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 |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
SwiftSage: A Generative Agent with Fast and Slow Thinking for Complex Interactive Tasks | Accept (spotlight) | Summary: The paper suggests an approach to solving interactive reasoning and planning tasks inspired by the dual-system theory of human cognition popularized by Kahneman. The approach, SwiftSage, combines a small seq2seq model as System 1 and a LLM as System 2. A heuristic module mitigates between the two. The authors ... | Rebuttal 1:
Rebuttal: Dear Reviewer MDXU,
Thank you very much for your thoughtful feedback and comments! We are pleased to learn that you find our method to be **solid** and that we address an important problem through **extensive experiments**. We also appreciate your valuable suggestions and questions. In the follow... | Summary: The paper proposes SwiftSage that predicts a high-level plan by Swift and prompts LLMs to conduct each high-level subgoal by Sage.
Swift is trained by imitation learning for action planning as fast intuitive thinking and Sage utilizes LLMs for better generalizable planning and exception handling as deliberate ... | Rebuttal 1:
Rebuttal: Dear Reviewer Sgwx,
We greatly appreciate your thoughtful review and are pleased to learn that you find the SwfitSage well-motivated, yielding strong performance over prior work. We value your four insightful questions (Q1-Q4) regarding the weaknesses of our approach and have provided detailed re... | Summary: This paper proposes a dual-module approach for ScienceWorld which combines an imitation learning agent based on t5, with an llm-based module which activates based on a heuristic to re-plan its actions, using a buffer of actions to be executed.
The approach is novel, sensible and well-motivated, drawing from th... | Rebuttal 1:
Rebuttal: Dear Reviewer CSRK,
Thank you very much for your thoughtful feedback and comments! We are glad to hear that you believe our method is well designed and significantly advances the state of the art in complex interactive reasoning, as well as finding our evaluation informative. We also appreciate y... | Summary: This paper proposes a new agent framework for action planning in complex interactive reasoning tasks. The authors leveraged an imitation-learning trained "light-weight" LLM for fast inference and an LLM (e.g. GPT4) for slow reasoning with proper designs of prompts and interaction between fast/slow modules. The... | Rebuttal 1:
Rebuttal: Dear Reviewer AXAQ,
Thank you very much for your thoughtful feedback and comments! We are glad to hear that you believe our paper is well motivated and addresses the important problem of interactive reasoning and planning. We appreciate your concerns and have addressed your two questions below.
... | Rebuttal 1:
Rebuttal: ### General response
Dear Reviewers and AC,
We greatly appreciate your insightful reviews and are delighted that all of you have acknowledged our paper's positive aspects. We briefly summarize the favorable feedback as follows:
- **Well-motivated, innovative, generalizable method**: Sgwx, MDXU... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Similarity, Compression and Local Steps: Three Pillars of Efficient Communications for Distributed Variational Inequalities | Accept (poster) | Summary: This paper investigates distributed variational inequalities (VI) problems in the context of a server-workers architecture. It proposes a method that combines three major communication-saving techniques: i) compression, ii) local steps, and iii) data similarity across workers. For strongly monotone VI and stro... | Rebuttal 1:
Rebuttal: We thank Reviewer __JXTW__ for the time, work and detailed review of our paper! Next, we will answer the issues raised.
But before the response, we note that __we added new results to the revised paper__ - see our general response! Reviewer __JXTW__ requested __lower bounds__. We made them and sh... | Summary: The paper presents a study of distributed algorithms for solving variational inequalities in a centralized distributed optimization setting. The authors propose three improvements: compression, data similarity, and local steps. Compression reduces the communication complexity by transmitting compressed version... | Rebuttal 1:
Rebuttal: We thank Reviewer __4qia__ for the work, time, and positive feedback! Below we will respond to the issues raised.
But before the response, we note that __we added new results to the new version of the paper - see our general response!__ We added __the extension of our algorithms to the stochasti... | Summary: This paper studies distributed variational inequalities. It combines three techniques – similarity, compression, and local updates – to reduce the total number of communication rounds. It establishes theoretical guarantees of communication complexity.
Strengths: 1 It combines three techniques – similarity, c... | Rebuttal 1:
Rebuttal: We thank Reviewer __kubg__ for the work, time, and positive feedback! Below we will respond to the issues raised.
But before the response, we note that __we added new results to the new version of the paper - see our general response!__ Reviewer __JXTW__ requested __lower bounds. We made them and... | Summary: The paper proposes two novel distributed algorithms, for solving variational inequalities and saddle point problems. These algorithms combine three techniques - local updates, compression, and similarity - to address the communication bottleneck in distributed optimization for these problems. Algorithm 1 uses ... | Rebuttal 1:
Rebuttal: We thank Reviewer __D8HM__ for the time and the work! Next, we will answer the issues raised.
But before the response, we note that __we added new results to the new version of the paper__ - see our general response! We added __the extension of our algorithms to the stochastic case__ (as requeste... | Rebuttal 1:
Rebuttal: Reviewer __JXTW__ requested lower bounds. We made them and showed that our result is optimal and unimprovable. Reviewers __D8HM__ and __4qia__ asked to extend the results of the paper to the stochastic case. This we also added. Because of the character limit, we give sketches of results and proofs... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Diversified Outlier Exposure for Out-of-Distribution Detection via Informative Extrapolation | Accept (poster) | Summary: The authors of this paper present an innovative methodology for out-of-distribution (OOD) detection. While existing methodologies typically involve the direct use of given OOD samples, this paper introduces a new approach that applies perturbations to OOD samples, allowing the model to experience a more divers... | Rebuttal 1:
Rebuttal: Thank you for your time devoted to reviewing this paper and your constructive suggestions. Here are our detailed replies to your questions.
> **Q1:** However, it would be important for the authors to provide empirical evidence that demonstrates the superiority of this adversarial perturbation ove... | Summary: The manuscript studies image-wide OOD detection in presence of auxiliary negative data. The negative data is often limited and therefore cannot fully encompass the distribution of inliers. Consequently, contemporary learning procedures fail to deliver classifiers resilient to outliers. To overcome this issue, ... | Rebuttal 1:
Rebuttal: Thank you for your time devoted to reviewing this paper and your constructive suggestions. Here are our detailed replies to your questions.
> **W1:** The manuscript does not discuss the effectiveness of the method when there is only a small auxiliary dataset available. It seems that the developed... | Summary: The paper tries to solve the problem that sampled auxiliary informative outliers may not be sufficient and diverse enough to recover the data boundary in the OOD detection setting. To achieve this, the authors propose a new learning objective with information extrapolation, where second term expands the surrog... | Rebuttal 1:
Rebuttal: Thank you for your time devoted to reviewing this paper and your constructive suggestions. Here are our detailed replies to your questions.
> **W1:** About Theorem 3.1
Thanks for the constructive comments. In this work, we follow POEM [Ming et al., 2022] to provide an explanation about DivOE und... | Summary: This paper improves the methods of utilizing auxiliary outlier data for fine-tuning. Specifically, it synthesizes informative outlier samples close to in-distribution at the decision boundary by adding noise to existing auxiliary data and utilizes them in learning. As a result, it shows improved results by app... | Rebuttal 1:
Rebuttal: Thank you for your time devoted to reviewing this paper and your constructive suggestions. Here are our detailed replies to your questions.
> **W1, Q1, W5, W3, Q5:** about DivOE, MixOE, DOE
Thank you for the special recommendation of MixOE and DOE for clarity . We would like to explain their dif... | Rebuttal 1:
Rebuttal: ## General Response
We appreciate all the reviewers for their thoughtful comments and suggestions on our paper.
We are very glad to see that the reviewers find our focused problem is **important** (R2,R3) within the OOD detection research, and the method is **novel** (R1,R2,R3,R4), **theoretical... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Dynamic Personalized Federated Learning with Adaptive Differential Privacy | Accept (poster) | Summary: This paper addresses the pressing issues of inflexibility and convergence difficulties in personalized federated learning under the influence of differential privacy constraints. The authors propose an innovative adaptive approach that prioritizes high fisher value local parameters during the personalization p... | Rebuttal 1:
Rebuttal: Dear Reviewer UEFt:
Thank you very much for your affirmation of our work, as well as the insightful concerns and questions you have raised. We have carefully considered each comment and provided responses.
**Weaknesses:**
**W1: More discussion on potential adverse social impact**
**A1:** Than... | Summary: This study generally explores the problem of personalized federated learning with differential privacy. Specifically, the paper proposes two novel methods known as DFP and AC for enhancing the learning capability of a model while reducing the impact of noise, and improving robustness against clipping. At its c... | Rebuttal 1:
Rebuttal: Dear Reviewer aCpX:
Thank you very much for your recognition of our work and for raising thoughtful questions and concerns about our work. We have carefully considered each comment and provided responses.
**Weaknesses:**
**W1: Complete complexity analysis**
**A1:** Our analysis of extra comput... | Summary: This paper proposes a dynamic personalization strategy using layer-wise Fisher information values, which enables flexible personalization and reduced noise impact. This paper further presents an adaptive method for enhancing clipping robustness by adjusting constraint strengths for different parameters, balanc... | Rebuttal 1:
Rebuttal: Dear Reviewer EfaD:
Thank you for your thoughtful review and for raising key concerns regarding our work. We aim to address your concerns in our detailed responses below, hoping to provide clarity and demonstrate the effectiveness of our proposed approach.
**Weaknesses:**
**W1: Experiments are... | Summary: This paper proposes an adaptive method for Personalized Federated Learning with Differential Privacy (DP-PFL), which aims to address the inflexible personalization and convergence difficulties that DP-PFL suffers from. Firstly, layer-wise Fisher information is used to measure the information content of local p... | Rebuttal 1:
Rebuttal: Dear Reviewer imA4:
Thanks for raising thoughtful questions and concerns. We sincerely hope this point-to-point response allows the reviewer to update the score.
**Weaknesses:**
**W1: Lacks technical novelty**
**A1:** We did not merely adopt statistical estimation theory into our work. Rather,... | null | NeurIPS_2023_submissions_huggingface | 2,023 | Summary: This paper proposes a novel method for Dynamic Personalized Federated Learning with Adaptive Differential Privacy. The method addresses the challenges of inflexible personalization and convergence difficulties in current DP-PFL methods by leveraging layer-wise Fisher information and introducing an adaptive app... | Rebuttal 1:
Rebuttal: Dear Reviewer dw3B:
Thanks for your insightful comments and questions. We sincerely hope our detailed response allows the reviewer to update the score.
**Weaknesses:**
**W1: Motivation of adaptive differential privacy and equations 8 and 9, Details about Section 3.2.2**
**A1:** Current Differ... | null | null | null | null | null | null |
A Partially-Supervised Reinforcement Learning Framework for Visual Active Search | Accept (poster) | Summary: This paper introduces a novel Reinforcement Learning (RL) framework for visual active search. In contrast to previous methods, the authors suggest integrating visual supervised signals to enhance performance. Additionally, a meta-learning based approach is proposed to allow the model to adapt to varying test c... | Rebuttal 1:
Rebuttal: Thank you for the comments and suggestions. We are encouraged to see your appreciation of the writing and presentation style, proposed methodology being intuitive and technically sound, and motivation of TTA and MQA being logically sound and well justified. We respond to all the comments below.
... | Summary: This paper proposes a Partially Supervised Reinforcement Learning to search for the objects of interest in geospatial images. The paper uses a meta-learning approach to learn initialization parameters of the prediction module jointly with a search policy. A variant of this approach, MPS-VAS, learns to choose... | Rebuttal 1:
Rebuttal: We thank the reviewer for the insightful comments and suggestions. We are encouraged to see that you find our proposed method is promising and novel. We respond to all the comments below.
> **Q1**: Can we train on a dataset (xView) and generalize to the classes on another dataset (DOTA)?
**A1**:... | Summary: The paper proposes a partially supervised reinforcement learning framework (PSVAS) for visual active search tasks (VAS). Specifically, it decompose previous end-to-end deep RL method for VAS into a task-specific prediction module and a task-agnostic search module. In such way, the prediction module can be supe... | Rebuttal 1:
Rebuttal: Thank you for the thoughtful comments and suggestions. We are inspired to see that you found our overall method is techniqually sound and empirically proved. We respond to all the comments below.
> **Q1**: How different balancing factor lambda between RL and BCE loss would affect the performance ... | Summary: The paper addresses visual active search problem for geospatial exploration. For an effective adaptation of search policies to out-of-distribution tasks, the paper proposes partially supervised reinforcement learning framework. A meta-learning approach is also proposed that focuses on jointly learning a policy... | Rebuttal 1:
Rebuttal: Thank you for your comments. We are encouraged to see that you appreciated our contribution. We respond to all the comments below.
> **Q1**: Not sure of the reproducibility of this work. Also, no code is provided in the supplementary.
**A1**: With anonymity being a primary concern, we have cho... | Rebuttal 1:
Rebuttal: We thank all the reviewers for their constructive and insightful feedback. We are delighted to see that our work is found novel and technically solid by all the reviewer. We’ve addressed all reviewer concerns to the best of our effort by adding detailed clarifications with supporting experimental ... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Cluster-aware Semi-supervised Learning: Relational Knowledge Distillation Provably Learns Clustering | Accept (poster) | Summary: This paper studies relational knowledge distillation (RKD) on semi-supervised learning setting. This is different from the previous theoretical results on knowledge distillation, which mostly focus on feature matching (not relational). They introduce a new notion called low clustering error, which quantifies t... | Rebuttal 1:
Rebuttal: We would like to thank the reviewer for appreciating the novelty and presentation of this work, as well as for the insightful comments and questions.
__[Intuition of Ass 4.2]__ Ass 4.2 can be intuitively summarized as assumptions on the __boundedness $\beta$__ and __reasonably large margins $\ga... | Summary: This paper points out that RKD over a population-induced graph given by a teacher model leads to low clustering error (Spectral Clustering), which provably benefits to semi-supervised learning and other downstream tasks. Consequently, it provides a theoretical analysis of relational knowledge distillation (RKD... | Rebuttal 1:
Rebuttal: Many thanks for appreciating our novelty, as well as for raising constructive suggestions and questions concerning our experiments.
Nevertheless, from a high-level perspective, we would like to kindly point out that this work is __a novel theoretical study on a family of existing methods__ that ... | Summary: The paper provides a comprehensive theoretical proof of the relationship between RKD and clustering from the perspective of spectral clustering. The theoretical proof presented in the paper is detailed and reasonable. Firstly, RKD is regarded as a spectral clustering problem of the teacher model, and then the ... | Rebuttal 1:
Rebuttal: We would like to first thank the reviewer for appreciating the theoretical contribution of this work, as well as for bringing up constructive questions and suggestions in the review. On the detailed concerns:
__[Motivations and improvements on presentation]__
We understand and value the reviewer... | null | null | Rebuttal 1:
Rebuttal: First, we would like to thank all the reviewers for their efforts and valuable suggestions. We will share an update on our experiments (in Appendix A) here globally while addressing individual questions and concerns posed in the reviews separately.
__[Additional experiments with multiple random s... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Spectral Co-Distillation for Personalized Federated Learning | Accept (poster) | Summary: The authors proposed a personalized federated learning (PFL) method, where the (dis-)similarity of models are represented through their Fourier spectra. The authors also proposed a bi-directional knowledge distillation between the generic and personalized models and a wait-free local training protocol that can... | Rebuttal 1:
Rebuttal: We are very encouraged by the reviewer's positive feedback. Below, we address the comments point-by-point.
> The use of Fourier spectra is motivated by previous studies on deep learning theories. As it stands, the method seems like a mathematically motivated heuristic. It would be excellent if ce... | Summary: This paper proposes a novel approach to address the challenge of data heterogeneity in personalized federated learning (PFL) by introducing a spectral co-distillation framework based on model spectrum information. The proposed method combines a wait-free local training protocol with bi-directional spectrum kno... | Rebuttal 1:
Rebuttal: We thank the reviewer for the overall positive feedback. Below we address the highlighted weakness point-by-point.
> The proposed approach may require additional computational resources to compute the spectral information of the models, which could increase the overall training time and cost. The... | Summary: This paper introduces a novel spectral co-distillation framework for PFL, leveraging a Fourier spectrum perspective to measure model similarities. With the assumption that local models share low-frequency features and the global model's Fourier spectrum includes averaged high-level features from all clients, t... | Rebuttal 1:
Rebuttal: We thank the reviewer for the valuable feedback. Below we address the reviewer's concerns point-by-point.
> [Lack of Theoretical Analysis] The authors highlight in lines 39-41 that existing works lack theoretical support and heavily rely on empirical observations. However, the core motivation beh... | Summary: This paper proposes a method for PFL, which aims to address the challenge of data heterogeneity. The main idea is to use spectral distillation to capture the relations between the generic global and personalized local models based on the similarity of spectral information. The paper introduces a co-distillatio... | Rebuttal 1:
Rebuttal: We thank the reviewer for the overall positive feedback. Below we address the highlighted weakness point-by-point.
> The clarity of this paper should be improved. Please be careful with using uncommon abbreviations, or provide explanations where appropriate. GM and PM should be clarified in the F... | null | NeurIPS_2023_submissions_huggingface | 2,023 | Summary: This paper introduces a spectrum based co-distillation for personalized federated learning; which involves the local model training with regularization/distillation with respect to global parameters's spectral representation and the global model training with regularization/distillation with respect to local p... | Rebuttal 1:
Rebuttal: We thank the reviewer for the comments.
> The technical contribution is relatively low
We have the following novel technical contributions, which have also been pointed by all the other reviewers.
1. We are the first to characterize model (dis-)similarity of personalized models (PMs) versus the... | Summary: 1. This paper proposes a new personal FL methodology: FedSCD. The authors facilitates the generic and personal models to capture global/local distributions through spectral distillation. A co-distillation framework is introduced to help generic and personal models to learn from each other. From the Fourier ana... | Rebuttal 1:
Rebuttal: We thank the reviewer for the detailed comments and suggestions.
> Why spectral co-distillation is not counter-intuitive
In existing PFL works, a common starting point is that the generic model (GM) and personalized models (PMs) share the same low-level features, while the high-level features are... | null | null | null | null |
Unsupervised Video Domain Adaptation for Action Recognition: A Disentanglement Perspective | Accept (poster) | Summary: The proposed method in this paper handles the spatial and temporal domain divergence separately via disentanglement.
Two sets of latent factors are considered (i) static information encoding and (ii) dynamic information encoding.
The transfer sequential VAE framework is proposed to model the generation of cr... | Rebuttal 1:
Rebuttal: We thank Reviewer qCdJ for devoting time to this review and drawing valuable comments.
---
> ***Q1:** "(i) Complexity analysis w.r.t. existing UDA methods in action recognition; (ii) the major difference/contribution compared to existing works; (iii) why bother to design such a complicated model?... | Summary: This paper proposes an unsupervised domain adaptive approach for video de-entangling based on unraveling spatial and temporal domain discrepancies. The proposed TranSVAE framework models the problem and optimizes the latent factors through multiple objective constraints. Experimental results demonstrate the ef... | Rebuttal 1:
Rebuttal: We thank Reviewer AobY for devoting time to this review and drawing valuable comments.
---
> ***Q1:** "The approach of disentanglement for video tasks is relatively mature."*
**A:** We would like to thank the reviewer’s comment.
- Firstly, we highlight that the primary idea of this paper is to p... | Summary: The paper proposes a Transfer Sequential VAE (TranSVAE) to model temporal and spatial domain divergence through feature disentanglement for unsupervised domain adaptation for action recognition from videos. The architecture consists of 5 components - encoder, LSTM, latent spaces, sampling and decoder. Video is... | Rebuttal 1:
Rebuttal: We thank Reviewer cyvj for devoting time to this review and drawing valuable comments.
---
> ***Q1:** "Summarizing the main contributions point by point would make the paper a better read."*
**A:** We have included the following summarization in the revision. The main contribution can be summari... | Summary: This paper focuses on video domain adaptation. The authors' primary focus is to eliminate the temporal invariant component of the features initially, followed by the implementation of an adversarial-based domain adaptation technique. To achieve disentanglement, they employ a method of temporally random shuffli... | Rebuttal 1:
Rebuttal: We thank Reviewer 3nXx for devoting time to this review and drawing valuable comments.
---
> ***Q1:** "Add an ablation study to specifically evaluate $L_{mi}$ and/or $L_{ctc}$, without the influence of other components, especially $L_{svae}$."*
**A:** Thanks for your suggestion. We highlight tha... | Rebuttal 1:
Rebuttal: We sincerely thank the PCs, ACs, and all the reviewers for devoting time and effort to this review.
---
We are glad to see that the reviewers are acknowledging that:
- *"The motivation of the method is clear and easy to understand and follow"* (Reviewer JyS2)
- *"The idea of handling spatial and ... | NeurIPS_2023_submissions_huggingface | 2,023 | Summary: The paper tackles the unsupervised domain adaptation (UDA) problem by viewing it from a disentanglement perspective. It suggests separating two sets of latent factors, one of which is responsible for domain-related spatial information and another encodes dynamic temporal information. As a result, a new model f... | Rebuttal 1:
Rebuttal: We thank Reviewer JyS2 for devoting time to this review and drawing valuable comments.
---
> ***Q:** "More discussions about related works in the disentanglement area and more outlined explanations of differences of the proposed TranSVAE from them, especially with the papers for video generation ... | null | null | null | null | null | null |
Agnostic Multi-Group Active Learning | Accept (poster) | Summary: This paper studies active learning in a multi-group setting, in which the learner performs adaptive sampling over groups of distributions for solving a binary classification task. The performance of the learner is measured by its maximal loss over the groups. The main goal is to extend the disagreement-based a... | Rebuttal 1:
Rebuttal: We thank the reviewer for the comments, and the close reading even of parts of our Appendix.
>Can the authors comment on the computational complexity issue above regarding the computation of $R_i$ and $\mathcal{H}_{i+1}$?
With regard to computational considerations, one can always construct an ... | Summary: This paper looks at disagreement-based active learning (DBAL) where
the goal is to minimize the maximum error over a (known) set of G
distributions ("groups"). One could learn each group individually,
but the optimum for each group may be very bad on other groups. By
combining the search, a decent sample com... | Rebuttal 1:
Rebuttal: We thank the reviewer for reading our work with an eye for algorithmic detail.
> In the easier regime of getting error $O(\epsilon)$ [Theorem 3], the dependence is theta as one would expect. But the bound there is $d G \theta \log^2(1/\epsilon)$, and it seems like the $\log^2(1/\epsilon)$ shoul... | Summary: The paper considers an active binary classification setting where the feature space is partitioned into several groups. The performance of the learner is given as their worst case performance across all groups. The effect of considering such a multi-group setting is that, it is no longer efficient to sample ex... | Rebuttal 1:
Rebuttal: We thank the reviewer for a thorough reading of our submission. We try to address some questions/concerns below.
>It would be of more interest to consider lower bounds in the active setting, even in the form of conjectures
We agree it would be nice to have active lower bounds. In general, tight ... | Summary: This paper studies multi-group learning from the active learning perspective. The authors first discuss why classical disagreement-based active learning methods cannot work and provide an example. Then, they propose a modified algorithm and provide theoretical guarantees. The label complexity bound is smaller ... | Rebuttal 1:
Rebuttal: We appreciate the review and the comments.
> How about uncertainty-based methods? What if the learner just labels the sample from the most uncertain group?
In general, one must be careful with uncertainty based approaches in the agnostic AL setting as they can easily lead to inconsistent strate... | Rebuttal 1:
Rebuttal: We appreciate all the reviews.
There were a few comments on the fact that the general agnostic algorithm (Algorithm 1), as stated, cannot be explicitly executed for infinite hypothesis classes.
In the case that one wishes to run the algorithm, one can always construct an epsilon-cover of the ... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
H2O: Heavy-Hitter Oracle for Efficient Generative Inference of Large Language Models | Accept (poster) | Summary: The author proposes to reduce the size of KV cache to accelerate the inference of the Transformer models. The size reduction is conducted through a mechanism called heavy hitter oracle, which treats attention score as the importance score. The authors conduct experiments with various tasks and observe consiste... | Rebuttal 1:
Rebuttal: We thank the reviewer for really insightful and detailed questions and suggestions. These help us strengthen experimental results and make more precise claims. We present details of our implementation as well as additional experiment results that we have added to the updated draft below.
**[Q1: T... | Summary: The paper presents H2O, a technique to identify key elements in the sequence whose KV values play the most important role in attention scores for generation of future tokens and proposes to only keep these in cache to improve upon memory requirement and latency during inference.
The proposed algorithm is then... | Rebuttal 1:
Rebuttal: Many thanks to the reviewer for acknowledging that the problem we explored is “important”, “has huge impact”, and also our algorithm is “impactful”. We provide detailed responses below to address the constructive feedback.
**[Q1: Inconsistent Notations]** Thanks for pointing it out. To fix the in... | Summary: This paper first identified an interesting observation that only a small portion of tokens contributes most of
the value when computing attention scores. And based on this observation, this paper proposed Heavy Hitter Oracle (H2O), a KV cache eviction policy that dynamically retains a balance of recent and H2 ... | Rebuttal 1:
Rebuttal: We thank the reviewer for strongly supporting our work and providing valuable feedback, which helps improve our work. We present pointwise responses for your constructive suggestions as the following:
[**Q1: Results on Encoder-Decoder Framework**] Thanks for the suggestion. We further evaluate o... | Summary: Paper addresses inefficiency in key-value matrix caching during attention computations of generation with transformer models by introducing a new pruning strategy to discard less useful information without significant degradation of performance. This strategy is based on the reliance of transformers' on observ... | Rebuttal 1:
Rebuttal: We are glad that the reviewer appreciates our work and efforts, and acknowledges we **tackle an important issue** and our approach **has many potentially critical practical implications**. We provide detailed responses below to address the questions.
**[Q1: More Evaluation of Diverse Generation]... | Rebuttal 1:
Rebuttal: We thank reviewers [AZfb,CXTK,N7jZ,nR7P] for their thoughtful and highly supportive feedback! We were glad that the reviewers found the problem **significant and impactful** [AZfb,CXTK,N7jZ, nR7P], the observations and theory **interesting and well-analyzed** [CXTK, nR7P], the methods **simple and... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Posterior Sampling with Delayed Feedback for Reinforcement Learning with Linear Function Approximation | Accept (poster) | Summary: This paper introduces and analyzes two "posterior sampling" algorithms for learning optimal policies in linear MDPs in a setting where feedback from the environment is delayed (by stochastic amounts of time). The algorithms, called Delayed-PSVI and Delayed-LPSVI (namely Delayed (Langevin) Posterior Sampling Va... | Rebuttal 1:
Rebuttal: We sincerely appreciate Reviewer TvPe's thoughtful review of our work. We are grateful for the detailed feedback, which provides valuable insights.
1. Common ground of delayed feedback definition with existing literature
By Definition 2, we call each trajectory (i.e. rollouts) $[s^t_h,a^t_h,r... | Summary: The paper discusses the challenges of delayed feedback in real-world systems for reinforcement learning and proposes an optimistic algorithm that utilizes noise perturbation via posterior sampling to effectively explore. The algorithm is shown to achieve a worse-case regret that depends on the delay and time h... | Rebuttal 1:
Rebuttal: We appreciate Reviewer 7RMs' thorough evaluation of our work. We thank the reviewer for providing valuable feedback that will contribute to the refinement of our work.
1. Clarification of technical novelty (theory)
Theory-wise, our main technical contributions and novelties are as follows:
(... | Summary: The authors study the frequentist regret of TS under stochastic delayed feedback with linear function approximation. They further proposed using LMC as a posterior sampling strategy to improve computational efficiency.
Strengths: 1. As far as I know, this is little work that develops frequentist regret bound ... | Rebuttal 1:
Rebuttal: We appreciate Reviewer U3pE's thorough evaluation of our work. We thank the reviewer for providing valuable feedback that will contribute to the refinement of our work.
1. Clarification on mathematical definitions and notation issues
We thank the reviewer for pointing out the unclear definiti... | Summary: This paper studies the problem of online reinforcement learning with delayed feedback. To address this problem, the authors proposed two Thompson sampling-based algorithms, Delayed Posterior Sampling Value Iteration (Delayed-PSVI) and Delayed-Langevin Posterior Sampling Value Iteration (Delayed-LPSVI). Delayed... | Rebuttal 1:
Rebuttal: We appreciate Reviewer p2ng's detailed evaluation of our work. We thank the reviewer for recognizing the significance of the delay feedback problem that we address and for providing valuable feedback that will contribute to the refinement of our work.
Below we provide our response to the question... | Rebuttal 1:
Rebuttal: We sincerely thank all reviewers for insightful comments on our work. We are grateful that reviewers recognize the significance of addressing delayed feedback in RL; find our proposed algorithms to be a valuable contribution to this domain; and acknowledge that the theoretical analysis is thorough... | NeurIPS_2023_submissions_huggingface | 2,023 | Summary: This paper presents a theoretical study of a posterior sampling algorithm for linear MDPs with _delayed rewards_. The authors propose two algorithms, Delayed-PSVI and Delayed-LPSVI, which are based on sampling from a Gaussian posterior distribution over weights induced by a Bellman style likelihood function. T... | Rebuttal 1:
Rebuttal: We appreciate Reviewer RbB5's thorough evaluation and positive assessment of our work. We thank the reviewer for providing valuable feedback that will contribute to the refinement of our work.
1. Clarification of relationship with double Q-learning
Thank you for the good question! Overall spe... | Summary: This paper studies the Thompson sampling methods for reinforcement learning problems with delayed feedbacks. The authors tackle the challenge of delayed feedback under the linear function approximation setting. They proposed two new Thompson sampling algorithms, using the exact sampling mechanism and also the ... | Rebuttal 1:
Rebuttal: We appreciate Reviewer UyNX's detailed evaluation of our work. We thank the reviewer for the thoughtful feedback and for recognizing the soundness of our work.
1. Clarification of technical contribution
Regarding the concern about the technical contribution, we would like to point out that:
(... | Summary: This paper considers the scenario where the feedback is delayed. The motivation for such a scenario is recommender/advertisement systems and clinical trials. They present a novel algorithm based on Thompson sampling idea: Delayed-PSVI. The authors present theoretical results for worst-case regret. They also sh... | Rebuttal 1:
Rebuttal: We appreciate Reviewer 8fzn's thoughtful review and positive assessment of our work. We are pleased to see that the reviewer acknowledges the significance of addressing delayed feedback scenarios in RL, and finds our proposed algorithm, Delayed-PSVI, to be a valuable contribution to this domain.
... | null | null |
Active representation learning for general task space with applications in robotics | Accept (poster) | Summary: The paper introduces an algorithm for active representation learning in multi-task settings. The algorithm generalizes to settings in which the “task space” is either discrete or continuous.
The overall setting involves an (optionally) nonlinear map between an input space X and an intermediate feature space v... | Rebuttal 1:
Rebuttal: ### Q1: The setting itself seems interesting, but not enough emphasis is put on why it is significant and worthwhile exploring. It seems that the approach effectively requires us to be able to at-will sample tasks within their “task parameter space”, which seems to be extremely hard to do for anyt... | Summary: This work considers the multi-task learning problem and proposes an algorithm to select related source tasks to tackle a known target task in an active learning like setup. The analysis is primarily for the linear function classes and handles discrete and continuous task spaces. The proposed algorithm operates... | Rebuttal 1:
Rebuttal: ### Q1.1: Experimental setup: To the best of my understanding, all the experiments in the manuscript (including appendix) compare passive to target-aware active learning strategies.
We want to correct the in our experiment 1: synthetic data, we provide comparison between passive and target-agnos... | Summary: This paper proposes a framework for active representation learning that allows learners to optimally choose which source tasks to sample from. The framework covers both task-aware and task-agnostic settings and is compatible with deep representation learning practices. The paper provides several instantiations... | Rebuttal 1:
Rebuttal: ### Q1: The manuscript would be considerably enhanced by the inclusion of more comprehensive empirical assessments, specifically pertaining to real-world robotic assignments, as opposed to merely simulated tasks. Such addition would serve to confirm the applicability and efficacy of the suggested ... | Summary: The paper studies active representation learning under the assumption that the source and target tasks are parameterized by a vector space (instead of a discrete set of tasks). The proposed algorithm works by first finding a subspace of the task space that spans the representation space and then iterates betwe... | Rebuttal 1:
Rebuttal: ### Q1: The problem statement and motivation were not described precisely or concisely. E.g., it's not clear to me why robotics is the only target for this work. The methods section was robotics-agnostic, but the introduction uses robotics as the motivating example and the paper is titled as a rob... | Rebuttal 1:
Rebuttal: # Scope of this paper :
As pointed out by some reviewers, active learning, multi-task learning and representation learning is a broad area. However, our work is particularly focus on the following settings:
(1) We focus on assumptions that all the tasks share the same representation, which means... | NeurIPS_2023_submissions_huggingface | 2,023 | Summary: The paper introduces and studies a theoretic and algorithmic framework for actively and optimally selecting the source tasks in multi-task learning, accommodating task-aware and task-agnostic settings.
The authors present cohesive theoretical guarantees, particularly demonstrated through the pure bilinear set... | Rebuttal 1:
Rebuttal: ### Q1: What is the trade-off and limitations of the proposed algorithm compared to previous methods?
Please refer to our response to all reviewers. In particular, in the “limitation” section we discuss its limitations in both assumptions and the solutions.
### Q2: When it comes to a single targ... | Summary: Which source task to sample from is an important issue in multi-task learning. This paper proposes a general and versatile algorithm and framework for active representation learning. The proposed algorithm can be adapted to arbitrary target and source task space, and can cover task-aware and task-agnostic sett... | null | null | null | null | null |
SimMMDG: A Simple and Effective Framework for Multi-modal Domain Generalization | Accept (poster) | Summary: This work focuses on addressing a special case of multimodal learning, i.e., multi-modal domain generalization, where one or more modalities will be absent during test phase.
This work propose a approach to solve multimodal DG from a disentanglement perspective. Modality-specific features contain modality-s... | Rebuttal 1:
Rebuttal: Thanks for your insightful reviews, and we appreciate your valuable suggestions! We address your concerns and questions as follows:
>**Q1**: However, the setting in this work is that the testing domain/modality exists in training data but some training modalities might be missing.
**A1**: In our ... | Summary: This paper addresses the challenge of domain generalization in a multi-modal setting. The core idea is that aligning features from multiple modalities in a shared embedding space can hinder generalization. To overcome this limitation, the paper proposes a more effective approach: it separates the features with... | Rebuttal 1:
Rebuttal: Thanks for your insightful reviews, and we appreciate your valuable suggestions! We address your concerns and questions as follows:
>**Q1**: The paper highlights an apparent contradiction in its argument regarding modality-specific information. On one hand, it acknowledges that certain modalities... | Summary: This paper presents a framework for domain generalization for models trained with multi-modal/ multi-source data: SimMMDG. The approach is motivated by the observation that each modality contains features shared with other modalities as well as features specific to that modality. The paper proposes to split pe... | Rebuttal 1:
Rebuttal: Thanks for your insightful reviews, and we appreciate your valuable suggestions! We address your concerns and questions as follows:
>**Q1**: The learning rate of CL is set to 3.0; this essentially increased the effective learning rate of the model significantly. I wonder if authors have tuned the... | Summary: This paper proposes a method for multi-model domain generalization. Rather than mapping features of different modality into the same embedding space, it splits features within each modality into modality-specific feature and modality-shared feature. Then it applies different learning strategies to different sp... | Rebuttal 1:
Rebuttal: Thanks for your insightful reviews, and we appreciate your valuable suggestions! We address your concerns and questions as follows:
> **Q1:** I am confused to the section 3.2.2 cross-modal Translation. The design is translating the i-th modality feature to j-th modality feature. However, like the... | Rebuttal 1:
Rebuttal: >**Q1**: More details on HAC dataset.
**A1**: The detailed dataset statistics such as the name of each action, the number of each action segment, and training-testing splits are in the supplementary material. Here, we provide more details on how we collect the dataset.
Our dataset was collected ... | NeurIPS_2023_submissions_huggingface | 2,023 | Summary: The paper tackles the problem of multimodal domain generalization in which the goal is to learn a model from multi-modal multi/single-source domain data that can generalize to unseen multimodal distributions. The main motivation is that mapping data from multiple modalities to a shared embedding space only har... | Rebuttal 1:
Rebuttal: Thanks for your insightful reviews, and we appreciate your valuable suggestions! We address your concerns and questions as follows:
>**Q1**: The paper seem to be missing the domain generalization aspect, like which pertinent DG problem in the context of multimodal data the paper is trying to solve... | null | null | null | null | null | null |
Feature Selection in the Contrastive Analysis Setting | Accept (poster) | Summary: The paper studied feature selection in the contrastive analysis setting. There are two types of datasets, the background dataset and the target dataset. The background dataset contains variations due to the uninteresting sources. The target dataset contains observations with variability from both ``interesting... | Rebuttal 1:
Rebuttal: We thank the reviewer for examining our work and providing their feedback. We’re thrilled to hear that the reviewer found our work “well-written” with “strong” motivation. In the space below we respond to the reviewer’s specific concerns:
**Evaluations beyond MNIST:** The reviewer noted that, des... | Summary: The paper aims to solve the problem of “salient” feature selection in a Contrastive Analysis setting, termed “Contrastive Feature Selection.” The authors' primary focus lies in uncovering distinctive features that highlight salient variations present in a specific target dataset that are absent in a standard ... | Rebuttal 1:
Rebuttal: We thank the reviewer for examining our work and providing their feedback. We’re thrilled to hear that the reviewer found that CFS represents a “novel attempt at feature selection” and that our results show “the superior performance of the proposed feature selection method”. We have responded to ... | Summary: The authors tackle the important problem of feature selection in a contrastive setting. Supervised and unsupervised feature selection problems have been extensively studied and have many high-impact applications. In this work, the authors design an algorithm for finding features that are more enriched in a tar... | Rebuttal 1:
Rebuttal: We thank the reviewer for examining our work and providing their feedback. In the space below we respond to the reviewer’s specific concerns.
**Train/test split:** The training points from the 80/20 test split were used to select features using CFS and train the subsequent classifiers, while the ... | Summary: They employed a contrastive analysis method for feature selection on the target dataset, which provides variation not found in the background dataset. Contrary to previous contrastive analysis that typically involves joint modelling of background and target datasets to identify shared variables, they drew from... | Rebuttal 1:
Rebuttal: We thank the reviewer for carefully examining our work and providing their feedback. We’re thrilled to hear that the reviewer found that our manuscript “clearly explained” why CFS achieved superior performance compared to previously proposed feature selection methods. We have responded to general ... | Rebuttal 1:
Rebuttal: We thank the reviewers for their thoughtful comments. We’re thrilled to hear that reviewers found CFS a “novel attempt at feature selection in a weak supervision setting”, “new and highly appropriate” (xS67), and “with practical interest” (nNUw). We’re also thrilled to hear that reviewers found ou... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Reusing Pretrained Models by Multi-linear Operators for Efficient Training | Accept (poster) | Summary: The authors propose a new approach, Mango, to use a small model as initializer to larger models in order to speed up the training of the larger model. Mango proposes to learn a mapping of weights from the smaller model to the larger model, not layer by layer or other restricted form of mapping, but a full mapp... | Rebuttal 1:
Rebuttal: We appreciate the reviewer's valuable comments!
> The most interesting regime for this class of method is with bigger transformers since they are the most power hungry. The paper would gain a lot of impact by showing that Mango works for larger models (for instance to initialize a ViT-G from a Vi... | Summary: This paper addresses the issue of high resource costs in training large AI models from scratch. To this end, the authors propose a novel approach that linearly correlates each weight of the larger target model to all weights of the smaller pre-trained model, employing multi-linear operators to decrease complex... | Rebuttal 1:
Rebuttal: We appreciate the thorough feedback provided by the reviewer!
> In the proposed method, the authors first concatenate weights across layers. However, given that the row count of $\mathbf{W}_j^{OUT}\in \mathbb{R}^{kI\times O}$ differs from that of other weight matrices, how is this discrepancy add... | Summary: This paper proposes a new method for reusing pre-trained models to a larger target model. They propose mango which decomposes the parameters into 4 smaller matrices and then learns to grow these parameters. They show their method requires less FLOPS to achieve as good as performance other baselines across visi... | Rebuttal 1:
Rebuttal: We appreciate the reviewer for the meaningful comments!
> Missing background on tensor ring matrix product - since this is the main method proposed
Thanks for the suggestion. We will add the background of TR-MPO in the "Preliminary" section.
> Is the acceleration in terms of FLOPS or wall clock... | Summary: This paper studied the problem of solving a proper initialization for large deep learning models on top of smaller pre-trained networks. The key idea is to concatenate the weights of pre-trained models to a large matrix, and map the matrix to the parameter space of the larger model to be trained with a multi-l... | Rebuttal 1:
Rebuttal: We thank the reviewer for the thoughtful comments!
> The writing needs to be improved. It is difficult to understand Eqs. (4-5), which are the key contributions of this paper.
Eqs. (4-5) are both sum of element-wise production among tensors. The subscripts (e.g., $b_{\ast}, i_{\ast}, o_{\ast}, l... | Rebuttal 1:
Rebuttal: We deeply appreciate the insightful and considerate feedback provided by all of the reviewers. We are glad that the reviewer kqva notes that efficient training is practically important, and all the reviewers (u4ot, kqva, ShVL, 9N2L, and cdCX) consider the proposed Mango can achieve significant ac... | NeurIPS_2023_submissions_huggingface | 2,023 | Summary: The contribution of this paper lies in proposing a new method to accelerate the training of large models by leveraging the correlations among all weights in the model, instead of just mapping partial weights as in existing methods. The authors' approach utilizes multi-linear operators to reduce computational a... | Rebuttal 1:
Rebuttal: We are thankful for the insightful comments of the reviewer!
> In the experiments of the paper, only DeiT-B, BERT-Base, and GPT-Base were used, and no further exploration was done on larger models.
We construct an experiment to train a large model ViT-G-1B. We plot the early result in the upload... | null | null | null | null | null | null |
DeepPCR: Parallelizing Sequential Operations in Neural Networks | Accept (poster) | Summary: This work introduces DeepPCR, a method for parallelizing sequential computations with an application in neural networks. The method is based on the assumption that the target sequence is Markovian, and if so, DeepPCR can leverage Parallel Cyclic Reduction to convert the sequence into a parallel version that ex... | Rebuttal 1:
Rebuttal: >W1 As one of the tradeoffs here is memory vs compute time, it would be great to see an analysis of space complexity with respect to the Jacobian blocks/model width, and a detailed memory profile breakdown at training time (which can be obtained using a profiler).
The reviewer raises an excellent... | Summary: This paper presents a method called DeepPCR for converting sequential operations in Deep Neural Networks into parallel ones, thus accelerating DNN training, inference, and denoising procedure in diffusion models. The key idea of DeepPCR is to interpret a sequential operation of L steps that satisfies the Marko... | Rebuttal 1:
Rebuttal: > W1 The general idea of interpreting a sequential operation as the solution of a large system of equations is not brand new [36]
Indeed this intuition is not new: as mentioned in the paper, the first theorisation of this comes from [30] (a work also referenced in [36]). Still, we do not claim th... | Summary: In this paper, the authors proposed an algorithm (DeepPCR) which parallelizes typically sequential operations used in inference and training of neural networks. DeepPCR is based on interpreting a sequence of L steps as the solution of a specific system of equations, which got recovered using the Parallel Cycli... | Rebuttal 1:
Rebuttal: > W1.1. The motivation is not very clear.
The main motivation behind the proposed algorithm lies in providing an additional venue for speeding-up training and inference in NNs, to be used in conjunction with already-established parallelisation methods: this is particularly useful in time-sensitiv... | null | null | Rebuttal 1:
Rebuttal: We are grateful to the reviewers for the comments on our work, and their appreciation for the soundness of the manuscript and for the impactfulness of our proposed algorithm. We are encouraged that all reviewers found that the paper is well-written, that we “provided a comprehensive result to demo... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Learning Environment-Aware Affordance for 3D Articulated Object Manipulation under Occlusions | Accept (poster) | Summary: This work formulates environment-aware affordance for articulated object manipulation, which incorporates both object-level per-point actionable priors and environment constraints. First, this work identifies limitations of previous works on object-level affordance learning, including over-simplified robot man... | Rebuttal 1:
Rebuttal: Thanks you! We have provided responses below.
> A clear definition of per-point affordance in problem formulation.
Thanks for your suggestion! We follow [1] to define per-point affordance as an actionability score of each point on the object. Interactable means literally movable towards the inte... | Summary: The paper introduces the environment-aware affordance task, taking into account both object-centric affordance and environmental constraints. More specifically, the proposed task is designed to incorporate various occlusion cases and the robot's position. In order to tackle the inherent challenge of combinato... | Rebuttal 1:
Rebuttal: Thanks for valuable comments! We have addressed them below. Hope to hear back from you if you have further questions!
> Whether the addition of simple heuristics can yield positive results.
We have conducted the corresponding Where2Act with the heuristic experiment with the successful rates of *... | Summary: Given a 3D point cloud and a robot, the method predicts per-point actionable affordance. More specifically it predicts whether a given point is pushable or pullable by the robot in a given position. The method takes into account environment occlusion. The paper presents a data-efficient training framework that... | Rebuttal 1:
Rebuttal: Thanks you! We have provided responses below.
> The action space is too small. Only pushing and pulling are considered.
We follow the settings of a series of affordance learning works [1, 3, 4] and use pushing and pulling as the action primitives, as they are the most common action primitives, a... | Summary: The paper presents a novel environment-aware affordance prediction framework for robot interactions (pushing/pulling) with parts of articulated objects (drawers and doors). The proposed approach consists primarily of two novel improvements: a robot-target conditioned occlusion field and a contrastive learning ... | Rebuttal 1:
Rebuttal: Thanks for your valuable comments! We have addressed them below. Hope to hear back from you if you have further questions!
> High-level concepts were repeated often.
Thanks for your suggestion! We will prune repetitive concepts to make space for crucial details.
> Clarifications on carry-over... | Rebuttal 1:
Rebuttal: We extend our gratitude to reviewers for their careful reading, meticulous feedback and valuable insights.
We are glad that reviewers unanimously agree that our work (1) proposes meaningful environment-aware affordance (zyYG: “a step up from previous work”, ntD4: “novel and insightful”, wNB2: “add... | NeurIPS_2023_submissions_huggingface | 2,023 | Summary: This paper introduces a framework for learning environment-aware affordance in occluded 3D object manipulation. It tackles combinatorial explosion and improves data efficiency. The authors showcase their superior framework using benchmarking multi-object full-robot environments in the SAPIEN simulator. Their c... | Rebuttal 1:
Rebuttal: Thanks for your valuable comments. We have addressed them below, and hope to hear back from you if you have further questions!
> Lack of real-world exp from different perspectives and scenarios.
We have provided **3 pairs of comparisons of affordance and their corresponding manipulation policies... | null | null | null | null | null | null |
GenS: Generalizable Neural Surface Reconstruction from Multi-View Images | Accept (poster) | Summary: This paper proposed a method for neural implicit surface reconstruction from multi-view images. The proposed method is generalizable without per-scene overfitting. The key component of the proposed method is a multi-view feature volume constructed from a multi-scale FPN network. During training process, the mu... | Rebuttal 1:
Rebuttal: We sincerely thank the reviewer XxbL for your time and valuable comments. Here, we will explain your concerns point by point. Due to words limitation, we put most experiments in the uploaded PDF.
**Q1: Missing the citation of [1], which also adopts a multi-scale strategy**
Sorry for missing this... | Summary: This paper proposed a generalizable Nerf method. The novelty of this method lies in a proposed multi-scale feature-metric consistency loss and a proposed view contrast loss. The feature-metric consistency loss encourages the features extracted from different images at the same surface point to be consistent wi... | Rebuttal 1:
Rebuttal: We sincerely thank the reviewer cro4 for recognizing our work and the valuable comments. Here, we will explain your concerns point by point.
**Q1: The grammar and sentence structure should be further polished.**
Thanks for pointing them out, and we appreciate these suggestions. We will seriously... | Summary: This paper aims to address the problem of multi-view reconstruction using neural implicit representation. The key to improving both efficiency and efficacy is a deep network that predicts a generalizable multi-scale feature volume encapsulating the geometry and appearance of the scene given a set of posed imag... | Rebuttal 1:
Rebuttal: We sincerely thank the reviewer EdqW for recognizing our work and the valuable comments. Here, we will explain your concerns point by point.
**Q1: The multi-scale mechanism has been explored in existing method.**
The multi-scale strategy is effective and has inspired and advanced many fields. We... | Summary: This paper proposes a generalisable neural surface reconstruction approach that is able to learn an SDF function from volume features and does not need per-scene training.
Strengths: 1. This paper uses a multi-scale volume to model the feature field.
2. The view contrast loss is novel to me which helps to enh... | Rebuttal 1:
Rebuttal: We sincerely thank the reviewer Tu79 for recognizing our work and the valuable comments. Here, we will explain your concerns point by point.
**Q1: Lacking comparison with latest advanced approaches: VolRecon [1] and ReTR [2].**
Thanks for your suggestion, and the reasons why we don’t compare wit... | Rebuttal 1:
Rebuttal: We sincerely thank all reviewers for their detail and constructive comments. We are glad that the reviewers appreciate the novelty (Tu79, EdqW, cro4), state-of-the-art and impressive results (cro4), reasonable ablation studies (EdqW), good technical soundness (cro4, XxbL), clear writing (XxbL). We... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Modality-Agnostic Self-Supervised Learning with Meta-Learned Masked Auto-Encoder | Accept (poster) | Summary: This paper presents Meta-learned Masked Auto-Encoder (MetaMAE), a novel modality-agnostic self-supervised learning (SSL) framework that leverages meta-learning to improve the transfer abilities of Masked Auto-Encoder (MAE). The authors reinterpret the mask reconstruction task of MAE as a meta-learning task and... | Rebuttal 1:
Rebuttal: Dear Reviewer Pe9M,
We sincerely appreciate your efforts and insightful comments to improve the manuscript. We address your comment below.
---
**[Q1] Significant difference in hyper-parameters for various downstream tasks, e.g., masking ratio, batch size, and decoder depth.** \
**[A1]** For you... | Summary: The authors of this paper address modality-agnostic self-supervised learning (SSL), a significant yet under-researched problem with various real-world applications. They explore the potential of the Masked Auto-Encoder (MAE) in this context and propose SSL framework, MetaMAE, which enhances MAE with meta-learn... | Rebuttal 1:
Rebuttal: Dear Reviewer 9KfC,
We sincerely thank you for your helpful feedback and insightful comments. We address your comments and questions below.
---
**[Q1] Explanation of the experimental setup for meta-learning-based gradient update, e.g., how the support and query set is constructed during trainin... | Summary: This paper studies self-supervised learning (SSL) for multiple modalities (e.g. tabular data, images, text, speech). Central to this paper are Masked Auto-Encoders (MAEs), a popular SSL training paradigm where models are tasked to reconstruct inputs from randomly masked subsets. The authors interpret this reco... | Rebuttal 1:
Rebuttal: Dear Reviewer b3vS,
We sincerely thank you for your positive feedback and insightful comment to improve the manuscript. We address your comment below.
---
**[Q1] Training a single model jointly on all the modalities and benchmarks to build a general model that could simultaneously perform all t... | Summary: This work interprets MAE as a meta-learning framework to formulate a modality-agnostic SSL framework. Specifically, they formulate the masked data and unmasked data (nearby tokens) as the support set and the query set, respectively. Moreover, the authors proposed to use gradient-based meta-learning on a decode... | Rebuttal 1:
Rebuttal: Dear Reviewer 2wfj,
We sincerely thank you for your helpful feedback and insightful comments to improve the manuscript. We address your comments and questions below.
---
**[Q1] Relationship between the proposed method (i.e., MAE+meta-learning) and modality-agnostic SSL.** \
**[A1]** We would li... | Rebuttal 1:
Rebuttal: Dear reviewers and AC,
We sincerely appreciate your valuable time and effort spent reviewing our manuscript.
As reviewers highlighted, we propose a novel (GRPA, 2wfj, Pe9M, 3Z8m) and interesting (2wfj, b3vS, 9KfC) approach with strong empirical results (ALL Reviewers) and extensive experiments (G... | NeurIPS_2023_submissions_huggingface | 2,023 | Summary: The paper titled "Modality-Agnostic Self-Supervised Learning by Meta-Learning" presents a novel framework called MetaMAE (Meta-learned Masked Auto-Encoder) for modality-agnostic self-supervised learning (SSL). The authors propose an innovative interpretation of Masked Auto-Encoder (MAE) as a meta-learning fram... | Rebuttal 1:
Rebuttal: Dear Reviewer GPRA,
We sincerely thank you for your positive feedback on our paper! We address your question below. If you have any other questions or suggestions, please let us know!
Best, \
Authors
---
**[Q1] What might help in handling diverse and complex data with this approach? What is ke... | Summary: This paper proposes MetaMAE, a novel framework to improve Masked Autoencoder (MAE) for modality-agnostic self-supervised learning. Unlike many SSL methods reliant on domain-specific inductive biases, MetaMAE enhances MAE with two meta-learning techniques: latent adaptation via gradient-based meta-learning to o... | Rebuttal 1:
Rebuttal: Dear Reviewer 3Z8m,
We sincerely appreciate your efforts and insightful comments to improve the manuscript. We address your comment below.
---
**[Q1] Why meta-learning improves MAE for modality-agnostic SSL?**\
**[A1]** We remind that the masked modeling of MAE can be interpreted as an *amortiz... | null | null | null | null |
BanditPAM++: Faster $k$-medoids Clustering | Accept (poster) | Summary: This paper studies a clustering problem (k-medoid) in which we have to select k centers among the datapoints in order to minimize the sum of distances of points to their closest center. For this problem, there is an intuitive algorithm which works well empirically (the PAM algorithm): (1) add greedily centers ... | Rebuttal 1:
Rebuttal: We thank Reviewer mktW for their thoughtful comments and helpful feedback. We provide a point-by-point response below.
Response to W1: We would like to clarify our contributions and their novelty.
BanditPAM++ improves upon BanditPAM via two novel techniques: the Virtual Arms (VA) technique and P... | Summary: The paper introduces BanditPAM++, a faster version of the BanditPAM algorithm for k-medoids clustering. It incorporates two improvements: the Virtual Arms technique, which reuses loss changes within each SWAP iteration, and the Permutation-Invariant Caching method, which samples reference points and reuses dis... | Rebuttal 1:
Rebuttal: We thank Reviewer s8tF for their thoughtful comments and helpful feedback. We provide a point-by-point response below.
Response to W1.1: We apologize for the confusion. The algorithm we refer to as BanditPAM++ is the version of BanditPAM with both the Virtual Arms (VA) technique and the Permutati... | Summary: This paper studies the $k$-mediods clustering problem. Given an instance $P$, the goal of the $k$-mediods problem is to find a subset $C \subseteq P$ of size $k$ such that the sum of the defined distances from data points to $P$ to the centers in $C$ is minimized. In previous heuristic solutions for the $k$-me... | Rebuttal 1:
Rebuttal: We thank Reviewer hYVb for their thoughtful comments and helpful feedback. We provide a point-by-point response below.
Response to W1, Q3, and Q4 (sub-Gaussianity): We agree that it is ideal to remove the sub-Gaussian assumption and discuss this limitation in the revised paper. However, we argue ... | Summary: The authors developed an algorithm (called BanditPAM++) which is claimed to be more efficient than banditPAM, the state-of-the-art algorithm for k-medioid (which is the version of the k-means problem where the centers must be input points). They prove that with high probability the two algorithms produce the s... | Rebuttal 1:
Rebuttal: We thank Reviewer zP71 for their thoughtful comments and helpful feedback. We provide a point-by-point response below.
Response to W1 and Q1: From a theoretical point of view, the computational complexity of BanditPAM++ is same as that of BanditPAM with respect to the number of datapoints $n$ (bo... | Rebuttal 1:
Rebuttal: Attached please find additional experimental results that we refer to in our rebuttals to reviewers.
Pdf: /pdf/1469b03f103daf19847cee0e033a8b31f4b53428.pdf | NeurIPS_2023_submissions_huggingface | 2,023 | Summary: The paper
* presents a new algorithm for the k-medoids problem called BandidPAM++, which is an extension of BanditPAM
* analyzes the algorithm theoretically showing that it produces essentially the same output as BanditPAM and uses O(n log n) distance computations
* evaluates the algorithm empirically comp... | Rebuttal 1:
Rebuttal: We thank Reviewer oPsi for their thoughtful comments and provide a point-by-point response below.
Weakness 1: We have now added detailed explanations of the pseudocode at the end of Section 3 and made better use of space. BanditPAM++ improves upon BanditPAM via two techniques: the Virtual Arms (V... | null | null | null | null | null | null |
Federated Spectral Clustering via Secure Similarity Reconstruction | Accept (poster) | Summary: The paper proposes a secure kernelized factorization method for federated spectral clustering on distributed data. It provides convergence guarantee of the optimization, reconstruction error bound of similarity matrix, and sufficient condition of correct clustering. It also presents guarantees of differential ... | Rebuttal 1:
Rebuttal: The authors sincerely appreciate the insightful comments made by the reviewer.
**Response to weakness 1:**
This issue gets involved with the unexpected drop-out of participated clients in each round, which is actually a common phenomenon in network communication. But here, we just want to claim ... | Summary: This paper solves unsupervised spectral clustering by reconstructing the similarity matrix under a Federated Learning (FL) and Differentially Private (DP) setting. The authors argue that since the similarity matrix requires computation across pairs of clients it might not be directly FL-compatible. Thus, the a... | Rebuttal 1:
Rebuttal: First, the authors sincerely thank you for recognizing our work. We provide the responses to your five questions as follows.
**Response to weakness 1:**
Thanks for the insightful comment. Our goal is to compute a kernel matrix $\mathcal{K}(\mathbf{X}, \mathbf{X})$ without sending the data to the... | Summary: This paper focuses on computing spectral clustering in the setting of horizontal federated learning. The main contributions of this paper are: (1) proposing a secure method for computing the adjacency matrix to achieve spectral clustering, (2) providing rigorous theoretical analysis to ensure the algorithm sat... | Rebuttal 1:
Rebuttal: The authors sincerely appreciate the insightful comments made by the reviewer.
**Response to weakness 1:**
As claimed by Chai et al.[1], the general framework of federated matrix factorization may still leak users' raw data. Therefore, extra measure needs to be taken for enhancing the security o... | Summary: This paper introduces a secure kernelized factorization approach of spectral clustering. The proposed method works on distributed datasets while maintaining privacy. The proposed technique approximates the kernel matrix required as in the spectral clustering, in an effort to preserve the privacy. This paper an... | Rebuttal 1:
Rebuttal: The authors thank the reviewer for the comments. Our responses to your comments and questions are as follows.
**Response to Weakness 1:**
First, due to the kernel mapping, the minimization problem for $\mathbf{Z}_p$ is nonconvex and has no closed-form solution. This implies that it is very diffi... | Rebuttal 1:
Rebuttal: We sincerely thank the area chairs and all reviewers. We have responded to the comments given by the reviewers carefully.
Here we summarize a few important points of our rebuttal or revision.
1. For Reviewer EUeG, we provide a formula to present a privacy-utility trade-off, namely, a trade-off be... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Promises and Pitfalls of Threshold-based Auto-labeling | Accept (spotlight) | Summary: The paper studies an important problem. It theoretically analyzes TBAL systems and develops tradeoffs between
the quantity of manually-labeled data and the quantity and quality of auto-labeled data. Empirical results on real/synthetic data also support their findings.
Their results provide two crucial insights... | Rebuttal 1:
Rebuttal: We sincerely appreciate your feedback and acknowledgment of the paper's strengths. Being the first theoretical study of TBAL systems, our research offers essential guarantees on auto-labeling error and coverage using learning theoretic tools. The experiments conducted not only validate our theoret... | Summary: This paper provides an analysis of Threshold-Based Auto-Labeling (TBAL) systems by deriving sample complexity bounds on the amount of human-labeled validation data required for guaranteeing the quality of the machine-labeled data. The paper provides two main insights: (1) large amounts of unlabeled data can be... | Rebuttal 1:
Rebuttal: We are delighted to receive your positive feedback and recognition of the strengths in our paper.
Our work represents the first attempt to derive sample complexity bounds on the validation data for the confident utilization of TBAL systems. Moreover, we are happy to know that the paper's clarity a... | Summary: Threshold-based auto-labelling (TBAL) is widely adopted in academic and industrial settings and this is the first work that studies TBAL from a theoretical standpoint, answering the questions of 1) Does TBAL works even if our models are not perfect and 2) what are the key requirements to having a successful TB... | Rebuttal 1:
Rebuttal: Thank you for your valuable feedback and for recognizing the strengths of our paper. We greatly appreciate your acknowledgment of the significance of TBAL, which is heavily used in practice. As you rightly pointed out, our paper is the first theoretical study of TBAL and provides an understanding ... | null | null | Rebuttal 1:
Rebuttal: Thank you all for the thoughtful and positive feedback on our paper. With your constructive suggestions, we have been able to further strengthen our work with additional experiments and improved clarity.
Your questions and comments have provided us with valuable insights and we answer each of th... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
ANPL: Towards Natural Programming with Interactive Decomposition | Accept (poster) | Summary: In this paper, the authors introduce Abstracted Natural Programming Language (ANPL), a programming language and system to improve Large Language Model (LLM) assisted programming. ANPL programs are structured in a top-down manner; a top level program called the "sketch" solves the task by specifying the data-fl... | Rebuttal 1:
Rebuttal: Thanks for your review comments! We provide detailed responses to all of your concerns. We'll revise our paper to make your concerns clear. We have a space limit so please let us know of any additional comments.
## Weaknesses
### Presentation
Thanks for the suggestions and we'll revise the paper.
... | Summary: This paper proposes ANPL, a programming system for user interaction with an LLM for code generation. The main idea is to have the user generate high-level dataflow and program structure, while directing the LLM to fill in the "holes." They present results of a user study of ANPL, where users are asked to write... | Rebuttal 1:
Rebuttal: Thank you for your review! Your concerns mainly involve some questions about the experimental setup. In response to this, we have provided comprehensive explanations and conducted an additional experiment on APPS and complicated programs. The results show that ANPL still outperforms baselines and ... | Summary: The paper proposes ANLP, a framework that interactively gets the definition of the data flow of a program from user input, and then fill the "holes" in the sketched program using LLMs. The authors conducted comprehensive human experiments on the Abstraction and Reasoning Corpus by letting human subjects solve ... | Rebuttal 1:
Rebuttal: Thank you for your review comments! Your concerns were about the differences between ANPL and other similar work like Parsel, as well as the applicability of ANPL to a broader range of complicated real-world tasks. For the former, we have provided a detailed explanation and examples in our respons... | Summary: ANPL proposes a Sketch guided program synthesis using Natural Language in an interactive fashion.The results show better interactive coding than base LLM.
Strengths: The paper shows potential in using sketch guided methods in conjunction with LM inference. It can be a good tool to use for debugging and potent... | Rebuttal 1:
Rebuttal: We thank you for your review comments! Your main concerns were about the generality of the programming framework and the adequacy of the references to related works.
We have tried our best to address your concerns by incorporating additional experiments and adding related works. Please refer to th... | Rebuttal 1:
Rebuttal: We thank the reviewers and ACs for their time and effort.
We are glad the reviewers acknowledge the following contributions:novel and reasonable for proposing a LLM programming system by introducing extensive (i.e. non 1-shot)user interactions(neh7, tgZy, ecZq); comprehensive human study and str... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
How to Turn Your Knowledge Graph Embeddings into Generative Models | Accept (oral) | Summary: This paper provides an new perspective for looking at the popular knowledge graph embedding (KGE) method for knowledge graph completion. Specifically, KGEs models are re-interpreted as circuits. The paper first prove the circuits-based equivalence of score functions of several KGE models, and, in order to regu... | Rebuttal 1:
Rebuttal: We thank the reviewer for the positive feedback and pointing out how GeKCs are indeed competitive with existing KGE models.
> Could other KGE models such as TransE, RotatE be viewed as circuits?
This is a super interesting question. Here are some ideas.
To our knowledge, distance-based KGE mode... | Summary: The authors show that standard KGE models, such as COMPLEX, can be represented as structured computational graphs called circuits. They propose a new interpretation of these graphs and their parameters to create efficient and expressive probabilistic models over triples in a KG, called generative KGE circuits ... | Rebuttal 1:
Rebuttal: We are grateful to the reviewer for valuing the significance of our contribution as well as its execution.
> “The constraint adding part can be extended into another work”
Thanks for recognizing the significance of our work for reliable decision making in safety-critical scenarios. We have some... | Summary: This paper converts score functions of knowledge graph embeddings into generative KGE circuits (GeKCs) by restricting activations to be non-negative or squaring their outputs.
The changes come with little or no loss of link prediction performance,
while enabling probabilistic interpretations, exact maximum-l... | Rebuttal 1:
Rebuttal: We thank the reviewer for all the comments. We believe we addressed all the concerns and are happy to follow up in the discussion phase.
>The difference between GeKCs and other KGEs is constraining scores or squaring?
Yes, although apparently simple, this modeling difference enables the semantic ... | Summary: This paper provides a reduction from several well-known knowledge graph embedding (KGE) methods (e.g., TuckER, CP, etc.) to *decomposable* circuits $\phi$ which can be efficiently computed in $O( (|E| + |R|) |\phi|)$ (note that this is mostly due to the decomposable nature of TuckER and its derivatives). The a... | Rebuttal 1:
Rebuttal: We thank the reviewer for appreciating the rigor, novelty and direction of our research! Such a simple idea can open up many valuable perspectives.
> There are other KGE methods (e.g. graph neural networks) that are not covered by this paper.
Indeed, connecting circuits with more general KGEs su... | Rebuttal 1:
Rebuttal: We thank all reviewers for the time spent reviewing the paper and recognizing the **significance** (“useful and elegant theoretical framework” – v5Gr, “solid theory” – P87h, “deserves a wide audience for its insights and rigorous analysis” – t9NK) and **novelty** (“Novel combination” – P87h, “maki... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Spectral Invariant Learning for Dynamic Graphs under Distribution Shifts | Accept (poster) | Summary: This paper studies the distribution shifts, i.e., out-of-distribution generalization, on graph data and focuses on distribution shifts on temporal dynamic graphs. To this end, the authors propose a model from the spectral domain and resort to the established techniques of invariant learning. Experiments on syn... | Rebuttal 1:
Rebuttal: We sincerely appreciate the insightful comments provided by the reviewer. We have carefully considered each point raised and would like to respond as follows:
> How the mixed loss $\mathcal{L}_{INV}$ is calculated and how the sampling for the variant patterns is conducted?
Thank you for your com... | Summary: Dynamic graph neural networks are still facing the difficulty in handling distribution shifts. This work discovers that in some cases, distribution shifts are unobservable in the time domain but not in the spectral domain. There are currently no studies to handle cases involving distribution shifts in the spec... | Rebuttal 1:
Rebuttal: We sincerely thank the reviewer for the valuable suggestions. We respond to the reviewer’s comments point by point as follows.
> It would be great to include the F1 scores as metrics for comparisons of different methods in node classification tasks.
Thank you for your suggestion. Following your ... | Summary: This paper addresses the problem of dynamic graph learning under distribution shifts. Distinct from the previous work that focuses on the time domain, the authors argue that there exists patterns that are not easily distinguished in the time domain while could be easier to recognize in the spectral domain. Som... | Rebuttal 1:
Rebuttal: We would like to express our sincere appreciation to the reviewer for providing us with detailed comments and suggestions. We have carefully reviewed each comment and offer the following responses:
> In line 87, the condition 2 assumes that the observed data is composed of invariant and variant p... | Summary: This paper finds that there are cases with distribution shifts observable in the spectral domain while difficult to distinguish in the time domain. Based on this characteristic, the authors propose Spectral Invariant Learning for Dynamic Graphs under Distribution Shifts (SILD) to handle distribution shifts on... | Rebuttal 1:
Rebuttal: We would like to express our sincere gratitude to the reviewer for providing us with detailed comments and insightful questions. We have carefully considered the reviewer's feedback and would like to address each point as follows:
> In Figure 2, the spectrums block (bottom-middle) seems a little ... | null | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
IMPRESS: Evaluating the Resilience of Imperceptible Perturbations Against Unauthorized Data Usage in Diffusion-Based Generative AI | Accept (poster) | Summary: This paper investigates the reliability of methods for imperceptibly perturbing images to degrade the performance of latent diffusion models for style mimicry or editing. The authors observe that imperceptible perturbations designed to hamper diffusion-based image generation systems often result in perceptible... | Rebuttal 1:
Rebuttal: # Response to Reviewer EvE8
Thanks for the constructive comments!
**Q1**: Figures 1/4 are hard to read
**A1**: We are sorry for the confusion. Figures 1 and 4 showcase images generated by a model that was fine-tuned on either clean, protected, or purified images. We've detailed this in the ca... | Summary: This paper proposes IMPRESS, a new purification method to remove imperceptible perturbations that are claimed to be able to prevent unauthorized data usage in diffusion-based image generation. Specifically, IMPRESS aims to produce a purified version of the perturbed image by minimizing the distance between the... | Rebuttal 1:
Rebuttal: # Response to Reviewer DYEa
Thanks for the constructive comments!
**Q1**: More Baselines
**A1** Please see **CA2** in our common response to all reviewers
**Q2**: The first to study Imperceptible Perturbations
**A2**: Thank you for your suggestion. What we want to say here is prior works ar... | Summary: This work is focused on testing the vulnerability of approaches that prevent unauthorized data usage in diffusion models. The paper shows that existing approaches for this problem are fragile, and leads to difference b/w original and diffusion reconstructed image. Then, based on the previous observation the pa... | Rebuttal 1:
Rebuttal: # Response to Reviewer ULFY
Thanks for the constructive comments!
**Q1:** More Complicated Baselines
**A1:** Please see **CA2** in our common response to all reviewers
**Q2:** The Impact of Different JPEG Compression Rates
**A2:** In our paper, we used the JPEG compression rate that provided ... | Summary: In this work, the authors showed some intellectual protection methods for diffusion methods via imperceptible perturbations could be removed, hence questing the effectiveness of such methods. The authors observed perceptible inconsistency between the pair of original samples and the reconstructed ones by diffu... | Rebuttal 1:
Rebuttal: # Response to Reviewer vDR3
Thanks for the constructive comments!
**Q1**: limited novelty
**A1**: We respectively disagree and believe that our work is novel in the following aspects:
1) The issue we delve into (vulnerabilities of imperceptible perturbations based defenses) is new and critica... | Rebuttal 1:
Rebuttal: # Response to Common Question for All Reviewers
We thank all the reviewers for your valuable comments!
**We have incorporated all experiments suggested by the reviewers:**
1. We've included a broader range of baseline methods, encompassing more potent traditional denoising techniques, combinati... | NeurIPS_2023_submissions_huggingface | 2,023 | Summary: Diffusion-based image generation models, such as Stable Diffusion or DALLE2, are able to learn from given images and generate high-quality samples following the guidance from prompts. However, such ability also brings unauthorized data usage problem. In response, several attempts have been made to protect the ... | Rebuttal 1:
Rebuttal: # Response to Reviewer UShu
Thank you for your valuable comments!
**Q1:**: Lacks theoretical support for NeurIPS
**A1:** Indeed our paper does not provide much theoretical results and we plan to leave the theoretical study as our future work. Nonetheless, we believe it still represent signific... | null | null | null | null | null | null |
On Imitation in Mean-field Games | Accept (poster) | Summary: The paper presents a novel solution concept Nash imitation gap (NIG) for imitation learning problems in Mean-field games. For both population-independent and population-dependent dynamics, the author provides the upper-bounds of NIG for behavioral cloning (BC), vanilla adversarial imitation learning (vanilla-A... | Rebuttal 1:
Rebuttal: We would like to thank the reviewer for the insightful comments. We would like to thank the reviewer to recognize the originality of our paper (¨a novel solution concept, NIG, which generalized the classic imitation gap, and shows the upper-bounds of NIG for several different settings¨), and the c... | Summary: This paper focuses on the imitation learning problem in mean-field games. The authors propose a new performance metric Nash imitation gap. They show that when the transition kernels of MDP are independent of the mean fields, the imitation learning problem reduces to that in the single-agent MDP. When the trans... | Rebuttal 1:
Rebuttal: We would like to thank the reviewer for the insightful comments and for recognizing that our work is the first one to provide theoretical guarantees, which shows how previous approaches have exponential dependency on the horizon. Thanks also for recognizing that we ¨propose a new adversarial algor... | Summary: The authors prove new upper bounds on the error in imitation learning for mean-field games. They then propose a new approach based on using mean-field control, which might provide a more reasonable imitation goal for imitation learning.
Strengths: The paper is clear in its description of connections and provi... | Rebuttal 1:
Rebuttal: We would like to thank the reviewer for the insightful comments. We would like to thank the reviewer also for recognizing the clarity of our work and how our paper ¨shows a gap in current literature and proceeds with an ample analysis of how to fill it¨.
We answer the raised questions below. Tha... | Summary: In this work, the authors propose a newly defined measure called the Nash imitation gap for imitation learning in mean-field games (MFG). They establish upper bounds for the Nash imitation gap of behavioral cloning and adversarial imitation learning in both population independent and dependent dynamics in MFG.... | Rebuttal 1:
Rebuttal: We would like to thank the reviewer for the insightful comments and for recognizing the clarity and the novelty of the work. We answer the raised questions below.
### Weaknesses
1. ¨*the lack of comprehensive experiments weakens the justification for using the Nash imitation gap as a measure for... | Rebuttal 1:
Rebuttal: We would like to thank the reviewers for the time they spent reviewing our paper and for their valuable feedback.
All the reviewers recognized the **clarity** and the **novelty** of our work, providing **new** and **first** theoretical guarantees for this setting. Moreover, MgoB, underlined how ... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Towards Consistent Video Editing with Text-to-Image Diffusion Models | Accept (poster) | Summary: Previous works on one-shot video editing might produce results of unsatisfied consistency with text prompt as well as temporal sequence. This paper presents that such inconsistency is caused by covariate shift in the feature space. Therefore, they propose Shift-restricted Temporal Attention Module (STAM) and F... | Rebuttal 1:
Rebuttal: __Q1__: The influence of covariate shift may be small. It is better to analyze the issue quantitatively.
__A1__: Thanks for pointing out. We agree with your point that the implementations have potentially alleviated the covariate shift. But, due to the resonance of multiple variables, the issue i... | Summary: This paper carefully analyzes the problems of current one-shot video editing and proposes two techniques: Shift restricted Temporal Attention Module (STAM) and Fine-coarse Frame Attention Module (FFAM), to address these issues. Several quantitative and qualitative experiments are conducted to verify the effect... | Rebuttal 1:
Rebuttal: __Q1__: It would be better to quantitatively analyze components besides qualitative visual comparison.
__A1__: Thanks for pointing out. We have addressed this by including the quantitative results in the Table 1 of the provided PDF file.
1. By replacing the TA module with the proposed STAM (+ST... | Summary: This paper claims that existing models suffer from inconsistent video editing results due to covariate shift caused by additional modules for learning temporal information. It proposes a novel model addressing this issue through two modules: Shift-restricted Temporal Attention (STAM) and Fine-coarse Frame Atte... | Rebuttal 1:
Rebuttal: __Q1__: Results in table1 and Figure 3 are not consistent with the main target and point of the proposed method.
__A1__: In our work, the consistency refers to (L44-46) both semantic alignment of text and edited video, and temporal consistency in cross frames.
In this context, the results and y... | Summary: The paper provides a comprehensive theoretical analysis of the covariate shift problem in current one-shot fine-tuning methods. In response, the authors propose a novel Shift-restrict Temporal Attention Module, designed to enhance the editing capability. Additionally, they introduce a Fine-coarse Frame Attenti... | Rebuttal 1:
Rebuttal: __Q1__: It would be valuable to generalize analysis to account for a broader range of situations.
__A1__: Thanks for pointing out. We recognize this limitation (Line 335-337), and explore two perspectives for generalization:
1. Distribution Transition Perspective: In the proof, we approach the c... | Rebuttal 1:
Rebuttal: We thank all the reviewers for their thoughtful reviews and insightful comments on our work!
It is encouraging to hear the feedback from reviewers that:
1. The paper is commendably easy to follow. It is well-written and presented in terms of organization, logic, commendable clarity and cohesi... | NeurIPS_2023_submissions_huggingface | 2,023 | Summary: This paper proposes $EI^2$ for the consistent text-driven video editing. To address the semantic disparity problem, it provides a detailed theoretical analysis and proposes Shift-restricted Temporal Attention Module (STAM) to tackle the covariate shift problem introduced by the temporal consistency module. To ... | Rebuttal 1:
Rebuttal: __Q1__: Lack of quantitative ablation study to validate the effectiveness of the STAM and FFAM.
__A1__: Thanks for pointing out. We have addressed this by including the quantitative results in the Table 1 of the provided PDF file.
1. By replacing the TA module with the proposed STAM (+STAM) in ... | null | null | null | null | null | null |
Elastic Decision Transformer | Accept (poster) | Summary: This paper tackles the "trajectory stitching" problem of Decision Transformer. It proposes a method called Elastic Decision Transformer (EDT) that dynamically adjusts the history at inference time such that longer history is maintained for optimal trajectories and shorter history is maintained for sub-optimal ... | Rebuttal 1:
Rebuttal: We thank the reviewer for their insightful comments. We address your comments in the following.
---
**Q**: "The applicability of the proposed method seems to be limited since it is only applicable to DT and its variants."
**A**: In this work, we primarily focus on improving DT and its variants.... | Summary: This paper proposes the Elastic Decision Transformer (EDT), an improvement over the existing Decision Transformer (DT). DT has had notable limitations with trajectory stitching, i.e., generating optimal or near-optimal trajectories from sub-optimal ones. EDT addresses this issue by modifying the history length... | Rebuttal 1:
Rebuttal: We thank the reviewer for their insightful comments. We address your comments in the following.
---
**Q**: A potential concern arises from the fact that the EDT's performance hinges heavily on the value maximizer. While effective in function, it tends to be less efficient in terms of training sp... | Summary: This paper proposes Elastic Decision Transformer (EDT), which helps Decision Transformer (DT) generalize offline data by trajectory stitching. Their changes to the DT are: 1) during training, adding an extra term to the objective to estimate the maximum Q values starting from a state (similar to the Implicit Q... | Rebuttal 1:
Rebuttal: We thank the reviewer for their insightful comments. We address your comments in the following.
---
**Q**: “Does this simply implement the idea behind the IQL algorithm in DT?”
**A**: Although both IQL and EDT employ expectile regression, they use it for different purposes. In IQL, expectile r... | Summary: This paper introduces the Elastic Decision Transformer (EDT), an extension of the Decision Transformer (DT) model that automatically determines the history length input during inference. Specifically, the EDT estimates the value of each candidate history length with the expectile return regression and chooses ... | Rebuttal 1:
Rebuttal: We thank the reviewer for their insightful comments.
---
**Q**: The design of the EDT appears ad-hoc.
**A**: The components proposed in EDT such as using expectile regression to estimate the maximum return and optimal history length search, are not particularly designed for specific cases. Inst... | Rebuttal 1:
Rebuttal: We thank all reviewers for their thoughtful comments and insights. We will revise our manuscript based on your feedback, and we have responded to each of your individual comments.
Summary of revisions:
* Fixing typo: “casual“ -> “causal“, “Figure 2“ -> “Table 2“ on line 234.
* Adding explanati... | NeurIPS_2023_submissions_huggingface | 2,023 | Summary: This paper proposes to vary the history length provided to a Decision Transformer to enable it to stitch suboptimal trajectories into better ones. The motivation behind this, the authors argue, is that the model can shift to an optimal trajectory best given only the current state (and not the history), while p... | Rebuttal 1:
Rebuttal: We thank the reviewer for their insightful comments. We address your comments in the following.
---
**Q**: “The main limitation of the method seems to be inference cost”
**A**: Thank you for pointing this out. We have conducted an ablation study to compare the wall clock time with the proposed ... | null | null | null | null | null | null |
Exploring the Promise and Limits of Real-Time Recurrent Learning | Reject | Summary: The authors investigate Realtime Recurrent Learning (RTRL) applied to recurrent RL, in contrast with traditional Backpropagation Through Time (BPTT). The authors' key insight is that they can significantly reduce time and space complexity by using single-layer elementwise recurrence. They propose an elementwis... | Rebuttal 1:
Rebuttal: We thank the reviewer for spending valuable time on reviewing our work. We also thank the reviewer for many positive and encouraging comments supporting the core philosophy underlying our work.
Please find our response as follows.
> *"My biggest complaint is that there is no data showing empirica... | Summary: This paper explores an alternative to back-propagation through time (BPTT) for training recurrent neural networks, called real time recurrent learning (RTRL). The alternative algorithm does not depend on past activations, and also does not need truncation (which limits past context) like in BPTT to make it tra... | Rebuttal 1:
Rebuttal: We thank the reviewer for spending valuable time on reviewing our work and for many positive and supportive comments.
Please find our response as follows:
> *"Ablate the dependence of M (episode rollout length) on R2AC."*
> *"… also to compare the best performance for both the training algorithm... | Summary: - The paper explores the use of RTRL in the case of scenario, where no approximation is needed.
- The proposed "method" is not novel per se i.e., Mozer et. al put forward the derivation of RTRL for element-wise recurrence, but to the best of reviewer's ability no one has formally wrote about the tractability o... | Rebuttal 1:
Rebuttal: We thank the reviewer for spending valuable time on reviewing our work. We also thank the reviewer for many positive comments (soundness, presentation, and contribution scores are all “3 good”).
We are not sure which aspects led the reviewer to rate our work as 4 with confidence 5 in the end, desp... | Summary: The paper modifies the LSTM architecture so that RTRL complexity, which is usually O(n^4), becomes O(n^2) where n is the number of recurrent units. This is achieved by constraining the recurrent weights to a diagonal structure. After performing some diagnostic tests, the paper presents experimental results usi... | Rebuttal 1:
Rebuttal: We thank the reviewer for spending valuable time on reviewing our work.
There are several misunderstandings to be resolved, but we believe our explanations will clarify all the concerns raised by the reviewer.
The main criticism breaks down into three points (quoting the reviewer):
1. > *“incrom... | Rebuttal 1:
Rebuttal: We thank all the reviewers for their valuable comments.
Overall, we believe that our response should resolve all main concerns raised by the reviewers.
If you find our explanations convincing, please consider increasing your scores.
For the moment, we will keep the originally submitted PDF as is,... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Symbol-LLM: Leverage Language Models for Symbolic System in Visual Human Activity Reasoning | Accept (poster) | Summary: The paper addresses activity reasoning, where the task is to detect human activities given visual inputs.
To achiever compositional generalization for activity understanding, the paper propose a new symbolic system that is capable of relational reasoning.
In the experiments, the proposed system improved naiv... | Rebuttal 1:
Rebuttal: We thank the reviewer for the feedback and hope to address the questions and concerns below.
### Logic Notations
The notations "$\models$" (semantic entailment) and "$\rightarrow$" (syntactic entailment) both describe the concept of one statement leading to another, but their emphasis is differen... | Summary: This work is in the area of human visual activity reasoning and tries to improve System-2 (logic, deliberative reasoning) integrated with System-1 (that deals with intuitivity, associativity, and correlation) to improve explainability, generalization and efficiency which is demonstrated through empirical evalu... | Rebuttal 1:
Rebuttal: We thank the reviewer for the feedback and hope to address the questions and concerns below.
### Method Details Explanation
Thanks for your advice. We will add more explanations in the final version.
1. Rule Extension
- It is explained in the main paper Sec. 3.2.3 and Alg. 1 (L6-10). The ini... | Summary: This paper studies the use of a symbolic system for understanding human activity. The author argues that a good symbolic system for human activity understanding should possess a diverse range of symbols that accurately identify the human activity, and the rules within the symbolic system should be logically co... | Rebuttal 1:
Rebuttal: We thank the reviewer for the feedback and hope to address the questions and concerns below.
### Method Details Explanation
1. Symbol Initialization
- The prompt is manually designed by us. Special constraints are considered in the prompt to output expected initial prompts. For example, we con... | Summary: In this paper, a symbolic-System-2 submodule is integrated with a System-1 module to perform human activity recognition. The work follows the intuitive idea, where a visual prediction module extracts visual symbols and a symbolic reasoning module performs logical reasoning to recognize activities. For the visu... | Rebuttal 1:
Rebuttal: We thank the reviewer for the feedback and hope to address the questions and concerns below.
## Generality
The method could be more general. We choose activity understanding as a good and important initial test bed because it is a difficult task with complex visual patterns and compositional natur... | Rebuttal 1:
Rebuttal: We sincerely thank all reviewers for their constructive comments and recognition of our work’s strengths:
- It is interesting (Reviewer K3QA), encouraging, and worth appreciation (Reviewer zui3) to leverage and extend the intuitive idea of the System-1 and System-2 combination. The contributions ... | NeurIPS_2023_submissions_huggingface | 2,023 | Summary: The paper addresses the problem of visual activity understanding and proposes a way to leverage large visual and language models to i) propose symbols from an image, ii) derive new symbols via logic entailment and perform classification. The results show that the augmented system with symbolic reasoning capabi... | Rebuttal 1:
Rebuttal: We thank the reviewer for the feedback. We will add more discussions about limitations brought by using pre-trained large models. We hope to address the questions and concerns below.
## Detailed Evaluation\&Analysis
We respectfully explain that using LLM is the optimal method within our consider... | Summary: The paper proposes to increase coverage and generalization of logic-based methods to model 'system-2' or deliberative reasoning for the task of human activity prediction in images. Specifically, the authors incorporate an LLM (GPT3.5) to obtain a larger number of symbols and inference rules related to human ac... | Rebuttal 1:
Rebuttal: We thank the reviewer for the feedback and hope to address the questions and concerns below.
## Computation Scalability
Symbol predictions will increase the computational cost as a trade-off for explainability and generalization. However, with the proper method design shown below, the computation... | Summary: The paper proposes a novel method for visual activity understanding using an enriched symbolic system. The authors focus on enhancing the integration of intuitive (System-1) and logical (System-2) reasoning processes in AI systems. They argue that the current System-2 reasoning methods suffer from limitations ... | Rebuttal 1:
Rebuttal: We thank the reviewer for the feedback and hope to address the questions and concerns below.
## Contribution
Though it is an active research area to incorporate commonsense knowledge into AI reasoning, this paper has its unique contribution toward integrating a rule-based symbolic logic reasoner ... | null | null |
Cheaply Estimating Inference Efficiency Metrics for Autoregressive Transformer Models | Accept (poster) | Summary: The authours proposed a metric called `idealized runtime` to evaluate the inference efficiency among different LLMs as if they were run on the same standard hardware/software system. The main contributions can be summarized as following:
* the `idealized runtime` for a specific query with prompt length `p` an... | Rebuttal 1:
Rebuttal: Thank you for your review! We respond to your main comments and questions below.
---
> The inference latency of a well known (open sourced) LLM architecture can be easily estimated by total ops of the model and divided by hardware effective (typically around 50% for LLMs on A100s) FLOP/s. The au... | Summary: The authors proposed idealized runtime, a metric for measuring inference efficiency in LLMs, which measures the performance of models as if they were executed on a given hardware and software platform. Idealized runtime efficiently can be extended to estimate the idealized energy and dollar cost. The metric al... | Rebuttal 1:
Rebuttal: Thank you for your review! We respond to your comment on weaknesses below.
---
> The models on different tasks and metrics exhibit vastly different Pareto Frontiers (Figure 9). The cause of this variance hasn't been fully explored and explained in the work. Can these datasets/tasks properly eval... | Summary: This paper aims to better understand the tradeoff between the inference efficiency and capability of LLM. It comes up with a metric called “idealized runtime” to fairly evaluate the inference efficiency of several popular LLMs, which is designed to factor out the irrelevant noises such as hardware, code stack,... | Rebuttal 1:
Rebuttal: Thank you for your review! We respond to your main comments and questions below.
---
> The concept of “idealized runtime” could be a bit unpractical because it requires one to implement and deploy the LLM on a very specific hardware and software setting, which may involve quite amount of proper ... | Summary: "This paper introduces a range of metrics designed to compare the tradeoffs between inference efficiency and capability of autoaggressive Language Model (LLM) based on Transformers. Alongside the raw run-time metric, which can be influenced by infrastructure optimization and performance variance, the paper pro... | Rebuttal 1:
Rebuttal: Thank you for your review! We respond to your main comments and questions below.
---
> There are a few minor formatting issues that require attention. For instance, in line 153, the notation of o^2 is confusing. The "2" is intended to refer to the footer, but it appears as if it denotes an expon... | Rebuttal 1:
Rebuttal: We thank all the reviewers for their time and effort in providing feedback on our paper, which will be useful in improving its quality.
All reviewers seem to agree that the problem addressed by the paper is important and well-motivated. However, the two lowest-score reviews seem to have serious m... | NeurIPS_2023_submissions_huggingface | 2,023 | Summary: LLMs are dominating the NLP space at the moment, but efficient inference run-time and cost estimations have not been explicitly defined and tested. Raw run-times are not particularly useful across multiple services providing LLMs, as there is typically high variance overhead via the presence of black box promp... | Rebuttal 1:
Rebuttal: Thank you for your review! We respond to your main comments and questions below.
---
> Some areas of the paper are a bit overexplained for this venue. Readers and reviewers of NeurIPS should, for example, be very familiar with autoregressive frameworks and how they generally work, or the forward... | Summary: This paper presents a new metric that attempts to more fairly evaluate machine learning model inference runtime performance cost, which the authors refer to as idealized runtime. The authors argue that the black-box nature of many large language models (LLMs) does not necessarily provide accurate estimates of ... | Rebuttal 1:
Rebuttal: Thank you for your review! We respond to your main comments and questions below.
---
> Floating point operation overhead (a critical paper weakness): in Section 3.1.1 the authors discuss some of the concrete details used for their idealized runtime equation. Of note is the focus on deriving expr... | null | null | null | null |
Relax, it doesn’t matter how you get there: A new self-supervised approach for multi-timescale behavior analysis | Accept (spotlight) | Summary: The authors propose BAMS, an architecture for learning agent behaviors by mapping observations to actions. BAMS utilizes two Temporal Convolutional Networks (TCNs) operating at different time scales, enabling modeling of both short-term and long-term trends in data through varying dilation rates. The short-ter... | Rebuttal 1:
Rebuttal: Thank you for your comments and questions! We provide point-by-point responses below.
> 1. “The design choices in the proposed method appear to be based on arbitrary experimentation rather than being guided by well-established principles. This absence of a solid theoretical foundation raises conc... | Summary: The authors propose a model called BAMS for unsupervised behavioral analysis. It combines two temporal convolutional neural networks with different timescales used to learn latents, as well as a loss to predicting a distribution of future actions. BAMS reaches strong performance on a novel, small-scale synthet... | Rebuttal 1:
Rebuttal: Thank you for your comments and questions! We provide point-by-point responses below.
> 1. “While the overall pipeline is novel, each component is rather well known and thus the overall novelty is somewhat limited.”
**Reply:** The method consists of two core components: (1) a HoA predictive los... | Summary: The paper introduces a self-supervised (autoregressive) learning system that processes behavioral data on two fixed time scales via TCNs. Combined, the TCN output it then used to predict behavioral histograms over a fixed time window. The innovative histogram encoding encourages the learning of behavioral dist... | Rebuttal 1:
Rebuttal: Thank you for your comments and questions! We provide point-by-point responses below.
> 1. System does not seem to be generally applicable. It focuses on analyzing repetitive behavioral patterns and general, global, behavioral-influencing agent properties. It seems to me that the tasks indeed are... | Summary: The authors propose a multi-task representation learning model for animal behavior consisting of an action-prediction objective and a multi-scale architecture with separate short and long term dynamics. The authors use this model to achieve state of the art performance on the MABe2022 Challenge. Furthermore, t... | Rebuttal 1:
Rebuttal: Thank you for your comments and questions! We provide point-by-point responses below.
> 1. While the connection between mouse and robot behavior is interesting, the connection should be made a bit more strongly. What would these behavior models do to aid robotics? This question should be explicit... | Rebuttal 1:
Rebuttal: We would like to thank all of the reviewers for their thoughtful feedback and helpful suggestions! We are delighted to hear that all of the reviewers appreciated the work and motivation, and were impressed with the fact that BAMS achieves state-of-the-art performance on a complex realworld multita... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Deep Graph Neural Networks via Flexible Subgraph Aggregation | Reject | Summary: This manuscript understands the expressive power of GNNs from a new perspective of subgraph aggregation, and reveals the potential reason for the performance degradation of traditional deep GNNs due to the overlap of aggregated subgraphs. The authors propose a sampling-based generalized residual module SNR and... | Rebuttal 1:
Rebuttal: Thank you very much for your valuable feedback. Below are our responses to your questions/concerns.
> More related works on dealing with oversmoothing and overfitting issues in deep graph neural networks should be reviewed.
>
Thanks for the useful comment. We will discuss more related work on o... | Summary: The paper propose a sampling-based module to enhance the expressive power of Graph Neural Networks (GNNs), which traditionally assume that information from the subgraph of the same hop is equally important for all nodes in the graph. They argue that this rigid assumption restricts the models' ability to captur... | Rebuttal 1:
Rebuttal: We thank the reviewer very much for the valuable feedback. Below are our responses to your questions/concerns.
>**W 1:**
Thank you for the comment. We have used the official DGL implementation of GCNII with the same parameter settings as in the original paper, and we have limited the number of ... | Summary: This paper studies the problem of over-smooth with GNNs. It is argued that the over-smooth problem is caused by the increased overlap of the sub-graph when the respective field of GNNs becomes larger and larger. In order to alleviate this problem, this paper proposed a method that random weighting the nodes w... | Rebuttal 1:
Rebuttal: Thank you very much for your valuable feedback. Below are our responses to your questions/concerns.
>**W1:**
Firstly, our work focuses on the over-smoothing problem of classical MPNNs, which are localized graph models based on the massage-passing mechanism, and are much more scalable than global... | Summary: This paper focuses on the alleviation of overfitting and over-smoothing of deeper GNNs. It is an interesting topic and the solution seems promising with a sample-based node level residual block. Extensive experiments on public datasets verifies the applicability.
Strengths: The paper analyses in a new perspe... | Rebuttal 1:
Rebuttal: Thank you very much for your valuable feedback. Below are our responses to your questions/concerns.
> The presentation is in great need to improve. There are some long sentences that hinder the understanding of the authors’s ideas, such as L54. A research article should describe facts with a 3rd ... | Rebuttal 1:
Rebuttal: We thank the reviewers for their extensive reviews, helpful comments and actionable suggestions for improving our manuscript. We have made a PDF that includes tables and figures of our extended experimental results. Please refer to our detailed rebuttal for interpreting the results.
Pdf: /pdf/5cf3... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Learning Rule-Induced Subgraph Representations for Inductive Relation Prediction | Accept (poster) | Summary: This paper proposes REST, a method for inductive link prediction over knowledge graphs. Compared to previous techniques that extract subgraphs around the target edge, REST avoids node labelling (thus saving substantial execution time) and only initializes the representation of the target link, setting other ed... | Rebuttal 1:
Rebuttal: We thank the reviewer for the positive and insightful comments. We respond to each comment as follows and sincerely hope that our rebuttal could properly address your concerns. If so, we would deeply appreciate it if you could raise your score. If not, please let us know your further concerns, and... | Summary: This paper proposes a subgraph-based method to address the inductive relation prediction task, where the model learns to predict links by learning implicit rules from subgraphs. Models proposed in previous works lack the ability to differentiable the relevant and irrelevant paths. To address this, the authors ... | Rebuttal 1:
Rebuttal: We thank the reviewer for the positive and insightful comments. We respond to each comment as follows and sincerely hope that our rebuttal could properly address your concerns. If so, we would deeply appreciate it if you could raise your score ("4: Borderline reject"). If not, please let us know y... | Summary: This paper tackles a problem in the sub-area of machine learning on graphs. It looks at approaches for inductive reasoning (i.e. ones the generalize to unseen vertices), identifies a problem with those approaches and suggests a novel manner to tackle is problem, exploiting the recurrant nature of message-pass... | Rebuttal 1:
Rebuttal: We thank the reviewer for the positive and insightful comments. We respond to each comment as follows and sincerely hope that our rebuttal could properly address your concerns. If so, we would deeply appreciate it if you could raise your score. If not, please let us know your further concerns, and... | Summary: This work proposes a single-source edge-wise GNN model for the inductive relation prediction task, called REST. The method includes a single-source initialization approach that reduce the irrelevant rules during learning and an RNN-based edge-wise message passing mechanism to model the sequential property of m... | Rebuttal 1:
Rebuttal: We thank the reviewer for the positive and insightful comments. We respond to each comment as follows and sincerely hope that our rebuttal could properly address your concerns. If so, we would deeply appreciate it if you could raise your score. If not, please let us know your further concerns, and... | Rebuttal 1:
Rebuttal: We thank all the reviewers for the insightful comments and constructive suggestions, which are very helpful for us to strengthen this submission.
We have submitted detailed responses to address the concerns raised by all the reviewers. Besides answering the technical questions,
1. we conduct more... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
On-the-Fly Adapting Code Summarization on Trainable Cost-Effective Language Models | Accept (poster) | Summary: This paper proposes an AdaCom approach to improve the code comment generation performance by on-the-fly model adaptation. For one code sample at test time, AdaCom first detects whether the model might have a compromised summarization performance and if yes, re-adapt the model on-the-fly by training the most c... | Rebuttal 1:
Rebuttal: We thank the reviewer’s appreciation of our work. We observe that your fundamental concern lies in the runtime overhead of our approach. In this rebuttal, we (1) clarify that improving smaller models is far more valuable and (2) show the inference trick for AdaCom to work on larger models.
We sin... | Summary: Language models trained on code tend to suffer in performance when applied to unseen code bases. This work investigates adapting smaller LMs to a new project context based on a brief fine-tuning phase that is invoked immediately before the model predicts a comment for a given method. Fine-tuning examples are s... | Rebuttal 1:
Rebuttal: We thank the reviewer’s comments for improving our work. We understand the importance of providing accurate and helpful statement and we will do our best to ensure that our responses are clear. It seems that your concerns on the soundness derive from the seemingly confused interpretation. We will ... | Summary: AdaCom is a system meant to adapt a trained model, on-the-fly, when making a prediction (specifically targeting code summarization models). It does this by identifying training examples that seem similar to the target example, lightly re-training the model to adapt better to those similar examples, and then ma... | Rebuttal 1:
Rebuttal: We thank the reviewer for the appreciation and support for our work. We address your concerns as follows:
# Q1: Where does the training contribution score come from? Is it the mutual influence?
The training contribution score is defined between a training sample *s_i* and a testing sample *s_j*,... | Summary: This work proposes AdaCom for boosting code comment generation by on-the-fly model adaptation. The authors started with the motivation that existing generators are heavily affected by conflicts within the dataset (CodeSearchNet). The proposed AdaCom method is divided into two stages: clustering the training sa... | Rebuttal 1:
Rebuttal: We thank the reviewer for appreciating our work. In response to your concern about the limited improvement on larger models, we would like to highlight the following points:
### Relative Improvement:
Larger models already exhibit relatively better performance compared to smaller models. Conseque... | Rebuttal 1:
Rebuttal: We thank the reviewers' comments for improving our work. We observe that the reviewers' concerns lie in (1) required new baselines, (2) confused technical differences with retrieval-based baselines, and (3) general presentation problems. In this rebuttal, we have fixed all the concerns and request... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Public Opinion Field Effect Fusion in Representation Learning for Trending Topics Diffusion | Accept (poster) | Summary: This paper proposes to model the public opinion field effect in representation learning of social networks. The authors propose to model social networks as heterogeneous graphs and incorporate three effects into a neural architecture. Experiments on three datasets demonstrate that the proposed approach outperf... | Rebuttal 1:
Rebuttal: We sincerely thank the reviewer for the positive comments on the novelty and importance of our research. We also appreciate the reviewer's suggestion on ethics statement, $k$-hop neighborhood, and more baseline models. We respond carefully to the reviewer’s questions and hope our answers could add... | Summary: The seminal models of opinion dynamics as information cascades on networks, such as DJ Watts 2002 or Watts and Dodds 2007, uses homogeneous graphs in which the nodes (representing people) are characterized by just one or two numbers (such as a bias number and a Boolean on adoption). This line of research provi... | Rebuttal 1:
Rebuttal: We sincerely thank the reviewer for positive comments on the novelty and high performance of our work. We respond carefully to the questions and hope our answers could address your concerns.
**[W1&L1]: Dynamic of trending topic diffusion**
Thanks a lot for your insightful suggestion for consi... | Summary: In this paper, the public opinion field effect is fused into a representation learning framework for trending topics diffusion. The proposed algorithm includes representation learning for both topic propagation and social circle influence. Experimental results on public opinion concern prediction and event cla... | Rebuttal 1:
Rebuttal: We sincerely thank the reviewer for the recognition of the novelty and importance of our work. We also appreciate the positive comment on the effectiveness and efficiency of our model. We respond carefully to the reviewer’s questions and hope our answers could address your concerns.
**[W1]: Signi... | Summary: In the manuscript, the authors consider two interesting novel factors for trending topics diffusion, which are public opinion field effect and social circle influence effect. They propose a representation learning framework to incorporate the two effects and conduct extensive experiments to validate the effect... | Rebuttal 1:
Rebuttal: We sincerely thank the reviewer for positive comments on novelty, good presentation and reasonable technique of our work. We also appreciate the reviewer's suggestion about giving more discussion on public opinion field effect. We respond carefully to the reviewer’s questions and hope our answers... | Rebuttal 1:
Rebuttal: We would like to thank all the reviewers for their insightful and invaluable comments, which really help to improve the quality of our paper. We also appreciate five reviewers' positive comments about the novelty and well-organization of our paper. Here, we response to four common questions and h... | NeurIPS_2023_submissions_huggingface | 2,023 | Summary: This paper proposes a Representation Learning method that considers field information in topic diffusion, i.e., situations with multiple topics and opinion leaders. It defines new concepts (public opinion network/field and public opinion field energy) and proposes a method for incorporating some observations i... | Rebuttal 1:
Rebuttal: We sincerely thank the reviewer for the positive comments on the strength of experiments and good presentation. We also appreciate the suggestion about more discussion on public opinion field effect and the fairness of the experiment. We respond carefully to the reviewer’s questions and hope our ... | null | null | null | null | null | null |
Neural Combinatorial Optimization with Heavy Decoder: Toward Large Scale Generalization | Accept (poster) | Summary: This paper proposes a neural network model in the form of a Light Encoder Heavy Decoder, as well as its training method for solving combinatorial optimization problems. In addition, this paper suggests an RPC method based on updating partial solutions as a way to improve solutions during the inference process.... | Rebuttal 1:
Rebuttal: Thank you very much for your time and effort in reviewing our work. We are glad to know you find our proposed heavy decoder model to learning the state of various node sizes is promising, have promising experimental results on large-scale TSP and CVRP instances, and the paper is well-written.
W... | Summary: The paper proposes a learning-based approach called the Light Encoder and Heavy Decoder (LEHD) to solve combinatorial optimization (CO) problems. The proposed method learns less from the static graphs (via a light encoder), and pays more attention to dynamically capturing the relationships between all availabl... | Rebuttal 1:
Rebuttal: Thank you very much for your time and effort in reviewing our work. We are glad to know you find our proposed method design is insightful and logical, the empirical results are promising especially for the generalization ability, and the paper is well-structured and easy to follow.
We address you... | Summary: The paper focuses on a class of NP-hard combinatorial optimization problems called Vehicle Routing Problems (VRP) which is of wide practical interest. This paper proposes a novel Light Encoder and Heavy Decoder (LEHD) model for enabling neural construction methods well handling large-scale routing problems. It... | Rebuttal 1:
Rebuttal: Thank you very much for your time and effort in reviewing our work. We are glad to know you find our work has a clear motivation and is easy to read, the proposed LEHD is novel and can generalize well to problems with various sizes, and has sufficient and promising experiment results especially on... | Summary: In summary, this work proposed a new solution that generalizes well on large-scale problem instances for routing problems. A data-efficient training scheme and solution construction mechanism are further proposed for efficient training. Experiments verify the effectiveness of the proposed method, and an impres... | Rebuttal 1:
Rebuttal: Thank you very much for your time and effort in reviewing our work. We are glad to know you enjoy reading our paper, find our method simple yet effective to tackle large-scale routing problems, and believe this work would be very insightful for the community.
We address your concerns as follows.... | Rebuttal 1:
Rebuttal: **General Response**
Dear AC and Reviewers,
We sincerely thank you for your time and effort in reviewing our work. We are glad to know the reviewers enjoy reading our paper or find it easy to follow (**Hd9p**, **6SQz**, **8Nbk**, **6Pbv**), our method is effective (**Hd9p**) /novel (**6SQz**) /... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Beyond Black-Box Advice: Learning-Augmented Algorithms for MDPs with Q-Value Predictions | Accept (poster) | Summary: This work considers the setting of using advice in Markov Decision Processes (MDPs). They consider the setting of using only action suggestions (termed black-box) or action and Q-values (grey-box) from a potentially erroneous advisor. They propose an algorithm that decides between the advice and a robust polic... | Rebuttal 1:
Rebuttal: Thank you so much for your insightful comments! Below are our responses to your specific comments and questions:
1. **Non-stationary Environment**:
Than you so much for providing more insights and potential application domains of our methods. In fact,
existing algorithms for non-stationary RL pr... | Summary: The authors study the tradeoff between consistency and robustness in the context of a single-trajectory time-varying Markov Decision Process (MDP) with untrusted machine-learned advice. Time-varying MDPs are relevant when dealing with environments where costs, rewards, and transition probabilities might change... | Rebuttal 1:
Rebuttal:
We sincerely appreciate your thoughtful feedback and valuable insights regarding our paper. Below we respond to your questions and concerns:
1. **Comparison to Existing Approaches**:
>*Insufficient comparison to existing approaches: The paper does not provide a detailed comparison of the propo... | Summary: The paper proposes a novel algorithm that combines a robust policy and a potentially untrusted machine learning based policy. The idea is to be able to achieve a trade-off between robustness and consistency or high performance entailed by the ML policy. The paper suggests a novel projection algorithm along wit... | Rebuttal 1:
Rebuttal: Thank you for your valuable feedback and insightful comments. Below are our responses to your specific comments and questions:
1. **Motivation**:
>*The aim is to derive a new policy that combines the two and be able to achieve consistency vs. robustness trade-off in real-time decision making, d... | Summary: This paper investigates ways to combine untrusted learned policy and Q value predictions with a trusted baseline fixed policy such that the resulting combination achieves a higher performance bound and lower regret than established methods for combining the two. In particular, this paper studies how the use of... | Rebuttal 1:
Rebuttal: Thank you so much for the thorough review and the valuable insights provided. Your positive evaluation from a experimentally-driven deep RL perspective of the core argument regarding the derived tradeoff between consistency and robustness using TD error information is very encouraging.
Regarding ... | null | NeurIPS_2023_submissions_huggingface | 2,023 | Summary: This paper provides a theoretical study of the tradeoff between consistency and robustness in Markov Decision Process for which machine learned advice is available. With extensive theoretical analyses, the paper leverages Q-value advice for improving the impact of machine-learned advice in these contexts. Seve... | Rebuttal 1:
Rebuttal: Thank you for your thoughtful comments! Below are our responses to your specific comments and questions:
1. **Empirical Validation**:
While our primary focus and main contribution center around the theoretical aspects, we fully agree with your suggestion that empirical validation can significan... | Summary: Authors study algorithms for Marcov Decision Processes receiving advice in the
form of a value-based machine-learned policy. Value-based policies decide
their actions based on minimizing an estimated Q-function, which predicts
the best possible outcome in the rest of the time horizon given a current
action. Su... | Rebuttal 1:
Rebuttal:
Thank you so much for the valuable feedback on our work! Below are our responses to the comments:
1. **Consistency-Robustness Tradeoff for Black-Box Setting**:
>*It is not clear whether consistency-robustness tradeoff with weaker predictions is necessary for any algorithm. They prove a lower b... | null | null | null | null |
Segment Anything in High Quality | Accept (poster) | Summary: The authors observe that the segmentation results of SAM model are unsatisfactory in several cases such as coarse mask boundaries and increect predictions. Thus, this paper presents HQ-SAM, whose objective is to improve the segmentation capability of SAM, with the introduction of the learnable high-quality out... | Rebuttal 1:
Rebuttal: We thank the reviewer for acknowledging our paper being the first step to improve SAM, with good motivation and presentation. Here we respond to your insightful questions.
> **Q1.** What’s the error correction module?
Error correction is not a module, but simply a direct element-wise sum between... | Summary: This paper proposes a high quality version of segments anything. It points out the issues of existing segment anything models in Figure 1 and proposes the corresponding solution. It proposes a high-quality output token and a global-local fusion module. The results are good at various open benchmarks. The visua... | Rebuttal 1:
Rebuttal: We thank the reviewer for acknowledging our paper with great/meaningful contribution, good writing and impressive results. Here we respond to your insightful comments and questions.
> **Q1.** Visualization of HQ-Output Token.
We provide visual comparison of our HQ-Output Token **vs.** SAM's comm... | Summary: This paper proposes HQ-SAM that produces segmentation masks high quality while maintaining SAM's original promptable design, efficiency, and zero-shot generalizability, by only introducing minimal additional parameters. A learnable High-Quality Output Token is injected into SAM's mask decoder. Instead of only ... | Rebuttal 1:
Rebuttal: We thank the reviewer for acknowledging that HQ-SAM is important, interesting and effective with extensive experiments evaluation. Here we respond to your insightful comments and questions.
> **Q1.** What would happen if SAM is also trained on HQSeg-44K dataset?
We provided the experiment of tra... | Summary: This paper proposes HQ-SAM, an improved SAM with higher-quality segmentation performance. HQ-SAM freezes the entire SAM model and appends additional learnable modules concurrently to SAM. With the HQ-output token and global-local feature fusion, HQ-SAM achieves better segmentation results and only costs margin... | Rebuttal 1:
Rebuttal: Thank you for acknowledging our work with an interesting and reasonable design, good performance and marginal training cost. Here we respond to your insightful comments and questions.
> **Q1.** Experiments on unfreezing the entire SAM, or using PEFT methods (LoRA, adapters) with performance-param... | Rebuttal 1:
Rebuttal: We appreciate the positive and insightful comments from all four reviewers. We are glad that they found our paper "well motivated and interesting” (E7oL, 2HQS, eEa5), “important and a great contribution” (2HQS, xJfb), “effective with extensive experiments” (E7oL, 2HQS, xJfb), “with marginal trai... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
PROTES: Probabilistic Optimization with Tensor Sampling | Accept (poster) | Summary: The authors present the algorithm PROTES for solving discrete optimization problems. The method uses probabilistic sampling to obtain a probability function of minimal (or maximal) values of a target function. This probability function takes the same shape as the target tensor. Alternatively, this could be des... | Rebuttal 1:
Rebuttal: Thank you for the detailed analysis of our work and extensive comments and remarks! We tried to take them into account and make appropriate improvements. Below we present (within the specified limit on the length of the sent message) our responses to the main points of your review.
1. The method ... | Summary: This paper presents a study on black-box optimization. The authors propose a new method based on sampling from the probability density defined in the tensor train (TT) format (extension effort). The experiments have shown that PROTES outperforms existing efforts.
Strengths: (Overall) This paper is well organi... | Rebuttal 1:
Rebuttal: Thank you for your comments, which allowed us to improve the presentation of the proposed approach! Below we present our responses to your questions on the theoretical part of the work and comments on additional numerical experiments performed.
1. The method section "Optimization method PROTES" i... | Summary: The paper proposes PROTES, a new method for multi-dimensinoal discrete black-box optimization based on adaptive probablistic sampling in the tensor train representation.
Strengths: - The technical method is simple and intuitive, with a clear presentation.
- Diverse benchmark problems and comprehensive compar... | Rebuttal 1:
Rebuttal: Thank you very much for your positive evaluation of our work and your comments! Below we present our responses; we hope that they will eliminate the shortcomings you have noted.
1. ... lack of intuition/discussion on why the proposed approaches outperform baselines ...
- A potential limitati... | Summary: This paper proposes a novel algorithm to estimate global extrema of discrete black-box target functions. The approach proposes learning a probability distribution, represented as a multidimensional tensor and efficiently stored using a TT-decomposition, that is trained to favor multi-indices that achieve high/... | Rebuttal 1:
Rebuttal: Thank you very much for the high appreciation of our work, very detailed comments and useful suggestions to improve its quality! Below we present our responses to the main points of your review.
1. "Gradient-free" is defined very late in the paper...
- You are absolutely right that this poin... | Rebuttal 1:
Rebuttal: Thank you very much for your comments, which allowed us to correct the shortcomings and refine the presentation of the proposed approach. We have improved the text of the manuscript, however, according to the rules of the conference, we are now unable to attach the updated version.
1. We practica... | NeurIPS_2023_submissions_huggingface | 2,023 | Summary: The paper targets the challenging gradient-free discrete multidimensional optimization. It is formalized as finding a vector of d indices whose associated entry in a d-dimensional tensor is minimized, where the tensor contains all the possible values for the black-box objective function. There are an exponenti... | Rebuttal 1:
Rebuttal: Thank you very much for your detailed comments and questions, which allowed us to refine the presentation of the proposed approach. Below we present our respective responses.
1. ... when and why the combinatorial optimization problem's structure is consistent with the low-rank probability assumpt... | null | null | null | null | null | null |
Projection-Free Methods for Stochastic Simple Bilevel Optimization with Convex Lower-level Problem | Accept (poster) | Summary: A nice and clean result for stochastic simple bilevel optimization. The manuscript extends the algorithm design from the deterministic simple bilevel optimization with convex lower-level problem to the stochastic setting and demonstrate the sample complexity, which seems to be near-optimal (ignoring the log fa... | Rebuttal 1:
Rebuttal: We thank reviewer oeu1 for the comments and feedback!
**Q1 Why do we use STORM and SPIDER rather than a simpler moving average estimator?**
**A1** This is an excellent point. Since our framework utilizes a projection-free approach for stochastic bilevel optimization, we have endeavored to incor... | Summary: The paper presents a novel approach for solving stochastic simple bilevel optimization problems. The authors develop new methods that locally approximate the solution set of the lower-level problem using a stochastic cutting plane and then employ a conditional gradient update with variance reduction techniques... | Rebuttal 1:
Rebuttal: We thank reviewer Xgt8 for the comments and feedback!
**Q1 The authors assume that the constraint set $\mathcal{Z}$ is compact and convex, which may not hold for many real applications.**
**A1** While this assumption may not apply to certain real-world scenarios, it remains a common assumption ... | Summary: This paper introduces novel projection-free optimization methods for solving a class of *simple* stochastic bilevel optimization problems. The proposed methods are based on the conditional gradient algorithm and use variance reduction techniques to approximate the solution set of the lower-level problem. The p... | Rebuttal 1:
Rebuttal: We thank reviewer iBKW for the comments and feedback!
**Q1 Can you describe how we can see this optimization problem from a game theory perspective?**
**A1** As the reviewer has correctly pointed out, a Stackelberg game (SCG) can be viewed as a general bilevel program in which a leader aims to m... | Summary: This paper proposes methods for solving nonconvex/convex-convex simple bilevel optimization problems. The proposed methods are proven to have better non-asymptotic convergence guarantee than existing works. Numerical experiments are done on two applications to verify the effectiveness of the proposed methods.
... | Rebuttal 1:
Rebuttal: We thank reviewer FAuV for the comments and feedback!
**Q1 There are several works on general bilevel optimization problems without lower-level strong convexity. The authors should provide more detailed discussions on how these existing works are/aren't applicable to the target formulation of thi... | Rebuttal 1:
Rebuttal: In response to Reviewer Xgt8, we included some additional plots in the attached PDF file comparing the studied methods in terms of runtime.
Pdf: /pdf/0382cfcd7138c78b47fcc76c8aab3e5abc569f70.pdf | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Input margins can predict generalization too | Reject | Summary: The paper presents a modification to existing optimisation-based margin computation technique in neural networks to be used as an empirical complexity measure. The measure is evaluated on the Predicting Generalization in Deep Learning (PGDL) 2020 competition dataset (now all public) and shows better performanc... | Rebuttal 1:
Rebuttal: We thank you for your review and comments. We shall first address your question, before providing a general rebuttal of the weaknesses you have pointed out.
>"The assurance that no tuning was done on the (now public) PGDL test set is noble... but in the end, the test data is now public, and while... | Summary: The authors propose a novel notion of constrained input margin to measure the generalization performance of neural networks. Compared to conventional input margin defined on $\mathbb{R}^n$, the proposed constrained input margin is constrained on the data manifold and estimated via g Principal Component Analysi... | Rebuttal 1:
Rebuttal: We thank you for your review and suggestions. We pose the following response to the two questions raised:
>"There is lack of a theoretical analysis on the contrained margins, as a theoretical analysis will make the metric more convincing to connect to model generalization."
Please see the global... | Summary: This paper deals with the relationship between input margins and generalization. Considering the decision boundary may be not inherently useful in classification, the authors develop a constrained margin based on directions extracted from input data. Principal Component Analysis (PCA) is used to identify the s... | Rebuttal 1:
Rebuttal:
We thank you for your constructive commentary and concerns. We have the following response to the four points raised:
>"The approximation in Equation (3) is rough. It may be useful on small datasets and easy tasks, but can be massively computational for PCA and not indicative on more complicated... | Summary: The paper empirically studies input margins as a predictive measure for the generalization of neural networks.
Strengths: The paper has a very interesting finding: the distance of the decision boundary of a neural network to the training samples is predictive of its generalization performance when such a dist... | Rebuttal 1:
Rebuttal: We thank you for your detailed review and insight. The case of a single separating hyperplane is an interesting thought experiment that requires careful consideration, and we are also grateful for the other minor corrections. We pose the following responses to each of the questions:
>"Concerning ... | Rebuttal 1:
Rebuttal: We would like to thank all reviewers for their commentary; we believe this has been a constructive exercise in improving the manuscript. We have addressed each reviewer’s comments individually. In addition, we would like to make two points:
1. There is a strong line of research that investigates... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
The Utility of “Even if” Semifactual Explanation to Optimise Positive Outcomes | Accept (poster) | Summary: This paper proposes a framework for "semi-factual" explanations - explanations given to an applicant who receives an accept decision, about what they could do which would allow them to maintain that decision. The framework and objective propose utilize a causal model of the action space, and consider the appli... | Rebuttal 1:
Rebuttal: We thank the reviewer for noting the relevance of our user study and how it sets the paper apart. We are truly grateful for your suggestions which have improved the paper, we now respond to your questions & concerns. Apologies for the brevity of some points, but space was limited.
***
### Review... | Summary: This paper looks into the mirror problem of counterfactual recourse. In the latter the goal is to determine what is needed to change a negative outcome to a positive such as in loans. In the former the goal is to determine to what extend a positive decision would remain and to what extend the end user could ha... | Rebuttal 1:
Rebuttal: We sincerely thank the reviewer for noting the novelty and simplicity (in a good way) of our idea and its relevance to the field. We now address all questions and concerns:
***
*The paper seems to take too much inspiration from the causal algorithm recourse paper to a point where it becomes just... | Summary: The paper formalizes the notion of semi-factual for positive outcomes. Given a black-box classifier, a semi-factual is a statement which informs a user about “even if” scenarios in which they would maintain the assigned positive prediction, increasing their “gain”. The authors formalize the problem for both a ... | Rebuttal 1:
Rebuttal: We greatly thank the reviewer for the extremely positive comments, noting the novelty of our work, and voting for acceptance. We now address all questions and concerns:
***
*The notation and the various proofs can be hard to read at first glance. For example, in Section 3.1., between lines 95 an... | Summary: This paper explores techniques for optimizing positive outcomes using semi-factuals. To do so, they propose the concept of "gain" and set up a small user test to demonstrate the usefulness of their technique.
Strengths: The authors offer an interesting and compelling framework for optimizing positive outcomes... | Rebuttal 1:
Rebuttal: We thank the reviewer for noting the compelling nature of our framework, and on account of voting for the acceptance of the work. Here, we address your specific questions and concerns.
***
*I think... the authors need to better motivate their work, especially in the Introduction... the language.... | Rebuttal 1:
Rebuttal: We would like to thank the reviewers for taking the time to read our paper and give constructive feedback. We have taken seriously the suggestions of all reviewers and either amended the manuscript accordingly to address the concerns, or discussed it here in our response.
Firstly however, we than... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Dynamo-Depth: Fixing Unsupervised Depth Estimation for Dynamical Scenes | Accept (poster) | Summary: This paper introduces a new unsupervised approach and a new network architecture for monocular depth estimation.
The paper is motivated by the necessity of new algorithms capable of learning to predict depth from unlabelled data while being able to perform reliable predictions in the regions of moving objects... | Rebuttal 1:
Rebuttal: Thank you for your support and constructive comments! We are encouraged that you find the paper well-written, the insights interesting, and the results convincing! We list our response to your questions below.
---
### Contexts Beyond Self-Driving
Our approach would work well in other general sett... | Summary: This paper aims to estimate the depth map in a dynamic scene environment.
To this end, the paper proposed 1) two separate module architectures that estimate rigid scene flow and residual scene flow and 2) a motion initialization method to enable a stable learning process.
However, the technical and performance... | Rebuttal 1:
Rebuttal: Thank you for taking the time to point out the references! We outline the differences between our approach and each of the references below. We have added many of these references to Table 8 (see below for details) and we also discuss how we propose to add additional evaluations to our final came... | Summary: This paper proposes a self-supervised depth estimation method. It addresses dynamic objects in the scene by jointly learning monocular depth, 3D independent flow field, and motion segmentation from unlabeled videos. The paper writing is very good, and the motivation/insights are clearly stated. The method is e... | Rebuttal 1:
Rebuttal: Thank you for your support and for finding our paper well-written and the proposed method effective. We are encouraged by your remarks on our evaluations. We share your concern regarding the overfitting and fewer-moving-object issue that makes KITTI a less effective benchmark compared to Waymo and... | Summary: This paper presents a method for monocular depth estimation in dynamic scenes, employing an unsupervised learning approach. While prior work utilizes networks for depth estimation, pose estimation, flow estimation, and mask estimation to decompose object motion and agent motion, the proposed method in this pap... | Rebuttal 1:
Rebuttal: Thank you for your thoughtful review and for finding the proposed method effective and our evaluation convincing. We list our response to your questions and concerns below.
---
### Fail to Surpass Existing SOTA on KITTI
Thank you for pointing this out. Although our approach performs only comparab... | Rebuttal 1:
Rebuttal: We thank the reviewers for their insightful comments and positive feedback. We are encouraged that they found our work well-written (*Reviewer EvaZ, 9pRa*), well-motivated (*Reviewer SBjP*), effective (*Reviewer 9riS, EvaZ, 9pRa*), and with proper evaluations (*Reviewer 9riS, EvaZ*).
We appreciat... | NeurIPS_2023_submissions_huggingface | 2,023 | Summary: This paper deals with monocular depth estimation by optical flow estimation.
The key is to decompose the flow as ego-motion-induced flow and independent flow.
To implement the idea, depth is predicted on the target image, and camera pose change from target to source is also estimated.
Then a flow network predi... | Rebuttal 1:
Rebuttal: Thank you for your constructive review and for agreeing with the motivation of our work. We list our response to your questions and concerns below.
---
### Modeling independent motion and ego-motion is not new
Compared to the mentioned papers, our method learns from **monocular** videos only **w... | null | null | null | null | null | null |
Additive Decoders for Latent Variables Identification and Cartesian-Product Extrapolation | Accept (oral) | Summary: This paper develops a theory to show how an additive decoder may be able to disentangle an image composed of several components. The proposed theory also shows how an additive decoder may produce novel images, possibly providing insights about the process performed by generative models.
Besides the theory, pa... | Rebuttal 1:
Rebuttal: **Regarding uncited relevant works:**
Object-centric representation learning is a vast field and our literature review did not cover all works. We believe many OCRL datasets have this nice additive structure and many OCRL approaches have the inductive bias of additive decoder, which underlines t... | Summary: The paper analyses the statistical identifiability of latent variables in an autoencoder with a so-called additive decoder. It is shown that under this class of decoders, the blocks of latent dimensions associated with the additive decoder can be identified. This result is further related to the ability of a m... | Rebuttal 1:
Rebuttal: **Connection to mixture models:**
We assume that by “mixture models” the reviewer refers to models of the form $p(x) = \sum_k p(c=k)p(x | c=k)$ where $c$ is the random component index. After some thinking, we do not see a clear connection between mixture models and our additive decoders. A possib... | Summary: This paper presents the identification theory for the additive mixing function. Specifically, they transfer the existing nonlinear ICA conditions from the distribution (i.e., sufficient variability) to the nonlinear mixing function (i.e., sufficient nonlinearity). Under this model, they make the connection to ... | Rebuttal 1:
Rebuttal: **On the restrictiveness of the sufficient nonlinearity assumption:**
At line 218, we refer to Appendix A.5 to provide intuition for why the sufficient nonlinearity assumption is likely to be satisfied when d_x >> d_z. However, we notice that this explanation is lacking from our submission. In th... | Summary: This paper extends the recently popular approach of constraining the nonlinear function to achieve identifiable disentanglement. Here the idea is that $\mathbf{f}$ is additive i.e. made of constituent functions that operate independently on non-overlapping partitions of the latent function. Identifiable disent... | Rebuttal 1:
Rebuttal: **Strength of the assumptions**
We agree that the additive assumption is rather strong and won’t make sense in many applications other than OCRL. That being said, we believe our work is a nice starting point for other works to investigate more expressive function classes.
**Issues raised by bloc... | Rebuttal 1:
Rebuttal: We would like to thank all reviewers for engaging with our work and asking important and thought-provoking questions.
We were glad to see that many reviewers appreciated the clarity, relevance and novelty of our work. For example, **Reviewer C5F5** said that “The paper touches on an important top... | NeurIPS_2023_submissions_huggingface | 2,023 | Summary: Motivated by the problem of disentanglement in generative models, this paper proposes a novel decoder architecture, so-called additive decoders, based on the addition of block-wise latent variables, where blocks of latents correspond to semantic factors.
An extrapolation property of their decoder is demonstra... | Rebuttal 1:
Rebuttal: **Practical relevance of Assumption 2**
Although we haven’t shown that Assumption 2 is necessary in the sense that it cannot be relaxed, we know that we cannot simply remove it and get the same guarantees since this would contradict the very well established non identifiability of linear ICA when... | null | null | null | null | null | null |
SGFormer: Simplifying and Empowering Transformers for Large-Graph Representations | Accept (poster) | Summary: Paper proposes a linear attention mechanism in the transformers for graph data. The attention becomes linear by eliminating the softmax in the dot product attention and multiplying K (transposed) and V matrices first followed by multiplication with Q matrix. The Q, K matrices are normalized by their respective... | Rebuttal 1:
Rebuttal: Thank you for the valuable feedback. Below we group related points in Weaknesses (W) and Questions (Q).
> **Q1 & W2: Optimal baseline results**
We have to respectfully point out the factual errors from the reviews. **The suggested work [1] does not experiment on the largest graph ogbn-papers100... | Summary: The paper presents a novel approach called Simplified Graph Transformers (SGFormer) for large-graph representations using Transformers. SGFormer achieves competitive performance with just one-layer attention, eliminating the need for positional encodings, pre-processing, and augmented loss. The paper provides ... | Rebuttal 1:
Rebuttal: We thank the reviewer for the positive comments and we are glad that you liked our approach.
> **(Q1): Why simple attention works? How the model learns to down-weight far-away nodes?**
Thank you for proposing this insightful inquiry. The principle of our model is that the one-layer all-pair mes... | Summary: This paper proposes Simplified Graph Transformers (SGFormer) that uses a one-layer attention to achieve competitive performance in node property prediction benchmarks.
Strengths: 1. The idea and the design of the simple global attention is novel.
2. The proposed SGFormer is efficient and scalable.
Weaknesses... | Rebuttal 1:
Rebuttal: Thank you for the valuable comments and insightful questions which can allow us to further illuminate the applicability of the theory.
> **(Q1): Applicability of the theory when subnets exist**
The theorems can be applied to the cases with FFN and non-linearity by some extension. For Theorem 1, ... | Summary: SGFormer adapts the transformer to large graphs by replacing the usual quadratic self-attention mechanism with a convex combination of linear attention and an arbitrary Graph Neural Network (GNN). The benchmark using node prediction tasks from Open Graph Benchmark (OGB) and achieve SOTA results. Moreover, the ... | Rebuttal 1:
Rebuttal: Thank you for the positive comments and constructive suggestions that can help to further improve our work.
> **(W1): Discussions of $\alpha$, $\beta$ and importance of graphs**
In Table 6 (in the appendix), we provide ablation studies for different values of $\alpha$ (the weight of GNN). The r... | Rebuttal 1:
Rebuttal: Dear Area Chairs and Reviewers,
We thank the reviewers for their time, thorough reviews and constructive suggestions. Overall, the reviewers deemed our work well written, the method novel and the results solid. Below we first restate our main contributions and distill the reviewers' comments to f... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Efficient and Learnable Transformed Tensor Nuclear Norm with Exact Recoverable Theory | Reject | Summary: This paper addresses the challenges associated with the tensor nuclear norm, which is a proxy for the low-rank property of tensors. Existing transformations for the tensor nuclear norm lack adaptability and theoretical guarantees due to fixed or non-invertible nonlinear approaches. This paper introduces a prac... | Rebuttal 1:
Rebuttal: We greatly appreciate your positive evaluation of our work. We sincerely thank you for your valuable feedback, which is incredibly important to us, and we truly appreciate the time and effort you've taken to review our work. As for your concerns, we give a detailed response which we hope can help ... | Summary: This paper presents a new approach for learning a data-adaptive and learnable column-orthogonal matrix (COM) transform for tensor nuclear norm (TNN) minimization. The authors show that the proposed transform can capture the low-rank structure of tensors more effectively than existing fixed or nonlinear transfo... | Rebuttal 1:
Rebuttal: We sincerely thank you for your valuable feedback, which is incredibly important to us, and we truly appreciate the time and effort you've taken to review our work. As for your concerns, we give a detailed response which we hope can help you fully understand our work. Please feel free to let us kn... | Summary: This paper introduces a new method of learning a transformation that can make the tensor nuclear norm better capture the low-rank structure of tensor data. The method is fast, data-adaptive, and has a theoretical guarantee of being reversible. The paper shows that the proposed method works well in experiments.... | Rebuttal 1:
Rebuttal: We greatly appreciate your valuable insights and feedback on our research. Your feedback is highly meaningful to us, and we truly value the effort you've invested in reviewing our work. We've thoroughly addressed your concerns to provide a comprehensive understanding of our study. As for your conc... | Summary: The authors propose a novel Adaptive TNN (ATNN) for tubal rank (t-SVD)-based low-rank tensor recovery and apply it to Tensor Robust PCA and Tensor Complete problems. Theoretical recovery guarantees have been established which is technically sound. However, tubal rank is known as an analog to matrix rank. Compa... | Rebuttal 1:
Rebuttal: We would like to express our sincere gratitude for your valuable comments on our work. Your feedback is incredibly important to us, and we truly appreciate the time and effort you've taken to review our work. As for your concerns, we give a detailed response which we hope can help you fully unders... | Rebuttal 1:
Rebuttal: Dear AC and reviewers,
Thank you so much for your valuable comments and we truly appreciate the time and effort you've taken to review our work. We are also pleased that the reviewers recognized the value of our work and provided positive feedback. Your feedback is very important to us, and accor... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Scaling MLPs: A Tale of Inductive Bias | Accept (poster) | Summary: This paper argues that MLP networks can perform well on challenging tasks when they are scaled up. Thus, they are able to compensate for their lack of inherent inductive bias when trained on a large amount of data with many model parameters. The authors show through experiments on popular image classification ... | Rebuttal 1:
Rebuttal: We thank the reviewer for the feedback and the interesting points raised. We will address them in the following.
**Image classification:** We agree with the reviewer and we will highlight that we focus on image classification. We largely do this since inductive bias is clearer and more intuitivel... | Summary: This paper presents a comprehensive study on the scaling capabilities of MLPs, challenging the conventional belief that this architecture is limited in its performance on vision tasks. The authors employ recent advancements in deep learning architectures to modernize MLPs while preserving their core characteri... | Rebuttal 1:
Rebuttal: We thank the reviewer for the thorough assessment of our work and the very positive feedback. We will address the questions in the following.
**Role of patch size:** We completely agree with the reviewer and also believe that breaking the images into patches plays a fundamental role for all these... | Summary: This paper studies scaling of MLPs on various image classification datasets with and without pretraining and data augmentation.
Strengths: MLP scaling trends are a welcome data point for practitioners and theoreticians alike.
Weaknesses: The paper does not carefully tune the learning rate, weight decay, and ... | Rebuttal 1:
Rebuttal: We thank the reviewer for the helpful feedback on our work. We will address the concerns in the following.
\
\
**Weaknesses:**
**Not tuning hyper-parameters:** We apologize for not clearly stating this in the paper but we did tune hyper-parameters employed upstream based on the upstream validati... | Summary: This paper studies the abilities of Vanilla MLPs and a variant of the MLP with residual connections, LayerNorm, and inverted bottlenecks, performs when scaling in size, compute, batch size, and dataset size. Many learning theory papers study MLP-based settings and claim to motivate their results by deep learni... | Rebuttal 1:
Rebuttal: We thank the reviewer for the very extensive feedback, and we are very happy to hear that the reviewer is excited about this research direction. We address the questions below.
**Mixer:** We agree, it is unclear how the token mixing (attention vs MLP) influences the inductive bias. While [1] does... | Rebuttal 1:
Rebuttal: We thank all the reviewers for their valuable feedback. We reply to each reviewer’s questions in the individual responses and have added tables and figures in the attached rebuttal pdf which we reference and explain in the responses. For convenience we also provide a detailed summary below.
\
\
Du... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Two Sides of The Same Coin: Bridging Deep Equilibrium Models and Neural ODEs via Homotopy Continuation | Accept (poster) | Summary: The paper shows that the fixed point equation solved by a Deep Equilibrium Model (DEQ) can be written as a Neural ODE.This is achieved by first introducing the equation of the solution path for the Homotopy equation, which is referred to as the Fixed Point Homotopy equation, and then deriving the appropriate h... | Rebuttal 1:
Rebuttal: Thanks for your valuable comments and suggestions. Here we address your detailed comments as below:
> **Q1**: In lines 205-210 the authors mention that neural ODE is a fixed-point problem with a “fixed-condition” $x$ with different initial point $z(t_0)$. I am not sure if I get that point. Presum... | Summary: This paper unifies DEQs and Neural ODEs via homotopy continuation. The authors propose a novel implicit model, HomoODE, which inherits the property of high accuracy from DEQs and the property of stability from Neural ODEs. Moreover, they develop an acceleration method by borrowing the idea from the DEQ-solver.... | Rebuttal 1:
Rebuttal: We gratefully appreciate the reviewer for recommending acceptance of our submission and thank his/her valuable suggestions to help improve it. We will answer all the questions that the reviewers concern about.
> **Q1**: This paper argues that "DEQs and Neural ODEs are two sides of the same coin",... | Summary: The paper uses homotopy continuation to show that Deep Equilibrium Models and Neural ODEs are effectively equivalent. They use the best properties of both worlds: DEQs for their higher accuracy and Neural ODEs for their stability to propose a new implicit method HomoODE.
Strengths: * Theoretical explorations ... | Rebuttal 1:
Rebuttal: We thank the reviewer for your valuable feedback and constructive comments! We itemize the weaknesses or comments you mentioned and answer to them.
> **Q1**: Reformulating DEQs as Neural ODEs have been done in a prior work [R4], which has not been mentioned or compared against
**A1**: Thank you ... | Summary: The work shows the connection between Deep Equilibrium Moldes (DEQs) and Neural Ordinary Differential Equations (NODEs) using homotopy continuation, a method to solve nonlinear equations. Based on the analysis, the authors suggest a new implicit model, HomoODE, which has high expressible power from DEQs and st... | Rebuttal 1:
Rebuttal: We would like to thank the Reviewer rMvz for the detailed and constructive comments. In the following, we have provided an item-by-item response to the comments.
> **Q1**: The reason why the authors choose image classification is not very clear. The practical contribution looks quite limited beca... | Rebuttal 1:
Rebuttal: We thank the reviewers for their careful reading and constructive and detailed suggestions to help us improve our manuscript. We sincerely appreciate that the reviewers consider unanimously that our work "effectively" (ZDnx) establishes an "inherent and insightful" (iBob) and "mathematically soli... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Emergence of Shape Bias in Convolutional Neural Networks through Activation Sparsity | Accept (oral) | Summary: This work is motivated by the observation that deep neural networks may have a bias towards texture in classification decisions. The paper proposes a top-k sparsity selection process that is shown to result in emergence of shape sensitive neurons while improving performance across a variety of tasks.
Strength... | Rebuttal 1:
Rebuttal: We appreciate the reviewer's kind comments and insightful feedbacks. We address the reviewers' concerns as follows:
- ResNet variant of Figure 4. We thank the reviewer for bringing up this point. We conduct experiments on ResNet-50 (See general response Figure 1 (a)) and also on ViT (General resp... | Summary: The paper proposes to use activation sparsity induced by the top-k operation to address the recently identified hard problem in DNNs, the texture bias. The operation is simple yet highly effective, as it shows good structural encoding capability (Sec 4.1), shape bias (even without training, Sec 4.2) and parts/... | Rebuttal 1:
Rebuttal: We thank the reviewer for the kind advice and constructive feedback. The reviewer consider our paper is technically solid and has novel contribution as well as good writing. We prepare the following feedback to address the reviewer’s remaining questions:
- Quantitative analysis of the visualizati... | Summary: Current deep learning models are known to have more texture bias for object recognition tasks while humans have more shape bias. They test the hypothesis that sparse coding may improve shape bias in deep learning models. Using Top-K operations, they show improved accuracy in pre-trained vison models as well as... | Rebuttal 1:
Rebuttal: We thank the reviewer for the insightful comments. The reviewer think our paper is novel and solid. Here we prepare the response to address the reviewer’s questions.
- Comparison with Ts w. Top-K only. As for the limitation of the response pages, we put the further experiments under our anonymou... | Summary: This paper proposes enforcing sparsity within a neural network through a top-K mechanism and demonstrates that this leads in an emergent fashion an emphasis on shape rather texture in the learned representations of the network. This in turn leads to improvements in robustness for network behaviour.
Strengths:... | Rebuttal 1:
Rebuttal: We thank the reviewer for the positive review and constructive feedback. The reviewer thinks our paper is technically solid with excellent contribution to a challenging goal in deep learning. We carefully prepare the following response aiming to address the reviewer’s questions:
- Top-K layer spe... | Rebuttal 1:
Rebuttal: We appreciate the reviewers’ kind responses and constructive comments. In the following, we provide further experiments to support our paper.
First, we generalize section 4.2 to ResNet-50 and ViT-B architectures. The results are documented in Figure 1 of the general response. The ResNet-50’s spa... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Equal Opportunity of Coverage in Fair Regression | Accept (poster) | Summary: Authors propose a new uncertainty-aware fairness – Equal Opportunity of Coverage (EOC) since current definition do not guarantee equal coverage rates across more fine-grained groups (e.g., low-income females) conditioning on the true label and is biased in the assessment of uncertainty.
Strengths: - paper is ... | Rebuttal 1:
Rebuttal: Thank you for your suggestions. We study fair regression under uncertainty in this paper. Our major contributions are summarized as follows:
We identified several important limitations in prior works: equal coverage is not guaranteed when conditioned on different values of $Y$ and; interval widt... | Summary: This paper studies fair machine learning under predictive uncertainty.
They propose a new notion of fairness named Equal Opportunity of Coverage (EOC), building upon "equalized coverage." While EC ensures that every demographic group receives the same level of prediction certainty, EOC ensures that every demo... | Rebuttal 1:
Rebuttal: We sincerely appreciate your thoughtful and constructive suggestions which help enhance the clarity and accessibility of this paper. Below are our replies to your questions/comments, and we will revise our ambiguous expressions to avoid confusion and misunderstanding.
1. The exchangeability of ca... | Summary: The authors propose a method to achieve same coverage rates between different groups conditioned on similar targets in addition to satisfying a marginal coverage rate on the entire population. The proposed objective is denoted as Equal Opportunity of Coverage (EOC) and the authors propose a two step approach d... | Rebuttal 1:
Rebuttal: We appreciate your valuable and informative suggestions and would clarify the questions and concerns in the following.
**Q1**: Unclear writing in Section 4.
**A1**: We will rephrase this paragraph and add a new Figure 3 illustrating how $\beta_m$ is enforced to be equal within the same bin (see ... | Summary: This paper studies the problem of algorithmic fairness using uncertainty quantification methods in regression settings. The main novelty of this work over previous work is taking the notion of “equalized coverage” and strengthening it by considering finer groups by additionally conditioning on the target varia... | Rebuttal 1:
Rebuttal: We extend our sincere appreciation for your valuable feedback and suggestions. Regarding your concerns, we would like to offer further clarification.
**Q1**: Modest improvement in interval width and degradation in coverage rates.
**A1**: The research problem of the trade-off between interval wid... | Rebuttal 1:
Rebuttal: We sincerely appreciate all the recognition and suggestions from all reviewers, which greatly contribute to enhancing both the quality and clarity of our paper. We have thoughtfully addressed several common concerns raised by the reviewers and would incorporate them into the next version of our pa... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Beyond Myopia: Learning from Positive and Unlabeled Data through Holistic Predictive Trends | Accept (spotlight) | Summary: In this paper, the authors investigate a phenomenon appearing in positive and unlabeled data learning. i.e. predictive trends is different for positive and negative classes in binary classification task. Based on this observation, the authors propose a holistic approach to capture the predictive trends of diff... | Rebuttal 1:
Rebuttal: Thanks for your insightful comments. We address your concerns as follows.
### Q1: Missing comparing algorithms in experiments
Thank you for your valuable suggestions. Actually, due to space limitations, we have included the comparing algorithms in Section F of the Appendix. We will also add a bri... | Summary: The paper focuses on addressing the positive and unlabeled learning problem. The authors make an observation that a simple resampling method can yield strong early-stage performance, which has been overlooked in previous literature. Building on this insight, the paper proposes a trend detection measure to tack... | Rebuttal 1:
Rebuttal: Thanks for your insightful comments. We address your concerns as follows.
### Q1: Further discussions on how this method applies to other areas
Please kindly refer to our responses to Question 3 from Reviewer vUs8.
### Q2: Whether the method can perform well with different positive priors?
As sho... | Summary: In this paper, the authors investigate the topic of positive and unlabeled learning, and solve it from an interesting perspective, namely trend score. The story begins with the empirical observations of different temporal output score trends of positive instances, unlabeled negative instances, and unlabeled po... | Rebuttal 1:
Rebuttal: Thanks for your insightful comments. We address your concerns as follows.
### Q1: Whether resampling is a must in the framework? What's the relationship between trend score and resampling?
We argue that resampling is essential in this framework. As we emphasized in Sections 2.2 and 2.3, a benign... | Summary: The paper addresses the positive and unlabeled learning problem and presents interesting and meaningful observations. Inspired by the observations, the authors propose an innovative approach that effectively discriminates between positive and negative examples within unlabeled data. The performance of this app... | Rebuttal 1:
Rebuttal: Thanks for your insightful comments. We address your concerns as follows.
### Q1: Why prediction trends for positive and negative examples are significantly different?
The divergence in trends between positive and negative examples can primarily be attributed to the strong fitting capabilities o... | null | NeurIPS_2023_submissions_huggingface | 2,023 | Summary: The paper proposed an interesting idea of using the trend detection technique for positive and unlabeled learning (PUL). The core problem of PUL is thus reformulated as recognizing trends in these scores. A new TPP-inspired measure is then utilized to solve the task. The idea of the paper is interesting; it is... | Rebuttal 1:
Rebuttal: # Reviewer cdNW:
Thanks for your insightful comments. We address your concerns as follows.
### Q1: It is unclear whether existing methods witness the same phenomenon.
A： Most PUL methods currently use a combination of regularization techniques and refinement on mislabeled examples brought by th... | null | null | null | null | null | null |
Is Distance Matrix Enough for Geometric Deep Learning? | Accept (poster) | Summary: This paper proposes k-DisGNNs, that can use distances alone to represent higher-order interactions, instead of the direct use of angles, torsional angles, and dihedral angles. The authors try to build a bridge between geometric deep learning and the traditional expressiveness of graph representation learning w... | Rebuttal 1:
Rebuttal: We thank Reviewer fGWC for the careful and detailed review. We address the comments below.
> W1: Prohibitive Computational Costs for k > 2: The computational demands ... escalate dramatically when k > 2, ..., when k is set to 3, the model becomes computationally infeasible for the QM9 dataset, ..... | Summary: The paper proposes $k$-DistGNNs, a sequence of k-(F)WL-like models that operate on the distance matrix of a geometric graphs. These models are able to compute higher-order geometric invariants, generalise existent models, and for certain values of $k$ are shown to be able to approximate any geometrically invar... | Rebuttal 1:
Rebuttal: We thank Reviewer 5v5F for the careful review. We address the comments below.
> Note that WL generalisations for Geometric GNNs were also studied in On the Expressive Power of Geometric Graph Neural Networks (ICML 2023), which is not mentioned in the text as far as I could see.
>
Thanks for the... | Summary: In this paper, the authors argue that the distance matrix associated with a geometric graph is sufficient to build powerful geometric neural networks.
In particular, the paper proposes a new family of graph architectures, named k-DisGNN, operating on a complete graph with node distances on the edges.
These arc... | Rebuttal 1:
Rebuttal: We thank Reviewer w5WJ for the insightful and detailed comments. We address the comments below.
> I am not sure I understand the difference between Eq. 4 and 5. It seems to me that the indices j and w are just swapped, but in the end the set contains the same tuples. Could you clarify this?
Than... | Summary: This work proposes neural networks for operating on distance matrices obtained from 3D/Euclidean point clouds. New architectures which generalise the k-WL/k-tuple higher order GNN framework (Morris et al.) to distance matrices are proposed, along with proofs of their universality/completeness for k=2 (scalar) ... | Rebuttal 1:
Rebuttal: We thank the Reviewer UsYi for their detailed and insightful review. We address the comments below.
> W1: Ambiguous claim re. Geometric Deep Learning
Thanks for the insightful comment. Indeed, in this paper, we mainly use the term geometric deep learning to refer to learning over points embedded... | Rebuttal 1:
Rebuttal: We sincerely thank all the reviewers for their careful reading, and the thoughtful and detailed comments! The **figures** and **tables** referred to in separate responses are included in the following supplementary **PDF** file.
Best,
Authors.
Pdf: /pdf/5bf2071072ce50ad486aa4154748e9487dcb5910.p... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
SOAR: Improved Indexing for Approximate Nearest Neighbor Search | Accept (poster) | Summary: The paper proposes a neat variant of spill trees for approximate nearest neighbor search (ANN). The idea is to reduce correlation between the set of points retrieved from each of the partition elements in a spill tree (where the partitions elements overlap or "spill over") by adding a penalty term in the assig... | Rebuttal 1:
Rebuttal: Thank you for your review. We’ll address the weaknesses and questions point-by-point:
### Weakness
* We agree, this assumption isn’t entirely accurate. However, we’ve found this assumption to still be very effective, because it’s somewhat of a worst-case; the hardest queries tend to be those far f... | Summary: This paper introduces an innovative technique for constructing a vector index to perform approximate nearest neighbor search, building upon the foundations of vector quantization and redundancy through spill trees. The authors propose a novel loss function aimed at identifying k-alternate centroids in the vect... | Rebuttal 1:
Rebuttal: Thank you for your review. We’ll address the weaknesses and questions point-by-point:
### Weaknesses
1. We’ve added Figure 2 in the rebuttal PDF to better describe the implementation details and space complexity. We also plan on adding algorithm blocks to the appendix to more fully describe our in... | Summary: The paper discusses how a data index is built for Approximate Nearest Neighbor Search (ANNS). However, existing methods train and compute redundant representations independently, leading to inefficiency and inaccuracy. The paper proposes a novel data indexing technique called SOAR for approximate nearest neigh... | Rebuttal 1:
Rebuttal: Thank you for your review. W1/W2 + Q1/Q2 are all with regards to datasets chosen for experiments, so we will address this first.
* Our initial paper involved experiments on four datasets, and we believe this is more than sufficient to prove our point. L2 and angular datasets are simply more common... | Summary: The paper revists the "vector quantization" technique for nearest neighbor search. In the VQ technique, points are represented by their closest centroid in a partition schema, typically based on k-means. Answering a query consists of finding the nearest centroids to the query and enumerating points represented... | Rebuttal 1:
Rebuttal: Thank you for your review. To answer your questions point-by-point:
1. We’ve added Figure 2 in the rebuttal PDF to better describe the implementation details and space complexity. Also see Section 3.5 and Appendix A.3 for details. We also plan on adding algorithm blocks to the appendix to more ful... | Rebuttal 1:
Rebuttal: Thank you all for taking the time to review our work. The two most common points of contention and confusion among the reviews were with regards to our experiments and with implementation details (especially memory footprint). We believe we have addressed both in our rebuttal. Our added material f... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Sorting with Predictions | Accept (poster) | Summary: This paper presents sorting algorithms augmented by prediction under two settings: (i) sorting with positional predictions and (ii) sorting with dirty and clean comparisons. In (i), the algorithm can use a predicted position $\hat{p}(i)$ of each element $i$, and in (ii), the algorithm is provided a faster but ... | Rebuttal 1:
Rebuttal: We thank the reviewer for their positive feedback and interesting comments. We address specific comments below.
- Derandomizing the dirty comparison algorithm is an interesting question, and we do not know how to do this. Indeed the height of the tree can be made O(log n) deterministically via ro... | Summary: This paper studies learning-augmented sorting, where an algorithm obtains access to a potentially erroneous prediction on the sorting of an array. The paper considers two prediction models. In the positional predictor, every item has a prediction for its place in the sorted array, and the error of item i is th... | Rebuttal 1:
Rebuttal: We thank the reviewer for their feedback and in particular pointing out additional related works. We inspected these works more closely, and while they may seem similar at first glance, they actually consider different problems, and the techniques from these papers cannot be used to conclude any o... | Summary: The paper studies the oracle-augmented sorting problem, in which, in addition to the array, an oracle is provided to return ‘reference answers’ for the sorting of the array. The oracle is presumably efficient – queries are assumed to take $O(1)$ time – but might be noisy, i.e., at times return incorrect answer... | Rebuttal 1:
Rebuttal: We thank the reviewer for their positive feedback and interesting questions and observations.
**Simplicity:** We believe that the simplicity of our algorithms is a strength as it makes them more widely appreciable, especially given that this is the initial investigation of the subject.
**Amortiz... | Summary: The paper proposes algorithms for sorting a set of items when unreliable extra information is available in the form of positional predictions or comparisons. The paper proposes a randomized algorithm for the former and deterministic algorithms for the latter problem.
Strengths: The paper proposes interesting ... | Rebuttal 1:
Rebuttal: We thank the reviewer for their comments and interesting suggestions. We believe we can address the perceived potential weakness of suitability for NeurIPS, and provide answers to this and other questions below.
**Suitability to the NeurIPS community:** We recognize that our paper is not a classi... | Rebuttal 1:
Rebuttal: We thank all reviewers for their valuable feedback and many interesting questions and suggestions. We provide answers in individual responses. It appears the main concerns besides otherwise mostly positive feedback were whether our work is suitable for the NeurIPS community (reviewer JwAo) and tha... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
PDE-Refiner: Achieving Accurate Long Rollouts with Neural PDE Solvers | Accept (spotlight) | Summary: This paper targets the prediction of future PDE states from given initial conditions with deep-learning based surrogate simulators over long rollouts. The authors propose a method titled PDE-Refiner that combines three aspects: 1) autoregressive solution rollouts similar to numerical PDE solvers, 2) direct nex... | Rebuttal 1:
Rebuttal: Thank you for your encouraging review and thoughtful feedback. We provide answers to your questions and concerns below.
**2D Kolmogorov Flow - the correlation computation shown for the ML surrogates is performed differently**
The evaluation for the 2D Kolmogorov Flow (lines 328-333) follows the ... | Summary: This paper focuses on the long-term simulation of time-dependent PDEs. The authors find that high frequencies in solutions are always neglected, impeding the long-term rollout performance. This paper presents the PDE-Refiner by introducing the diffusion model to refine the model prediction. PDE-Refiner can boo... | Rebuttal 1:
Rebuttal: Thank you for your valuable review and helpful suggestions, especially the question on the low data regime was very interesting. We respond to your questions and concerns below.
**1. From Figure 1(b), we can find that the PDE-Refiner will smooth the amplitude of high frequencies. This should be d... | Summary: This work introduces a PDE-Refiner model to address the challenge of accurately predicting the process of time-dependent partial differential equations (PDEs) over long rollouts. The model can be divided in two parts: prediction (based on FNO or UNet) and multi-step refinement (based on the diffusion model). B... | Rebuttal 1:
Rebuttal: Thank you for helpful review and thoughtful feedback. Please find below our response to your questions and concerns raised.
**Analytical support for 2D case for the statement “the primary errors in predicting a one-step solution arise from inaccuracies in modeling the dynamics of these low freque... | Summary: Recently, there has been active research on using deep neural networks as surrogates to predict the temporal evolution of solutions in time-dependent partial differential equations (PDEs). Improving the accuracy of long-term predictions for surrogate models that take previous time steps' solutions as inputs an... | Rebuttal 1:
Rebuttal: Thank you for your constructive review and feedback. We provide below clarifications to your questions and concerns raised.
**Clearer explanation of how Gaussian noise enables denoising across sigma values / Equation (4), shouldn't the $u(t)$ after 'NO' be $\hat{u}^k$ instead?**
The refinement p... | Rebuttal 1:
Rebuttal: We would like to thank all reviewers for their constructive and valuable feedback.
We are encouraged to see that reviewers highlight this study as interesting (YYxA, JgV1, xf2d), the good presentation of the paper (YYxA, xf2d), and the thorough ablation studies (K6r8, JgV1, xf2d). For example, r... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Streaming Factor Trajectory Learning for Temporal Tensor Decomposition | Accept (poster) | Summary: In traditional tensor decomposition, the factor of decomposition is static. Therefore, it was not possible to fit time-evolving data well. This paper develops SFTL using a Gaussian process to capture the time-developing trajectory of the factors. Their naive implementation would have increased computational co... | Rebuttal 1:
Rebuttal: Thanks for your insightful and constructive comments. Here are our responses. C: comments; R: responses
>C1: "There are some known tricks to reduce the computational complexity of the Gaussian process from O(n^3), ... I think it is possible to achieve fast computations without Hartikainen's elega... | Summary: The authors highlight that real-world tensor data often includes time information, representing the timestamps of interactions between objects in different modes. However, most existing decomposition methods for tensors with temporal information only estimate static factors for each object and cannot capture t... | Rebuttal 1:
Rebuttal: Thanks for your comments and questions. Here are our clarification and responses. Please see global response R1 for our clarification of Fig. 1. C: comments; R: responses
> C1: "the real-world applications are not explained, so it is also not clear to understand what are the application areas of... | Summary: This paper studies the dynamic tensor factorization under the streaming data setting. In particular, the authors propose to use a state-space Gaussian process (GP) to parameterize the CP or Tucker factors as functions of time. To efficiently learn the GP and handle streaming data, an equivalent SDE and a condi... | Rebuttal 1:
Rebuttal: Thanks for your valuable comments. Here are our responses. C: comments; R: response
> C1: "Although the problem setting is new in the tensor decomposition field, most techniques used in this work seem standard, such as state-space Gaussian process and expectation propagation."
R1: We do agree th... | Summary: The work proposes a new approach for the decomposition of a tensor with a temporal dimension. The developed method can apply to streaming data where decomposing the entire tensor from scratch would be prohibitive.
The key idea is to model each entry (or parameter) in the decomposition factor as a function of... | Rebuttal 1:
Rebuttal: Thanks for your detailed and insightlful comments! Here are our responses. Please see global responses R2 & R3 for your questions regarding regression baselines and Tensor Train and Tensor Ring decompositions. C: comments; R: responses.
>C1: "Parameters of the decomposition factor are sampled fro... | Rebuttal 1:
Rebuttal: We thank all for the reviewers for their careful, valuable and constructive comments! We post our global response to a few questions here. C:comments; R: responses
>C1: more explanation/clarification for Figure 1.
R1: Figure 1 explains the key step in our development of the online filtering al... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Towards a Unified Framework of Contrastive Learning for Disentangled Representations | Accept (poster) | Summary: In this paper the authors propose a new dissimilarity measure function for contrastive representation learning. They provide the identifiability of the factors of variations in four different contrastive objectives. Specifically, they describe the connection between the learned representations and true factors... | Rebuttal 1:
Rebuttal: We appreciate that our contribution is valued, and we are grateful for the comments on where we can provide further elaboration.
**Identifiability in practical scenarios**
The scenario described by the reviewer comes closest to our experiments with 3DIdent. However, instead of using data augment... | Summary: This paper presents an analysis focusing on the disentanglement of the characteristics of internal representations obtained by contrastive learning. The aim is to recover the internal representation of the true latent space up to some transformation by placing certain assumptions on the data generation process... | Rebuttal 1:
Rebuttal: We thank the reviewer for the detailed assessment and for pointing out where further comparison and discussion may be needed.
**Gap between theoretical claims and realistic experimental settings**
It is true that the assumed conditional (Eq. 1) is not fully consistent with data augmentation, and... | Summary: This paper studies a unified framework of contrastive learning. The authors extend the theoretical result of [1] to embrace more contrastive losses, adapt it to marginal distributions beyond uniform distribution, and further relax some conditions that are required in [1]. The experimental results support the t... | Rebuttal 1:
Rebuttal: We thank the reviewer for the valuable feedback and are glad that the reviewer recognizes the novelty and benefit of our work.
Q1. In Section 4.5 (line 268), we analyzed the impact of larger dimensions on disentanglement performance, which requires more GPU memory. We decided to reduce the batch ... | Summary: This paper proves two identifiability theorems for a broad family of contrastive learning methods, without imposing the commonly used independence constraint. The theoretical findings are validated on several benchmark datasets, and the theoretical and practical limitations are fully discussed.
Strengths: 1.... | Rebuttal 1:
Rebuttal: We appreciate the positive assessment. We will discuss the intuition behind the theory in more detail in the revised version and add footnotes on the used mathematical concepts.
We also thank the reviewer for pointing out typos.
**Why β=2 is excluded in the conditions of Theorem 2?**
Excluding β... | Rebuttal 1:
Rebuttal: We thank all reviewers for their constructive feedback and overall positive evaluation. Below we describe the main concerns and how we intend to address them. Additional points are discussed in the individual responses to the reviewers.
**Confusion about the generative process**
We will add a cl... | NeurIPS_2023_submissions_huggingface | 2,023 | Summary: The paper develops a new theoretical framework for learning disentangled representations with contrastive methods. Building on the theory that contrastive methods approximately invert the data generating process, the paper extend it by proposing a new dissimilarity measure. The proposed framework supports a br... | Rebuttal 1:
Rebuttal: We thank the reviewer for the positive evaluation. We appreciate the nice summary of our research and the hints on where we can further expand the intuition behind the design choices. Please find below a detailed response to the questions raised.
Q1. Yes, the optimal $\alpha$ and $\tilde{\alpha}$... | Summary: Authors propose a generalized contrastive learning framework extending the guarantees for disentanglement learning. The paper also links contrastive learning objective with identifiability results in both week and strong identifiability setting.
Strengths: - The paper is mostly well written and easy to follow... | Rebuttal 1:
Rebuttal: We thank the reviewer for the detailed feedback and advice on how to improve the quality of the work. We will include the mentioned limitation in the revised version and are happy to provide clarification on the raised issues.
The reviewer assessed the present work from the perspective of causal ... | null | null | null | null |
Advice Querying under Budget Constraint for Online Algorithms | Accept (poster) | Summary: This paper considers learning augmented online algorithms where there is a bound on the number of predictions that can be queried. Three problems are considered within this framework: Ski-rental, Secretary Problem, and Non-clairvoyant scheduling.
For the ski-rental problem, they consider a model where the al... | Rebuttal 1:
Rebuttal: We thank you for reviewing our submission and for your feedback. Below, we address the weaknesses and questions that you raised.
## Weaknesses
- We explain how to choose the time of querying the prediction in the discussion following Lemma 2.1 (paragraph starting in Line 165). Given a learning-au... | Summary: The paper studies learning-augmented algorithms in a setting where predictions are a scarce resource and algorithms need to decide when to request them. The framework is instantiated over (i) ski-rental with a single prediction, (ii) the secretary problem with B predictions, (iii) preemptive non-clairvoyant sc... | Rebuttal 1:
Rebuttal: We thank you for reviewing our paper, and for your positive evaluation of our work. We are very happy that you appreciated the paper and enjoyed reading it. Below we address the weaknesses and questions you raised.
## Weaknesses
**1.** Regarding the paper by Im et al., it is indeed a closely rel... | Summary: Authors consider ML-augmented algorithms with a limited number of predictions.
They consider three problems: ski rental, secretary, and non-clairvoyant
scheduling. Their result on non-clairvoyant scheduling assumes perfect
predictions, while for ski rental and secretary problems they assume
knowledge of the pr... | Rebuttal 1:
Rebuttal: We thank you for reviewing our submission and for your feedback. Below, we address the weaknesses and questions that you raised.
## Weaknesses
See the common answer.
## Questions
- The paper by Antoniadis et al. studies the problem of dynamic power management (DPM), which is a generalization of ... | Summary: In this paper the authors re-consider classic online optimization problems in the area of learning-augmented algorithms, under the setting where the algorithms that have access to a limited number of predictions. Additionally, the setting that is considered assumes that as the algorithm progresses the predicti... | Rebuttal 1:
Rebuttal: We thank you for the time and effort spent on our submission and for the positive evaluation of our work. We address below the weaknesses and questions that you raised.
## Weaknesses
See the common answer.
## Questions
- In the secretary problem with perfect advice, as you explained in your com... | Rebuttal 1:
Rebuttal:
We would like to thank the reviewers for their feedback and thoughtful comments toward improving our paper. We address below some general comments and weaknesses that were raised by the reviewers.
### Ski-rental: Tightness of the results
We do not provide tightness results for the ski-rental pro... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Selective Sampling and Imitation Learning via Online Regression | Accept (poster) | Summary: In this work, the authors make several contributions. Firstly, they propose a novel selective sampling algorithm (Algorithm 1) that utilizes a regression oracle and provide regret bounds and query complexity analysis. Secondly, they present a lower bound that demonstrates the unavoidable dependency of query co... | Rebuttal 1:
Rebuttal: Thank you for the detailed review and for your questions. Before providing the answers to your questions, we first recap our notation for the imitation learning setting which we will repeatedly rely upon in the rebuttal.
- **Noisy Expert**: We assume that the learner interacts with an expert who... | Summary: The paper proposed a selective sampling scheme on an online regression oracle over a function class $\mathcal{F}$ and time horizon $T$ for multiclass setting, which has the regret guarantee $Reg^{\ell_{\phi} } (\mathcal{F}, T)$, and we assume the label $y_t$ is sampled through a link function $\phi$. The algor... | Rebuttal 1:
Rebuttal: Thank you for your valuable feedback. We will be sure to incorporate your suggestions in the final version of the paper. Please find our response to your concerns below.
### How does $\delta$ error propagate?
Our algorithm can be easily modified if $\Delta\_t(x\_t)$ could only be solved approxi... | Summary: This paper studies selective sampling and imitation learning. In this setting a series of states arrives online, and a learner has the ability to choose to query expert labels/policy in order to correctly label the states/perform as well as an expert policy. The goal is to minimize the regret (i.e. the drop in... | Rebuttal 1:
Rebuttal: Thank you for your valuable feedback. We will be sure to incorporate your suggestions in the final version of the paper. Please find our response to your concerns below.
### Solving optimization problem in equation (5) for Algorithm 1 and equation (6) for Algorithm 2
Please find our shared re... | Summary: The paper studies algorithms for selective sampling with applications to imitation learning.
In the former setting the authors propose an algorithm build around an online regression oracle and derive bounds on the overall regret and number of queries.
These are provided for both a general (potentially adversar... | Rebuttal 1:
Rebuttal: Thank you for your valuable feedback. We will be sure to incorporate your suggestions in the final version of the paper. Please find our response to your concerns below.
> Limited Experimental Evaluation
Our motivation in this work is to theoretically understand regret and query complexity boun... | Rebuttal 1:
Rebuttal: ## Solving optimization problem in equation (5) for Algorithm 1 and equation (6) for Algorithm 2 computationally efficiently
Our work is theoretical in nature, and our focus in the paper is thus to understand the statistically efficient (query complexity), and to develop algorithms, for selectiv... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Bifurcations and loss jumps in RNN training | Accept (spotlight) | Summary: The paper studies the dynamical systems properties of piecewise-linear recurrent neural networks (PLRNNs). Specifically:
- Section 3.1 (and its accompanying Appendix A.1-A.5) analytically categorizes the fixed points of 2-dimensional PLRNNs, and analyzes its bifurcations by explaining how the fixed points (and... | Rebuttal 1:
Rebuttal: We thank the referee for her/his constructive comments and for bringing up a couple of important points we agree need to be addressed.
**Weaknesses**
*”Improving presentation”*: We will 1) take up the referee’s suggestions and add a pg. to the Introduction which primes more clearly for the impor... | Summary: The paper analyses a class of ReLU-RNNs (PLRNNs) from the perspective of bifurcations in dynamical systems. This gives rise to new insights into the training process of RNNs (as specific bifurcations are associated with exploding or vanishing gradients -- a major issue in training RNNs on long-term dependency ... | Rebuttal 1:
Rebuttal: We thank the referee for the very constructive and fair feedback, and the supportive response!
**Weaknesses**
*”Limited significance for design of training algorithms”*: Although the exploding or vanishing gradient problem has been studied for a while now, viewing the training process from the p... | Summary: The main focus of the paper is piecewise-linear RNNs, that is ReLU-based RNNs called PLRNNs, and how certain bifurcations that take place during their training phase will have an adverse effect on their gradients and hence will cause sudden jumps in the loss functions. This helps shed light onto the difficulty... | Rebuttal 1:
Rebuttal: We thank the referee for the very supportive and constructive feedback!
**Weaknesses**
*”Theoretical guarantees for algorithm”*: Obtaining rigorous theoretical guarantees is challenging in the current case (and sometimes, historically, theoretical understanding of convergence and other propertie... | Summary: In this paper, the authors provide a thorough analysis of the bifurcation structure of RNNs and its relationship with discontinuities of the loss function during the training process. The authors also present a novel algorithm for detecting fixed points and k-cycles in ReLU-based RNNs, providing a new tool for... | Rebuttal 1:
Rebuttal: We thank the referee for the very supportive and overall positive assessment of our work!
**Weaknesses**
*“Limited discussion of bifurcation types”*: Yes, true, originally there were two types of bifurcation for which we could offer rigorous proofs, so we focused on them in our presentation. To ... | Rebuttal 1:
Rebuttal: We thank all four referees for their generally positive evaluation and supportive response! We appreciate the detailed reading of our manuscript, and the many helpful comments pointing out where our paper needs clarification or extension of results. Besides the detailed reply to all issues raised ... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Convergence of Actor-Critic with Multi-Layer Neural Networks | Accept (poster) | Summary: The paper proposes an analysis of the the actor critic setting with function approximator and it proves the convergence using deep neural networks with an arbitrary number of hidden layers.
Strengths: The claims are ambitious.
Weaknesses: The analysis of the paper sidesteps many key elements from the literat... | Rebuttal 1:
Rebuttal: We welcome the opportunity to contextualize our results with respect to the earlier literature showing examples of divergence with non-linear function approximations. However, we do want to note to the reviewer that they misunderstand the state of this research area. Please see our responses below... | Summary: The paper studies non-asymptotic convergence rate of actor-critic algorithm. The critic network is parameterized by multi-layer neural network whose activation function is assumed to be smooth, excluding ReLU operator. The actor is assumed to be general smooth non-linear function.
The convergence rate for me... | Rebuttal 1:
Rebuttal: > The main weakness...
Our contribution is not a simple combination of two papers. Even superficially, note that our paper deals with actor-critic, yet our Lyapunov function is not the same as Olshevsky et al. (which relied on square norm). Further, Olshevsky et al. was for the i.i.d. sampling c... | Summary: The authors prove a bound on the approximation error of sample-based actor critic learning on a more general and realistic setting, namely allowing for a NN approximator of any depth. This is a very non-trivial result and an interesting contribution to the literature.
Note: I have not gone through the full pr... | Rebuttal 1:
Rebuttal: > The paper is very well written and the ideas are clearly communicated. I especially appreciate the tables and figures to aid in explaining the approach and how it fits into the literature.
> The main theorem is a very interesting theoretical result.
Thank you for the encouraging review!
> Y... | Summary: This is a theoretical paper which studies actor-critic reinforcement learning algorithms in which both the actor and the critic are represented using deep neural networks with more than one hidden layer. The previous research addressed linear representations and neural networks with only one hidden layer. This... | Rebuttal 1:
Rebuttal: Thank you for the detailed review and questions! We'll clarify these points in the final version of the paper. Meanwhile, our point-by-point replies are below.
> In the paper, the authors emphasize the need to stay close to the initial conditions of the critic (e.g. in Sec. 2.4). This makes sens... | Rebuttal 1:
Rebuttal: Dear Reviewers,
We would like to use the "global" review to provide two simulations asked for by some of the reviewers.
*First*, one reviewer (implicitly) asks for evidence supporting our Assumption 2.7. Our attached pdf actually
shows evidence supporting the stronger assumption
$$ || Q(w_1)... | NeurIPS_2023_submissions_huggingface | 2,023 | Summary: This is an RL theory paper with an improved convergence bound for actor-critic methods
Strengths: To be frank, I'm not a theory person at all and I have no idea why this paper is assigned to me. I cannot really judge the theoretical contribution of this work. By assuming all the statements are correct, I per... | Rebuttal 1:
Rebuttal: We appreciate the reviewer's frankness! But we would like to to disagree: we think applied RL people are qualified to review RL-theory papers.
Indeed, we do not want to be in a state where only RL-theory people are reviewing each other's papers: that is likely to lead to the community becoming i... | Summary: This paper presents a convergence analysis of Actor Critic method with multiply layer networks.
Strengths: The convergence analysis of AC with multi-layer networks is important, as AC methods with neural networks plays the core role of the success of DRL.
Weaknesses: The paper is difficult to follow and th... | Rebuttal 1:
Rebuttal: Thank you for bringing these points up! Our responses are below.
> What are the differences for AC methods with single hidden layer networks and multiple layer networks, especially for proving the convergence? In the current version, these differences are not well explained. To improve the clari... | Summary: The paper is a well written and clear result demonstrating the convergence of the actor critic algorithm. To my knowledge, this is the first example of global convergence of the actor critic algorithm using neural network parametrization. It is a good extension of the actor critic sample complexity analysis gi... | Rebuttal 1:
Rebuttal: > Since a finite state space is being assumed here, the comparison to existing results such as Wang et. al (2019) and Cayci et. al. (2022) does not seem to be valid. Both these works assume an infinite state space. Since the constant c1 is a multiple of the cardinality of the state space, an infin... | Summary: This paper establishes the convergence of single-timescale actor-critic with neural networks representing the value and policy with > 1 layer, strengthening over prior results in the linear setting and two-scale approaches.
Strengths: I will preface my review by saying that I have little background on conve... | Rebuttal 1:
Rebuttal: > I found that despite this lack of background, this paper was an absolute pleasure to read. The paper is very well-written, and the authors do a great job of motivating the problem and the technical approach. Each assumption is well-motivated, and the two tools -- nonlinear gradient splitting and... |
Goal-conditioned Offline Planning from Curious Exploration | Accept (poster) | Summary: The paper presents a method that using a dataset of rollouts and a dynamic model create a graph like structure in-order to train goal conditioned policies. The nodes of the graph are selected based on it high local value function, and connected using a model based mechanism offline that avoid the myopic behavi... | Rebuttal 1:
Rebuttal: We thank the reviewer for their comments and suggestions, which we address sequentially as follows.
# Contributions
We have added to the Appendix a detailed algorithm box describing the practical algorithm, and how the different components are combined. In the context of this response, we would l... | Summary: This paper tackles the problem of performing goal-conditioned control using data collected from curious exploration. They first perform an empirical evaluation of several methods for learning value functions from this offline data, and then identify an issue with estimated value functions called $\mathcal{T}$-... | Rebuttal 1:
Rebuttal: We thank the reviewer for their assessment and detailed feedback. In this response, we comment on Figure 5, on the generality of the method, and provide clarifications on graph-based aggregation.
# Figure 5
We appreciate the suggestion for extending Figure 5 to include other environments from the... | Summary: This paper investigates the issue of mitigating estimation artifacts in the value function of off-policy RL. The authors adopt an unsupervised scheme to collect exploratory data, which is subsequently used to train both off-policy and offline RL algorithms. In their analysis, the authors identify a geometric c... | Rebuttal 1:
Rebuttal: We are grateful for the detailed comments and clear listing of points, which we address sequentially.
Please let us know if further clarification or evidence is needed.
# Performance comparison
As noted by the reviewer, performance in the four environments is not comparable to that reported in ... | Summary: This paper addresses the challenge of extracting goal-conditioned behavior without the need for continued environment interaction. The authors tackle this challenge by leveraging unsupervised exploration to collect diverse experiences. The study explores the suitability of different off-policy actor-critic alg... | Rebuttal 1:
Rebuttal: We thank the reviewer for their detailed feedback and clear assessment. In this response we further clarify our technical contributions, provide a discussion and experiments regarding the generality of our method and answer the reviewer's question.
# Technical contributions
While both model-based... | Rebuttal 1:
Rebuttal: We thank all reviewers for their constructive feedback, which helped us to improve our paper. We ran additional baselines, added evaluations on more environments, and provided individual responses.
Here, we summarize the added validation, and then address the applicability of our method in general... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Weakly-Supervised Concealed Object Segmentation with SAM-based Pseudo Labeling and Multi-scale Feature Grouping | Accept (poster) | Summary: This paper proposes to tackle Concealed Object Segmentation tasks by generating high-quality pseudo segmentation masks using state-of-the-art vision foundation models--SAM. SAM takes prompts as provided sparse annotations (e.g. points or scribbles) to produce figure-ground pixel labels, which are used to trai... | Rebuttal 1:
Rebuttal: Thanks for the valuable comments. Unless otherwise specified, all experiments are conducted on the *COD10K* dataset for the COD task to save space.
**Q1: Literature review.**
We focused on concealed object segmentation and mainly discussed and compared existing methods in this field. However, we... | Summary: This paper tackles the problem of Weakly-Supervised Concealed Object Segmentation (WSCOS). The first strategy is using SAM, which is a vision foundation model for generic segmentation. To handle the miss-segmentation of SAM, the authors propose techniques using multi-augmentation and entropy-based scoring. In ... | Rebuttal 1:
Rebuttal: Thanks for the valuable comments. Unless otherwise specifically stated, all experiments are conducted on the *COD10K* dataset for the COD task to save space.
Q1: Fair comparison with the same backbone.
Q1-A: In Table 3, ''Baseline'' denotes using SAM to generate only one segmentation mask from ... | Summary: This paper investigates Concealed Object Segmentation based on the recently proposed foundation model SAM, which provides dense annotation for learning. To alleviate the la-bel noise caused by imperfect segmentation results of SAM on concealed objects, they adopt pixel-level and image-level uncertainty-based s... | Rebuttal 1:
Rebuttal: Thanks for the valuable comments. Unless otherwise specified, all experiments are conducted on the *COD10K* dataset for the COD task to save space.
**Q1: ''Weakly-supervised'' claim is not appropriate.**
We follow SCOD [29] and use the term ''weakly-supervised'' to indicate that training data on... | Summary: The submission presents a method for Weakly-Supervised Concealed Object Segmentation (WSCOS) to address the challenges of segmenting objects that are well blended with their surroundings using sparsely-annotated data.
+ The method introduces a multi-scale feature grouping module that groups similar features a... | Rebuttal 1:
Rebuttal: Thanks for the valuable comments. If not specifically stated, all experiments are conducted on the COD task with *COD10K* for space limitation.
**Q1: Performance gains are from SAM, not from the weakly-supervised strategy.**
While SAM has shown impressive performance for natural scene images, it... | Rebuttal 1:
Rebuttal: We extend our sincere gratitude to all the reviewers (**R1**-76do, **R2**-yhjR, **R3**-ka5E, and **R4**-DmES) for their insightful and considerate reviews, which help us to emphasize the contributions of our approach. We are encouraged to hear that the reviewers found the work is well-motivated wi... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
ProPILE: Probing Privacy Leakage in Large Language Models | Accept (spotlight) | Summary: This paper addresses the pressing issue of potential privacy breaches resulting from the use of large language models (LLMs). The authors propose a novel tool named ProPILE that empowers data subjects, or the owners of personally identifiable information (PII), to evaluate the risk of their personal data being... | Rebuttal 1:
Rebuttal: We thank the reviewer for thorough review and valuable comments. We are glad that the reviewer recognized our work as "timely" and "innovative". The comments also helped us to elaborate and clarify our paper. We hope that our response address the reviewer's concern and clarify any uncertainty.
_... | Summary: The paper presents ProPILE, which is a probing tool designed to help identify if the PII of the user can be revealed by probing the LLM-based services. The authors consider two attributes of PII: linkability and structurality, and two types of access: black box and white box. The authors also provide two quant... | Rebuttal 1:
Rebuttal: We appreciate the reviewer for insightful review and suggestions. We noticed that some of the weaknesses can be addressed by the Appendix. We will reorganize the paper for more visibility.
__Weakness 1__
> The black-box probing is one way to generate a prompt. In many cases, users don't know wh... | Summary: This paper introduces ProPILE, a tool to assess the potential leakage of Personally Identifiable Information (PII) in Large Language Models (LLMs). The work is significant given the recent widespread use and growing concerns around privacy issues of LLMs. The authors demonstrated its use with the OPT-1.3B mode... | Rebuttal 1:
Rebuttal: We appreciate the reviewer's thorough review. We are glad that the reviewer acknowledged novelty, significance, concretness, and effectiveness of our work. We hope that this response can clarify any uncertainty.
__Weakness 1__
> The paper acknowledges that the construction of the evaluation dat... | Summary: The authors propose a privacy attack against large, pre-trained language models to evaluate PII leakage. The idea is that a data subject queries the model using some of their PII to discover whether the model can reconstruct the remaining PII not contained in the query. This attack measures the ability of a mo... | Rebuttal 1:
Rebuttal: We thank the reviewer for thoughtful comments and insightful suggestions. Additional study on the suggestion also deepened our understanding and added another dimension to our study. We are also glad that the reviewer recognized the importance, meaningfulness, and clear presentation. We hope that ... | Rebuttal 1:
Rebuttal: To all reviewers and ethics reviewers,
We appreciate all reviewers for acknowledging the importance of our work and providing us with constructive feedback. We hope that this rebuttal addresses any uncertainties and concerns. All responses will be incorporated into the revised paper.
A few revi... | NeurIPS_2023_submissions_huggingface | 2,023 | Summary: This paper proposes a novel probing tool to probe potential privacy leakage in large language models. The proposed method includes both black-box and white-box probing. The authors demonstrate its effectiveness on the OPT-1.3B model with the publicly available Pile dataset. The experimental results show that t... | Rebuttal 1:
Rebuttal: We thank the reviewer of thorough review and insightful comments. We are glad that the reviewer enjoyed reading and acknowledged importance, novelty and presentation of our work. The weaknesses also helped us to deepen our arguments and improve the readability.
We make sure to update the paper wit... | null | null | null | null | null | null |
Neural-Logic Human-Object Interaction Detection | Accept (poster) | Summary: This paper introduces a novel Human-Object Interaction (HOI) detection method called Logic-Guided Transformer Reasoner (LoTR). Unlike standard Transformer-based HOI detectors that accept pre-composed human-object pairs, LoTR leverages Transformers to infer plausible interactions among entities, thus enabling t... | Rebuttal 1:
Rebuttal: **Q: Applicability to visual tasks beyond HOI detection.**
**A:** Thank you for your interest in the generability of LoTR. We firmly believe that the core principles of our logic-guided reasoning framework have the potential to be extended and applied to various computer vision tasks. This adapt... | Summary: The paper introduces a novel approach to tackle the problem of human-object interaction detection. It presents two main contributions. Firstly, there is a revisited self-attention mechanism within the decoder that focuses on the interaction queries. This aspect has received limited attention in the field. Seco... | Rebuttal 1:
Rebuttal: **Q1: Generalize to more existing Transformer-based HOI detectors.**
**A1:** Sorry for this confusion. Our approach involves three distinct elements as input to the interaction decoder $D^p$: human, object, and action. In this context, the action item globally encompasses all possible actions occ... | Summary: This paper introduces LoTR: Logic-Guided Transformer Reasoner for Human-Object Interaction (HOI) detection that adapts the self-attention mechanism in standard Transformers to infer feasible interactions over entities, guided by affordances and proxemics.
Strengths: none
Weaknesses: 1-The meanings of the no... | Rebuttal 1:
Rebuttal: **Q1: Clarify $\textbf{W}^{v'}$ and $\textbf{X}_i$ in Eq.2.**
**A1:** Sorry for this confusion. Here $\textbf{W}^{v'}$ is the weight of a linear layer that adapts the outputs of the attention mechanism. For $\textbf{X}_i$ in Eq.2, it refers to the soft weight assigned to each **pre-constituted** ... | Summary: The paper addresses the problem of human-object interaction (HOI) detection. The authors introduce a new "triplet-reasoning attention" mechanism that enumerates all possible "human-action-object" triplets. Meanwhile, the authors propose a new logic-guided loss function, which enables an interaction detector wi... | Rebuttal 1:
Rebuttal: **Q1: Clarify the dimension of $\textbf{Q}^h$ + $\textbf{Q}^a$.**
**A1:** Sorry for this confusion. We omit a powerful feature named "broadcasting" which is widely used in Pytorch and Numpy.
Broadcasting allows to perform element-wise operations between matrices of different shapes without explic... | Rebuttal 1:
Rebuttal: # To all reviewers
We express our sincere gratitude to all reviewers for their valuable time and thorough assessment of our manuscript. In response, we have meticulously addressed each concern raised, providing point-to-point clarifications which shall be integrated into the new version of our ma... | NeurIPS_2023_submissions_huggingface | 2,023 | Summary: The paper presents LoTR, a method for detecting human-object interactions (HOI) that utilizes Transformer as a reasoning mechanism to model the interactions between people, objects and their interactions. By feeding the self-attention with different feature combination as key query and value, LoTR can infer tr... | Rebuttal 1:
Rebuttal: **Q1: Claims may be too strong and lack sufficient support.**
**A1:** Thank you for pointing this out! The related parts will be revised as follows:
"**First**, the principal focus of them is to optimize the samples with known concepts, ignoring a large number of feasible combinations that were... | null | null | null | null | null | null |
Adaptive Privacy Composition for Accuracy-first Mechanisms | Accept (poster) | Summary: This paper presents extensive background on ex-post-privacy as well as some new analysis providing a new method for leveraging this framework for the design of differentially private interactive protocols. Finally, this framework is instantiated in a method for releasing private counts while targeting an accur... | Rebuttal 1:
Rebuttal: Thank you for the thoughtful review. The goal of the paper is indeed not any particular application to be made precise. Adaptive composition bounds in DP aim to quantify the privacy loss when we make repeated, possibly dependent, private queries (and multiple queries are the norm, nobody would co... | Summary: This paper tries to tackle the following basic question in differential privacy: suppose I desire to answer as many as possible counting queries with low **relative error** with a predetermined privacy budget (\eps,\delta). What is the best mechanism I can use?
Building on prior works on "ex-post" private mec... | Rebuttal 1:
Rebuttal: As stated as a general takeaway from the reviewer feedback, we will include the results on the basic, yet more general, composition for ex-post private mechanisms in the next revision. We presented the composition of (approximate) concentrated DP mechanisms with Brownian mechanisms as we feel tha... | Summary: ‘Accurary-first’ mechanisms try to achieve the strongest possible privacy guarantee that satisfies a target level of accuracy. One such example is a noise reduction mechanism, which scales up the privacy cost / decreases the noise level until the target utility is achieved. Noise reduction mechanisms only pay ... | Rebuttal 1:
Rebuttal: While formally introduced in Definition 4.4., we will certainly add a gentler introduction to ex-post or noise-reduction mechanisms like the Brownian mechanism to the paper. For now, an informal description is that it is a recently devised generalization of the Gaussian mechanism: one first adds a... | Summary: This paper focuses on the study of adaptive compositions of ex-post differential privacy, where privacy guarantees depend on the outcome of the algorithm. The utilization of ex-post DP enables analysts to achieve a more favorable privacy-utility trade-off. The paper introduces a tool for composing ex-post DP a... | Rebuttal 1:
Rebuttal: We do not have a formal lower bound argument for why these composition bounds are tight, but they at least match the “standard” (non-outcome-dependent) privacy loss composition that depends on the variance of the noise that is added. In traditional DP, if we add noise with standard deviation $1/\... | Rebuttal 1:
Rebuttal: We thank all the reviewers for their thoughtful comments and helpful suggestions. Many reviewers pointed out that the main paper should have included the basic, yet more general, composition that included zCDP and ex-post privacy mechanisms, which was instead pushed to the Supplement (where it wa... | NeurIPS_2023_submissions_huggingface | 2,023 | Summary: One of the challenges any attempt to implement DP preserving techniques, is the inherent tradeoff between the level of guaranteed privacy and accuracy levels. While most of the analysis done by privacy experts focused on the perspective of maximizing accuracy subject to an upper bound on the privacy loss, many... | Rebuttal 1:
Rebuttal: We removed the basic, yet more general, composition result from the main body of the paper due to space limitations, but it states that the realized privacy parameters add up. We wanted to highlight the more “advanced” composition bounds that allow for tighter privacy loss bounds but will include... | null | null | null | null | null | null |
Adaptive Data Analysis in a Balanced Adversarial Model | Accept (spotlight) | Summary: The paper considers balanced adversaries that do not have access to the underlying distribution and instead interact with an oracle adversary that chooses the distribution and samples points from it. The authors show that with this more realistic adversarial setting, they can improve the previous known lower b... | Rebuttal 1:
Rebuttal: Thank you for reviewing our paper. In the following we would like to respond to the specific points you make:
Regarding the weakness:
Thank you for the suggestion. We will add a comparison table.
Regarding your question:
We are following Steinke and Ullman [2015a] who used 2000n as a large enoug... | Summary: Adaptive data analysis considers the setting where a mechanism and an analyst play a game against each other, where for $k$ sequential rounds, the analyst issues queries to a mechanism which holds a dataset of size $n$ sampled from a distribution, the mechanism returns a response, and the analyst chooses the n... | Rebuttal 1:
Rebuttal: We thank you for your positive review. In the following we would like to respond to specific points you make.
Regarding the weakness:
In our model, the sampler $A_1$ is publicly known (similar to the Bayesian ADA model of Elder [2016] as we note in Section 1.2). This way, $A_1$ captures knowledg... | Summary: The paper proposes to consider balanced adversaries for ADA. Under the public-key assumptions, the paper proves that
Strengths: 1. The paper assumes a more practical threat model for ADA.
2. The theoretical results in this paper seem to be well-supported. I am not an expert in this research direction, so I ... | Rebuttal 1:
Rebuttal: Thank you for reviewing our paper. We will implement your suggestions to make our paper more broadly accessible.
Statistical queries are basic building blocks in ML and statistics. Our result is about the possibility of accurately answering adaptively chosen statistical queries and hence has cons... | Summary: The paper studies limitations of "adaptive data analysis" with a focus on "computational assumptions" that are used for proving such limitations (i.e., lower bounds).
A mechanism $M$ answers $m=m(n)$ queries (almost) correctly, if it succeeds in the following game for every distribution $D$ and PPT adversary ... | Rebuttal 1:
Rebuttal: We thank you for the comments. In the following we would like to respond to specific points you make.
W1: In our model, $A_1$ is publicly known. It defines a publicly known prior, similarly to the Bayesian ADA model of Elder [2016] (mentioned in Section 1.2). This way, $A_1$ captures the common ... | null | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Counterfactually Fair Representation | Accept (poster) | Summary: The paper presents a new approach for achieving counterfactual fairness. Specifically, builds on past work on counterfactual fairness and proposes an improved algorithm that does not discard any input features. The paper achieves this by sampling from the latent background variables to aggregate a set of coun... | Rebuttal 1:
Rebuttal: Thank you for your comment. Here are our answers to your questions.
Question 1: As we can see in algorithms 1 and 2, $u$ is generated based on conditional distribution U given X=x and A=a. As a result, $u$ is generated differently for each sample. As we discussed in the paper and the appendix, w... | Summary: The paper introduces an innovative training technique that guarantees adherence to the counterfactual fairness measure. The method comprises three distinct steps: counterfactual samples generation, counterfactually fair representation generation, and fair model training.
In contrast to previous approaches tha... | Rebuttal 1:
Rebuttal: Thank you for your comment. Here are our answers to your questions.
Question 1: $Y$ in the law school dataset represents the (normalized) first-year average GPA (FYA). The range of it in the dataset is from –3.45 to 3.48 (We will add more information about the dataset to include the range of $Y$... | Summary: The paper proposes a new algorithm to learn counterfactual fairness representation. The idea behind this is first to generate the counterfactual examples and use the symmetric function to generate the symmetric function. Empirical results models trained with this method can better satisfy counterfactual fairne... | Rebuttal 1:
Rebuttal: Thank you for your comment. Here are our responses to your questions.
Question 1: Our goal in our paper is to have a perfect fair algorithm, so we did not include tunable parameters to adjust between accuracy and fairness like the weights used in the regularization-like method (Note that using a... | Summary: The authors claim to:
* present a novel and efficient method for generating counterfactually fair representations through the use of a symmetric function.
* theoretically show that ML models trained with such representations can achieve perfect/exact CF.
* extend their method to path-dependent counterfactual ... | Rebuttal 1:
Rebuttal: Thank you for your comment. Here are our responses to your questions one by one.
Question 1: We do not have any requirements on the type of features. As you can see, we provided an algorithm that generates a new representation using the original feature and a counterfactual feature with a symmet... | null | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
CoDA: Collaborative Novel Box Discovery and Cross-modal Alignment for Open-vocabulary 3D Object Detection | Accept (poster) | Summary: This paper addresses the challenge of open-vocabulary 3D object detection (OV-3DDet). The proposed framework consists of two modules: the 3D Novel Object Discovery (3D-NOD) module and the Discovery-Driven Cross-Modal Alignment (DCMA) module. The 3D-NOD module utilizes a detector trained on annotations of base ... | Rebuttal 1:
Rebuttal: **Q1 Comparison with existing method [15] and [R1]:**
Regarding the comparison with [15], please refer to Tab. B of the rebuttal PDF. As for the comparison with [R1] (i.e., CLIP$^2$), we conducted comparison experiments on SUN-RGBD. More specifically, we utilize 3DETR to generate pseudo boxes and... | Summary: This paper aims to address the problem of detecting objects from an arbitrary list of categories within a 3D scene, which is less explored area in the literature. It proposes a unified framework that simultaneously addresses the fundamental problems of localizing and classifying novel objects, under the condit... | Rebuttal 1:
Rebuttal: **Q1 Model complexity:**
During testing, our method only adds the CLIP-text encoder and one more distillation head to 3DETR. Therefore, the parameters are just slightly more than 3DETR [19] backbone (6.99M vs. 5.81M), which demonstrates that our designs will not bring many extra computations.
**... | Summary: This paper proposes an end-to-end open-vocabulary 3D object detection (OV-3DDet) framework which can learn to localize and classify novel objects simultaneously. To achieve OV-3Det, the author designed a 3D Novel Object Discover (3D-NOD) strategy which utilizes both the 3D box geometry priors and 2D semantic ... | Rebuttal 1:
Rebuttal: **Q1 Minor grammar issues:**
Thank you very much for your careful review. We will fix the minor issues in the revision.
**Q2 Novelty of DCMA:**
We design the Discovery-Driven Cross-Modal Alignment (DCMA) in conjunction with the 3D-NOD to achieve joint learning of 3D object box discovery and cro... | Summary: This paper is about addressing open-vocabulary 3D object detection, which describes a method aiming for two key challenges, localization and classification for novel objects in point-cloud. The key components are 3D novel-object discover module and the discovery-driven cross-modal alignment module. The first m... | Rebuttal 1:
Rebuttal: **Q1 Novelty of DCMA:**
Our proposed Discovery-Driven Cross-Modal Alignment(DCMA) has two key differences from the contrastive learning in [15]:
(i) Joint Learning with 3D Novel Object Discovery (3D-NOD): DCMA is simultaneously learned with our 3D-NOD, enabling generalization to a broader range ... | Rebuttal 1:
Rebuttal: We thank all the reviewers for their detailly suggestions and encouraging comments: **"The idea makes sense"** and **"is well-motivated"** (Reviewer zhog, Reviewer goao). **"Discard the reliance on any additional 2D detection model"** (Reviewer Qxx5, Reviewer zhog). Bring **"superior performance i... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Bottleneck Structure in Learned Features: Low-Dimension vs Regularity Tradeoff | Accept (poster) | Summary: This work studies the leading order expansions in $L$ of the representation cost for a large $L$, where $L$ is the depth of the model. The work concludes that there is a low-dimension and regularity tradeoff as one varies the depth
Strengths: I think the main strength of the work is the novelty of the message... | Rebuttal 1:
Rebuttal: Thanks for the thoughtful review. Let us answer your remarks/questions:
1. This a good remark and it is true that we did not address it in
details. Actually for functions $f$ where the Jacobian and Bottleneck
rank match $\mathrm{Rank}\_{J}(f;\Omega)=\mathrm{Rank}\_{BN}(f;\Omega)=k$
one can easily... | Summary: This paper looks at the representation cost $R(f)$ of networks as the $L_2$ norm of all of the parameters. The paper claims that for deep neural networks, computing the minimum cost representation of a function is not tractable. Hence, the look at the representation cost normalized by the number of layers $L$ ... | Rebuttal 1:
Rebuttal: Thanks for the thoughtful review.
We plan to modify the structure of the paper for the final version
to better explain how the results are related. We will move the discussion
of the second correction $R^{(2)}$ to right after the discussion
of the first correction $R^{(1)}$. At first we expected ... | Summary: This paper introduces two corrections to previous work that relates the representation cost of infinite-depth nonlinear networks to the bottleneck rank. These two corrections are obtained by Taylor expansion of the representation cost. In addition to the previously-known rank term $R^{(0)}(f)$ that is directly... | Rebuttal 1:
Rebuttal: Thanks for the positive review.
We plan to change the stucture of the paper and talk about the $R^{(2)}$
term earlier. The reason why we discussed $R^{(2)}$ towards the end
of the paper is that it ended up playing little role in the overall analysis. At first,
inspired by the linear network case... | Summary: This paper extends the theoretical framework around the representation cost of DNNs as defined in previous work. The authors introduce two corrections to the existing infinite depth description, namely two regularity measures $R^{(1)}$ and $R^{(2)}$, which balance against the dominating rank bias $R^{(0)}$. Th... | Rebuttal 1:
Rebuttal: Thanks for the thoughtful review. To answer your remarks/questions:
- When we use the term regularity of a function, we mean properties
that imply that the outputs of the function change little if the inputs
are changed a little bit, e.g. continuity, Lipschitzness, or ($k$-times)
differentiabilit... | Rebuttal 1:
Rebuttal: Before we answer to each of the reviewers individually, we describe
here two main changes that we plan to do for the final version, the
first change affects the structure of the paper, while the second
one changes some of the mathematical tools used in the proofs, yielding
stronger Theorems:
1. W... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
FaceDNeRF: Semantics-Driven Face Reconstruction, Prompt Editing and Relighting with Diffusion Models | Accept (poster) | Summary: The paper introduces FDNeRF, a generative method that reconstructs Face NeRFs with semantic editing and relighting capabilities. FDNeRF utilizes 3D GAN inversion and a trained 2D latent-diffusion model to enable zero-shot learning, eliminating the need for explicit 3D data. With designed loss functions and mul... | Rebuttal 1:
Rebuttal: We thank you for your thoughtful reviews and suggestions. We are encouraged that you appreciate the two strengths of our work.
As explained in responses to common concerns, we extended our method to 360 human heads based on PanoHead. We also utilize PTI for the inversion process. The results of F... | Summary: The authors proposes an approach for EG3D-based image editing in this paper.
To be specific, they utilize modified Score Distillation Sampling (SDS) on a pretrained EG3D along with reconstruction and ID loss for semantic editing (e.g., change age, eyeglasses, and hair) and relighting from a single image.
T... | Rebuttal 1:
Rebuttal:
We thank you for your thoughtful reviews and suggestions. We are encouraged that you appreciate the two strengths of our work.
Following your suggestions, we conduct more comparisons and evaluate them quantitatively using ID similarity. The results are given in the pdf file as shown in Fig.1 and... | Summary: This article primarily introduces a method called Face Diffusion NeRF (FDNeRF) for 3D face reconstruction and editing. The key feature is the use of pre-trained 3D GAN and 2D latent diffusion models as priors, allowing users to operate and construct Face NeRF without requiring explicit 3D data. FDNeRF also inc... | Rebuttal 1:
Rebuttal: We thank you for your thoughtful reviews and suggestions.
For your concerns about editing beyond the photorealistic human face domain, our diffusion-guided NeRF generator latent vector optimization pipeline can generalize to more domains by replacing EG3D with other NeRF generators. Fig.2 demonst... | Summary: This work proposes a 3D generative model-based optimization pipeline for 3D editing from a single image which is a useful and challenging task. The proposed pipeline achieves text and lighting-driven editing by reusing several existing pre-trained models. By carefully designing the losses trade-off, it achieve... | Rebuttal 1:
Rebuttal: We thank you for your thoughtful reviews and suggestions. We are encouraged that you appreciate the three strengths of our work. Following your suggestions, we compared our work with InferfaceGAN and StyleClip quantitatively and qualitatively in Fig.1, Fig.3, Tab.1, and Tab.2. To show the generali... | Rebuttal 1:
Rebuttal: We thank all reviewers for their insightful comments and appreciation of our work. Below we address the issues. The referred figures and tables are in a separate PDF file. Following the given review sequence, we use R1, R2, R3, R4 to denote reviewer JGgm, M2ze, 1wAz, 4TUt respectively.
### [R1... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Hodge-Aware Learning on Simplicial Complexes | Reject | Summary: In this paper, the authors first use Dirichlet energy minimizations on simplicial complexes (SCs) to interpret their effects on mitigating the simplicial oversmoothing. Then, through the lens of spectral simplicial theory, they show the three principles promote the Hodge-aware learning of this architecture, in... | Rebuttal 1:
Rebuttal: We appreciate the reviewer's feedback. Please find below our replies.
> @ #1: applying SCCNN for node/graph classifications, comparing with sota GNN
We understand your interest in how SCCNN performs in node/graph classifications.
These are important problems but do not fit the goal of this paper.... | Summary: The authors identified the limitations of non-Hodge aware learners on simplicial complex (SC) data and proposed a convolutional structure that 1) decomposes the upper and lower k-Laplacian and 2) takes the inter-simplicial couplings into account. The paper has presented a justification for the performance by a... | Rebuttal 1:
Rebuttal: We appreciate the reviewer's careful and critical feedback. Here we address the detailed concerns.
> @ Weakness #1: about complexity of $n_k$
It is not practical to use SCCNN under this scenario due to the high complexity of $(k+1)$-cliques ($k$-simplices), $n_{k}=\mathcal{O}(n^{k+1})$. This also... | Summary: This paper aims to propose a convolutional architecture which incorporates the Hodge theory. Specifically, the proposed architecture incorporates the following three properties: uncoupling the lower and upper simplicial adjacencies, accounting for the inter-simplicial couplings, and performing higher-order con... | Rebuttal 1:
Rebuttal: We thank the reviewer for reviewing our manuscript and point out some textual and grammar inconsistencies in it.
> @ "Errors:..." in the Weakness
Our apologies from the grammar errors, which were due to continuously shortening and updating the manuscript till minutes before the deadline. We have... | Summary: This paper introduces a novel architecture designed to operate on simplicial data, drawing its foundation from the Hodge decomposition. This decomposition ensures that features associated with a simplicial complex of order $k$ can be represented by three distinct quantities: a curl-free quantity, a divergence-... | Rebuttal 1:
Rebuttal: We really appreciate the concise summary of the paper from the reviewer!
> @ Weakness #1: about why we included stability analysis and further exposition of Thms 6 and 7.
We appreciate your thorough suggestions. Following your suggestion, we will focus the results on the main points (comparisons... | Rebuttal 1:
Rebuttal: In this global response, we first express our gratitude to the reviewers for providing us with their insightful and critical feedback.
The positive feedback from the reviewers acknowledges the strengths of our paper in the following areas:
- Studying an important problem: Our paper delves into the... | NeurIPS_2023_submissions_huggingface | 2,023 | Summary: The authors propose a general convolutional architecture with principles of uncoupling the lower and upper simplicial adjacencies, which accounting for the inter-simplicial couplings, and performing higher-order convolutions for learning of higher-order structure and simplicial signal.
Strengths: - The author... | Rebuttal 1:
Rebuttal: We appreciate reviewer's comments on the strengths of our paper. Here we address your four detailed questions.
> @ #1 about the orientation
The orientation of a simplex is set as the lexicographical ordering of its vertices (line 79), as commonly done in simplicial literature [15, 16, 17, 27, 4... | null | null | null | null | null | null |
Networks are Slacking Off: Understanding Generalization Problem in Image Deraining | Accept (poster) | Summary: This paper explores the generalization problem of image deraining. The authors claim that deep networks display a tendency to learn the less complex element in the task of separating image content from additive degradation. Overall, this paper contributes some new ideas and new insight to the community, but th... | Rebuttal 1:
Rebuttal: We thank reviewer FkGF (R#4) for his review. We also thank R#4 for acknowledging our conclusions and contributions. Our response to your concerns is as follows:
`In Weakness 1): "I am curious whether the latest Transformer-based image deraining method still has the same conclusion."`: Thank you f... | Summary: This paper proposed to explore the problem of generalization problem in image deraining. The authors find that deep networks tend to overfit the degradation patterns present in the training set, which leads to poor generalization performance for unseen degradation patterns.
Strengths: 1. Writing fairly well.
... | Rebuttal 1:
Rebuttal: We thank Reviewer w1Er (R#3) for his\her efforts. We also thank R#3 for finding our conclusions interesting and novel. Our response to your concerns is as follows:
`In Weakness 1): "Why did the author create a new dataset without conducting experiments on previous existing benchmarks."`:
We appre... | Summary: This paper is working on understanding generalization problem in image deraining. Several interesting findings are: 1) training with fewer background images leads to better deraining effect; 2) the relative complexity between the background and rain determines the network behavior; 3) a more complex background... | Rebuttal 1:
Rebuttal: We thank Reviewer e5xd (R#2) for his\her efforts. Thank you for finding our conclusions interesting. Our response to your concerns is as follows:
`In weakness 1) "the overall contribution is limited, i.e., authors didn't provide a good solution to solve the problem." and "I would expect more cont... | Summary: This paper makes observations on the generalization problem of deraining networks by conducting extensive experiments, on the perspective of training data. This paper claims to find out some underexplored insights that less training data with less complex background can lead to better rain removal ability at t... | Rebuttal 1:
Rebuttal: We thank Reviewer UxXt (R#1) for his\her efforts. We next responded to R#1's comments. To save space, we set some abbreviations: generalization (GNL), background (BG).
R#1's resistance to data reduction likely arises from a misinterpretation of our intentions. Through experiments, we aim to chall... | Rebuttal 1:
Rebuttal: Reply to all reviewers and area chairs for short abstracts
Dear Reviewers and Area Chairs,
We thank all reviewers and regional chairs for their valuable time. We sincerely appreciate all comments. They gave positive, high-quality comments to our paper and provided a lot of constructive feedback.... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
CADet: Fully Self-Supervised Out-Of-Distribution Detection With Contrastive Learning | Accept (poster) | Summary: This paper consists of two mostly unrelated parts. In the first part, the authors aim to detect whether two sets of samples are drawn from the same distribution. To do this, they apply the unbiased estimate of the MMD, in the manner suggested by Sutherland et al., to produce a conformal p-value. Then they p... | Rebuttal 1:
Rebuttal: 1a. We consider that works that tuned their methods on calibration OOD datasets and evaluated on different test benchmarks are still tuning on OOD samples since these test and calibration benchmarks typically involve similar types of distribution shifts, applied from similar inlier distributions. ... | Summary: This paper studies the OOD detection of new classes and adversarial attacks with learned similarities via self-supervised contrastive learning in conjunction with MMD two-sample test. To enable MMD applicable for OOD detection on single samples, they improve the idea by using augmentations to create a set. Exp... | Rebuttal 1:
Rebuttal: We thank the reviewer for their time and feedback.
1- As mentioned L195-L198, CADet is designed to leverage the similarity function of self-supervised contrastive learning (SSCL). Indeed, the CADet score is computed based on the cosine similarity between the representations of random transforms o... | Summary: The authors propose to handle the OOD detection problem, exploring self-supervised contrastive learning to detect samples from previously unseen classes and adversarially perturbed samples. They use self-supervised contrastive learning with the maximum mean discrepancy (MMD) to test if two sets of samples orig... | Rebuttal 1:
Rebuttal: We thank the reviewer for their time and feedback.
**Weaknesses:**
1- The reviewer is correct that we only evaluate on homogenous benchmark. Evaluating in a realistic setting is extremely difficult, as even data collection can pose legal challenges. Nonetheless, the general motivation for trying... | Summary: The paper proposes Maximum Mean Discrepancy with Clean Calibration (MMD-CC), as an improvement of MMD when the number of samples is small. Moreover, it introduces CADet, a novel anomaly detector for both OOD and adversarial detection inspired by MMD.
Strengths: In line with recent works, unsupervised anomaly... | Rebuttal 1:
Rebuttal: We thank the reviewer for their time and feedback.
1- In [2], the Mahalanobis detector uses a score based on a logistic regression of layer-wise Mahalanobis distances. The regression weights are trained on validation outliers. Training directly a classifier on outliers has long been discarded by ... | null | NeurIPS_2023_submissions_huggingface | 2,023 | Summary: The paper presents a novel method for Out-of-Distribution (OOD) detection called CADet (Contrastive Anomaly Detection). The authors leverage self-supervised contrastive learning and the maximum mean discrepancy two-sample test(MMD) to assess whether two sets of samples have been drawn from the same distributio... | Rebuttal 1:
Rebuttal: We thank the reviewer for their time and feedback.
**Q1:** It is a very pertinent question and recent works such as "ViM: Out-Of-Distribution with Virtual-logit Matching" by Wang et al. shows more recent architectures like ViT have distinct and promising features when it comes to OOD detection. ... | null | null | null | null | null | null |
Fractal Landscapes in Policy Optimization | Accept (poster) | Summary: This paper studies the dynamics of policy gradient training in
reinforcement learning, and argues that the policy gradient often does
not exist due to the lack of regularity of the training objective. The
authors derive a threshold on the "maximum Lypanunov exponent", a
quantity they introduce which captures t... | Rebuttal 1:
Rebuttal: Thank you for the detailed comments and suggestions. Please see the answers below. We should clarify that the results are not meant to be negative (and will revise some of the statements) but to illustrate a source of difficulty in policy optimization that has not been discussed much in existing w... | Summary: The paper points out that the value function of some MDPs is not differentiable or even fractal in the infinite horizon setting.
This raises the question of the well-formedness of policy gradient methods.
A method for statistically estimating the Hölder exponent is presented to find out whether the training ... | Rebuttal 1:
Rebuttal: Thank you very much! We are delighted that you are interested in our paper!
The question you raised is very insightful, and we would like to share some opinions about it.
In Example 4.3, we showed that if no additional assumptions are added, using stochastic policies may lead to a more non-smoot... | Summary: This paper proposes a method to understand the optimization landscape of the policy gradient method in the search for a non-smooth or fractal landscape. The method is based on the chaos theory and non-smooth analysis of the optimization objective. In particular, the paper proposed a method to estimate the Höld... | Rebuttal 1:
Rebuttal: Thank you for your positive feedback. We're glad to hear that you liked our paper!
$\textbf{The conclusion in Section 4.3}$: We claimed that unless further assumptions are made, the worst case of stochastic policies is worse than that of deterministic policies, as shown in Example 4.3. Therefore... | Summary: The paper tries to argue that in a certain class of continuous state-action MDPs, the policy gradient landscape can be non-smooth. It does so by showing that the Holder exponent of the policy objective $J(\theta)$ is less than one, when the MDP's maximum Lyapunov exponent is greater than $-\log \gamma$. The pa... | Rebuttal 1:
Rebuttal: We appreciate the reviewer for the detailed comments and questions. We focus on addressing the major issues, and will further improve the writing based on the list of minor issues.
1. The main result in this paper is that given a policy parameter $\theta$, the objective function $J(\theta)$ is $... | null | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Adversarial Self-Training Improves Robustness and Generalization for Gradual Domain Adaptation | Accept (poster) | Summary: This paper investigates the adversarial gradual domain adaptation. The authors apply adversarial learning in the self-training approaches in gradual domain adaptation, forming the adversarial self-training approach (AST). Empirically, the authors show that AST improves both adversarial accuracy and clean accur... | Rebuttal 1:
Rebuttal: We thank the reviewer for their constructive comments and positive assessment. We are glad that the empirical improvement and theoretical analysis are appealing to you.
**Q1: In the Introduction, the contributions could be better summarized.**
**A1:** Thank you for this important suggestion. The... | Summary: The paper proposes an Adversarial Self-Training method to improve the robustness and clean accuracy under the scenario of gradual domain adaptation. More specifically, the paper uses networks to generate pesudo label for gradually changing domain and update the model with the pesudo label. Experiments on Rotat... | Rebuttal 1:
Rebuttal: Thank you for your constructive feedback on our work.
**Q1: The proposed method, AST, seems to be an application of [1,2] under the scenario of gradual domain adaptation (GDA), which lacks of novelty.**
**A1:** Thank you for raising the concern. As you mentioned, some AST similar methods were pr... | Summary: This paper proposes adversarial self-training (AST) that incorporates adversarial training with gradual self-training. Empirical results validate that AST can enhance both adversarial robustness and generalization ability in the scenario of the gradual unsupervised domain adaption. The authors provide empirica... | Rebuttal 1:
Rebuttal: We thank the reviewer for their constructive comments and positive assessment. We are glad that you find the empirical observation interesting and surprising.
**Q1: What if the results were evaluated on a more complex network say ResNet18?**
**A1:** Thank you for raising the concern. As per your... | Summary: In this paper, the authors analyze the adversarial robustness of the gradual self-training method with adversarial self-training (AST). AST first predicts labels on the unlabeled data and then adversarially trains the model on the pseudo-labeled distribution. The authors introduce an adversarial self-training ... | Rebuttal 1:
Rebuttal: We thank the reviewer for their detailed and constructive comments.
**Q1: They should provide a more detailed discussion of the implications of the milder assumptions.**
**A1:** Thank you for your constructive suggestion. In the work of [A], the authors assume that the new model can always fit t... | Rebuttal 1:
Rebuttal: ## Common Response (Additional Experiments)
We thank all the reviewers for their detailed and constructive comments. As per suggestions from the reviewers, we include some additional experiments here. **The results can be found in the attached PDF file**.
* **Larger number of intermediate domains... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Where2Explore: Few-shot Affordance Learning for Unseen Novel Categories of Articulated Objects | Accept (poster) | Summary: This paper first proposes a new task, cross-category few-shot learning for articulated object manipulation. And it further proposes a framework that first explictly measures the semantic similarity of local geometries, interacts with the objects of novel categories according to the similarity, uses the intera... | Rebuttal 1:
Rebuttal: We sincerely thank you for your efforts and your thoughtful feedback. Thank you for the valuable questions and suggestions!
Here, we will address the questions and comments in the same order as they appear in the reviews. Further questions are welcomed!
> Statistics, eg. std, are not provided in... | Summary: The paper focuses on the task of predicting affordances given a single-frame observation of an articulated object. The authors utilize few-shot learning to address the novel object adaption problem. However, previous approaches to few-shot learning require per-instance interaction during test time, which can b... | Rebuttal 1:
Rebuttal: We sincerely thank you for your efforts and your thoughtful feedback. Thank you for the valuable questions and suggestions!
Here, we will address the questions and comments in the same order as they appear in the reviews.
> There are not enough descriptions of the hard-coded intersection sets. Th... | Summary: This paper studies the problem of interacting with novel articulated objects. More specifically, a few-shot learning based approach is proposed to adapt to unseen instances at test time. The methodology explicitly leverages similar local geometries, which may be shared across objects from distinct categories. ... | Rebuttal 1:
Rebuttal: We sincerely thank you for your efforts and your feedback. Thank you for the valuable questions!
Here, we will address the questions and comments in the same order as they appear in the reviews.
> The robot arm is abstracted away. The input point cloud is assumed to be cleanly segmented out.
We ... | Summary: The paper focuses on few-shot affordance learning and articulated object manipulation. It introduces a framework named Where2Explore, which consists of the affordance module and the similarity module. Both modules share the same architecture, a Where2Act [24]-like network based on PointNet++ [28]. While the af... | Rebuttal 1:
Rebuttal: We sincerely thank you for your efforts and your thoughtful feedback. Thank you for the valuable questions and suggestions!
Here, we will address the questions and comments in the same order as they appear in the reviews. Please feel free to follow up if you have more questions.
> The method is... | Rebuttal 1:
Rebuttal: We thank you all for your efforts and insightful comments! We thank reviewers for appreciating our motivation: "The methodology is well motivated", "The motivation to explore more efficient few-shot learning is intriguing", and design: "The concept of leveraging affordance similarities across cate... | NeurIPS_2023_submissions_huggingface | 2,023 | Summary: The paper explores a few-shot learning strategy to predict the affordance of objects in novel categories. The key idea of the paper is to use a similarity module to transfer the knowledge from known categories to novel categories. Then the affordance prediction on the novel object can be learned with just a fe... | Rebuttal 1:
Rebuttal: We sincerely thank you for your time and your positive feedback. Thank you for the valuable questions! Here, we will address the raised questions and comments in the same order as they appear in the reviews. Please Feel free to follow up if you have more questions.
> is the similarity module exa... | null | null | null | null | null | null |
An Optimal and Scalable Matrix Mechanism for Noisy Marginals under Convex Loss Functions | Accept (spotlight) | Summary: The paper studies a timely problem of improving the scalability of differentially private query releases using a matrix mechanism. The central approach is to carefully design a set of the residual basis for the representation of workload queries, such that the workload queries could be represented as Kronecker... | Rebuttal 1:
Rebuttal: **The optimality statement Theorem 4.4 puzzles me, as it seems only about an optimal number of noisy answers. In that sense, all previous matrix mechanism-based solutions are also optimal. (As they answer the same number of noisy queries as ResidualPlanner in this paper.) In that sense, I wonder w... | Summary: INTRODUCTION:
This paper presents significant advancements in the matrix mechanism, a technique used for releasing differentially private linear queries or count statistics. Releasing these types of queries, for instance, “count of people living in NYC who are male” (also known as a 2-way marginal query), is a... | Rebuttal 1:
Rebuttal: **Thus, the paper could greatly benefit from providing more intuitive explanations of their techniques. For instance, the definition and context of the Kronecker product are noticeably absent, which detracts from the clarity of the relevant sections.**
Sorry for the omission, we will add that. In... | Summary: The paper studies scalability of matrix mechanism when answering marginal queries. The paper looks at a restricted yet large enough class which they call regular loss function, satisfying three key properties. Scalability of the matrix mechanism is often poor as it requires heavy SVD computation to compute the... | Rebuttal 1:
Rebuttal: **This construction only works in full-rank structure, so they perform a Gaussian elimination to get linear independence.**
This is a confusion due to our writing. We propose a residual basis (formed by all possible R matrices as defined in Section 4.2). This basis is linearly independent, but n... | Summary: This work presents a scalable and optimal, under some conditions, matrix mechanism, which can compute private marginals. Releasing private marginals is one of the most fundamental problems in private data analytics, and as mentioned in the paper, an important primitive in synthetic data generation. This work... | Rebuttal 1:
Rebuttal:
**Why the general linear Gaussian Mechanism is necessary and why not just use an identity covariance and matrix B?**
Great question. In theory, it is the same, but as a practical matter, if the covariance is the identity, then B will have irrational entries (a general formula for B even seems tr... | Rebuttal 1:
Rebuttal: We thank the reviewers for carefully reading the paper and the insightful comments. The questions pointed out where we need additional clarification and explanation. As suggested, these can be added by moving Table 3 to the supplementary material.
Since we have enough space in the comments for eac... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Causal normalizing flows: from theory to practice | Accept (oral) | Summary: The authors propose a normalizing flow (NF) model that incorporates causal information, either via a partial order or complete DAG. They derive identifiability conditions for causal estimands and implement their algorithm on a range of simulated and real-world datasets. The resulting NF can be used to compute ... | Rebuttal 1:
Rebuttal: We thank the reviewer for their valuable comments and for the supportive words. Please see our comments to the reviewer’s comments below:
> I’m unsure how to interpret $\sum_{n=1}^\infty G^n$ . Why the infinite sum? The examples in Fig. 4 all appear finite?
The short and honest answer is that th... | Summary: n this paper, the authors derive and demonstrate the usefulness of normalizing flows (NF) in causal inference. As a first building block of their approach, the authors show that structural causal models (SCMs) can be expressed as triangular monotonic increasing maps. By this reduction, they show that SCMs can... | Rebuttal 1:
Rebuttal: We thank the reviewer for their valuable feedback, detailed summary, and positive comments. Please see our comments regarding the raised weaknesses below:
> The paper would benefit of an example of how a learned causal NF behaves, how the interventions influence the output, and how it matches an ... | Summary: This paper proposes causal normalizing flows, a method to learn a structural causal model (SCM) using normalizing flows (NFs), when the causal order (and potentially more information about the graph) is available. They consider a general SCM where we have observed variables X which follow the equations X_i = f... | Rebuttal 1:
Rebuttal: We thank the reviewer for their valuable comments, as well as the highly detailed summary of our work. Before commenting on the concerns raised by the reviewer, we would like to clarify a few points that may have been overlooked or misunderstood based on the review’s summary:
> [Identifiable] mea... | Summary: This work explores the use of normalizing flows for causal reasoning. The authors demonstrate that causal models can be identified from observational data using autoregressive normalizing flows. They discuss design choices, learning strategies, and the implementation of the do-operator to handle interventional... | Rebuttal 1:
Rebuttal: We thank the reviewer for their positive comments, and we are flattered to know that the reviewer had a good time reading our work. While there are no specific questions, we would like to add a few clarifications below that we hope helps the reviewer better contextualize our work:
> From my under... | Rebuttal 1:
Rebuttal: We thank all the reviewers for their useful comments, which will ultimately help us improve the manuscript. We also want to thank them for their reassuring words towards our work. To name just a few, reviewers acknowledged our effort towards good writing and readability:
> I thoroughly enjoyed rea... | NeurIPS_2023_submissions_huggingface | 2,023 | Summary: This work explores the use of normalizing flows (NFs) in causal inference. The authors demonstrate that causal models can be identified from observational data using autoregressive NFs. They investigate design choices and implement the do-operator in causal NFs to answer interventional and counterfactual quest... | Rebuttal 1:
Rebuttal:
We thank the reviewer for their valuable feedback, which points out details that we will clarify in the updated manuscript. Please see our responses addressing the specific concerns below:
> The number and characteristics of the given dataset are limited when compared to practical applications. ... | null | null | null | null | null | null |
Lossy Image Compression with Conditional Diffusion Models | Accept (poster) | Summary: In this paper, the authors propose a novel architecture of lossy image compression using conditional diffusion models. They use an encoder to map images into contextual latent variables and train a conditional diffusion model under the guidance of these contextual latent variables as the decoder. Moreover, th... | Rebuttal 1:
Rebuttal: > Some implementation details are not illustrated clearly
We have supplied the decoding time of the model and architecture details in Appendices A and C. Additionally, Appendix G contained visualizations of image transitions during the decoding process. We chose 17 as the number of steps to strik... | Summary: This paper presents a method of using conditional diffusion as a stochastic decoder for neural image compression. The approach learns a discrete latent representation of an image and conditions the diffusion process on this representation. Compression is achieved by learning a distribution over the latent repr... | Rebuttal 1:
Rebuttal: > not novel enough
Please see general response 1
> The authors should cite this paper on the same topic
Thank you for pointing this out; we also noticed these important related work and will cite and discuss it in the final paper version. Note that some works are concurrent and appeared after t... | Summary: This paper proposes a diffusion-based image codec. It has two major components: a VAE-based encoder and a diffusion-based decoder. The VAE-based encoder generates the image latents, which serve as the conditioning signal for the diffusion-based image reconstruction. Both components are trained end-to-end with ... | Rebuttal 1:
Rebuttal: > It is rarely seen that a diffusion model needs to be trained with a perceptual loss.. More insights need to be provided
Please see our general response 2
> It is unclear whether one could use more decoding steps without the perceptual loss to achieve the same perceptual quality as with the per... | Summary: In this paper, a conditioned diffusion model is proposed to improve the perceptual quality of learned image compression. Different from widely used VAE-like nonlinear transform coding frameworks, the proposed model adopts a diffusion process to synthesize most of the texture, conditioned on a learned hyperprio... | Rebuttal 1:
Rebuttal: > The approach seems somewhat empirical... provide some more theoretical analysis to convince.
Neural image compression is largely an empirical field, so our paper does not stand out in this regard. Yet, we made sure our optimization objective follows from a rigorous lower bound on a latent varia... | Rebuttal 1:
Rebuttal: # General Response:
We thank all reviewers for their valuable feedback. Before we address them individually, we would like to address common questions in their reviews.
> 1. Motivation, novelty, and advantages of our work
Our diffusion model provides a new approach towards “perceptual compress... | NeurIPS_2023_submissions_huggingface | 2,023 | Summary: This paper proposes an end-to-end learned image compression method based on diffusion model, where it consists of a VAE for latent feature encoding and a conditional diffusion decoder. Concerning theoretical parts, the authors also derive the variational upper bound to the diffusion model’s negative data log l... | Rebuttal 1:
Rebuttal: > The method itself is reasonable but without a clear motivation.. What’s the key advantages
Please see our general response 1 above.
> Although with X-parameterization..decoding time is still significantly longer than existing LIC methods.
Our aim is to investigate the potential of diffusion ... | null | null | null | null | null | null |
RADAR: Robust AI-Text Detection via Adversarial Learning | Accept (poster) | Summary: The paper proposes a framework called RADAR for training a robust AI-text detector using adversarial learning. The framework consists of a paraphraser and a detector, both based on large language models (LLMs). The paraphraser aims to generate realistic contents to evade AI-text detection, while the detector a... | Rebuttal 1:
Rebuttal: We thank the reviewer for the time and expertise you have invested in these reviews. We are delighted to receive the positive feedback that our method is simple but effective.
---
* **Comment 1**: The generation quality after post-processed by the AI paraphraser is my major concern. Since the ... | Summary: The paper introduces a new framework called RADAR, which aims to address the gap in existing AI-text detection methods. RADAR utilizes adversarial learning to jointly train a robust AI-text detector. The framework consists of a paraphraser and a detector, both trained through adversarial training. The paraphra... | Rebuttal 1:
Rebuttal: We thank the reviewer for the time and expertise you have invested in these reviews. We are delighted to receive the positive feedback that our research problem is both interesting and important, the proposed method is reasonable, and the experimental results are promising.
Below we provide poi... | Summary: This paper proposes an adversarial learning framework, called RADAR,
for improving the robustness of AI-text detector. Given a target
large language model (LLM) that generates AI-text, the framework
trains a detector model and a paraphraser model both are also based on
an LLM simultaneously. The performance ... | Rebuttal 1:
Rebuttal: We thank the reviewer for the time and expertise you have invested in these reviews. We are grateful to receive your appraisal that our work provides a straightforward but effective way of training a robust AI-text detector. This is indeed our original aspiration.
We reply to individual points... | Summary: This paper proposes RADAR, a detection method for AI-generated text that is based on adversarial learning. Inspired by the conceptual idea of generative adversarial networks, RADAR consists of a paraphraser and a detector module. The former aims to generate textual content that evades an AI content detection s... | Rebuttal 1:
Rebuttal: We thank the reviewer for the time and expertise you have invested in these reviews. We are delighted to receive the positive feedback that our work provides a solid contribution to the field, especially that the paper is well-written, the technical aspects are well-explained, the analysis is thor... | Rebuttal 1:
Rebuttal: ## Rebuttal References
[R1] PyTorch actor-critic example. (we are not allowed to post any link, so please google it to find the link)
[R2] PyTorch REINFORCE example. (we are not allowed to post any link, so please google it to find the link)
[R3] Tom Hosking and Mirella Lapata. Factorising Mean... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Learning Large-scale Neural Fields via Context Pruned Meta-Learning | Accept (poster) | Summary: This paper proposes a way to sample the subset of the dataset to improve convergence speed and the eventual accuracy in the meta-learning framework, especially for neural fields. They also propose a bootstrap correction to resolve the well-known myopia issue. Moreover, gradient re-scaling during meta-testing a... | Rebuttal 1:
Rebuttal: Dear reviewer 9F3K,
We sincerely appreciate your efforts and insightful comments to improve our manuscript. We respond to each of your comments one-by-one in what follows.
---
**[W1] Comparison with other data pruning. e.g., loss-based top-K selection.**
First, we carefully remind the reviewer... | Summary: This paper presents an efficient optimization-based meta-learning technique for large-scale neural field training. It introduces automated online context point selection, resulting in significant memory savings and improved model quality. The authors also propose a bootstrap correction to enhance the meta-lear... | Rebuttal 1:
Rebuttal: Dear reviewer 3CuA,
We sincerely appreciate your efforts and insightful comments to improve the manuscript. We respond to each of your comments one-by-one in what follows.
---
**[W1] Clarification on the motivation for using optimization-based meta-learning.**
The motivation for using optimi... | Summary: This paper addresses the issues of high memory usage and myopia problem in optimization-based meta-learning for large-scale neural field training. Three techniques are proposed: 1) online context pruning improves memory efficiency by using a selected context set in the inner learning problem, 2) bootstrap corr... | Rebuttal 1:
Rebuttal: Dear reviewer EMYi,
We sincerely appreciate your efforts and insightful comments to improve the manuscript. We respond to each of your comments one-by-one in what follows.
---
**[W1] GradNCP for non-linear last layer architectures.**
We follow the setup of common design choice for NFs [1,2,3... | Summary: The paper proposes a set of new techniques for optimizing neural fields with meta-learning. The new techniques includes using gradient norms to prune less important examples, bootstrapping to correct for myopia of the inner loop optimization, and gradient rescaling to correct for different context sizes at tra... | Rebuttal 1:
Rebuttal: Dear reviewer hvG5,
We sincerely appreciate your efforts and comments to improve the manuscript. We respond to your comment in what follows.
---
**[S2] GradNCP could be beneficial for meta-learning applications beyond NFs.**
Thank you for your insightful comment! We primarily focused on resolv... | Rebuttal 1:
Rebuttal: Dear reviewers and AC,
We sincerely appreciate your valuable time and effort spent reviewing our manuscript.
As reviewers highlighted, we believe our paper tackles an interesting and important problem (EMYi, 3CuA), and provide an efficient yet effective (all reviewers) framework for meta-learnin... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Neural Oscillators are Universal | Accept (poster) | Summary: Various machine learning architectures inspired by oscillators perform well in a variety of tasks. The authors study the approximation ability of neural networks based on oscillators. The main contribution of this paper is the construction of a neural oscillator framework combining nonlinear dynamics and a lin... | Rebuttal 1:
Rebuttal: We start by thanking the reviewer for your appreciation of the merits of our paper and your welcome suggestions to improve it. Below, we address the concerns raised by the reviewer and thank the reviewer in advance for their patience in reading our detailed reply.
1. Our main motivation for stud... | Summary: This paper introduces an abstract framework of neural oscillators and proves the universal approximation theorem of multi-layer neural oscillators, which shows that neural oscillators with a specific structure have the capability of approximating causal and continuous maps between spaces of continuous function... | Rebuttal 1:
Rebuttal: We start by thanking the reviewer for reading our paper and for your comments. Below, we address the concerns raised by the reviewer and thank the reviewer in advance for their patience in reading our detailed reply.
1. The main contribution of this paper is fundamentally an approximation theoret... | Summary: The authors present a universal approximation theorem for a system of 2nd order ODEs they call "neural oscillators." At the heart of the construction is the fact that a forced harmonic oscillator with input computes a windowed sine transform of the input, up to a constant (Lemma 3.4). The other major piece (Le... | Rebuttal 1:
Rebuttal: We start by thanking the reviewer for your appreciation of the merits of our paper and your welcome suggestions to improve it. Below, we address the concerns raised by the reviewer and thank the reviewer in advance for reading our detailed reply.
1. Regarding the reviewer's question about more de... | Summary: This work presents a comprehensive review of diverse learning architectures based on ordinary differential equations (ODEs) and explores their relationship with coupled oscillator systems. Furthermore, the paper provides a proof showcasing the universality of the coupled oscillator architecture. The universali... | Rebuttal 1:
Rebuttal: We start by thanking the reviewer for your appreciation of the merits of our paper and your welcome suggestions to improve it. Below, we address the concerns raised by the reviewer and thank the reviewer in advance for their patience in reading our detailed reply.
1.Regarding the reviewer's conc... | Rebuttal 1:
Rebuttal: At the outset, we would like to thank all four reviewers for their thorough and patient reading of our article. Their criticism and constructive suggestions will enable us to improve the quality of our article. All the changes outlined below will be incorporated in a camera ready version (CRV) of ... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
CAT-Walk: Inductive Hypergraph Learning via Set Walks | Accept (poster) | Summary: The paper proposes an hyperedge learning method suitable for modelling temporal hyperedges. The approach consists of a novel hyperedge-centric random walk, a permutation-invariant pooling method and a technique for anonymizing both the identity of the nodes and the identity of the hyperedges.
Different than... | Rebuttal 1:
Rebuttal: Thank you so much for your review! We really appreciate it! Please see below for our response to your comments:
>*what is the connection between the SetWalk method and a classical graph-based random walk applied on the line expansion?*
----
**Response:** Thank you for mentioning it. We have discu... | Summary: The paper proposes an inductive approach to temporal hypergraph learning. The method includes two main components: a hypergraph random walk called SET-WALK and a pooling strategy termed SETMIXER, for extracting higher-order causal patterns. In addition, the authors design a simple neural network for hyperedge ... | Rebuttal 1:
Rebuttal: Thank you so much for your review! We really appreciate it! Please see below for our response to your comments:
----
> *The method contains several components ... It would be more convincing if each component could be evaluated in all datasets and all settings in Table 1.*
----
**Response:** Th... | Summary: The paper proposes a walk-based learning method for temporal hyperedge prediction. Several conceptional contributions are made to walks in (temporal) hypergraphs which are compared to existing techniques such as clique or star expansion to represent hypergraphs by simple (weighted) graphs. A corresponding neur... | Rebuttal 1:
Rebuttal: Thank you so much for your review! We really appreciate it! Please see below for our response to your comments:
----
>*While the conceptional contributions are nicely motivated, this is not the case to the same extent for the design choices ...*
----
**Response:** Thank you for mentioning it. Ye... | Summary: The paper introduces a novel way of defining random walks on hypergraph. The random walks consist in a sequence of hyperedges, where the transition probabilities take into account both the temporal closeness of the hyperedges and the importance of the nodes through which the hyperedges are connected.
The resu... | Rebuttal 1:
Rebuttal: Thank you so much for your review! We really appreciate it! Please see below for our response to your comments:
----
> *The theorems 1 and 2 are stated in an informal and imprecise way. While I understand the need not to flood the main text with mathematical details, one would have expected to fi... | Rebuttal 1:
Rebuttal: We want to thank all the reviewers for their time and comments. We are also grateful for all the constructive feedback for improving the paper.
We have attached a file with two tables. The first table reports the results of previous ablation studies and the two newly asked ablation studies (high... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Adaptive Quasi-Newton and Anderson Acceleration Framework with Explicit Global Convergence Rates | Reject | Summary: This paper presents first-order iterative methods for solving unconstrained minimization problem. The connection between the proposed methods, quasi-Newton and Anderson accelaration methods are illustrated, which gives an insight of the motivation. Under certain assumptions, the methods exhibit explicit gobal ... | Rebuttal 1:
Rebuttal: # Author's response
We appreciate the reviewer’s positive feedback. We will do our best to answer the reviewer’s concerns.
## Weaknesses
> **Too many assumptions and requirements in the theoretical part. Can those requirements be verified?**
See **SDP, A)** for more details. Those requirements... | Summary: The paper proposes a generic framework for developing new iterative schemes for smooth optimization in the deterministic setting. The derived new methods show some similarity to quasi-Newton methods and Anderson acceleration, while using a backtracking line search for estimating the Lipschitz constant and havi... | Rebuttal 1:
Rebuttal: # Author's response
We thank the reviewer for their thoughtful evaluation of our paper. While we genuinely appreciate their feedback, we must respectfully address some concerns regarding inaccurate statements made in the review.
### A) Scalability for large-scale problems.
> **The algorithm mi... | Summary: The authors introduce a generic framework for developing novel quasi-Newton and Anderson/Nonlinear acceleration schemes, offering a global convergence rate in various scenarios, including accelerated convergence on convex functions. They also provide empirical results in the numerical experiments.
Strengths: ... | Rebuttal 1:
Rebuttal: # Author's response
We thank the reviewer for the reviews and detailed feedback.
We begin by answering what we believe to be the most critical points of the review.
## A) Points 6 and 10: Clarity of algorithm
> **10)** The most significant weakness [..] is the lack of formal algorithm for the ... | Summary: In the paper, the authors propose a new framework that connects Quasi-Newton methods with Anderson acceleration. By exploiting the Cubic Regularization technique, the authors achieve new competitive convergence rates comparable to first-order methods in the worst case and second-order methods in the best case.... | Rebuttal 1:
Rebuttal: # Author's response
We thank the reviewer for the insightful comments. All the reported typos in points 1 to 9 are corrected, and we are reviewing the paper in more detail to correct the remaining ones.
We also thank the reviewer for finding the results presented in the paper both original and s... | Rebuttal 1:
Rebuttal: # Author rebuttal
We thank all the reviewers for their insight and precise comments, and we appreciate seeing that the reviewers see the paper as innovative and hold great potential for further advancements in optimization methods (FnQv), providing a solid theory (4Qfc).
Since some comments in t... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Nearly Tight Bounds For Differentially Private Multiway Cut | Accept (oral) | Summary: The authors give a novel and simple diferentially private algorithm for the min $s$-$t$ cut and multiway $k$-cut problems.
Strengths: Originality:
Work seems original.
Simple but novel!
Quality:
Quality of the paper is high.
The authors give interesting upper and lower bounds for the min $s$-$t$ cut and mult... | Rebuttal 1:
Rebuttal: > What are some computational complexity implications of your work?
Thank you for this great question. There are several implications. *First*, our Alg 1 uses a non-private min $s$-$t$ cut algorithm in a black-box manner. Hence, it essentially translates to any popular computation setting without... | Summary: The paper presents differentially private algorithms for the s-t mincut and multiway cut
problems under an edge privacy model with weights. It shows that a simple algorithm gives
an O(n) additive error bound for the s-t mincut problem, and obtain a near tight lower bound.
This is used for obtaining a bound for... | Rebuttal 1:
Rebuttal: > The main ideas in the paper are quite simple. ...
We agree that a separate related work section is beneficial to have all the related references in one place, and we will make one for the final version of our paper. However, there is very limited work on differentially private min-cut and min s... | Summary: This paper presents edge-DP algorithms for min s-t cut and multiway k-cut. The s-t cut algorithm injects fake edges with weights drawn from the exponential distribution, then returns the min s-t cut of the modified graph. The k-cut algorithm is based on $O(\log k)$ calls to the min s-t cut algorithm. They show... | Rebuttal 1:
Rebuttal: > The proof of Lemma 4.1 does not match exactly the steps of Algorithm 2. ...
We agree with your comment that Lemma 4.1 does not match exactly the steps of Algorithm 2. Our intent is to represent Algorithm 2 in the simplest form, and Lemma 4.1 provides an alternative view on Algorithm 2, which we... | null | null | Rebuttal 1:
Rebuttal: We thank all the reviewers for their constructive comments.
Two of the reviewers asked for clarification of the analysis in Lemma 3.3, so in addition to our comments below, we provide more explanations that we will include in the final version.
For intuition and brevity, assume that $\epsilon = ... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
DiffAttack: Evasion Attacks Against Diffusion-Based Adversarial Purification | Accept (poster) | Summary: This paper aims to generate adv-samples that can effectively attack diffusion-based purification. Two key strategies are proposed: (a) deviated reconstruction loss which is used as a surrogate loss to relieve gradient obfuscation existing in the long computational graph; (b) segment-wise forward-backward strat... | Rebuttal 1:
Rebuttal: > Q1: The speed of segment-wise forward-backward compared with the original forward-backward.
Due to the page limits, in line 361-362, we mention that we provide comparisons of runtime between DiffAttack and [a] (standard forward-backward) in Appendix D.3. The results demonstrate that DiffAttack ... | Summary: This paper proposes a new adaptive attack to unify the robustness evaluation of diffusion-based purification defenses, including both score-based and DDPM-based diffusion models. The new attack mainly includes two novel techniques to overcome the challenges of evaluating diffusion-based purification defenses. ... | Rebuttal 1:
Rebuttal: We thank the reviewer for the valuable suggestions, and we are glad that the reviewer finds our work novel, timely, with strong results.
> Q1: Related work for memory-efficient back-propagation
Thanks for the suggestion. Prior works [a,b] propose the technique of gradient checkpointing for backp... | Summary: This paper proposes a framework, DiffAttack, to attack against diffusion-based adversarial purification defenses. To address the issue of gradient vanishing/exploding, the paper introduces the concept of deviated-reconstruction loss and theoretically analyzes its relationship with density gradient estimation. ... | Rebuttal 1:
Rebuttal: > Q1: Detailed explanation of the benefits of deviated-reconstruction loss to alleviate gradient vanishing/exploding.
Thanks for the suggestion. We will add the following explanations in Section 3.2. The deviate-reconstruction loss is applied at the intermediate time steps, and thus, the regulari... | Summary: The paper introduces an adversarial attack focused on diffusion-based purification techniques. In diffusion-based purification, a (malicious) input is first noised and then denoised by a diffusion model before being fed through the classifier. The paper aims at solving the issues that arise when using backprop... | Rebuttal 1:
Rebuttal: > Q1: Related work of the gradient checkpointing technique.
Thanks for the suggestion. We will add discussions of the related work in our revision. To the best of our knowledge, prior works [a,b,c] propose the technique of gradient checkpointing to perform gradient backpropagation with memory eff... | Rebuttal 1:
Rebuttal: We thank all the reviewers for their comments and valuable feedback. We are glad that the reviewers find our work novel, effective, and well-written. We have made the following major updates following the reviews’ suggestions to further improve our work.
1. We add related work of memory-efficient... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
SegRefiner: Towards Model-Agnostic Segmentation Refinement with Discrete Diffusion Process | Accept (poster) | Summary: This paper develops a method based discrete diffusion process to refine coarse segmentation masks with input images in the context of semantic image segmentation. In this work, the refinement problem has been formulated as an iterative data generation process. To work with binary masks, as adapted for most seg... | Rebuttal 1:
Rebuttal: # Response to Reviewer Xhma
### **Q1. Clarification of Method Description**
Thank you for your valuable and thoughtful comments and we will rewrite the Method section to enhance clarification. Due to the confusion caused by "source and target masks", we will replace them with "fine and coarse mask... | Summary: In this paper, the authors propose a model-agnostic solution called SegRefiner to enhance the quality of object masks generated by different segmentation models. The approach introduces a novel perspective on segmentation refinement, treating it as a data generation process. By leveraging a series of denoising... | Rebuttal 1:
Rebuttal: # Response to Reviewer oFWK
### **Q1. Value of $\alpha$ in the Objective Function**
Regarding your concern, we do the following ablation study on instance segmentation with coarse masks from MaskRCNN (ResNet-50). The first row of the Table below refers to not using texture loss for training, and t... | Summary: This manuscript presents a model-agnostic solution for segmentation refinement, which implements the refinement process of the coarse masks through a series of denoising diffusion steps. Comprehensive experiments on various segmentation tasks are conducted, including semantic segmentation, instance segmentatio... | Rebuttal 1:
Rebuttal: # Response to Reviewer cXkh
### **Q1. (a) Why using diffusion**
As mentioned in Line 40 and Fig. 1 in the paper, we found there are diverse types of errors presented in the coarse masks. This is the reason for the underperformance of existing methods because attempting to address all these diverse... | Summary: This work proposes diffusion-based image segmentation refinement approach, SegRefiner, using a newly designed discrete diffusion process. SegRefiner performs model-agnostic segmentation refinement that takes in coarse output masks from any segmentation model as an input and refines them through a series of dif... | Rebuttal 1:
Rebuttal: # Response to Reviewer ygQQ
### **Q1. About Eq. (9) and Eq. (11)**
**Firstly, for Eq. (9),** it's derived from the Bayes' theorem：
\begin{equation}\begin{split}
q(x_{t-1}|x_t, x_0) = \frac{q(x_t|x_{t-1}, x_0)q(x_{t-1}|x_0)}{q(x_t|x_0)}.
\end{split}\end{equation}
According to Eq. (6) and Eq. (7), ... | Rebuttal 1:
Rebuttal: ## General Response
We have uploaded a PDF file here, which includes:
* **Table 1:** (to Reviewer cXkh Q4) An additional experiment to evaluate SegRefiner on DIS-5K dataset with the commonly used metrics for Dichotomous Image Segmentation.
* **Figure 1:** (to Reviewer cPQx Q2) Visualizations of h... | NeurIPS_2023_submissions_huggingface | 2,023 | Summary: This paper proposes a novel method for improving the quality of object masks generated by different segmentation models. The method, called SegRefiner, is model-agnostic, meaning that it can be used with different segmentation models. SegRefiner works by interpreting segmentation refinement as a data generatio... | Rebuttal 1:
Rebuttal: # Response to Reviewer cPQx
### **Q1. The coarse masks used for training**
We very appreciate your insightful comments and will clarify how the coarse masks are generated in earlier sections and in the overview figures in our next version.
### **Q2. More examples of inference**
More additional vi... | null | null | null | null | null | null |
Unified 3D Segmenter As Prototypical Classifiers | Accept (poster) | Summary: In this paper, the authors propose to use prototype paradigm to develop a generic framework for semantic segmentation, instance segmentation and panoptic segmentation. The proposed ProtoSeg use prototype association to distribute points to different prototypes and dynamic update prototypes according to the ass... | Rebuttal 1:
Rebuttal: We thank reviewer Mj9E for the valuable time and constructive feedback. We provide a point-to-point response below.
### Weaknesses 1: Limitation
Sorry for the confusion. We discussed some limitations of this work in the supplementary material, which is related to model efficiency.
### Weaknesses... | Summary: This paper introduces PROTOSEG, a prototype-based model that aims to unify semantic, instance, and panoptic segmentation on point cloud. It treats these tasks as a classification problem with varying levels of granularity and achieves high performance on multiple benchmarks.
Strengths: 1. The research topic a... | Rebuttal 1:
Rebuttal: We thank reviewer Mj9E for the valuable time and constructive feedback. We provide point-to-point response below.
### Weaknesses 1: Parameters, flops, and speed consumption
We detail the complexity analysis in the supplementary material. Although the Sinkhorn-Knopp iteration is utilized, in prac... | Summary: This work aims to unify different 3D segmentation tasks, i.e., semantic, instance, and panoptic segmentation on point clouds. The authors propose unified prototypical classifiers --- ProtoSeg -- on top of a Transformer backbone, for multi-task 3D segmentation. The main idea is to establish prototype associatio... | Rebuttal 1:
Rebuttal: Thanks reviewer VQFt for the valuable time and constructive feedback. We provide point-to-point response below.
### W1: Discussions with multi-task image segmentation methods
Thank you for your reminder. As you mentioned, we plan to introduce a new subsection in Section 4 to address multi-task ima... | Summary: The paper presents a unified 3D segmenter to unify different segmentation tasks, including semantic, instance, and panoptic segmentation, for 3D point clouds. It proposes to define the segmentation task as prototype classification and designs the prototype association and prototype update process.
Strengths:... | Rebuttal 1:
Rebuttal: We thank reviewer MoHU for the valuable time and constructive feedback. We provide a point-to-point response below.
### Weaknesses 1: Challenges and the corresponding methods of unifying segmentation tasks
Thank you for your insightful review. We greatly value your observation and will revise Sec... | Rebuttal 1:
Rebuttal: Thank you so much for your careful review and suggestive comments. We will revise our paper according to your comments. The major changes are as follows:
+ We will detail the challenges and methods of unified segmentation tasks, introduce a diagram of the prototyping process, and detail the use o... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Refining Diffusion Planner for Reliable Behavior Synthesis by Automatic Detection of Infeasible Plans | Accept (poster) | Summary: In this work, authors propose a new metric called “restoration gap” that is able to measure the reliability of plans emerging from diffusion-based planners. They then suggest using this metric along with guidance, “restoration guidance”, to encourage the diffusion-based planner to generate more reliable plans.... | Rebuttal 1:
Rebuttal: **Q: Comparison to rejection-based methods, especially in scenarios where the problem context remains fixed.**
**A:** We respectfully disagree with the reviewer’s comment on comparing our approach to rejection-based methods. It is important to note that rejection-based methods rely on strong assu... | Summary: This paper investigates the potential of diffusion models in addressing long-horizon sparse reward tasks. While diffusion models are effective generative models, they are not inherently guaranteed to generate feasible plans. To overcome this limitation, the study proposes an approach to refine unreliable paths... | Rebuttal 1:
Rebuttal: **Q: What if the diffusion model is not well-trained? Can you comment on the accuracy of the restoration gap in such case?**
**A:** We acknowledge that inaccuracies might arise especially in a low-data regime if the diffusion model is not good at estimating score functions for certain trajectorie... | Summary: The paper proposes a novel metric named restoration gap for evaluating trajectory quality generated by diffusion models. It further proposed restoration gap guidance to refine the diffusion generation. The restoration gap predictor also provides explainability of the model. Experiment results indicate improved... | Rebuttal 1:
Rebuttal: **Q: While the restoration gap is intuitive in the given examples (2D map with walls), it is unclear how the gap works in more complex environments or environments with fewer hard constraints.**
**A:** In more complex environments with fewer hard constraints, determining infeasibility might not b... | Summary: This paper introduces a novel approach to refining unreliable plans generated by diffusion models in long-horizon, sparse-reward tasks. While diffusion-based planning has shown promise, the generative nature of diffusion models can result in infeasible plans, limiting their usefulness in safety-critical applic... | Rebuttal 1:
Rebuttal: **Q: How does the proposed approach address the potential issue of rare but valid trajectories being considered abnormal due to the distribution of the expert dataset?**
**A:** We appreciate the reviewer’s insightful comment regarding the accuracy of the restoration gap. Indeed, the inaccuracy co... | Rebuttal 1:
Rebuttal: We would like to sincerely thank all reviewers for their time and effort in reviewing our paper, and for their constructive feedback and comments. In this thread, we answer commonly asked questions. Please find below our response to your insightful comments and suggestions.
**Q: Lack of substanti... | NeurIPS_2023_submissions_huggingface | 2,023 | Summary: The paper proposes to refine the predictions of a diffusion-based planner by identifying unfeasible plans prior to execution. The newly introduced "restoration gap" measures the reconstruction error for a given sample under some noisy perturbation, and it is argued that this gap is a reasonable proxy for feasi... | Rebuttal 1:
Rebuttal: **Q: I disagree a bit on the high-level motivation for this work. I don’t think there is anything about the restoration gap that specifically targets “unfeasible” plans.**
**A:** In low-data regime, we acknowledge the scenario where even feasible plans can appear highly OOD, leading to a substant... | null | null | null | null | null | null |
VanillaNet: the Power of Minimalism in Deep Learning | Accept (poster) | Summary: This paper aims to minimize the complex design of modern deep models. The authors propose an architecture, VanillaNet, which eliminates high depth, residual connections, non-linear activations, self-attentions, and normalization layers, but during training, the model still use some components, such as activati... | Rebuttal 1:
Rebuttal: **Q1:** Merging operation.
**A1:** Sorry for the unclear statement. The inference architecture of VanillaNet can be found in the supplementary material. During the training time, each convolution layer is doubled into two convolutions with an activation between them. After training, the two convo... | Summary: The paper presents VanillaNet, a neural network architecture that emphasizes simplicity by avoiding high depth, shortcuts, and complex operations like self-attention. The proposed VanillaNet performs comparably to well-known deep neural networks and vision transformers.
Strengths: **Originality**: To the best... | Rebuttal 1:
Rebuttal: **Q1:** More parameters and FLOPs.
**A1:** Thanks for the nice concern. The inference speed in GPUs and the number of parameters or FLOPs are two different aspects to measure the performance of neural networks. According to the “no free lunch” rule, it is nearly impossible to be satisfactory to b... | Summary: The paper introduces VanillaNet, a neural network architecture that emphasizes simplicity in design while maintaining remarkable performance in computer vision tasks. VanillaNet achieves this by avoiding excessive depth, shortcuts, and intricate operations like self-attention. The authors propose a "deep train... | Rebuttal 1:
Rebuttal: **Q1:** More parameters and FLOPS.
**A1:** Thanks for the nice concern. The inference speed in GPUs and the number of parameters or FLOPs are two different aspects to measure the performance of neural networks. According to the “no free lunch” rule, it is nearly impossible to be satisfactory to b... | Summary: This paper introduces VanillaNet, a neural network architecture that emphasizes simplicity in design. By discarding commonly used ingredients, e.g., high depth, shortcuts, and self-attention mechanism, VanillaNet is refreshingly concise. VanillaNet series have only 5-13 layers. The authors devise the “deep tra... | Rebuttal 1:
Rebuttal: **Q1:** The number of parameters.
**A1:** Thanks for the suggestion. Although the deep training technique will increase the training cost, it would not affect the inference cost, which is much more important when implementation.
**Q2:** Inference speed:
**A2:** Thanks for the nice concern. We t... | null | NeurIPS_2023_submissions_huggingface | 2,023 | Summary: The authors propose a minimalist neural network backbone architecture as an alternative to current ResNets and ViTs. The minimalism is focussed on faster inference, avoiding complex operations with higher time complexity when implemented on modern GPUs. In their proposed architecture, individual operations are... | Rebuttal 1:
Rebuttal: **Q1:** Missing details in series activation function.
**A1:** Sorry for the unclear claim. The scale and bias terms are learned parameters. We initialize them using Kaiming initialization [1]. We also test other initialization. As shown in the Table below, the initialization affect little final ... | null | null | null | null | null | null |
Fast and Simple Spectral Clustering in Theory and Practice | Accept (poster) | Summary: The paper is concerned with the spectral method for clustering vertices of graph. Specifically, the authors find that the classical spectral clustering methods need eigendecomposition of the graph Laplacian matrix which is computationally expensive, and that the authors propose to use a power method to bypass ... | Rebuttal 1:
Rebuttal: Thank you for your positive review. We address your questions below.
**Code.**
The code to reproduce our experiments is included with the submission in the supplementary material.
When the paper is published, we will make the code publicly available online.
**Editorial points.**
Thank you for yo... | Summary: This paper proposed a new spectral-based graph clustering algorithm. The main novelty of this algorithm is a fast implementation of vertex embedding, built on the power method.
Previous algorithms achieved a low dimensional embedding via two steps:
1) compute k eigenvectors of the graph Laplacian matrix;
2... | Rebuttal 1:
Rebuttal: Many thanks for your detailed and thoughtful review and questions.
**Balancedness assumption.**
We address the balancedness assumption in our top-level comment.
Furthermore, in the attached PDF, we include the results of the requested SBM experiment.
We set $p = 0.01$ and $q = 0.00001$ as sugges... | Summary: The paper deals with the problem of spectral clustering in networks. The paper proposes a fast spectral clustering method by utilizing the power method and random projection. The paper also provides theoretical guarantee of cluster recovery.
Strengths: The main strengths of the paper are -
(1) The paper pro... | Rebuttal 1:
Rebuttal: Many thanks for your time and thoughtful review. We respond to your specific questions below.
**Question 1.**
We address the effect of the network sparsity in our top-level comment.
**Question 2.**
This is a good question, and it is an example of a place where theory and practice diverge.
In the... | Summary: The paper proposes a new spectral clustering algorithm that instead of computing $k$-leading eigenvectors, the authors proposes to use power method to embed the nodes into a $\log(k)$ dimension space, which saves computational time when $k$ is large. Theoretical guarantees on misclassification rate with const... | Rebuttal 1:
Rebuttal: Thank you for your time and thoughtful comments.
We respond to each of your questions below.
**Further study of the Letter dataset.**
For computing the eigenvectors, we use the Implicitly Restarted Arnoldi Method (IRAM) provided by the eigsh method in the Python scipy library, which itself uses t... | Rebuttal 1:
Rebuttal: We thank the reviewers for their time and thoughtful reviews.
Here we respond to questions which were shared by more than one reviewer, and we respond to other questions individually with reference to figures in the attached PDF.
**Balanced clusters assumption.**
The balanced cluster assumption i... | NeurIPS_2023_submissions_huggingface | 2,023 | Summary: The paper proposes a new fast spectral clustering algorithm. The main idea of the algorithm is to, rather than embed the vertices according to the k eigenvectors of the graph Laplacian, use the power method to efficiently approximate a random projection of the spectral embedding. Then, as in the classical spec... | Rebuttal 1:
Rebuttal: Thank you for your thoughtful review. We respond to your questions regarding the experimental evaluation below.
**Comparison with additional methods.**
Table 1 in the attached PDF includes additional experimental results for the KASP algorithm proposed by Yan et al. [34].
We find that on real-wor... | null | null | null | null | null | null |
Approximate Allocation Matching for Structural Causal Bandits with Unobserved Confounders | Accept (poster) | Summary: The paper presents two primary contributions for the causal bandit problem with unobserved confounders, general graphs, and discrete RVs.
1) a lower bound on the regret any algorithm can obtain for this problem.
2) an upper bound on the regret of a newly proposed algorithm that is able to generalize across in... | Rebuttal 1:
Rebuttal: Thank you for your valuable feedback and supportive comments. Your questions will help us improve the clarity and presentation of the paper. Please see our responses below.
“size of graphs in experiments”: We agree with the reviewer that extending our simulations to larger graphs will help demon... | Summary: This paper studies the structural causal bandits in which the causal graph contains latent confounders. The paper contains a lower-bound on attainable regret as well as an algorithm that achieves instance-dependent upper bound.
Strengths: The paper studies an established problem, building primarily on [20] an... | Rebuttal 1:
Rebuttal: Thank you for your valuable feedback. Please see our responses below.
In L228 and L291, the upper and lower bounds depend on the topology of the causal graph. The complicated space of general causal models is not amenable to a closed-form result or an approximated value. Instead, the factors appe... | Summary: The authors present a method for studying structural causal bandits with unobserved confounders. They show that their method, which unlike the baseline (POMIS), includes two tuning parameters, has superior empirical performance on three synthetic experiments (three synthetic SCMs).
Strengths: The main review... | Rebuttal 1:
Rebuttal: Thank you for your valuable feedback and constructive comments. We hope that our responses will address the reviewer's concerns.
"title is a bit misleading": Please note that we are not claiming that ours is the first causal bandit algorithm with unobserved confounders. Our goal with the title wa... | Summary: This paper considers online learning in a structural causal bandit whose stochastic environment is govened by causal relations represented by an SCM. The paper gives a logarithmic asymptotic regret lower bound characterized by an optimization problem, and inspires on that, provides an algorithm to untilize the... | Rebuttal 1:
Rebuttal: We first thank the reviewer for valuable feedback and supportive comments. Please see below for our response.
“highlight extra contributions of this paper”:
The major contribution of this paper is utilizing a canonical SCM formulation, which partitions the space of causal models, to characteri... | Rebuttal 1:
Rebuttal: We thank the reviewers for their valuable feedback and AC for the assistance during the review process. We address each review’s comments in separate rebuttals. In response to reviewer BoCS, the attached PDF file includes additional experiment results of Thompson Sampling (TS) in comparison with t... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
OneNet: Enhancing Time Series Forecasting Models under Concept Drift by Online Ensembling | Accept (poster) | Summary: The paper proposes a method for online learning based on ensemble learning. The method proposed by the authors leverages cross-variable and cross-time dependency. In addition, the authors propose a reinformed learning method that dynamically adjusts the weights used to combine the models.
Strengths: See Quest... | Rebuttal 1:
Rebuttal: We sincerely apologize for any confusion caused by the lack of clarity in our paper. We appreciate your feedback and would like to address your comments to provide a clearer understanding of our work.
----
**Concern 1 Online learning literature not based on deep learning**
Regarding the literat... | Summary: Motivated by the observation that a single model may not consistently achieve optimal performance in time series forecasting, the author proposes a method, which combines two stream architectures (cross-time and cross-variable models) using online convex programming. However, due to the limitations of online c... | Rebuttal 1:
Rebuttal: Thank you for your valuable feedback and positive evaluation of our paper. We appreciate your thorough review and are glad that you find the paper easy to follow and well-presented.
----
**Concern 1 About the motivation**
Regarding the motivation, we understand your concern about the clarity of... | Summary: This paper addresses the problem of concept drift in online Time Separate forecasting. A method OneNet is proposed that uses temporal correlations and inter-variables dependencies. OneNet is compared with other state of the art methods and OneNet is demonstrated to be superior. Furthermore, through empirical s... | Rebuttal 1:
Rebuttal: Thank you for your feedback and positive comments. We appreciate your thorough evaluation on our paper. We are glad that you find the mathematical formulation, the training of two independent forecasts, and the use of long-term historical and short-term variations as strengths of our work.
We agr... | Summary: The paper introduces an approach called Online ensembling Network (OneNet), aimed at online updating of time series forecasting models. This approach is crafted to handle the concept drift problem efficiently using streaming data. The authors note that while some algorithms designed for online time series fore... | Rebuttal 1:
Rebuttal: Thank you for your valuable feedback and comments on our paper. We appreciate the positive feedback on the readability and the importance of the considered problem in the real world. We also acknowledge the points you raised regarding the technical novelty and the need for theoretical analysis in ... | Rebuttal 1:
Rebuttal: We thank the Reviewers for the insightful comments and detailed feedback. We are delighted that reviewers find our paper has **good motivation** (QKSm), **easy to follow** (QKSm, r9Pc), **excellently -presented and organized** (UB5d, QKSm), **enough technical novelty** (QKSm, 7dRB), and **Impressi... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Near-optimal learning with average Hölder smoothness | Accept (poster) | Summary: The authors proposed a notion named average Hölder smoothness, which is a generalization of average Lipschitz smoothness. They gave sample complexity bounds of learning the averge Hölder smooth function class in both the realizable and agnostic setting with corresponding algorithms. They also gave lower bounds... | Rebuttal 1:
Rebuttal: We thank the reviewer for the insightful review and helpful comments.
- Regarding computational efficiency, we agree this is very plausible. In fact, after the initial submission we were able to prove a generalization of Proposition 3.2 in the agnostic case, with an excess error decaying as $\frac... | Summary: This paper introduces the notion of "average Hölder smoothness", generalizing the notion of Holder smoothness.
Authors motivate the introduction of this class, explaining that the Holder constant is a supremum over many slopes that can eventually not be representative of the general behavior of this slope.
Aft... | Rebuttal 1:
Rebuttal: We thank the reviewer for the insightful review and helpful comments.
- Thank you for spotting the bug for small $N$, we will fix this in the revised version.
- Regarding the proposed alternative notion of local smoothness: Though this is a natural idea, unfortunately it does not suffice to contro... | Summary: This paper generalizes the notion of average Lipschitz smoothness proposed by Ashlagi et al. by extending it to Hölder smoothness. Eventually, the class of the average holder smooth function is much larger than the class of the worst-case holder smooth function. Further, the paper provides upper and lower risk... | Rebuttal 1:
Rebuttal: We thank the reviewer for the insightful review and helpful comments.
- In the revision, we will provide a cleaner explanation of the algorithm in the main text.
- Regarding the exponential dimension dependence, this is very typical of non-parametric regression methods, as exemplified e.g. by the ... | null | null | null | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Does Visual Pretraining Help End-to-End Reasoning? | Accept (poster) | Summary: - The authors explore whether self-supervised visual pretraining can allow end-to-end methods to be more general purpose.
- The paper proposes a self-supervised representation learning framework for this task, based on slot tokens, and evaluates performance on multiple benchmarks.
Strengths: - The paper is ... | Rebuttal 1:
Rebuttal: We thank Reviewer PrZv for their thorough review. We seek to address their listed considerations below:
> **On synthetic evaluation**
We primarily evaluate the reasoning capabilities for IV-CL over established benchmarks for spatiotemporal reasoning, which are largely synthetic due to the diffi... | Summary: The authors investigate whether unsupervised pre-training yields representations that are suitable for solving reasoning tasks. To that end, they propose a new transformer-based model inspired by masked autoencoders that is pre-trained on CATER. The authors find that the proposed model outperforms supervised p... | Rebuttal 1:
Rebuttal: We thank Reviewer e8rc for their review, and seek to address their listed questions and feedback:
> **On relevant work**
We do agree that the “[A] Solving Reasoning Tasks with a Slot Transformer” is relevant and related to our approach. We would like to highlight that our IV-CL approach is more... | Summary: This paper explores with the need of explicit visual abstraction for visual reasoning tasks.Instead, this paper explores the use of implicit representations learned through a slot attention module.The proposed approach uses slot attention in a frame prediction frame-work that aims to compress visual input info... | Rebuttal 1:
Rebuttal: We thank Reviewer t1ps for supplying useful feedback in response to our work, as well as relevant prior work, which we address:
> **CATER moving camera performance**
We conduct the moving camera experiments by adopting the same pretrained checkpoint as for the static camera experiments (which wa... | Summary: This paper investigates if an end-to-end trained neural network is able to solve challenging visual reasoning tasks. A new representation learning framework named implicit visual concept learner (IV-CL) is proposed. It consists of two main components. First, a single image is compressed into a small set of tok... | Rebuttal 1:
Rebuttal: We thank Reviewer zViR for the in-depth, constructive feedback. Below we seek to address the reviewer’s concerns:
> **On explicit information from supervision**
The reviewer suggests that “without any supervision, [the learned representation] is not controllable and stable, and difficult to gen... | Rebuttal 1:
Rebuttal: We thank the reviewers for their thorough, thoughtful, and constructive comments. We are happy to hear that the reviewers consistently agree that the proposed work is interesting, well-presented, and presents novel insights. In particular, the reviewers highlighted that the proposed method is te... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Vulnerabilities in Video Quality Assessment Models: The Challenge of Adversarial Attacks | Accept (spotlight) | Summary: This paper studies the adversarial robustness of NR-VQA models. A new loss function called Score-Reversed Boundary Loss is designed to push the estimated quality score of adversarial video far away from its MOS towards a specific boundary. It also proposes a patch-based random search method for black-box atta... | Rebuttal 1:
Rebuttal: Thank you for the review. We address your mentioned weakness and questions below. In this response, the performance of VQA models is evaluated by SRCC, PLCC, and $R$. Due to the limited space in the response, only SRCC is listed in the following Tables. Note that all these three metrics exhibit si... | Summary: This paper proposes a patch-based random search method for white-box attacks and black-box attacks for NR-VQA models. A Score-Reversed Boundary Loss is proposed to push the adversarial video's estimated quality score far away from its MOS towards a specific boundary. And the robustness of NR-VQA models is eval... | Rebuttal 1:
Rebuttal: Thank you for the review. We address your mentioned weakness, questions, and limitations below.
**Regarding the comparison with [1]:** Thank you very much for reminding us of [1] which has been neglected in searching for the related works. We will include this paper in the related work in the re... | Summary: The paper studies the adversarial robustness of No Reference Video Quality Assessment Algorithms. The paper introduces a new loss function, "Score-Reversed Boundary Loss," and demonstrates its usefulness in designing adversarial attacks for NR VQA algorithms. Overall the paper is well written, contains adequat... | Rebuttal 1:
Rebuttal: Thank you for the review. We address your mentioned weakness, questions, and limitations below.
**Regarding JND constraint based on the principles of Vision Science:** Thanks for the valuable comment. In our work, the JND constraint is designed based on the principles of Vision Science, and the ... | Summary: The adversarial attacks against video quality assessment models are proposed in both white box and black box scenarios.
Strengths: 1. The problem is important and practical.
2. The attempt to avoid subjective testing is a good try.
3. The attacks seem to work as in Tables 1 and 2.
Weaknesses: 1. The auth... | Rebuttal 1:
Rebuttal: Thank you for the review. We address your mentioned weakness and questions below.
**Regarding the association of JND constraint in our work with $L_2$ and $L_\infty$ norm:** Thanks for the insightful comment. The authors agree that in extreme situations, only applying $L_2$ norm may lead to modif... | Rebuttal 1:
Rebuttal: According to reviewers C8CZ, mjyw, xq3Y, and NoLq, the examples of the original video frames and the corresponding adversarial video frames are presented in the supplementary PDF under both white-box and black-box settings.
Pdf: /pdf/200090a7834dff8534218d9779660aa0f8e68e99.pdf | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Private estimation algorithms for stochastic block models and mixture models | Accept (spotlight) | Summary: In this paper, the authors investigate graph clustering with the Stochastic Block Model (SBM) and metric clustering with the Gaussian Mixture Model (GMM) within a novel privacy-preserving framework. They propose differentially private (DP) algorithms for both problems, which offer significantly improved perfor... | Rebuttal 1:
Rebuttal: We thank the reviewer for their constructive feedback and suggestions.
**comment:**
> It might be useful to briefly discuss why the current framework does not achieve $\log(k)^{1/2 + c}$ separation in polynomial time.
**response:** On a high level, our framework converts non-private algorithms ... | Summary: This paper introduces a general framework for designing efficient private estimation algorithms from robust non-private algorithms. Their approach, based on a certain stability analysis for convex optimization, is applied to weak and exact recovery in the stochastic block model (SBM) and for learning mixtures ... | Rebuttal 1:
Rebuttal: We thank the reviewer for their constructive feedback and suggestions.
**comment:**
> The choice of objective function in the Gaussian mixture application wasn't clear to me - where does the all ones matrix come from?
**response:** We choose the objective function $||W-J||\_F^2$ (where $J$ is th... | Summary: The paper studies private parameter estimation in certain high-dimensional statistical inference tasks. In particular, community detection in stochastic block models (where the privacy is with respect to edge changes) and estimation of unknown means in isotropic Gaussian mixture models.
Although these problems... | Rebuttal 1:
Rebuttal: We thank the reviewer for their constructive feedback and suggestions.
**comment:**
> Line 45: "However, for the more challenging task of graph clustering, node-DP is sometimes impossible to achieve." I do not follow. For the SBM, I believe that node-DP is possible to achieve since the informatio... | Summary: The paper considers the problems of weak and strong recovery for stochastic block models, as well as the problem of estimating the means of Gaussian Mixture Models with sphrerical Gaussians, under the constraint of differential privacy. Their contributions include:
* A computationally efficient algorithm that... | Rebuttal 1:
Rebuttal: We thank the reviewer for their constructive feedback and suggestions.
## Comparison with Arbas-Ashtiani-Liaw (AAL)
AAL [1] is a subsequent and independent work to ours. Their technique (as well as many previous papers, e.g. [2,3]) is based on the Subsample-And-Aggregate framework, while our appr... | null | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.