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 |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
FineMoGen: Fine-Grained Spatio-Temporal Motion Generation and Editing | Accept (poster) | Summary: Existing methods often struggle to generate intricate motion sequences that align with fine-grained descriptions. Addressing this, the authors introduce FineMoGen, a diffusion-based motion generation and editing framework capable of synthesizing intricate, detailed motions that align with user instructions. Fi... | Rebuttal 1:
Rebuttal: We sincerely thank the reviewer for the detailed feedback and comments. We respond to the concerns below:
> **Q1**: The method illustration should be improved, such as the difference between MEA and SAMI, the usage of $G’$, etc.
**A1**: Thanks for your suggestion. We add a new figure in the atta... | Summary: The paper proposes a method for diffusion-based motion generation and editing framework for synthesizing fine-grained motions. It takes fine-grained spatio-temporal description. The framework supports zero-shot and fully-supervised generation scenario. The utilize MOE for generating global template. They have ... | Rebuttal 1:
Rebuttal: We would like to express our sincere thanks for your insightful review and constructive suggestions. We reply to each of your comments here:
> **Q1**: The natural description of the fine-grained motion may not be best utilized in some situations.
**A1**: We concur that employing the generated fi... | Summary: The paper targets an interesting problem of fine-grained spatial-temporal motion generation and editing. It introduces a new transformer structure with spatio-temporal mixturee attention. Also, a large scale dataset, called HuMMan-MoGen, has been introduced to train the proposed algorithm. Attractive experimen... | Rebuttal 1:
Rebuttal: We sincerely appreciate your constructive and helpful suggestions. And we carefully reply to each of the reviewer's comments below.
> **Q1**: Is HuMMan-MoGen diverse enough?
**A1**: The HuMMan dataset comprises 500 actions, strategically curated to encompass foundational movements. From this co... | Summary: The paper introduces a compositional text-to-motion synthesis approach that allows for control up to body part level. For fine-grained spatio-temporal editing, the motion of each body part from a predetermined set (e.g., head, arms, legs, etc.) can be described. Furthermore, for zero-shot motion control, the m... | Rebuttal 1:
Rebuttal: We sincerely appreciate your recognition of our work’s strengths and insightful suggestions. Here we respond to each of your comments in detail below.
> **Q1**: The inference setting is not clear. The qualitative results have artifacts.
**A1**: The initial examples represent the ground truth dat... | Rebuttal 1:
Rebuttal: We express our gratitude to all the reviewers for their valuable and constructive feedback. We are glad at reviewers’ recognition of our work's strengths:
**1)** The problem setting is interesting and practical (R-C9Az, R-DsXZ, R-CFoF, R-4S8a).
**2)** The decoupling of spatial independence and... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Structured Neural-PI Control with End-to-End Stability and Output Tracking Guarantees | Accept (poster) | Summary: This paper considers the problem of learning feedback controllers for multi-input, multi-output (MIMO) dynamical systems. At the core of the paper, a PI-like nonlinear control structure is proposed. Specifically, the controller consists of two NN, one which receives the set-point error (denoted as "P-part") an... | Rebuttal 1:
Rebuttal: We appreciate the detailed comments and thank the reviewer for the time spent on this paper.
**(W1) We thank the reviewers for identifying imprecise statements. We have updated them and that would be reflected in the revised paper. Below we focus on some clarifications**:
- W1-2: Steady-state b... | Summary: This paper proposes a method for optimal control of dynamical systems using a neural network-based controller that guarantees stability and output tracking. The key structure of the proposed neural proportional-integral (PI) controller is based on the strict monotonicity of the proportional and integral terms,... | Rebuttal 1:
Rebuttal: We are grateful to the reviewer for thoroughly reading our paper and providing encouraging
comments. Below, we provide a point-by-point response.
**The multiple equilibria problem**
- The multiple equilibria come from different setpoints of output to track. For example, a group of vehicles may ... | Summary: This paper presents an approach to control multi-input, multi-output dynamical systems, introducing a neural network-based PI controller offering provable stability and output tracking guarantees. The neural Proportional-Integral controller is based on strict monotonicity parameterized by strictly convex neura... | Rebuttal 1:
Rebuttal: We are grateful to the reviewer for thoroughly reading our paper and providing encouraging comments. We agree with the reviewer that carefully tuned PI controllers can work well for a large range of applications, and the controller by design has been shown to have good stability properties. **Cla... | Summary: The paper "Structured Neural-PI Control with End-to-End Stability and Output Tracking Guarantees" presents a well-motivated and technically sound approach to address the lack of provable guarantees in neural network-based controllers. The combination of structured neural-PI control, equilibrium-independent pas... | Rebuttal 1:
Rebuttal: We are grateful to the reviewer for thoroughly reading our paper and providing encouraging
comments. Below, we provide a point-by-point response.
**The applicability and potential limitations in extending the method to other domains.**
- Theoretically, this method works for systems that satisfy... | Rebuttal 1:
Rebuttal: We thank the reviewers for the detailed comments. We notice there are some common concerns and respond to these questions in the general response below.
We have attached a **one-page PDF containing the figures and updated numerical results**. Because of the space limits, all the numerical resul... | NeurIPS_2023_submissions_huggingface | 2,023 | Summary: This paper propose a neural-PI controller based on the newly designed strictly convex neural network. Thanks to the EIP condition of the considered systems, the neural-PI controller naturally satisfies the stability and output tracking guarantee with the special structure of the SCNN. The efficacy of the propo... | Rebuttal 1:
Rebuttal: We are grateful to the reviewer for thoroughly reading our paper and providing encouraging comments. Below, we provide a point-by-point response.
**Can this method outperform the existing methods (RL such as PPO or MPC)?**
- The Neural-PI controller can be trained by almost all the RL algorithm... | null | null | null | null | null | null |
ChatGPT-Powered Hierarchical Comparisons for Image Classification | Accept (poster) | Summary: The article presents an approach for refining the classification performance of a vision-language model (i.e. CLIP), using descriptions derived from a large-language model (i.e. GPT-4). Different from previous attempts in this direction [26], directly extracting a discriminative description for each of the cla... | Rebuttal 1:
Rebuttal: We appreciate the reviewer's recognition of our approach's innovative concept of classification through specific descriptions, insights into our visualizations (Figs. 4, 5, and 6), and acknowledgment of the framework's adaptability to new large-language and vision-language models.
### [Q1] Improve... | Summary: This paper presents a hierarchical comparison method for zero-shot open-vocabulary image classification. By leveraging CLIP and Large Language Models (ChatGPT here), the authors build class hierarchies by recursively comparing and grouping classes. This hierarchy enables image classification by comparing image... | Rebuttal 1:
Rebuttal: We extend our heartfelt appreciation to you for acknowledging the high interpretability of our proposed method, which derives inspiration from human conceptual organization, and your recognition of the efficient utilization of ChatGPT's extensive pretraining knowledge and the comprehensive evaluat... | Summary: This paper shows an example of introducing the power of LLMs in the zero-shot classification. This paper proposes a framework (or a process flow) rather than proposing a new mathematical algorithm. However, as shown in the experimental results section, it is demonstrated that the classification accuracy is abo... | Rebuttal 1:
Rebuttal: We extend our sincere gratitude to the reviewer for their recognition of our work's innovation in integrating grouping and Language Model-generated descriptions to enhance zero-shot classification performance.
### [Q1] Impact on the choice of large language models
We deeply appreciate the valuab... | Summary: This paper argues that CLIP is limited due to the semantic ambiguity of similar classes. To overcome this, the paper proposes to construct a class hierarchy with large-language models and introduces a novel image classification framework based on hierarchical comparisons that compare embeddings at each level. ... | Rebuttal 1:
Rebuttal: We express our gratitude to the reviewer for their valuable insights and recognition of our motivation, innovative hierarchical classification framework, and the significance of our experimental results and interpretability. In the subsequent sections, we address the concerns raised by the reviewe... | Rebuttal 1:
Rebuttal: We would like to express our sincere gratitude to all the reviewers for their invaluable time and expertise invested in the meticulous review of our paper. Your thoughtful feedback has immensely contributed to refining our work. We are also encouraged by the recognition of our paper's strengths:
... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
4D Panoptic Scene Graph Generation | Accept (spotlight) | Summary: This paper offers a novel perspective on understanding a 4D environment via the introduction of 4D Panoptic Scene Graph (PSG-4D). PSG-4D transforms 4D sensory data into nodes, representing entities with precise location and status data, and edges that capture temporal relationships. The authors have developed ... | Rebuttal 1:
Title: Additional Analysis
Comment: We sincerely thank the reviewers for their detailed feedback and recognition of our paper's potential impact. We appreciate the positive remarks regarding the significance, novelty, and presentation of our work. Below, we address your concerns:
**1. Detailed Analysis in ... | Summary: This paper introduces the concept of the 4D Panoptic Scene Graph (4D-PSG) as an approach for achieving a high-level visual understanding of 4D scenes. Additionally, the authors have developed two datasets annotated with 4D-PSG information. Furthermore, the authors propose a PSG4DFormer method for constructing ... | Rebuttal 1:
Rebuttal: Thank you for your thoughtful feedback. We sincerely appreciate the time you have taken to review our work and provide constructive insights.
**1. Novelty Concerns:**
We understand your concerns about the perceived linear progression from 3DSG to our proposed PSG4D. However, it's crucial to eluci... | Summary: The author introduces a new task called the 4D Panoptic Scene Graph (PSG-4D), which focuses on predicting 4D panoptic scene graphs from RGB-D video sequences. To support this task, the author provides a unique PSG-4D dataset, encompassing diverse viewpoints through two subsets: a third-view synthetic subset (P... | Rebuttal 1:
Rebuttal: #
Thank you for your constructive feedback. We deeply appreciate the time and effort you invested in reviewing our work, and we would like to address each of your raised concerns.
**1. Relevance to NeurIPS:**
While we understand the viewpoint on the relevance of our work to specific venues, we ... | Summary: This paper introduces a new task of scene graph generation that predicts 4D panoptic scene graphs from RGB-D video sequences. This paper also provides a high-quality PSG-4D dataset including PSG4D-GTA with third-view synthetic data and PSG4D-HOI with egocentric real-world data. As the third contribution, this ... | Rebuttal 1:
Rebuttal: We genuinely appreciate your comprehensive feedback on our paper, "4D Panoptic Scene Graph Generation." Based on your comments, we'd like to further clarify our intentions and the processes involved:
**Regarding Weaknesses:**
1. **Datasets based on existing datasets**:
- The motivation behin... | null | NeurIPS_2023_submissions_huggingface | 2,023 | Summary: This paper proposes a new task, named 4D panoptic scene graph generation (PSG-4D). At the same time, a dataset, including PSG4D-GTA and PSG4D-HOI, is collected for this task, and a benchmark is provided to evaluate the performance of this task. PSG-4D offers a more comprehensive scene understanding, bringing m... | Rebuttal 1:
Rebuttal: Thank you for your detailed review and for recognizing the contributions of our work on "4D Panoptic Scene Graph Generation". We are grateful for your feedback and would like to provide additional information to address your concerns.
**Regarding Weaknesses:**
1. **Sample size of the dataset**:
... | null | null | null | null | null | null |
Adversarial Training for Graph Neural Networks: Pitfalls, Solutions, and New Directions | Accept (poster) | Summary: This paper presents a comprehensive analysis of the transductive setting, highlighting the limitations of achieving perfect robustness solely by memorizing the clean graph. The authors argue that the observed improvements in robustness from previous adversarial training methods stem from unintentional leakage ... | Rebuttal 1:
Rebuttal: We want to thank the reviewer for the constructive review!
## W1: Relationship of transductive and inductive learning with robust graph diffusion
The use of adaptive spectral GNNs is **directly connected** to the discussion on the flaw of the transductive setting. With our refined model choice, ... | Summary: This paper shows fundamental limitations of transductive adversarial training of GNNs for node classification and advocates that adversarial training of GNNs should be in an inductive setting. Therefore, this paper proposes a more flexible message-passing scheme using learnable diffusion to gain robust represe... | Rebuttal 1:
Rebuttal: We want to thank the reviewer for the constructive feedback and for noting that we study an important problem while giving interesting observations and insights useful for the community.
## W1 (a): Technical novelty
We appreciate the reviewer recognizing our effective solution to implementing lo... | Summary: This paper focuses on adversarial training techniques for Graph Neural Networks (GNNs) in the context of inductive semi-supervised learning. It addresses the shortcomings of previous approaches in graph learning: perfect robustness is possible for transductive learning setting under evasion attack. The authors... | Rebuttal 1:
Rebuttal: We want to thank the reviewer for the positive review!
## W1: Clarifying the evasion attack scenario in the beginning
>Limited to evasion attack, should clarify this in a early stage of presentation
Thank you for pointing this out! Currently, we clarify this in line 67 (Section 2). In a revised... | Summary: The paper aims to fix adversarial training in graph neural networks (with respect to perturbations) by revealing limitations in prior work (from theoretical insufficiencies to experiment flaws), proposing a novel attack scheme and a superior architecture (robust diffusion) that makes adversarial training a lea... | Rebuttal 1:
Rebuttal: We want to thank the reviewer for the positive feedback!
## W1: Introducing adversarial perturbations
>Readers unfamiliar with adversarial perturbations may have a hard time understanding the paper, which may warrant a short paragraph explaining the threat model for both numerical data and graph... | Rebuttal 1:
Rebuttal: We thank the reviewers for the constructive feedback and for acknowledging that our work is intriguing and addresses key challenges (bw8r), is insightful and useful for the community (fjD8 & zNuu), as well as provides a theoretical analysis that is supported by elegant intuition (ZTec).
We addre... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
D4Explainer: In-distribution Explanations of Graph Neural Network via Discrete Denoising Diffusion | Accept (poster) | Summary: D4Explainer contributes to the important problem of explainable Graph Neural Networks (GNNs) and offers in-distribution GNN explanations for both counterfactual and model-level scenarios. The incorporation of generative graph distribution learning into the optimization objective allows D4Explainer to generate ... | Rebuttal 1:
Rebuttal: We sincerely appreciate your valuable feedback, which significantly improves the quality of this work. We would like to answer the questions you raised as follows.
> The experiments were only conducted on small graphs (<=10 nodes)
We prudently clarify that the above comment is incorrect. Please ... | Summary: The authors propose to apply discrete denoising diffusion as a technique for generating counterfactual and model-level graph explanations. From my understanding, the proposed method focuses solely on incorporating graph structure while **disregarding node features**. Additionally, the authors claim that the pr... | Rebuttal 1:
Rebuttal: We sincerely appreciate your valuable time and efforts in reviewing this paper. We would like to answer potential concerns as follows.
>Previous generative works, eg. CLEAR. "[Only edge deletion is allowed as in the previous paradigm[36,13,17] is a wrong claim.”
We apologize for the misstatement ... | Summary: The increasing application of graph neural networks (GNNs) has sparked significant interest in understanding their explainability. In this paper, the authors introduce D4Explainer, a method that provides in-distribution GNN explanations for both counterfactual and model-level explanation scenarios. By integrat... | Rebuttal 1:
Rebuttal: We thank the reviewer for constructive feedback and positive comments on paper writing, architecture efficiency, and extensive experiments. We address the potential concerns as follows.
>Certain technical details are glossed over. Some equations may not be elaborated in detail.
Thanks for bringin... | Summary: A diffusion-based approach for generating instance-level
counterfactual explanations and model-level explanations of GNN
predictions.
Strengths: The approach improves over existing counterfactual-based solutions by
allowing both addition and removal of nodes, thus better retaining
in-distribution of explanati... | Rebuttal 1:
Rebuttal: We thank the reviewer for the positive comment on the proposed model and extensive experiments. We sincerely appreciate the feedback and suggestions you provided, which help us to further improve the quality of this paper. We have addressed each point and potential concern you raised as follows.
>... | null | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Estimating Koopman operators with sketching to provably learn large scale dynamical systems | Accept (poster) | Summary: This paper utilizes theory developed around random projections to develop fast/efficient numerical methods for approximating the Koopman mode decomposition of data sets with long-time trajectories, that additionally come with learning rates that are optimal.
Strengths: 1. This paper developed new fast/effici... | Rebuttal 1:
Rebuttal: We would like to thank the reviewer for the helpful comments. We address the main ones below, while minor comments and typos will directly be fixed in the revised version.
**Motivation for long trajectories**
We stress that our contribution allows to tackle large datasets, which can in practice ... | Summary: This paper proposes sketching algorithms for approximating three different kernel-based estimators of Koopman operators from large scale temporal data. New non-asymptotic error bounds are derived.
Strengths: - The paper leverages the concept of sketching to speed up the estimation of Koopman operators withou... | Rebuttal 1:
Rebuttal: We would like to thank the anonymous reviewer for their evaluation of this paper, and for their kind comments. We will address the main questions below.
**Complexity with respect to the dimension of the state space?**
Yes, essentially the complexity with respect to state space dimensionality is ... | Summary: This paper improves the computational efficiency of existing kernel method for Koopman operator approximation, including KRR, PCR, RRR based on the Nystrom estimator. In this way, the authors reduce the original computational complexity from $\mathcal{O}(n^3)$ to $\mathcal{O}(n\sqrt{n})$. They provide theoret... | Rebuttal 1:
Rebuttal: We would first of all like to thank the reviewer for taking the time to read this paper and helping us improve its current state. We will address the main points raised below.
**Novelty of random projections & comparison to random projections for generative models and neural ODEs.**
Random proje... | Summary: The paper talks about estimating Koopman operators. Sketching (random projections), specifically the Nyström method is used as a computational tool to make things tractable. Furthermore, it is shown that these approximates have the same convergence rate as their slower counter parts. Further error bounds are p... | Rebuttal 1:
Rebuttal: We thank the reviewer for their helpful comments, which we address below.
**Novelty and comparison with refs. [3], [6] and kernel DMD.**
References [3] and [6] deal with the Nystrom KRR model, with [3] being more concerned with theoretical considerations (in a slightly different setting from us:... | Rebuttal 1:
Rebuttal: We thank all the anonymous reviewers for their thoughtful comments.
We provide here a pdf file with additional experimental results, and address directly the comments of the reviewers in the individual rebuttals.
In particular, the experiments concern the addition of Nystrom KRR to a comparison w... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Two-Stage Learning to Defer with Multiple Experts | Accept (poster) | Summary: This paper studies generalization bounds for learning to defer, with its main focus on "two-stage" algorithms that first fit a classifier using a surrogate classification loss, then separately fit a rejector that decides which expert, if any, to defer the classification to.
By and large, these two-stage algori... | Rebuttal 1:
Rebuttal: Thank you for your insightful comments. We have carefully addressed all the questions raised. Please find our responses below.
**1. The empirical results are missing baselines / comparisons to other surrogate methods in the literature.**
**Response:** To the best of our knowledge, our work is th... | Summary: This paper studies a two-stage scenario for learning to defer with multiple experts. For both the score-based setting and the predictor-rejector setting, the authors design a new family of surrogate loss function . The authors show that the surrogate losses are realizable H-consistent and Bayes-consistency of... | Rebuttal 1:
Rebuttal: Thank you for your appreciation of our work. We will take your suggestions into account when preparing the final version. Below please find responses to specific questions.
**Weaknesses: I am satisfied with the theoretical part. However, I would like to remark that I am not an expert in this fiel... | Summary: This paper provides bounds on the difference between the target loss and its surrogate loss, for two-stage learning to defer to multiple experts. The bounds guarantee that a small surrogate risk leads to a small target risk. The authors also show that in realizable cases, achieving the minimum surrogate risk r... | Rebuttal 1:
Rebuttal: Thank you for your appreciation of our work. We will take your suggestions into account when preparing the final version. Below please find responses to specific questions.
**Weaknesses:**
**1. Some technical parts were a little unclear. (See the questions below.)**
**Response:** Thank you for ... | Summary: The paper proposes two stage learning algorithms for learning to defer framework. Learning to defer framework consists of two components, a rejector and a classifier. Previous literature has focussed on jointly training both the classifier and the rejector. This paper, instead, focuses on the two stage setup i... | Rebuttal 1:
Rebuttal: Thank you for your encouraging review. We will take your suggestions into account when preparing the final version. Below please find responses to specific questions.
**Weaknesses:**
**1. While theoretical results is the main focus of this paper, I believe experimental section could have been mo... | null | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Machine learning detects terminal singularities | Accept (poster) | Summary: The authors trained a feedforward neural network to detect $\mathbb{Q}$-Fano varieties in dimension 8, among toric varieties defined by weights in a bounded region. The resulting classifier achieves a high (95%) accuracy. The authors used this classifier to build a large dataset of (probably) $\mathbb{Q}$-Fano... | Rebuttal 1:
Rebuttal: Q1: *It's not very clear whether, at the end of the day, the classifier built in this work is actually going to make an impactful contribution in the area. This weakness is outweighted by the strengths, in my opinion.*
A1: Understanding the classification of Q-Fano varieties is a fundamental open... | Summary: The authors study properties of algebraic geometric objects, four complex dimensional Fano toric varieties,
using mathematical data science methods, i.e. statistical analysis of a large database of examples combined
with mathematical domain knowledge. They find a fairly accurate neural network classifier to i... | Rebuttal 1:
Rebuttal: Q1: *The authors study properties of algebraic geometric objects, four complex dimensional Fano toric varieties…*
[Small correction: all our varieties are over the complex numbers, so these are 8-dimensional complex Fano toric varieties.]
*The paper is very unsystematic statistically. Since the... | Summary: This paper applies machine learning methods to predict whether an algebraic variety is Q-Fano. Q-fano varieties are Fano varieties that are Q-factorial and have terminal singularities. Checking terminally is computational challenging. To train a machine learning classifier, the authors build a balanced dataset... | Rebuttal 1:
Rebuttal: Q1: *…the authors could include more introductory materials, for example, in the appendix.*
A1: We agree that including additional introductory material in an appendix would be an excellent idea, and will do so in a future revision. The planned addition will cover three points. It will contextual... | Summary: The paper describes an ML assisted approach to the classification of Q-Fano varieties. Checking whether the singularities of a given Q-Factorial Toric variety is very challenging by mathematically established methods, involving intensive combinatorial computations. This paper applies supervised machine learnin... | Rebuttal 1:
Rebuttal: Q1: *The poor OOD performance is in my opinion the biggest obstacle to the claims of the paper, raising the troubling possibility that the better performance of the classifier comes from some 'finite size effect' due to the particular weight bound [...] Is it reasonable to expect that a NN classif... | Rebuttal 1:
Rebuttal: We thank the referees for their careful reading of the manuscript, and for their challenging and insightful questions. We have responded to each referee separately. In light of the referees' comments, we propose in a future revision to:
* add an expanded mathematical background section in the Supp... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
High dimensional, tabular deep learning with an auxiliary knowledge graph | Accept (poster) | Summary: The authors present a method to leverage prior knowledge in the form of a knowledge graph for tabular deep learning, illustrating performance on a number of biological prediction problems in which features correspond to gene-level measurements. In this case, the knowledge graph is a biomedical knowledge graph ... | Rebuttal 1:
Rebuttal:
> **Overall response**
We **thank the reviewer** for their excellent review! The reviewer states that PLATO is **“non-obvious and effective”** while also being **“intuitive and clearly explained.”** The reviewer highlights PLATO’s novelty in “[combining] ideas from knowledge graph embedding and... | Summary: The paper introduces PLATO, a novel machine learning approach for high-dimensional tabular data, where the number of features far exceeds the number of samples (d>>n). This scenario is often encountered in fields like physical sciences and biology. PLATO uses a knowledge graph (KG) to train the first layer of ... | Rebuttal 1:
Rebuttal:
> **Overall response**
We thank the reviewer for pointing out that **“the mechanism to induce KG regularization is novel.”** The reviewer further notes that our study’s **“setting matters especially in a scientific context”** and that our **“general method for knowledge graph integration could b... | Summary: The paper explores tabular ML tasks in scenarios where data has more features than samples and auxiliary structured data is available in the form of knowledge graphs. One of the central assumptions is that each feature corresponds to a node in the knowledge graph. A model called PLATO is proposed which combine... | Rebuttal 1:
Rebuttal: > **Overall response**
We thank the reviewer for their **incredibly thoughtful review**! The reviewer says that the “**presentation is excellent**,” and the idea of “leveraging auxiliary KGs to aid in tabular data tasks” is “**interesting and useful.**”
---
> **The first MLP layer is a linear ... | Summary: This paper introduces PLATO, an approach handles tabular deep learning challenges when the number of features (d) significantly outnumbers the available instances (n). The authors propose the integration of an auxiliary knowledge graph into the learning process to help the model better understand the relations... | Rebuttal 1:
Rebuttal:
> **Overall response**
We **thank** the reviewer! The reviewer highlights our “**original approach to** a longstanding issue in machine learning: **the ‘curse of dimensionality’.”** The reviewer says PLATO is a “**potentially transformative technique** to address [the] challenge of high-dimensi... | Rebuttal 1:
Rebuttal: > **Overall response**
We are extremely grateful for the highly detailed and constructive feedback on our manuscript! The reviewers describe PLATO as a **“potentially transformative technique to address [the] challenge of high-dimensionality in tabular data”** (E2W5), say that **“the mechanism t... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Domain Adaptive Imitation Learning with Visual Observation | Accept (poster) | Summary: This paper aims to solve the domain-adaptive imitation learning problem, where the demonstrations are collected in a source domain, and the agent needs to learn policies in a target domain. Beyond that, the tasks are with high-dimensional visual observations. The authors have proposed a new framework based on ... | Rebuttal 1:
Rebuttal: Thank you for your valuable feedback.
**1. Contribution of this paper**
In summary, our contribution is that we propose a novel feature extraction model in the domain-adaptive IL with visual observations. We significantly improved the quality of behavior feature extraction by effectively elimina... | Summary: Imitation learning from visual observations for domain adaptation. Given expert and non expert demonstrations in a source domain, and non expert demonstrations in a target domain, learn a reward and policy in the target domain via image consistency and domain and behavior feature consistency.
Strengths: Domai... | Rebuttal 1:
Rebuttal: Thank you for your valuable feedback.
**1. Assumptions about the relation between the source and target domains**
It may be difficult to express the general assumption about the relation between domains clearly and mathematically. But, we think that the general assumption about the relation betw... | Summary: The paper attempts to solve the problem of domain shift in Imitation Learning with image observation. The author proposes D3IL architecture which is based on dual feature extraction (behavior and domain specific features from input) and dual cycle consistency (image level and feature level cycle consistency). ... | Rebuttal 1:
Rebuttal: Thank you for your valuable feedback.
**1. Experiment on more challenging domain mismatch**
As the reviewer mentioned, our method was evaluated primarily on simulated environments. During this short rebuttal period, we could not apply our method to real-world scenarios, but to assess the appli... | Summary: This work introduces a method of learning a visual behavior encoding for reinforcement learning tasks, D3IL.
From a TPIL behavior feature encoder baseline, D3IL introduces dual feature extraction (including both behavior and domain feature encoders) and dual cycle-consistency (cycle consistency on image and f... | Rebuttal 1:
Rebuttal: Thank you for your valuable feedback.
**1. Answer to Quality section**
We claim that learned domain encoders are essential for enhancing the feature extraction and policy performance in the domain-adaptive IL problem with visual observations. To demonstrate this, we conducted an experiment wher... | Rebuttal 1:
Rebuttal: Thank you for your valuable feedback. In the Common Response, we address inquiries from multiple reviewers and provide responses requiring images and plots. Please refer to the attached PDF for detailed explanations.
**Experiment on more challenging domain mismatch**
To assess the effectiveness ... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
CORNN: Convex optimization of recurrent neural networks for rapid inference of neural dynamics | Accept (poster) | Summary: The motivation behind this paper is to replace FORCE/RLS fitting of biologically-flavored RNNs to data by a faster and more robust alternative training scheme. This is accomplished by 1) altering the training objective to a cross-entropy loss; 2) using an approximate pre-computable Hessian to facilitate the us... | Rebuttal 1:
Rebuttal: We thank the Reviewer for their expert review, careful reading, and identification of typos, which we have now fixed. We especially appreciated the expert parsing of our approach into 5 distinct components; this showcased how well the Reviewer understood and appreciated our research. With the Revi... | Summary: The paper introduces a new method for training RNNs that allows fast convergence which could make realtime parameter update possible.
Strengths: - The method conceptually is simple and intuitive.
- It tackles an important problem which is the time consuming nature and computing intensivity of ML models
- It a... | Rebuttal 1:
Rebuttal: We thank the Reviewer for their careful reading of our manuscript. The suggestion to better tie our results back to Figure 1 is an excellent one, which we implemented. You can find below our detailed, point-by-point responses to each question raised.
- any equation in the main text has to be defi... | Summary: The paper proposes a convex loss for training data constrained recurrent neural networks (dRNNs). This is achieved by performing the optimization in the $d_{t,i}$ space instead of $r_{t,i}$ space (as defined in the paper) and using cross-entropy instead of $l_2$ loss. The authors discuss that this choice makes... | Rebuttal 1:
Rebuttal: We appreciate the time the Reviewer put into this detailed review. As they noted, developing tools for real-time experiments and targeted intervention of neural activity is a significant and timely problem. The reviewer also commented that the motivation was clear, logical, and easy to follow. In ... | Summary: The authors introduce a method to produce a digital twin of a biological neural network that can be inferred from measurements of neuronal activity. The major advantage of the model is that it can be trained in real time to gain the basic dynamics of neural population. In the future this could allow advance H... | Rebuttal 1:
Rebuttal: We thank the Reviewer for their comments and endorsement of our work. We carefully reviewed each comment and have provided below our responses detailing how we changed the manuscript to incorporate the Referees’ suggestions. We would also greatly appreciate it if the Reviewer would like to propose... | Rebuttal 1:
Rebuttal: We thank the AC for considering our work and all reviewers for their expert reports, which we used to refine the presentation of our manuscript and add new experiments. We note that two reviewers thought the paper had excellent impact in its area and one thought there were no major concerns. The c... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
iSCAN: Identifying Causal Mechanism Shifts among Nonlinear Additive Noise Models | Accept (poster) | Summary: This paper focuses on identifying mechanism shifts in related structural causal models without learning the structures of the individual SCMs. It builds on the recent advances in score-matching based causal discovery and consider nonlinear additive noise models, unlike the prior work that assumed linear Gaussi... | Rebuttal 1:
Rebuttal: Thanks for your valuable feedback! We next respond to your questions, and hope our answers will make you feel more positive about our work. Please do not hesitate to ask any follow-up questions.
### Response to comments from Weaknesses section
> "Invariant noise assumption (Assumption B) and shi... | Summary: The authors propose a two-step algorithm for identifying causal mechanism shifts. The algorithm is motivated by a characterization of the nodes with mechanism shifts and the procedure is based on several existing tools. The proposed algorithm outperforms the baseline.
Strengths: 1. Theorem 1 provides a new pe... | Rebuttal 1:
Rebuttal: Thanks for your comments! We next respond to your concerns, and hope our answers will make you feel more positive about our work. Please do not hesitate to ask any follow-up questions.
> Assumption B (invariant noise distribution) makes the setting quite restrictive. There is no discussion or exp... | Summary: This work addresses the problem of estimating differences in the causal models in two (or more) environments, over a single set of variables with the same causal ordering. The authors are interested in discovering those variables that undergo a functional or structural mechanism shift. As their main insight, t... | Rebuttal 1:
Rebuttal: Thank you for your valuable input. We have new experiments on additional settings, and other baselines, as per your suggestions. Please see response to all reviewers.
### Response to comments from Weaknesses section
> "Novelty: This work follows ..."
We firmly disagree with the statement that “... | Summary: The paper proposes an efficient method to identify causal mechanism shifts across various interventional environments. The method's objective is to identify changes not only in the mechanism itself but also in the causal structure across the environments, which has been evaluated in artificial and real-world d... | Rebuttal 1:
Rebuttal: Thanks for your valuable feedback! We next respond to your questions, and please do not hesitate to ask follow-up questions.
### Response to comments from Weaknesses section
Next we only respond to the two points that are not repeated in the Questions section.
> "The overall complexity is relat... | Rebuttal 1:
Rebuttal: We thank all reviewers for their time and effort put into our work. We truly appreciate your feedback and we are happy to see that your comments will lead to a stronger version of our manuscript. Overall, there is a positive feeling about the significance of the problem, our contributions, and pre... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Transition-constant Normalization for Image Enhancement | Accept (spotlight) | Summary: This paper presents a novel Transition-Constant Normalization (TCN) for various image enhancement tasks. It has several formats and is plug-and-play for existing image enhancement networks. In the experimental part, proposed method show the extensibility and superiority of the proposed method on both quantitat... | Rebuttal 1:
Rebuttal: Thanks for your time and efforts and here we present the rebuttal.
**1. Discussion with BN.**
Firstly, BN is a kind of normalization that has the best capability of keeping information representation property [1]. Therefore, BN can transit almost all information and thus can be employed for som... | Summary: The authors introduce a series of normalization formats by incorporating the transition-constant constraint design. The motivation behind the work is presented, extensive experiments across multiple image enhancement tasks are carried out.
Strengths: i. The idea is interesting. With the introducing of invert... | Rebuttal 1:
Rebuttal: Thanks for your time and efforts and here we present the rebuttal.
**1. About adding a residual for normalization.**
In fact, adding a residual path is not a reasonable manner: Firstly, it adds the original information to the normalized one directly, which does not consider their relationship ... | Summary: This paper proposes an invertible feature normalization module based on the design of coupling layers in flow-based models [1]. The authors experimentally evaluate the proposed model on various image enhancement tasks such as low-light image enhancement, HDR, image dehazing, as well as medical segmentation and... | Rebuttal 1:
Rebuttal: Thanks for your time and efforts and here we present the rebuttal.
**1. About motivation and contribution.**
The meaningful of our work: (1) Since IN can normalize different lightness effectively, we have shown that equipping IN as TCN with transition constant ability is useful for image enhance... | Summary: In this paper, a transition-constant normalization (TCN) is introduced to harness the potential of applying normalization techniques for image enhancement. The TCN is designed with two rules, adhering to the principle of invertible information transmission and dividing the features into two streams with consis... | Rebuttal 1:
Rebuttal: Thanks for your time and efforts and here we present the rebuttal.
**1. Would the TCN’s ability limited?**
The answer is no and here are the reasons: Firstly, the partial normalized features are integrated into the original one evenly in each channel, and thus each channel feature of the TCN’s o... | null | NeurIPS_2023_submissions_huggingface | 2,023 | Summary: 1. A new normalization technique is introduced that transits information constantly for image enhancement.
2. The introduced TCN covers a range of formats and is compatible with existing method.
3. The introduced technique exhibits extensibility to various scenes and tasks through experiments.
Strengths: It i... | Rebuttal 1:
Rebuttal: Thanks for your time and efforts and here we present the rebuttal.
**1. About the relationship of transition-constant ability and image enhancement.**
First, for architectures such as CNN and transformer, they convey information with the driven of reconstruction losses, thus the whole part of th... | null | null | null | null | null | null |
$L_2$-Uniform Stability of Randomized Learning Algorithms: Sharper Generalization Bounds and Confidence Boosting | Accept (poster) | Summary: This paper considers generalization bounds for SGD with a notion of randomization on learning algorithms. The paper considers a $L_2$-uniform stability framework and establishes generalization inequalities under such a framework. Additional algorithms and analysis using confidence-boosting are discussed.
Stre... | Rebuttal 1:
Rebuttal: Thank you for your insightful review and detailed comments for improvement. We sincerely hope that the main concerns raised in the review can be clarified by the following responses.
> **Your comment:** It may be helpful to add in formal definitions and clear up typos/confusions.
**Our respons... | Summary: This paper develops new and potentially sharper (as compared to the state of the art) data and randomness dependent (high probability type) generalization bounds using a kind of new and more restrictive notion of algorithmic stability, which the authors call L2-uniform stability. The focus is in particular on ... | Rebuttal 1:
Rebuttal: Thank you for your insightful review and encouranging comments.
> **Your comment:** By the way, are the results presented in Sections 3/4 the first ones on the generalization of that type of algorithms?
**Our response:** Yes, we for the first time establish in Theorem 2 a generalization bound ... | Summary: Stability is a useful tool for obtaining sharper generalization bounds in learning theory. In prior work, the notion of "on-average" stability is as follows. A randomized learning algorithm is stable if for all datasets $S$, adjacent datasets $S'$ (i.e. differing by a single point), and all data points $Z$, th... | Rebuttal 1:
Rebuttal: Thank you for your insightful review and positive evaluation of our work.
> **Your comment:** I think that the subtly of the contribution of this paper needs to be spelled out more clearly.
**Our response:** Thanks for your concise summary of the crucial differences between our first-moment ge... | Summary: Algorithmic stability refers to the sensitivity of learning algorithms to the training set. Roughly speaking, if a learning algorithm is more stable, then its generalization gap (i.e., its training loss is highly indicative of its population loss) can be better controlled. In the literature, various notions of... | Rebuttal 1:
Rebuttal: Thank you for your insightful review and appreciation of our work.
> **Your comment:** Is it possible, perhaps under additional assumptions, to get exponentially high probability generalization bounds from on-average uniform stability? What are some technical barriers that stand in the way of boo... | null | NeurIPS_2023_submissions_huggingface | 2,023 | Summary: A new notion of algorithmic stability is introduced for randomized algorithms, which is intermediate to previously studied notions and leads to new strong generalization bounds. The need for a new notion is justified due to known lower bounds under previously studied notions by related to on-average stability,... | Rebuttal 1:
Rebuttal: Thank you for your insightful review and appreciation of our work.
> **Your comment:** No lower bounds are provided, so tightness of the results is not clear.
>
**Our response:** Per your comment, we plan to provide some additional analysis regarding tightness of our bounds in the next version... | null | null | null | null | null | null |
Smooth Flipping Probability for Differential Private Sign Random Projection Methods | Accept (poster) | Summary: This paper develops a series of differentially private (DP) Random Projection and Signed Random Projection algorithms. By using randomized responses and the idea of smooth sensitivity, the new proposed algorithms are shown to outperform previous results while still maintaining the same amount of privacy guaran... | Rebuttal 1:
Rebuttal: Dear Reviewer hnU4:
Thanks for your feedback. We appreciate your interests in our work and the ideas behind the algorithms. We also hope that the theoretical and empirical results in our paper (the adoption of optimal Gaussian mechanism and the smooth flipping mechanism) could be helpful to the c... | Summary: The authors study the important problem of preserving privacy in matrix data. They study a family of fundamental operations on data which is Random Projection. This is useful for dimensionality reduction, NN search etc. Based on RP one can also obtained signed RP (SRP) which are useful (as in SimHash) to find... | Rebuttal 1:
Rebuttal: Dear Reviewer TuXU:
We appreciate your encouraging feedback on our theoretical and empirical results, and the significance of the DP sketching problem. We too believe that DP sketching methods (e.g., the RP-based algorithms in our paper) are interesting, important, and could have broad applicatio... | Summary: The paper introduces a series of differential privacy (DP) random projections (RP) algorithms based on the robustness of the sign flipping probability in RP. This approach is promising as it seeks to enhance the existing methods in both theoretical guarantees and empirical performance.
Strengths: 1. The anal... | Rebuttal 1:
Rebuttal: Dear Reviewer bvWd:
Thanks for your feedback on our paper.
As we mentioned at line 195, The concrete exponentially decaying form of the flipping probabilities in Algorithm 4 is leads to the "smoothness" of the flipping probability: for any neighboring $u$ and $u'$, $P^{(u)}\leq e^{\epsilon/k}P^{... | Summary: This paper studies the problem of computing differentially private sketches of high dimensional vectors via random projections. The contributions are:
1. Algorithm DP-RP-G-OPT: An algorithm that uses the Gaussian mechanism to privatize random projections using Gaussian and Rademacher matrices, but adds more ca... | Rebuttal 1:
Rebuttal: Dear Reviewer ywEG:
Thanks for your feedback on our submission.
W1 and Q1. As Reviewer mentioned, the proposed private sketching algorithms are designed to protect attribute-level privacy. That is, our method prevents the identification of a change in any element in the vector in the sense of D... | null | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Facing Off World Model Backbones: RNNs, Transformers, and S4 | Accept (poster) | Summary: This paper presents a novel architecture for world-models in deep RL, which is based on S4, a generative sequence model. Furthermore, the paper presents an evaluation to two other world-model architectures: RSSM's and transformer-based world-models.
Strengths: - Integrating more powerful models into existing ... | Rebuttal 1:
Title: Response to Reviewer kwZk
Comment: We sincerely appreciate your valuable insights! We hope we have addressed your concerns about incremental contribution, discussion about TECO, and evaluation with MBRL agent in our "global" response. We address your remaining concerns below. Please let us know if yo... | Summary: Model-Based RL algorithms rely on learning efficient world models for learning optimal behaviors. Existing agents like Dreamer rely on an RNN for learning world models which have limited memory capacity. S4 has shown promising results when compared with RNNs and Transformers in terms of both complexity and eff... | Rebuttal 1:
Title: Response to Reviewer vwzx
Comment: We sincerely thank you for your constructive feedback! We hope we have addressed your concerns about baseline comparison in our "global" response. S4WM is significantly better than Clockwork VAE on DMLab in the TECO benchmark. S4WM is also significantly better than ... | Summary: This paper studies the backbones of world models. An S4-based world model S4WM is proposed by replacing RNNs in RSSM with S4 Blocks, which is known to be good at long-term memory. An empirical study has been conducted to compare three major backbones, RNN, Transformers, and S4, on speed, memory usage, and perf... | Rebuttal 1:
Comment: I thank the author for their response. I appreciate the effort made on the preliminary evaluation of S4WM for MBRL (planning).
However, I respectfully ask for a further discussion on Weakness 1 above.
---
Rebuttal 2:
Title: Response to Reviewer TTTx
Comment: We would like to express our apprecia... | Summary: This paper introduces S4WM, the first world model based on Structured State Space Sequence (S4) models for MBRL agents. Specifically, the proposed world model leverages a stack of S4 blocks for temporal prediction in latent space. CNN encoder and decoder are used to map the image to the latent space. Compared ... | Rebuttal 1:
Title: Response to Reviewer ALhP
Comment: We sincerely appreciate your valuable feedback! We hope we have addressed your concerns about incremental contribution and evaluation with MBRL agent in our "global" response. We address your remaining concerns below. Please let us know if you have additional commen... | Rebuttal 1:
Rebuttal: We thank all reviewers for their insightful and constructive feedback. We are encouraged that they appreciate our contribution of introducing S4 into the field of world models to improve long-term memory capacity (@TTTx, @vwzx). Many reviewers praised our thorough and high-quality experiments that... | NeurIPS_2023_submissions_huggingface | 2,023 | Summary: This work aims to better model based RL agents by focusing on developing robust world models. They propose S4WM which uses the S4 model as backbone and incorporates actions as context in the hidden state. They evaluate on prediction tasks for image sequences through latent imagination and show experimental res... | Rebuttal 1:
Title: Response to Reviewer hTvu (1/2)
Comment: We sincerely thank you for your positive feedback! We hope we have addressed your concerns about baseline comparison in our "global" response. S4WM is significantly better than Clockwork VAE on DMLab in the TECO benchmark. S4WM is also significantly better tha... | null | null | null | null | null | null |
TaskMet: Task-driven Metric Learning for Model Learning | Accept (poster) | Summary: The authors propose using task information to modify a metric that the prediction model uses for training and include task information. This way, they encourage the model to generate more accurate predictions specifically for inputs relevant to the downstream task.
Strengths: - The manuscript is well-structu... | Rebuttal 1:
Rebuttal: Thank you for the review of our paper and detailed comments. We are happy to hear you enjoyed the paper. Thanks also for pointing out the other minor problems about the typos and figures. We will correct them. Here are some clarifications to your comments, please let us know if there is anything ... | Summary: The authors propose an approach to loss re-shaping for a supervised prediction model with the goal of improving the performance of a larger system that performs a downstream task using the predictions of the prediction model. Such shaping is necessary because reducing loss on the prediction task may only rough... | Rebuttal 1:
Rebuttal: Thank you for the detailed review and comments on our paper. We were surprised by your decision to reject the paper based on the clarity and experimental weaknesses, as we have evaluated the idea of learning a metric on standard experimental settings in the most recent DFL and end-to-end model lea... | Summary: The paper proposes a novel method, called TaskMet, for learning a metric that can be used to improve the performance of prediction models on downstream tasks. The key idea is to use the task loss to guide the learning of the metric. This is done by using a gradient-based approach to optimize the metric paramet... | Rebuttal 1:
Rebuttal: Thanks for the review of our paper. We’re happy you found the paper well-written, well-motivated, and easy to follow. Here are some clarifications about the hyperparameters and baselines. To further help with reproducibility in the field we will also publish the entire source code necessary to rep... | Summary: This paper presents a method for end-to-end learning of metrics to train prediction models. It proposes a concept of task-driven metric learning. The main idea is to let the model focus on task-relevant features and dimensions in the prediction space. In addition, the resulting prediction model can be more int... | Rebuttal 1:
Rebuttal: Thanks very much for the review! We are delighted you liked the idea and found the results promising. Here are a few responses to your comments, please let us know if there is anything else:
> The writing of the methodology is a bit difficult to understand. It would be great if the writing can be... | Rebuttal 1:
Rebuttal: We extend our gratitude to the reviewers for investing their time and expertise into our paper. The feedback provided has been immensely valuable, and we greatly appreciate the positive remarks regarding the novelty and impact of our proposed method. While we include more detailed responses inline... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
KAKURENBO: Adaptively Hiding Samples in Deep Neural Network Training | Accept (poster) | Summary: The paper introduces an approach to reduce training time of deep neural
networks on large datasets by excluding samples on an epoch basis with a
minimal reduction of the accuracy. In each epoch, a scheduled fraction of the
samples with the lowest loss is excluded, unless its prediction is incorrect
or its soft... | Rebuttal 1:
Rebuttal: **5.1. Single setup per trial issue:**
We confirm that the results of our implementation are stable by running experiments 3 times for CIFAR-100 and ImangetNet-1K. The new result is reported in Table 1 (extra-result in the rebuttal). As for our original results, due to resource limitation, we cou... | Summary: The paper proposes to accelerate DNN training by dynamically selecting a subset of the training set. The method records the per-sample loss during training, and hides (a heuristically set proportion of) samples with the least loss and correct prediction, then trains the model with an adjusted learning rate. Th... | Rebuttal 1:
Rebuttal: **4.1. Discussion on the hyperparameters ($\tau$, $F_e$):** Higher prediction confidence threshold $\tau$ leads to a higher number of samples moved back, i.e., fewer hidden samples at the beginning of the training process. At the end of the training process, when the model is well-trained, more sa... | Summary: This paper proposes a framework called KAKURENBO to accelerate DNN training by adaptively hiding samples. In the proposed framework, sample of low loss are hidden (i.e, not used) in the next training epoch, so that the total training steps are reduced. To compensate for the reduction in optimization steps, the... | Rebuttal 1:
Rebuttal: **3.1. The hidden factor F needs to be manually adjusted based on different datasets and training schedules**. We consider F as a hyperparameter, similarly to other methods in this domain. In our method, F is the maximum fraction, which can be adapted and changed during the training process.
**3.... | Summary: This paper introduces a method called KAKURENBO to increase the efficiency of training deep neural networks by reducing sample complexity while maintaining performance accuracy. The method encourages adaptive sample selection for each training epoch based on empirical criteria such as prediction accuracy (PA),... | Rebuttal 1:
Rebuttal: **More complexity using a single GPU:**
It is worth noting that KAKURENBO has computation overheads for updating the loss and prediction (Step D in Figure 1 in our main manuscript) and sorting the samples based on the loss (Step A in Figure 1 in our main manuscript). The performance gain from hid... | Rebuttal 1:
Rebuttal: We thank all reviewers for their constructive comments and questions.
**1. Single setup per trial issue (Reviewer Zijr)**
We confirm that the results of our implementation are stable by running experiments 3 times for CIFAR-100 and ImangetNet-1K. The new result is reported in Table 1 (extra-resul... | NeurIPS_2023_submissions_huggingface | 2,023 | Summary:
Summary: This paper proposes a method of dynamically selecting training data during the training process to improve the efficiency of the model training. Specifically, the authors optimize the training process by dynamically adjusting sample weights, learning rates, and the upper limit of sample numbers. Whil... | Rebuttal 1:
Rebuttal: **1.1. Difference between this paper and data distillation**
Data distillation reportedly introduces high overhead and is thus used mainly as a means to improve the model's generalization ability, and not reducing the amount of work. In comparison, the work in this paper aims to reduce the work by... | null | null | null | null | null | null |
Guide Your Agent with Adaptive Multimodal Rewards | Accept (poster) | Summary: The paper proposes an interesting and novel idea of using large pretrained mulitmodal models to compute image-text alignment
score to use as a reward to train return-conditioned policies using a Decision Transformer (DT). They propose a novel
imitation learning framework called Multimodal Reward Decision Tra... | Rebuttal 1:
Rebuttal: Dear Reviewer HCX5,
We sincerely appreciate your valuable comments. We found them extremely helpful in improving our draft. We address each comment in detail, one by one below.
---
**[W1, L1] Limited experiments.**
In response to your concern, we have investigated the efficacy of our method wi... | Summary: The authors propose a formulation of the reward-conditioned decision transformer but do so using a learned reward function estimated by a pretrained CLIP style model which maps expert demonstrations and textual descriptions of these demonstrations into the same domain. They evaluate their method on the coinrun... | Rebuttal 1:
Rebuttal: Dear Reviewer nhjK,
We sincerely appreciate your valuable comments. We found them extremely helpful in improving our draft. We address each comment in detail, one by one below.
---
**[W1] Clarification of Figure 1 in the submitted draft.**
Figure 1 in the submitted draft presents an illustrati... | Summary: This paper presents a framework called Multimodal Reward Decision Transformer (MRDT) that uses the visual-text alignment score from pre-trained vision-language models (after careful fine-tuning using the in-domain data) as the reward signals in visual-based reinforcement learning. Specifically, the authors pro... | Rebuttal 1:
Rebuttal: Dear Reviewer 9gB5,
We sincerely appreciate your valuable comments. We found them extremely helpful in improving our draft. We address each comment in detail, one by one below.
---
**[W1] Limited experiments.**
In response to your concern, we have investigated the efficacy of our method within... | Summary: This work tackles the goal misgeneralization problem in goal-conditioned RL agent setups. They propose to leverage pre-trained multimodal models, i.e., CLIP, to serve as a reward function that simultaneously estimate both text and image modalities for deciding an engineered reward for each time step. The autho... | Rebuttal 1:
Rebuttal: Dear Reviewer qrbz,
We sincerely appreciate your valuable comments. We found them extremely helpful in improving our draft. We address each comment in detail, one by one below.
---
**[W1] Effect of pre-trained models.**
To verify the effectiveness of pre-trained multimodal representations, we ... | Rebuttal 1:
Rebuttal: Dear reviewers,
We sincerely appreciate your time and effort in reviewing our manuscript. As the reviewers highlighted, our work proposes a simple but effective framework (qrbz, HCX5, 9gB5) for imitation learning with a novel adaptive reward that leverages pre-trained multimodal representations (... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
On the Pareto Front of Multilingual Neural Machine Translation | Accept (poster) | Summary: This paper frames the sampling ratio selection problem of MNMT as an optimization problem, and introduces a double power law to control the sampling ratio for different languages. The author argue that traditional method assumes unlimited training data which does not hold true for low resource settings, thus b... | Rebuttal 1:
Rebuttal: Dear reviewer, we really appreciate your effort in reviewing our paper, and thank you for your helpful comments. We are glad that you recognize our work 1) addresses an important problem in this direction 2) with good organization of the paper 3) solves the problem with a well-grounded method. We ... | Summary: This paper explores the influence of language pair sampling weights on the ultimate performance of Multilingual Neural Machine Translation (MNMT), specifically through the lens of the Pareto front concept. Existing literature suggests that under certain conditions, diverse sampling weights can yield solutions ... | Rebuttal 1:
Rebuttal: Dear reviewer, we really appreciate your effort in reviewing our paper, and thank you for your helpful comments. We are glad that you recognize our work 1) delves into a more realistic scenario of MNMT 2) provides an insightful explanation for the collapse 3) The DPL is effective in predicting the... | Summary: This paper investigates the multi-task scaling laws for multilingual machine translation (mNMT) in the presence of *limited data* for at least one of the tasks/language pairs.
Extending Fernandes et. al. (2022) work on the data-rich scenario, the authors found that their proposed multi-task scaling law doesn'... | Rebuttal 1:
Rebuttal: Dear reviewer, we really appreciate your effort in reviewing our paper, and thank you for your helpful comments. We are glad that you recognize our work 1) having the potential to have a lot of practical impact in balancing languages/tasks during the training of models 2) analysis relating sharpne... | Summary: This paper proposes to model the scaling law of multilingual machine translation using the double power law, which considers both the sampling ratio of a language and the size of the corpus. The motivation is that simply increasing the sampling ratio of low-resource language might lead to decreased performance... | Rebuttal 1:
Rebuttal: Dear reviewer, we really appreciate your effort in reviewing our paper, and thank you for your helpful comments. We are glad that you recognize our analysis of Pareto front well motivated and our method effective. We address your questions below:
> W1/W2/Q1: The experiment is restricted to maximu... | Rebuttal 1:
Rebuttal: # General Response for all Reviewers
Dear Reviewers and ACs,
We sincerely appreciate the reviewers' effort in reviewing our paper and the AC for reviewing and organizing discussions! We are encouraged to hear that our analysis of Pareto front in MNMT very interesting (R1,R3), well motivated(R2),... | NeurIPS_2023_submissions_huggingface | 2,023 | Summary: This work proposes a power law for multilingual machine translation, a magic formula and constants to predict the performance of multilingual machine translation, that trades the model capacity and generalization based on the number of parameters and the training data sizes. The law is derived from the observa... | Rebuttal 1:
Rebuttal: Dear reviewer, we really appreciate your effort in reviewing our paper, and thank you for your helpful comments. We are glad that you recognize our analysis of Pareto front very interesting, the simplicity of the derived power law and the effectiveness of its application. We address your questions... | null | null | null | null | null | null |
CSLP-AE: A Contrastive Split-Latent Permutation Autoencoder Framework for Zero-Shot Electroencephalography Signal Conversion | Accept (poster) | Summary: This research paper introduces a novel framework called the contrastive split-latent permutation autoencoder (CSLP-AE) for EEG signal conversion, with the primary aim of disentangling task-related content and subject-specific style. This technique is based on disentangled representation learning which has been... | Rebuttal 1:
Rebuttal: We thank the reviewer for the careful and constructive critique acknowledging that the proposed approach is an innovative contribution to the field with visual clarity, detailed and coherent explanations, and a comprehensive comparative analysis of the proposed method.
### Re. Clear and comprehens... | Summary: Electroencephalography (EEG), a non-invasive neuroimaging technique, is often compromised by high noise levels and subject variability, making consistent signal extraction challenging. To mitigate this, the authors introduce a novel framework, the Contrastive Split-Latent Permutation Autoencoder (CSLP-AE). Thi... | Rebuttal 1:
Rebuttal: We thank the reviewer for the careful assessment of the manuscript acknowledging that the paper is innovative in being among the first to separate style and content in EEG signals providing favourable performance and that it may have potential for wider adoption within biomedical signal processing... | Summary: The authors hypothesize that converting EEG signals between tasks and subjects requires extracting latent representations that capture both content and style. They introduce a novel framework called CSLP-AE, which utilizes contrastive learning to guide the latent representations and achieve EEG conversion. The... | Rebuttal 1:
Rebuttal: We thank the reviewer for the careful review and constructive comments and for acknowledging that the paper uniquely frames a voice conversion problem as an EEG analysis problem presenting a new framework that leverages state-of-the-art deep learning models.
### Re. Limited generalizability and la... | Summary: The paper proposes the use of swappable auto-encoder embedding along with contrastive learning in the embedding space to improve the performance of subject and task classification on evoked potentials in EEGs beyond using just an auto-encoder with contrastive learning in the embedding space. The supervised per... | Rebuttal 1:
Rebuttal: We thank the reviewer for the constructive comments and careful assessment finding that the paper is well written and clear, results show consistent improvements on a large dataset (for EEG), and that the discussions are both extensive and insightful.
### Re. Generality beyond ERPs.
We include two... | Rebuttal 1:
Rebuttal: We thank the reviewers for their efforts in reviewing this manuscript and providing constructive and insightful comments. Individual responses to each reviewer’s questions and comments have been prepared and are provided below where appropriate.
Multiple reviewers address the fact that results h... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Factorized Tensor Networks for Multi-Task and Multi-Domain Learning | Reject | Summary: This paper proposes a factorized tensor network (FTN), that adds task/domain-specific low-rank tensors to shared weights, to perform lightweight MTL/MDL. Specifically, the low-rank tensors can form a weight additive to the shared backbone weight so that each task can have its own parameters to achieve good acc... | Rebuttal 1:
Rebuttal: We would like to first thank the reviewer for the detailed analysis on our method and pointing out the concerns with missing references and experiments. We discuss all of them in detail below.
1. **Re method**
- **FTN and MTL/MDL methods.**\
First, we would like to mention that many of t... | Summary: The paper presents the Factorized Tensor Network (FTN) as a solution to the challenge of learning multiple tasks/domains using a single unified network. The authors claim that FTN achieves comparable accuracy to independent single-task/domain networks with a smaller number of additional parameters. The method ... | Rebuttal 1:
Rebuttal:
We would like to first thank you for your detailed and insightful comments. We appreciate that you listed the strengths of our paper and pointed out the efficiency of our approach in terms of minimum parameter requirement without losing accuracy. We also find it interesting and appropriate that y... | Summary: The paper tackles multi-task learning (MTL) and multi-domain learning (MDL). The author proposes to a general method to add task/domain-specific weights to the common shared backbone. Specifically, they add the factorized task/domain specific tensors which introduce very low storage cost for each task and doma... | Rebuttal 1:
Rebuttal: We would like to first thank you for your detailed and insightful comments and we are glad that you were able to follow the paper easily. While we acknowledge that models like BigGAN may be considered old-fashioned, we believe they still serve as a good baseline for image generation tasks. Our mai... | Summary: The paper tackles the problem of adapting a single pretrained network to multiple tasks or domains while introducing as little parameters as possible. More specifically, first a backbone model is trained which contains a shared backbone. Then, the proposed method (Factorized Tensor Networks, **FTN**) is appli... | Rebuttal 1:
Rebuttal: We would like to first thank you for your detailed and insightful comments and nicely summarizing the contributions and strengths of our paper. Below we provide our response to your comments and questions.
****
1. **Re missing related work (LoRA):** \
Thank you for pointing out this paper. We ... | Rebuttal 1:
Rebuttal:
We thank all the reviewers for their careful, insightful, and constructive comments. We very much appreciate that reviewers recognized our work as a simple and effective strategy that leads to significant improvements. Our proposed factorized tensor network (FTN) can achieve accuracy comparable t... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Pick-a-Pic: An Open Dataset of User Preferences for Text-to-Image Generation | Accept (poster) | Summary: This paper introduces Pick-a-pic, a large and open dataset of human preference over model-generated images. Using this dataset, a scoring function (PickScore) based on CLIP is trained. The paper finds that PickScore has better correlation with real user preferences compare to existing scoring methods. Furtherm... | Rebuttal 1:
Rebuttal: Weaknesses:
1. It would be good if the paper could provide a more comprehensive analysis of the prompts used by the users in Pick-a-pic. How diverse are the prompt? What is the word distribution difference between Pick-a-pic and existing datasets such as COCO or DreamBench?
Answer: We agree that ... | Summary: This paper introduces a dataset of text-to-image prompts, along with user preferences over the generated images. This is done by building a web application where multiple images are generated for a given prompt, and the user is asked to select their preferred image. As a result of collecting this Pick-a-Pic da... | Rebuttal 1:
Rebuttal: Questions:
1. My only major question is regarding the "superhuman" performance achieved by the PickScore model. Given that the ground truth is captured from human preferences, I do not fully understand what achieving superhuman performance on this benchmark means (and the description in the paper... | Summary: The study gathers a dataset of user preferences by soliciting human choices regarding more preferred images in response to text prompts given a text prompt. Besides, the authors propose a metric called PickScore by fine-tuning CLIP-H on this dataset. The experiments show that PickScore demonstrates better alig... | Rebuttal 1:
Rebuttal: Weaknesses:
1. Some experimental results are confused, e.g., the correlation value between FID and human judgment is -0.9, which approaches the lower bound of correlation. Does it mean there is a negative correlation between FID and human judgment?
Answer: Regarding the specific example and que... | Summary: This paper contributes a new dataset of human preferences for text-to-image (T2I) generations. It then trains a CLIP-based score function to that can predict human preferences between two generated images. The paper then shows how this can be used to better align a T2I model to human preferences.
Strengths: -... | Rebuttal 1:
Rebuttal: Weaknesses:
1. While the proposed dataset itself is a significant enough contribution towards acceptance, the paper's technical contribution is relatively trivial and calls for further discussion.
Answer: While we appreciate the reviewer's note that the dataset itself is a significant enough co... | Rebuttal 1:
Rebuttal: We thank the reviewers for their thoughtful feedback. We are encouraged that they found our paper to be a very solid (UW8i), timely (UW8i, c5Um), and significant (c5Um) contribution, and our experiments to be thorough (UW8i). We are glad that they consider our open-sourced and large dataset of hum... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Implicit Bias of Gradient Descent for Two-layer ReLU and Leaky ReLU Networks on Nearly-orthogonal Data | Accept (poster) | Summary: The paper studies convergence and implicit bias in two-layer (leaky) ReLU networks trained with gradient descent. They focus on the case where the initialization scale and the step size of GD are sufficiently small, and the training data are nearly orthogonal. They show two main results:
- For leaky ReLU net... | Rebuttal 1:
Rebuttal: Thank you for your positive feedback and thoughtful comments! Below are our answers to your comments and questions.
----
**Q1**. In the introduction, the authors argue that “This finding suggests that ReLU networks possess superior learning ability compared to leaky ReLU networks”. This sentence... | Summary: The paper investigates implicit bias, especially in terms of the stable rank of the weight matrix (which is the square of the ratio between the Frobenius and spectral norms), of gradient descent on the logistic loss for binary classification tasks and two-layer neural networks, where the first layer is trained... | Rebuttal 1:
Rebuttal: We appreciate your valuable comments and feedback! We address your comments and questions below.
----
**Q1**. It is not very clear to me what difficulties exactly needed to be overcome to step up from that previous work to this paper.
**A1**. The technique employed in this paper differs signifi... | Summary: This paper studies training two-layer leaky-ReLU/ReLU networks with gradient descent. The authors show that for approximately orthogonal data, GD converges and the weight matrices in the leaky-ReLU network have a stable rank of one asymptotically under GD, while those in the ReLU network can have a stable rank... | Rebuttal 1:
Rebuttal: Thank you for your insightful feedback! Here's our response to your comments and questions.
----
**Q1**. There is little discussion on the upper bound c in Theorem 4.3. How large it can be? How does it depend on the training data?
**A1**. The upper bound c in Theorem 4.3 can be considered an ab... | Summary: This study analyzes two-layer ReLU and leaky ReLU networks trained by gradient descent in a classification problem. It is shown that when training data is mutually nearly-orthogonal, by taking an initialization scale and a learning rate sufficiently small, a leaky ReLU network converges to a parameter with sta... | Rebuttal 1:
Rebuttal: Thank you for your positive feedback and valuable comments! We address your comments and questions as follows.
----
**Q1**. The near-orthogonality assumption seems to be strong.
**A1**. We have based our work on the assumption of nearly-orthogonal data following Vardi et al., (2022) and Frei et... | Rebuttal 1:
Rebuttal: We want to thank all the reviewers for their positive and valuable feedback. In the uploaded pdf, we include additional empirical results to address reviewer c8EH’s concern. We have conducted experiments for long epochs (1000). We find that the stable rank for the leaky ReLU network with large slo... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
On the Sublinear Regret of GP-UCB | Accept (poster) | Summary: The paper proposes a new analysis for the popular GP-UCB algorithm. Under this new approach, it is shown that GP-UCB achieves a sub-linear regret (though not optimal), partially resolving the question of whether GP-UCB can achieve optimal regret.
Strengths: The paper proposes a new self-normalized martingale ... | Rebuttal 1:
Rebuttal: [*1. I can see that there is clearly certain improvement over the existing result. However, my concern is whether the result warrants a full publication. From what I understand, the important trick is to analyse in the Hilbert space representation, which allows for a tighter analysis and a cleaner... | Summary: This is a very well written paper with nice, clean results on self-normalised concentration for seperable RKHS. The paper claims two contributions:
1. A new self-normalised concentration result for seperable RKHSs.
2. Improving the regret of GP-UCB by tuning the regularisation parameter.
Strengths: The writ... | Rebuttal 1:
Rebuttal: [*1. The result claimed in point 1 is actually well known, and has been for over a decade. See Theorem 4.1 in [1]. Indeed, the authors of the submitted manuscript state: "To the best of our knowledge, Fact 6 is the only existing result on self-normalised, time-uniform concentration in RKHS's [sic]... | Summary: This paper addresses the kernelized bandit problem and focuses on improving the regret bounds of the Gaussian Process Upper Confidence Bound (GP-UCB) algorithm. The authors introduce novel techniques to achieve nearly optimal regret rates for GP-UCB. These improvements surpass previous state-of-the-art analyse... | Rebuttal 1:
Rebuttal: [*1. For example, the authors refer to the Matern kernel as a commonly used kernel, yet to someone not as familiar with the specific literature it is harder to interpret the result. Perhaps consider providing a bit more background on it. (e.g., it was not mentioned in [4]) Why is it commonly used ... | Summary: This paper re-investigates the upper bound analysis of the popular GP-UCB algorithm (in its vanilla version), focusing on the cumulative regret for Matern kernel (and more generally, kernels with polynomial decay). The study reveals that the previous non-sub-linear regret bounds can be surpassed.
Strengths: ... | Rebuttal 1:
Rebuttal: [*1. The technical contribution of this paper appears to be non-significant, as it essentially extends a well-known result from linear bandits (Abbasi-Yadkori et.at, 2011), by fusing several exsiting results/facts.*]
We believe the simplicity of our arguments are what make the technical contribut... | Rebuttal 1:
Rebuttal: We would like to thank all reviewers for their thoughtful comments and feedback on our paper. We have responded to each of your comments, and we hope to address any follow-up questions you may have in the discussion.
Several reviewers have pointed out the existence of a bound in the thesis of Yas... | NeurIPS_2023_submissions_huggingface | 2,023 | Summary: This paper addresses the open question of Vakili et al, on the possibility of obtaining rate-optimal regret bounds for kernelized bandits under the RKHS assumption. Authors present a slightly different RKHS martingale bound, which allows them to obtain a regret of the form $\rho\sqrt{\gamma_TT}+ \gamma_T\sqrt{... | Rebuttal 1:
Rebuttal: We would like to thank the reviewer for the overall positive reviews and constructive feedback.
[*1. This paper does not resolve the open question of Vakili et al. As far as I know, the key question within the field is to whether we can get a bound on the regret in the RKHS setting, as we do in ... | Summary: This paper investigates the performance of the Gaussian Process Upper Confidence Bound (GP-UCB) algorithm in kernelized bandit problems, focusing on the minimization of regret. The authors present a novel, self-normalized concentration inequality and a technique to regularize in proportion to the kernel's smoo... | Rebuttal 1:
Rebuttal: We would like to thank the reviewer for their overall positive reviews. | null | null | null | null |
Evolutionary Neural Architecture Search for Transformer in Knowledge Tracing | Accept (poster) | Summary: This paper proposes to search for the optimal Transformer-based architecture for the knowledge tracing task. The authors have effectively constructed a search space that captures both global and local context modeling to accurately capture students' forgetting behavior. To further improve the efficiency of the... | Rebuttal 1:
Rebuttal: ## Response to Weakness 1
Although we have reported the search cost of our approach and compared approaches in **Figure 3** of ***the previous Appendix***,
Figure 3 only reports each approach's overall cost and the training cost of the final architecture,
where the cost of the supernet training... | Summary: This paper modifies the Transformer architecture and employs Neural Architecture Search (NAS) to tackle knowledge tracing tasks
Strengths: Writing is good and the method looks sophisticated.
Weaknesses: As I am not specialized in the field of knowledge tracing, I cannot accurately assess the novelty of this ... | Rebuttal 1:
Rebuttal: ## Response to Weakness
We would like to clarify the novelty of our work. While it is true that NAS has been extensively researched, its application in the field of KT is indeed novel. The main challenge in KT is to model the students' knowledge states accurately, which involves capturing both t... | Summary: This paper presents neural architecture search (NAS) for transformer in the context of educational application, specifically knowledge tracing. The authors employed two level NAS:
one at a local level to encorporate forgetting behavior and architecture serach at global level for exercise and response related e... | Rebuttal 1:
Rebuttal: ## Response to Weakness 1
Actually, we have discussed the comparison of the proposed approach ENAS-KT and existing NAS approaches for other tasks in terms of approach design in the *Related work* of the submitted paper, where the approach design comparison between ENAS-KT and NAS-Cell is also pres... | Summary: This paper introduces neural architecture search (NAS) for Transformer in the field of Knowledge Tracing (KT) for the first time. The authors propose a Transformer architecture that combines local and global paths, as well as a search space, to address the issue of students’ forgetting behavior in the knowledg... | Rebuttal 1:
Rebuttal: ## Response to Question 1
According to your suggestion, we have added more variants to the original Table 3 (forming **Table II** of ***global_response***) to show the effectiveness of searching and the contribution of each component to the final model's performance.
Firstly, two variant archit... | Rebuttal 1:
Rebuttal: We sincerely appreciate the valuable feedback provided by all the reviewers. We have carefully addressed their questions and concerns in our response, aiming to provide satisfactory answers.
Here we uploaded a file named "global_response" to show some necessary results and comparisons, which cont... | NeurIPS_2023_submissions_huggingface | 2,023 | Summary: The architecture of a transformer model for a knowledge tracing task is searched for using evolutionary computation. The found solutions perform well.
Strengths: - ANN design is difficult
- KT is important
- Simple NAS method
Weaknesses: - Interpretability of the solution, i.e. ANN
- Limited discussion re... | Rebuttal 1:
Rebuttal: ## Response to Question 1
*The 'fusion' method* refers to that **a hierarchical fusion module** is employed in our approach to **aggregate** the input features fully.
In **the hierarchical fusion module**, the pairwise concatenation between input features is first performed to obtain some inter... | null | null | null | null | null | null |
A Dual-Stream Neural Network Explains the Functional Segregation of Dorsal and Ventral Visual Pathways in Human Brains | Accept (poster) | Summary: This paper analyzes which ANN structures are more suitable for conversion to SNN. Based on the analysis results, the RepConv model is chosen as the baseline ANN model. The ANN-to-SNN conversion is performed using Value-range (VR) encoding and Clamp and Quantization (CQ) training, resulting in a high-performanc... | Rebuttal 1:
Rebuttal: We are grateful for your positive evaluation, constructive feedback, and insightful comments.
***
1\. Regarding the questions on why the learning rate for stage3 was set lower than that for stages 1 and 2, the answer is that our empirical findings led us to this decision, as a reduced learning rat... | Summary: The main contributions is a dual stream neural network to model human visual processing. The network is composed of two CNNs (where and what) to model the dorsal and ventral pathways. The authors demonstrate that they can predict human gaze behavior and neural responses using their model, and gain new insights... | Rebuttal 1:
Rebuttal: We are glad that the reviewer recognizes the novelty, clarity, and contribution of our paper and offers constructive comments. Our responses to the weaknesses and questions are as below.
***
__Applications in computer vision__
The main focus of this paper is on modeling the human visual system, ... | Summary: This paper focuses on investigating differences between the human ventral and dorsal visual streams. The authors develop a dual-stream model that has different retinal inputs to support “what” processing (object recognition) or “where” processing (guiding the fixations). The authors then test how well the mode... | Rebuttal 1:
Rebuttal: We are glad that you recognize some strengths, interests, and novelties of our work. We also respect your critical evaluation and feedback.
***
1. __Overreach claims.__ We respect your concern regarding the potential overreach of our initial claim. We have revisited our wording to more accurately... | Summary: The author introduces a dual-path parallel neural network designed to simulate the hypothesized functions of the ventral and dorsal streams in the human visual system. This network consists of a WhereCNN, which has a wide view to learn spatial attention and direct gaze, and a WhatCNN, which has a local view to... | Rebuttal 1:
Rebuttal: Thanks for your summary and constructive feedback.
***
1\. __Novelty__. While we concur that the ventral and dorsal visual streams have been individually studied before, our work is novel in its integrative approach. We model and link the ventral and dorsal streams as two integral parts of a who... | Rebuttal 1:
Rebuttal: We appreciate the time, effort and feedback from all reviewers. We are glad that the review comments are overall positive and seem to merit acceptance. We have also carefully addressed all concerns from the reviewers, by providing new results (shown in the pdf file), discussions, and clarification... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Neural Image Compression: Generalization, Robustness, and Spectral Biases | Accept (poster) | Summary: This paper proposes to adopt spectral tools to analysis OOD issues of NIC. On the proposed corrupted dataset, the paper investigates the error of both conventional and learning-based image compression approaches and provides some interesting conclusions, showing robustness and generalization of NIC and suggest... | Rebuttal 1:
Rebuttal: Thank you for your thoughtful review of our submission. In addition to our inline responses to your comments below, please note that our global response highlights new experiments **E1** and **E2** and their insights **I1** - **I4** as well as clarifies the significance of our work in **C1** - **C... | Summary: This paper introduces a benchmark for evaluating the out-of-distribution (OOD) performance of neural image compression (NIC) models. The authors observe distinguished spectral biases of NIC models by comparing them with JPEG2000. Then the authors listed all the spectral biases along with a theoretical analysis... | Rebuttal 1:
Rebuttal: Thank you for your thoughtful review of our submission. We believe that there is some misunderstanding in the goal and contribution of our work which we first clarify. Namely, existing tools are incapable of highlighting inter- and intra-class similarities and differences for classical codecs and ... | Summary: This paper analyzes neural image compression models relative to classical (hand-engineered) methods using a new dataset and new tools based on spectral analysis. Specifically the paper presents (from Section 1, "Main Contributions"):
(1) a new dataset covering different kinds of distortions used to compare met... | Rebuttal 1:
Rebuttal: Thank you for your thoughtful review of our submission. We are grateful to hear an expert in this subfield found our work to be fascinating while being mainly concerned about the fit of our work at NeurIPS. We appreciate you stating clear reasoning behind your rating as it gives us an opportunity ... | Summary: This paper addressed the problem of image compression. Neural-based methods have started outperforming traditional codec methods in image compression, but there are no comprehensive datasets and informative tools to evaluate neural-based methods in real-world settings. This paper proposed a benchmark to study ... | Rebuttal 1:
Rebuttal: Thank you for your thoughtful review of our submission. In addition to our inline response to your comment below, please note that our global response highlights additional experiments **E1** and **E2** and their insights **I1** - **I4** as well as clarifies the significance of our work in **C1** ... | Rebuttal 1:
Rebuttal: We thank the reviewers for their valuable feedback. We are grateful that reviewers found our contributions for benchmarking neural image compression models to be “comprehensive” (5yiv), “interesting” (e3od, h4BN), “important” (h4BN), “fascinating” (e3od), “crucial to bridge the real-world adoption... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Data Quality in Imitation Learning | Accept (poster) | Summary: The paper analyze several theoretical property of the expert demonstration distribution to the performance of the learnt policy in imitation learning. The paper claims that people should provide consistent and transition and state diversity.
Strengths: The paper defines several quantity to clear express the m... | Rebuttal 1:
Rebuttal: We thank the reviewer for their thoughtful review.
**Straightforward results**
We agree with the reviewer that several of the conclusions are intuitive; however, we disagree that they are trivial or straightforward. See our main response comment (Straightforward Results).
The reviewer also stat... | Summary: This paper considers how to evaluate the quality of a given dataset for an imitation learning task. It is well-known that IL suffers from distribution shift. While much work on IL focuses on improving algorithms, the authors propose that another important approach to making IL more effective is to curate bet... | Rebuttal 1:
Rebuttal: We thank the reviewer for their thoughtful review, and we are glad the reviewer found both the theoretical and empirical analysis important and interesting.
**Reducing repetition**
We are sorry that the reviewer felt the paper was repetitive! Our intention was to clearly state the takeaways, but... | Summary: This paper studies imitation learning (IL) from a theoretical perspective. It proposes the idea that we should curate the data to improve performance, rather than solely relying on designing more robust algorithms, which is a more common practice in the field. The authors propose data quality for a particular ... | Rebuttal 1:
Rebuttal: We thank the reviewer for their thoughtful review, and we are excited that the reviewer found our insight about the algorithm's role in data quality to be very nice.
**Theoretical Assumptions**
1. *Why should $\beta$ be small in Lemma 4.2?*
$\beta$ corresponds to the in-distribution loss of the... | Summary: The paper addresses a key issue in Imitation Learning (IL) called as distribution shift. The paper formalize data quality and data curation for IL. The paper propose two fundamental properties to evaluate/measure data quality as "action divergence" and "transition divergence". The authors investigate these two... | Rebuttal 1:
Rebuttal: We thank the reviewer for their thoughtful review, and we appreciate that they agree of the importance of the ideas presented in the paper.
**The work doesn't seem to fit for a high quality conference like NeurIPS.**
We disagree that this work is not fit for NeurIPS. As noted by several other re... | Rebuttal 1:
Rebuttal: We thank all the reviewers for their detailed and constructive feedback. We are excited that the majority of reviewers agree about the importance of the data quality problem in imitation learning and find our analysis sound.
We outline the points of important feedback that came up several times i... | NeurIPS_2023_submissions_huggingface | 2,023 | Summary: Data quality is an important aspect of offline robotics because there is not a large quantity of data. In particular, in imitation learning, it has been shown that naive training policies on low-quality data can lead to distribution shift issues. Most researchers focused on dealing with this issue using algori... | Rebuttal 1:
Rebuttal: We thank the reviewer for their thoughtful review, and we are glad the reviewer felt our empirical evidence supported our claims.
**Lack of novelty**
The theorem the reviewer cites as similar to 4.1 in [1] looks similar, however it makes the strong assumption of the f-divergence satisfying the t... | Summary: This paper systematically studied the data quality assessment problem in imitation learning, proposing new metrics based on test-time distribution shift.
Strengths: The paper is generally well-organized and well-written. The analysis seems to be in fair detail.
Weaknesses: One of the drawn conclusions seem... | Rebuttal 1:
Rebuttal: We thank the reviewer for their thoughtful review, and we appreciate that the reviewer felt our analysis was in fair detail.
**The conclusion on state diversity seems straightforward.**
We agree that this conclusion is intuitive; however, we disagree that it is trivial: state diversity has been ... | null | null | null | null |
On Differentially Private Sampling from Gaussian and Product Distributions | Accept (poster) | Summary: The paper discusses the problem of generating a sample from a distribution that is close to an unknown distribution P, under the constraint of differential privacy. The authors present new differential privacy sampling algorithms for multi-dimensional Gaussian distributions, with different assumptions on the a... | Rebuttal 1:
Rebuttal: * While it would be nice to have a more generic algorithm, we note that it is usually quite difficult to nail down the exact sample complexity in DP for a generic family of distributions. To give a sense of this difficulty, note that it required more than a dozen papers (cited in our paper) focusi... | Summary: This work studies the problem of differentially private sampling. In this setting, the goal is the following: Given $n$ iid samples from an unknown distribution (e.g., a Gaussian distribution with an unknown mean), design a private algorithm that produces a fresh sample from the distribution the original sampl... | Rebuttal 1:
Rebuttal: The polynomial dependence on $1/\\alpha$ is necessary in the product distribution case; as indicated in Table 2, it was already shown in [34] that the sample complexity for $(\\epsilon,\\delta)$-DP is at least $\\Omega(d/\\alpha\\epsilon)$. As you said, it is indeed a pretty interesting situation,... | Summary: The paper considers the problem of sampling from distributions under differential privacy. In particular, the specific formulation presented in the paper assumes the existence of a dataset that has been drawn i.i.d. from a distribution $P$ that belongs in a family of distributions $\mathcal{F}$ and the goal is... | Rebuttal 1:
Rebuttal: We wish to correct the misunderstanding here: Our bounds are completely *independent of the $R$ parameter* and in fact *do not* require any assumption on $R$. Please refer to Theorems 1.2, 1.4, and 1.6 for our bounds.
It seems like the misunderstanding is perhaps due to focusing on the bounds in ... | Summary: Raskhodnikova, Sivakumar, Smith, and Swanberg (NeurIPS '21) recently introduced the problem of privately approximate sampling: Given i.i.d. draws from a member D of a known family of distributions, produce one sample whose distribution is close to D in TV distance. This is an easier task than the well-studied ... | Rebuttal 1:
Rebuttal: We thank the reviewer for their review & comments. Please find our answers to the questions below.
- It is possible to get some bounds in the agnostic case using our algorithms, but these bounds are not particularly good. In particular, let $n$ be the sample complexity in our algorithms (e.g., fr... | null | NeurIPS_2023_submissions_huggingface | 2,023 | Summary: This paper considers a very interesting problem: Design a $\mathit{differentially~private}$ algorithm that given samples from unknown distribution $D$ from a class of distributions ${\mathcal D}$, output a sample from a distribution $D'$ where $D'$ is close to $D$ in total variation distance. One direct approa... | Rebuttal 1:
Rebuttal: We thank the reviewer for their review & comments. | null | null | null | null | null | null |
Balancing memorization and generalization in RNNs for high performance brain-machine Interfaces | Accept (spotlight) | Summary: This paper explores the utility of different classes of neural decoders for online brain machine interfaces. The authors first show that recurrent network decoders outperform transformers, CNNs and Kalman filters in an offline setting. They then validate that the superiority of RNN-based decoders also holds in... | Rebuttal 1:
Rebuttal: We thank the reviewer for their questions and helpful suggestions.
**“ReFIT” is not explained.**
We propose modifying line 40 to briefly explain ReFIT:
“Alternatively, [30] found that a convolutional feedforward network using using intention-based recalibration (ReFIT) outperformed a ReFIT Kalm... | Summary: The work analyzes the performance of different decoders for finger movement prediction from monkeys in different offline and online settings. The authors evaluate recurrent networks, convolutional networks, transformers and Kalman filters. They find recurrent networks to perform best both for online decoding a... | Rebuttal 1:
Rebuttal: We thank the reviewer for their questions and helpful suggestions.
**ReFIT is first used without an explanation.**
We propose modifying line 40 to briefly explain ReFIT:
“Alternatively, [30] found that a convolutional feedforward network using intention-based recalibration (ReFIT) outperformed ... | Summary: The authors study the performance of various models (GRUs, LSTMs, TFMs, FNNs, KFs) in reading out neural population activity during one and two finger tasks in non-human primates. Overall, they find that RNNs (GRUs & LSTMs) outperform other models during online tasks and serve as the most promising model for r... | Rebuttal 1:
Rebuttal: We thank the reviewer for their questions and feedback.
**What are the practical limitations of extending the claims here to more naturalistic BMIs? (More DoFs)**
We address this question in the global response and propose an addition to the discussion. Briefly, more DoFs may require more train... | Summary: This paper uses an RNN for motor control task. Specifically, it uses the RNN to decode finger movements using neural data recorded from NHPs.
Strengths: The following are the strengths of this paper:
-Online decoding using RNNs.
-Evaluation and comparison with other neural network algorithms.
-Evaluation w... | Rebuttal 1:
Rebuttal: We thank the reviewer for their questions.
**Do you see a difference between SBP and spike thresholding?**
In this work we did not directly compare SBP to spike thresholds but used SBP as the feature for all decoders. Anecdotally, in experiments prior to this study, Monkey N was unable to compl... | Rebuttal 1:
Rebuttal: We thank the reviewers for their support of this work and their helpful questions and suggestions. Our work demonstrates state-of-the-art control of a finger brain-machine interface, where we show the importance of closed-loop decoder testing and benefits of decoder memorization. The majority of r... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Dynamic Pricing and Learning with Bayesian Persuasion | Accept (poster) | Summary: The paper studies a dynamic pricing model augmented by advertisement campaigns. The problem is framed in a Bayesian persuasion framework, where the authors focus on the online learning version of the model in which the agent is not aware of buyer the demand function from which its types are drawn.
Strengths: ... | Rebuttal 1:
Rebuttal: Thank you for your comments, we will incorporate your comments and presentation suggestions in the revision. We will respond the questions as follows:
**Q1 (Revelation principle on binary signals):**
If buyer’s type were fixed, then binary-signal would indeed suffice, as known from previous liter... | Summary: This work focuses on advertising and selling markets under the Bayesian persuasion framework. On the one hand, this paper models the effect of the advertisement strategy on customers’ beliefs and purchase decisions. On the other hand, the authors also address the problem of online learning the optimal advertis... | Rebuttal 1:
Rebuttal: Thank you for your comments, we will incorporate your comments in the revision. We start with addressing the technique comment:
**Q1 (Tightness of $O(T^{2/3})$ and $O(T^{3/4})$ for discrete and continuous quality spaces):**
Our regret $O(T^{2/3})$ for discrete quality space is tight, please see t... | Summary: The authors consider an online single product dynamic pricing product, where a buyer's purchase decision is not only influenced by the offered price from the selling, but also depends on the advertising strategy used by the seller. The authors formalized the latter via the model of Bayesian persuasion framewor... | Rebuttal 1:
Rebuttal: Thank you for your comments, we will incorporate your comments in the revision. We below respond the questions as follows:
**Q1 and Q3 (The seller’s knowledge of buyer’s posterior belief $\mu_t$):**
The reviewer is correct that the seller knows the (realized) buyer's posterior belief $\mu_t$. We ... | Summary: The paper extends the traditional dynamic pricing model (e.g. Kleinberg and Leighton) to also include product quality. In each period, we get both a random type of the buyer as well as a random quality of the product. The seller can employ an advertising strategy, which is a mapping from product quality to a s... | Rebuttal 1:
Rebuttal: Thank you for your comments, we will incorporate your comments and presentation suggestions in the revision. We below mainly address the technique comment:
**Weakness 3 (Why not Explore-then-Exploit algorithm):**
We believe the regret achievable by any "Explore-then-Exploit" algorithm is $\tilde... | null | NeurIPS_2023_submissions_huggingface | 2,023 | Summary: In this work, the authors study a novel setup of repeated posted-price auctions, where a single seller deals with a single myopic buyer (or, equiv, iid new buyer in each round). In each round, a new good is offered and has characteristics that affect buyer valuation (so, goods are not equal between rounds). Th... | Rebuttal 1:
Rebuttal: Thank you for your comments, we will incorporate your comments and presentation suggestions in the revision, and also adequately discuss the relevance to the pointed literature. We below respond the important questions:
**Weakness 3 and Q7 (Complexity of the algorithm):**
We kindly remind the rev... | null | null | null | null | null | null |
Exact Bayesian Inference on Discrete Models via Probability Generating Functions: A Probabilistic Programming Approach | Accept (oral) | Summary: This paper proposes a new approach to exact inference in probabilistic programs. Existing approaches to exact inference attempt to compute marginal and normalized probabilities by efficiently computing the necessary (finite) sums, or using computer algebra to recognize and solve tractable integrals. By contras... | Rebuttal 1:
Rebuttal: > The novel capabilities (and limitations) with respect to the line of work by Klinkenberg et al. could be more clearly described. It would also be good to cite and explain the differences with their recent LAFI abstract
The ideas alluded to in their LAFI abstract are now available in a more fles... | Summary: This paper proposes a new probabilistic programming language called statistical guarded command language (SGCL) and an exact Bayesian inference method for any discrete statistical model that can be expressed in SGCL. Empirical evaluatios on a variety of different discrete probabilistic models demonstrate the e... | Rebuttal 1:
Rebuttal: > The paper seems to miss discussions on probabilistic circuits which is a large body of work concerning probabilistic models with tractable inference.
Thank you for pointing this out. Probabilistic circuits (Choi et al., 2020) are indeed a family of probabilistic models that allow for efficient ... | Summary: The paper presents a restricted, first-order probabilistic programming language in which inference for infinite-support latent variables, with continuous or discrete sample spaces, admits a closed form. For experiments they compare against importance sampling, Lightweight M-H, Random Walk M-H (RMH), sequentia... | Rebuttal 1:
Rebuttal: > making explicit the if-and-only-if conditions that must be met for generating functions to support closed-form inference
Please refer to our general rebuttal for the “only-if” aspect (whether additional constructs could be added to the programming language while still preserving a closed form).... | Summary: A probabilistic programming framework for exact inference in discrete probabilistic programs with infinite support is introduced. The framework levegares probability generating functions as a first class attribute of discrete distributions. The probabilistic programming language is formally defined, inference ... | Rebuttal 1:
Rebuttal: > The PPL is implemented in Rust, a language which, while it certainly has its merits, less popular for statistical modelling and Bayesian inference, which may make adoption of the approach more difficult.
It is correct that Rust is not used in Bayesian inference very much. However, it would not ... | Rebuttal 1:
Rebuttal: First of all, we would like to thank the reviewers for their thoughtful reviews and interesting questions about our work. We are delighted that the reviewers think that our approach is “significant” (JaNH), “novel and very useful” (SkQ5), that our analysis is “rigorous” (c91c, JaNH), that our impl... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
RiskQ: Risk-sensitive Multi-Agent Reinforcement Learning Value Factorization | Accept (poster) | Summary: This paper studies risk-sensitive action selection in cooperative multiagent reinforcement learning (MARL) under the centralized training but decentralized execution (CTDE) paradigm, based on the value decomposition class of methods. The principle underlying value decomposition is Individual Global Max (IGM), ... | Rebuttal 1:
Rebuttal: We want to express our sincere gratitude to the reviewer. Such a thoughtful and in-depth review can help us greatly improve the quality of this work. We hope that with our response, the reviewer can kindly reconsider the final opinion. We answer the comments as follows.
**Limitations are not thor... | Summary: The authors introduce the Risk-sensitive Individual-Global-Max (RIGM) principle as a generalization of the Individual-Global-Max (IGM) and Distributional IGM (DIGM) principles and propose RiskQ satisfies the RIGM principle. The method is evaluated by both theoretical analysis and experiments.
Strengths: 1. T... | Rebuttal 1:
Rebuttal: We want to thank the reviewer for the effort and time in reviewing this paper. We will take your valuable feedback into account to improve our work. We hope that the reviewer will have a better understanding of our work and a higher confidence in the reviewer's assessment.
***Representation capac... | Summary: The authors propose RiskQ, a value function factorization method for MARL that considers the risk factor. Risk is a concept present in reinforcement learning, and the authors propose a generalisation of the two popular principles IGM and DIGM: the RIGM, that considers the factor of risk in MARL.
Strengths: I... | Rebuttal 1:
Rebuttal: We thank the reviewer for your time and effort in reviewing this submission. We will describe the definition of DEC-POMDPs, the VaR metric, the DRM, the risk-averse SMAC, and the risk-sensitive action selection in more detail and include them in the revised paper. The source code of RiskQ is inclu... | Summary: This paper propose a factored distributional value learning for multi-agent reinforcement learning. The paper also propose a theoretically desirable factored risk-distortion property to which the proposed method adhere.
Strengths: The theoretical argument establishes an intuitively sound criterion for evaluat... | Rebuttal 1:
Rebuttal: We want to express our sincere gratitude to the reviewer. Such a thoughtful and in-depth review can help us greatly improve the quality of our work.
**The empirical improves are marginal ... require risk-sensitive cooperations**
As shown in Fig. 2 of our work, RiskQ performs ***significantly be... | Rebuttal 1:
Rebuttal: We faithfully thank all reviewers for their insightful comments and valuable feedback. The reviewers acknowledge our work as an extension of popular principles (X7fQ, rLaN, EN1R), an intuitively sound criterion (e297,EN1R), a good contribution (rLaN, EN1R), novel (EN1R), promising (X7fQ), with go... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Differentiable Blocks World: Qualitative 3D Decomposition by Rendering Primitives | Accept (poster) | Summary: The paper presents a method for producing simple "blocks world" 3D geometry for a scene from posed RGB images. Compared with multi-view 3D reconstruction approaches that produce point clouds or nerfs, the emphasis is on producing a simple, robust and (somewhat) object-aligned 3D representation. The representat... | Rebuttal 1:
Rebuttal: We thank the reviewer for their valuable feedback. In the following, we address the main concerns raised in this review. We will incorporate relevant suggestions in the revised version.
***Weaknesses: "The system apparently produces tight fits only for actual "blocks world" scenes", "The usual c... | Summary: This paper presents an optimization-based method to recover 3D textured primitives from image observations. The core of the method is a fully differentiable and parsimonious 3D scene representation consisting of a background dome, ground plane, and a collection of multiple superquadrics, where the transformati... | Rebuttal 1:
Rebuttal: We thank the reviewer for their valuable feedback. In the following, we address the main concerns raised in this review. We will incorporate relevant suggestions in the revised version.
***W1: Additional baselines using geometry from NeuS***\
A: In the following table, we report DTU results for M... | Summary: This paper (DBW) presents a way to fit superquadric blocks to a scene, starting from multiple images with calibrated cameras. Unlike recent other methods, they don't start with 3D data in the form of point clouds or meshes. The compositional nature of the recovered scene is shown by manipulating the recovered ... | Rebuttal 1:
Rebuttal: We thank the reviewer for their valuable feedback. In the following, we address the main concerns raised in this review. We will incorporate relevant suggestions in the revised version.
***Weaknesses: "The DRC's is the approach in my view that should motivate DBW presented in this paper. [...] I ... | Summary: This paper focuses on parsing a scene into mid-level 3D representations made of a small set of textured primitives. The main contributions are as follows:1) Unlike existing primitive decomposition methods which requires 3D data, this approach relies directly on images as input.
2) The proposed method reconstru... | Rebuttal 1:
Rebuttal: We thank the reviewer for their valuable feedback. In the following, we address the main concerns raised in this review. We will incorporate relevant suggestions in the revised version.
***W1, Q1: "The motivation in this paper should be made clear. Why using a differentiable rendering based optim... | null | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Faster approximate subgraph counts with privacy | Accept (poster) | Summary: The paper considers the problem of releasing information about graphs (specifically, subgraph counts) under differential privacy. This has been a widely studied problem, due to the challenging nature of ensuring that the noise added is bounded. Many techniques (e.g., smooth sensitivity) have been developed w... | Rebuttal 1:
Rebuttal: W1. The paper takes great pains to cite recent work by Blocki et al [4] which as far as I can tell is currently an unpublished technical report. It summarizes the difference as studying approximate sensitivity. Still, I think it would be useful to provide more contrast. For instance, the empirical... | Summary: The paper focuses on the problem of counting the number of non-induced embeddings of a subgraph while ensuring differential privacy. The objective is to develop efficient DP algorithms for subgraph counting that outperform previous approaches in terms of running time.
To accomplish this goal, the authors give... | Rebuttal 1:
Rebuttal: W1. The running time of the DP algorithm for (arbitrary) subgraphs with edges is $O(n^{2l})$, which is still quite inefficient.
*Response.* We completely agree with the reviewer's assessment about the practicality of the result for general subgraphs $H$.
However, we note that from a theoretical p... | Summary: This paper extends the study of local sensitivity and smooth sensitivity for the problem of subgraph counting. Previous work has identified local sensitivity and its derivates like smooth sensitivity as a tool to calibrate noise to the given input instead of global sensitivity. However, this comes with two cha... | Rebuttal 1:
Rebuttal: W1. The bound for general graphs and paths is impractical, as the degree of the running time polynomial scales linearly with the size of $H$.
*Response.* We completely agree with the reviewer's assessment about the practicality of the result for general subgraphs $H$.
However, we note that from a... | Summary: This paper provides an algorithm to approximately estimate the smooth sensitivity for counting triangles in a graph, while preserving edge-level differential privacy. Compared with the existing solution for smooth sensitivity, the proposed algorithm demonstrates efficiency improvement in terms of computation. ... | Rebuttal 1:
Rebuttal: W1.It would be better if the authors can provide a table comparing their method with existing solutions. The comparison should include time complexities, tasks (e.g., triangle counting, clique counting), sample complexities, and maybe utility guarantees.
*Response:* Thanks for the suggestion.
We... | null | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
3D-IntPhys: Towards More Generalized 3D-grounded Visual Intuitive Physics under Challenging Scenes | Accept (poster) | Summary: This paper proposed a combination of a NeRF-style perception module and a 3D point-based dynamics module to model challenging physical phenomena involving fluid, granular materials, and rigid objects. The perception module learns spatial-equivariant representations and transforms the environment into points. T... | Rebuttal 1:
Rebuttal:
Thank a lot for identifying our work and providing valuable comments. Below, we try to address the raised questions:
> The training of NeRF only promises the quality of object surfaces and does not impose any constraint on the interior of an object. As a result, the object occupancy field may... | Summary: The paper proposes to learn intuitive physics from visual input by learning to infer the 3D scene structure and dynamics. The claim is that this explicit 3D reasoning generalizes better than existing implicit representations, and can be learned from videos without any explicit 3d supervision. To this end, the ... | Rebuttal 1:
Rebuttal: > The main limitations is it's unclear how the method will generalize to real-world scenarios.
Thanks for your advice about real-world experiments. Our paper is among the first to embed explicit 3D information and strong inductive bias into 3D intuitive physics learning.
Here we want to emphas... | Summary: This paper proposes a method that learns to approximate physical interactions of primitive scenes from a set of videos through the use of of a NeRF and a dynamics prediction model. A static NeRF (based on PixelNeRF) is used to reconstruct a per time-step point cloud which is then used to supervise the dynamics... | Rebuttal 1:
Rebuttal:
Thank you for reviewing our work and providing valuable comments. Below, we try to address the questions you raised:
> No videos of interpolated or extrapolated dynamics are provided. MSE itself is not a good measure of the good a simulation is.
We kindly request the reviewers to watch the... | Summary: Authors propose an approach to learn physics based model of real world objects from multi-view video. Authors argue that 3D representation of the scene is important to reason about the physics and use per-frame Nerf to extract the object point cloud from every frame of the multi-view video. They build a graph ... | Rebuttal 1:
Rebuttal:
Thank you for reviewing our work and providing valuable comments. Below, we try to address the questions you raised:
> Proposed method consists of two steps, extracting the scene point cloud using Nerf [41] and computing dynamics using message passing [7,36,37]. Isn’t this similar to [35] wh... | Rebuttal 1:
Rebuttal: # General Response
We thank the reviewers for their valuable comments on our paper. We are excited to see that reviewers (Reviewer cs73, 493i, R6LT, kxU1, SEVF) identified the novelty of our technical contribution, acknowledged the better performance of our methods over baselines (Reviewer cs73, ... | NeurIPS_2023_submissions_huggingface | 2,023 | Summary: The paper introduces a 3D point-based dynamic model for visual intuitive physics reasoning. Specifically, The model adopts a conditional NeRF as the perception module to extract 3D point clouds from multi-view images. It then employs a graph neural network that operates on the extracted points for forward dyna... | Rebuttal 1:
Rebuttal:
Thanks for the careful review and valuable comments. Here we will address the questions raised up by the reviewer:
> Reliance on instance-level masks. While SAM can provide instance-level masks (as in Supp. 148-152), how the mask quality can impact the performance
The instance-level masks i... | null | null | null | null | null | null |
NU-MCC: Multiview Compressive Coding with Neighborhood Decoder and Repulsive UDF | Accept (poster) | Summary: The paper studies the problem of shape reconstruction from single-view RGBD images and addresses two key limitations in a major prior work, MCC: 1) inefficient decoder of quadratic time complexities, and 2) suboptimal implicit representation for pointclouds.
To improve computational efficiency, the authors p... | Rebuttal 1:
Rebuttal: **Comparison to PoinTr:** There are fundamental differences between PoinTr and our NU-MCC. First, PoinTr is based on an explicit 3D representation and can only predict a fixed number of points. In contrast, our method allows an arbitrary number of output points, providing better flexibility and hi... | Summary: This paper approaches the task of 3D reconstruction given a single RGBD image. It builds off of recent SOTA approach MCC, adding four main contributions: (1) applying attention on only anchor points instead of each query point and using weighted interpolation, (2) using linear projection for “fine” features at... | Rebuttal 1:
Rebuttal: Thank you for finding our proposed method compelling. We address your concerns below.
**F1-score metrics do not match the prior work:** As explained in the footnote in our paper (page 8), there is a bug in MCC codebase F1 calculation. The bug can be traced on their GitHub (as of the writing of th... | Summary: The paper introduces a novel method, NU-MCC, for reconstructing 3D objects from single-view RGB-D images. The proposed approach builds upon the existing MCC method by incorporating two important advancements. Firstly, it introduces a neighborhood decoder to efficiently manage a large number of query points, wh... | Rebuttal 1:
Rebuttal: We thank the reviewer for recognizing the value and promise of our method. We address your concerns as follows.
**Analysis of the effect of different anchor point numbers:** To study this effect, we retrain our model with different numbers of anchor points. Given the constrained timeframe of the ... | Summary: This paper embarks on the task of 3D reconstruction using single-view RGB-D inputs. Building on Multiview compressive coding (MCC), it proposes two enhancements: 1) incorporating nearest neighbors in the decoder to minimize query time, and 2) the utilization of a repulsive unsigned distance function (UDF) as t... | Rebuttal 1:
Rebuttal: Thank you for the overall positive opinion to the contributions introduced in our paper. We address your concerns below.
**Ablation study on neighborhood points:** We conduct an additional ablation study: aggregating different numbers of anchor points and seen points in the proposed neighborhood ... | Rebuttal 1:
Rebuttal: First and foremost, we thank all the reviewers for the time and effort in reviewing our paper. We believe all of the reviewers’ comments and suggestions provide useful feedback to improve the quality of our paper. We hope to be able to clarify some concerns and doubts in this rebuttal and discussi... | NeurIPS_2023_submissions_huggingface | 2,023 | Summary: The paper proposes a novel framework for generating 3D object from single view RGB-D input. Based on the previous work MCC, this work implements a neighborhood decoder that is more efficient and a repulsive udf representation that has higher quality than the previous occupancy field.
Strengths: The major impr... | Rebuttal 1:
Rebuttal: Thank you for the appreciation of the novelties introduced in our paper. We address your questions below.
**Generation resolution:** Since the proposed NU-MCC learns a continuous neural field (UDF), we can query an arbitrary number of query points at inference. Note that the point generation proc... | null | null | null | null | null | null |
Robust Multi-Agent Reinforcement Learning via Adversarial Regularization: Theoretical Foundation and Stable Algorithms | Accept (poster) | Summary: The paper proposes a training objective Lipschitz regularization for obtaining adversarial robustness in cooperative multi-agent RL. To overcome training instablities, the paper reframes the regularization as a Stackleberg game. The paper also shows how the method can be applied to Mean-field approximations of... | Rebuttal 1:
Rebuttal: Dear Reviewer YAu4,
- **Weakness 1: Clarity of figures.**
Thank you for pointing this out! We will increase the font size on the figures to improve the legibility. As for "Evaluation Curves," this refers to the reward achieved when we evaluate the policy in the perturbed environment. We also just... | Summary: The paper introduces a robust Multi-Agent Reinforcement Learning (MARL) framework called ERNIE (adversarially Regularized multi-ageNt reInforcement lEarning). The goal of ERNIE is to address the issue of policy robustness in MARL algorithms. The key contributions of the paper are as follows:
1. Robust MARL Fr... | Rebuttal 1:
Rebuttal: Dear Reviewer fpJY,
Thank you for your very detailed and helpful review! We believe that by addressing your critiques, we have significantly improved the paper. In particular, we clarify the originality of our method, add comparisons with robust MARL methods, and provide ways to clear up the writ... | Summary: Robust RL is attracting great attention, and lately there has been great advancements both in theory and empirically. However, there is only limited work in robust multi-agent reinforcement learning. Under some assumptions, the paper shows the existence of policies with nice properties. Building upon this, the... | Rebuttal 1:
Rebuttal: Dear Reviewer J2jx,
Thank you for your very insightful review! We provide our responses below. We believe that by including discussions on the papers and works you mentioned, the quality of our paper is greatly improved.
- **Question 1: Smoothness terminology.** Indeed, the smoothness in the opt... | Summary: This paper proposes to develop a robust MARL framework based on Lipschitz smoothness of the learned policy (smooth policies). The authors prove the existence of smooth and nearly-optimal policies and propose a framework (ERNIE) to learn robust MARL policy. Experiments are conducted on the two environment: Traf... | Rebuttal 1:
Rebuttal: Dear Reviewer NyS9,
Thank you for constructive comments! We believe the weaknesses you pointed out are addressed by our rebuttal. In particular, we make the following changes: (1) Clarify the experimental setting for action robustness (2) Adding new experiments in robotic simulation tasks (3) Cl... | Rebuttal 1:
Rebuttal: Dear Reviewers,
Thank you for all your helpful comments and critiques! We believe addressing them has significantly improved our work. In particular, we have made several additions to our work:
1. New experiments in multi-agent robotic control.
2. Comparison with new robust MARL algorithms (for... | NeurIPS_2023_submissions_huggingface | 2,023 | Summary: This paper proposes a method to train robust policies in a multi-agent reinforcement learning setting via adversarial regularization. The authors showed that robust policies can be achieved by enforcing Lipschitz continuity with respect to states for robustness to noisy states and with respect to actions for n... | Rebuttal 1:
Rebuttal: Dear Reviewer mUHD,
Thank you for the constructive comments! We provide responses to your critiques and questions below.
- **Weakness 1: ERNIE on mean-field actions.**
Thank you for the suggestion. The results can be seen in Figure 1 of the rebuttal material. We find that applying ERNIE to mea... | null | null | null | null | null | null |
Entropy-dissipation Informed Neural Network for McKean-Vlasov Type PDEs | Accept (poster) | Summary: This work focuses on deriving a uniform-in-time bound to the KL divergence between the NN-derived solution $\rho_t^f$ and the true solution $\bar{\rho}_t$, where $\rho_t^f$ is obtained via a push forward of samples of $\bar{\rho}_0$ with velocity parameterized by the neural network $f$. The neural network $f$ ... | Rebuttal 1:
Rebuttal: We thank the reviewer for making valuable comments. In the following, we address them one by one. Our response will be incorporated in our revision.
> Addressed in the last paragraph of section 4, the similarity in methodology to Shen et al. [2022] ... And perhaps discuss the generalization to M... | Summary: This paper proposes a new method for learning solutions to partial differential equations of the McKean-Vlasov type, and presents a theoretical analysis of the proposed method. In particular, for a certain class of equations, it is shown that the error bound does not increase exponentially in time. This is a s... | Rebuttal 1:
Rebuttal: We thank the reviewer for pointing out the need to highlight the application of our research and to compare with classical method. In the following we address the concerns one by one. Our response will be incorporated in our revision.
> What are the practical applications of the McKean-Vlasov-typ... | Summary: The authors propose a novel method for solving the McKean-Vlasov Equation (MVE), a Partial Differential Equation (PDE) that models the dynamics of stochastic particle systems with mean-field interactions. The new approach, named Entropy-dissipation Informed Neural Network (EINN), leverages neural networks to e... | Rebuttal 1:
Rebuttal: We thank the reviewer for pointing out interesting related works. In the following, we address the concerns one by one. The following discussion will be incorporated in our revision.
> The lack of discussion on the neural operator framework for solving PDEs. This approach has been a considerable ... | Summary: The paper proposed an entropy-dissipation informed neural network for solving certain PDE systems. Using energy dissipation, the paper designed a special structure that propagates information from previous dynamics. They provided a theoretical guarantee for this framework, including the control of the KL diver... | Rebuttal 1:
Rebuttal: We thank the reviewer for carefully reading our paper and for making helpful suggestions. In the following, we address the comments/questions one by one. The responses will be incorporated in our revision.
> The authors specifically take advantage of the special structure of this type of PDEs. On... | Rebuttal 1:
Rebuttal: This pdf file contains the additional experiments on the Porous Medium Equation and a preliminary study of the influence of the width of the NN on EINN's performance.
Pdf: /pdf/55a2e3bb08b15fb07f5d1dcffcec791c3d092a5b.pdf | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Faith and Fate: Limits of Transformers on Compositionality | Accept (spotlight) | Summary: This paper studies the compositional reasoning capability of Transformer models on three kinds of tasks, namely multi-digit multiplication, logic grid puzzles, and a dynamic programming problem. The compositional complexity of a given task, or the number of reasoning steps, can be easily controlled by varying ... | Rebuttal 1:
Rebuttal: We thank reviewer jnqc for their encouraging and detailed feedback. We are glad they highlighted we provide **“new insights in analyzing the successes and failures of Transformers”** using state-of-the-art models, which implies that **“the observed limits are likely to be universal for existing tr... | Summary: This work studies the ability of LLMs for compositional tasks. They propose a formalization called computation graphs. With this formalization, they further propose three different tasks: multi-digit multiplication, Einstein’s puzzle, and a dynamic programming problem. They test multiple LLMs on these three ta... | Rebuttal 1:
Rebuttal: We thank reviewer cZLE for their helpful feedback. We appreciate that Reviewer cZLE found our evaluation study **“comprehensive”** and **“detailed”**, and the designed tasks **“valuable”** for future work. We are also thrilled that Reviewer cZLE found the error analysis to be **“detailed”** and **... | Summary: This paper contributes to the literature on analyzing the limitations of Transformer models in compositional tasks by presenting an analysis based on computation graphs. Based on these computation graphs, some interesting new observations are made, which I believe are worth publishing.
I found the paper a gre... | Rebuttal 1:
Rebuttal: We thank Reviewer euRW for their comprehensive, insightful, and constructive feedback! We’re thrilled to hear that they found our paper **“a great read”**, **“very interesting”**, containing **“interesting new observations”** and **“worth publishing”**.
### On Grokking
We thank Review euRW for br... | Summary: This paper develops empirical and theoretical studies on the limitations of Transformers for solving compositional tasks. It defines computation graphs and quantifies compositional complexity with the graph metrics and uses relative information gain to predict surface patterns learned by Transformer. It conduc... | Rebuttal 1:
Rebuttal: We thank reviewer fn75 for their feedback. We are excited that reviewer fn75 stated our paper includes **“extensive experiments”** and **“profound analysis”** and found the idea of using computational graphs **“inspiring”**. As a result, we would like to seek clarification on some conflicting comm... | Rebuttal 1:
Rebuttal: ## Overall Response
### Discuss Additional Related Work
We appreciate reviewer euRW and reviewer cZLE for bringing additional relevant works to our attention! Given the page limit constraint, we were unable to discuss those works in our initial submission. However, we will incorporate them into t... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Direct Preference-based Policy Optimization without Reward Modeling | Accept (poster) | Summary: This paper investigates learning a policy directly from preferences, i.e., pairwise comparison between trajectory segments. By defining a metric between policy and trajectory segments, the authors show that this problem resembles contrastive learning problems, and they propose to learn a policy by minimizing t... | Rebuttal 1:
Rebuttal: We thank the reviewer for the encouraging comments and constructive feedback. We would like to address your comments below.
**1. Discussion and comparison with some very important alternative methods are missing**
Thank you for pointing out some relevant baselines for our work. While [1] does mo... | Summary: In the context of Preference-based RL (PbRL), the authors propose to learn a preference model that ranks trajectory segments without assuming that the model decomposes into a sum of rewards in an MDP sense. The motivation behind this choice is that the human teachers themselves might not follow a model of the ... | Rebuttal 1:
Rebuttal: We express our gratitude to the reviewer for the inspiring remarks and valuable feedback. We would like to address your comments below.
**1. Some references are missing.**
Thank you for pointing out an important previous work related to our paper. For a detailed discussion on related works, plea... | Summary: This paper presents DPPO, a method for offline Preference-based Reinforcement Learning (PbRL). Rather than using an Offline RL algorithm, DPPO uses a supervised learning objective inspired from contrastive learning. A distance metric is defined, and the policy is trained to have lower distance with preferred s... | Rebuttal 1:
Rebuttal: We extend our appreciation to the reviewer for the positive remarks and insightful feedback. We would like to address your comments below.
**1. Is contrastive learning important?**
In common response 3, we incorporated the baseline algorithms' evaluation results when combined with our smoothness... | Summary: This paper presents an algorithm called DPPO that can learn a policy from paired trajectory comparisons without having to learn a reward function or do online rollouts. PBRL algorithms typically learn a reward function with a Boltzmann comparison model and then optimize it with an online RL algorithm, like PPO... | Rebuttal 1:
Rebuttal: We express our gratitude to the reviewer for their positive remarks and insightful feedback. We would like to address your comments below.
**1. ... examples include LLMs and real robots—although neither are used in this paper ...**
We concur with the reviewer that domains such as LLMs and real r... | Rebuttal 1:
Rebuttal: We are grateful to the reviewers for their supportive and helpful feedback. We plan to address some of the shared concerns raised by multiple reviewers in this section, while individual responses will deal with more specific issues. Moreover, we'd like to emphasize that section C of our appendix s... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
A Unified Approach to Domain Incremental Learning with Memory: Theory and Algorithm | Accept (poster) | Summary: This paper proposes a framework for domain incremental learning that unifies ideas from previous methods in this area. A theoretical analysis is first performed to formally state the problem of domain incremental learning. The theoretical formalization identifies three ways to bound the risk of error on a part... | Rebuttal 1:
Rebuttal: Thank you for the constructive and encouraging comments as well as the insightful questions. We are glad that you find our work ``"original"`` at both ``"theoretical and experimental levels"``, our paper ``"very well-written and easy to follow"``, and acknowledge that our UDIL is ``"the first comp... | Summary: In this paper, the author proposed a unified theoretical framework in order to reveal how the tasks in the domain incremental learning are related. Moreover, the framework unifies various existing methods from the perspective of domain generalization bound and it showed that the unified generalization bound ca... | Rebuttal 1:
Rebuttal: Thank you for the constructive and encouraging comments as well as the insightful questions. We are glad that you find our unified framework ``"powerful"`` and able to ``"explain a variety of previous works"``, our theory ``"sound"``/``"impressive"``, and our paper ``"well-organized"``. Below we a... | Summary: This submission investigates domain incremental learning problem, where the model is adapted to a sequence of new tasks which have different distributions. Starting from some (domain related) generalization bound theory, the combination of three different bounds could be used to address domain incremental lear... | Rebuttal 1:
Rebuttal: Thank you for the constructive and encouraging comments as well as the insightful questions. We are glad that you find our motivation ``"clear"``/``"reasonable"``, our method ``"decent"``/``"creditable"``, our theoretical analysis ``"admirable"``, and our submission ``"a good paper"`` with ``"no ... | null | null | Rebuttal 1:
Rebuttal: # General Response
We thank all the reviewers for their valuable and constructive comments. We are glad they find the problem we address ``"original"``/``"of high relevance"`` (4VUe, Ps1S), our proposed unified framework ``"admirable"``/``"reasonable"``/``"powerful"``/``"original"`` (Ps1S, jN5u,... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
A Metadata-Driven Approach to Understand Graph Neural Networks | Accept (poster) | Summary: Edit: Based on the new experiments, I am increasng my score slightly
They conduct a multivariate regression on the metadata of datasets and performance of multiple GNN models and examine the correlation between GNN performance and data properties.
They show that Gini-Degree, Edge Homogeneity, and In-feature... | Rebuttal 1:
Rebuttal: We appreciate the reviewer's acknowledgment of the novelty and importance of this study. We address your concerns and questions in detail below.
**W1: [Characterization of universal factors.]**
A: This is a good point. These so-called "universal" factors are only salient for the GNN models we co... | Summary: This paper offers an intriguing analysis aimed at analyzing the performance of Graph Neural Networks (GNNs). The analysis explores various graph-theoretic properties of the graph datasets, which may impact the performance of GNNs. The authors investigate 15 graph properties across 15 datasets and examine the c... | Rebuttal 1:
Rebuttal: We appreciate the reviewer's acknowledgment of the soundness and clarity of our paper. We address your individual concerns in detail below.
**W1: [Real-world datasets with varying Gini-Degree.]**
A: We would like to clarify that the real-world datasets used in Section 3 do have varying Gini-Degr... | Summary: Graph neural networks (GNNs) have demonstrated significant success across various problem areas. In this study, the authors propose a metadata-driven approach to examine the sensitivity of GNNs to graph data properties, spurred by the growing availability of graph learning benchmarks. A multivariate sparse reg... | Rebuttal 1:
Rebuttal: We appreciate the reviewer's acknowledgment of the clarity of our paper, the value of our proposed metadata-driven analysis, and our theoretical contribution. We address your concerns in detail below.
**W1: [Selection of the GNN model we use for regression analysis.]**
A: We agree that many poss... | Summary: The authors analyze the sensitivity of GNNs to graph properties such as degree distribution, homophily, etc. Using a multivariate sparse regression analysis (based on Lasso), they find a set of properties that are universal in that they effect all kinds of GNNs and some properties that are local in that they e... | Rebuttal 1:
Rebuttal: We appreciate the reviewer's acknowledgment that we present a good analysis of the effect of the degree distribution. We address your concerns in detail below.
**W1: [Requirement of bounded total degree for datasets.]**
A: We clarify that a bounded total degree does not necessarily imply that ma... | Rebuttal 1:
Rebuttal: Dear Reviewers,
We appreciate your thorough evaluation of our paper and the valuable feedback you have provided. We have carefully considered your comments and responded individually. We would also like to summarize below our response to the major concerns raised in the reviews.
*Clarification o... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Gradient-Based Feature Learning under Structured Data | Accept (poster) | Summary: Summary: The authors consider single index learning with a spiked input distribution, rather than the Gaussian typical in most works in this setting, showing that a modified SGD can learn where vanilla SGD fails and showing better sample complexity under the assumption that the spike is aligned with the true h... | Rebuttal 1:
Rebuttal: We thank the reviewer for their thoughtful feedback and positive assessment.
**Initialization;** We remark that our main objective was to investigate the different complexity regimes arising from structural variations in data, rather than designing optimal algorithms for the learning task at han... | Summary: In this work, the authors investigate the learning of single-index models with a gradient-based algorithm. They analyze structured inputs considering a spiked matrix model for the data. They show that an appropriate normalization of the weights in the training algorithm can help recover the target direction. T... | Rebuttal 1:
Rebuttal: We appreciate the reviewer’s thoughtful evaluation and detailed comments. Following the reviewer’s suggestions, we will improve the general readability in the final version. We address the reviewer’s questions below.
***
**Q1**. We appreciate the reviewer's feedback and we will rephrase the senten... | Summary: This paper investigates the learning dynamics of single index models when the input data has a structured covariance. It demonstrates that the commonly used spherical gradient flow fails to learn the target direction even when the spike and target directions are identical. However, the paper introduces an appr... | Rebuttal 1:
Rebuttal: We thank the reviewer for their valuable feedback and positive evaluation. We address the reviewer's concerns below.
***
**Impractical training procedure**:
While we agree with the reviewer's statement, we emphasize that two-step procedures have become mainstream in the theory of neural networks, ... | null | null | null | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
On the Trade-off of Intra-/Inter-class Diversity for Supervised Pre-training | Accept (poster) | Summary: This paper provides empirical study and theoretical analysis of the class-to-sample ratio in supervised pre-training datasets.
Strengths: 1. This papers reveals an important conclusion - the optimal class-to-sample ratio is invariant to the size of the pre-training dataset. This can be utilized as a guidance ... | Rebuttal 1:
Rebuttal: In the general rebuttal, we added new experiments regarding the three weakness points/questions: using different pre-training dataset (**Figure R1**), using different model architecture (vision transformer, **Figure R2**), and using larger downstream datasets or datasets from other domain (paintin... | Summary: The paper presents a detailed analysis on the intra-class diversity and inter-class diversity when you have a fixed dataset size. The authors develop some theoretical rule based on empirical results.
Strengths: - Provides a rigorous study on pretraining using ImageNet, that makes the empirical motivation clea... | Rebuttal 1:
Rebuttal: >W1/Q1: *All of the experiments are based on using ImageNet as the pretraining dataset. This makes me uncertain of whether the result relates to ImageNet only, as there are particularities in relation to dataset acquisition, labelling, diversity between samples in a class etc. that makes ImageNet ... | Summary: This work studied the impact of the trade-off between the intra-class diversity (the number of samples per class) and the inter-class diversity (the number of classes) of a supervised pre-training dataset. The authors found that with the size of the pre-training dataset fixed, the best downstream performance c... | Rebuttal 1:
Rebuttal: >W1: *The study was limited to fine-tuning on classification tasks only, other types of tasks, such as detection and segmentation, were not covered.*
This is a good point, our current study(theory) focuses on(supports) classification tasks and we would like to leave the exploration of other type... | null | null | Rebuttal 1:
Rebuttal: We thank all of the reviewers for their thoughtful feedback. Based on reviewers’ feedback and questions, we add several experiments and the results can be found in the attached PDF. In particular, we plot the curves of test error rate on downstream tasks across different class-to-sample ratio (sim... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Constraint-Conditioned Policy Optimization for Versatile Safe Reinforcement Learning | Accept (poster) | Summary: 1. This paper introduces the Conditioned Constrained Policy Optimization (CCPO) framework, consisting of two modules: (1) Versatile Value Estimation (VVE) for approximating value functions under unseen threshold conditions, and (2) Conditioned Policy Inference (CPI) for encoding arbitrary constraint thresholds... | Rebuttal 1:
Rebuttal: We thank the reviewer for your time, valuable feedback, and acknowledgment of our contribution. We address the concerns as follows.
> Q1: How reasonable/strong is assumption 2? Especially if latent vectors are highly dimensional. In experiments, what model do you use to estimate the latent vector... | Summary: The paper addresses Safe Reinforcement Learning with varying safety constraint requirements (Versatile Safe Reinforcement Learning). The paper proposes to learn a versatile Q function, which is decomposed as the product of feature functions of state and action, and the latent features of the constraint thresh... | Rebuttal 1:
Rebuttal: We thank the reviewer for your time and valuable feedback. We address the concerns as follows.
> W1: The writing is unclear, especially in the pretraining phase...
Thank you for your valuable suggestions. While we placed the detailed algorithm diagram in the appendix due to page constraints, we ... | Summary: Safe reinforcement learning approaches have so far focused on the case where
cost thresholds are static and known a priori. In contrast, this work addresses
RL environments where cost constraints may change over time. To do this, the
proposed algorithm learns a policy which is conditioned on the given cost
thr... | Rebuttal 1:
Rebuttal: We gratefully thank the reviewer for your time, valuable feedback, and acknowledgment of our contribution. We address your concerns as follows.
> W1: The given example of a car operating in highway vs. urban environments provides some intuition, but I'm not clear on what the cost signal is that w... | Summary: The authors introduce the versatile safe RL problem, in which we want to maximize the expected reward, subject to an expected constraint, but for a range of constraint thresholds $\epsilon\in \mathcal E$ (compared to the traditional safe RL problem, which only aims to find a solution for one $\epsilon$). This ... | Rebuttal 1:
Rebuttal: We thank the reviewer for your time and valuable feedback. We address the concerns as follows.
> Q1: On Page 4, Line 134 - what is $K_f$ is it a constant for $f \in r, c$ Similarly, on line 146, I understand that $Poly(\epsilon, p)$ is a $p$-degree polynomial of $\epsilon$, but shouldn't it have ... | Rebuttal 1:
Rebuttal: Dear reviewers,
We sincerely appreciate your valuable and constructive feedback on our paper. In addition to addressing each reviewer's comments, we'd like to highlight the new examples and experiments we added during the rebuttal phase. You can find the figures we reference in the attached PDF f... | NeurIPS_2023_submissions_huggingface | 2,023 | Summary: The study investigated a versatile, safe reinforcement learning problem and proposed the Conditioned Constrained Policy Optimization (CCPO) algorithm. CCPO may be helpful in ensuring safety under the settings of various constraint limits. Moreover, the experiment results demonstrate that CCPO works well.
Stre... | Rebuttal 1:
Rebuttal: We thank the reviewer for your time and valuable feedback. We address the concerns as follows.
> W1: The code the study provided is not convenient to run; we have to download the code file by file. Could you provide the whole code? Such that we can examine the algorithm better.
We apologize if t... | null | null | null | null | null | null |
PaintSeg: Painting Pixels for Training-free Segmentation | Accept (poster) | Summary: This paper proposed to interactively use inpainting (I-step) and outpainting (O-step) to achieve foreground masks without training.
Formally, the I-step refines the false positive masks while the O-step tackles false negative ones. The authors proposed Contrastive Potential to evaluate the differences between ... | Rebuttal 1:
Rebuttal: We thank the reviewer for the time and effort to review our paper. Our answers to the questions are as follows.
---
**1. The method is not efficient. Lack of comparison of efficiency.**
We acknowledge that PaintSeg is not efficient. We report the computing time for each component of AMCP tested... | Summary: The author proposes a training-free segmentation technique AMCP, which leverages off-the- shell diffusion model to refine the corase segmentation prompt iteratively. AMCP contains three main components: Inpainting step, outpainting step and adversarial updating. The paper gives detail ablation of AMCP and demo... | Rebuttal 1:
Rebuttal: We thank the reviewer for the time and effort to review our paper. Our answers to the questions are as follows.
---
**1. The segmentation ability stems from the diffusion model while the sampling speed of the diffusion model obstacles for real-time applications.**
We acknowledge that PaintSeg i... | Summary: The following work proposes an iterative approach to converting any off-the-shelf inpainting/outpainting capable models to perform segmentation. The proposed method alternates between two approaches to separating the segmentable foreground object from the background. In the I(npainting) step, they mask out the... | Rebuttal 1:
Rebuttal: We thank the reviewer for the time and effort to review our paper. Our answers to the questions are as follows.
---
**1. Does PaintSeg require supervision?**
In the entire pipeline, PaintSeg does not require any supervision. We find it intriguing to demonstrate that satisfactory segmentation ca... | Summary: This paper presents a new unsupervised method for object segmentation, named PaintSeg. Given the initial mask, the proposed PaintSeg adopts adversarial masked contrastive painting (AMCP) to generate the foreground and background masks through the contrastive relations between the original image and the painted... | Rebuttal 1:
Rebuttal: We thank the reviewer for the time and effort to review our paper. Our answers to the questions are as follows.
---
**1. Whether PaintSeg is sensitive to the generative models?**
We consider PaintSeg is not sensitive to generative models. PaintSeg segments objects by contrasting original and in... | Rebuttal 1:
Rebuttal: We sincerely thank the reviewers for their valuable time and efforts.
We include the additional figure (Fig A) and pseudo code (Algo A) in the PDF file along with additional tables in the response for each question. We hope those responses could clarify your confusion, and we are more than happy ... | NeurIPS_2023_submissions_huggingface | 2,023 | Summary: This paper designs a novel unsupervised segmentation method (AMCP) without training via image painting. Its main process is to mask the image based on the specified prompt, recover the image using a generative model, and determine the foreground mask and background mask of the image by comparing the original a... | Rebuttal 1:
Rebuttal: We thank the reviewer for the time and effort to review our paper. Our answers to the questions are as follows.
---
**1. The role and necessity of updating near the foreground-background boundary (adversarial mask updating).**
We would like to further clarify the necessity of adversarial mask u... | null | null | null | null | null | null |
Mnemosyne: Learning to Train Transformers with Transformers | Accept (poster) | Summary: This paper proposes to use transformers to learn optimization policies for training deep learning models. Here, the sequence corresponds to the state sequence (representing the internals of the model) in the trianing process. The intuition is that when we collect many sequences from many training processes, we... | Rebuttal 1:
Rebuttal: Thank you very much for the review and all the comments.We address them in detail below. We are happy to hear that the Reviewer appreciates the amount of work conducted in this work and evaluates it as sufficient for NeurIPS acceptance.
> *cannot quite interpret the associate memory part*
Our th... | Summary: The paper introduces Mnemosyne, a new learnable optimizer that can train neural networks without task-specific tuning. The authors claim that it employs minimal resources and handles larger parameter sets, outperforming LSTM optimizers and matching state-of-the-art ones. Its efficacy is demonstrated in the fin... | Rebuttal 1:
Rebuttal: Thank you very much for the review and all the comments. We address them in detail below.
> *The method's effectiveness on broader model training remains unknown as it only trains or fine-tunes a subset of Transformer layers*
We would like to clarify that our experiments involve in particular pr... | Summary: The paper proposes to use transformers to learn to optimise. The proposed architecture relies on the performer architecture and leverage attention to encode long-term dependencies. The authors discuss two variants, a coordinate-wise and tensor-wise approach, to account for the dependencies between the paramete... | Rebuttal 1:
Rebuttal: >*choices driven mainly by practical considerations…implications of the theoretical analysis*
See "Implications of theoretical analysis on design choices" in the global rebuttal.
> *discussion of other choices*
We chose Adam for non-learnable and LSTM for learnable optimizer baselines due to th... | null | null | Rebuttal 1:
Rebuttal: # New experiments: Comparison with other learnable optimizers
In the rebuttal we provide the comparison with two other classes of learnable optimizers:
1. Those using powerful [S4 memory units](https://arxiv.org/pdf/2111.00396.pdf) based on state space machines as an alternative to LSTMs (and Tran... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Towards In-context Scene Understanding | Accept (spotlight) | Summary: The authors propose a self-supervised approach to adapt models learned via SSL to in-context learning via retrieval. To obtain the best performance, the authors propose to add attention pooling to encourage features attention (even though within an image) and supervised training for retrieval. Through an exten... | Rebuttal 1:
Rebuttal: We thank the reviewer for their thoughtful comments.
Major
1.
* “The features are retrieved independently of each other and there is no mechanism to consolidate the retrieved features into a whole scene output.” We agree with the reviewer that the neighbors are retrieved independently for e... | Summary: This paper investigates in-context learning in computer vision. It proposes the Hummingbird model, which uses a pretraining protocol and a nearest neighbor retrieval mechanism. This model avoids the need for task-specific parameters or finetuning, offering fast and data-efficient adaptation to various scene un... | Rebuttal 1:
Rebuttal: We thank the reviewer for their thoughtful comments.
Weaknesses
* (1) Thank you for the suggestion, we will add a figure showing model components of Hummingbird in the final version. (2) Nearest neighbors are patch-level, we will edit the figure to clarify. (3) The input prompt images are not cu... | Summary: This paper presents a novel in-context learning paradigm based on NN retrieval, which can be adapted for downstream dense vision tasks without requiring fine-tuning or specialized decoders.
Strengths: 1. The proposed method addresses a promising regime for reducing task-specific training requirements with pro... | Rebuttal 1:
Rebuttal: We thank the reviewer for their thoughtful comments.
Weaknesses
1. We focused on the pixel-level dense tasks in this paper due to their relative novelty in the nearest neighbors domain, however, the framework can be used in a straightforward manner for other tasks. For instance, image-level task... | Summary: The paper presents a non-parametric nearest neighbor retrieval method designed to efficiently adapt a pretrained model for dense scene understanding tasks. The fundamental elements include a self-supervised pretraining method, which incorporates attention across images, and a spatial attention-pooling mechanis... | Rebuttal 1:
Rebuttal: We thank the reviewer for their thoughtful comments.
Weaknesses
* We demonstrate the generality of Hummingbird representations by evaluating the model on two different tasks (semantic segmentation and depth prediction), of which one is a pixel-level classification task and the other is a regress... | Rebuttal 1:
Rebuttal: Thank you for your time and effort in reviewing our work.
Please see the attached PDF with a graph showing the overhead of the nearest neighbors lookup vs. the length of the memory bank on an A100 GPU.
Pdf: /pdf/85bcbd16485a448f504759d79f28198b1c2d246d.pdf | NeurIPS_2023_submissions_huggingface | 2,023 | Summary: This submission proposed a new representation and pre-training method, named Hummingbird for its fast adaption ability. Hummingbird could perform inference simply by nearest neighbor retrieval from a prompt of annotated features. Unlike the previous pre-training and then fine-tuning/linear probing protocol, th... | Rebuttal 1:
Rebuttal: We thank the reviewer for their thoughtful comments.
Weaknesses
1. We agree that further scaling Hummingbird (e.g. to ViT-H) would require scaling the dataset in tandem. For this, we could leverage standard extra-large datasets such as ALIGN, and adapt the supervised retrieval objective to the mu... | null | null | null | null | null | null |
StyleGAN knows Normal, Depth, Albedo, and More | Accept (poster) | Summary: This paper proposes a novel method for obtaining "intrinsic images" for an RGB image, where in this context intrinsic images mean mainly depth, normal, segmentation, albedo, and shading maps. To this end, the authors investigate how a pre-trained StyleGAN generator model can be used to output the respective in... | Rebuttal 1:
Rebuttal: Thank you, Reviewer qbrH, for your insightful feedback and acknowledgment of the strengths in our work. We are particularly heartened by your recognition of our proposed approach as an “interesting and valuable method” for investigating the information contained within generative models like Style... | Summary: The paper proposed the method of extracting intrinsic images (Normal, Depth, Albedo, Shading, and Segment) from a pre-trained StyleGAN model without any pre-training or finetuning. It is empirically shown to be better than the SOTA for getting images intrinsic and robust to relighting effects, unlike SOTA.
S... | Rebuttal 1:
Rebuttal: We thank Reviewer 1Cfh for their detailed review and the insights provided. We're grateful to the reviewer for recognizing the novel insights our work brings, especially how the StyleGAN W-space captures intrinsic scene properties without explicit instruction. We are glad to know the reviewer appr... | Summary: The paper uses StyleGAN latent space properties to extract properties like normals, depth, segmentation, albedo and shading. The method does not train a new model to achieve this. The method relies on the properties of w+ space to predict these properties. In practice, the method uses the off-the-shelf predict... | Rebuttal 1:
Rebuttal: We thank the reviewer for their comprehensive review and the effort they took in understanding our work. We acknowledge their concerns and hope to clarify them with this response.
**Abstract:** We appreciate feedback on the abstract and will strive to rewrite it to be more conventional and inform... | Summary: The paper finds that for a GAN generator that is solely trained on images, intrinsic information such as depth, albedo, normal, etc. are implicitly learned during GAN’s training, even though such information is not provided as supervision. More surprisingly, such intrinsic information can be easily retrieved b... | Rebuttal 1:
Rebuttal: Thanks for taking the time to review our paper and for providing a detailed and constructive assessment, particularly thanks for the “fascinating finding”, “technically sound”, “major result”. We’ve tried to address the issues below.
**On the evaluation criteria and dataset diversity:** We apprec... | Rebuttal 1:
Rebuttal: We are grateful for the comprehensive feedback provided by all four reviewers. Here's a synthesized response to the main concerns raised across reviews:
**Scope and Contribution Clarification:** Our primary contribution is to demonstrate that StyleGAN inherently “knows” important scene properties... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Episodic Multi-Task Learning with Heterogeneous Neural Processes | Accept (spotlight) | Summary: The article is innovative in addressing the problem of data insufficiency by introducing the mutli-task learning method into the meta-learning paradigm. In this process, the author identify the issue of task-relatedness between heterogeneous tasks in a single episode and propose a solution called Heterogeneous... | Rebuttal 1:
Rebuttal: *We appreciate Reviewer DuFB for the positive review and helpful feedback. We are glad the review found that our paper is innovative and impressive.*
---
**1. More experimental results on new benchmarks or datasets.**
Thanks for this suggestion. We’ve performed new experiments on the Office31 d... | Summary: This paper proposes HNP: Heterogeneous Neural Processes, an approach that combines multitask learning with metalearning to address the problem of insufficiency of data in the multitask learning setting. The paper introduces episodic multitask learning as a way to exploit task relatedness from meta-training to ... | Rebuttal 1:
Rebuttal: *We thank Reviewer xREG for the positive review and constructive feedback. We are glad that the review found that the quality and clarity of our paper are remarkable; the paper is very well-written and easy to follow.*
---
**1. Influence of the number of tasks.**
Thanks for the insightful quest... | Summary: The paper tackles the problem of learning tasks with insufficient data within the context of multi-task meta-learning. Specifically, the method proposed, namely Heterogenous Neural Processes (HNPs), is designed to leverage information amongst tasks in a single episode to improve performance while retaining lea... | Rebuttal 1:
Rebuttal: *We thank Reviewer LXqe for the positive review and helpful comments. We are glad that the review found that HNPs are interesting, the architectures are well-motivated, and the paper is well-written.*
---
**1. Real-world examples or benchmarks in multi-task meta-learning.**
Episodic multi-task ... | Summary: Learning multiple tasks with limited data poses a significant challenge. Addressing this issue, the paper proposes a novel approach that leverages heterogeneous information across tasks by combining the benefits of multi-task learning and meta-learning. This is achieved through the introduction of hierarchical... | Rebuttal 1:
Rebuttal: *We appreciate Reviewer 21Dc the positive review and helpful feedback. We are glad that the review found that leveraging meta-learning and multi-task learning is intriguing and well-grounded.*
---
**1. Related works [1-3] and MTPs.**
Thanks for this suggestion.
We highlight the differences ... | Rebuttal 1:
Rebuttal: *We thank all reviewers for their insightful comments and supportive suggestions. These feedbacks are constructive in improving this paper’s quality. Here, we provide a global response to summarize the primary merits of this paper and concerns to address as follows.*
---
## **I. Merits of this W... | NeurIPS_2023_submissions_huggingface | 2,023 | Summary: The paper proposes an NP-based approach to episodic multi-task learning called HNPs that uses transformer-based hierarchically arranged inference modules to model class-specific distributions. As a result, the predictions are obtained by a dot product of the input features and the sampled class-specific latent... | Rebuttal 1:
Rebuttal: *We thank Reviewer YPdL for the positive review and constructive feedback. We are glad that the review found that our method is well-motivated and well-explained.*
---
**1. More experimental comparison and discussions.**
We’ve reimplemented recent multi-task and meta-learning methods, such as,... | Summary: The paper introduces a meta-learning paradigm to perform few-shot multi-task learning. It introduces a local latent representation $\omega$, which effectively combines heterogeneous information from various tasks. Additionally, the paper suggests using transformer-structured inference modules to infer both hie... | Rebuttal 1:
Rebuttal: *We thank Reviewer n4KJ for the positive review and thorough feedback. We are glad that the review found that our method is novel, the proposed meta-knowledge and latent representation are well motivated.*
---
**1. Performance with more “shot”.**
Thanks for this question. We’ve performed addit... | null | null | null | null |
Non-stationary Experimental Design under Linear Trends | Accept (poster) | Summary: The paper introduces a new experimental design for non-stationary experiments under the assumption that the expected potential outcomes are linear in time.
The proposed design consists of two phases: a switchback experiment that alternates at each time point and a UCB-type algorithm for multi-armed bandits. P... | Rebuttal 1:
Rebuttal: Thank you for your valuable feedbacks! We refer our response to the linear assumptions, the extension to higher order polynomial trends, and the switchback design to the "global response" on top of page. We detail responses to the other comments as follows.
a. About the Pareto optimality.
Sorry ... | Summary: This work is concerned with designing an adaptive experiment in the presence of a linear or polynomial trend in the potential outcome. The authors describe the natural tension between estimation estimation error and regret and describe the Pareto optimal objective. They then introduce an optimistic procedure t... | Rebuttal 1:
Rebuttal: Thank you for your valuable feedback and helpful comments on our work. We sincerely appreciate your careful read of our paper, and we are grateful for the opportunity to address your questions and concerns. Below, we provide our detailed response:
1. The Selection of $\alpha$.
We agree with your... | Summary: This paper studied the problem of non-stationary experimental design. The authors adopt a linear model assumption, where the reward associated with each arm exhibits a linear relationship with time. They concern two objectives: accurate estimation of dynamic treatment effects and regret. They propose a two-sta... | Rebuttal 1:
Rebuttal: We sincerely appreciate your helpful comments and careful read of our work. Thank you for providing us with an opportunity to address your concerns.
1. Comparing with [A].
We completely agree that [A] was the first to introduce Pareto optimality to characterize a similar trade-off. Allow us to c... | null | null | Rebuttal 1:
Rebuttal: We thank all the reviewers for your time, valuable feedback, and insightful comments. In addition to the responses to each reviewer, we would like to use this part to provide clarifications on three specific comments.
a. Assumption on linear trends.
Linear trends are indeed a strong assumption, ... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Token-Scaled Logit Distillation for Ternary Weight Generative Language Models | Accept (poster) | Summary: The work proposes modifications for Quantization-Aware Training (QAT) that utilize knowledge distillation and takes into account token-scaled logit distillation for preventing overfitting. They show the motivation for their method through a deep analysis and present the empirical merits of this method via lang... | Rebuttal 1:
Rebuttal: Thank you to the reviewer for your valuable and constructive comments. To address the queries posed, our explanations are provided below.
**Comment 1) Expanding TSLD Evaluation Across NLU/CSQA Tasks
(Major Comments 2,5)**
**Answer 1)** To validate the effectiveness of TSLD, we broadened our comp... | Summary: This paper proposes to improve the quantization of generative language models via an adaptively scaled logit distillation method.
With an intuitive illustration in Figure 1 and validation results with attention heatmap discrepancy, the paper identified a cumulative quantization error introduced by the masked ... | Rebuttal 1:
Rebuttal: We are grateful for the thorough and thoughtful reviews provided by the reviewer. Below are our responses to the raised questions.
**Comment 1) Paper organization (Weakness 1)**
**Answer 1)** We concur with the reviewer's feedback regarding the shortcomings in the logical flow, structure, and co... | Summary: This paper proposed the token-scaled logit distillation method to counteract accuracy loss caused by using Quantization-Aware Training（QAT） methods for generative language models . (1) The author proposed the logit distillation method to overcome the limitation in token prediction recovery by modifying the int... | Rebuttal 1:
Rebuttal: We sincerely appreciate the insightful feedback provided by the reviewers. In response to the raised questions, our answers are as follows.
**Comment 1) Clarification of Baseline Settings and Expanding TSLD Evaluation Across Tasks and Models (Weakness 1, 2)**
**Answer 1)**
FP 32 baseline settin... | Summary: This research paper focuses on Generative Language Models (GLMs) and the challenges of their deployment due to their large model sizes. The authors introduce a novel approach to tackle the challenges faced by Quantization-Aware Training (QAT) in fine-tuned GLMs.
**Problem**: GLMs, despite their impressive per... | Rebuttal 1:
Rebuttal: We sincerely appreciate the insightful feedback provided by the reviewer. In response to the raised questions, our answers are as follows.
**Comments 1) Expanding TSLD Evaluation (Weakness)**: To address the concern about the limited datasets we initially used, we've expanded our evaluation to in... | Rebuttal 1:
Rebuttal: We sincerely thank the reviewers for their invaluable feedback. To effectively respond to your inquiries, we have attached a PDF containing new experimental results and figures. Detailed responses to each question raised by the reviewers can be found in the respective sections, with references to ... | NeurIPS_2023_submissions_huggingface | 2,023 | Summary: The article introduced a new approach for LLM Quantization Aware Training (QAT).
Strengths: The paper is well written and the proposed idea is described clearly.
Weaknesses: I have some concerns about the evaluation in the paper. I
First, the number of tasks used for accuracy/perplexity comparison between ... | Rebuttal 1:
Rebuttal: We appreciate the reviewer’s valuable feedback. Responses to your comments are provided below.
**Comment 1) Expanding TSLD Evaluation Across NLU/CSQA Tasks and Models**
>Weakness 1) “the number of tasks used for accuracy/perplexity comparison between float and the proposed method is too small. Th... | null | null | null | null | null | null |
PI-FL: Personalized and Incentivized Federated Learning | Reject | Summary: This paper proposes and analyzes a mechanism for incentivizing data contributions to federated learning systems in a context where contributors consider both monetary incentives for data contributions and the ability to experience the benefits of personalized systems.
The paper provides background on incentiv... | Rebuttal 1:
Rebuttal: **Answer 1.**
We deeply appreciate the reviewer's insightful questions, and we would like to address them in a consolidated response.
In the rapidly evolving landscape of AI, several AI marketplaces such as ModelPlac [1], GravityAI [2], AWS AI Marketplace [3], and AI Marketplace [4] are emerging... | Summary: This paper proposed PI-FL, an incentive-aware federated learning method that produces clustered personalization of client models. This work uniquely considers the client contribution assessment in the multi-objective setting of clustered personalization. The clients are also given the freedom to choose the clu... | Rebuttal 1:
Rebuttal: **Answer 1.**
We appreciate the reviewers for raising this important issue. In our future work, we intend to dive deeper into this topic, Specifically, we will explore variable pricing methods. In this paper, we have followed other state of art token-based incentive federated learning frameworks ... | Summary: In this paper, the authors propose a novel clustering-based pFL combined with a token-based incentive mechanism to address incentive provision in pFL. Specifically, the proposed method clusters clients based on their cluster preference (the distribution similarity), which maximizes their contribution to the cl... | Rebuttal 1:
Rebuttal: Ck* is the same as Ck, and theta_i originates explicitly from algorithm 2 in line 30. Theta represents the utility of the client, as defined in equation 6. To eliminate any possible confusion, we will promptly modify the symbols and rectify Ck* without delay.
Furthermore, for the readers who may ... | Summary: This paper presents PI-FL, a new cluster-based personalized Federation Learning (pFL) approach. The new idea is the first to propose to consider motivation and personalization as interrelated challenges and address them through incentive mechanisms that promote personalized learning. PI-FL let clients provide ... | Rebuttal 1:
Rebuttal: We extend our sincere appreciation to the reviewer for providing their valuable comments. We would like to address any misunderstandings and concerns regarding our work.
To address the reviewer's concerns about the convergence speed, we have uploaded new evaluations (Figures 1 and 2), which demon... | Rebuttal 1:
Rebuttal: Additional evaluation results as per the reviewer's recommendations.
Pdf: /pdf/02185ac587dfcc900772ba1af0b6407686a9aa83.pdf | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Markovian Sliced Wasserstein Distances: Beyond Independent Projections | Accept (poster) | Summary: The authors focus on the study of sliced Wasserstein distances.
It consists in projecting measures onto Euclidean lines, such that OT between these projected measures has low complexity because it amounts to sort samples.
The projections are unit vectors of the d-dimensional sphere of R^d, and in practice on... | Rebuttal 1:
Rebuttal: We would like to show our appreciation for your valuable feedback and comments.
**Q16**: To my mind, the use of a Markov chain does not seem to bring a significant improvement... all W2 distances are of order of $10^{-5}$.
**A16**: This is due to the small scale of the initial supports. The sc... | Summary: This paper introduces Markovian Sliced-Wasserstein distances, which imposes a Markovian structure on the projecting directions. The author impose different distributions on the transition kernel.
Strengths: - The paper is well written and easy to follow. The idea is well-presented.
- Each important theoretica... | Rebuttal 1:
Rebuttal: We'd like to express our gratitude for the constructive feedback. We're here to provide responses as outlined below. We are eager to participate in further discussions.
**Q13**: The experiments in generative modeling are not very convincing, would you have the time to do some other experiments? I... | Summary: This paper presents a novel variant of the sliced Wasserstein (SW) distance, addressing the challenges of redundancy and loss of metricity in other SW distance variants, such as Max-K-SW. The proposed method introduces the use of Markov chain-based projection directions selection to overcome these issues.
The... | Rebuttal 1:
Rebuttal: We begin by expressing our gratitude for the insightful feedback and comments provided in the review. In response, we offer the following explanations and answers to your inquiries. We welcome any further questions and are open to engaging in a more in-depth discussion on the matter.
**Q9**: One ... | Summary: The authors propose a variant of the sliced Wasserstein distance that samples the projection directions from a Markov chain instead of uniformly from the hypersphere. They propose two Markov chain designs: the first chooses the next projection vector to be orthogonal to the previous one and the second is deter... | Rebuttal 1:
Rebuttal: First, we would like to thank the reviewer for the time and constructive feedback.
**Q1**: In high dimensions, two random vectors will be almost orthogonal with high probability. Therefore, oMSW should be basically equivalent to SW.
**A1**: Regarding the concentration result, two uniformly drawn... | Rebuttal 1:
Rebuttal: We would like to begin by extending our gratitude to the reviewers for their valuable time and feedback. In our response, we will outline a summary of additional experiments in the rebuttal PDF and provide answers to questions frequently raised by the reviewers. Further inquiries have been address... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Streaming PCA for Markovian Data | Accept (spotlight) | Summary: The submission considers the task of *streaming* PCA, estimating the largest eigenvector of a covariance matrix in an online manner as new data points arrive. The focus is on theoretical analysis of a classical method, Oja's Algorithm, in the hitherto unconsidered non-IID setting.
A seemingly strong restrict... | Rebuttal 1:
Rebuttal: Thank you for commenting on the novelty of our analyses and the fact that they can find applications in other online problems. We provide answers to specific queries below.
1. *Regarding wider interest*:
Lines 19-32 mention the connection to distributed PCA where the Markov chain arising from a r... | Summary: This paper studies the problem of streaming PCA, where the data points are revealed to the algorithm one at a time and the goal is to estimate the top eigenvector of the underlying covariance matrix.
Crucially, the data points are sampled from a Markov chain and thus do not satisfy the typical assumption of be... | Rebuttal 1:
Rebuttal: Thank you for your kind words regarding our results' novelty and the importance of the problem.
1. *Regarding writing*: We will correct the typographical errors and incorporate the suggestions. We will also add a simplified version of Theorem 1 with reasonable values substituted for the parameter... | Summary: This paper focuses on a particular algorithm for PCA known as the Oja's algorithm.
This is an iterative algorithm that operates in a streaming manner.
The goal is to, when the data stream is Markovian, estimate the principal component of the covariance matrix of the unknown covariate vectors.
The main quest... | Rebuttal 1:
Rebuttal: Thank you for your encouraging words about removing the logarithmic dependence in our rate.
1. *Regarding the example Markov chain*: Thank you for this suggestion. Due to the interpretable relationship between the spectral gap and the mixing time, we do use a similar Markov chain for our experim... | Summary: This paper analyzes the performance of Oja'a algorithm for streaming PCA when the samples come from an irreducible, aperiodic, and reversible Markov chain. In this work the authors avoid using downsampling to reduce dependency, and their result improves on that proposed in [3] by a poly-logarithmic multiplicat... | Rebuttal 1:
Rebuttal: Thank you for your kind words regarding the novelty of our analyses and wider applications for online problems.
1. *Regarding the estimation of the mixing time*: [A] provides an offline algorithm for estimating the mixing time of Markov chains. They show that at least a constant times $\frac{\ta... | Rebuttal 1:
Rebuttal: We want to first thank all the reviewers for their valuable suggestions and insightful feedback. We believe we have addressed nearly all of their main technical questions. We give brief sketches to show how to extend our analysis to **non-reversible** Markov chains, the **non-zero mean** data sett... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
BCDiff: Bidirectional Consistent Diffusion for Instantaneous Trajectory Prediction | Accept (poster) | Summary: This paper proposed a diffusion-based model specifically for instantaneous trajectory prediction. It leverages the consistency of the prediction for unobserved historical trajectory and the future trajectory to improve the accuracy of the future trajectory. Regarding the noise in the prediction wh, especially... | Rebuttal 1:
Rebuttal: >Q1: The authors let the system access the unobserved historical trajectories during the training time as one of the supervision signals. It would be necessary to rule out the influence of unobserved historical trajectories when evaluating the boost brought by the system and prove the performance ... | Summary: In this paper, the authors propose a diffusion framework for instantaneous trajectory prediction (only observing two frames in history). Specifically, they develop two coupled diffusion models by designing a mutual guidance mechanism that can bidirectionally and consistently generate unobserved historical traj... | Rebuttal 1:
Rebuttal: >Q1: The proposed solution is time-consuming (4 seconds for predicting trajectories once). This problem appears to be particularly serious, especially in the area of autonomous driving. The long computation time should be caused by the model design (diffusion model and using an iterative way to ge... | Summary: This paper focuses on the trajectory prediction in an instantaneous observation by generative diffusion models. Specifically, the authors propose a bidirectional consistent diffusion framework, named BCDiff, to address the issue of temporal information scarcity. BCDiff employs two coupled diffusion models are ... | Rebuttal 1:
Rebuttal: >Q1: The compared baselines are not the newest methods. The authors should add more baselines with high accuracy, such as SocialVAE, SIT, MemoNet. These baselines could be trained under the setting of the instantaneous trajectory prediction, such as observing 2 steps and predicting the next 12 ste... | Summary: The paper presents a bi-directional consistent diffusion approach for handling instantaneous trajectory prediction, i.e. trajectory prediction without requiring a large number of observation time. It relies on two coupled diffusion models to generate historical and future trajectories. The forward diffusion mo... | Rebuttal 1:
Rebuttal: >Q1: The task of instantaneous trajectory prediction is not standard. It would value for the approach to be evaluated for the standard trajectory prediction protocol. Then it would be allowed to be compared with a wide range of methods. This is a particularly interesting point since the diffusion ... | Rebuttal 1:
Rebuttal: We really thank the reviewers for their valuable feedback. We are encouraged as reviewers found that our idea is novel and interesting (ytbj), promising (VvSx), solid (WQWs), and intuitive (JXan), our experiments show good results (GWuJ, ytbj, JXan) and are thorough (JXan), our paper is easy to fo... | NeurIPS_2023_submissions_huggingface | 2,023 | Summary: The authors propose a bidirectional consistent diffusion framework for instantaneous trajectory prediction. Specifically, two coupled diffusion models are employed to predict unobserved historical and future trajectories jointly. In addition, a gating mechanism is presented to deal with the noisy situation at ... | Rebuttal 1:
Rebuttal: >Q1: I am concerned about the consistency between the motivation and experimental setting.
A1: Thanks for raising this concern which helps us to clarify the experimental settings. We guess this concern is caused by the misunderstanding on the "commonly used leave-one-out (LOO) setting." Here, th... | null | null | null | null | null | null |
Hyperbolic VAE via Latent Gaussian Distributions | Accept (poster) | Summary: The paper proposes a pseudo-Gaussian distribution over hyperbolic spaces, which is suitable for constructing VAEs with hyperbolic latent spaces. This pseudo-Gaussian distribution arrives by replacing the usual Euclidean distance with a KL-divergence (this locally approximates the hyperbolic geodesic distance).... | Rebuttal 1:
Rebuttal: We sincerely thank for providing constructive feedback. We answer the questions about i) analysis on numerical stability, ii) the parameter $c$ and $\beta$ of VAE, and iii) the approximation performance of KL divergence.
---
**Q1) Quantitative analysis on the numerical stability of hyperbolic di... | Summary: This paper proposes to utilize a set of Gaussian distributions as the latent space of a VAE model. One benefit of doing so is that this forms a hyperbolic space under the Fisher information metric. A key insight is that this Gaussian manifold corresponds to the Poincare half-plane model of hyperbolic geometry.... | Rebuttal 1:
Rebuttal: We sincerely thank the careful comments for the novelty of our method. We address the questions about the i) novelty on numerical stability, ii) drawbacks of HWN, and other minors.
---
**Q1) PGM normal is another novel way to stabilize hyperbolic distribution.**
We thank for introducing a work ... | Summary: In this paper, the authors propose a novel Gaussian manifold generative autoencoder (GM-VAE) whose latent space is a statistical manifold formed by univariate Gaussian distributions. They also introduce a pseudo-Gaussian manifold normal distribution on the latent, which is easy to sample from and improves trai... | Rebuttal 1:
Rebuttal: We acknowledge with sincere help for improving our work. We provide the answers for the questions about i) the comparable results in model-based RL and ii) large-scale datasets.
---
**Q1) The experimental results of the GM-VAE in the model-based RL task are only comparable to those of competitor... | Summary: Some VAE methods have been proposed to model the latent space as a hyperbolic space for modeling the structured representation of data. This paper proposes a hyperbolic distribution by using KL divergence to measure the geodesic distribution of Poincare normal, by using KL divergence as the local approximation... | Rebuttal 1:
Rebuttal: We thank for the sincere feedback. We address the issues about i) the modeling capability with local approximation, ii) visualization of our method and iii) analysis on numerical stability.
---
**Q1) It is not clearly studied and discussed how the proposed pseudo Gaussian manifold normal distri... | Rebuttal 1:
Rebuttal: We appreciate all reviewers for valuable comments which can improve our work. We reply to each reviewer with independent comments. We also provide supplementary material to complement the response to reviewers. The supplementary material consists of a PDF with figures and tables, and answers to fr... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Core-sets for Fair and Diverse Data Summarization | Accept (poster) | Summary: The study presents improved core-set construction algorithms for maximizing diversity under partition constraints.
Two diversity measures were considered: sum-of-pairwise distances and sum-of-nearest-neighbor distances.
The algorithms are validated through real-world application to time-sensitive message data,... | Rebuttal 1:
Rebuttal: Thank you for your positive feedback mentioning that our paper provides significant insights, effective algorithms, sound theoretical results supported by experiments, and, indeed, it can be of broader interest of the NeurIPS community and the conference attendees. Also thank you for all the feedb... | Summary: A set of metric data points is comprised of m disjoint groups. The manuscript seeks for each group a subset of some given size such that the union of subsets the minimal pairwise distance (or sum of pairwise / nearest neighbour distances). For constant factor approximation of minimising the sum of pairwise di... | Rebuttal 1:
Rebuttal: We thank you for your review and positive feedback.
Precisely as you say, the goal in our paper is to provide an intermediate summary (i.e., a core-set) such that the union of the core-sets contains a good solution relative to the whole data. Later on, as a post-processing, one can use any exact ... | Summary: This paper gives coresets for two different fair diversity measures (FDM). Fairness is defined as a partition matroid where the number of items from each part is fixed. The authors give a coreset for FDM under the sum-pairwise distance. They also give a coreset for sum-NN-distance. The proposed algorithms are ... | Rebuttal 1:
Rebuttal: We thank you for reviewing our paper and for the kind words on the paper merit.
We want to clarify that the existing O(1) approximation is in the “standard offline setting” of the problem and it **does not** give a “core-set construction algorithm” (Please refer to Table 1). This is because thos... | Summary: This paper investigates data summarization problems. Given a set of points in metric space, the goal is to select k points that are diverse. Diversity could be measured as the sum of squared distances, the sum of nearest neighbor distances, etc.
Constant approximation algorithms for these problem are known. ... | Rebuttal 1:
Rebuttal: We thank you for reviewing our paper.
We emphasize that our core-set results **do imply** streaming and distributed algorithms for the considered problem. As mentioned in our introduction [in the paragraph right before Section 1.1], the application of core-sets to streaming and distributed models... | Rebuttal 1:
Rebuttal: We sincerely thank all four reviewers for the positive feedback and their valuable comments. We provided a response individually to all the reviewers. Please let us know if there are additional questions that we can address during the discussion phase.
Pdf: /pdf/a9cb988e13b0222633f9088a4d56159a11b... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
FedL2P: Federated Learning to Personalize | Accept (poster) | Summary:
This study primarily concentrates on refining hyperparameters in federated learning, especially concerning non-IID data. The researchers highlight two key hyperparameter groups—Batch Normalization (BN) and selective update hyperparameters—that yield impressive outcomes.
BN hyperparameters help mitigate featu... | Rebuttal 1:
Rebuttal: We thank the reviewer for the detailed constructive feedback. We further evaluate our approach on a speech dataset, clarify our setups, discuss the position, costs and limitations of our approach.
> Q4.1) Generalizability of our method to different FL scenarios
Our method can handle both label p... | Summary: This paper proposes a new personalization strategy for the Federated Learning setting. Their key idea is to train a meta-network that takes as input the client's data and outputs optimal hyperparameters for that data. These optimal hyperparameters can then be used to personalize the pretrained model. The hyper... | Rebuttal 1:
Rebuttal: We thank the reviewer for their valuable feedback. We provide clarifications and evaluate on a speech dataset, including the case where some clients are not used in learning the meta-nets.
> Q3.1) Clarify client/data split.
In our CIFAR10 experiments, 20% of the train data is used to construct t... | Summary: The paper introduces a novel approach that utilizes meta-learning to learn hypernets that aim to determine the optimal fine-tuning hyperparameters for individual clients. The proposed idea is intriguing and holds promise. The results presented in the paper are compelling. However, the paper would benefit from ... | Rebuttal 1:
Rebuttal: Thank you for the constructive feedback and valuable time spent on our paper. We will revise the paper accordingly to address the points you have made to improve clarity.
> Q2.1) The proposed method is limited to one certain type of personalized federated learning, i.e., fine-tuning local models.... | Summary: The paper introduces a method for fine-tuning a client's model based on a pretrained model using its local data. The proposed personalization approach involves fine-tuning the learnable weights and biases (assume, there are $M$ such layers) as well as the means and standard deviations (SDs) of batch normalizat... | Rebuttal 1:
Rebuttal: We thank the reviewer for their time and for providing constructive feedback to improve our work. We appreciate the questions and all the detailed minor comments and suggestions and will adjust them accordingly in the revised paper, fixing algorithmic, notation, and text ambiguities and errors.
>... | Rebuttal 1:
Rebuttal: We would like to express our appreciation and gratitude to the reviewers for the detailed comments and helpful feedback to further clarify and improve our paper. We will revise the paper accordingly, adding the new evaluations, clarifying ambiguities, and fixing notation & spelling errors. We have... | NeurIPS_2023_submissions_huggingface | 2,023 | Summary: The paper proposes FedL2P, a federated learning framework for learning personalized strategies in the federated meta-learning problem. The authors introduce meta-nets to estimate hyper-parameters for personalized fine-tuning, such as batch normalization parameters and learning rates, based on client-specific d... | Rebuttal 1:
Rebuttal: Thank you for the detailed comments and feedback. We hope our response below will address the reviewer’s concerns and questions. We will revise the paper acoordingly to include these discussions.
> Q5.1) Calculating meta-gradients involves time-consuming computations of second-order derivatives. ... | null | null | null | null | null | null |
Scaling Riemannian Diffusion Models | Accept (poster) | Summary: This paper considers the diffusion model on Riemann space. They propose several strategies for numerically computing the heat kernel on Riemannian symmetric spaces. They also propose two techniques to sample from their proposed heat kernel approximations, including a rejection sampling method and a probability... | Rebuttal 1:
Rebuttal: We thank the reviewer for their review and hope to address any concerns the reviewer may have.
* On novelty of section 3 techniques: we do not claim novelty for the heat kernel computation methods, as these have existed in the physics and applied mathematics fields for decades. Rather, as mention... | Summary: Developing diffusion model on Riemannian manifolds has received attention recently. However, the current proposals did not scale well to high dimension as they used sliced-score matching or denoising score matching with bad approximations of the heat kernel. This work focuses on using better approximations of ... | Rebuttal 1:
Rebuttal: We would like to thank the reviewer for their positive comments!
* On transition kernel: we will update our final version of the paper to include a more detailed explanation for how the transition kernel is derived from the SDE. In particular, we will emphasize that it comes from considering the ... | Summary: The theory of stochastic differential equations on vector spaces forms the basis for recent diffusion generative models, and recent efforts have attempted to extend this framework to stochastic differential equations on Riemannian manifolds (Riemannian diffusion models). From a practical computational perspect... | Rebuttal 1:
Rebuttal: We would like to thank the reviewer for their thorough and interesting comments. We are glad that the reviewer appreciated our presentation, and we hope to address any concerns.
## Addressing weaknesses:
* On results: we emphatically agree that showing real-world applications of geometric machine... | Summary: In this work the authors leverage properties of symmetric spaces to get better approximation of the heat kernel and therefore of the denoising score matching loss for training diffusion models on compact manifolds.
In particular, using the concept of maximal torus and roots of the symmetric space, the Laplacia... | Rebuttal 1:
Rebuttal: We would like to thank the reviewer for their detailed review and insightful comments. We are glad that the reviewer found the paper enjoyable to read, and we hope to address any concerns the reviewer has.
* On the generalization to spheres: we will include the generalization to split-rank symmet... | null | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Meta-learning families of plasticity rules in recurrent spiking networks using simulation-based inference | Accept (poster) | Summary: The authors propose a novel meta-learning approach based on Bayesian filtering for discovering synaptic plasticity rules that entrain a E-I recurrent spiking neural network to satisfy a set of predefined criteria (e.g. the network firing rates occupy a reasonable regime at steady-state, with appropriate variab... | Rebuttal 1:
Rebuttal: We thank the reviewer for their insightful review. We have already addressed the more major concerns above, but elaborate here on the more technical and philosophical points that we couldn’t cover in the common rebuttal:
### 1/ Local optimization vs fSBI:
The reviewer raised the point that samp... | Summary: The authors present the filter Simulation-Based Inference (fSBI) method, aimed at automating the discovery of families of rules that establish plausible dynamics within large spiking networks. This study seeks to overcome the challenges presented by human intuition and the overly simplified assumptions that ha... | Rebuttal 1:
Rebuttal: We thank the reviewer for their positive assessment of our method and its usefulness.
### 1/ Compute requirements:
While we agree that this method is not computationally cheap, this is also true of other meta-learning methods. Notably, our study can be considered as a form of pre-training of plas... | Summary: The manuscript introduces a novel way of sampling plasticity rules that obey certain plausibility constraints. Specifically, the authors focus on networks of excitatory and inhibitory spiking neurons and use a simulation-based inference (SBI) technique to filter plasticity rules based on the behavior of the ne... | Rebuttal 1:
Rebuttal: We thank the reviewer for the thoughtful review, please find below our attempt to address the points raised that are not in the common rebuttal.
### 1/ Biases in fSBI:
The reviewer's concern about biases of the method and the influence of the filtering order is valid, as fSBI may suffer from the ... | Summary: How neurons update their synaptic strengths is an important and active area of research in neuroscience, with potential applications in ML. Much work has been done using analytically tractable mean-field models, in large part because the search space for possible rules is intractably large. Here, the authors ... | Rebuttal 1:
Rebuttal: We thank the reviewer for the constructive feedback, and for noting the potential of the approach. Please find a detailed response to the points that are not addressed in the common rebuttal below.
### Choice of metrics and their plausible ranges:
We addressed most of this point in the common... | Rebuttal 1:
Rebuttal: We thank the reviewers for their constructive comments and their positive appraisal of our study, calling it "a good study that contributes to the field" (r. vJPD), "a nice extension of simulation-based inference" (r. 5no9), with "insightful” analysis” (r. jNn6), and the discovery of non-mean-fiel... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Pseudo-Likelihood Inference | Accept (poster) | Summary: This method perform simulation-based inference with a hybrid of previous methods, in this case integrating a neural likelihood-type step within a SMC-ABC algorithm for performing posterior inference. It considers use of distance metrics such as the Wasserstein and the Maximum Mean Discrepancy to compare simula... | Rebuttal 1:
Rebuttal: ### Answer to Reviewer 7dVy
We thank the reviewer for the extensive feedback.
**Remarks and Questions:**
> [explain] computational steps used to estimate the relevant divergences.
We will add:
*We follow [1] to approximate the MMD for two discrete probability measures, whereas entropy-regularize... | Summary: This paper proposes a Pseudo-Likelihood Inference approach that brings neural approximation into ABC. This new approach can condition on a variable number of observations unlike previous neural-based SBI approaches.
Strengths:
* The incorporation of Sequential Monte Carlo into SBI is a strength. SMC seems l... | Rebuttal 1:
Rebuttal: ### Answer to Reviewer dHap
We thank the reviewer for the insightful questions and feedback.
**Remarks and Questions:**
> Is there a better way to highlight that $N\neq M$ for PLI?
Indeed, choosing $M > N$ is a strength of PLI and ABC that we could have highlighted more. We expect both methods... | Summary: This paper proposes an approch to deal with posterior inference when the likelihood is intractable, which is often the case for problems involving simulators of physical processes.
The paper investigates Approximate Bayesian Computations (ABC) methods, and it proposes incorporating neural approximations throug... | Rebuttal 1:
Rebuttal: ### Answer to Reviewer 9BA6
We thank the reviewer for the valuable feedback. We appreciate the *"overall positive opinion"* on the methodology but would like to summarize the main contributions of PLI again as the reviewer *"find[s] it difficult to assess the novelty of this work"*
We want to mot... | Summary: This paper introduces Pseudo-Likelihood Inference (PLI) as a novel method inside the simulation-based inference (SBI) family. SBI methods are used to estimate model parameters when the likelihood function is unknown or intractable. Existing SBI methods, such as Approximate Bayesian Computation (ABC) and Sequen... | Rebuttal 1:
Rebuttal: ### Answer to Reviewer AzJv
We thank the reviewer for the valuable feedback and the relevant references that we will include in the final version of the paper.
**Remarks and Questions:**
> The changes induced by replacing the KL divergence with the MMD and the Wasserstein distance should further ... | Rebuttal 1:
Rebuttal: ### Answer to all Reviewers
We thank the reviewers for their insightful feedback. We are pleased that the **novelty and significance** of the paper are recognized by the reviewers.
- **XuP4:** A number of different non-trivial steps are performed [...] and I believe this deserves commendation
- *... | NeurIPS_2023_submissions_huggingface | 2,023 | Summary: This work aims at proposing a sequential Monte Carlo scheme to perform approximate Bayesian computation in the context of simulation based inference. Previous works focused on directly modeling posterior distributions whereas this approach presents a novel method to approximate likelihoods utilizing kernel den... | Rebuttal 1:
Rebuttal: ### Answer to Reviewer XuP4
We thank the reviewer for the valuable feedback regarding our proposed method.
**Remarks and Questions:**
> I understand the necessity of [introducing IPMs over the KL] due to the likelihood being unavailable in SBI; however, this does undercut a lot of the impact that... | null | null | null | null | null | null |
On the spectral bias of two-layer linear networks | Accept (poster) | Summary: The paper explores the convergence of gradient descent for two-layers linear neural networks.
Strengths: The paper substantially advances the knowledge on the topic.
Weaknesses: Missing reference: E. Boursier et al, Gradient flow dynamics of shallow ReLU networks for square loss and orthogonal inputs.
Pleas... | Rebuttal 1:
Rebuttal: We thank the reviewer for the valuable feedback.
> Missing reference: E. Boursier et. al., Gradient flow dynamics of shallow ReLU networks for square loss and orthogonal inputs. Please mention this reference as they already highlight the low-rank structure of the optimal solution in the nonlinea... | Summary: The manuscript studies the evolution of weights in a two layer *linear* NN under gradient flow subject to certain initialization. Results are provided that show how different initialization scales result in qualitatively different behavior of the learned weights.
Strengths: The main points of the manuscript ... | Rebuttal 1:
Rebuttal: We thank the reviewer for the valuable feedback and interesting questions which help us in clarifying and improving the paper.
> However, it seems that at least point (ii) in Thm 3.1 may crucially depend on the 0 initialization of $W_2$.
Certainly this is the case for analysis of the random fe... | Summary: This paper explores the behavior of two-layer fully connected networks with linear activations trained using gradient flow on the square loss. The authors reveal that the optimization process is influenced by the initial parameter values and exhibits an implicit bias. Notably, in the case of small-scale initia... | Rebuttal 1:
Rebuttal:
We thank the reviewer for the valuable feedback and interesting questions which help us in clarifying and improving the paper.
> Restrictive setting: In this work, the theoretical analysis is conducted on a two-layer linear network, which is too restrictive. Many existing works have more genera... | Summary: The paper provides an analysis of an implicit bias in the hidden layer for two-layer linear networks under a small-scale initialization regime.
Strengths: This paper shows the hidden layer is biased in a small-scale initialization under over-parameterized two-linear networks.
Weaknesses: The novelty and re... | Rebuttal 1:
Rebuttal: We thank the reviewer for the valuable feedback.
> The novelty and results significance are not enough given there are a lot of classic papers in this area such as https://arxiv.org/pdf/1806.08734.pdf.
We respectfully disagree with the reviewer's opinion that "the novelty and significance are not... | Rebuttal 1:
Rebuttal: We thank the reviewers for the detailed feedback and positive evaluations. We address general concerns of the reviewers below.
## Networks with more depth. [Reviewers ynT5, zebK, mL5m ]
A first thought to tackle this problem is to simply extend our result by initializing the first layer as an ... | NeurIPS_2023_submissions_huggingface | 2,023 | Summary: The paper studies a fairly simple model of deep learning, with a view to understanding the effect of initialisation scale on the spectral structure of the trained network. The model is a two-layer network with out nonlinearity, trained via gradient flow and square loss.
The authors are able to make explicit s... | Rebuttal 1:
Rebuttal: We thank the reviewer for the valuable feedback and interesting questions which help us in clarifying and improving the paper.
> What's the intuitive reason for using orthogonal init?
The orthogonal initialization captures the high-dimensional asymptotics of initialization with i.i.d. Gaussian ... | null | null | null | null | null | null |
C-Disentanglement: Discovering Causally-Independent Generative Factors under an Inductive Bias of Confounder | Accept (poster) | Summary: The paper proposed Confounded-Disentanglement a method that makes use of inductive bias of a confounder by leveraging labels/knowledge from domain expertise. The aim is to recover the true/causal generative factors for the observed data which can recover a more meaningful latent representation that can be furt... | Rebuttal 1:
Rebuttal:
We thank Reviewer S7Fk for the effort of reviewing our paper. We are encouraged to see that the reviewer appreciates the quality and the significance of our work.
Most of the concerns and questions we believe are due to the confusions from various definitions of disentanglement in existing wor... | Summary: This paper is motivated by the assumption that a few semantically significant generative factors generate real-world data. Then they study causal independence in generative methods by introducing a new framework, C-Disentanglement, that discusses the identifiability issue of generative factors regarding the in... | Rebuttal 1:
Rebuttal: We want to thank Reviewer j8n4 for the effort of reviewing our paper and the recognition of the novelty and significance of the proposed method. We hope that our response and additional results could clarify doubts toward our work.
---
> [Q1] Notation issue
We would like to thank reviewer j8n4... | Summary: This paper delves into the domain of representation learning, focusing on the discovery of meaningful generative factors within a latent space. The authors propose the C-Disentanglement framework, which incorporates the inductive bias of confounders through labels or knowledge obtained from domain expertise. T... | Rebuttal 1:
Rebuttal:
We greatly appreciate Reviewer TdLu for the valuable feedback and suggestions. We are encouraged that Reviewer TdLu finds the problem setting is fundamental and important, and our paper in good clarity and quality. We address the concerns and discuss the insightful suggestions raised by the revie... | Summary: The paper proposes a new framework called Confounded-Disentanglement (C-Disentanglement) to learn disentangled representations by using inductive bias of confounding. Instead of the common statistically independent assumption for latent distribution which is usually not true in real-world datasets, the propose... | Rebuttal 1:
Rebuttal: We greatly appreciate Reviewer Mytf for the insightful feedback and suggestions. We are encouraged that Reviewer Mytf finds our work solid and promising. We address the concerns and questions raised by the reviewer below. Should Reviewer Mytf have any other questions or concerns, please do not hes... | Rebuttal 1:
Rebuttal:
### Summary of Rebuttal Discussion
We thank all reviewers for the valuable feedback and insightful questions. We are particularly encouraged that they consider our proposed framework and method fundamental, important and novel.
---
### Review Highlights
Reviewer Mytf finds our work "**techn... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Squared Neural Families: A New Class of Tractable Density Models | Accept (spotlight) | Summary: The paper introduces SNEFY, a tractable probabilistic model obtained by squaring the Euclidean norm of a neural net and normalizing it w.r.t. a base measure $\mu$ (cf. Equation 1).
The paper is mainly focused on studying settings in which one can obtain a closed form tractable solution for the normalizing cons... | Rebuttal 1:
Rebuttal: Our thanks for your thoughtful review and suggestions.
**Weaknesses:**
- " *Sampling seems to be possible only by rejection sampling, and is therefore very computationally expensive. Could you elaborate more about sampling?* "\
There are some interesting approaches (beyond rejection sampling) and... | Summary: The authors introduce a class of normalized probability densities that are obtained by squaring a single-layer neural network. The class of densities, dubbed Squared Neural Family (SNEFY), offers a number of attractive properties including: i) they can be understood as a generalization of exponential family di... | Rebuttal 1:
Rebuttal: Thank you for your in-depth review. We were very happy to read that you found that our model is novel, addresses a problem that is of wide interest to the community, admits attractive properties, and is well-placed in the context of related methods. We agree entirely with the limitation of samplin... | Summary: This paper introduces a parametric family of densities which admit tractable normalizing constants when a related neural-network kernel can be efficiently computed (e.g., in closed form).
The model is parameterized by the weights of a one-hidden-layer network, is closed under conditioning and admits tractable ... | Rebuttal 1:
Rebuttal: Thank you for your kind review and useful feedback. In addition to the general rebuttal, we address your individual concerns below. Please let us know if you would like to discuss further.
**Strengths:**
- You are right to highlight a strength of our model is that it scales linearly in the number... | Summary: The authors present a new class of neural network based density models. This new family, SNEFYs, is based on a Gaussian Process paradigm (kernel methods) and contrasts with other competitors like flow based density estimators by imposing less restrictions on the network dimensionality. SNEFYs are shown to be a... | Rebuttal 1:
Rebuttal: Thank you for your positive review. We are glad that you believe this area of interest (new classes of density estimators) will very likely spur a lot of related activity. We agree that this area is relatively under-explored, and are happy to hear that you found the paper well-written. We address ... | Rebuttal 1:
Rebuttal: **General response to reviewers:**
We thank the reviewers for their careful, detailed feedback and overall positive evaluation. We address commonly raised concerns here and respond to individual queries separately. We look forward to further engaging with the reviewers during the reviwer-author d... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Anytime-Competitive Reinforcement Learning with Policy Prior | Accept (poster) | Summary: This paper addresses the problem of anytime constrained RL. Contrary to traditional constrained RL where the constraints are expressed as expected costs, this work looks at a stronger criteria that the cost should be bounded for any episode. The cost is defined relative to the performance of an existing safe p... | Rebuttal 1:
Rebuttal: We appreciate your comments and are happy to address them as below.
`Experiments.`\
Thank you for your suggestions, we will move some important experiment results to the main body of the paper.
Since the motivation is from mission-critical applications, we will add another experiment on real-tim... | Summary: Existing work on Constrained Markov Decision Processes (CMDPs) typically formulate a problem where the expected cumulative reward
is maximized under a constraint where the expected cost over random dynamics is less than some threshold. Such existing work has a drawback that the safety constraint is violated i... | Rebuttal 1:
Rebuttal: We appreciate your comments and are happy to address them as below.
`The differences between Ref-1 and this paper.` \
Ref-1 solves the MDP with almost surely safety constraint by introducing Sauté MDP where constraints are incorporated into the reshaped objective. The approach is verified to have... | Summary: The paper studies a novel problem of Anytime-Constrained Markov Decision Process (A-CMDP), where the goal is to optimize the expected reward while guaranteeing that the cumulative cost in each round of any episode does not exceed a relaxed version of the cost of a policy prior. The paper proposes a projection-... | Rebuttal 1:
Rebuttal: We appreciate the comments of the reviewer and address them as below.
`The reason that the safe policy prior can be queried on any state`\
The safe policy prior is a mapping from a state to an action. In the ACD policy, policy prior is queried on true states to get a virtual action (which does no... | Summary: This paper proposed a novel problem called Anytime-constrained Markov Decision Process. Unlike CMDP, at each step, a constraint should be satisfied. A projection-based algorithm is proposed to solve the Anytime-constrained MDP. The authors gave a perturbation-based regret analysis for the algorithm.
Strengths... | Rebuttal 1:
Rebuttal: We appreciate the comments of the reviewer and address them as below.
`Are the assumptions too strict?`\
As discussed in limitations, the any-time constraint is significantly stricter than the expected constraints or the constraints with high probability, so it cannot be guaranteed without furthe... | null | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Directed Cyclic Graph for Causal Discovery from Multivariate Functional Data | Accept (poster) | Summary: The paper develops a causal discovery method for directed cyclic graph. The proposed method is a two-step approach which utilizes a Bayesian approach to reduce the dimension of functional data and performs the causal structure learning on the learned embeddings. Experiments on the synthetic and real-world data... | Rebuttal 1:
Rebuttal: **[Q1]** Following your suggestions, we have now compared the proposed method FENCE with DAG-GNN and NOTEARS (we could not find publicly available code for CSIvA yet). Both DAG-GNN and NOTEARS were applied to the functional principle components. The results are reported in the table below. We hav... | Summary: This paper discusses the inference of causal relationships between multivariate functions through the proposal of a causal discovery model. The model involves embedding the functional nodes into a lower-dimensional space and separating the structural equation model into two components: the projection onto the ... | Rebuttal 1:
Rebuttal: **Response to your question**:
In Hayden et al. (2006), it was observed that alcohol-dependent subjects exhibited frontal asymmetry, distinguishing them from the control group. Our own investigation aligns well with these results, as we have identified denser connectivity across various brain reg... | Summary: The authors propose a causal model, i.e. a linear structural equation model for multivariate functional data. The authors' proposed model does not require the dependency structure to conform to a DAG, allowing for modeling cyclic cause-effect relationships. Given certain assumptions on exogenous variables and ... | Rebuttal 1:
Rebuttal: **Response to your questions**:
**[Q1]** Following your suggestion, we have carried out misspecification analysis for our model by considering nonlinearly generated data. We have considered a scenario with number of samples ($n$) $= 100$, number of nodes ($p$) $= 6$ and evenly spaced time grid ov... | Summary: In this paper, the author proposes an operator-based non-recursive linear structural equation based novel causal discovery framework that identifies causal relationships among functional objects in the presence of cycles and additional sampling noises. Furthermore, author demonstrates the effectiveness of prop... | Rebuttal 1:
Rebuttal: **Response to your questions**:
**[Q1]** We obtained the fLiNG's code from the authors of that paper. Our research team has also reached out to Yang (2022) but we were unable to obtain their code. The paper by Lee (2022) mainly focused on the statistical theories and hence the code is not availab... | Rebuttal 1:
Rebuttal: We would like to thank all reviewers for their constructive comments.
This is a global response to several points mentioned by different reviewers while reviewing our work.
1. **Assumptions of the model**: Some of the reviewers pointed out limitations regarding some of the assumptions that we ha... | NeurIPS_2023_submissions_huggingface | 2,023 | Summary: The paper considers the problem of learning causal structure from multivariate functional data that may involve cyclic interactions.
It employs an adaptive mapping to a lower dimensional space that retains the relevant causal information, in combination with a Bayesian framework that obtains posterior estimate... | Rebuttal 1:
Rebuttal: **Response to your questions**:
**[Q1]** Bongers introduces simple structural causal models (SCMs), which allows certain cyclic settings. An SCM is simple if any subset of its structural equations can be solved uniquely for its associated variables in terms of the other variables that appear in t... | null | null | null | null | null | null |
Cognitive Steering in Deep Neural Networks via Long-Range Modulatory Feedback Connections | Accept (poster) | Summary: This paper explores adding long-range modulatory feedback connections to deep CNNs (specifically AlexNet, evaluated on ImageNet). It explores 2 ways of incorporating the feedback: a default mode and a cognitive steering mode. The results show improvements on ImageNet accuracy, adversarial robustness, and a com... | Rebuttal 1:
Rebuttal: # Strengths:
>*The paper is high-quality: it's well-written, clear, the visualizations seem to have been very thoughtfully prepared, etc. … I think this is a very significant question (how to incorporate long-range feedback into deep neural networks), which has been studied for many years but hasn... | Summary: The paper presents a novel, brain-inspired modulatory feedback circuitry (long-range modulation, LRM) for regular feedforward DNNs. The multiplicative modulatory pathways can be conditioned on a) higher-level activations (computed in the initial forward pass or subsequent modulatory passes, Sec 2.1, 2.2) to im... | Rebuttal 1:
Rebuttal: # Strengths:
> *[Originality] The paper is sufficiently novel in my opinion, with key architectural features well motivated by experimental psychology & neuroscience evidence and reasonably different from existing RNN & predictive coding based architectures.*
Thanks!
# Weaknesses:
> *How exactly... | Summary: This work introduces a recurrent modulation module that can be added to visual models so that the top layer can project to and influence the earlier layers. The authors find that the model with the feedback projection layers outperforms the baseline feedforward model in both the categorization performance and ... | Rebuttal 1:
Rebuttal: # Strengths:
> *The paper is well-written and easy to read. The significant improvement of the feedback-augmented model compared to the baseline model is also interesting. The steering analysis is done on both side-to-side and overlaying settings.*
Thanks!
# Weaknesses:
> *The idea of adding fee... | Summary: Authors study ways of incorporating cognitive steering in vision neural network models. They add a top-down feedback mechanism to Alexnet with which they report improved performance. Further, they test other steering mechanisms, including prototypes, language-based signals etc. These tests are over image compo... | Rebuttal 1:
Rebuttal: # Strengths:
> *The paper is interesting, the experiments and the controls are convincing. Cognitive steering in deep CNNs is novel as far as I know, especially with language signals. Some parts of the paper are well written.*
Thanks!
# Weaknesses:
> *The paper lacks benchmarking. There are seve... | Rebuttal 1:
Rebuttal: # Summary of Strengths:
There was near consensus about the **novelty** of introducing cognitive steering capacities into deep neural network models through long-range modulatory feedback connections (R1,R3,R4; though see R2), with experimental results that are **convincing**, **interesting**, an... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Variational Monte Carlo on a Budget — Fine-tuning pre-trained Neural Wavefunctions | Accept (poster) | Summary: The authors propose a pre-trained deep learning wavefunction ansatz for variational Monte Carlo (VMC) solutions of the Schrödinger equation. The ansatz is trained on a chemically diverse set of (different conformations of) molecules. The zero-shot accuracy of the proposed method is a significant improvement ov... | Rebuttal 1:
Rebuttal: Thank you for your review! Regarding your specific comments:
### Weaknesses
> While the paper compares results with previous methods, these comparisons consider only metrics like the number of optimization steps, but not the raw computational cost in terms of CPU/GPU time, which is arguably more... | Summary: This paper concentrates on applying deep learning techniques to electronic structure computations. Unlike the strategy adopted in [1], the authors introduce a pretraining process to enhance the few-shot or zero-shot performance of Deep Learning Variational Monte Carlo (DL-VMC) methods. As their initial step, t... | Rebuttal 1:
Rebuttal: Thank you for your feedback! Regarding your specific questions and potential weaknesses:
### Weaknesses
**W1**
We ran additional experiments, increasing the nr of fine-tuning steps up to 32k, and depict the result in Fig. 2 of the rebuttal.
We find that with increasing nr of fine-tuning steps... | Summary: The paper proposes a meta-learning model for neural wavefunctions. The neural model is first pre-trained by solving the variational optimization problem on many molecules. To adapt the model for a new problem (such as a new molecule), relatively few fine-tuning steps are needed to arrive at an accurate solutio... | Rebuttal 1:
Rebuttal: We thank the reviewer for the review and feedback regarding our submission.
### Weaknesses
> Relative energies, which depend on small energy differences, are not very accurate without fine-tuning on that specific system. The transferability is not as good for relative energies.
Predicting acc... | Summary: This paper tackles the problem of solving quantum Monte Carlo optimization on multiple molecules. A PhisNet model is used to encode geometric information from nuclei and other GNN models are then used to embed electrons. Additionally, the authors propose a method to generate more diverse nuclear geometries by ... | Rebuttal 1:
Rebuttal: We thank the reviewer for the feedback and the effort to improve our work. We hope we can address in the following rebuttal all raised questions and comments.
### Weaknesses
> The improvements of the proposed method mainly comes from an improved architecture. To my understanding, the novelty se... | Rebuttal 1:
Rebuttal: We thank all reviewers for their constructive feedback and appreciate the overall positive reception of our work (“excellent accuracy on challenging systems”, “extensive Empirical Study”), as well as its presentation (“experiments and figures are high quality”, “clear and well-written”).
Several ... | NeurIPS_2023_submissions_huggingface | 2,023 | Summary: The paper proposes a neural network wave function model which generalizes to different molecules and geometries. The neural network includes electron model, orbital model, both with message passing, Slater determinants, and an energy estimate for geometry sampling. The network is pretrained on 98 molecules and... | Rebuttal 1:
Rebuttal: Thank you very much for your review! Regarding your specific questions and potential weaknesses:
### Weaknesses
>I'm not so familiar with CC-2Z, -3Z etc methods the paper compares against. Not sure how the
proposed method compares with something like Slater-Jastrow variational Monte Carlo (whic... | null | null | null | null | null | null |
A Privacy-Friendly Approach to Data Valuation | Accept (spotlight) | Summary: This paper proposes an approach for data valuation that tries to be privacy-preserving by design. More precisely, the proposed method aims at computing the data valuation of a point through its Shapley value for the k-nearest neighbours classifier (knn) in a differentially-private manner.
Strengths: The pape... | Rebuttal 1:
Rebuttal: We thank the reviewer for the very positive feedback!
**Q1. [Limitation of KNN-Shapley?]** *“The paper currently lacks a discussion on the limitations of the knn Shapley values compared to other methods for computing Shapley values in a more generic context. For instance, is there situations in w... | Summary: The authors propose a relaxation of the KNN based Shapley valuation framework based on thresholding. By doing so, they are able to reduce the sensitivity associated with shapley valuation, and consequently apply DP to the mechanism.
Strengths: 1. First formalization of TKNN based shapley valuation.
Weaknesse... | Rebuttal 1:
Rebuttal: We thank the reviewer for the positive feedback!
**Q1. [Primary contribution]**
**A:** Our contributions in this work are three-fold:
1. **Uncovering the privacy risks of data valuation technique:** To the best of our knowledge, this work is the first formal study on the privacy risks associat... | Summary: In this paper, the authors suggest a modified version of KNN-Shapley that aims to enhance the privacy of the data valuation procedure. They introduce TKNN-Shapley, which focuses only on neighbors within a $\tau$ vicinity of the utilized data point instead of a fixed number of neighbors (k), and DP-TKNN-Shapley... | Rebuttal 1:
Rebuttal: We thank the reviewer for the positive feedback about our experiment!
**Q1. [Theoretical Contributions in this work]** *“The method ... is limited.”*
**A:** We thank the reviewer for the comment. There are in fact at least two highly non-trivial theoretical contributions (in our opinion) in this... | Summary: The paper studies the privacy concerns surrounding KNN-Shapley, a prevalent method for data valuation. Authors introduce TKNN-Shapley, which employs the concept of differential privacy to protect privacy. Empirical tests demonstrate that TKNN-Shapley performs on par with KNN-Shapley in assessing data quality, ... | Rebuttal 1:
Rebuttal: We thank the reviewer for the positive feedback about our presentation and methodology!
**Q1. [Privacy risk of KNN-Shapley]** *“The privacy risk of KNN-Shapley ... can lead to a membership inference attack (MIA). Additionally, KNN is not easily ... less likely. Furthermore, in Figure 1, the assum... | Rebuttal 1:
Rebuttal: We thank all of the reviewers for their detailed and valuable comments. We are glad that our work receives a majority of positive reviews. We considered the reviews carefully and modified our paper accordingly. We have answered other questions in the individual responses. Here’s our global respons... | NeurIPS_2023_submissions_huggingface | 2,023 | Summary: The paper addresses the problem of evaluating and assigning a value to the quality of training data for KNN in a privacy preserving way. Therefore, it first introduces an adaptation of the KNN-Shapley algorithm, namely TKNN-Shapley which due to its non-recursive way of implementation facilitates integrating di... | Rebuttal 1:
Rebuttal: We thank the reviewer for the positive comments!
**Q1 [Selection of the threshold in TKNN]**
**A:** In our experiment, we found that the performance of TKNN-Shapley is quite robust to the selection of the threshold $\tau$. In Section 6, we set the threshold $\tau = -0.5$ (note that cosine distan... | Summary: The paper proposes a new variant of KNN-Shapley named TKNN-Shapley from which they derive a privacy-preserving data value approach DP-TKNN-Shapley. They conduct experiments to showcase that their approaches obtain better privacy-utility tradeoff and comparable performance to baselines.
Strengths: - Data valu... | Rebuttal 1:
Rebuttal: We thank the reviewer for the positive comments!
**Q1. [Improving the motivation section]** *“I believe the authors need to make the following changes: 1) provide a simple version of the membership inference attack (MIA) algorithm in the main paper (Section 3) and present basic attack results (Ta... | null | null | null | null |
Debiased and Denoised Entity Recognition from Distant Supervision | Accept (poster) | Summary: The paper presents DesERT, a novel self-training framework for distant-supervised named-entity recognition (NER). It addresses the challenges of noisy distant labels and introduces two types of biases that impact performance. DesERT incorporates a dual-form self-training approach and a debiased module to impro... | Rebuttal 1:
Rebuttal: Thanks very much for your valuable comments and suggestions! Our detailed responses are as follows.
**Q1. The method proposed in 4.3 is a direct application of [Chen et al., 2022], and the dual co-guessing framework is also a direct extension in [Liang et al., 2020] by adding two model branches. ... | Summary: The paper proposes a novel self-training framework, DesERT, for distantly-supervised named entity recognition (NER) that addresses two types of biases omitted by prior work: structural noise in distant labels and the inherent bias introduced by the self-training framework. The paper shows that the proposed fra... | Rebuttal 1:
Rebuttal: Thanks very much for your insightful comments and suggestions! Our detailed responses are as follows.
**Q1. The framework seems a simple mixture of existing techniques and their application on the NER task. The two observed omitted bias of previous works resembles the data bias the training bias ... | Summary: The paper studies the distantly-supervised named entity recognition (NER) problem and proposes a framework called DesERT. Motivated by the bias patterns in the distant labeling noise, DesERT conducts clean token selection to split the training set into a clean set and an unlabeled set and then trains the encod... | Rebuttal 1:
Rebuttal: Thanks very much for your comments and suggestions. Below we address the feedback and comments in detail:
**Q1. Problem scope and method design.**
A. We originally tend to focus on the setting of distantly-supervised NER, which is a practical and fundamental task in NLP and benefits many downstr... | Summary: This paper presents a new method named DesERT for distantly supervised named entity recognition (NER), which starts with analyzing the noise of and biases in distant supervision which is used in self-training frameworks: 1) skewed data distribution, 2) inherent error patterns in pseudo-labels. They decompose t... | Rebuttal 1:
Rebuttal: Thanks very much for your insightful comments and suggestions! Our detailed responses are as follows.
**Q1. The presentation can be further improved with more illustrative examples and figures.**
A. Thanks for your suggestions! We provided our visualized illustration of Dataset Taxonomy in Appen... | Rebuttal 1:
Rebuttal: We sincerely appreciate all reviewers for their great efforts in providing constructive and valuable feedback. We are pleased that the reviewers find our work **well-written** (**R1,R3**), **theoretically solid** (**R1,R2**), and **experimentally comprehensive** (**R1,R2,R4,R5**). Moreover, we are... | NeurIPS_2023_submissions_huggingface | 2,023 | Summary: The paper introduced DesERT, a new method that improves the distant supervision training framework for DSNER (distant supervision NER) using two highly effective components: 1. To solve the skewed label distribution (entity vs. non-entity tokens) in DSNER, DesERT proposes to decompose learning into two parts: ... | Rebuttal 1:
Rebuttal: Thank you very much for your comments and suggestions! We are happy you enjoyed the paper. Below are our detailed responses:
**Q1. The results come without a sense of variance. Are the performance improvements significant? What is the standard deviation of your approach compared to the baselines?... | null | null | null | null | null | null |
Toward Better PAC-Bayes Bounds for Uniformly Stable Algorithms | Accept (poster) | Summary: This paper gave an improved PAC-Bayes bound for uniformly stable randomized algorithms. It introduced new assumptions and proof techniques. The paper also applied the general theorem to concrete applications (SGD and RCD) and got better or novel PAC-Bayes bound for these two algorithms.
Strengths: 1. The gen... | Rebuttal 1:
Rebuttal: Thank you for your constructive comments and suggestions.
**Q1**. The prior and posterior are over the hyperparameters but not on the hypothesis class.
**A:** Thanks for the suggestion. We build on the idea of [29] to consider the prior and posterior on the hyperparameters instead of that ... | Summary: The work aims to provide stronger generalization guarantees for stochastic algorithms through the analysis of stability and PAC-Bayes. Such algorithms include SGD and RCD. The work uses a milder assumption in the analysis under the same problem setting as [29]. The results can be extended beyond strong convexi... | Rebuttal 1:
Rebuttal: Thank you for your constructive comments and suggestions.
**Q1**. It is unclear why you need PAC-Bayes analysis to guarantee all distributions $\mathbb{Q}$. Isn't it enough to consider a fixed distribution like in [15]?
**A:** Thank you for the intuitive question. Since the PAC-Bayes bound ... | Summary: In this paper, the authors study PAC-Bayesian bounds on the generalization error of uniformly stable algorithms. They consider algorithms in which the randomness comes from a particular hyperparameter $\theta$, e.g., stochastic gradient descent (SGD), in which the randomness comes from selecting the indices us... | Rebuttal 1:
Rebuttal: Thank you for your constructive comments and suggestions.
**Q1**. It would be better to clarify why would one introduce this seemingly artificial distribution over the hyper-parameters.
**A:** Thanks for the suggestion. We will add more discussions on the distribution of the hyper-paramete... | Summary: In this paper, the authors extend the work of London [29] and prove PAC-Bayesian generalization bounds for uniformly stable algorithms. Compared to [29], the present work removes some required assumptions in the work of [29], such as the hyperparameter stability. To do so, they required a new assumption called... | Rebuttal 1:
Rebuttal: Thank you for your constructive comments and suggestions.
**Q1**. L173: In your Definition 1 and one in [29], the expectation over $\mathbb{p}$ is not in the same place.
**A:** Thanks for the careful reading and indicating this. $\mathbb{E}[\beta_\theta]$ in our submission is different from $\be... | null | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Decision Tree for Locally Private Estimation with Public Data | Accept (poster) | Summary: In this paper, the authors propose a scheme for training decision trees on a combination of public and private data. By using local differential privacy the algorithm guarantees that adversaries learn little about the private data from the outcomes of the algorithm. Under assumptions and a specific splitting r... | Rebuttal 1:
Rebuttal:
Q1:
Assumption 3.2 is key to deriving error bounds but I am unsure if this holds. Does assumption 3.2 hold when considering decision trees that make discontinuous partitions?
A1:
The reviewer is correct.
We apologize for the misleading statement of Assumption 3.2.
Assumption 3.2 is posed on t... | Summary: The paper studies nonparametric regression under local differential privacy (LDP) constraints with the aid of public data. The paper proposes an algorithm, locally differentially private decision tree (LPDT), which uses public data to construct the splitting criteria for the decision tree and private data to c... | Rebuttal 1:
Rebuttal:
Q1: The importance of public data has been established ... but less surprising.
A1:
We agree with the reviewer that the idea of using public data is less surprising.
However, we would like to emphasize our motivation for utilizing public data.
Most public data usage, such as gradient precond... | Summary: This work studies non-parametric estimation under local differential privacy with public data. The authors propose a locally private decision tree algorithm and show that it is min-max optimal under some regimes. The authors also test the algorithm on real and synthetic datasets.
Strengths: 1. The idea to lev... | Rebuttal 1:
Rebuttal: Q1: The proposed algorithm does not seem to improve too much ... the same sample complexity.
A1: Please see the R2 of the global response for the theoretical improvements.
Q2: Given the first point, the fact that the proposed algorithm does not work without public data appears to be a significa... | Summary: This paper proposes a locally differentially private (LDP) decision-tree (DT) regressor algorithm that takes advantage of public data for utility improvement.
The proposed algorithm, LPDT, uses Randomized Response (RR) and the Laplace mechanism to protect the tree partition step.
This paper also introduces a n... | Rebuttal 1:
Rebuttal: Q1: The paper would greatly benefit from including high privacy regimes $\epsilon$ <=1 and the utility gain of using public data.
A1: Please see R1 of the global response for small $\varepsilon$ and Experiment 4 for the utility gain.
Q2: In lines 73 \& 74 ... How worse would that be in high pri... | Rebuttal 1:
Rebuttal: We express our gratitude to the reviewers for their insightful questions and valuable comments. Reviewers are encouraged to refer to their individual rebuttals, where we direct them to the relevant sections in the text provided below.
R1: We did not incorporate the results of $\varepsilon < 1$ d... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
GRAND-SLAMIN’ Interpretable Additive Modeling with Structural Constraints | Accept (poster) | Summary: This paper introduces hierarchical structure constraints to improve the sparsity of selected features in interaction-aware GAM. The authors use soft relaxations to enable gradient descent during this procedure. Theoretical analyses are given to show statistical properties when using soft tree shape functions i... | Rebuttal 1:
Rebuttal: We thank reviewer for his comments and the time spent in reviewing the paper.
1) **Response to Limited Contribution**
a) **Discussion on GA^2M** We want to clarify that our approach is very different from GA^2M with tree-shaped functions proposed by Lou et al. (2013) and later packaged as EBM. ... | Summary: The present paper presents a method to incprorate pairwise interaction effects into differentiable generalized additive models while retaining scalability.
The authors pose the fitting of the GAM with pairwise interactions as a combinatorial optimization problem with indicator variables and hierarchy constra... | Rebuttal 1:
Rebuttal: We thank reviewer for his comments and the time spent in reviewing the paper.
1) **Multiple runs across seeds** Tables F.1 and F.2 in the Supplement correspond to 10 runs/seeds for the optimal trials reported in Table 2 and Table 3. We will use F.1 and F.2 in the main text in the camera-ready ver... | Summary: The authors introduce GRAND-SLAMIN’, a framework for end-to-end learning of sparse Generalized Additive Models (GAMs). GRAND-SLAMIN’ utilizes soft neural decision trees as base learners and incorporates support for second-order interactions and structural constraints. Sparse variable selection is achieved via ... | Rebuttal 1:
Rebuttal: We thank reviewer for his comments and the time spent in reviewing the paper.
1) **Visualization** We provide a new visualization study to further highlight an important contribution of our paper.
In particular, our framework can support models with structural constraints. Hence, we study the... | Summary: The authors proposed an GAM under structure constraints. By incorporating herarical constraint, the proposed method presents better performance than two stage methods. Theoritical analysis are also provided to bound the generalization error.
Strengths: A GAM method was proposed, which can be learnt in an end... | Rebuttal 1:
Rebuttal: We thank reviewer for his comments and the time spent in reviewing the paper.
1) **Sparsity** The goal of our method is to recover a subset of $p^2$ effects. We achieve this goal with two ideas.
For small to medium sized datasets, we recover an active subset via an end-to-end learning optimizati... | Rebuttal 1:
Rebuttal: We thank all reviewers for their comments and the time spent in reviewing the paper.
Please see our attached figure rebuttal document.
Pdf: /pdf/67ac09c8914b0e20d05bba28f9bde57d51a67a01.pdf | NeurIPS_2023_submissions_huggingface | 2,023 | Summary: This paper deals with interpretable machine learning. Authors propose a new interpretable model based on generalized additive modeling. Namely they propose a scalable solution for sparse interaction modeling. They compare against related work on interpretable benchmarks.
Strengths: + Novel contributions: nove... | Rebuttal 1:
Rebuttal: We thank reviewer for his comments and the time spent in reviewing the paper.
1) **Choice of datasets** A central theme of our work is a sparse selection of main/interaction effects. Our selection of datasets was motivated to consider tabular classification datasets which have a sufficient number... | null | null | null | null | null | null |
Invariant Anomaly Detection under Distribution Shifts: A Causal Perspective | Accept (poster) | Summary: This paper tackles the problem of anomaly detection under a distribution shift. The authors describe the setting in detail, where the training is drawn from multiple subsets of examples, and test data is drawn from different subsets. The subset of each training example is known. The goal is to detect anomalies... | Rebuttal 1:
Rebuttal: We thank the reviewer for their time and thorough review of our work, as well as the comments and proposed suggestions.
**1. (i) Yet in fact, a similar setting was presented in Red PANDA when treating a nuisance feature as a sub-domain.**
We agree with the reviewer that nuisance features when ac... | Summary: The paper focuses on the making anomaly detection models robust to various kinds of distribution shifts over normal/benign samples. This goal is achieved by formalizing the anomaly detection problem from the information theory perspective and introducing the requirements it brings for developing AD models that... | Rebuttal 1:
Rebuttal: We thank the reviewer for the positive evaluation, constructive feedback, and thorough review of our work. We will address the pointed weaknesses jointly with raised questions.
**I had hard time to understand the purpose of Figure 2.**
The goal of Figure 2, was to help visualise the formulation g... | Summary: This paper aims to promote AD through the lens of causality. To this end, the authors propose to minimize the distribution discrepancy of the leaned representations among environments. Some experiments are conducted to verify the effectiveness of the proposed method.
Strengths: 1 Exploring AD through a causal... | Rebuttal 1:
Rebuttal: We thank the reviewer for their comments and suggestions to improve this work.
**1. I suggest the authors to give a corresponding figure using features learned with the proposed method.**
We agree with the reviewer's suggestion and have produced a TSNE plot of representations produced with diffe... | Summary: The paper tackles the problem of anomaly detection under distribution shifts from the perspective of causal inference. Causal inference aims to induce invariance across environments. The authors derive a regularization term for anomaly detection to achieve this goal. The resulting detector is robust to domain ... | Rebuttal 1:
Rebuttal: We thank the reviewer for their time and constructive feedback.
**1. (A) There is a lack of justification of using the proposed solution.**
We agree with the reviewer that our work would benefit from a more comprehensive background of other causal methods.
We verified that the state-of-the-art ... | Rebuttal 1:
Rebuttal: We thank all reviewers for their time and constructive feedback when reviewing this work.
In this global response, we will provide details on the additional experiments performed and address the general concern of partial conditioning on only W=0.
**1. Experiments on Waterbird**
To address the... | NeurIPS_2023_submissions_huggingface | 2,023 | Summary: The paper proposes a formalism for detecting anomalies in the presence of domain and covariate shifts. Data X is assumed of being decomposable into X_a and X_e where the latent factors X_a determine if X is an anomaly while X_e represents style features. To obtain invariant features, a regulariser is proposed.... | Rebuttal 1:
Rebuttal: We thank the reviewer for the positive evaluation of the paper, and their comments and suggestions for improvement.
**1. A similar framework for anomaly detection in the presence of style changes is proposed in [Smeu et al. 2022].**
We thank the reviewer for identifying this framework. Although ... | null | null | null | null | null | null |
GlucoSynth: Generating Differentially-Private Synthetic Glucose Traces | Accept (poster) | Summary: Instead of the existing DP GAN framework, which cannot capture the relationship in time-series data well when temporal length is not the only relationship but also different event types and do not perform well when more strict DP (differential privacy) condition is required, the paper proposes new architecture... | Rebuttal 1:
Rebuttal: We thank the reviewer for their time and review. Responding to Weakness 2 about DP: This is a good point. The privacy is at the trace level, which we claimed was user-level because we assume that each patient has only one trace in the dataset. We will add a clarifying statement about this assumpti... | Summary: This paper proposes a DP generation model, i.e. GlucoSynth, particularly for glucose data. To address a few challenges that prior works fail in, GlucoSynth proposes a causality learning block, to learn the dependency between motifs, rather than merely learning the temporal dependencies. The generator is basica... | Rebuttal 1:
Rebuttal: We thank the reviewer for their time and review. Responding to Weakness 2: The reviewer is correct that our model is time and memory-intensive. Even so, since you only have to train the model once and then can use the produced synthetic data over and over, we believe this is an okay trade-off. We ... | Summary: The authors introduce a new differential privacy scheme where other, non-related signal recordings are used to anonymise the signal provided that they have a motif that seems to predict the time signal forwards in time. For this a time dependent vector is used to contain multiple signals at the same time prop... | Rebuttal 1:
Rebuttal: We thank the reviewer for their time and review. Responding to the Weaknesses: Thanks for catching these corrections. We will correct the supplemental material submission and adjust the language in the abstract to read: "Previous methods for time series data synthesis ... " We have also responded ... | Summary: This manuscript presents GlucoSynth to generate synthetic glucose traces. The main idea is to conserve relationships among motifs within the time series and to preserve the privacy of the glucose traces. Specifically, the manuscript proposes a motif causality block to learn the causal relationships amongst the... | Rebuttal 1:
Rebuttal: We thank the reviewer for their time and review. Responding to Weakness 1: The main point of Figure 2 was to show that glucose motifs are not temporally-dependent, and provide insight about why methods that only focus on conserving temporal relationships fail. The motivation of the work was to dev... | null | NeurIPS_2023_submissions_huggingface | 2,023 | Summary: In this paper, the problem of generating differentially private synthetic glucose traces using a GAN framework, by conserving the relationships among glucose events or motifs within the traces, as well as the temporal dynamics.
Strengths: This paper presents GlucoSynth, an approach that generates private and ... | Rebuttal 1:
Rebuttal: We thank the reviewer for their time and review. For our privacy integration, we use a straightforward application of prior differential privacy methods to our framework. Since we have not developed any new differential privacy mechanism, and are using established privacy techniques in a standard ... | null | null | null | null | null | null |
Kronecker-Factored Approximate Curvature for Modern Neural Network Architectures | Accept (spotlight) | Summary: This paper extends Kronecker-Factored Approximate Curvature (K-FAC) into training generic neural networks, especially for transformers and graph neural networks. This is done by taking advantage of the weight sharing mechanism. The authors propose two flavors of K-FAC (expand and reduce), depending on the dime... | Rebuttal 1:
Rebuttal: We thank the reviewer for their time and their comments.
>The performance of transformers is evaluated on ImageNet. Since transformers are mostly used for language modeling, it is better to evaluate your optimizers on some language tasks (a very big dataset is unnecessary).
We believe that trans... | Summary: This paper introduces two different settings of linear weight-sharing layers to fit for the second-order optimization method Kronecker-Factored Approximate Curvature (K-FAC). Experiments on GNNs and ViTs achieve less number of steps compared with the first-order baseline.
Strengths: Experiments include GNNs a... | Rebuttal 1:
Rebuttal: >How about the comparison with other efficient optimization methods other than weight sharing?
We believe there is a fundamental misunderstanding. **Linear weight-sharing is a property of the layers of the model architecture and not of the optimisation method.** In our paper, we show how we can a... | Summary: This paper propose two methods called KFAC-reduce and KFAC-expand to conduct the secon-order based training for neural networks. The key idea is to approximate the Gauss–Newton matrix with low-rank matrix decomposition. The experiments on several tasks show that the propsed method can achieves faster convergen... | Rebuttal 1:
Rebuttal: We thank the reviewer for their time and comments.
>The proposed method is only verified on limited number of tasks and networks (4 configurations). The author should show more results with various configurations.
>The method is only compared with two type of optimizers. For example, SGD is the ba... | Summary: The researchers propose a framework for applying the Kronecker-Factored Approximate Curvature (K-FAC) method to neural networks with linear weight-sharing layers. They identify two different settings of linear weight-sharing layers, leading to two variations of K-FAC: expand and reduce. They demonstrate that b... | Rebuttal 1:
Rebuttal: We thank the reviewer for their time and appreciate that they also found the concept of linear weight-sharing layers to describe modern neural network layers useful.
>Minor issues in presentations, such as the plots are (...) really hard to read, and organization of the propositions are too clust... | null | NeurIPS_2023_submissions_huggingface | 2,023 | Summary: The paper formalizes two variants of the K-FAC optimisation method for a family of neural network architectures that can be expressed as linear layers with weight sharing. Experiments are conducted to validate that the proposed method can be applied to different kinds of neural network architectured speeding-u... | Rebuttal 1:
Rebuttal: We thank the reviewer for their time and their feedback.
>I miss a better contextualitzation, both from a theoretical and experimental point of view, of the proposed method with other existing approaches, cited in the related work, that also make use of K-FAC and that have also been applied to si... | Summary: This paper considers linear layers with weight-sharing, which is used in many neural network architectures, such as transformers, convolutional and graph neural networks. Two variants of K-FAC, K-FAC –expand and K-FAC-reduce are considered for two different settings of linear weight sharing layers. The authors... | Rebuttal 1:
Rebuttal: We thank the reviewer for their time and the detailed comments.
>One of my concerns is the applicability of second-order methods, particularly for modern architectures such as the LLM model. As noted in the AlgoPerf paper [1], the use of KFAC requires detailed architectural information, which may... | null | null | null | null |
Spiking PointNet: Spiking Neural Networks for Point Clouds | Accept (poster) | Summary: This paper presents the Spiking PointNet, a novel solution for processing point cloud data. The authors propose a "train-less, learn-more" strategy, which leverages the Surrogate Gradient (SG) method to train the SNN over a single time step, while deploying multiple time steps during the inference phase to opt... | Rebuttal 1:
Rebuttal: We appreciate your time in reviewing our paper and your acknowledgment of our work on point cloud, our novel "train-less, learn-more" approach, and our clear writing complemented by understandable illustrations. Please find our point-by-point responses to your queries as follows.
**Q1**: Section ... | Summary: To investigate whether SNN can be suitable for 3D data, this paper proposes a Spiking PointNet for point cloud classification tasks. The authors mainly focus on two key issues: (i) Training SNNs with large time steps; (ii) Reducing the extensive memory and computation cost for the proposed spiking point model.... | Rebuttal 1:
Rebuttal: Thank you for taking the time to review our paper and for recognizing our contributions in point cloud research, our effective method, and clear writing. We will address your concerns and questions in the following response.
**Q1**: The two key issues (i.e., training large timesteps, and reducing... | Summary: The authors replaced the ReLU activation function of the original PointNet with the LIF neurons and named it spiking PointNet. The work offers trained-less but learning-more tricks and corresponding analysis. The effectiveness of spiking PointNet is evaluated on the ModelNet10/40 datasets, specifically in poin... | Rebuttal 1:
Rebuttal: Thank you for your efforts in reviewing our paper and your recognition of our effective method and well-written submission. The response to your questions is given piece by piece as follows.
---
**Q1**: If I understand correctly, the authors only replaced ReLu with LIF. If so, whether the first ... | Summary: In this paper, the authors introduce the spiking neural network into 3D data processing, the point cloud data. In specific, the authors analyze the training difficulty when using SNNs for the point cloud recognition and find that number of timesteps should not be increased in the training phase. Their framewor... | Rebuttal 1:
Rebuttal: Thanks for your efforts in reviewing our paper and your recognition of our interesting findings, efforts for point cloud, and the well-performance method. The response to your questions is given piece by piece as follows.
---
**Q1**: The authors claimed theoretical analysis in Section 3.4. Howe... | Rebuttal 1:
Rebuttal: We thank all reviewers’ time and effort in reviewing our paper. While we will respond to the comments piece by piece, we would like to first provide a general response to the common concern about the inference speed of spiking PointNet.
The advantage of SNNs is to reduce the power consumption, n... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
OpenShape: Scaling Up 3D Shape Representation Towards Open-World Understanding | Accept (poster) | Summary: The paper proposes a new system for learning a joint embedding space for text, images, and 3D shapes (point clouds). It starts from pre-trained CLIP embeddings for text and images, and learns one more embedding function for 3D shapes in a self-supervised manner, using a contrastive loss. Several additional imp... | Rebuttal 1:
Rebuttal: Thank you for your insightful comments and valuable suggestions. We will revise our paper based on your feedback. Here are our responses to your comments:
**Limited qualitatively surprising findings**
We kindly refer the reviewer to the supplementary for more exciting results, where we showcase ... | Summary: The paper studies the problem of multi-modal learning of text, image and shapes. Shapes are represented by pointclouds and the learning is driven by the standard contrastive loss. The main technical difference of this work to prior ones lies in the scale of data and corresponding training strategies. The autho... | Rebuttal 1:
Rebuttal: Thank you for your insightful comments and valuable suggestions. We will revise our paper based on your feedback. Here are our responses to your comments:
**Definition of "zero-shot"**
Thanks for pointing this out. We agree that the current "zero-shot" evaluation does not exclude all shapes in t... | Summary: In this paper, the authors propose a joint learning framework for multi-modal representations among text, image, and point clouds. Specifically, the authors fix the pretrained CLIP language and image encoder, and align the point cloud representation by the proposed multi-modal representation alignment techniqu... | Rebuttal 1:
Rebuttal: Thank you for your insightful comments and valuable suggestions. We will revise our paper based on your feedback. Here are our responses to your comments:
**Vision-language encoder**
Thanks for your suggestion. We will add the details of the point cloud backbone and pretrained vision-language en... | Summary: This work learns a multi-modal representation among text, images, and point clouds based on a scaling 3D dataset. The results demonstrate remarkable performance in point cloud zero-shot classification, retrieval, and captioning tasks. In order to construct this dataset, the authors combined four commonly used ... | Rebuttal 1:
Rebuttal: Thank you for your insightful comments and valuable suggestions. We will revise our paper based on your feedback. Here are our responses to your comments:
**The difference between the proposed method and ULIP**
While we follow the multi-modal contrastive learning framework for representation al... | Rebuttal 1:
Rebuttal: Dear Reviewers,
Thank you for dedicating your time to review our paper and for offering insightful feedback. We deeply appreciate your efforts to help enhance the quality of our research. We are also pleased to note that all reviewers were supportive of our work:
(a) Acknowledge our remarkable p... | NeurIPS_2023_submissions_huggingface | 2,023 | Summary: OpenShape explores scaling-up strategy for learning joint representations of texts, image, and 3D point clouds. It proposes to construct larger-scale 3D datasets, filter and enrich paired texts for pre-training, which achieves SOTA results on zero-shot 3D classification and retrieval benchmarks.
Strengths: 1.... | Rebuttal 1:
Rebuttal: Thank you for your insightful comments and valuable suggestions. We will revise our paper based on your feedback. Here are our responses to your comments:
**GPT-4 is expensive to access**
We also tried GPT-3 in our experiments but found it less effective at reasoning and finishing the filtering t... | null | null | null | null | null | null |
Finite-Time Logarithmic Bayes Regret Upper Bounds | Accept (poster) | Summary: This paper proves logarithmic (in horizon $n$) regret bounds for Bayesian bandits using the BayesUCB algorithm. They specifically study the settings of Gaussian MABs and Gaussian linear bandits. This improves on prior works which show the more standard $\sqrt{n}$ bounds on regret.
Strengths: - Generally the w... | Rebuttal 1:
Rebuttal: We would like to thank the reviewer for valuable feedback and detailed questions. We address your concerns below. If you have any additional concerns, please reach out to us to discuss them.
**Q1: Novelty in techniques**
Please see the **global rebuttal**.
**Q2: Parameter $\varepsilon$**
This ... | Summary: This paper considers a Bayesian setting of the multi-armed bandit problem and provides finite-time regret bounds. For the Gaussian bandits with independent priors, the authors provide regret analyses for BayesUCB and UCB1 algorithms. Both bounds scale as $O(c_h \log^2 n)$ and match the lower bound by Lai (1987... | Rebuttal 1:
Rebuttal: We would like to thank the reviewer for valuable feedback. We address your concerns below. If you have any additional concerns, please reach out to us to discuss them.
**Q1: Not all constants in the upper bound match asymptotically**
We believe that you mean $\sigma^2$ in the leading term of Cor... | Summary: The authors consider the Bayesian bandits problem and provide first finite-time logarithmic regret upper bounds by analyzing the BayesUCB algorithm from a new perspective, departing from previous studies. Their theoretical results demonstrate the significance of the prior. Additionally, the authors showcase th... | Rebuttal 1:
Rebuttal: We would like to thank the reviewer for valuable feedback and noting the significance of our results. We address your concerns below. If you have any additional concerns, please reach out to us to discuss them.
**Q1: Key challenges in analyzing the finite-time case**
Please see the **global rebu... | Summary: This paper derives a finite-time regret bound for BayesUCB in Bayesian bandits. In Gaussian bandits, the bound grows linearly in $c_h \log^2(n)$, where $n$ is the time horizon, and $c_h$ is a constant determined by the prior. The bound matches that given by Lai (1987). A regret bound of BayesUCB in linear band... | Rebuttal 1:
Rebuttal: We would like to thank the reviewer for appreciating the usefulness of our results and their writing suggestions. $A_t$ is indeed a random variable with values in $\mathcal{A}$. The reviewer is also right that $C_{t, a}$ is a half of the width of the confidence interval centered at the posterior m... | Rebuttal 1:
Rebuttal: We would like to thank all reviewers for feedback, and appreciating the novelty of our work and its potential impact. One comment that appeared in most reviews is that the technical novelty of the work was never clearly presented in one place. We summarize it below and address all other comments i... | NeurIPS_2023_submissions_huggingface | 2,023 | Summary: The paper studies the multi-arm bandit (MAB) problem, which is an online learning problem where a learner interacts with an environment over multiple rounds. In each round the learner would make an action, and each different action will incur a reward sampled from some distribution parametrized by an unknown $... | Rebuttal 1:
Rebuttal: We would like to thank the reviewer for valuable feedback and appreciating the significance of our results. For technical novelty, please see the **global rebuttal**. If you have any additional concerns, please reach out to us to discuss them.
We also wanted to add the following note. We improve ... | null | null | null | null | null | null |
Does Invariant Graph Learning via Environment Augmentation Learn Invariance? | Accept (poster) | Summary: The paper proposes a new framework called Graph invAriant Learning Assistant (GALA) to address the problem of learning invariance among data from different environments for out-of-distribution generalization on graphs. The authors argue that the performance of graph neural networks (GNNs) can be seriously degr... | Rebuttal 1:
Rebuttal: Thank you for your time in reviewing our work. Please see our detailed responses to your comments and suggestions below where we use reference numbers in our manuscript due to the character limit.
> W1 Problem definition
We complemented the problem definition in Sec. 2:
The goal of OOD generali... | Summary: This paper investigates OOD problems on graphs without environmental information. The authors show that it is impossible to learn invariant graph representations without additional assumptions. The authors propose a novel graph OOD method namely GALA based on a new assumption. An auxiliary model is used to mak... | Rebuttal 1:
Rebuttal: Thank you for taking the time in reviewing our work. Please see our detailed responses to your comments and suggestions below where we use reference numbers in our manuscript due to the character limit.
> W1 Notations
We understand it might be difficult for readers who are not familiar with the ... | Summary: This paper discusses the challenges of learning invariance among data from different environments for out-of-distribution generalization on graphs. The authors propose a new framework called Graph invAriant Learning Assistant (GALA) that incorporates an assistant model to differentiate the variations in spurio... | Rebuttal 1:
Rebuttal: Thank you for your constructive suggestions and your positive feedback! Please see our detailed responses to your comments and suggestions below.
> W1 Using two markers wrt. confidence interval.
We revised our table and add shadows to highlight entries with the best performance and a confidence ... | Summary: Invariant graph learning methods aim to extract invariant patterns among different training environments for OOD generalization on graphs. A plethora of approaches focuses on augmenting environment information from training samples as the environment label is difficult to obtain. This work provides both theore... | Rebuttal 1:
Rebuttal: Thank you for your insightful comments and your support of our work! Please see our detailed responses to your comments and suggestions below where we use reference numbers in our manuscript due to the character limit.
> W1 How to guarantee the environment assistant model only learns environment ... | Rebuttal 1:
Rebuttal: Dear reviewers,
Thank you for your comments and suggestions. To summarize, all reviewers agree that our theoretical analysis is clear, significant, and interesting. The motivated solution GALA is novel (waKT, m4Xg, voKo, voKo). The empirical study is comprehensive and extensive to demonstrate th... | NeurIPS_2023_submissions_huggingface | 2,023 | Summary: The paper proposes Graph invAriant Learning Assistant (GALA), targeting at out-of-distribution (OOD) generalization on graphs leveraging invariant graph representation learning. Authors argue that environment augmentation cannot achieve this without additional assumptions, leading to the development of minimal... | Rebuttal 1:
Rebuttal: Thank you for your time in reviewing our paper and your positive feedback! We hope our response below would make you more confident about the evaluation of our work.
> W1 Computational overhead.
We have revised our manuscript to include the analysis. The additional overhead of GALA compared to a... | null | null | null | null | null | null |
Sample Complexity of Goal-Conditioned Hierarchical Reinforcement Learning | Accept (poster) | Summary: The paper presents a theoretical analysis of the sample complexity in goal-conditioned hierarchical reinforcement learning (HRL) and establishes a lower bound using hierarchical decomposition to quantify it. Additionally, the paper empirically validates the theoretical results by examining the sample complexit... | Rebuttal 1:
Rebuttal: We thank you for the time you dedicated to evaluating our work.
We thank you for recognising that we are working on a “critical problem in HRL”. Our bound highlights the importance of the design of the subgoal space. In particular, the high level action space,which in our framework corresponds t... | Summary: The paper takes an important step toward quantifying the benefits achieved due to hierarchical decomposition of an MDP by deriving lower bound on sample complexity of goal-conditioned HRL algorithms. The paper also proposes a novel hierarchical Q-learning algorithm that exploits goal-based hierarchical decompo... | Rebuttal 1:
Rebuttal: Thank you for your enthusiasm about our work and for your detailed feedback.
In what follows we address the concerns raised during your review.
i) “*The assumptions regarding the scope of the derived bounds restricted to a tabular setting need to be clarified in the Introduction.*”
At the moment... | Summary: Provides a sample bound on the complexity of goal-conditioned HRL algorithms based on the two MDPs they are decomposed into, and a Q learning algorithm to leverage these findings. Formulates HRL as a two-level problem, where the upper level passes actions to the lower level policy. The work proves a lower boun... | Rebuttal 1:
Rebuttal: Many thanks for your feedback and for your interest in our work. In the following, we try to clarify your concerns.
Figure 1 is actually composed of two independent figures, the leftmost schema describes the HRL framework and the rightmost plot is simply an illustration of the MDP decomposition i... | null | null | Rebuttal 1:
Rebuttal: We are grateful for the time and the effort the reviewers dedicated to our work, which will strengthen our article. We are also pleased to see that several of the reviewers were enthusiastic about our work, especially on the writing: “clean proof with a highly understandable sketch and a strong in... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Unbounded Differentially Private Quantile and Maximum Estimation | Accept (poster) | Summary: The paper considers the problem of differentially private computation of quantiles for the data. The core idea is to leverage the ‘AboveThreshold’ algorithm as a subroutine to guess the desired item value such that the rank of the item value is about qn for target quantile q. Authors also discuss the applicat... | Rebuttal 1:
Rebuttal: We thank the reviewer for their feedback and apologize for not stating the runtime of our result more clearly. To be very clear, our algorithm only requires ONE pass through the data, so the runtime is O(n) in the data size, and does not require sorted data. We point the reviewer to section 4.2 fo... | Summary: The paper tries to improve upon Sparse Vector Techniques for differential privacy to obtain better estimates of the maximum and quantiles.
Strengths: The paper covers an important problem, finding a differentially private estimate of the max. This is often needed in one of the most common aggregations, i.e. ... | Rebuttal 1:
Rebuttal: We thank the reviewer for pointing out the incorrect reference (should be Thm 2 and 3 for the VLDB version) in Lemma 3.1. Our reference in line 618 is also for Lemma 3.4 in the arxiv version of Cesar and Rogers and needs to be updated.
We felt that the flow and clarity of the introduction was be... | Summary: The authors propose a differentially private algorithm for differentially private quantile estimation that can handle unbounded data setting. The authors revisited the AboveThreshold subroutine in Sparse Vector Technique and proved new privacy guarantee results. The authors can verify the asymptotic runtime co... | Rebuttal 1:
Rebuttal: We thank the reviewer for taking the time to review and give detailed feedback on our work.
Response to questions:
1. We felt that verification of runtime asymptotics was less necessary given the simplicity of the code (see Section D.3 in the supplemental file), but agree that experimental runti... | Summary: The paper considers the problem of computing the quantile q of a dataset with differential privacy. The prior standard technique for this problem was the exponential mechanism with the loss of an output T being the difference between the number of examples less than T and qn. An issue with the exponential mech... | Rebuttal 1:
Rebuttal: We thank the reviewer for thoroughly reading our work, giving extensive and detailed feedback, and are very pleased to hear that they enjoyed doing so.
As the reviewer mentioned we largely left a more detailed examination of beta settings to future work, partly to demonstrate that a reasonable g... | null | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Reconciling Competing Sampling Strategies of Network Embedding | Accept (poster) | Summary: The paper first introduces two desirable properties for network embedding. The discrimination property expresses how well the model is able to distinguish node pairs that are similar from those which are dissimilar. The monotonicity property expresses how the similarity scores obtained from the embedding follo... | Rebuttal 1:
Rebuttal: **Thanks: First, we want to thank you again for your *very detailed comments and many valuable suggestions*, which are very helpful for us to further improve the readability and quality of our paper in the revised version.**
**Response to Q1**:
This is a good observation that the deviation of $... | Summary: This work targets on network embedding under different strategies. The authors first theoretically demonstrate the dilemma in existing network embedding algorithms, especially the sampling-based framework, that the sampling-based embedding performance can have an error gap due to the competing sampling strateg... | Rebuttal 1:
Rebuttal: **Q1: Why is this approach called 'SENSEI’?**
**Response**: Thanks for asking us the question about the name of our model. Here we adopt 'SENSEI' for its similarity to the word 'sense'. The word 'sense' has two levels of meanings: (1) the first level meaning is 'notice'. Our work notices the comp... | Summary: Graph embedding methods typically involve negative sampling, where a set of nodes are sampled as positive samples (usually nodes of close proximity, such as edges) and a set of nodes are sampled as negative samples. While early work tended to randomly sample negatives from non-edges, more recent work has adopt... | Rebuttal 1:
Rebuttal: **Q1: The analysis in Section 2 is quite general with respect to the similarity function $s(u,v)$. Sigmoid and cosine functions are mentioned as examples. What requirements must the similarity function satisfy in order for your analysis to be valid? Would, for example, a Euclidean distance-based s... | null | null | Rebuttal 1:
Rebuttal: We sincerely thank all reviewers for their valuable time and insightful feedbacks which are very helpful for further improving the quality of our paper. We are grateful that the reviewers appreciate the novelty of the idea (`fuvb`), the importance of our work for "unifying existing different sampl... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Transfer learning for atomistic simulations using GNNs and kernel mean embeddings | Accept (poster) | Summary: In this work, the authors investigate the use of kernel ridge regression with feature maps of atomic environments from pre-trained graph neural network models for predicting energies of atomistic systems. They show that their method, mean embedding kernel ridge regression (MEKRR), outperforms or is competitive... | Rebuttal 1:
Rebuttal: > The authors only evaluate their method against one GNN, SchNet; there is a whole zoo of modern GNNs purpose-built for learning potential energy surfaces that incorporate additional physical priors and are fast and easy to train.
As discussed above we do not simply test our method against a GNN,... | Summary: The paper proposes a transfer learning algorithm that leverages the ability of graph neural networks (GNNs) together with kernel mean embeddings. A flexible kernel function that incorporates chemical species information is proposed to improve performance and interoperability. Empirical results in a series of r... | Rebuttal 1:
Rebuttal: > … no substantial theoretical evidence or intuition is provided to motivate this claim.
While we do not provide formal evidence we believe the experimental results speak for themselves. Theoretical justification could be made using results from transfer learning and kernel ridge regression. A ... | Summary: The authors propose to use embedding of GNNs pretrained on large and diverse molecular datasets (in their case OC20) to construct kernel embeddings for efficient transfer learning.
Strengths: This paper was a delight to read. It solves a real problem with a clever idea, implemented well, and I learned someth... | Rebuttal 1:
Rebuttal: Thanks for your positive comments.
> The authors mention meta-learning a few times in the intro section, but meta-learning is a big field and can mean many things. This is hard to understand for the reader without looking up the references.
We will clarify this point further in our revision. Our... | Summary: This paper presents a transfer learning algorithm for interatomic potentials, leveraging GNNs and kernel mean embeddings. It overcomes data and computational challenges, achieving generalization and transferability on realistic datasets.
Strengths: The benchmark provided in this paper is relatively comprehens... | Rebuttal 1:
Rebuttal: >…it would be ideal for the authors to make the complete code available.
The link to the code can be found in the supplementary material (https://anonymous.4open.science/r/mekrr-76F1/).
> Novelty of the proposed method and comparison to to existing GNN works…
We will expand the contribution sec... | Rebuttal 1:
Rebuttal: We thank the reviewers for their useful comments, which we will incorporate in our revision. Before addressing each review in detail, in this global message we answer some key suggestions/questions from the reviewers that we implemented and summarize our contributions.
Note that we have updated T... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Honesty Is the Best Policy: Defining and Mitigating AI Deception | Accept (spotlight) | Summary: In this paper, the authors develop a theory of deception for learning agents in games. Defining deception as intentionally causing a false belief that is not believed to be true, the authors distinguish their theory from others available in the literature. Firstly, with regard to belief, the authors operationa... | Rebuttal 1:
Rebuttal: Thanks for your praise and feedback! Below we try to clarify how our definition of intention does indeed fix substantive problems with the HKW definition and address your other comments.
## Fixing substantive problems with HKW
We say that we fix a problem with the HKW definition, that an agent ... | Summary: The authors propose a criterion for deception in structural causal graph/structural causal games. This requires first defining belief and intent; some parts of these definitions are taken or adapted from previous work. The authors then apply their deception criterion to an RL toy model and an LLM finetuning ex... | Rebuttal 1:
Rebuttal: Thanks for your support for the paper and appreciation for the importance of the research! We agree that the experiments section is the weakest part of the paper and we’re keen to improve it.
1. We will try to make everything much clearer in the updated version
2.
* Agree that the table can be ... | Summary: This paper develops a formal account of deception applicable to AI systems which may not have interpretable representations of "mental states" (e.g. neural networks), but which nonetheless can be viewed as agentic systems. Deception is defined as intentionally causing another agent to have a false belief which... | Rebuttal 1:
Rebuttal: Thanks for your careful engagement with the paper, praise for the theory, and constructive feedback! We agree that the experiments are the least compelling section and we will fix the issues you point out in the updated paper.
## Is GPT-truth claiming ignorance really false and/or deceptive?
* F... | Summary: This paper deals with safety and cooperation with AI systems, it addresses the problem that AI agents may deceive to achieve their goals.
Acknowledging the lack of 'overarching theory of deception for learning agents in games', the authors propose a novel formal definition of deception (Definition 3.7) which ... | Rebuttal 1:
Rebuttal: Thanks for your careful comments on the paper and good ideas for experiments! In the updated paper, we will incorporate your suggestions throughout Section 3 and include new experiments.
# Statistical analysis and code sharing
We agree that statistical analysis and code sharing is an important... | Rebuttal 1:
Rebuttal: Thanks for the engagement with the paper, which offers the first formal definition of deception applicable to AI agents, a solution proposal, and supporting experiments. We’re excited to incorporate your comments, by trying to clarify the denser parts and updating the experiments. First we discuss... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Conditional Mutual Information for Disentangled Representations in Reinforcement Learning | Accept (spotlight) | Summary: The paper mentions that current disentanglement techniques in RL require features to be independent and cannot disentangle correlated features. Therefore the paper proposes an auxiliary task for RL methods based on conditional mutual information (CMI) which can learn to disentangle observations with correlated... | Rebuttal 1:
Rebuttal: We thank you for your review and insightful comments. We are glad that you find the paper well written, and the description of our proposed method clear. We are also pleased that you appreciate our experimental results and ablations. We provide a more detailed response to the comments you raised b... | Summary: The paper proposes an auxiliary task based on minimising mutual information in the learned latent representation. While several approaches have adopted similar loss functions for learning representations, the difference here is that the mutual information is conditioned on the history of past (latent) states a... | Rebuttal 1:
Rebuttal: Thank you for your time and feedback. We are glad that you find the writing clear and the approach well motivated and elegant. We are also pleased that you appreciate the experimental results to demonstrate the effectiveness of our approach. We would like to address specific comments you raised in... | Summary: This paper presents a novel approach to disentanglement in RL, focusing on the challenge of spurious correlations in high-dimensional observations.
The authors argue that conventional disentanglement techniques, which rely on minimizing mutual information between features, can struggle with the task of disen... | Rebuttal 1:
Rebuttal: We thank you for your review and insightful comments. We are glad you find the writing clear and the narrative compelling. We are also pleased that you highlight the elegance of CMID method, and that you appreciate the experimental results. We would like to address comments you described in more d... | Summary: Reinforcement learning often suffers from feature correlations from the limited number of data and feature coverage which hinders the generalization to other environments. To this end, this paper proposes to add an auxiliary task for RL algorithms to learn a disentangled representation of high-dimensional obse... | Rebuttal 1:
Rebuttal: We thank you for your time and feedback. We are glad you find the idea and motivation clear, and that you appreciate the experimental setting we used to demonstrate performance. We would like to address the specific comments you raise below.
> “The experiment results only focus on SVEA. Authors... | Rebuttal 1:
Rebuttal: We would like to thank all the reviewers for their time and feedback. We have provided an individual response to each reviewer to address the key points raised in each review. For some of the responses we have provided additional results, which can be found in the attached one-page pdf.
Pdf: /pdf/... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
ImageBrush: Learning Visual In-Context Instructions for Exemplar-Based Image Manipulation | Accept (poster) | Summary: This paper proposed a novel scheme for image editing via exemplar-based.
An image edit will be represented as an in-context instructions.
To that end, the novel framework is trained to perform image editing as image in-painting task.
Experiment results indicate the advantages of this scheme over typical text-g... | Rebuttal 1:
Rebuttal: Thank you for your thoughtful comments.
**_1. Some parts need to be revised to avoid over-claiming/ clearer_**
1.1-1.2. We greatly appreciate the constructive advice provided on our writing. To ensure accuracy and clarity, we will carefully tune down the claim and provide a more comprehensive in... | Summary: The authors propose a novel, visual-based method for image editing, using a pair of example images as instructions for users to specify the desired operations. The proposed idea uses the details and nuance in the example pairs to accomplish complex editing goals. The authors formulate this approach as a diffus... | Rebuttal 1:
Rebuttal: Thank you for your thoughtful comments.
**_Difficulty in finding example pairs._**
As stated in the common response, we highlight that the practical application of our proposed system is to offer an efficient tool for users to automatically streamline their desired editing operation to a new-com... | Summary: This paper introduces a novel image manipulation framework called ImageBrush, which allows for accurate and efficient image manipulation without the need for complex language descriptions. The framework consists of a diffusion-based generative model coupled with a hybrid context injection strategy, which facil... | Rebuttal 1:
Rebuttal: Thank you for your thoughtful comments.
**_Visual instruction collection seems a challenge._**
As stated in the common response, we highlight that the practical application of our proposed system is to offer an efficient tool for users to _automatically streamline their desired editing operation... | Summary: In this paper, the authors propose a new method that can perform visual prompting via a pair of exemplar images. The authors introduce and train a new model for this task and evaluate the model performance on various datasets, showing strong empirical performance and a wide range of applications. They also pro... | Rebuttal 1:
Rebuttal: Thank you for your thoughtful comments.
**_Weakness_**
1. We further include additional comparison results against VisualPrompt in our appended PDF document. Fig.4 illustrates the qualitative results, revealing that ImageBrush yields plausible and realistic edits whereas VisualPrompt tends to pr... | Rebuttal 1:
Rebuttal: We express our gratitude to the reviewers for their valuable insights and suggestions. In this section, we address a common concern raised by the reviewers regarding the task formulation.
In this study, we introduce a novel protocol for image manipulation based on visual instructions. As indicate... | NeurIPS_2023_submissions_huggingface | 2,023 | Summary: The paper proposes a novel image manipulation technique based on diffusion and visual context images. The authors argue that context from language can only communicate abstract ideas while giving visual context can bring across inexpressible concepts like art style or complex appearances.
The novel technique ... | Rebuttal 1:
Rebuttal: Thank you for your thoughtful comments.
**_How the different context image embeddings are combined._**
We apologize for the lack of clarity in the description of the prompt encoder and will refine this section for improved clarity. Yes, the different context image embeddings are indeed concatena... | null | null | null | null | null | null |
Does Localization Inform Editing? Surprising Differences in Causality-Based Localization vs. Knowledge Editing in Language Models | Accept (spotlight) | Summary: The paper provides an experimental study on localization claims of causal tracing and editing methods such as ROME. Existing literature uses localization results to decide which layer to edit for knowledge editing in large language models. The paper shows that, surprisingly, localization and edit success are n... | Rebuttal 1:
Rebuttal: Thanks for your review! We have a few comments below.
> The Appendix contains additional experiments on zero ablation. Another technique one can try is mean ablation and resampling ablation, over a known dataset such as CounterFact. This may be more principled than zero ablation. See the Causal S... | Summary: Focusing on recently popular model editing methods, the authors study whether localizing methods based on causal mediation analysis play a decisive role. Based on the sample-wise analysis of the tracing effect and rewriting score, they find some surprising results that model editing performances (rewriting sco... | Rebuttal 1:
Rebuttal: Thanks for your review! We have a few comments below.
> The main conflict between this paper with ROME is the viewpoint to switching layers…
It’s true we disagree with the perspective in the ROME paper about how to interpret [39]. The paper [39] shows drops up to 6 percentage points of MNLI accu... | Summary: This paper investigates whether causal localization, which "locates" factual associations within LLMs, provides locations in models which when modified have the greatest effect on those particular factual associations. Counter to intuitions and previous claims in the literature. the authors find that causality... | Rebuttal 1:
Rebuttal: Thanks for your review!
> Line 206 does not appear to relate to the neighborhood score, which measures the change in prediction for similar subjects; rather, it seems to be referring to a metric in which…
We meant for this line to describe the purpose of the neighborhood metric from the ROME pa... | Summary: This paper investigates the assumption that we can edit the located component to modify the facutal knowledge stored in the language model. The assumption might be false due to the empirical findings of this paper: (1) many knowledge facts stored outside the located parts of the model, and (2) the near-zero c... | Rebuttal 1:
Rebuttal: Thanks for your review! We have a few comments below.
> Can you explain why your metrics differ from Rome?
Certainly. (Note: we also check that our results hold with original ROME metrics, see Line282 and Fig 14.)
So we make two changes from the ROME metrics, which are both designed to make t... | Rebuttal 1:
Rebuttal: We would like to thank all of the reviewers for their detailed reading of the paper and thorough feedback. We are so glad to see such a positive reception for the work! Hopefully this paper can help foster much more research in the community about localization and model editing, as well as add nua... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Recaptured Raw Screen Image and Video Demoiréing via Channel and Spatial Modulations | Accept (poster) | Summary: The paper proposed a neural algorithm to remove the moire pattern in an image/video when capturing a display. Since it is observed moire patterns are simpler in raw signals, the proposed method starts from raw data and applies channel and spatial modulations targeting moire pattern properties in the raw domain... | Rebuttal 1:
Rebuttal: Thanks for your nice summary and positive comments on our contributions and dataset. In the following, we give detailed responses to your comments. We only give the outline of your questions due to the limited characters.
> About the comparison with [30].
Thanks! As you mentioned, [30] point... | Summary: This paper aims at tackle the problem of recaptured raw screen image/video demoiréing. A deep demoireing network with a two-branch structure is proposed, where one branch is a cross-channel convolution for shared feature enhancement and the other is a color group convolution with learnable parameter modulation... | Rebuttal 1:
Rebuttal: Thanks for your nice summary and positive comments on our contributions and datasets. In the following, we give detailed responses to your comments.
> The proposed network model does not show an advantage in model size.
We agree with you that our method does not have advantage in model size. In ... | Summary: This paper proposes an image and video demoiréing network tailored for raw inputs. It introduces a color-separated feature branch, and it is fused with the traditional feature-mixed branch via channel and spatial modulations. It builds the first well-aligned raw video demoiréing (RawVDemoiré) dataset and achie... | Rebuttal 1:
Rebuttal: Thanks for your nice summary and positive comments on our datasets and SOTA performance. In the following, we give detailed responses to your comments.
> First of all, the network structure is not very novel. The modulation of channels is similar to channel attention mechanisms, and cross-channel... | Summary: This paper addresses the problem of image/video demoireing through raw image data. The paper shows some observations, such as, the moire pattern in different color channels of the raw inputs have different properties, and pursues how to used this properties for raw image/video demoireing. It proposes to use ch... | Rebuttal 1:
Rebuttal: Thanks for your nice summary and positive comments on our novelty and contributions. In the following, we give detailed responses to your comments.
> Although the paper shows some nice observations, there is no in-depth analysis why these observations happens, e.g., why the R, G1, G2, and B chann... | Rebuttal 1:
Rebuttal: Dear Reviewers,
We sincerely thank you all for your careful review and helpful suggestions, which significantly helped to improve our paper. In the following, we give separate responses to your comments one by one. The figures used for further illustration are presented in the pdf file, and they ... | NeurIPS_2023_submissions_huggingface | 2,023 | Summary: This paper proposed an image and video demoiréing network tailored for raw inputs. The authors constructed the first real raw video demoiréing dataset with temporal and spatial aligned pairs and designed a network tailored with two branches including one branch to share feature enhancement and the other branc... | Rebuttal 1:
Rebuttal: Thanks for your nice summary and positive comments on our datasets. In the following, we give detailed responses to your comments.
> -Although the experimental results demonstrate the effectiveness of the collected dataset, compared with UHDM*, the performance of the proposed network has no signif... | null | null | null | null | null | null |
Equivariant Neural Operator Learning with Graphon Convolution | Accept (spotlight) | Summary: This paper proposes an architecture for learning mappings between signals/functions, which is equivariant to rigid body transformations. The architecture is based on a spectral graph convolutional network based on a novel design of a graphon/graph shift operator (or sequence of graphons/graph shift operators).... | Rebuttal 1:
Rebuttal: # Rebuttal to Reviewer SJUC
We appreciate your recognition of our work's **underlying physical connection** as well as our **comprehensive experiments**. We are willing to address your concerns and questions. Ambiguity regarding SE(3)/SO(3) is addressed in the **rebuttal to all reviewers**.
##... | Summary: This work presents a new model architecture that combines the coefficient learning scheme with coordinate-based residual operator layer for learning mapping between continuous function in 3D Euclidean space. The authors further show the proposed model can be interpreted as graph convolution on a graphon. Exper... | Rebuttal 1:
Rebuttal: # Rebuttal to Reviewer cyPX
We appreciate your recognition of our work's **novel equivariant operator learning** scheme and **rigorous mathematical formulation and proof**. We also thank you for acknowledging our **extensive experiments** on large-scale datasets. We will address your concerns and... | Summary: This paper presents a novel framework for learning mappings between continuous functions in the 3D Euclidean space. First, the framework combines the coefficient learning scheme and the residual operator layer to maintain the sensitivity of the model to equivariance. In addition, the proposed approach utilizes... | Rebuttal 1:
Rebuttal: # Rebuttal to Reviewer sQju
We appreciate your recognition of our **novel design of an equivariant operator learning** scheme and **detailed theoretical interpretation**. We also thank you for acknowledging our **extensive experiments** on large-scale datasets and **superior performance** over th... | Summary: This paper introduces a method for learning SE(3) equivariant continuous functions in R^3 given an approximation of a continuous input and structural information in the form of a graph. The authors represent continuous inputs using ideas from coefficient learning (i.e., learning functions as the sum of learned... | Rebuttal 1:
Rebuttal: # Rebuttal to Reviewer Wmsq
We really appreciate your recognition of our **originality** in designing the novel equivariant operator learning scheme and the good **quality** of the overall work. We also thank you for acknowledging the **significance** of our work in scientific domains. We will ad... | Rebuttal 1:
Rebuttal: # Rebuttal to All Reviewers
Dear Reviewers, we would like to first express our sincere gratitude for your valuable reviews to help make our work more comprehensive and solid. We thank your recognition of the innovation of our work, and we are willing to address your concerns and questions. Here w... | NeurIPS_2023_submissions_huggingface | 2,023 | Summary: The paper introduces an rchitecture combining a coefficient learning scheme with a residual operator layer for learning mappings between continuous functions in 3D Euclidean space designed to achieve SE(3)-equivariance. The proposed method can be seen as convolution on graphons, which are dense graphs with inf... | Rebuttal 1:
Rebuttal: # Rebuttal to Reviewer cHMR
We really appreciate your high recognition of the high quality of our work. We will address your concerns and questions as follows.
## Q1 SE(3)/SO(3) equivariance
Clarification regarding SE(3)/SO(3) is made in the **rebuttal to all reviewers**. We thank you for poin... | null | null | null | null | null | null |
Timewarp: Transferable Acceleration of Molecular Dynamics by Learning Time-Coarsened Dynamics | Accept (spotlight) | Summary: This paper addresses high-throughput MD simulations.
The proposed sampling algorithm simulates the MD of 10^5-10^6 fs timestep by a pre-trained model of a normalizing flow.
The normalizing flow can be used as a proposal distribution in an MCMC method.
The experimental results show the pre-trained model trained... | Rebuttal 1:
Rebuttal: We thank the reviewer for their insightful review and questions. We now address their questions individually. To stay within the character limit for our rebuttals, we often cite parts of the questions.
> No comparison with conventional MD-acceleration methods ...
We do not compare to other MD-a... | Summary: In this paper the authors introduce Timewarp, a method that relies on a normalizing flow as a proposal distribution for enhanced sampling of the Boltzmann distribution. The key results include an acceleration in the MD time step from ~1 fs to 10^5 fs and transferability between peptide systems.
Strengths: T... | Rebuttal 1:
Rebuttal: We thank the reviewer for their insightful review, questions, and suggestions. We appreciate their judgement that our paper is an important contribution. We now address their questions individually.
To keep the rebuttal within the character limit, we often only cite parts of each question.
> The ... | Summary:
This paper presents a Boltzmann generator for small molecular systems in implicit solvent. The generator works demonstrates transferability to test systems of the same class (dipeptides or tetrapeptides). It is the first Boltzmann generator that operates in Cartesian coordinates and demonstrates transferab... | Rebuttal 1:
Rebuttal: We thank the reviewer for their insightful review and questions. We appreciate their judgement that our paper is a significant contribution to this conference. We now address their questions individually.
> The correct operation of this generator requires a Metropolis Hastings step, which will li... | Summary: In this work, the authors present a method for speeding up MD simulations. The speeding up of MD simulations is needed to be able to simulate processes that occur over longer timespans (eg. protein folding). The method proposed by the authors, Timewarp, does so by learning the transition probability of MD simu... | Rebuttal 1:
Rebuttal: We thank the reviewer for their insightful review, and appreciate their assessment that the paper has the potential to be very significant. We now address their comments individually.
> Unfortunately, there's no code available and the datasets are not open sourced.
We will release the code and d... | Rebuttal 1:
Rebuttal: We thank all the reviewers for taking the time to review our paper.
A number of authors have asked for code and datasets to be released. We will release the code upon publication under an open source license. We will also publicly release the datasets that we used to train and evaluate the models.... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Information-guided Planning: An Online Approach for Partially Observable Problems | Accept (poster) | Summary: The paper presents an extension to the POMCP algorithm for online planning under partial observability, called Information-Based POMCP (IB-POMCP). The distinguishing feature of IB-POMCP is that it uses information entropy to enhance the tree search procedure. One effect of this extension is that IB-POMCP is be... | Rebuttal 1:
Rebuttal: We thank your feedback and comments towards acceptance.
### 1.“Argue for the significance of the contributions.”
- **Weak constraints to run**: we refined action selection (via I-UCB) and proposed a novel framework (IB-POMCP) that *does not demand prior data collection or the modelling of latent ... | Summary: This paper presents IB-POMCP, a novel algorithm for online planning under partial observability, that uses entropy estimation to guide the tree search. The main concept is to replace the traditional explore-exploit balancer (UCB1) with (an IUCB score) which is a weighted mixture of value, exploration score, an... | Rebuttal 1:
Rebuttal: We appreciate your insightful feedback and acceptance of our proposal. Your thorough review demonstrates the depth of your evaluation and the meaningful consideration.
### 1.”The paper assumes a reader has a lot of prior knowledge in the area of POMDP.” and “Notation may be not entirely clear for... | Summary: This paper proposes a novel approach for online planning under uncertainty. It leverages the information entropy to guide a tree search process.
Strengths: * This paper is welled structured.
* The paper conducts both empirical and theoretical analysis.
Weaknesses: N/A
Technical Quality: 3 good
Clarity: 2 ... | Rebuttal 1:
Rebuttal: Thank you for your assessment and for considering our paper for acceptance.
We are delighted and encouraged by your positive evaluation and enthusiasm for the work we propose. | Summary: The paper proposes several improvements to the POMCP algorithm. The improvements introduce adaptive sample size, entropy regularized tree exploration and action selection.
Strengths: The research considers an important problem of online planning in POMDP --- dealing with sparse rewards. A few improvements to ... | Rebuttal 1:
Rebuttal: Thank you for your insightful feedback and detailed review.
### 1.“Authors argue that MCTS is based on UCB1. UCB1 is just an option”
- To avoid confusion, we will clarify that UCB1 is an option by writing that it is "widely used" and adding some variations in the Related Works, such as PUCT (Alph... | Rebuttal 1:
Rebuttal: We thank all the reviewers for their valuable feedback. In this global response, we summarise all the major points of our work based on this rebuttal phase’s feedback.
### Novelty
- **Particle filter update**
- Our root's particle filter update strategy is based on the behaviour of tree sea... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Generative Modelling of Stochastic Actions with Arbitrary Constraints in Reinforcement Learning | Accept (poster) | Summary: This paper introduces an algorithm for learning with large discrete action spaces where only some actions are valid for any given state in ways that are hard to programmatically exclude. The paper proposes to use normalizing flows to learn a policy across such action spaces without having to model the full act... | Rebuttal 1:
Rebuttal: We thank the reviewer for the comments and feedback.
We will write an alternate abstract and present it to the reviewer in the discussion phase (we are just not sure if that is allowed in this first round of rebuttal).
In allocation problems, stochastic allocation, if one exists, is often prefer... | Summary: The paper proposes a new approach to solve RL problems with large categorical action spaces and state-dependent constraints. The approach consists of two main components: a flow-based policy network that uses a discrete normalizing flow model to represent the stochastic policy, and an invalid action rejection ... | Rebuttal 1:
Rebuttal: We thank the reviewer for the comments and feedback.
* MASK comparison: The MASK approach can work when the mask can be specified easily. For constraints that vary by state (need a mask per state) and where the number of actions is large specifying the mask is costly - particularly, for our main ... | Summary: The submission investigates reinforcement learning in the context of large categorical action spaces. The submission proposes parameterizing the policy using a state conditional normalizing flow in these settings. The submission derives a rejection-sampling-based method for handling valid action constraints de... | Rebuttal 1:
Rebuttal: We thank the reviewer for the comments and feedback.
We agree that that autoregressive model can also produce stochastic policies. We will modify this statement. We have implemented an autoregressive model (unfortunately, we did not find any publicly available code) and results are in the attache... | Summary: This paper studies challenges in Reinforcement Learning (RL) problems that involve finding an optimal policy where the action space is categorical and large in which current RL approaches do not have a good performance. In addition, these problems often require validity constraints on the actions, which are di... | Rebuttal 1:
Rebuttal: We thank the reviewer for the comments and feedback.
We have implemented an autoregressive model (unfortunately, we did not find any publicly available code) and results are in the attached pdf; also present are results for the marginal/factored approach.
Figure 1 in the attached PDF shows that... | Rebuttal 1:
Rebuttal: We thank all the reviewers for their comments and feedback. We use this space to attach the one-page pdf with extra experiments. We discuss the results in specific responses to the reviewers who asked for the experiments.
Pdf: /pdf/5c96218eba1fbccf320ddb87570736a0b054ea6c.pdf | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
f-Policy Gradients: A General Framework for Goal-Conditioned RL using f-Divergences | Accept (poster) | Summary: This paper studies goal-conditioned reinforcement learning and proposes a framework based on f-divergence. The authors show that minimizing the f-divergence between the agent’s visitation and the proposed goal distribution leads to the optimal policy and discuss the practical implementations, connections with ... | Rebuttal 1:
Rebuttal: We thank the reviewers for lending their expertise to reviewing our paper, as well as for their constructive critique and suggestions for improvement.
We have provided a detailed analysis of the use of the Dirac distribution in our common rebuttal. We request the reviewer to go refer to it if cla... | Summary: The paper focuses on goal-conditioned reinforcement learning (RL), where the goal-conditioned reward function is often sparse. Consequently, standard goal-conditioned RL algorithms suffer from the exploration problem, thus requiring the need to shape rewards. The paper proposes a a new formulation that minimiz... | Rebuttal 1:
Rebuttal: We thank the reviewers for lending their expertise to reviewing our paper, as well as for their constructive critique and suggestions for improvement.
We have addressed several of the concerns in the main rebuttal as well.
Weaknesses:
(1) $\gamma$ is used later in 4.2 when deriving the practical... | Summary: The paper proposes to minimise the f-divergence $J(\theta) = D_f(p_\theta(s) \| p_g(s))$ between the desired state-distribution $p_g(s)$ and the current one $p_\theta(s)$. Clipped likelihood-ratio policy gradient is used to estimate the gradient. Kernel density estimation used to estimate the state visitations... | Rebuttal 1:
Rebuttal: We thank the reviewers for lending their expertise to reviewing our paper, as well as for their constructive critique and suggestions for improvement.
We have addressed several of the concerns in the main rebuttal as well.
Strengths:
(1) Comparison to harder environments: We have provided additi... | Summary: This paper proposes a novel framework for goal-conditioned RL called f-Policy Gradients (f-PG), which minimizes the f-divergence between the agent's state visitation and the goal distribution to provide dense learning signals for exploration in sparse reward settings. The paper derives gradients for various f-... | Rebuttal 1:
Rebuttal: We thank the reviewers for lending their expertise to reviewing our paper, as well as for their constructive critique and suggestions for improvement.
We address each of the concerns raised by reviewer as follows:
Weakness:
(1) F-divergence is not symmetric but we are not sure how that will affe... | Rebuttal 1:
Rebuttal: We thank the reviewers for their reviews. We observe that the reviewers have unanimously agreed to the novelty of the f-divergence perspective of goal conditioned RL. They agree that this method is a promising approach for solving the exploration challenge in sparse reward RL.
We respectfully subm... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
SwapPrompt: Test-Time Prompt Adaptation for Vision-Language Models | Accept (poster) | Summary: Authors propose SwapPrompt for the CLIP-based test-time adaptation task. SwapPrompt consists of two loss functions, swapped prediction loss and pseudo label loss, with a prompt EMA strategy. Experimental results on 14 datasets validate that SwapPrompt outperforms several baselines.
Strengths: 1. The experimen... | Rebuttal 1:
Rebuttal: Thanks for your valuable feedback and suggestions. We have tried to address all your concerns as follows:
**W1:** "The proposed method is ... in NeurIPS conference."
**A1:** Sorry for the insufficient explanation on the novelty of our proposed method. We would like to re-emphasize the novelty an... | Summary: This work presents a framework to perform test-time prompt adaptation for pre-trained vision-language models. In test-time settings, one has no access to labeled test data. Therefore, the authors employ self-supervised contrastive learning techniques on the test domain. In particular, this works proposes an ex... | Rebuttal 1:
Rebuttal: Thank you so much for recognizing the value of our work and for the constructive comments, which have helped us a lot to improve the paper's quality and clarity.
**W1-5:** Please have a native english speaker review the grammar.
**A1-5:** We appreciate your highlighting of the grammatical error... | Summary: This paper studies an interesting and practical task “test-time adaptation for large vision and language models”. The authors propose a novel SwapPrompt method by introducing self-supervised contrastive learning into prompt tuning. The proposed dual prompts paradigm and swapped prediction mechanism enhance per... | Rebuttal 1:
Rebuttal: We would like to thank the reviewer for the constructive and detailed comments, we have tried to address all your concerns one by one.
**W1&W2:** "Two important ... in the authors’ response."
**A1&A2:** Thanks for your valuable comments. We will first explain why "swap prediction" works: inspi... | Summary: The authors propose SwapPrompt, a novel test-time prompt adaptation method that leverages self-supervised contrastive learning to enhance prompt adaptation for pre-trained vision-language models. SwapPrompt employs an online prompt and an EMA updated target prompt, along with a swapped prediction mechanism, to... | Rebuttal 1:
Rebuttal: Thanks for your valuable feedback and suggestions. We have tried to address all your concerns as follows:
**W1:** There is no comparison for robustness to natural distribution shifts which is done in TPT.
**A1:** If we understand your comment correctly, we have already shown the results of robus... | null | NeurIPS_2023_submissions_huggingface | 2,023 | Summary: This paper proposes the test-time prompt adaption method "SwapPrompt" for the vision-language model. The author leverages contrastive learning strategy to learn the prompts on multiple augmented views of unseen data. To better adapt to out-of-distribution tasks, the learnable target and online prompts are appl... | Rebuttal 1:
Rebuttal: Thanks for your valuable feedback and suggestions. We have tried to address all your comments and suggestions on Weaknesses and Questions as follows:
**W1:** The related works might lack some existing works. For example, one of the similar works that the author didn't mention [1] also applied the... | null | null | null | null | null | null |
Training Private Models That Know What They Don’t Know | Accept (poster) | Summary: This paper explores the combination of differential privacy (DP) and selective classification. It examines the performance variations of existing selective classification methods when equipped with DP-SGD. Additionally, the paper demonstrates the complex correlation between selective accuracy and differential ... | Rebuttal 1:
Rebuttal: We thank the reviewer for their feedback on our work and address individual concerns below.
> One potential weakness of this work lies in the exclusive use of DP-SGD for all the differential privacy implementations, which could be a subject of questioning. There exist various methods to enforce d... | Summary: This paper investigates the relationship between selective classification (SC) and differential privacy (DP).
First, the authors consider various SC approaches and if they incur additional privacy/utility costs under the constraint of DP.
They then investigate the effect of DP on SC algorithms and find that DP... | Rebuttal 1:
Rebuttal: We thank the reviewer for their detailed assessment of our work and address individual concerns below.
> Grouping of SC techniques
The main reason for a method to fall into the composition category is that these methods require multiple passes over the data (and hence lead to a higher needed pri... | Summary: This paper explores the interaction between differential privacy and selective classification. Several methods for training selective classifiers are considered and evaluated for their compatibility with DP-SGD. Experimental results are given that show decreasing privacy parameters leads to selective classifie... | Rebuttal 1:
Rebuttal: We thank the reviewer for their feedback on our work and address individual concerns below.
> I think additional attention to presentation is necessary to clarify the main contributions of the paper. [...]
To the best of our knowledge, and as the reviewer acknowledges themselves in the _Strength... | Summary: The paper studies selective classification with differential privacy (DP). In particular, the authors conduct a comprehensive empirical study of different selective classifiers under DP. The key findings and contributions are:
* An existing SC method that uses checkpoints is suitable under DP.
* A novel accur... | Rebuttal 1:
Rebuttal: We thank the reviewer for their strong positive assessment of our work! We provide responses inline below.
> One minor point that could be improved is the explanation of the proposed metrics (Definition 3.1) in lines 255-260. It takes me some time to digest why the authors define the upper bound ... | Rebuttal 1:
Rebuttal: We thank all reviewers for their assessment of our work and have performed a few changes to improve our paper as a result of their feedback.
First and foremost, we were happy to see that **reviewers found our paper well written, that our idea of examining the connection between differential priv... | 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.