequential setting, which lacks exchangeability. To resolve this, the paper suggests using online qu... | Summary: This paper aims to provide asymptotically valid conformal prediction (CP) regions for the time series prediction problem. Similar to the adaptive conformal inference (ACI) framework [11], the main idea is to choose the appropriate quantile of the non-conformity scores (or equivalently tuning the miscoverage ra... | Rebuttal 1:
Rebuttal: Thank you for the accurate summary and the thoughtful review. We enjoyed engaging with your comments.
The main response (in a separate comment) contains a summary of the extensive improvements we have made to the paper, including new experiments, finite-sample theoretical statements, automatic pa... | Rebuttal 1:
Rebuttal: We are grateful to the four expert and engaged reviewers, who took a clear interest in the paper and suggesting ways to improve it. Thank you!
All four said the practical benefits are substantial, as evidenced by comprehensive experiments. Three lean towards acceptance.
We hope to respond to the... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Residual Alignment: Uncovering the Mechanisms of Residual Networks | Accept (poster) | Summary: This work discovers the phenomenon of Residual Alignment (RA) in ResNets, whereby the top left and right singular vectors of residual Jacobians align with each other and in between different residual blocks. Through extensive experimental verification as well as novel theoretical frameworks and derivations, th... | Rebuttal 1:
Rebuttal: We are grateful to the reviewer for the time and effort they must have dedicated to providing such a thorough and constructive review. Below, we respond to each point that was raised.
> RA1 and RA3 are specific to classification, and all the tasks considered fall within such a context. If the fin... | Summary: The paper "Residual Alignment: Uncovering the Mechanisms of Residual Networks" explores the underlying mechanisms and success factors of the ResNet architecture, which has gained significant popularity in deep learning. The authors conduct an empirical study by linearizing the residual blocks of ResNet using R... | Rebuttal 1:
Rebuttal: We would like to extend our appreciation to the reviewer for their diligent and thorough evaluation of our paper. In what follows, we address carefully each raised point.
> Fig 1 caption: “s true label and connected to form a trajectory” every connected line goes from input to output? Shouldn’t ... | Summary: The paper tries to analyze the remarkable performance of ResNet architecture and they find the residual alignment phenomenon. The phenomenon is general and they also proposed a mathematical model call the Unconstrained Jacobian Models to theoretically analyze it.
Strengths: The authors find an interesting phe... | Rebuttal 1:
Rebuttal: We express our sincere gratitude to the reviewer for their inspiring questions and careful review of our paper. Below, we respond to each point that was raised.
> I think the weaknesses of this paper have they only discussed such a phenomenon. It will be great if they can utilize such phenomenon... | Summary: This paper investigates the ResNet architecture, a popular deep-learning model known for its improved performance through skip connections. The authors aim to uncover the underlying mechanisms behind its success. They conduct an empirical study by linearizing the residual blocks of ResNet using Residual Jacobi... | Rebuttal 1:
Rebuttal: We express our sincere gratitude to the reviewer for the time and effort they put into the assessment of our paper as well as for their insightful comments and critiques. In what follows, we address each of the points raised.
> Linearization of the Residual Jacobian and analysis of its SVD decomp... | Rebuttal 1:
Rebuttal: Enclosed, please find a page detailing further experiments we carried out in response to the reviewers' inquiries.
Pdf: /pdf/1458732aa1e93aa0e397be5072deb50edc716bbb.pdf | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Stabilized Neural Differential Equations for Learning Dynamics with Explicit Constraints | Accept (poster) | Summary: Neural ODEs which learn a particularly dynamics can be unstable during evolution of the learned dynamics. This papers studies instabilities which take solutions outside of the constraint space of a given evolution. The authors show that a simple added control term to the learned evolution can improve stability... | Rebuttal 1:
Rebuttal: >There are a host of tools in the Riemannian optimization literature...For example, there exists many algorithms for learning flows or vector fields on manifolds (e.g. [1], [2]) and performing transport on those manifolds.
We thank the reviewer for bringing our attention to these very interesting... | Summary: This paper proposes a general framework to stabilize NODE models. By introducing a new stabilization term, it can learn vector fields that satisfy various conditions. The authors evaluate the effectiveness of the proposed method on several physical systems represented by ODEs.
Strengths: - This paper proposes... | Rebuttal 1:
Rebuttal: > Lack of clarity on how the problem setup differs from HNN and LNN.
We thank the reviewer for drawing our attention to this shortcoming and we have updated the related work section of the manuscript to further clarify how our method and the problem setup relates to and differs from HNNs and LNNs... | Summary: The authors propose a new method called stabilized neural differential equations (SNDEs) that enforce arbitrary manifold constraints for neural differential equations. The proposed method is applicable to any type of ODE, hybrid models (UDEs), and can incorporate any type of manifold constraints. The authors' ... | Rebuttal 1:
Rebuttal: > Despite providing a thorough overview of related work that also incorporate constraints, experimentally the authors only compare to NODE and SONODE which are known not to extrapolate well beyond the training regime due to the limited constrains/bias for the differential function/vector field. If... | Summary: This paper aims to learn ODE dynamics where the dynamics has some extra equation constraint. For example, a double pendulum with energy conservation. In the problem setting, a neural network is used to learn the time derivative of the system, the dynamics is then rolled out with a classical solver. To ensure t... | Rebuttal 1:
Rebuttal: > The baseline method involves no stabilization. Is there other simple method that can ensure the stabilization? For example, since the conservation equation is known explicitly, can one use one degree of freedom only to satisfy the constraints? I.e., if the output is -dimensional, the network can... | Rebuttal 1:
Rebuttal: We thank the reviewers for their thoughtful reviews and insightful feedback. We are glad they found the problem we study important (**RTPu**) and relevant (**kVz8**) while appreciating the novelty (**kVz8**) and simplicity (**RTPu**, **pbb8**) of our approach. We are encouraged that they found our... | NeurIPS_2023_submissions_huggingface | 2,023 | Summary: The authors of this paper proposed to augment neural ODEs with an additional stabilization term that is derived from the pseudo-inverse of a Jacobian matrix to learn dynamics of a system from data. The authors provided theoretical grantees to show that their augmented version is still capable of learning the o... | Rebuttal 1:
Rebuttal: > The idea behind this paper seems simple and the results seems promising, which given it's simplicity, could really be an impactful contribution. Hence, if the authors could also demonstrate the gains of this method on another variation of vanilla NODE and also perhaps an actual application, I be... | null | null | null | null | null | null |
P-Flow: A Fast and Data-Efficient Zero-Shot TTS through Speech Prompting | Accept (poster) | Summary: The paper proposes P-Flow that can achieve high speaker similarity performance and fast inference speed on the zero-shot TTS Task. To improve the speaker similarity score, P-Flow uses a speech prompted text encoder to generate speaker-conditional text representation for speaker adaptation. To achieve fast inf... | Rebuttal 1:
Rebuttal: **Q1,3. Regarding timbre instability**
After careful analysis, we believe that the event observed by the reviewer is a change in the fundamental frequency range. While timbre is defined to be F0 invariant [1], it is possible that a change in fundamental frequency range can be perceived as a chang... | Summary: This work propose a flow-matching zero-shot TTS model called P-Flow. It is fast and data-efficient in comparison with Vall-E.
Strengths: I like the idea of flow matching, and it seems a new fashion for generative tasks. I do believe flow-matching will benefit the speech generation community.
The idea of promp... | Rebuttal 1:
Rebuttal: We agree that our base model architecture is similar to existing Non-AutoRegressive (NAR) TTS methods (Glow-TTS and Grad-TTS). However, we believe our work provides valuable insight into the current state of zero-shot TTS research. The recently proposed language modeling-based zero-shot TTS model,... | Summary: The paper introduces P-Flow, a novel zero-shot text-to-speech (TTS) model that addresses the limitations of existing large-scale neural codec language models. P-Flow utilizes speech prompts for speaker adaptation and consists of a speech-prompted text encoder and a flow-matching generative decoder. The text en... | Rebuttal 1:
Rebuttal: **Motivation for using a flow matching decoder**
The precursor papers on conditional flow matching [1, 2] provided evidence of the advantages of flow matching versus DDPM, especially flow matching's ability to simulate simpler trajectories while retaining good sample quality with fewer ODE steps ... | Summary: The paper proposes a zero-shot TTS model, P-Flow, which combines a speech-prompted text encoder and a flow-matching generative decoder. The encoder generates a speaker-conditional text representation using the target speech prompt and text, while the decoder utilizes conditional flow matching to model the cond... | Rebuttal 1:
Rebuttal: **W1 and Q3 - P-Flow with long prompt**
Concerning P-Flow for long prompts, we began by taking inspiration from the VALL-E and SPEAR-TTS, which demonstrated effective voice cloning using only 3 seconds of voice data. Based on this, we fixed our speech prompt length at 3 seconds for all experiment... | Rebuttal 1:
Rebuttal: Thank you for taking the valuable time to review our paper and providing helpful feedback. All the questions will greatly contribute to improving our paper.
We have provided responses to the weaknesses and questions mentioned by each reviewer. Additionally, we have included figures and tables fo... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
FAMO: Fast Adaptive Multitask Optimization | Accept (poster) | Summary: The paper addresses an important issue with (most) prior MTL optimization techniques which requires the computation of all per-task gradients, during training, for obtaining the update direction. This results in a $\mathcal{O}(K)$ requirement in space and time where $K$ is the number of tasks. The authors prop... | Rebuttal 1:
Rebuttal: We thank the reviewer for the constructive feedback and comments. We address your concerns and questions in the following.
---
**Weakness:**
**1. More discussion is needed with MGDA.**
Thanks for the suggestion, we will make the difference and connection to MGDA more clear in the final version... | Summary: This paper proposes a new multi-task balancing method with claimed O(1) efficiency, which is much more efficient than previous gradient manipulation methods.
The core idea is to let the multiple losses decrease at roughly the same speed.
Strengths: The method is more efficient than previous gradient manipu... | Rebuttal 1:
Rebuttal: We thank the reviewer for the constructive feedback and comments. We address your concerns and questions in the following.
---
**Weakness:**
**1. FAMO does not manipulate gradients but reweighs them. Re-weighting methods exist like UW/DWA. Should refactor the abstract, intro, method.**
Existin... | Summary: This paper proposes a dynamic convex combination multi-task learning losses reweighting so that it can decrease different task losses more balancedly while having little computational overhead than simple task loss average. Specifically, they formulate an optimization problem to find a new “gradient” direction... | Rebuttal 1:
Rebuttal: We thank the reviewer for the constructive feedback and comments. We address your concerns and questions in the following.
---
**Weakness:**
**1. Lack of significant contribution. Why STL is good?**
The 2 contributions of FAMO are: 1) we propose that we should do equal-rate descent instead of ... | Summary: This paper proposed a novel multi-task optimization method aimed at mitigating conflict between task gradients without inducing the substantial time slowdown that typically comes with specialized multi-task optimizers.
The proposed method, called FAMO, aims at improving the worst-case rate of improvement acros... | Rebuttal 1:
Rebuttal: We thank the reviewer for the constructive feedback and comments. We address your concerns and questions in the following.
---
**Weakness:**
**1. Compare FAMO against amortized MGDA.**
Please refer to Figure 1 above. We see that since amortized MGDA (like MGDA) is seeking equal descent for eac... | Rebuttal 1:
Rebuttal: ## Common Response to All Reviewers (with Additional Results)
---
We sincerely thank all reviewers for their comments and valuable suggestions. Per the reviewers' request, **we conduct additional experiments on NYU-v2 and summarize the results in the attached PDF. We will respond to each reviewe... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Implicit variance regularization in non-contrastive SSL | Accept (poster) | Summary: This article builds on the theory of non-constrastive (nc) self-supervised learning (SSL), with methods such as BYOL or SimSiam. Contrary to the existing litterature, they study directly the cosine similarity loss used rather than a Euclidian loss on the eigenspace, using NTK dynamics. They show that collapse ... | Rebuttal 1:
Rebuttal: Thanks for your questions and suggestions. It resulted in an exciting new figure showing that our theory does apply to general settings. Please find below our point-by-point response
> *“I find the use of the result of Tian et al. that the predictor eigenspace aligns with the one of the correlati... | Summary: The learning dynamics of non-contrastive self-supervised learning is an important problem to understand how these methods avoid collapse without using negative samples. In this paper, the authors provide a rigorous analysis of this problem on a simple linear network with Gaussian inputs, especially the differe... | Rebuttal 1:
Rebuttal: Many thanks for enabling a further discussion of the finer points of our theoretical analysis. Please find below our point-by-point response.
> *“For real-world datasets and practical network structures, it is not clear if those insights are still valid.”*
Thanks for the question. Our empirical ... | Summary: This work analyzes the learning dynamics of non-contrastive SSL approaches such as BYOL and Simsiam. Based on the proposed theory, the authors analyze the how the stop-grad and predictor module affect the learning dynamics. Importantly, the authors design a theoretically inspired loss and gain improvement on c... | Rebuttal 1:
Rebuttal: Thanks for your feedback. It helped us significantly improve our numerical results. Please find below our point-by-point response.
> *“Lack of basic introduction to the used techniques, e.g., the neural tangent kernel (NTK).”*
We are sorry this was missing in the initial submission. We will rest... | null | null | Rebuttal 1:
Rebuttal: We thank the reviewers for the detailed and insightful feedback, which allowed us to further improve the article. Importantly, we improved our numerical results. The comments further motivated us to reconsider the scope of our article by focusing more on the theory part.
A common concern raised... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Efficient Post-Processing for Equal Opportunity in Fair Multi-Class Classification | Reject | Summary: This paper considers fairness in multi-class classification under the notion of parity of true positive rates - an extension of binary class equalized odds - which ensures equal opportunity to qualified individuals regardless of their demographics. We focus on algorithm design and provide a post-processing met... | Rebuttal 1:
Rebuttal: We thank the reviewer for the time, comments, and support! We hope that the following addresses the concerns:
- **(W2, Q3) "novelty of the proposal is not clear... elaborate and descrive the technical challenges of the contribution?"**
We would like to kindly remind the reviewer of our main con... | Summary: This paper proposes a novel post-processing approach to reduce the true positive rate parity for multi-class classification problems. It is shown on two real world data to outperform an existing baseline in terms of accuracy and true positive rate parity.
Strengths: 1. The proposed approach is novel and techn... | Rebuttal 1:
Rebuttal: We thank the reviewer for the time, comments, and support! We hope that the following addresses the concerns:
- **(Q1) "Why not compare this approach to ["Learning Fair Representations"]?"**
The group fairness criterion considered in [1] is for statistical parity, $A\perp Y$, not equal opportun... | Summary: The work proposes a post-processing algorithm to achieve the equal opportunity constraint in multi class classification. The proposed algorithm takes arbitrary Bayes rule estimate and only requires additional unlabelled data. The authors derive finite-sample guarantees and perform empirical evaluation to suppo... | Rebuttal 1:
Rebuttal: We thank the reviewer for the time, detailed comments, and support! We hope that the following addresses the concerns:
- **(Q1, Q4) "explicitly present the form of the optimal fair classifier"**
Thank you for the suggestion! We provide below the explicit expression (although not closed-form, wh... | Summary:
This paper studies algorithmic fairness in multiclass classification setting. The fairness notion considered is parity of true positive rates (TPR) which is the multi class analog to equalized odds. The paper gives a post-processing algorithm which, given a score function, outputs a fair classifier. The paper... | Rebuttal 1:
Rebuttal: We thank the reviewer for the time, comments, and support! We hope that the following addresses the concerns:
- **"What are the assumptions on group structure? Must they be disjoint?"**
We do not impose any requirement on the group structure; the practitioner shall set $\mathcal A$ to include a... | null | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Inference Attacks Against Face Recognition Model without Classification Layers | Reject | Summary: This paper introduces a novel inference attack algorithm for face recognition models that do not have classification layers. The proposed attack consists of two stages: membership inference and model inversion. The membership inference attack analyzes the distances between intermediate features and batch norma... | Rebuttal 1:
Rebuttal: Thanks very much for constructive comments.
1、To our knowledge, there is currently no research on model inversion attacks against face recognition models without classification layers, i.e., there is no method to generate any training images in such a scenario. We are the first to explore this p... | Summary: The authors propose a membership inference attack against face recognition (FR) models in the white-box scenario where membership information is known for some records and white-box model access is available, but without access to a classification layer. The attack utilizes information stored in batch-norm sta... | Rebuttal 1:
Rebuttal: Thanks very much for careful and detailed comments.
1. In your reference [1], they used face embeddings for the attack, but they did not achieve model inversion attacks. (Coincidentally, this article was published in April, whereas our work was also completed in April.) In [2], we mentioned it an... | Summary: The paper presents a novel method for inference attacks against face recognition method. In particular, it advocates two-stage inference attack, where the first stage identify the membership and the second stage involves model inversion attack that recover the input from embedding. Experimental evaluation show... | Rebuttal 1:
Rebuttal: Thanks very much for your insightful reviews.
1. As far as we know, there is no standard evaluation for model inversion attacks without classification layers. Our work is the first to focus on model inversion attacks under such a scenario, i.e., there is no method to generate any training images ... | Summary: In this submission, the authors advocate an inference attack composed of two stages for practical FR models. The first stage analyzes the distances between the intermediate features and batch normalization parameters. The second stage reconstructs data using a pre-trained generative adversarial network (GAN) g... | Rebuttal 1:
Rebuttal: Thanks very much for constructive and insightful suggestions.
1. It is true that the way of using BN to perform membership inference attacks has been explored, but this paper focuses on the model inversion attack in the field of face recognition without classification layers, which is the very fi... | Rebuttal 1:
Rebuttal: As far as we know, model inversion attack against face recognition models without classification layers is a challenging task and is first proposed in this paper. This is a more practical scenario and there is currently no method to recover any training images in the scenario. We are the first to ... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
New Complexity-Theoretic Frontiers of Tractability for Neural Network Training | Accept (poster) | Summary: The paper studies the neural network training problem for ReLU and linear networks. They show new settings where the problem can be solved in polynomial time. They provide two main contributions:
- For ReLU networks, they give a polynomial-time algorithm for training constant-size networks where the out-degre... | Rebuttal 1:
Rebuttal: Regarding our results on ReLU networks, the class of networks where our algorithmic upper bound applies can indeed be considered as restrictive - especially when one views it from the perspective of practical applications. However, from a fundamental graph-structural viewpoint, we believe that bei... | Summary: This paper studies the computational complexity of empirical risk minimization (ERM) for neural networks, that is, given an architecture and training data points, find weights and biases of a global minimum of the training error. The problem is well-known to be NP-/ER-/W[1]-hard already in very easy special ca... | Rebuttal 1:
Rebuttal: We are grateful for the encouraging (and very helpful) feedback, and completely agree that obtaining an understanding of the fundamental complexity of neural network training is an important part of machine learning. The article improves our understanding of the boundaries of tractability of the p... | Summary: The paper studies algorithms for exact minimization of quadratic loss over ReLU and linear neural networks when the total number of neurons is a fixed constant. The results can be divided into two parts:
* A polynomial time algorithm for a fixed size ReLU/linear network, when it has the following structure: 1... | Rebuttal 1:
Rebuttal: Regarding the perceived weaknesses of our results, the class of networks where our algorithmic upper bound applies can indeed be considered as restrictive, but from a fundamental network-structural viewpoint we believe them to be a significant step beyond the previous state of the art. All results... | Summary: This paper studies the computational complexity of training linear and ReLU neural networks. Under assumptions such as the squared loss and the out-degree of every hidden neuron being exactly one, the authors prove that there exists an algorithm such that the global optimal solution can be computed in polynomi... | Rebuttal 1:
Rebuttal: We are indeed well aware that the tractability results obtained in the article are of theoretical interest, and that real-life networks are unlikely to satisfy the stated structural properties. Even the tractability for general bounded-size ReLU-NNT architectures obtained in Section 5 are not like... | Rebuttal 1:
Rebuttal: We thank all reviewers for their feedback and insightful comments. Detailed responses to individual questions and potential misunderstandings are provided as separate comments to each review. | NeurIPS_2023_submissions_huggingface | 2,023 | Summary: This paper investigates the theoretical complexity of finding the global optimal solution of ReLU networks and linear networks. Specifically, it first considers a specially designed neural network, each hidden neuron of which has only one outgoing edge. By using the homogeneity of the ReLU function, it reparam... | Rebuttal 1:
Rebuttal: We would like to point out two potential misunderstandings related to the purpose and scope of the article. Indeed, we do obtain polynomial-time algorithmic upper bounds under the assumption that certain parameters of the input (such as the size of the architecture) are fixed to be constants. Howe... | null | null | null | null | null | null |
Optimal Treatment Regimes for Proximal Causal Learning | Accept (poster) | Summary: The authors present a new optimal individual treatment regime (ITR) within the proximal causal inference framework, which avoids the strong assumption of no unmeasured confounding. Instead, one assumes the effect of the unmeasured confounders flows exclusively through proxy variables, as defined through outcom... | Rebuttal 1:
Rebuttal: We sincerely appreciate the constructive suggestions from the reviewer.
W1: The proposed extension of the ITR function class appears quite incremental. The value of the proposed ITR follows directly from application of the tower rule. The paper would be greatly strengthened if the authors can sh... | Summary: The goal is to learn an optimal individual treatment rule (ITR) where the data suffer from unobserved confounding but where the researcher has a treatment proxy and an outcome proxy.
While the general problem has been studied before by Qi et al (JASA 2023), this paper’s contribution is to broaden the class o... | Rebuttal 1:
Rebuttal: We sincerely thank the constructive suggestions from the reviewer.
W1: The theoretical contribution is somewhat incremental.
R: We appreciate the reviewer's feedback and their assessment of our theoretical contribution. We acknowledge the importance of continually advancing the field, and we ar... | Summary: The paper discusses the optimization of treatment rules in the context of observational data and under assumptions of proximal inference. Various theorems are introduced, and a real data analysis performed using a healthcare example.
Strengths: Below is a list of perceived weaknesses.
The paper is overall s... | Rebuttal 1:
Rebuttal: We first thank the reviewer for a careful reading of our work.
W1: It was not clear to me how the empirical results compare to competing methodological baselines from other approaches (I don't believe the different values presented in the figure represent different algorithmic approaches).
R: In... | Summary: Most estimation methods for individualized treatment rules (ITRs) assume no unmeasured confounders for valid causal inference. However, such an assumption can be unreasonable, such as when estimating ITRs from observational data. Previous work has applied proximal causal inference to estimate ITRs when this as... | Rebuttal 1:
Rebuttal: We sincerely appreciate the constructive suggestions from the reviewer.
W1: Theoretical guarantees are much weaker than those of previous methods.
R: We express our gratitude for your invaluable suggestions regarding the exploration of convergence rates and finite sample error bounds associate... | Rebuttal 1:
Rebuttal: We extend our heartfelt gratitude for all the comments provided by the reviewers on our work. Within this global rebuttal, we address recurring concerns raised by multiple reviewers. Our response is organized into three key parts: refining the theoretical analysis, presenting updated experimental ... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Multi-Swap k-Means++ | Accept (poster) | Summary: This paper studies local-search algorithms for k-means clustering. The goal here is to obtain a local-search algorithm which (1) give a close to 9-approximation ratio (which is best possible for local search algorithms) and (2) is practical. In the past literature, there has been many local search algorithms d... | Rebuttal 1:
Rebuttal: We thank the reviewer for their time and effort put into reviewing the paper.
Answer to the questions:
We did not manage to prove the tightness of our analysis, neither for $p = 1$ (26.64..), nor for $p$ approaching infinity (10.48..). This is an interesting open question. However, it seems to u... | Summary: This paper studies the standard $k$-means problem. Given a $k$-means instance $(P,k)$, the goal is to find a set $C$ of centers with size at most $k$ such that the sum of the squared distances from $P$ to $C$ is minimized. For the $k$-means problem, Lattanzi and Sohler (ICML 2019) proposed an elegant combinat... | Rebuttal 1:
Rebuttal: We thank the reviewer for their time and effort put into reviewing the paper. We start by addressing the weaknesses suggested by the reviewer and then we answer the specific questions.
Answers to weaknesses:
Weakness 1. Although the algorithm is an extension of the one in Lattanzi and Sohler, th... | Summary: This paper proposes a new k-means algorithm: multi-swap local search (MSLS) which combines local search and k-means++, to achieve a constant approximation guarantee with efficient time complexity. Specifically, the local search framework includes a step that selects alternative centers for optimizing the cost ... | Rebuttal 1:
Rebuttal: We thank the reviewer for their time and effort put into reviewing the paper.
We would like to clarify that some exponential dependence on p is likely needed for any local-search based algorithm. In particular, if one could devise an exhaustive local search algorithm with polynomial dependence on... | Summary: The following results are given in the paper:
1. A tighter analysis of the local search algorithm of Lattanzi and Sohler. The paper shows a constant approximation guarantee for their algorithm.
2. The paper extends the approach of Lattanzi and Sohler to multi-swap local search (where more than one point is sw... | Rebuttal 1:
Rebuttal: We thank the reviewer for their time and effort put into reviewing the paper.
We acknowledge that the proof technique is similar to the one by Kanungo et al. on a high level. However, the exhaustive local search technique of their paper leads to their algorithm being prohibitive for practical ap... | Rebuttal 1:
Rebuttal: While we respond to each reviewer individually, we attach here a revised version of our paper where we incorporate things discussed with reviewer 2sgw.
We note that while this attached version is slightly longer than 9 pages, we only use it to refer to it in this rebuttal phase, and will be made ... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Automated Classification of Model Errors on ImageNet | Accept (poster) | Summary: This paper proposes a pipeline to automatically categorize ImageNet classification misclassifications. Previous work [1] defined and categorized those misclassifications into 4 types manually and this is the main motivation of automating the process. The results show that their pipeline has an accuracy of 60% ... | Rebuttal 1:
Rebuttal: We thank Reviewer $\Rh$ for their review and answer their questions below.
**Q: Can you discuss the motivation for your method? Where is the novelty in automating the categorization proposed by [1]?**
Please see our general response for a more detailed motivation. Further, we believe the large ... | Summary: The paper introduces an automated analysis on the model errors in ImageNet.
The analysis sequentially categorizes into six types of errors with increasing
criticality, unveiling the true model failures. The results are contextualized
with previous work.
Strengths: - This work is of high interest to the comput... | Rebuttal 1:
Rebuttal: Thank you very much for the kind review. We are encouraged that you see our work as a valuable contribution to the community. Regarding your question about using the WordNet hierarchy to form superclasses, we refer you to the general response and are happy to answer any further questions.
---
Re... | Summary: The paper introduces an automated pipeline to classify ImageNet's errors from 4 categories: (i) fine-grained categories; (ii) fine-grained OOV; (iii) Non-prototypical instances; and (iv) Spurious correlations. The main contribution of the paper is this automated pipeline with which various errors can be catego... | Rebuttal 1:
Rebuttal: We thank Reviewer $\RH$ for their insightful feedback, helpful suggestions, and interesting questions which we address below.
**Q: Can the proposed method be extended to ImageNet variants, such as ImageNet-A. What are the error distributions in such sets?**
In principle, our pipeline can be ext... | Summary: This paper aims to create an automated pipeline for defining the types of errors made by ML models on ImageNet. Error categories are from previous work, and the paper compares the results between the previous human based pipeline and the newly automated version. The authors use the pipeline to evaluate over 10... | Rebuttal 1:
Rebuttal: We thank reviewer $\Rr$ for their review and interesting questions, which we answer below.
**Q: Can you provide an in-depth analysis of the results obtained using your automated pipeline, e.g., by manually categorizing model failures?**
While we include some more high-level observations (e.g. t... | Rebuttal 1:
Rebuttal: $\newcommand{RG}{\textcolor{green}{GgcH}}$
$\newcommand{Rr}{\textcolor{blue}{rfV8}}$
$\newcommand{RH}{\textcolor{purple}{HYyY}}$
$\newcommand{Rk}{\textcolor{orange}{kbqQ}}$
$\newcommand{Rh}{\textcolor{teal}{hvnS}}$
We thank all reviewers for their work and are delighted that they appreciate the s... | NeurIPS_2023_submissions_huggingface | 2,023 | Summary: The paper proposes an automatic error classification pipeline for ImageNet models. In particular, they categorize mistakes into overlapping classes, multi-object images, fine-grained and fine-grained with out-of-vocabulary (OOV), non-prototypical, the examples influenced by spurious correlations and unexplaine... | Rebuttal 1:
Rebuttal: We thank Reviewer $\RG$ for their detailed review, insightful feedback, helpful suggestions, and interesting questions. We will fix the mentioned typos and update the Figure captions to provide further clarity. Below we answer the reviewer’s remaining questions.
**Q: Can you highlight the main no... | null | null | null | null | null | null |
(Amplified) Banded Matrix Factorization: A unified approach to private training | Accept (poster) | Summary: Applying matrix factorization mechanism and balancing tradeoffs of the mechanism in differential privacy is a long living issue. This work constructs MF mechanism with banded matrices for both centralized and federated training setting across all privacy budgets. For federated setting, this work is compatible ... | Rebuttal 1:
Rebuttal: Thank you for your detailed review.
>**(1) Advantages and disadvantages of BandMF in FL:**
Most importantly, the only prior unbanded approach to be applicable to cross-device FL is the single-epoch work, which we significantly outperform Empirically, we show this on SO-NWP, a large-scale, non-... | Summary: In this paper, the authors show how Matrix Factorization (MF) can subsume prior
state-of-the-art algorithms in both federated and centralized training settings, across all privacy
budgets. They apply the key technique: MF mechanisms with banded matrices. For both the
cross-device federated learning setting and... | Rebuttal 1:
Rebuttal: Thank you for your review and positive feedback.
Our main results for the cross-device FL setting are written in sections 3 and 4. To make this clearer, we have rewritten our contributions section in the introduction to clearly state the contributions separately for cross-device FL and for centra... | Summary: This paper proposes banded matrix factorization for differential privacy. It shows this new mechanism can be effectively applied to centralized and federated-learning settings (where individuals can choose when and how many times to participate in training). The main technical part is the $b$-minsep-participat... | Rebuttal 1:
Rebuttal: Thank you for your detailed review, and for providing line numbers in your review. We greatly appreciate it!
>**$\hat{b}$ and b:**
Thank you for highlighting this issue. We now informally define $\hat{b}$-banded matrices in the abstract, and clarify the difference between b-min-sep and $\hat{b... | Summary: The authors present a novel mechanism explicitly designed for differentially private training. The mechanism considers the sensitivity of different participation schemes in the context of fixed datasets during differentially private training. The key contributions of this work can be summarised as follows:
1. ... | Rebuttal 1:
Rebuttal: Thank you for your positive feedback, comments, and questions.
>**Make Paper Self-Contained:**
Thank you for these suggestions. We used the additional page to add in an informal proof to Theorem 2 and to add algorithm 4 to the main-text, with a description of how it works. We also have expanded... | Rebuttal 1:
Rebuttal: Thank you for your time in reviewing our work.
>**On intuition and improved organization.**
We have revised the paper to briefly describe the matrix mechanism in the first paragraph, and have additionally included a new figure to help provide intuition for the advantage of MF-DP-FTRL, as well as... | NeurIPS_2023_submissions_huggingface | 2,023 | Summary: The paper studies the problem of how to optimize the Matrix factorization (MF) mechanisms so that the effect of random noise can be minimized. The MF mechanisms can be applied in addition to the well-known DP-FTRL or similar online DP algorithms in machine learning training. This technique decomposes the query... | Rebuttal 1:
Rebuttal: Thank you for your comments, suggestions and questions.
>**MF may be ambiguous**
This is an excellent point that we missed. We have added the following footnote on the first page: “
“We use matrix factorization for the algorithmic design of DP mechanisms, and then use these mechanisms in genera... | null | null | null | null | null | null |
Leveraging Pre-trained Large Language Models to Construct and Utilize World Models for Model-based Task Planning | Accept (poster) | Summary: This paper proposed to apply GPT LLMs to perform PDDL model construction from natural language description and correction with feedbacks, and then use classic planners or LLM planners to construct plans from the PDDL models. Experiments were conducted in the synthetic domains to verify the proposal and found t... | Rebuttal 1:
Rebuttal: > 1. Clarification on the problem setting (Have you deployed the plans to real-world fetch manipulator? If you did, how does it work? What will be the solution to connect PDDL plans to real world execution?)
We are looking at the **model-based task planning** problem, where the main goal is to ge... | Summary: This paper proposes a translation approach to using LLMs for planning. Instead of relying on LLM prompt completion to generate plans, this approach parses a written description of the domain into a valid PDDL domain description. The user can then opt to use the PDDL domain description with an automated PDDL pl... | Rebuttal 1:
Rebuttal: > 1. Some papers in the LLM + symbolic planning
In fact, we cite both papers and discuss them in detail in our submission:
(a) LLM+P: starting from line 346, we explained the fact that LLM+P only uses LLMs for translating user instruction into PDDL format. In contrast, our work does much more t... | Summary: There's been a fair amount of recent work on using LLMs to directly write plans for planning problems. In contrast, this work divides the planning process into two stages where first an LLM is used to translate natural language descriptions of actions to formal PDDL actions, then goals can be planned towards u... | Rebuttal 1:
Rebuttal: > 1. Involvement of humans: the human feedback seems necessary given the errors GPT-4 introduces, and presumably this method wouldn't work well at all without that step of human correction.
As in any use case of LLMs, LLMs are only **approximately omniscient**, and without external intervention, ... | Summary: The submission investigates whether LLMs are able to create PDDL domains and problems description in three domains, two standard planning benchmarks, and one more challenging domain. The paper considers two related to the title: a) direct construction of the PDDL with an LLM, b) correcting errors in PDDLs. The... | Rebuttal 1:
Rebuttal: > 1. Both the planning benchmarks and the household domain might actually be easy for an LLM than domains in practical applications. That should be discussed further in the paper.
We would like to highlight that our evaluation domains pose significant challenges for LLMs (if we employ them direct... | null | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
A Novel Framework for Policy Mirror Descent with General Parameterization and Linear Convergence | Accept (poster) | Summary: This paper proposes a policy update scheme based on mirror descent for general function approximation. The paper also provides corresponding theoretical analysis on the sub-optimality gap and computation cost. Overall, the reviewer find the theory part not strong enough to prove the advantage of the proposed n... | Rebuttal 1:
Rebuttal: Please find below an answer to the issues you have raised regarding the significance and soundness of our work. Due to space constraints, please find the bibliography for this reply in the reply to Reviewer qVps.
**Characterization of $\epsilon_\mathrm{approx}$, $\nu_\mu$ and $C_v$ in Theorem 4.3... | Summary: While policy mirror descent methods have been theoretically analyzed in many past works to establish strong convergence guarantees, questions remain regarding the implementation of function approximation in this class of algorithms. The work analyzes general approximators with PMD, obtaining linear convergence... | Rebuttal 1:
Rebuttal: Thank you for your positive and insightful remarks. Please find the answers to your questions in Experimental results for AMPO and
below.
**Q1: $\omega$-potential class.** We are aware of two previously used mirror maps that cannot be recovered using $\omega$-potentials: $h(x) = \frac{1}{2}x^\top... | Summary: This paper extends the recently introduced policy mirror descent method of tabular setting to function approximation setting. The essential development seems to reside in exploiting the mirror descent update in the dual space and projecting the dual representation of the policy into a "realizable" representati... | Rebuttal 1:
Rebuttal:
Thank you for your positive and insightful remarks. Please find the answers to your questions in Experimental results for AMPO and
below.
**Tsallis divergence.** We thank the reviewer for pointing out the missing prior work [1] regarding the use of Tsallis entropy. We include it in Line 218 in t... | Summary: This is a theoretical work and the main claim is the proof of linear convergence rate for policy mirror descent (PMD) algorithms with general policy parametrization. This contrasts with previous results (Lan 2022, Xiao 2022, Cen et al 2021, Zhan et al. 2021, Cayci et al. 2021, etc.) that proved convergence ra... | Rebuttal 1:
Rebuttal:
Thank you for your review and remarks. Please find the answers to your questions in Experimental results for AMPO and below.
**Interpretation of the theory.**
An interpretation of our theory can be provided by connecting AMPO to the Policy Iteration algorithm, which shares the linear convergence... | Rebuttal 1:
Rebuttal:
**Thank you** We would like to thank the reviewers for taking the time to review our paper and for their careful remarks.
In particular, we thank Reviewer **qVps**, **rns2**, **oNAx** and **zS6s** for recognizing the importance of our contributions, considering that our work is substantial and p... | NeurIPS_2023_submissions_huggingface | 2,023 | Summary: This paper presents a novel framework for analyzing and understanding policy optimization based on mirror descent that can leverage any general parameterizations of the policy class. The new framework induces a new update rule based on mirror descent, and the authors proved the new framework enjoys sublinear a... | Rebuttal 1:
Rebuttal: Thank you for your positive and encouraging remarks. Please find the answers to your questions in Experimental results for AMPO and below.
**Application of AMPO to online/offline RL and other settings.**
AMPO can be applied to online and offline RL settings, and has a great potential to be applie... | null | null | null | null | null | null |
Quasi-Monte Carlo Graph Random Features | Accept (spotlight) | Summary: Kernel methods are important in many ML applications, but they suffer from scalability issues. In the Euclidean setting, random features methods (Rahimi & Recht NeurIPS 2017) can be used to "sketch" the kernel matrices and speed up computations. In the graph setting, however, no such random feature seems to ha... | Rebuttal 1:
Rebuttal: We thank the reviewer for reading the manuscript, and are delighted that they note the work’s novelty, clarity and strong experimental contributions. We agree that it is remarkable that such a conceptually simple technique to correlate walker lengths can lead to such rich behaviour and demand such... | Summary: The author(s) extend the work of Choromanski [2023] and introduce quasi monte carlo graph random features.
To estimate the 2-regularized Laplacian kernel, they now use random walk features for each node which have coupled probabilities of terminating after a certain number of steps.
This increases diversity in... | Rebuttal 1:
Rebuttal: We thank the reviewer for their careful reading of the manuscript and thoughtful feedback. We are delighted that they note this novel mechanism to increase the diversity of random walks is simple to implement and offers strong theoretical guarantees and empirical performance – in fact not only for... | Summary: This work proposes an efficient random-walk sampling approach that accelerates the estimation of the feature mapping for 2-regularized Laplacian kernels. The random walk sampling is based on antithetic termination, which is a variance reduction technique and sounds novel when being applied to the estimation o... | Rebuttal 1:
Rebuttal: We sincerely thank the reviewer for their encouraging comments, especially pertaining to the novelty of the mechanism, the quality of the writing and the exhaustive nature of the experimental section. Below, we address minor points of misunderstanding and answer the reviewer’s questions.
1. *Sco... | Summary: The paper introduces a method for graph random features (GRFs). Graph random features is the graph equivalent (i.e. kernels defined on graphs) of kernel matrix approximation using Random Features (Rahimi & Recht). Recently Chromanski et al. had proposed a method for GRF which uses a series of random walks.
I... | Rebuttal 1:
Rebuttal: We thank the reviewer for their reading of the manuscript and positive comments. We are delighted that they recognise the scheme’s originality, strong practical impact and broad downstream applications. They note that the work is well-motivated and the manuscript is clear. We address their questio... | Rebuttal 1:
Rebuttal: We are grateful for the reviewers’ careful readings of the manuscript and thoughtful feedback. We are delighted to receive such positive comments about the elegance of the mechanism and our exhaustive theoretical and empirical analysis. They also make several helpful suggestions which have further... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
One-Line-of-Code Data Mollification Improves Optimization of Likelihood-based Generative Models | Accept (poster) | Summary: This paper proposes the use of data mollification to improve sample quality from likelihood-based generative models. The authors observe that likelihood-based models are worse-performing than score-based models due to manifold overfitting (i.e., learning the manifold but not the distribution on it), and poor d... | Rebuttal 1:
Rebuttal: **Q15. My biggest concern is that the image datasets considered might be too simple ... One potential way to do this is to study, e.g., ImageNet 64x64...**
Thank you for your comment. Unfortunately, prior to submission, we faced limitations in compute capacity, preventing us from exploring such a... | Summary: In this paper, the authors address the limitation of likelihood-based Generative Models (GMs) in achieving high-quality samples compared to state-of-the-art score-based Diffusion Models (DMs). They propose a novel approach to enhance the generation quality of GMs by incorporating data mollification, a techniqu... | Rebuttal 1:
Rebuttal: **Q11. The approach described in the paper lacks explicit discussion regarding its relationship to noise augmentation. However, towards the end of section 3, it appears that the authors have chosen a temperature factor of 0.7, which leads to a substantial mollification towards the end of the train... | Summary: The paper proposes an optimization method that improves the training of likelihood-based generative models. The central idea is that in the early phases of the optimization, the learner tries to model a smoothed version of the likelihood function. As the optimization proceeds, the smoothening decreases (and th... | Rebuttal 1:
Rebuttal: **Q4. ... I do not see how this particular optimization method could be improved to achieve these results.**
We agree with Reviewer that our method does not necessarily help likelihood-based generative models **(GMs)** reach the performance level of diffusion models **(DMs)**, as this heavily rel... | Summary: The paper proposes a idea to borrow one of the strengths of score-based DMs, which is the ability to perform accurate density estimation in low-density regions and to address manifold overfitting by means of data mollification. We connect data mollification through the addition of Gaussian noise to Gaussian ho... | Rebuttal 1:
Rebuttal: **Q1. The first concern of this paper is its novelty and it is very critical for conference like Neurips. It basically states the same thing as in "Generative modeling by estimating gradients of the data distribution", a NeurIPS 2019 oral paper. Due to the identical proposition and analysis of gen... | Rebuttal 1:
Rebuttal: Dear Reviewers,
We thank the Reviewers for careful reviews of our paper and for insightful comments, which have helped us improve the paper significantly. We are encouraged by the endorsements in the initial reviews that: 1) **the paper is very well written** (Reviewer sZKb and yxxe); 2) **the pr... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Decision Stacks: Flexible Reinforcement Learning via Modular Generative Models | Accept (poster) | Summary: This paper focuses on solving offline RL with generative models. Concretely, it proposes a method to modularize the joint distribution of time-induced trajectories and use separate generative models to represent observation module, reward module, and action module. Evaluations are conducted on D4RL benchmark w... | Rebuttal 1:
Rebuttal: Thanks for the valuable comments and very detailed insights! We address your questions below.
> Q1: “Novelty of the proposed method is limited. The improved performance can also be attributed to other confounding factors, such as larger models due to reward and action modules being factored out. ... | Summary: This paper highlights a drawback in prior frameworks, such as Decision Transformer and Diffuser, where the absence of modular hierarchies among different tokens results in limited expressivity and flexibility. To overcome this issue, the paper introduces Decision Stacks (DS), a modular algorithm designed for l... | Rebuttal 1:
Rebuttal: Thanks for the valuable comments and very detailed insights! We address your questions below.
>Q1: “One major concern regarding DS is the high complexity, as it requires training three generative models...the performance improvement compared to other baselines that use only one generative model is... | Summary: This paper proposes to tackle the problem decision making using a stack of different generative models. The first generative model constructs a conditional distribution over observations. A subsequent generative model constructs a conditional distribution over rewards, with a final generative model constructs ... | Rebuttal 1:
Rebuttal: Thank you for your valuable comments and very detailed insights! We address your questions below.
>Q1: The results illustrated in the approach are a bit toy -- with the largest gains in the Maze2D environment. The Mujoco control environment has somewhat limited gains and the constructed POMDP env... | Summary: This paper proposes to disentangle the different modalities (reward, observations, and actions) utilized in offline goal-conditioned reinforcement learning instead of the standard temporal single-module structure. The paper contributes three independent generative modules that have the benefit of parallelizabl... | Rebuttal 1:
Rebuttal: Thank you for the valuable comments and very detailed insights! We address your questions below.
>Q1: “...the training cost comparison to the other algorithms.”
We provide a comparison in Table 4 of the rebuttal PDF, showcasing the per-iteration cost. DS's training time is influenced by the choic... | Rebuttal 1:
Rebuttal: We thank all the reviewers for providing insightful feedback and constructive suggestions, including recommendations for new ablation studies. In alignment with these insightful suggestions, we have conducted and provided 4 new experimental results that clarify Decision Stacks's novelty and addres... | NeurIPS_2023_submissions_huggingface | 2,023 | Summary: The paper proposes using different generative models (transformer or diffusion) rather than the same model (like in trajectory transformer/decision diffuser) for observation prediction, reward estimation, and action prediction in “model-based” offline RL. They demonstrate that this flexibility improves perform... | Rebuttal 1:
Rebuttal: Thank you for your valuable comments and very detailed insights! We address your questions below.
>Q1: Similar performance to DD on D4RL Gym tasks.
1. This similarity, rather than reflecting a limitation of our method, highlights the near-saturated state of the D4RL benchmark, where many methods ... | null | null | null | null | null | null |
Towards Foundation Models for Scientific Machine Learning: Characterizing Scaling and Transfer Behavior | Accept (poster) | Summary: This work explores and demonstrates how a "pretraining+finetuning" paradigm may be leveraged for neural networks applied to PDEs. Specifically, for three different systems governed by distinct PDEs, the authors study how different factors influence performance: (1) downstream dataset scale; (2) model scale; (3... | Rebuttal 1:
Rebuttal: We thank the reviewer for their detailed feedback, thorough assessment and valuable comments/suggestions regarding our paper.
### R4-1: W1: Additional results for SYS-3/SYS-2 to be more systematic
That is a good point. We should mention that SYS-3 (helmholtz) was chosen specifically as a system ... | Summary: This paper investigates the finetunability, transfer, and generalization properties of a foundational model applied to the field of scientific machine learning, as functions of finetuning dataset size, pretrained model size, and value of physics parameters governing the equations of the system.
Strengths: Th... | Rebuttal 1:
Rebuttal: We thank the reviewer for their positive feedback and constructive criticisms and address specific questions and suggestions below:
### R3-1: Exploring only one architecture:
We agree with this limitation of using just the FNO and make this point in the paper. We are happy to expand on this more.... | Summary: This paper analyzes the transfer performance of the pre-trained foundation model for scientific machine learning applications described by PDEs. A wide range of settings is considered: various sizes of downstream datasets, pretrained model, out-of-distribution data, as well as multi-tasks and applications. Con... | Rebuttal 1:
Rebuttal: We appreciate the positive feedback and constructive criticisms from the reviewer. While we are happy there are no major concerns, we address some of the specific suggestions below.
### R2-1: Exploring only one architecture:
We agree this is a major limitation of our work, and we call this out i... | Summary: This paper explores the behavior and performance of neural operator models on multiple partial differential equation (PDE) systems in the transfer learning setting. The authors investigate the impact of various factors on the performance of neural operator models, including model size, downstream dataset size,... | Rebuttal 1:
Rebuttal: We thank the reviewer for their thorough assessment and kind words regarding our work. The suggestions they provide are very helpful, and we address specific comments below:
### R1-1: Motivation for understanding the transfer learning (TL) for PDEs in SciML:
We agree that the introduction can sta... | Rebuttal 1:
Rebuttal: We thank all reviewers for their insightful comments and suggestions. We attach additional figures for the model scaling experiments for Helmholtz and the additional data points for the extreme OOD test-case addressing R4-1 comments in Fig 1 and Fig 2, respectively. We refer the reviewers to the i... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
CycleNet: Rethinking Cycle Consistency in Text-Guided Diffusion for Image Manipulation | Accept (poster) | Summary: The paper proposes CycleNet, a method for unpaired image-to-image translation using pretrained diffusion models. The main idea is to reconstruct the conditional images through a reverse process. The idea seems reasonable, and the paper is well-organized, with sufficient experimental results.
Strengths: The id... | Rebuttal 1:
Rebuttal: We are happy that the reviewer found our paper "well-organized, with sufficient experimental results.". We appreciate their constructive feedback on our efforts.
### Question 1
> Q1: The concept of cycle construction has been extensively studied in the field, and this work follows a similar idea,... | Summary: The paper introduces CycleNet, a new method that enhances image manipulation by incorporating cycle consistency into diffusion models. The paper addresses the challenge of unpaired image-to-image translation and aims to provide a consistent and intuitive interface for this task.
Strengths: 1.Originality: The ... | Rebuttal 1:
Rebuttal: We are happy that the reviewer finds our paper “well-written”, and that “the results also support the claims of the paper”. We are more than grateful for the reviewer’s feedback, many of which has been integrated into our paper! We sincerely hope there will be engaged communication during our disc... | Summary: This work aims at addressing the task of unpaired I2I translation with pre-trained diffusion models. Inspired by CycleGAN, authors incorporate cycle consistency into diffusion model to regularize process of image translation and proposed CycleNet. In addition it also contributes a multi-domain I2I translation ... | Rebuttal 1:
Rebuttal: We are happy that the reviewer found it “interesting and reasonable to introduce cycle consistency into diffusion models”. We thank the reviewer for recognizing our novelty, our “extensive experiment”, and our contributions (out-of-domain generalization, new multi-domain dataset). We appreciate th... | Summary: This paper presents a method that tackles the challenge of consistent image-to-image (I2I) translation using diffusion models (DMs) and unpaired image conditioning and text prompts. The proposed approach incorporates Cycle Consistency Regularization to ensure cycle consistency and Self-Regularization to genera... | Rebuttal 1:
Rebuttal: We thank the reviewer for recognizing our dataset contributions and our model’s capability to preserve the structures of input images. We are happy that much of the constructive feedback has been integrated into our paper!
### Question 1
> Q1: The experimental results are quite strange. In Figure... | Rebuttal 1:
Rebuttal: ### General response 1: Hue-Shifting and Sub-Optimal Performance on ManiCups
We thank the reviewers for pointing out this hue-shifting issue. We address this with a simple but effective solution at inference time and the updated results are as follows.
**Method**: Previously, when we are samplin... | NeurIPS_2023_submissions_huggingface | 2,023 | Summary: This paper introduces cycle consistency in diffusion models to achieve regularization in image manipulation. It allows out of domain image generation with text prompt modification, and can be trained with very little data (~2K) with minimal computational requirements (1GPU). When we apply pre-trained DM models... | Rebuttal 1:
Rebuttal: We thank the reviewer for recognizing our novelty (“there has not been much regularization in diffusion models”) and contributions (out-of-domain generalization, minimal data, and compute). We appreciate their constructive feedback on our efforts.
### Question 1
> Q1: While enforcing cyclic consi... | null | null | null | null | null | null |
State-wise Constrained Policy Optimization | Reject | Summary: This paper discusses an important topic about safe reinforcement learning, which explores the state-wise issue. It is a significant problem because, in the real world, state-wise constraints are one of the most common and challenging constraints in safety-critical applications. Most safe RL methods focus on cu... | Rebuttal 1:
Rebuttal: **W1**: In continuous control challenges, such as robot locomotion tasks, performance convergence tends to cluster around a similar magnitude, as rewards have constrained upper bounds due to episodic nature. As long as the goal can be achieved within an episode, performance remains consistent acro... | Summary: This work tackles the problem of state-wise safety in the reinforcement learning problem. To this end it introduces the framework of Maximum Markov Decision Process and an algorithm State-wise Constrained Policy Optimization (SCPO) to solve the problem. Numerical results illustrate the performance of the autho... | Rebuttal 1:
Rebuttal: **Q1**: Thank you for bringing this to our attention! We mentioned from line 141 to line 144 that enforcing constraints on every state-action transition pair in SCMDP, i.e. eq. (5), is challenging. As an alternative, we propose constraining the maximum state-wise cost along the trajectory, whic... | Summary: This paper introduces a novel approach to solve safe RL tasks. This approach is based on a new method SCPO and corresponding MMDP framework. Authors show the efficacy of this method by providing mathematical guarantees. Additionally, authors give useful practical implementation tips to improve reproducibility ... | Rebuttal 1:
Rebuttal: Thank you for your valuable feedback on our paper. We sincerely appreciate your time and effort in reviewing our work. Please find our answers to your questions below.
**Q1**: “Not necessarily a weakness, but including intuitive explanation of proposition 1 and 2 would be helpful to increase over... | Summary: This paper introduces a novel framework called Maximum MDP to address the problem of state-wise constrained policy optimization, namely the authors considers limiting the expected maximum state-wise cost rather than the cost for each state. Similar to the TRPO/CPO framework, the authors derived a worse-case co... | Rebuttal 1:
Rebuttal: We sincerely thank you for your comprehensive comments on our paper and please find our answers to your questions below.
**W1**: We totally agree that our Theorem 1 doesn't really depend on MMDP. Thanks very much for pointing that out! We will definitely add an explanation about this in the latte... | Rebuttal 1:
Rebuttal: We sincerely thank all reviewers for all the detailed and helpful reviews!
We have revised the paper to address all the important issues raised in the reviews. Among all the reviews, there are some important questions about our paper we would like to highlight the answers here.
**Q1**: Alternat... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
A Regularized Conditional GAN for Posterior Sampling in Image Recovery Problems | Accept (poster) | Summary: This paper introduces a regularization term, comprising an L1 penalty and a standard-deviation reward, for the conditional GAN with Wasserstein loss. The objective is to ensure the generation of high-quality samples for a specific input observation y. The proposed approach is evaluated on two image recovery ta... | Rebuttal 1:
Rebuttal: - **Weakness**:
Novelty is a bit limited. Only added a new reward term.
**Response**:
Please see the global rebuttal.
- **Weakness**:
Only linear inverse problems were demonstrated. How would it perform on nonlinear ones?
**Response**:
Our method should work equally well for nonlinear inv... | Summary: This work uses conditional GANs for solving image recovery problems including face-inpainting and accelerated MRI. In particular, a regularization scheme is proposed that tackles the lack-of-diversity issue related to posterior sampling from cGANs. The authors show that the proposed regularization allows sampl... | Rebuttal 1:
Rebuttal: - **Weakness**:
The experiments did not demonstrate fairness or classifier calibration advantages due to posterior sampling. Also, it's not clear how a radiologist would benefit from a pixel-wise standard-deviation map.
**Response**:
Our goal for this paper was to design a fast and accurate po... | Summary: The paper presents a new method based on conditional GANs (cGANs) for generating posterior samples for solutions of inverse problems. The motivation comes from the fact that standard approaches to inverse problems typically use point estimates, which due to the perception-distortion tradeoff will lead to blurr... | Rebuttal 1:
Rebuttal: - **Weakness**:
At 8x acceleration, it is possible to find a number of features and fine details that are missing, so I doubt these would be accepted for clinical use.
**Response**:
8x acceleration is perhaps too aggressive for clinical practice, but keep in mind that we are zooming the plots si... | Summary: This paper proposed a cGAN (conditional generative adversarial network) method for general image inverse problems (IR). Conventional cGAN methods for IR usually suffer from mode collapse that the generators learned to ignore latent inputs and are unable to generate diverse image reconstruction samples. Motivat... | Rebuttal 1:
Rebuttal: - **Weakness**:
Adding extra regularization to the GAN loss is not completely new.
**Response**:
Please see the global rebuttal.
- **Weakness**:
No discussion of Monte-Carlo dropout or deep ensembles.
**Response**:
Posterior samplers aim to accurately sample from the true posterior $p(\boldsy... | Rebuttal 1:
Rebuttal: We sincerely thank the reviewers for the time spent reading and reviewing our manuscript. The feedback is very useful and will help to improve the quality of the final paper if it is accepted. In this global rebuttal we respond to questions that were raised by multiple reviewers.
- **Weakness**... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Joint Learning of Label and Environment Causal Independence for Graph Out-of-Distribution Generalization | Accept (poster) | Summary: This work addresses the issue of graph out-of-distribution (OOD) generalization by considering the data generation assumption of the presence of a causal subgraph and a spurious subgraph within each graph, and the assumption of label and environment causal independence. The paper proposes an adversarial traini... | Rebuttal 1:
Rebuttal: Thank you for your thoughtful feedback and comments. We have incorporated your suggestions to refine the robustness and clarity of our paper. Below, we provide detailed answers to your questions in order to address your concerns listed in weaknesses.
## Questions:
1. We deeply appreciate your re... | Summary: This paper proposes a novel adversarial approach called LECI, for OOD generalization. The paper leverages label and environment information to identify the causal and invariant graph, thereby enhancing the accuracy and reliability of causal and invariant subgraph identification. The authors conduct extensive e... | Rebuttal 1:
Rebuttal: We sincerely appreciate the effort and time you have dedicated to reviewing our paper. We understand that some aspects may have led to certain misconceptions, and it is our foremost intention to address and clarify these points. We hope this further elucidation will assist in reevaluating our work... | Summary: The paper proposed a novel learning strategy to incorporate label and environment causal independence (LECI) for tackling the graph OOD problem. Enforcing such independence properties can help exploit environment information and alleviate the challenging issue of graph topology shifts. With good empirical resu... | Rebuttal 1:
Rebuttal: We appreciate your valuable comments and feedbacks! We have carefully considered each one and have endeavored to address your concerns. We believe that our responses will provide the necessary clarifications, and we look forward to your feedbacks.
## Weaknesses:
> I am puzzled for "graph-specifi... | Summary: This paper proposes a new graph learning model to incorporate label and environment causal independence (LECI) for solving the graph OOD generalization problem. The authors first introduce two causal independence properties to distinguish causal and spurious subgraphs. Enforcing environment causal independence... | Rebuttal 1:
Rebuttal: We appreciate your helpful feedbacks!
## Weaknesses:
> LECI assumes the environment information is known, which is a tight requirement. The authors argue that the environment labels can be accessed by applying simple groupings or deterministic algorithms, but in this case, the obtained labels ar... | Rebuttal 1:
Rebuttal: # Rebuttal of LECI
## The strengths of this paper
We would like to extend our gratitude to all the reviewers for their invaluable time and insightful feedback. It is encouraging to observe that our contributions and strengths resonated positively with you:
- Novelty (Reviewer tVF4, YxXK, Csim)
... | NeurIPS_2023_submissions_huggingface | 2,023 | Summary: This papers considers the graph OOD generalization problem with given environment information. It proposes a novel invariant learning method based on adversarial loss which is derived from independence conditions of the causal graph. Experimental results show the proposed method is very effective.
Strengths:... | Rebuttal 1:
Rebuttal: Thank you for your valuable time and insightful comments!
## Weaknesses:
> In fact, the proposed method ... the claim may not be 100% true.
**Response:** As highlighted on line 564 of the appendix, the mapping from $(X,A)_C \mapsto X_C$ is non-unique. While it's conceivable to adapt a feature-... | null | null | null | null | null | null |
Online List Labeling with Predictions | Accept (spotlight) | Summary: This paper is about the classic Online List Labeling problem in the recent model of algorithms with predictions (a.k.a. learning-augmented algorithms).
In this problem n orderable items (say, integers) arrive over time. We need to keep them in the sorted order in an array of size c*n, for some constant c. Whe... | Rebuttal 1:
Rebuttal: Thank you for the thoughtful review. We greatly appreciate your comments on writing and will incorporate them in the final version of the paper.
**Question:** would it be feasible to prove bounds in terms of average L1 prediction error per item, instead of L_inf error? In other words, is it rea... | Summary: The author presented a novel learning-augmented algorithm for the online list labeling problem, providing solid theoretical guarantees in terms of the error in the predictions. They also investigated the stochastic error model and bound the performance in terms of the expectation and variance of the error. Fin... | Rebuttal 1:
Rebuttal: Thank you for the thoughtful review.
**Question:** What other data structures will likely benefit from an approach similar to the one you used in this work?
**Answer:** We believe that the model and general algorithmic ideas will be useful for some other data structures. We have initial res... | Summary: Author study a fundamental online problem which is also important
in practice: online list labeling.
The introduce rank predictions for this problem and define prediction error
in a natural way. They achieve O(1) consistency and a robustness comparable
with the best possible classical online algorithm.
They sh... | Rebuttal 1:
Rebuttal: Thank you for the thoughtful review.
**Question:** consistency is $O(1)$ but not really close to $1$.
**Answer:** If the predictions are perfect, then every element is inserted and never moves. This gives consistency $1$. We will clarify this in the paper.
---
Rebuttal Comment 1.1:
Comment... | Summary: The paper considers the online list labelling problem: A set of $n$ elements arrive online and have to be inserted into an $c\cdot n, c>1$ constant, large array while the array must at all times be sorted. Every time an already inserted element is "shifted" in order to maintain the order, it incurs a cost of $... | Rebuttal 1:
Rebuttal: We thank the reviewer for the helpful comments. We address your questions below.
**Question:** saying that the data structure is "optimal for any prediction error" is not true. For one I think the authors use "optimal" to mean "best possible" which differs from the convention that optimal ref... | null | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Thin and deep Gaussian processes | Accept (poster) | Summary: Techniques such as Deep Kernel Learning and extensions such as Deep GPs have been proposed in order to overcome the flexibility constraints associated with standard shallow GPs. However, although these techniques are better suited to model non-stationary data, they are still susceptible to issues such as patho... | Rebuttal 1:
Rebuttal: Thank you for your detailed feedback. We have run additional experiments to answer your questions (below). Besides including the additional results and complexity analysis to the final manuscript, we will also add a discussion on which domains we expect TDGP to be most useful, as suggested.
>The... | Summary: This paper extends the CDGP to address two weaknesses in deep Gaussian processes. The solution is similar intuitively to a residual connection. Instead of letting each layer depend only on the last layer, they let each layer depend on the last layer and the input. This allows the model to induce a manifold and... | Rebuttal 1:
Rebuttal: Thank you for your suggestions to improve our manuscript. We agree that elaborating further on the interpretation of the experiments can make it more accessible to the general public. In the next paragraphs, we elaborate further on the importance of both types of representation (manifold/latent sp... | Summary: This paper considers the problem of non-stationary kernel design for Gaussian processes. Two main approaches for this problem are deformation kernels and length-scale mixture kernels. On one hand, deformation kernels, eg deep GPs, have found great success while trading off expressivity and the ability to learn... | Rebuttal 1:
Rebuttal: Thank you for the comments.
We would like to highlight that the bias term is a simple design choice (used profusely throughout ML) that suffices to show TDGP is at least as expressive as CDGP. Both CDGP and TDGP with no bias are limiting cases of TDGP with bias.
> Have you tested the method wit... | Summary: The paper presents a new deep Gaussian process (DGP) model, the thin and deep GP (TDGP), which does not suffer from a diminishing signal as the number of layers increases. The crux of the TDGP model is the covariance function that, for each layer $\ell$, acts on a linear combination of a (non-linear) transform... | Rebuttal 1:
Rebuttal: Thank you for the review. We are glad you found our paper to be interesting, significant, and of high quality. We appreciate your concerns and questions and hope you will find them addressed in the following:
## Approximate inference and kernel choice (W2, Q3)
We derived our ELBO in Appendix B.3 f... | Rebuttal 1:
Rebuttal: We thank the reviewers for their thoughtful feedback, as we believe addressing them will broaden the audience for this paper. We are also glad that most reviewers feel our ideas are clearly presented and constitute a solid and principled contribution to this problem space.
In particular, the cons... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Pre-RMSNorm and Pre-CRMSNorm Transformers: Equivalent and Efficient Pre-LN Transformers | Accept (spotlight) | Summary: The paper analyses the two common types of normalization layers in Transformers: LayerNorm and RMSNorm (root-mean-square-norm). LayerNorm scales all vectors to be of the same norm, while changing the vectors' "directions". RMSNorm, in contrast, keeps the same direction, but just rescales the vectors to the sam... | Rebuttal 1:
Rebuttal: Thanks a lot for your insightful comments. Due to the word limit, we have to redirect you to our response to other reviewers. We sincerely appreciate your understanding.
### **Question 1**
The original Transformer block is
* $x_{l+1}=x_l+F_l(x_l)$
* $x_{l+2}=x_l+F_l(x_l)+F_{l+1}(x_l+F_l(x_l))$
... | Summary: This study explores the relationship between Pre-RMSNorm and Pre-LN Transformers, demonstrating that these two variants can be theoretically reparameterized into one another. Additionally, the authors introduce a novel Transformer variant called Pre-CRMSNorm, which reduces one hidden dimension while maintainin... | Rebuttal 1:
Rebuttal: Thank you sincerely for your insightful feedback. We appreciate the opportunity to address your questions in this response.
### **Weakness 3, reparameterization and equivalence**
We would like to initially clarify these two terms. Let $f$ be a Pre-LN Transformer with parameter $\theta$, and $g$ b... | Summary: This paper proposes modifications to the popular transformer architecture's normalization mechanism in order to improve efficiency without sacrificing performance. It starts out with the baseline architecture Pre-LN (LayerNorm), and derives two new architectures: Pre-RMSNorm and Pre-CRMSNorm, which are inspire... | Rebuttal 1:
Rebuttal: Thank you so much for your insightful reviews. We are truly grateful for your acknowledgment of our paper's excellence in terms of soundness, presentation, and contribution.
We totally agree that our relative efficiency improvement is modest. As you correctly pointed out, our quantitative analysi... | Summary: The paper proposes two novel modifications to the Pre-Layer Normalization (Pre-LN) Transformers, introducing the Pre-Root Mean Square Normalization (Pre-RMSNorm) and Pre-Compressed Root Mean Square Normalization (Pre-CRMSNorm) Transformers. The authors aim to improve computational efficiency by simplifying Lay... | Rebuttal 1:
Rebuttal: Thank you so much for your insightful comments. We appreciate the opportunity to address your comments in this response.
### **Question 1, more empirical results on various domains and data types**
**Domains.**
We present experimental results on ViT and GPT3, covering the field of computer visio... | Rebuttal 1:
Rebuttal: We sincerely appreciate the constructive feedback provided by all the reviewers. We are truly grateful for their acknowledgment of our paper's contribution and presentation. Engaging with the reviewers is a privilege we highly cherish. Should any queries arise, please do not hesitate to reach out ... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
SparseProp: Efficient Event-Based Simulation and Training of Sparse Recurrent Spiking Neural Networks | Accept (poster) | Summary: This paper introduces an efficient Spike-based recurrent network learning simulation method based on an event-driven implementation.
The method is based on two main contributions. The first consists in a change of temporal reference frame that uses the phase of a neuron rather than its absolute time, the seco... | Rebuttal 1:
Rebuttal: We thank the reviewer for their insightful and encouraging assessment and agree that SparseProp could be a game changer for the spiking network community, especially, for very large, sparse networks.
Here are our more detailed responses:
> Minor: caption of Fig 4 : repetition of « higher densiti... | Summary: Simulation of RSNN (recurrent spiking neural network) is a difficult task, whose limits prevent large scale (say 1e.9 neurons) simulations. Often, in practice, one neuron is connected to only a few others, the sparsity. The present paper leverages this sparsity to upscale the simulations limits of RSNN. Sparsi... | Rebuttal 1:
Rebuttal: We thank the reviewer for their time and helping us improve our work. We will integrate these ideas into our manuscript with care. Here are our detailed responses:
> General comment on citations: the authors could give more references, for example: -line 59: several classical papers showing the o... | Summary: This paper presents an efficient event-based algorithm named *SparseProp* for both inference and learning of sparse SNNs. It gets rid of the discretized simulation of ODEs in neuronal dynamics and utilizes phase representation for those neuron models with closed-form solutions between spike times. A priority q... | Rebuttal 1:
Rebuttal: We thank the reviewer for the detailed and insightful feedback and request for clarifications, which substantially improved the publications.
Here are our more detailed responses to the questions:
> The whole study seems to assume a positive external input, i.e. in Line 155/544 and a special set... | Summary: This paper proposes SparseProp, an event-based algorithm for efficient simulation and training of sparse Spiking Neural Networks (SNNs). By leveraging the sparsity of the network, SparseProp reduces the computational costs of operations from O(N) to O(log(N)) per network spike. This method avoids iterating thr... | Rebuttal 1:
Rebuttal: We thank the reviewer for the thoughtful comments, which substantially improved the manuscript. As requested, we are working towards evaluating SparseProp on real-world scenarios, and we added clarifications on parallelization and comparison to SOTA SNN algorithms that use the Euler method to simu... | Rebuttal 1:
Rebuttal: We deeply appreciate the feedback and constructive critiques provided on our submission. Here, we provide a consolidated response to address the concerns and suggestions raised.
1. **Reference and Citations:** In response to the recommendation for added references, we've updated the manuscript to... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
What Distributions are Robust to Indiscriminate Poisoning Attacks for Linear Learners? | Accept (poster) | Summary: This paper explores poisoning attacks on linear classifiers. Specifically, the authors explore why poisoning attacks may have high efficacy on one dataset but low efficacy on another. The authors relate this phenomenon to two potential factors: (1) (projected) constraint set size and separability. The author... | Rebuttal 1:
Rebuttal: We first respond to concerns about the experiments and conclusions we draw. To utilize space, we also respond to the last comment from the **Questions**. For other unaddressed comments, we will reflect them in the revised version.
1. **Statement on results in Table 1 is trying to make a causatio... | Summary: This paper delves into the impact and effectiveness of indiscriminate data poisoning attacks on machine learning models, particularly linear ones. It highlights significant variations in attack effectiveness across different datasets and seeks to understand whether this is due to inherent dataset robustness or... | Rebuttal 1:
Rebuttal: 1. **No clear practical implications from this paper**
We appreciate that the reviewer agrees that our work might inspire the future investigation on the inherent robustness of datasets. As for the possible practical future implications, we present some preliminary results in the discussion secti... | Summary: Inspired by the variant performance of indiscriminate data poisoning attacks against different datasets, this paper studies the inherent robustness of datasets against them. Using an optimal attack designed for a binary Gaussian dataset, the authors discover three properties, namely class separability, variati... | Rebuttal 1:
Rebuttal: We appreciate the reviewer for acknowledging our work for providing the initial but valuable step towards understanding the impact of distributional properties on robustness of linear learners against poisoning attacks. Please see our individual responses below.
1. **Set title to be more focused*... | Summary: This paper explores the issue of data poisoning attacks on machine learning models, focusing on linear models. The authors find that the effectiveness of state-of-the-art poisoning strategies varies across different datasets. They introduce definitions of optimal poisoning attacks for finite-sample and distrib... | Rebuttal 1:
Rebuttal: 1. **Is the observation of "linear learners being robust to poisoning on some datasets" a novel observation or previously reported?**
Yes, we are the first to explicitly report the results that linear models can resist the current state-of-the-art poisoning attacks on some datasets by running the... | Rebuttal 1:
Rebuttal: We continue to respond to reviewer oRg1 in the global response due to space constraints and also attach the new result tables as a single PDF. Comments from other reviewers are addressed in the individual rebuttals.
2. **For results on data sanitization defenses, to justify "mostly by ...", a lo... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Revisiting Adversarial Robustness Distillation from the Perspective of Robust Fairness | Accept (poster) | Summary: The paper addresses the issue of amplified class-imbalanced performance during adversarial robust distillation (ARD). The authors propose a fair-oriented ARD method that incorporates a class re-weighting mechanism for adversarial training. The method is applicable to various ARD techniques and consistently ach... | Rebuttal 1:
Rebuttal: Thank you for your constructive comments. We have taken great care to address all your concerns, providing necessary clarifications when needed. We genuinely hope that our responses comprehensively answer your questions.
> **(Comment 1): The reviewer suggests investigating whether existing fair-o... | Summary: This paper focuses on the issue of robust fairness in adversarial robustness distillation. This paper proposes the Fair Adversarial Robustness Distillation (Fair-ARD) framework, which utilizes a more refined difficulty metric and an adaptive class re-weighting approach, enabling the student model to learn the ... | Rebuttal 1:
Rebuttal: Thank you for your constructive comments. We have taken great care to address all your concerns, providing necessary clarifications when needed. We genuinely hope that our responses comprehensively answer your questions.
>**(Comment 1): The novelty is somewhat limited ...**
**Answer 1:**
To beg... | Summary: Adversarial robustness distillation transfers knowledge from a teacher model to a student model and improves the overall robustness of the student in terms of resisting adversarial attacks. This work studies robust fairness in the task of adversarial robustness distillation, which focuses on the adversarial ro... | Rebuttal 1:
Rebuttal: Thank you for your constructive comments. We have taken great care to address all your concerns, providing necessary clarifications when needed. We genuinely hope that our responses comprehensively answer your questions.
> **(Comment 1): The experiments are mostly done on CIFAR-10/100, a relative... | Summary: This paper deals with the issue of fairness in adversarial training in image classification with distillation. The idea is that due to model capacity gap between the teacher and the student model, not all classes are equally capable of retaining the teacher model's robustness and thus the distilled student mod... | Rebuttal 1:
Rebuttal: Thank you for your constructive comments. We have taken great care to address all your concerns, providing necessary clarifications when needed. We genuinely hope that our responses comprehensively answer your questions.
> **(Comment 1): Due to lack of theoretical grounding, how reliable are the ... | Rebuttal 1:
Rebuttal: Figure 10: The weights of each class in Fair-ARD using ResNet18 on CIFAR-10 across epochs.
Table 14: The average (Avg.) and worst-class (Worst) robustness and their standard deviations for
various algorithms using ResNet18 on CIFAR-100.
Table 15: The average (Avg.) and worst-class (Worst) robust... | NeurIPS_2023_submissions_huggingface | 2,023 | Summary: This work studies the problem of fairness in the context of distillation of adversarially robust models. Specifically, it is shown that the accuracy disparity between classes (leading to “unfairness”) is _further exacerbated_ for student models, attributed to the a) capacity gap between the student and teacher... | Rebuttal 1:
Rebuttal: Thank you for your constructive comments. We have taken great care to address all your concerns, providing necessary clarifications when needed. We genuinely hope that our responses comprehensively answer your questions.
> **(Comment 1): An essential baseline assumption is that all classes are ba... | null | null | null | null | null | null |
Computing Optimal Equilibria and Mechanisms via Learning in Zero-Sum Extensive-Form Games | Accept (poster) | Summary: The authors proposed a learning-based framework to solve for optimal equilibria in n-player general-sum extensive-form games. The key idea is to leverage Lagrangian relaxation and convert such games to 2-player zero-sum extensive-form games which can then be solved using known learning methods that solve for m... | Rebuttal 1:
Rebuttal: Thanks for the review!
> Would it be possible to provide additional details on optimality as well especially in general-sum games such as Sheriff where we know the value of max-welfare equilibria (e.g. https://www.cs.cmu.edu/~gfarina/2020/efcce-aaai20/coarse-correlation.aaai20.pdf)?
Sheriff game... | Summary: The paper focuses on addressing the challenge of computing optimal equilibria in extensive-form games. The authors introduce the revelation principle, which transforms the problem into a linear programming (LP) task.
They propose using Lagrange relaxations to solve the LP, treating the resulting saddle-point... | Rebuttal 1:
Rebuttal: Thanks for the review!
> My only concern is about the revelation principle. The reduction in the paper relies on the revelation principle, which is a fundamental concept. However, it is not clear regarding the specific conditions under which a fixed pure strategy d_i exists for different players ... | Summary: This paper studies the computation of optimal equilibria in multi-player extensive-form games via no-regret learning algorithms. The key idea is to take the constrained LP formulation of the optimal equilibrium problem proposed by Zhang and Sandholm and consider the saddle point formulation, which can then be ... | Rebuttal 1:
Rebuttal: Thanks for the review!
> I have some concerns about the LP formulation (G).
There is an important detail that might have been missed here. The cited no-swap-regret dynamics compute one equilibrium, but this paper is concerned with optimizing over the space of equilibria (for example, to compute ... | Summary: This paper presents a novel approach to compute optimal equilibria in multi-player extensive-form games through the use of Lagrangian relaxation as a two-player zero-sum extensive-form game. Building upon the mediator augmentation game framework, the proposed computing approach significantly contributes to the... | Rebuttal 1:
Rebuttal: Thanks for the review!
> The paper lies heavily on notation that lacks intuitive explanations. To enhance reader understanding, a schematic representation of the mediator-augmented game framework should be included in the main body.
Thank you for the suggestion; we will incorporate this and othe... | null | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Would I have gotten that reward? Long-term credit assignment by counterfactual contribution analysis | Accept (spotlight) | Summary: This paper lays out a goal of addressing a weakness in HCA that HCA can confuse the contributions of actions to reaching a state, increasing the variance of gradient estimation. Instead, the COCOA method proposed in this paper generalizes HCA to use any feature that is predictive of rewards, such as the reward... | Rebuttal 1:
Rebuttal: Thank you for your comprehensive remarks and for the useful feedback that we incorporated in a revised version of our paper. We address your main concerns below.
## 1. SOTA Control Variate baseline.
We thank the reviewer for suggesting an extra baseline that leverages hindsight information for a... | Summary: This paper proposes Counterfactual Contribution Analysis (COCOA), an RL credit assignment approach inspired by the Hindsight Credit Assignment (HCA) family of algorithms.
The paper notes that in some instances, the previously proposed State-HCA approach can degrade to as high a variance as REINFORCE due to u... | Rebuttal 1:
Rebuttal: Thank you for your encouraging review and the useful feedback that helped us improve our work.
## 1. Adjusting the claims of COCOA
We acknowledge the foundational work of Harutyunyan et al. 2019 in proposing a general class of credit assignment algorithms using hindsight probabilities conditioned... | Summary: This paper introduces COCOA, a new family of hindsight credit assignment methods that build on HCA, which uses hindsight importance weights for the policy gradient estimator to reduce its variance. HCA uses state-conditioned importance weights, i.e., the ratio of $p(a_t|s_t, s_{t+k})$ to $\pi(a_t|s_t)$. COCOA ... | Rebuttal 1:
Rebuttal: Thank you for your thoughtful comments and the useful feedback that helped us improve our work. We address the points you raised one-by-one below.
## 1. Comparison to HCA-return
We agree that a dedicated comparison to and discussion of the return-conditioned variant of HCA was missing and we add... | Summary: The paper presents a policy gradient estimator using credit assignment measure on influence of specific actions toward future rewards. The paper offers theoretical analysis on resulting algorithm such as (1) policy gradient estimator is unbiased, (2) variance of the estimator in relation to some existing metho... | Rebuttal 1:
Rebuttal: Thank you for the useful feedback that helped us improve our work. We address the points you raised below.
## 1. Contribution compared to HCA and importance of our theoretical results
Although Harutyunyan et al. 2019 proposed HCA as a family of algorithms leveraging hindsight probabilities condit... | Rebuttal 1:
Rebuttal: We thank all reviewers for their constructive comments and useful suggestions that we believe have helped us to significantly improve our paper. Here we summarize the changes we did addressing the main concerns of the reviewers, linking to the corresponding detailed responses.
## Main changes
1... | NeurIPS_2023_submissions_huggingface | 2,023 | Summary: The paper introduces a novel credit assignment algorithm for reinforcement learning known as COCOA (Counterfactual Contribution Analysis). The proposed method extends the concept of Hindsight Credit Assignment (HCA) and aims at making the learning process more sample efficient. By leveraging counterfactual rea... | Rebuttal 1:
Rebuttal: Thank you for the encouraging review and for the useful feedback that we incorporated in a revised version of our paper. We reply to your questions point-by-point below.
## 1. Comparison to SOTA credit assignment methods
We now include HCA-return in our experiments, as well as a new method Counte... | Summary: This paper proposes Counterfactual Contribution Analysis (COCOA) to improve credit assignment in the reinforcement learning problem by building upon Hindsight Credit Assignment.
Strengths: This paper is written in a clear manner and allows non-expert audiences to understand its difference and contribution in ... | Rebuttal 1:
Rebuttal: We thank the reviewer for his review and pointing us to the typo. | Summary: This paper presents family of algorithms namely Counterfactual Contribution Analysis (COCOA), which is based on measuring contribution of an action by asking counterfactual question. The new measure of contribution performs better than existing methods such as HCA in terms of lower variance. They define reward... | Rebuttal 1:
Rebuttal: Thank you for your positive review and instructive suggestions that have helped to improve the paper. Below we provide responses to your raised questions and concerns.
## Computational cost
We provide the wall clock time of all methods after 3000 update steps on the linear key-to-door environme... | null | null |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.