Split (1)

train · 7.63k rows

title string	paper_decision string	review_1 string	rebuttals_1 string	review_2 string	rebuttals_2 string	review_3 string	rebuttals_3 string	review_4 string	rebuttals_4 string	global_rebuttals string	dataset_source string	conference_year int64	review_5 string	rebuttals_5 string	review_6 string	rebuttals_6 string	review_7 string	rebuttals_7 string	review_8 string	rebuttals_8 string
Equivariant flow matching	Accept (poster)	Summary: The manuscript builds on recent progress in simulation-free loss functions for continuous normalizing flows which allow to scale CNFs to significantly larger dimensions and can be thought of as a generalizion of continuous time diffusion models. Specifically, the manuscript extends the recently proposed condit...	Rebuttal 1: Rebuttal: We thank the reviewer for their insightful review, and appreciate their assessment that our paper is a valuable contribution in a very exciting and rapidly evolving field. We now address their comments individually. > Unfortunately, the conditional flow matching paper is not the most readable and...	Summary: This paper studies equivariant flow matching, an extension to the flow matching paradigm for training continuous normalizing flows by regressing parametric vector fields to conditional vector fields. Building upon prior work, conditional vector fields can be derived from the distribution of the 2-Wasserstein o...	Rebuttal 1: Rebuttal: We thank the reviewer for their detailed review and questions. To keep the rebuttal within the character limit, we often only cite parts of each question, but always answer the whole one. > How do you align the rotations after performing the Hungarian algorithm? With the Kabsch algorithm, see al...	Summary: The authors propose equivariant flow matching, which provides a way of incorporating syymetries into flow matching objectives. Specifically, they propose to replace the squared Euclidian distance cost used in the flow matching objective, with (approximately) the minimum squared Euclidian distance over all poss...	Rebuttal 1: Rebuttal: We thank the reviewer for their detailed review and questions. We now address their comments individually. > The performance results are mixed - the flow trained with vanilla OT is sometimes better. The difference for equivariant OT compared to normal OT is only marginal for systems with few s...	Summary: This paper extends on existing works on Flow Matching with minibatch OT solutions to the case of invariant cost functions. In particular, where the invariance is given by an SO (permutation + rotation) group. This is mainly a method of correcting the minibatch bias, since (non-equivariant) minibatch OT will st...	Rebuttal 1: Rebuttal: We thank the reviewer for their detailed review and questions. We now address their comments individually. To keep the rebuttal within the character limit, we often only cite parts of each question. > Eq 13 and 14 are not easy to solve ... Eq 13 is indeed intractable, but Eq 14 gives a good trac...	Rebuttal 1: Rebuttal: We thank the reviewers for their time reviewing our paper and for their insightful questions and suggestions. We present results for most of the suggested additional experiments and evaluations below and in the attached pdf. Moreover, we address some common questions and clarifications as well. ...	NeurIPS_2023_submissions_huggingface	2,023	null	null	null	null	null	null	null	null
Frequency-Enhanced Data Augmentation for Vision-and-Language Navigation	Accept (poster)	Summary: Authors propose a Frequency enhanced Data Augmentation (FDA) technique to improve generalization/robustness to unseen environments in the vision-language navigation task. This work provides an initial analysis that shows high frequency information impacts performance. Results are provided on R2R, RxR, and CVDN...	Rebuttal 1: Rebuttal: Answer to Weakness-1: Thanks for your comments. Figure 2 reveals important insights into the behavior of baseline models: HAMT, DUET, and TD-STP. These models show relatively high SR under low-frequency perturbations but experience significant SR drops under high-frequency perturbations. This ...	Summary: In this paper, the authors propose a Frequency-enhanced Data Augmentation (FDA) for VLN models to enhance their ability to capture high-frequency information, thus improving their performance. First, at each step, they take the view as Reference Image and randomly sample an image as Interference Image. Then th...	Rebuttal 1: Rebuttal: **Q1: I wonder why the authors choose to incorporate the high-frequency components from randomly sampled interference images into the original image, rather than using a fixed interference image, or utilizing the high-frequency components of the image itself, or randomizing certain high-frequency ...	Summary: In this paper, the authors investigate the VLN task from a Fourier domain perspective, which aims to enhance the vision-language matching ability of agents. A Frequency-enhanced Data Augmentation (FDA) technique is proposed to improve the capability of capturing high-frequency information, which is seen as an ...	Rebuttal 1: Rebuttal: Q1: The adopted HAMT, DUET, and TD-STP are all transformer-based agents. However, the LSTM-based agents with FDA also need to be verified. A1: Thanks for your valuable suggestion. We conduct experiments on two LSTM-based models: Agent [1] and Follower [2]. The first baseline model, Agent, ach...	Summary: This paper aims at investigating the Fourier domain of vision-and-language navigation (VLN). They propose Frequency-enhanced Data Augmentation (FDA) to improve the visual-text matching processing based on high-frequency visual information. Their method can achieve promising results efficiently on various VLN d...	Rebuttal 1: Rebuttal: Q1: Though the performance looks effective, what is the motivation to rely on Fourier representation instead of the widely-used learned visual features (maybe I missed it, but I cannot find a clear answer in the draft)? A1: Thank you for your question. Our method is essentially a combination ...	Rebuttal 1: Rebuttal: Dear Chairs and Reviewers, We would like to express our heartfelt gratitude to you for managing this paper and dedicating your valuable time to providing insightful comments. We also extend our sincere appreciation to the reviewers for recognizing the significance of our work. We have taken metic...	NeurIPS_2023_submissions_huggingface	2,023	Summary: This paper investigates the sensitivity of existing state of the art VLN models to high frequency perturbations in the image features, and identifies that this could be a way to improve the models performance by making them more robust. It also provides evidence to show that adding high frequency information e...	Rebuttal 1: Rebuttal: Q1: Lacking in implementation details regarding choice of model components such as image encoder, text encoder, etc. A1: Thank you for your concern. Our model components closely match baseline models (TD-STP, DUET, HAMT). For R2R, TD-STP, DUET, and HAMT use ViT-B/16 for image encoder and BERT...	null	null	null	null	null	null
What Planning Problems Can A Relational Neural Network Solve?	Accept (spotlight)	Summary: This paper focuses on formalizing what class of planning problems can be solved by a relational neural networks. It aims to bridge the gap between the expressivity of relational neural networks and the complexity of planning problems. To do this, a serialized goal regression rule and search algorithm is defin...	Rebuttal 1: Rebuttal: Thank you for your helpful comments! Q1: Notations in algorithm 1 and 2. A1: Thank you for all the suggestions. We will update our notations. For algorithm 1, we will include $\mathcal{A}$ as input, define `goal_set` as the set of goals to achieve during the current backward search proce...	Summary: The paper explores the application of relational neural networks, seen as circuits, in representing policies for discrete planning problems. Specifically, it introduces a circuit complexity analysis that categorizes planning problems into three classes based on how circuit width and depth grow, by establishing...	Rebuttal 1: Rebuttal: Thank you for your helpful comments! Q1: Quantitative results: A1: Our main objective in this paper is to formally clarify the nuances of when relational neural network circuits can realize a policy for specific planning problems, and to understand the size of the circuits necessary to a...	Summary: This paper discusses the expressivity of classical relational architectures and its implication for solving planning problems. In particular, the authors identify characteristics of the environments that make planning hard. They prove a theorem that bounds the size of a network sufficient to solve the problem ...	Rebuttal 1: Rebuttal: Thank you for your helpful comments! Q1: Connections to practical problems. A1: Thank you for the suggestion. We have included a new discussion of 6 popular problems that have been widely used in AI planning communities and discuss the circuit depth and breadth of them. We summarized the...	Summary: The paper presents a mapping from tractable but incomplete algorithm classical planning into circuits that can be represented by a class of neural networks called relational neural networks RelNN including transformers architecture and graph neural networks. The paper is divided into two parts. The first part...	Rebuttal 1: Rebuttal: Thank you for your helpful comments! Q1: Replication of experiments. A1: Thanks for the suggestion. We will extend Appendix C to include more details. Also, we have released an anonymized version of the code with the supplement (since NeurIPS rebuttal does not allow for URLs, please chec...	Rebuttal 1: Rebuttal: We thank all reviewers for their thoughtful and constructive comments. Some of the reviewers have suggested applying the theorems and analyses developed in this paper to a broader set of problems. Here, we list 6 popular problems that have been widely used in AI planning communities and discuss th...	NeurIPS_2023_submissions_huggingface	2,023	null	null	null	null	null	null	null	null
AD-PT: Autonomous Driving Pre-Training with Large-scale Point Cloud Dataset	Accept (poster)	Summary: This paper studies model pre-training for autonomous driving by utilizing a large-scale point cloud dataset, which largely facilitates the generalization performance of pre-trained models. To reconstruct such a dataset, scene- and instance-level distribution diversity are carefully enhanced. The overall pretra...	Rebuttal 1: Rebuttal: Response to Reviewer 5kGB Thanks for your review and suggestions. **Q1: Although the AD-PT presented an improved performance on downstream tasks with limited training data (e.g., 20% data amount on Waymo), it is interesting to illustrate the results with more training samples in downstream tasks...	Summary: This paper focuses on the point-cloud pre-training problem in the autonomous driving scenarios. Particularly, the authors regard the pre-training task as a semi-supervised learning problem and generate pseudo-labels for massive unlabeled data with few labeled frames. In the pseudo-label generation phase, they ...	Rebuttal 1: Rebuttal: Q1: More results of using the points, pillars, voxel-based methods to demonstrate the generalization ability of AD-PT. A1: - As mentioned by the reviewer, there are several different types of 3D object detection backbones, e.g. pillar-based, point-based, and voxel-based. We verify the g...	Summary: The paper proposes an autonomous driving pre-training method, where a large-scale pre-training dataset is built and a generalizable representation is created. To further improve the performances on the down-stream tasks, an advanced pre-training technique of semi-supervised method is proposed. Moreover, an unk...	Rebuttal 1: Rebuttal: ## Response to Reviewer hHpC We sincerely appreciate the Reviewer's efforts and comments. We have also tried our best to clarify the novelty of the proposed method and supplement more experimental results and discussions here. **Q1: Novelty is a bit limited, since the proposed techniques look li...	Summary: This paper aims at large-scale point cloud dataset pre-training. The authors propose AD-PT method to build pre-training dataset with diverse data distribution and learn generalizable representations. In details, they design a diversity-based pre-training data preparation procedure and unknown-aware instance l...	Rebuttal 1: Rebuttal: ## Response to Reviewer 8so5 Dear Reviewer XpxU, Thanks for your review, we provide more experimental results and explanations of your question. Q1: Table 3 cannot show the superiority of the proposed method. The performance gain with SS-PT methods such as BEV-MAE is marginal. A1: - D...	Rebuttal 1: Rebuttal: ## General Response: Dear AC and reviewers, Many thanks for your valuable comments and constructive suggestions to improve the quality of our work. Here is a summary of what we have done in the rebuttal phase. - We conduct several new experiments to cover the concerns of the reviewers. -...	NeurIPS_2023_submissions_huggingface	2,023	null	null	null	null	null	null	null	null
RGMIL: Guide Your Multiple-Instance Learning Model with Regressor	Accept (poster)	Summary: The paper presents a new approach for Multiple Instance Learning (MIL) called Regressor-Guided MIL (RGMIL). The contribution of RGMIL is a new aggregation approach Regressor-Guided Pooling (RGP). In this approach, the aggregation part of the MIL model is split into N branches (for N + 1 classes, assuming class...	Rebuttal 1: Rebuttal: Thanks a lot for your many valuable comments. It’s great that you appreciate our originality and investigation of the MIL model. Comment 1：Weaknesses-Originality [R] We will revise the writing thoroughly according to all comments. Comment 2: Weaknesses-Quality [R] We agree with the importan...	Summary: This paper tries to overcome the drawback of low instance-level prediction performance in the existing Multiple Instance Learning (MIL) models. According to the authors, the low instance-level performance is because the existing techniques focus mostly on analyzing the relationship between instances and aggr...	Rebuttal 1: Rebuttal: Thank you a lot for appreciating our efforts in addressing the use of multiple-instance learning (MIL) for instance-level predictions. Comment-1: Weaknesses (about applications) [R] In our method, RGP enhances the instance-level performance explicitly by transferring the learning task to the i...	Summary: This paper presents a new model, the Regressor-Guided Multiple Instance Learning network (RGMIL), which addresses the challenge of providing discriminative instance-level representation in multi-classification MIL scenarios. It introduces an aggregator, the Regressor-Guided Pooling (RGP), that enhances the ins...	Rebuttal 1: Rebuttal: Thank you a lot for appreciating our efforts in addressing the use of multiple-instance learning (MIL) for instance-level predictions. Comment-1: Weaknesses [R]The detailed information of the metrics can be found in section C in supplementary. We will give more details on the UNBC experiments a...	Summary: This paper introduces _Regressor-guided MIL network (RGMIL)_ as a solution to address the challenges in Multiple instance learning, particularly in the multi-class classification scenario for pain-estimation. RGMIL introduces a novel aggregator, Regressor-Guided Pooling (RGP) in place of currently widely used ...	Rebuttal 1: Rebuttal: Thank you for appreciating the effectiveness and novelty of our approach. Comment-1: About writing and descriptions. [R] Thanks for your suggestions, and we will revise our writing so that the proposed method can be easily understood. Comment-2：Missing dataset details [R] We construct the U...	Rebuttal 1: Rebuttal: About RGMIL: RGMIL can address the general multi-class classification problem in MIL scenarios by learning discriminative instance-level features. It holds a different view that we can solve MIL problems by learning on instance level like human does. Inherently coming with interpretability, RGP s...	NeurIPS_2023_submissions_huggingface	2,023	null	null	null	null	null	null	null	null
Revisiting Visual Model Robustness: A Frequency Long-Tailed Distribution View	Accept (poster)	Summary: In this paper, the authors investigate the high-frequency components (HFC) in images from a long-tailed perspective. They revisit the relationship between HFC and model robustness and demonstrate that the cause of the model’s under-fitting behavior is attributed to the limited information content in HFC. A Bal...	Rebuttal 1: Rebuttal: We sincerely appreciate your positive comments for our work. We set out below our responses to each of the questions. - **Q1: To my end, the re-sampling/re-weighting method in the long-tailed situation generally improves the recognition ability of the tails at the expense of a portion of heads, d...	Summary: This paper tackles the problem of visual models' robustness from the view of frequency components. Inspired by the long-tailed characteristic observed in frequency spectrum, the authors empirically prove that standardly trained models are sensitive to HFC. After analyzing its reasons, the authors proposed a sa...	Rebuttal 1: Rebuttal: Thanks a lot for your recognition of the novelty of our work！We set out our answers to each of your questions. - Q1: Baseline comparisons: BaSS is a new sampling strategy, does there exist any other sampling methods for comparison? PGD-AT is from 2017 and is rather an old method. - **Other...	Summary: Using the long-tail features in the spectrum, this paper investigates the model's sensitivity to HFC and discovers patterns of underfitting behavior due to limited information content. The authors propose the Balance Spectrum Sampling (BaSS) strategy, which enhances model learning on HFC. Experimental results ...	Rebuttal 1: Rebuttal: We sincerely thank you for affirming the innovativeness of our article and the positive comments about the writing and experiments and we respond to your questions and suggestions below: - **Q1: The paper lacks a discussion on the existing literature and its relevance to the topic in details, whi...	Summary: This paper studies the effect of power imbalance across frequencies in natural image datasets on the accuracy and robustness of neural net classifiers, and empirically argues that the model is underfitted on frequencies with lower power concentration. Then, the paper proposes an augmentation strategy to improv...	Rebuttal 1: Rebuttal: Thank you to the reviewers for suggestions that help improve the quality of our work，we make the following responses (e.g., R1 is the response to Q1): - R1:In Line113-117, we have revised it as:"Previous analysis indicating that a few low-frequency components form the image's main content, whi...	Rebuttal 1: Rebuttal: Dear Area Chair and Reviewers, We would like to thank all reviewers for their thoughtful comments and insights that greatly helped improve our work. We have carefully reviewed and responded to every question from all reviewers, providing a clearer explanation of the content of the article in ques...	NeurIPS_2023_submissions_huggingface	2,023	Summary: This paper addresses the problem of model robustness, which is of great importance in the machine learning field. The authors revisit this problem and discuss the relationship between high-frequency components and model robustness. They verify the effectiveness of the proposed method on several datasets compar...	Rebuttal 1: Rebuttal: We are grateful to the reviewer for your time and effort in evaluating our manuscript. We have responded to each question in detail below： - **Q1:Have you considered experimenting with other types of attacks to evaluate the performance of the proposed method in different scenarios? Additionally, a...	null	null	null	null	null	null
How to Select Which Active Learning Strategy is Best Suited for Your Specific Problem and Budget	Accept (poster)	Summary: Some recent works have shown that typical examples are more beneficial than more uncertain ones under low budget setting in active learning. However, there has been no prior works that address how to distinguish whether a budget is in low or high budget regime, or how to determine which approach should be take...	Rebuttal 1: Rebuttal: Thanks much for the detailed and precise list of questions. We wish to emphasize that some brevity in the manuscript was called for because the general framework is not original - it is adopted with all its assumptions from the work of [1]. This study demonstrated that the model, despite its simpl...	Summary: This paper considers the selection of active learning methods under different budgets. Previous studies have shown that different active learning methods perform differently under different labeling budgets, and incorrect selection can lead to model performance inferior to that of random sampling baselines. To...	Rebuttal 1: Rebuttal: $B_{low}$ and $B_{high}$, and the connection between the theoretical analysis and the practical methods Indeed, the theoretical framework presented in this study relies on a simplified model. This makes possible a rigorous analysis and the derivation of closed-form solutions, unlike the invol...	Summary: The paper studies an algorithm selection problem for active learning under different labeling budgets. A theoretical analysis is conducted The proposed algorithm uses a derivative approach by removing small sets of labeled examples to approximate the effectiveness of different active learning algorithms. Empir...	Rebuttal 1: Rebuttal: Connections between the theoretical analysis and the actual algorithm presented: The theoretical analysis presented in our work conveys two main take-home messages: 1. The effectiveness of a derivative-like test, highlighted in Section 2.2 and emphasized in Equation 1, proves to be valuable...	Summary: The literature has shown that the optimal type of active learning (AL) strategy depends on the type of budget. Methods based on uncertainty sampling are most effective when the budget is large, and methods based on typicality work better in the low budget setting. However, what qualifies "large" or "low" depen...	Rebuttal 1: Rebuttal: We thank the reviewer for the references; indeed a discussion of low and high budget, independently of active learning, will put the present work in greater context. We will modify our related work section accordingly, and cite the appropriate references. Specific comments raised by the reviewer...	Rebuttal 1: Rebuttal: We thank all the reviewers for the significant time and effort they made in reviewing our work. In the attached PDF, we added an experiment suggested by Reviewer GcEN, in which we use other active learning strategies, namely ProbCover [1] and CoreSet [2], as the decision rules for SelectAL. Speci...	NeurIPS_2023_submissions_huggingface	2,023	null	null	null	null	null	null	null	null
Evaluating Post-hoc Explanations for Graph Neural Networks via Robustness Analysis	Accept (oral)	Summary: This paper presents a new metric, OOD-resistant Adversarial Robustness (OAR), for evaluating post-hoc explanation methods for Graph Neural Networks (GNNs). Inspired by adversarial robustness, OAR performs random perturbations on the complementary part of the explanation result, reducing the impact of the Out-O...	Rebuttal 1: Rebuttal: Dear Reviewer vGHJ: Thank you for the thoughtful feedback! Your constructive criticism has been invaluable in refining our work. Below, we give point-by-point responses. Hope our responses could address your concern! >*W1 & W2: Concern about the training process in OAR, while SimOAR addresse...	Summary: The paper proposes a novel evaluation metric called OOD-resistant Adversarial Robustness (OAR) to address the issue of assessing the credibility of explanations provided by graph neural networks (GNNs). The paper criticizes existing evaluation methods that fail to consider out-of-distribution (OOD) data and ca...	Rebuttal 1: Rebuttal: Dear Reviewer uvv7: We gratefully thank you for your valuable comments! Here we meticulously give point-by-point responses to your comments. Hope that our responses could address your concerns! >W1: Figure 1 is kind of hard to follow. It could be better if authors can refine it. Thanks ...	Summary: This paper presents a new evaluation metric OAR for evaluating post-hoc explanation methods for GNNs. OAR has two main advantages in that it does not need ground-truth labels and in that it prevents finding spurious explanations based on out-of-data-distribution phenomena. The paper focuses on graph classifica...	Rebuttal 1: Rebuttal: Thank you very much for your valuable comments. We hope that the following reply could address your concerns! >W1: Concerns about the Introduction and the missing related work [1-5]. Thanks for bringing this point to us. Following your suggestions, we have implemented the following changes ...	Summary: This paper studies the explainability evaluation of GNNs, and proposes a new evaluation paradigm. Inspired by adversarial robustness, it uses a generative model (VGAE) to fulfill the explanation subgraph, randomly perturbs the fulfilled parts, and uses the OOD scores of perturbed graphs to measure the importan...	Rebuttal 1: Rebuttal: Dear Reviewer QMrv: Thank you for the thoughtful feedback! Your constructive criticism has been invaluable in refining our work. Below, we give point-by-point responses to your comments. Hope that our responses could address all your concerns! >*W1 & Q1: Does the adversarial robustness ...	Rebuttal 1: Rebuttal: Dear Reviewers: We gratefully thank you for your valuable comments! We were encouraged to hear that our work has clear and well-written presentations (by all Reviewers), well-designed and interesting technical contributions (by Reviewer QMrv and vGHJ), **extensive and sufficient experimen...	NeurIPS_2023_submissions_huggingface	2,023	null	null	null	null	null	null	null	null
On Sparse Modern Hopfield Model	Accept (poster)	Summary: The authors introduce sparse Hopfield networks that feature sparse retrieval dynamics corresponding to sparsemax attention mechanisms and resulting in sparse patterns that are more robust to noise. They prove fast convergence analogous to modern Hopfield networks and show how the sparse Hopfield model has a ti...	Rebuttal 1: Rebuttal: ### The revised draft is in the same Dropbox folder (in Supplementary Material PDF). --- >Reviewer's comment: The papers originality is fair since sparse computation is a well-known topic in Deep Learning and the presented work is merely an incremental improvement on modern Hopfield networks....	Summary: The authors introduce sparse Hopfield model, which are memory-associative models used to store and retrieve patterns. Theoretically, the authors connect sparse Hopfield with sparse attention mechanism and empiricially the authors show how their method can outperform the state-of-the art. Strengths: The author...	Rebuttal 1: Rebuttal: ### The revised draft is in the same Dropbox folder (in Supplementary Material PDF). --- >Reviewer's Comment: The paper in terms of language is very hard to follow for layman readers Response: Thank you for your feedback. We understand the need for making the paper more accessible to read...	Summary: This paper proposes the sparse modern Hopfield network. It studies the new proposed model from both perspectives, theoretical and empirical, validating the approach. Strengths: * The proposed model and the introduction of sparsity in the modern Hopfield network seem novel * The theoretical analyse is well rec...	Rebuttal 1: Rebuttal: ### The revised draft is in the same Dropbox folder (in Supplementary Material PDF). --- >Reviewer's Comment: The paper is very hard to read, with a mathematical exposition that seems more complex than necessary. It lacks basic intuitions about the proposed approach. Response: We acknowle...	Summary: The paper proposes a new model from the Dense Associative Memory family that uses Sparsemax function for its energy. This model is studied and compared to the model with the softmax-based energy from both theoretical and empirical perspectives. The work proposes that the sparse model outperforms the model with...	Rebuttal 1: Rebuttal: ### The revised draft is in the same Dropbox folder (in Supplementary Material PDF). --- > Q1 & It would be nice if the authors could summarize in a concise and crisp way... Response: Thank you for your comment. Please see below summary of the advantages of our sparse model over its dense cou...	Rebuttal 1: Rebuttal: Dear Reviewers, We thank the reviewers for the insightful questions and reviews. We have answered all the questions and addressed all the problems in detail in rebuttal and revision. The latest revision is readily available in the reproducibility code folder. Any changes or modifications made f...	NeurIPS_2023_submissions_huggingface	2,023	null	null	null	null	null	null	null	null
POP-3D: Open-Vocabulary 3D Occupancy Prediction from Images	Accept (poster)	Summary: This paper proposes an approach, POP-3D, to achieving open-vocabulary 3D occupancy prediction from images by aligning 3D features with those 2D features from pre-trained MaskCLIP+. Specifically, in contrast to previous occupancy prediction methods, such as TPVFormer, this framework uses two heads to predict cl...	Rebuttal 1: Rebuttal: We thank the reviewer for their feedback. Method novelty We would like to thank the reviewer for the comment. Although it is true that POP-3D uses image-text aligned features provided by MaskCLIP+, we argue that exploiting such features is an interesting and non-trivial starting point to ex...	Summary: This work presents a system that is able to achieve open-vocabulary 3D semantic volumetric predictions with multi-view image inputs. Given a set of images, the system learns a 3D voxel grids which decodes into binary occupancy and visual-linguistic embeddings. The embeddings are used for retrieval and semantic...	Rebuttal 1: Rebuttal: We thank the reviewer for the feedback. Comparison to other open-vocabulary methods: LeRF, Semantic-Abs, OpenScene, CLIP-FO3D Thank you for suggesting relevant works which tackle a similar problem to ours. We will integrate a discussion about those works in our related work section. For no...	Summary: The paper presents a novel approach for predicting open-vocabulary 3D semantic voxel occupancy maps from 2D images. The objective is to enable 3D grounding, segmentation, and retrieval of free-form language queries, which is challenging due to the ambiguity between 2D and 3D representations and the difficulty ...	Rebuttal 1: Rebuttal: We thank the reviewer for their feedback and suggestions. We address here each comment/question. Evaluating open-vocabulary capabilities on hundreds of object categories We would like to highlight that open-vocabulary approaches for image-based 2D tasks (semantic segmentation, object detecti...	Summary: This paper proposes a camera-only open-vocabulary 3D occupancy prediction method, named POP3D. POP3D consists of a 2D-3D encoder and 3D language heads, which can combine open-vocabulary segmentation with 3D occupancy prediction together. The proposed method can be trained in a self-supervised manner by leverag...	Rebuttal 1: Rebuttal: We thank the reviewer for the feedback. We address here the concerns individually. Details about ground-truth generation We happily provide more details about the ground-truth generation. We use a single LiDAR sweep (i.e., only a single LiDAR point cloud from a single time step without point...	Rebuttal 1: Rebuttal: We thank the reviewers for their valuable comments, which will allow us to improve the quality of the paper. We are glad to note that they appreciated the clarity [nph2, yGBj, Vrgc, q7fa] of the manuscript, the well-defined motivation of the tackled problem [yGBj, q7fa], the originality [*np...	NeurIPS_2023_submissions_huggingface	2,023	Summary: This work presents a novel approach for zero-shot 3D occupancy prediction for autonomous driving applications. The key idea consists of three parts: - a 2D-3D encoder to create a 3D voxel feature grid, based on TPVFormer - An MLP-based voxel occupancy predictor that is class agnostic - an MLP-based feature enc...	Rebuttal 1: Rebuttal: We thank the reviewer for the feedback and suggestions. We address here each comment/question. Datasets for evaluation In this work, we leverage TPVFormer for the multi-camera 2D to 3D projection. To show the effect of our contribution compared to a fully-supervised setting in a fair way, w...	null	null	null	null	null	null
Automatic Clipping: Differentially Private Deep Learning Made Easier and Stronger	Accept (poster)	Summary: The paper proposes a new alternative version of the DP-SGD algorithm where the gradients are normalized. This new method allows for a proof of convergence of the gradient when we add a small stability constant to the rescaling factor. This work eliminates the need for a costly 2D gridsearch between learning ra...	Rebuttal 1: Rebuttal: We thank the reviewer for the detailed comments! We will address the comparison to re-scaled DP-SGD in (De et al.) from the aspects of (1) convergence and hence performance, (2) better understanding of general DP optimizer, (3) saving the hyperparameter tuning efforts, (4) extensive experiments. ...	Summary: The author(s) proposed a new gradient clipping technique for differential private training algorithms. It was shown that the new clipping technique is more robust to hyper-parameters and can save the time for hyper-parameter tunning. Convergence of the proposed method under a gradient symmetric assumption is d...	Rebuttal 1: Rebuttal: Thank you for your enlightening review. We agree that Theorem 4 is not directly comparable to prior DP works, because our proof techniques are novel and settings are different. We hope the reviewer would agree that our setting (including the symmetric gradient noise, not symmetric gradient) is mor...	Summary: The paper discusses a way to control the sensitivity of DP-SGD by a method that does not require clipping. Strengths: - The motivation for the paper is clear - The authors make a compeling argument that suggest why their method should work - They have sound theoretical guarantees and experiments - The idea i...	Rebuttal 1: Rebuttal: Thank you! Happy to extend the discussion any time.	Summary: This paper introduces 'automatic clipping' to replace the usual clipping operation in DP-SGD. The challenge of usual clipping is to choose a good threshold 'R' especially for deep learning models. The authors claim that automatic clipping maintains the same level of privacy and computational efficiency as exis...	Rebuttal 1: Rebuttal: Thank you for your review. We will address the your comments point-by-point. Misleading justification & Overclaim: the dot product is not normalized and it is indeed aligned with good convergence in the following sense. Under our non-convex but Lipchitz smooth setting (see Appendix C.2), $L_{t+1}...	null	NeurIPS_2023_submissions_huggingface	2,023	Summary: The script proposed an automatic clipping methods for various DP algorithms. The problem is important because the performance of DP models are sensitive to the choice clipping threshold, yet there is no theorical guidance for tuning it. Strengths: The writing is clear and the paper is easy to follow. The i...	Rebuttal 1: Rebuttal: Thank you for appreciating our work! We will add index to equations as suggested. We expect the optimal clipping, as we already stated in line 140, to have a better performance BUT it will not lead to DP guarantee. The reason is due to the term $\sum_j g_j$, which changes when adding or removing ...	null	null	null	null	null	null
Using Imperfect Surrogates for Downstream Inference: Design-based Supervised Learning for Social Science Applications of Large Language Models	Accept (poster)	Summary: The author presents a complete and thorough study in their research regarding semi-supervised learning of social science application targeting bias and coverage control. The work include clear mathematical proof of the purposed approach, and experimental results demonstrating the well alignment of theory and a...	Rebuttal 1: Rebuttal: Thank you for the review! We are glad you found the theoretical and empirical work in the paper enjoyable! Clarify the RMSE gain of DSL over GSO: First, based on suggestions from reviewers, we further clarified the RMSE gains of our estimator compared to existing alternatives. In particula...	Summary: In order to exploit the advantages of Large Language Models (LLMs) in computational social science (CSS) area, this paper proposed a novel DSL estimator, which employed a doubly-robust procedure to combine surrogate labels with a few gold-standard labels. By using the proposed method, not only LLMs can be used...	Rebuttal 1: Rebuttal: Thank you for the review! We are glad you enjoyed the theory and the extensive experiments. Clarify the RMSE gain of DSL over GSO: First, based on suggestions from reviewers, we further clarified the RMSE gains of our estimator compared to existing alternatives. In particular, we show that...	Summary: This paper studies the use of surrogate labels in model training and it's bias, coverage and performance properties. Practically, the premise is to use a small number of gold data annotations for a classification task, then construct surrogate labels that have desirable bias, coverage and performance propertie...	Rebuttal 1: Rebuttal: Thank you for the review! We are glad that you enjoyed the proofs, experiments, and premise. Below, we quote the weaknesses you identified and respond. (1) *The key weakness is that the application presented as a motivation for this paper is not tested throughout any of the experiments. The...	Summary: The authors consider using outputs of LLMs for labeling in downstream analysis in computational social science while guaranteeing statistical properties (asymptotic unbiasedness and uncertainty quantification). They show that directly using LLM-produced labels yields bias and invalid confidence intervals, then...	Rebuttal 1: Rebuttal: Thank you for the kind words about the paper! We are glad that you found the setting and methodology well-described and motivated. (1) *Perhaps the biggest weakness of the paper is that [...] it prioritizes unbiasedness and coverage, however, DSL does not outperform GSO which is a very simpl...	Rebuttal 1: Rebuttal: Thank you to all the reviewers for the careful engagement with our work. In this paper, we proposed how to use imperfect LLM annotations in downstream regression analyses, while guaranteeing asymptotic unbiasedness and proper uncertainty quantification, which are fundamental to social science rese...	NeurIPS_2023_submissions_huggingface	2,023	null	null	null	null	null	null	null	null
DynGFN: Towards Bayesian Inference of Gene Regulatory Networks with GFlowNets	Accept (poster)	Summary: In this work, the authors propose an extension of GFlowNets to enable posterior sampling of cyclic graph structures and apply their approach to infer gene regulatory networks (GRN). While existing methods are able to infer cyclic graph structures or sample from Bayesian posteriors over DAGS, they cannot do bot...	Rebuttal 1: Rebuttal: We thank the reviewer for their comments, insightful feedback, and helpful suggestions to improve the manuscript. Please refer to our feedback summary and the attached document regarding questions/clarifications on Fig. 1, Section 5 and Fig. 2, and Fig. 3. In the following, we address individual c...	Summary: The authors tackle the well-known problem of Gene Regulatory Network (GRN) inference from time series data. Starting from an estimate of RNA velocity, GRN inference is framed as a causal discovery task, with two specificities: the inferred graphs may be cyclic, and a probabilistic distribution over candidate g...	Rebuttal 1: Rebuttal: We thank the reviewer for their comments and insightful feedback. We address your comments in what follows. >The paper builds on GFlowNets for the graph sampler and Hypernetworks to generate the structural causal model parameters. I found it slightly difficult to understand everything given that ...	Summary: Authors introduced a principled methodology called DynGFN, which effectively identified cyclic structures and concurrently modeled the Bayesian posteriors over directed acyclic graphs (DAGs). Leveraging RNA velocity, the authors formulated a dynamic system that unveiled the underlying gene regulatory networks....	Rebuttal 1: Rebuttal: We thank the reviewer for their comments and insightful feedback. Please refer to our summary to all reviewers and the attached document regarding questions about cyclic dependencies in our synthetic and real experiments, as well as for clarifications regarding Fig. 1. We also introduced a new toy...	Summary: The authors propose to learn dynamical systems with cyclic dependencies. They factorise the generative model using a variant of the GFlowNet model, a HyperNetwork and MLPs. Strengths: Positives include: 1. extending the formulation from DAGs to cyclic graphs 2. introducing a per-node posterior formulation to...	Rebuttal 1: Rebuttal: We thank the reviewer for their comments and insightful feedback. Please refer to our feedback summary and the attached document regarding questions about cyclic dependencies in Fig. 2 and Fig. 3. We also include further discussion regarding the toy example in the feedback summary. We now address ...	Rebuttal 1: Rebuttal: We would like to thank all the reviewers for their time and insightful feedback when reviewing our paper. We appreciate the constructive criticisms and suggestions which will serve to improve the overall quality of our paper. Our method focuses on Bayesian structure learning over cyclic graphs f...	NeurIPS_2023_submissions_huggingface	2,023	null	null	null	null	null	null	null	null
Accessing Higher Dimensions for Unsupervised Word Translation	Accept (poster)	Summary: The authors argue that relying solely on low-dimensional vectors misses out on better denoising methods and valuable world knowledge present in high dimensions. Therefore, they propose coocmap, a method that can exploit high-dimensional co-occurrence counts. Through extensive experiments, the authors show cooc...	Rebuttal 1: Rebuttal: Thank you for the helpful review and corrections, which will help us improve the revision. > Line 12: "similar data", the author should specify similar in terms of what? We will specify similar in domains. > The authors should provide more intuition on why clip and drop are very helpful for har...	Summary: This paper proposes using high-dimensional co-occurrence statistics for unsupervised word translation rather than relying on low-dimensional vector embeddings. They propose using coocmap (uses association matrix of cooccurrences) and combining this with regularization operations (clip/drop) that eliminate nois...	Rebuttal 1: Rebuttal: Thanks for the review and feedback. We are glad that you liked the content of the paper and we will try harder to make the paper more articulate in future revisions. > Is the following an accurate characterization of the work: low-dimensional representations are not expressive enough for this tas...	Summary: This paper introduces coocmap, an approach for unsupervised word translation that uses high-dimensional co-occurrence statistics instead of lower-dimensional word embeddings. The approach is generally analogous to vecmap, alternating between distance computations and a matching phase. Coocmap is often more dat...	Rebuttal 1: Rebuttal: Thank you very much for the review and concrete corrections. We will make the corrections in the final version. > Given that unsupervised word translation has been introduced many years ago, but arguably has limited practical applications, the authors should more clearly motivate their work. The...	Summary: This paper proposes coocmap, a method for unsupervised word translation that uses high-dimensional co-occurrence statistics instead of low-dimensional vectors. The authors show that relying on low-dimensional vectors can lead to suboptimal denoising methods and overlook useful world knowledge in high dimension...	Rebuttal 1: Rebuttal: Thanks for the review, we are glad to hear that you found the paper easy to follow. In particular, we claim that we are the first to do unsupervised word translation with just co-occurrences and there are several difficulties with doing this which we identified in the paper. More details in the an...	Rebuttal 1: Rebuttal: We thank all reviewers for the insightful and helpful reviews. We will aim to improve the draft using the feedback and make it more clear. A few common points are emphasized here and other points and questions are addressed in individual responses. 1. We are the first to use just co-occurrences t...	NeurIPS_2023_submissions_huggingface	2,023	Summary: In this paper, the authors target to solve unsupervised word translation, also called lexicon or dictionary induction, using their proposed method, coocmap, a simplified version of the conventional way, vecmap. Different from vecmap, coocmap can estimate word mapping without using rotation weights for row vect...	Rebuttal 1: Rebuttal: Thanks for the review with many concrete points that may help us avoid possible confusions. We first address the weaknesses and then the questions. We emphasize that the most substantial proposed weakness about unseen word is wrong. > Not using the rotation matrix means coocmap cannot handle unse...	null	null	null	null	null	null
Gacs-Korner Common Information Variational Autoencoder	Accept (poster)	Summary: The authors present a concept of common information that assesses and distinguishes shared information from unique information between two random variables. This notion, which is a variant of the Gacs-Korner common knowledge, is more optimizable and experimentally approximable when employing sample data. ...	Rebuttal 1: Rebuttal: We thank the reviewer for their comments and address their questions below. > The authors experimentally demonstrating that adding unsupervised viewpoints improves disentanglement; which is likely due to limitations in single-sample analysis; active interaction, not passive observation, fosters b...	Summary: The basis of the proposed work is a practical means to decompose the information contained in two random variables into common and unique components. With a couple tweaks to a vanilla VAE setup, data that has been paired such that certain factors of variation have a constant value within the pair can be used ...	Rebuttal 1: Rebuttal: We thank the reviewer for their comments and address their concerns below. To understand the difference between GK information and mutual information and why GK information is desirable, consider the following settings. Case 1: Let $X_1 = C + N_1$ and $X_2 = C + N_2$, where $C$ and $N_i$...	Summary: The paper proposes the Gracs-Korner Common information to measure common information between two random variables and propose a variational relaxation to compute the practical loss. The paper then derives an objective that eases disentanglement requirement. Strengths: 1. The paper proposes the Gracs-Korner Co...	Rebuttal 1: Rebuttal: We thank the reviewer for their comments and address their concerns below. > The proposed method is only compared with basic VAE where as the advantage of the proposed method is not significant. The purpose of our study was to show formally and empirically that we can partition the latent repre...	Summary: This paper proposes a new way to partition common and unique information in multi-view data, by leveraging and optimizing the objective of common information defined by Gacs and Korner. Authors extended the deterministic setup to stochastic scenario and used a variational autoencoder to realize the objective. ...	Rebuttal 1: Rebuttal: We thank the reviewer for their feedback and address their concerns below. Q1: We elaborate on the difference between common information and mutual information in our response to [Reviewer iAUJ](https://openreview.net/forum?id=e4XidX6AHd&noteId=HqfjeyTDaR) (see Case 1 and Case 2 in particular). ...	Rebuttal 1: Rebuttal: We thank all the reviewers for their thoughtful comments and suggestions, which we believe will lead to a clearer and improved manuscript. We have responded to each of the reviewer's questions individually. Please see the attached pdf for figures and a table that we refer to in our responses to ...	NeurIPS_2023_submissions_huggingface	2,023	null	null	null	null	null	null	null	null
CLadder: Assessing Causal Reasoning in Language Models	Accept (poster)	Summary: The paper collects a new dataset, named Cladder, that aims at evaluating the causal reasoning abilities of language models. The datasets contain 10,000 questions asking about boolean questions that require intensive causal reasoning. The dataset is constructed based on the idea of a causal inference engine, wh...	Rebuttal 1: Rebuttal: We thank the reviewer for the overall positive comments about the meaningfulness of the problem (“The paper works on an interesting problem”), dataset contribution (“a useful dataset consisting of 1,0000 examples”), “well-designed” “data creation process” that “covers multiple rungs and multiple c...	Summary: The paper introduces a large dataset, CLADDER, designed to evaluate the ability of LLMs to perform causal inference. The dataset includes 10K binary questions about variable association and treatment efficacy along with a ground truth explanation. Questions are answered in the context of an hypothetical closed...	Rebuttal 1: Rebuttal: We thank the reviewer for providing very detailed comments and an elaborate summary of the work. We appreciate the reviewer’s comments that “The paper nicely builds on the framework proposed by Pearl with a dataset of 10K questions covering a large variety of causal graph structures and estimands....	Summary: The author proposes a math-word-style formal causal reasoning dataset, CLADDER, which is the first one that can be used for formally evaluating the causal reasoning/inference ability of LLMs. The authors also proposed a new prompting strategy, CausalCoT, which improves the zero-shot performance of GPT-4 by 2.4...	Rebuttal 1: Rebuttal: We thank the reviewer for their positive feedback that “The paper is well-written and easy to follow”, the dataset is valuable as “It is the first formal causal inference/reasoning dataset in a natural language, math-word-problem-like format”, which constitutes “a nice addition to the current logi...	Summary: The research question is whether large language models are capable of doing formal causal reasoning as formalized by the approach with SCMs popularized by Pearl. The authors provide a benchmark of 10k examples to evaluate such abilities; they describes in detail the design choices for this benchmark. They then...	Rebuttal 1: Rebuttal: We thank the reviewer for recognizing that “The design principles behind the benchmark are well-explained”, the dataset is diverse with “A variety of topologically distinct treatment-effect pairs” “covering thus multiple causal graphs”, the experiments are multi-faceted with “Non-sensical and anti...	Rebuttal 1: Rebuttal: Firstly, we would like to thank all reviewers for a wealth of feedback. Four out of five reviewers recommended acceptance, and we are encouraged by the large number of thoughtful comments and suggestions, which demonstrate not only that all reviewers largely understood our focus, but also that the...	NeurIPS_2023_submissions_huggingface	2,023	Summary: This paper evaluates the causal reasoning ability of language models and proposes a structured prompting approach to improve this ability. For this goal, the authors create a benchmark. The benchmark includes synthesized causal graphs and related queries. The answers to queries are computed based on a causal r...	Rebuttal 1: Rebuttal: We appreciate the reviewer’s comment that this work “focuses on an important problem”, “The created benchmark is potentially a valuable evaluation resource”, and our CausalCoT is “a novel structure that is based on causal reasoning steps”. > Is this totally zero-shot? or do you do in-context lear...	null	null	null	null	null	null
Estimating and Controlling for Equalized Odds via Sensitive Attribute Predictors	Accept (poster)	Summary: This work provides tight and computable bounds on the EOD violation of a classifier in a setting where the sensitive attributes are unknown. In addition, a post-processing technique is proposed to provably yield classifiers that maximize prediction power while achieving minimal worst-case EOD violations with r...	Rebuttal 1: Rebuttal: Comment/Concern 1: Assumption of the proposed approach is that there exists a dataset of $(X,A)$ to learn $\widehat{A} = h(X)$. This could limit the use of the approach and should be discussed Answer: We completely agree. All the contributions of our work rely on there being a dataset of $(X,...	Summary: The paper explores the fairness problem in machine learning when sensitive attributes (e.g., demographics and gender) are unavailable for practitioners. The authors focus on the Equalized-Odds (EOD), a well-known definition of fairness for classifiers. The authors provide bounds for EOD violation in a setting ...	Rebuttal 1: Rebuttal: Comment/Concern 1: More real-world experiments to convince us that Assumption 1 is followed in many scenarios and your method helps to improve fairness. Answer: Thank you for the suggestion. We have performed additional experiments to show that Assumption 1 holds in multiple scenarios and sho...	Summary: The paper proposes a tight upper bounds for the equalized odds violation of a predictor in a setting without sensitive attributes. It also presents a post-processing correction method to control the worst-case equalized odds violation and presents results on a variety of synthetic and real datasets. Strength...	Rebuttal 1: Rebuttal: Comment/Concern 1: Limitations of Assumption 1 and how they might be addressed Answer: Absolutely. For clarity, we re-write Assumption 1 below: For $i,j \in \\{0,1\\}$ the classifiers $\hat{Y} = f(X)$ and $\hat{A} = h(X)$, with errors $U_i = P(\hat A = i, A \neq i)$, rates $r_{i,j} = P(\hat...	null	null	Rebuttal 1: Rebuttal: We would like to thank all the reviewers for the insightful questions about our work. All the reviewers had questions about and wanted to see a verification of Assumption 1. For clarity, we re-write Assumption 1 below: For $i,j \in \\{0,1\\}$ the classifiers $\widehat{Y} = f(X)$ and $\widehat{A}...	NeurIPS_2023_submissions_huggingface	2,023	null	null	null	null	null	null	null	null
Likelihood-Based Diffusion Language Models	Accept (poster)	Summary: This paper is concentrated on training a Diffusion LM. The authors chose Variational Lower Bound loss used in Diffusion LM as a core of the proposed framework named Plaid, as well as some improvements on the training procedure and architecture of a model (usage of categorical reparametrization, use of trained ...	Rebuttal 1: Rebuttal: We thank you for a very thorough review. We’d like to respond to some of your points below: > For the autoregressive model, it is convenient to expect that lower NLL leads to better samples, but is there such a convenience for Diffusion LMs [...]? I strongly believe that not. E.g., if one drama...	Summary: This paper formalises variation deffusion models for text and makes several algorithmic contributions to such models. They show that these models are able to model the likelihood of text well, and do several other evaluations to validate the likelihood approach, quantify the computational requirements, perfor...	Rebuttal 1: Rebuttal: Thank you for the thoughtful review. > [Appendices] Our appendices are included in the supplementary material that is uploaded. > The model specifications in sections 3.3 - 3.5 are very terse; perhaps more citations or a longer explanation in an appendix would be helpful to many readers. Secti...	Summary: This paper aims to close the gap between autoregressive and diffusion-based language models on standard perplexity-based language modeling benchmarks. To achieve this goal, the paper proposes several algorithmic improvements for maximum-likelihood training, and studies the scaling laws of the diffusion models ...	Rebuttal 1: Rebuttal: Thank you for your review! We'd like to respond to a few points below: > The basic algorithm is mostly following VDM (Kingma et, al.), and most of the proposed improvements are also seen from existing literature. Therefore, the novelty from the algorithmic side is somewhat limited. We are the fi...	Summary: This paper propose a variational diffusion language model which is based on variational diffusion model. The work is related on some works such as combineing vae with language model. Strengths: The application of variational diffussion model to language model is make sense at the some times, and the experimen...	Rebuttal 1: Rebuttal: Thank you for your review. We would like to clarify some points in our work, as the review seems to have some misconceptions. - our work does use a variational characterization of diffusion models, but has very little in common with VAEs. - our work does not propose an autoregressive language ...	null	NeurIPS_2023_submissions_huggingface	2,023	Summary: The paper explores the likelihood-based training of diffusion language models as an alternative to autoregressive models like GPT-2. Several algorithmic improvements are proposed for maximum likelihood training of diffusion LMs, including learned noise schedule, learned embeddings, categorical reparameterizati...	Rebuttal 1: Rebuttal: Thank you for your helpful comments. We give some detailed comments on your questions below. > [Sample evaluation of Plaid 1B] We have performed a study on Amazon Mechanical Turk, comparing samples from Plaid-1B to samples from the GPT2 124M model. We generate unconditional samples of length 128...	Summary: This research studies diffusion-inspired language models with a primary focus on narrowing the perplexity gap between autoregressive and diffusion-based LMs. To achieve this goal, the paper systematically explores various design choices within diffusion-based language models, addressing questions such as the b...	Rebuttal 1: Rebuttal: Thank you for your careful and thorough review! We’d like to respond to some of your points below: > [Limited Novelty] We are the first to find an algorithmic recipe for compute-efficient likelihood-based text diffusions. As you correctly point out, this recipe leverages lots of prior work, but...	null	null	null	null
Multinomial Logistic Regression: Asymptotic Normality on Null Covariates in High-Dimensions	Accept (poster)	Summary: The paper aims to study the statistics of the multinomial logistic MLE on null covariates in a multiclass classification setting in the asymptotic proportional regime, i.e., in the limit in which the size of the dataset $n$ and of the dimension of the covariates $p$ are both sent to infinity but with finite ra...	Rebuttal 1: Rebuttal: Thank you for the insightful and encouraging feedback. Here, we respond to your comments point by point. > Q1: Out of curiosity, the authors state that they expect their results to hold for covariates with distribution having "sufficiently light" tails. Could they comment on this? Do they exp...	Summary: This paper studies the asymptotic distribution of multinomial logistic regression in the high dimensions. Under the assumption of a Gaussian design (and a few other assumptions), it establishes and characterizes the asymptotic normality of the coefficient of a null feature. This extends previous results for hi...	Rebuttal 1: Rebuttal: Thank you for your insightful feedback to help us improve our paper. Here, we respond to your comments point by point. > Q1: Per my first comment on the "Weaknesses", can authors comment on the extent to which can be weakened? A1: The question regarding the possibility to weaken $H_0$ to...	Summary: In this paper, the authors provide asymptotic characterization of the behavior of the maximum likelihood estimator (MLE) of multinomial logistic model (with more than two classes), in the high-dimensional regime where the dimension and the sample size of data go to infinity at the same rate. Under some techni...	Rebuttal 1: Rebuttal: Thank you for your valuable and encouraging feedback to help us improve our paper. Here we provide answers to your concerns point by point. > Q1: While I am almost fine with Assumption 2.1 and 2.2 (just being curious, is some upper bound on the spectral norm of the covariance $\Sigma$ needed?...	Summary: This paper studies the asymptotic distribution of the MLE of the multinomial logistic regression model when the sample size and the number of parameters are of the same order. The validity of the asymptotic theories is evaluated through extensive simulation studies. The paper is overall very well written and ...	Rebuttal 1: Rebuttal: Thank you for the insightful and positive feedback. > Q1: Related to my question in the weakness section, one could use the proposed tests to exclude some noise covariates in the multinomial regression and see how much improvements can be achieved in terms of classification errors. Preferabl...	null	NeurIPS_2023_submissions_huggingface	2,023	null	null	null	null	null	null	null	null
MathNAS: If Blocks Have a Role in Mathematical Architecture Design	Accept (poster)	Summary: The paper introduces MathNAS, a novel approach to Neural Architecture Search (NAS) that utilizes mathematical programming techniques. MathNAS aims to improve the efficiency and accuracy of architecture search by dividing the search space into smaller building blocks and predicting network performance based on ...	Rebuttal 1: Rebuttal: Thank you for your thoughtful review of our work. --- # Clarification about Section 2.2 ## Clarifications on Definition and Calculation of Equation 3 The answers here correspond to parts a, b, and c of the original question. Both theoretically and empirically, we have shown that the effects of...	Summary: The paper proposes a way to perform neural architecture search (NAS) by solving an integer program to identify the constituents of a block-modular neural net, having pre-computed estimations of performance for the separate blocks by a novel method. The resulting "MathNAS" approach is demonstrated in various ex...	Rebuttal 1: Rebuttal: Thank you for your thoughtful review of our work. --- # Polynomial-time Search Complexity As you surmised, the polynomial time referenced here does not pertain to the duration required to solve the ILP problem. Instead, **it denotes the time complexity of the MathNAS algorithm in identifying the...	Summary: This paper introduces MathNAS, a mathematical programming NAS algorithm. MathNAS maps network performance to block performance, enabling the prediction of accuracy, latency, and energy consumption of large networks based on their constituent modules. This algorithm reduces the search space from n^m to m*n, all...	Rebuttal 1: Rebuttal: Thank you for your thoughtful review of our work. Here we respond point-to-point to your questions. --- # Search Time We explain the search times mentioned in the paper in the common response. To ensure a fair comparison, as you rightly pointed out, we will include the training time for the M...	Summary: This paper introduces MathNAS, a blockwise NAS framework. It begins by estimating the performance of the building blocks within the search space and then predicts the overall performance of a network based on the performance of its individual building blocks. This approach effectively reduces the complexity of...	Rebuttal 1: Rebuttal: Thank you for your thoughtful review of our work. --- # Originality and Contribution #### Regarding the comparison of the effectiveness of MathNAS and blockwise methods, we responded in the common response. At the same time, we supplemented the experimental comparison between MathNAS and the b...	Rebuttal 1: Rebuttal: Thank you all for your thoughtful reviews. Here, we respond to some common concerns across all reviewers. --- # Paper Citations, Novelty, and Supplementary Experiments Existing blockwise methods such as DNA [1], DONNA [2], and LANA [3] use block distillation techniques to block the teacher model,...	NeurIPS_2023_submissions_huggingface	2,023	Summary: This paper presents a divide-and-conquer neural architecture search methodology that estimates DNN performance based on the individual performance of each block within a DNN. Accuracy is estimated based on the perturbed accuracy of replacing a block within the NN, and latency/energy are also used within a line...	Rebuttal 1: Rebuttal: Thank you for your thoughtful review of our work. --- # Citations We have meticulously reviewed the mentioned works and discussed the distinctions between MathNAS and them in the common response. Furthermore, we compared the accuracy prediction performance of MathNAS with these works. The expe...	null	null	null	null	null	null
Large Language Models of Code Fail at Completing Code with Potential Bugs	Accept (poster)	Summary: In general, the code completion task that is commonly used to evaluate LLMs assume a clean, non-buggy starting point. However, in real-world software development, this assumption is not always valid. In this papers, the authors investigate code completion given a buggy prefix. For this, they derive two benchma...	Rebuttal 1: Rebuttal: We thank the Reviewer for the detailed review and constructive suggestions. We are encouraged that the Reviewer finds our task interesting and important with our studies and analysis providing interesting and cool findings. We appreciate your acknowledgment of the thorough evaluation of our insig...	Summary: The paper studies the ability of large language models for code to handle prompts which contain “buggy” code. To this end, it introduces two datasets: (i) buggy-HumanEval, which is constructed by injecting synthetic wrong binary operator bugs into HumanEval problems, and (ii) buggy-FixEval, which is constructe...	Rebuttal 1: Rebuttal: We thank Reviewer Cn5Q for the detailed review and constructive suggestions. We are encouraged that the Reviewer find our paper well-written and easy to follow with illustrated examples. We appreciate your acknowledgment of the thorough evaluation of our insights and the contribution of new bench...	Summary: This paper identifies and analyses a limitation of the LLMs for code, that they fail for completing code with potential bugs. This is a problem because LLMs are deployed in real world software projects which may contain many bugs. The authors attempt to bridge this gap between test environment and real world b...	Rebuttal 1: Rebuttal: We thank the Reviewer for the detailed review and constructive suggestions. We are encouraged that the Reviewer finds our problem interesting and practical, and our paper is well-written and well-demonstrated. We appreciate your acknowledgment of the contribution of new benchmark datasets and re...	Summary: Understanding the buggy code completion abilities of Large language models (LLMs) is a vital topic considering the increasing trend of adopting LLMs for programming assistance purposes. This paper studies the buggy code completion problem on different LLMs and datasets and takes a step to investigate three pos...	Rebuttal 1: Rebuttal: We thank the Reviewer XA1K for the detailed review and constructive suggestions. We are very excited that you enjoyed reading our work and found our problem timely, vital, and motivating, as well as our manuscript well-written. We also appreciate your acknowledgment of the novelty and contribution...	Rebuttal 1: Rebuttal: We are thrilled that the reviewers appreciate our paper, and we want to extend our gratitude for their insightful feedback and constructive comments. These invaluable suggestions will undoubtedly enhance the quality of our work. First and foremost, it's encouraging to note that the reviewers have ...	NeurIPS_2023_submissions_huggingface	2,023	Summary: This paper introduces a new code-completion problem: potential bugs exist in the completion prefixes. The authors constructed two benchmark datasets, one is created by manually injecting bugs from the HumanEval dataset, and the other is based on actual bugs in the FixEval dataset. Three baseline methods are p...	Rebuttal 1: Rebuttal: We thank Reviewer 2xun for the detailed review and constructive suggestions. We appreciate your acknowledgment of the novelty, challenge, and practical setting of our problem and the contribution of new benchmark datasets, baseline methods, and our comprehensive analysis. Please find our answers t...	null	null	null	null	null	null
Aligning Gradient and Hessian for Neural Signed Distance Function	Accept (poster)	Summary: The authors propose a new smoothing loss to better regularize the estimated Signed Distance Field (SDF) from unoriented point clouds. By aligning the SDF's gradient and its Hessian, the quality of reconstructed surfaces improves significantly. Compared with commonly used Eikonal loss, the proposed loss is more...	Rebuttal 1: Rebuttal: We thank the reviewer for considering our method novel and theoretically reasonable. Q1: Explanation of Sec 3.3 and Eq. 4. We are sorry for the unclear writing. We will further improve the readability in the revision. Recall that one of the basic properties of SDF is the Eikonal condition (l...	Summary: The paper under review proposes a method for signed distance function (SDF) reconstruction from point cloud data without the use of normal data at the points. The paper proposes a new term as a regularization to complement the EIkonal loss, which is a necessary condition for an SDF. The purpose of this regul...	Rebuttal 1: Rebuttal: Thanks for the valuable comments. We will incorporate all feedback into the revised version. Q1: Different metric compared to DiGS To our knowledge, there are two commonly used evaluation settings under SRB, one used by DiGS [CVPR 2022] and the other proposed by Shape as Points [NIPS 2021]. ...	Summary: The authors provide a novel regularizing term for training of implicit neural SDF representations, motivating the function gradient to be a zero eigenvector of the Hessian near the zero level set (the surface). They demonstrate that it can be used to reduce the appearance of ghost geometry and increase quality...	Rebuttal 1: Rebuttal: We appreciate that the reviewer finds our paper novel. We address additional comments below. Q1: How to handle unoriented point clouds? Yes, our alignment term encourages consistent orientations of normal vectors. Based on our $C^{\infty}$ activation function (sine and softplus), when we req...	Summary: The paper considers the task of surface reconstruction from point clouds without normals and uses neural signed distance functions. They consider previous smoothness losses, and come up with their own by taking the derivative of the Eikonal equation, thus guaranteeing that their loss constrains a fundamental p...	Rebuttal 1: Rebuttal: We thank the reviewer for constructive feedback. In the following, we will address the main concerns carefully and seriously. Q1: Compute IoU under ShapeNet Following your advice, we report the IoU statistics under ShapeNet, see Tab. 1 in the global response box. At the same time, we wil...	Rebuttal 1: Rebuttal: We sincerely appreciate all valuable comments and suggestions, which helped us to improve the quality of the article. We have carefully considered all your advice and addressed the questions raised, see details in the separate response and the attached PDF files. Specifically: - We clarify the mo...	NeurIPS_2023_submissions_huggingface	2,023	Summary: This paper introduces a novel approach to learning the SDF directly from point clouds without the use of surface normals. The key insight behind this is that aligning the gradient of the SDF with its Hessian allows for better training of the distance field. Through extensive experiments, the paper demonstrates...	Rebuttal 1: Rebuttal: We sincerely thank the reviewers for their appreciation and invaluable comments. We will incorporate all feedback into the revised version. Q1: Limitations and negative social impact We shall point out the following limitations in this revision: (1) it is weak in handling sparse input such a...	null	null	null	null	null	null
Private (Stochastic) Non-Convex Optimization Revisited: Second-Order Stationary Points and Excess Risks	Accept (spotlight)	Summary: This paper provides a private algorithm for (Stochastic) Non-Convex Optimization using gradients, differences between two successive gradients, and the exponential mechanism for privacy (added to the gradient). It leverages the SpiderBoost algorithm and its private version to find (privately) stationary points...	Rebuttal 1: Rebuttal: Thanks for your time reviewing and commenting on our paper. Thanks for pointing out the typo. It should be $\mathrm{drift}_{t+1}=\mathrm{drift}_t+ \eta^2 \|\|\nabla_t\|\|^2$ in line 14. Please let us know if you have further questions or concerns. --- Rebuttal Comment 1.1: Title: Respond to authors...	Summary: This paper addresses the problem of private stochastic optimization for general non-convex functions. For smoothed functions, the authors propose a combination of the Gaussian mechanism and the variance reduction technique (spider algorithm) to find the SOSP of the loss function. Their results for the empirica...	Rebuttal 1: Rebuttal: Thanks for your kind reviews and comments. In our future work, we will both explore the tightness of our bounds, and evaluate our algorithms empirically. --- Rebuttal 2: Title: Respond to authors Comment: Dear reviewer, please read authors' response to your review and reply to them regarding ...	Summary: The paper introduces a non-convex optimization framework that integrates two types of gradient oracle and an application of the exponential mechanism to find the global minima of non-convex objectives, requiring minimal assumptions and showcasing remarkable population risk bound performance. The paper also hig...	Rebuttal 1: Rebuttal: Thank you for your comments. We will add a detailed conclusion section in the paper. In more detail, our conclusion will review the main contributions of the paper and then propose some future directions which we feel are important/interesting, including: - Can we close the gap between our upper a...	Summary: This paper studies differentially private second-order optimization problems, where the loss function is assumed to be second-order smooth. The objective is to identify second-order stationary points (SOSP) while ensuring privacy guarantees. The authors present results for both empirical risk and population ri...	Rebuttal 1: Rebuttal: Thanks for your positive reviews and questions. Following are the responses to the specific questions: Q1. We use Lemma 3.4 for escaping the saddle point directly, which requires us to add “extra” Gaussian noise whenever the gradient norm of the current point is small (meaning it is possibly a sa...	Rebuttal 1: Rebuttal: General Reply: We thank all the reviewers for their reviews and comments. We will diligently address all the presentation issues raised by the reviewers and enhance the writing quality in future versions, like providing more explanations on the intuitions. In the following, we respond to the s...	NeurIPS_2023_submissions_huggingface	2,023	Summary: The paper studies the problem of non-convex optimization under differential privacy constraints. The goal of the paper is to find first and second order stationary points of the function. The proposed algorithm which is inspired by SpiderBoost [1] is a sampling based method that generates a private estimator t...	Rebuttal 1: Rebuttal: Thanks for your valuable reviews. We use $P^*$ as the input dataset can be of arbitrary size. For example, we consider a set D of size n in the paper, meaning $D \in P^n$. Q1: We discuss our techniques and the challenge of previous techniques in Subsection 1.2 (our techniques) and related work (...	null	null	null	null	null	null
Hypernetwork-based Meta-Learning for Low-Rank Physics-Informed Neural Networks	Accept (spotlight)	Summary: The paper presents a novel low-rank PINN architecture to solve parametrized partial differential equations (PDEs). Based on a hyper network framework they compute solutions to PDEs in a rank-adaptive way. Their approach is able to overcome for certain test systems known failure modes of PINNs. The method is ex...	Rebuttal 1: Rebuttal: Q1: In my opinion the paper would benefit from a more prominent motivation, why the proposed approach is more robust against failure modes. Is it connect to the motivation in appendix D.2 and the remark in appendix B? A1: Your understanding of our motivation is correct; we were motivated ...	Summary: In this paper, a meta-learning method is proposed for learning parameter-dependent dynamical systems. In particular, similarly to the reduced order method, the proposed method uses a low-rank neural network to construct a model, and the parameter-dependent coefficients are learned separately by a neural networ...	Rebuttal 1: Rebuttal: **Q1: Although it may not be a weakness, it is difficult to understand theoretically that learning singular vectors does not improve performance when the matrix is decomposed into lower ranks. Based on the idea of model order reduction, the singular vectors should correspond to the important modes...	Summary: This work aims to improve on the efficiency and performance of adapting physics-informed neural networks (PINN) when dealing in many-query settings, given that the present field is limited in computationally-demanding retraining on new PDE parameters. It specifically focuses on learning low-rank versions of PI...	Rebuttal 1: Rebuttal: **Q1: The lack of explanation on the meta-learning inclusion to the framework in the main paper is a bit confusing if a reader isn't coming from a meta-PINN perspective as there are no definitions given in the standard form of meta-train and meta-test environments or details on the context/query s...	Summary: The paper proposes an approach for solving PDEs by using a low-rank parameterized NN. While traditional PINNs are restricted to a single PDE instance, the authors use a hypernetwork to be able to solve PDEs with different parameters. The hypernetwork predicts the weights matrices of a PINN network in a low-ran...	Rebuttal 1: Rebuttal: Q1: “low-rank PINNs containing only hundreds of model parameters” (l. 10): the parameter comparison seems unfair, since only the parameters learned during phase 2 are counted (not hypernetwork or basis vectors parameters) A1: Thank you for the comment. We will make the statement more prec...	Rebuttal 1: Rebuttal: Global comment #1: [Additional results of computational cost] Assuming that we solve PDEs for the same number of PDE parameters without our method and other baselines, the computation needed for each forward pass in Phase 1 is slightly higher than that of training all individual vanilla PINNs...	NeurIPS_2023_submissions_huggingface	2,023	null	null	null	null	null	null	null	null
Characterizing Out-of-Distribution Error via Optimal Transport	Accept (poster)	Summary: In this work, the authors aim to predict model’s performance on out-of-distribution (OOD) data without access to labels during testing. Specifically, they identify pseudo-label shift, which gauges the difference between the predicted and the true OOD label distributions, to indicate the under-estimation issue...	Rebuttal 1: Rebuttal: We thank the reviewer for their valuable review of our work! In particular, we are excited that our motivation reads clear to the reviewer and our code submission alongside this contribution improves the reproducibility of our method. ### On Weakness 1 We thank the reviewer for pointing out the...	Summary: Authors develop a methodology for estimating the accuracy of a classifier in the case where the distribution of the test set is different from the one used in the training data. Their methodology is based on computing a certain Wasserstein distance between suitable distributions on labels. Authors argue this d...	Rebuttal 1: Rebuttal: We thank the reviewer for their thorough review and are encouraged to hear that our work was perceived as a good and strong contribution, providing a highly innovative perspective on the problem of OOD detection. Further, we appreciate the reviewer’s comment that our experiments are compelling in ...	Summary: This paper proposed Confidence Optimal Transport (COT) and a variant, Confidence Optimal Transport with Thresholding (COTT) based on pseudo-label shift to characterize the error between predicted and true OOD label distribution. Strengths: 1. The idea of using pseudo-label shift to characterize error is natu...	Rebuttal 1: Rebuttal: We thank the reviewer for the thoughtful feedback and for the recognition that our idea is natural and straightforward, and our experiment results are reasonable. We address concerns and questions below. ### On Weakness 1 Previous works have established that no theoretical guarantee is possible...	Summary: The paper proposes to compute OOD error estimates using a Wassertein distance between the target label distribution and the distribution of confidence vectors. This is in contrast to AC-MC which compare the pseudo-label distribution and the confidence distribution. The paper motivates COT (confidence optimal t...	Rebuttal 1: Rebuttal: We thank the reviewer for their thoughtful feedback and for the compliment that the experiments demonstrate clear advantages of our method, the presentation of the paper is clear, and we tackle an important problem of estimating OOD error without a complicated solution. We would like to address yo...	Rebuttal 1: Rebuttal: We would like to thank all reviewers for their time and effort in reviewing our work. We are glad that our work receives the following compliments from our reviewers: - It shows “clear advantage of COT and COTT for predicting OOD error” (1Jqy) over “most other metrics that are sensitive to miscali...	NeurIPS_2023_submissions_huggingface	2,023	Summary: The manuscript studies the challenges caused by out-of-distribution (OOD) data in machine learning models. They identify "the difference between the predicted and true OOD label distributions" (called pseudo-label shift) as the main reason many existing works underestimate OOD errors. They propose a new approa...	Rebuttal 1: Rebuttal: We thank the reviewer for the thoughtful review and for the recognition that our methods “have a significant advantage” over other metrics, our experiments are extensive, and we provide performance guarantees. We would like to address the concerns and questions raised by the reviewer as follows: ...	null	null	null	null	null	null
GMSF: Global Matching Scene Flow	Accept (poster)	Summary: Previous scene flow estimation methods require complicated coarse-to-fine or recurrent architectures as a multi-stage refinement. In contrast, this paper proposes a simpler single-scale one-shot global matching to address the problem. To this end, this paper decomposes the feature extraction step via a hybrid ...	Rebuttal 1: Rebuttal: We thank the reviewer for their positive comments and helpful suggestions. Below are our clarifications for the questions. Q1: Related work. We thank the reviewers for the suggestion of the unmentioned related work. We will add the following discussion in the related work. Different from th...	Summary: The paper presents GMSF, a transformer-based method that matches dense features to estimate the scene flow from point clouds. The proposed method uses a transformer-based architecture that matches two point clouds and calculates the scene flow leveraging the cross- and self-attention modules. The paper present...	Rebuttal 1: Rebuttal: We thank the reviewer for their positive comments and helpful suggestions. Below are our clarifications for the questions. Q1a: Ablation study of the KNN component. For the Point Transformer in Figure 2, we follow the setting in Zhao $et al.$ [1] and set $k$ to 16. Thanks to the reviewer's s...	Summary: This work aims to address the task of scene flow estimation for 3D point clouds. The authors propose a hybrid architecture based on local-global-cross transformers. Given as input a source and a target point cloud, first, the local transformer extracts geometric features within a patch, then the global transfo...	Rebuttal 1: Rebuttal: We thank the reviewer for their positive comments and helpful suggestions. Below are our clarifications for the questions. Q1a: Limited technical contributions. The transformer architecture and probabilistic matching approach have been studied and proven efficient previously in many NLP task...	Summary: This paper proposes a hybrid local-global-cross transformer scene flow estimation model, achieving the state-of-the-art results on FlyingThings3D and KITTI Scene Flow datasets. Strengths: 1. This work has a mature model design based on Transformers, achieving superior results on the Flyingthings3D and KITTI S...	Rebuttal 1: Rebuttal: Thanks for the comments and suggestions. We improve the paper based on the two weaknesses pointed out. We show how we improve the paper below: Q1: Experiments on the Waymo-open dataset. As suggested, we conduct additional experiments on the Waymo-open dataset for scene flow. The dataset cont...	Rebuttal 1: Rebuttal: We thank the reviewers for their valuable feedback and questions. We are happy that the reviewers find our proposed approach, and the motivation "intuitive and novel" (D7Yn). The architecture is "straightforward without bells and whistles" (YAmp), "mature" (QN7Z), and "simple and effective" (vPo9)...	NeurIPS_2023_submissions_huggingface	2,023	Summary: The paper proposes GMSF for scene flow estimation from point clouds. As far as the authors are aware, GMSF is the first to address scene flow estimation with global matching - GMSF is formulated as a single-scale one-shot global matching. GMSF incorporates a novel local-global-cross transformer architecture to...	Rebuttal 1: Rebuttal: We thank the reviewer for their positive comments and helpful suggestions. Below are our clarifications for the questions. Q1a: Ablation study on the number of transformer layers. Please refer to the global response to all reviewers B. **Q1b: Ablation study on decoupling global transformers...	null	null	null	null	null	null
Compositional Foundation Models for Hierarchical Planning	Accept (poster)	Summary: The paper presents Hierarchical Planning with Foundation Models (HiP), a framework that addresses long-horizon decision-making in novel environments, using multiple modalities of data and reasoning at various levels of hierarchy. The key elements of this framework involve the use of a large language model for ...	Rebuttal 1: Rebuttal: We thank reviewer zvNJ for their constructive feedback. We now answer the following concerns raised in the review. >The baselines compared are all relatively weak. It's more appropriate to compare with foundation robotics models, such as SayCan, Gato[1], palm-e[2]. The authors argue that compared...	Summary: This work presents a new approach of combining foundation models from different modalities to into hierarchical system to solve long-horizon robotics problem. The performance of the system is evaluated on two long-horizon tasks and largely out-performed other baselines. Strengths: The approach of this work i...	Rebuttal 1: Rebuttal: We thank reviewer z2ah for evaluating our work. We now answer the following concerns raised in the review. >What does dataset stats look like? We described our evaluation environments in Section 3.1 of our original submission and provided further details about our datasets for task planning, vis...	Summary: The paper introduces a framework called Hierarchical Planning with Foundation Models (HiP), which is designed to improve decision-making in new environments with long-term goals. HiP uses hierarchical reasoning to plan subgoals, visually reason about plans, and execute actions through visual-motor control. The...	Rebuttal 1: Rebuttal: We thank reviewer 3JTC for their constructive feedback. We now answer the following concerns raised in the review. >Novelty Prior works using foundation models for robotics typically learn a single large model on Internet/robotics data. However, LLMs cannot construct visually grounded plans and ...	Summary: This paper presents a way to leverage large foundation models to perform long-horizon hierarchical planning tasks. Specifically, the authors leverage an LLM for subgoal generation from the text instruction, a video generation model for generating a plan for each subgoal, and an action prediction model that out...	Rebuttal 1: Rebuttal: We thank reviewer 7a1H for their constructive feedback. We now answer the following concerns raised in the review. >unclear if the some approximations fully capture the target distribution. For example, I'm not sure if Eq. (3) is a good approximation of (2) since Eq. (3) doesn't have the reachabi...	Rebuttal 1: Rebuttal: We thank the reviewers for their thoughtful suggestions. We want to start by addressing the common concerns brought up by the reviewers and dive into the remaining points in the individual responses. ### Evaluation on more realistic robotic domain: We evaluate HiP on an additional third dom...	NeurIPS_2023_submissions_huggingface	2,023	null	null	null	null	null	null	null	null
Equivariant Single View Pose Prediction Via Induced and Restriction Representations	Accept (poster)	Summary: This paper shows that algorithms which learn 3D representations from 2D images must satisfy certain consistency properties, which are equivalent to SO(2)-steerable constraints. To this end, a differentiable induction layer is proposed to map signals on the plane into signals on the sphere. The method is evalua...	Rebuttal 1: Rebuttal: We thank the reviewer for the thorough feedback. We address the questions posed by the reviewer below. $\textcolor{red}{ \text{Weaknesses: } }$ **The implementation part is too short. It is unclear how the implementation details reflect the theoretical parts. Could the authors elaborate the c...	Summary: This is a theory paper. The paper presents the general form of H equivariant linear maps that lift the signals defined on R2 to signals defined on S2, which is defined by convolution with kernel constraints. The main paper focuses on the case where H=SO(2) and G=SO(3). Under such a setting, the authors evaluat...	Rebuttal 1: Rebuttal: $\textcolor{red}{\text{Weaknesses}}$ **One main limitation of the proposed theory is that the authors only extensively demo the case of SO(2)—SO(3), but lack more general application settings on other groups, for example, at least SE(2) and SE(3), which has wider application settings for vision. ...	Summary: This work tackles the challenging problem of achieving SO(3) equivariance over 2D projected images for pose estimation tasks. They propose a unified 2-step framework by adding an induced layer to turn SO(2)-equivariant representations into spherical signal, then using SO(3)-equivariant convolution to get the f...	Rebuttal 1: Rebuttal: We thank the reviewer for an especially detailed and thorough review. We provide a point-by-point response to the comments and questions below: $ \textcolor{red}{ \text{ Weaknesses: }} $ **The claimed major contribution of induced layer seems to mainly come from this Image to Sphere(Klee et al....	Summary: The paper studies learning equivariant SO3 representations from 2D images. The authors first propose a generalized theory that shows SO2 equivariance constraint for image to spherical signals. Then they propose a construction algorithm called induction layer to implement equivariance. The proposed method outpe...	Rebuttal 1: Rebuttal: Thank you for the feedback. Regarding your questions about our work, here is our response: $\textcolor{red}{ \text{Weaknesses: }} $ It is not clear how to construct the induction layer from the paper. More details may be provided on how to implement the layer, e.g. how to solve equation 2 ...	Rebuttal 1: Rebuttal: We thank the reviewers for their comments on our manuscript. We are pleased that the majority of reviewers recognize the fact that our theory provides a unifying perspective for a set of relevant computer vision tasks: - Reviewer Nd1u: "the paper proposes a unifying theory that encompases previous...	NeurIPS_2023_submissions_huggingface	2,023	Summary: The paper focuses on the task of 3d pose prediction. The paper proposes a set of consistency constraints for learning 3d representation and shows that few of the previously proposed neural architectures follow these constraints. Further using their constraints they propose a new architecture with an induction ...	Rebuttal 1: Rebuttal: Thank you very much for the useful feedback. Please find our point-by-point response below. $\textcolor{red}{ \text{ Weaknesses: }} $ I couldn't find any ablations/analysis in the paper We have added two ablations studies. The first study compares the performance an architecture that use...	null	null	null	null	null	null
Few-shot Generation via Recalling Brain-Inspired Episodic-Semantic Memory	Accept (poster)	Summary: In this article the authors propose to augment state of the art few-shot generative models with a memory module inspired by the brain. The memory module combines episodic and semantic memory. The authors demonstrate through extensive experimentations the benefits of their module on few-shot generation: when ta...	Rebuttal 1: Rebuttal: Thanks for your constructive comments and suggestions, which will be helpful for us to further improve the quality of our paper. Following your comments, we have carefully fixed these mentioned typos and also rewritten several parts of our paper for improving the experience of reading. Q1: Concer...	Summary: This paper introduces a novel approach to few-shot generation inspired by the human memory mechanism. The authors propose a Variational Structured Memory module (VSM) to store and recall episodic and semantic memories. They also introduce a bionic memory updating strategy to model the conversion between these ...	Rebuttal 1: Rebuttal: We appreciate your constructive comments and suggestions, which are helpful for us to improve the quality of our paper further. The concerns have been addressed below. W1: Concern about Computation Efficiency A1: Thanks, we will answer this question in our response to Q3. W2: Acronym Overlap: ...	Summary: The paper proposes a variational structured memory module (VSM) that can be combined with both VAE and Diffusion type models to improve their performance on few-shot generation tasks. The memory structure is both hierarchical and structured into episodic and semantic memory components such that the model can l...	Rebuttal 1: Rebuttal: Thanks for your constructive comments and suggestions, which are helpful for us to further improve the quality of our paper. The concerns have been addressed as below. W1: Concern on computational overhead of applying VSM. A1: Thanks for your constructive suggestion. We have provided th...	Summary: This work is focused on the problem of few-shot generation using a memory mechanism. A few priors works have leveraged semantic memory or short-term (episodic) memory for such tasks. In this work, a new architecture is proposed named Variational Structured Memory (VSM) that leverages both types of memory mecha...	Rebuttal 1: Rebuttal: We appreciate your constructive comments and suggestions, which are helpful for us to further improve the quality of our paper. The concerns have been addressed as below. W1: Some design choices of the algorithm can be explained better. A1: Thanks, we will explain your mentioned question...	Rebuttal 1: Rebuttal: We really appreciate all the reviewers for their constructive and helpful comments, which can greatly help us to improve the quality of our paper. For the mentioned typos, grammar mistakes, unclear notations and missing citations, we promise that we will carefully fix them in our revised paper. F...	NeurIPS_2023_submissions_huggingface	2,023	Summary: The authors propose to incorporate 'episodic' and 'semantic' memory inspired by the human memory system into existing generative modeling frameworks to improve few-shot generation abilities. Specifically, they propose a variational structured memory module (VSM) and show its incorporation into existing VAE bas...	Rebuttal 1: Rebuttal: Thanks for your constructive comments and suggestions. Q1. VSM for static image few-shot generation. A1: Thanks for your valuable question. Actually, the motivation of our work is to develop a series of memory-augmented generation models to imitate the creative process of humans. For instance,...	null	null	null	null	null	null
Deep Patch Visual Odometry	Accept (poster)	Summary: This paper proposes a deep learning-based method for monocular visual odometry, which equips two leading-edge advantages: (i) a deep feature-based patch representation for keypoints encoding local context, and (ii) a novel recurrent architecture designed for patches along with a differentiable bundle adjustmen...	Rebuttal 1: Rebuttal: The paper makes claims that the proposed method outperforms all prior work in several evaluation datasets, but only several works are listed which is not enough, can authors give more comparisons with SOTA? For example, the DytanVO[1], RAM-VO[2], DF-VO[3] Our claim was intended to mean that w...	Summary: The author propose a learning-based approach for visual odometry. It is essentially a sparse version of DROID-SLAM, without the loop closure capability. Instead of predicting the dense flow field relating the different frames as in DROID-SLAM, DVPO operates on sparse image patches, which are extracted at rando...	Rebuttal 1: Rebuttal: **In DROID-SLAM, predicting the dense flow allowed to optimize for full dense depths along with the camera poses with accuracy, providing advantages over both direct and indirect approaches. However, the proposed approach DVPO is based on jointly estimating the motion of sparse patches and refinin...	Summary: In this paper, the author propose a new method to solve monocular visual odometry (VO) in and end-to-end fashion. In contrast to previous work - DROID-SLAM, which utilizes dense optical to build correspondences across frames, the proposed method leverages sparse patches to avoid redundancy of dense pixels and ...	Rebuttal 1: Rebuttal: **Limited novelty. The key idea of the paper is relatively straightforward. If we review the classic monocular visual VO/SLAM systems, we can find most of them are based on sparse keypoints (ORB-SLAM) or semi-dense pixels (DSO) to avoid too much computation. The recurrent module designed to update...	Summary: This paper proposes a deep learning based visual odometry (VO) system which estimates the 6DoF poses of each frame and outputs a sparse reconstruction of the scene from an input monocular video sequence. The pipeline extracts deep features from the sampled image patches and updates deep optical flow of the pat...	Rebuttal 1: Rebuttal: The novelty is limited, the iterative update mechanism and differential bundle adjustments are already used in DROID-SLAM, even with the introduction of 1D Temporal Convolution and SoftMax Aggregation modules in the update operator. We do not claim the idea of an iterative update mechanism no...	Rebuttal 1: Rebuttal: Message to all reviewers regarding the novelty of DPVO: Our novelty is sparse matches integrated with neural iterative updates and differentiable bundle adjustment, which has not been done before and is nontrivial to design. While sparse matches have been used in prior work, they were not i...	NeurIPS_2023_submissions_huggingface	2,023	null	null	null	null	null	null	null	null
SmoothHess: ReLU Network Feature Interactions via Stein's Lemma	Accept (poster)	Summary: For smooth functions the joint interaction between pairs of features and their impact on the function output can be modeled and quantified via second order partial derivatives. The Hessian of a neural network can therefore provide insight as to how interactions between pairs of features impact predictions made...	Rebuttal 1: Rebuttal: Thank you for taking the time to review our work and for the positive assessment. Below, we respond to the weaknesses and questions brought up. --- ## Weaknesses W1: Choosing the covariance As you mention, SmoothHess gives the user more direct control over smoothing compared to the compe...	Summary: The goal of this paper is to model the point-wise interactions between input features in a ReLU network. ReLU networks are piece-wise linear, which inhibits using Hessian to model feature interaction (since it’s zero almost everywhere). This paper proposes SmoothHess, a new method to compute feature interactio...	Rebuttal 1: Rebuttal: We appreciate your enthusiasm about, and careful reading of, our work. We are especially glad to note your appreciation of the connection between SmoothHess and SmoothGrad. Below, we respond to the weaknesses and questions you have brought up. --- ## Weaknesses **W1: Generalization to larger m...	Summary: This paper introduces SmoothHess, a method to define a smooth approximate Hessian for networks where many components of the Hessian may be zero (e.g. ReLU networks). This can aid in tasks where interactions between features are needed, but where those components of the exact Hessian are zero. SmoothHess is def...	Rebuttal 1: Rebuttal: Thank you for taking the time to review our work. Below we respond to the weaknesses and questions brought up. --- ## Weaknesses W1: Impactfulness and non-zero Hessian We kindly wish to clarify any misunderstandings about the Hessian we are computing and its use in our work. We are inter...	Summary: The paper propose to compute Hessian of a ReLU neural network by a convolution of the network function with a Gaussian distribution function. It's equivalent (or cover) some prior work, e.g., SmoothGrad. The method is generally easy to understand, without modifying the network structure. A non-asymptotic bound...	Rebuttal 1: Rebuttal: Thank you for taking the time to review our work. Below we respond to the weaknesses and questions brought up. --- ## Weaknesses W1: Limitation of scope to ReLU We appreciate the comment as we feel this is an important point to clarify in our work. Our method is not limited only to ReLU n...	Rebuttal 1: Rebuttal: We would like to thank the reviewers for their thoughtful questions and comments, which have helped us to improve our work. --- ## Recurring Questions Below we respond to recurring questions among the reviewers. We also give individual responses to all the reviewer questions under each revi...	NeurIPS_2023_submissions_huggingface	2,023	Summary: The work proposes an approach for measuring feature interactions for neural networks that have no useful second order gradients or Hessians (notably those employing ReLU). The approach convolves a smooth normal with the network as a proxy with usable Hessians and then offers a means of estimating them using sa...	Rebuttal 1: Rebuttal: Thank you for reviewing our work. We've addressed your main concerns in our official rebuttal. Other concerns have been addressed in a comment due to the character limit. --- W1 Experimentation Focus We emphasize that, while we agree as to the potential for use in optimizations, Smoo...	Summary: The author introduces a novel method to approximate the Hessian of a smoothed loss function at a specific point. By incorporating a variation of Stein’s lemma, they cleverly estimate the gradient over a smoothed surrogate of a ReLU network at that point. To gauge the method’s effectiveness, they measure its pe...	Rebuttal 1: Rebuttal: We appreciate your positive assessment of our work. Below, we address the weaknesses and questions that you have brought up. ## Weaknesses --- W1: Regarding the statement of Theorem 3. Thank you for your careful reading of the statement of Theorem 3. We agree with the reviewer’s suggest...	null	null	null	null
Generalization in Neural Operator: Irregular Domains, Orthogonal Basis, and Super-Resolution	Reject	Summary: This paper studies the generalization error of neural operators that contain kernel operations. Under the basic setting of neural operators (such as FNO), this paper establishes upper bounds of the excess risk of neural operators. How to apply NO to irregular domains, and the error analysis of super resolution...	Rebuttal 1: Rebuttal: Thanks for your valuable comments. Questions are addressed as below. >Q1: Novelty? A1: See general response. >Q2: Theorems are not impressive and are unclear... A2: Network parameter norm larger than 1. First, it aligns with the effectiveness of our commonly used L2 regularization....	Summary: A theoretical analysis of neural operators (NOs) is presented, which provides further insight into the construction and performance of NOs. The theoretical insights are validated by numerical experiments. Strengths: A thorough theoretical analysis of NOs is presented, with significant improvements in the tig...	Rebuttal 1: Rebuttal: We would like to thank the reviewer for the valuable comments. Your questions are addressed as follows. >Q1: The summary of the overall model is not the most clear (Equation 1), which suggests the non-linear activation is applied first, rather than following the kernel and linear transforms. **A...	Summary: The image super-resolution (SR) is a recurrently used task, nowadays, since the SR images can improve the accuracy of downstream tasks like object detection. Many proposals rely on heavy Deep Learning models or lightweight models based on efficiently designed architectures. This work studies neural operators b...	Rebuttal 1: Rebuttal: Thanks for your valuable comments. Your questions are addressed as follows. >Q1: Despite a short evaluation performed that shows a better performance of the proposal, I feel it is not sufficient to validate the generalization ability in the super-resolution task. To this end is necessary to valid...	Summary: In this paper, the authors propose to analyze neural operators from the orthogonal bases in the kernel operators, which helps to guide designing kernel operators and choosing grid points, analyzing generalization and super-resolution capabilities, and adapting neural operators to irregular domains. Strengths...	Rebuttal 1: Rebuttal: Thanks for your valuable comments. Your questions are addressed as follows. >Q1: I feel there are still gaps between the overall claims, the theoretical results in Section 3, and the implication and experiments in Sections 4\&5. For example, the "NOs on Unbounded Domain" section is not what I wou...	Rebuttal 1: Rebuttal: We would like to thank all reviewers for their valuable comments. Here, we present general responses for understanding the manuscript better. Your specific questions are answered respectively. >P1 (reviewer 59nv): On the contribution. We present two primary theoretical contributions, we innovat...	NeurIPS_2023_submissions_huggingface	2,023	Summary: The authors consider a new perspective to neural operators, by examining the role of orthogonal bases. The kernel operators are constructed such that their eigenfunctions are predefined orthogonal bases, with the eingenvalues as trainable parameters. That is, a neural operator can be seen as a mapping from inp...	Rebuttal 1: Rebuttal: Thanks for your valuable comments. Your questions are addressed below. >Q1: Some of the results are fairly obvious. A1: The implications of our theory for practical applications, particularly in grid type selection but not grid size, carry profound significance. Conventionally, generate...	Summary: The paper attempts to provide theoretical analyses of Neural Operators (NOs) mainly considering the following aspects: 1. The generalization bound of NOs; 2. The discretization error of NOs; 3. The "super-resolution" (trained on low-resolution grid then evaluate on the high-resolution grid) error of NOs on the...	Rebuttal 1: Rebuttal: Thanks for your valuable comments. Your questions are addressed as follows. >Q1: Some of the discussion in the experiment section is relatively vague, especially in section 6.3. As revealed by Theorem 3.3, the integration scheme along with basis selection has a crucial effect on the super-resolut...	null	null	null	null
Actively Testing Your Model While It Learns: Realizing Label-Efficient Learning in Practice	Accept (poster)	Summary: The paper introduces the Active Testing while Learning (ATL) framework, which efficiently collects testing samples for online model risk estimation during the active learning process. The online risk estimation enables early termination of active learning. Additionally, ATL establishes a connection between Act...	Rebuttal 1: Rebuttal: Q1: Bias introduced by the feedback mechanism. Thank you very much for raising this insightful point! As we agree with your comment, we here further elaborate on the problem following the answer in Q5. The key objective of feedback is to improve model learning without sacrificing the risk e...	Summary: This paper considers the active testing procedure to assess the risk. The keys are 1) a proposal of a quiz using importance sampling, 2) combining the historical quizzes to make a test set for the final model test, and 3) a proposal of a feedback algorithm to strengthen the model using the labeled test data. E...	Rebuttal 1: Rebuttal: Q1: Explicitly calculated risk and limitation from AL algorithm. We would like to clarify that our design of framework is AL-agnostic, and we do not assume any fixed distribution for AL sampling. We will further discuss the AL strategies in Q2 and Q3 below. As for the test risk specification,...	Summary: The paper provides a relatively new flavor to recent active learning works, by considering active testing. In particular, they evaluate the learned model on the fly on adaptively acquired set, called quizzes. Model learns based on those quizzes and part of it is added, as a feedback from the quiz. At the end, ...	Rebuttal 1: Rebuttal: Q1: AL strategy specification. Please see Q2 in the general response. Q2: Sampling using $q_t$ and distribution estimation. As we explained in Q1 in the general response, the test sampling is done in a sequential way. In each test sampling round, we compute $q({\bf x})$ over the entire ...	Summary: This paper proposes a framework for integrating active learning and active testing in an online fashion. The proposed algorithm incrementally sample training and testing examples for each batch and can leverage all of the testing examples so far for evaluation. Moreover, the sampling distribution for testing i...	Rebuttal 1: Rebuttal: Q1: The unbiased risk estimate. Thank you very much for this insightful question! Please see the answer to Q4 and Q5 in the general response. Adding to Sections 3.2 to 3.4 in the paper, we discussed in detail the improved multi-source intermediate estimator (which impacts the variance-optimal...	Rebuttal 1: Rebuttal: In this general response, we address a set of important questions that commonly occur in multiple reviewers' comments, avoiding repeating the same response in each individual rebuttal. Q1: Clarification of the problem setting. (To all reviewers) In this paper, we aim to construct an inte...	NeurIPS_2023_submissions_huggingface	2,023	null	null	null	null	null	null	null	null
Counterfactual-Augmented Importance Sampling for Semi-Offline Policy Evaluation	Accept (poster)	Summary: The paper proposed a novel method for off-policy evaluation with offline data that entails human input. Specifically, the paper assumes access to human annotations for counterfactual state-action pairs as counterfactual immediate reward (in bandits) or Q values (in RL) under the behaviour policy (as a relaxati...	Rebuttal 1: Rebuttal: We thank Reviewer rPY7 for the time spent reading our paper (including the appendix) and their helpful feedback. We are also encouraged by their overall positive assessment of our paper. Below we address the reviewer's questions and concerns. **1. The proposed method does not address state distr...	Summary: - Propose a setting in OPE where experts can annotate the value of counterfactual actions. - A new importance sampling scheme for the setting is proposed. - Theoretical analysis and experiments confirm superiority over the unweighted case, etc. Strengths: 1. The proposed setting of expert annotating the count...	Rebuttal 1: Rebuttal: Thank you for taking the time to read and evaluate our work. Below, we provide clarifications and answers to your questions. 1. “The method itself and the baselines are fairly naive” (see also: [overall response](/forum?id=dsH244r9fA&noteId=jKEOBV6cyp), [71Lt](/forum?id=dsH244r9fA&noteId=tJF...	Summary: This paper offers a new OPE estimator leveraging additional annotations of counterfactual actions. The paper provided an extremely thorough and comprehensive discussion of the theoretical property of this estimator. This helps us develop an understanding of how it operates. This paper also discussed fully the ...	Rebuttal 1: Rebuttal: We are encouraged that the reviewer appreciated our clear/thorough theoretical analyses, as well as well-documented experiments. We also want to point out that we have provided full code implementation in the supplement. Below, we address the feedback from the reviewer: **1. Regarding human expe...	Summary: The paper considers OPE with annotation to improve the performance of OPE. The paper explains why naively incorporating such information may lead to a biased estimate. And the paper provides a theoretical analysis of the bias and variance. Finally, the paper demonstrates the performance of the proposed method ...	Rebuttal 1: Rebuttal: Thank you for taking the time to read and evaluate our work. Below, we list specific updates on which we plan to focus our efforts, as well as provide answers to the questions raised. 1. “The clarity of the paper needs to improve” Thank you for your detailed suggestions on improving the cla...	Rebuttal 1: Rebuttal: We thank all reviewers for taking time to read our paper and providing valuable feedback. Overall, reviewers found our paper to be well-written (71Lt, noPk, fzzo), addressing an interesting (71Lt), important (Qcjd) and timely problem (rPY7), and proposing a novel solution (rPY7) with comprehensi...	NeurIPS_2023_submissions_huggingface	2,023	Summary: This paper studies offline policy evaluation in reinforcement learning, where the behavior policy, some interaction trajectories, and some human annotated trajectories (modified from interaction trajectories) are available. The paper points out that, since the number of human annotated trajectories are not the...	Rebuttal 1: Rebuttal: We thank the reviewer for their overall positive assessment of our paper and valuable feedback, including providing additional references. We are encouraged that the reviewer found our paper to be well-written and recognized its potential impact. We address the main questions below: **1. How to ...	null	null	null	null	null	null
The Memory-Perturbation Equation: Understanding Model's Sensitivity to Data	Accept (poster)	Summary: This paper proposes measuring sensitivity to data in some problems in machine learning through what they refer to as the “memory perturbation equation” (MPE), derived using the Bayesian learning rule. They argue that sensitivity properties of machine learning models are not as well understood as they could be ...	Rebuttal 1: Rebuttal: >Q1: For the title, I don’t think “model’s” is the most suitable choice of term. I suggest using either “model” or “models’”....“MPE” also stands for most probable explanation in Bayesian modeling, so this could be a point of confusion for some, without context. A1: Thanks for the suggestions. We...	Summary: This paper studies the problem of the model’s sensitivity to its training data (e.g., the counterfactual of how the model's performance will change if trained without certain data), commonly referred to as the "data influence" problem. This paper presents the "memory perturbation equation (MPE)" to study the m...	Rebuttal 1: Rebuttal: >Q1: Despite strong correlations to the Influence Function have been shown, the comparison to data influence methods in general isn't comprehensive…. It would be nice to see how it relates to or compares to broader works, such as [1] or TracIn [2].. A1: This is a good point. The MPE unifies many ...	Summary: In this paper, the authors introduce the Memory Perturbation Equation (MPE), a generalization of sensitivity measures of models that uses the recently proposed Bayesian Learning Rule (BLR) as a foundation. Thanks to its Bayesian foundation, the MPE can be used for used for non-converged models and for non-diff...	Rebuttal 1: Rebuttal: >Q1: I think a small exposition on the BLR would due wonders, especially because the paper cites specific equations from the BLR paper. To make the paper more readable, a small section in the appendix with details relevant to the paper and its experiments would do wonders. A1: Thanks for the sugg...	Summary: The paper studies the sensitivity of machine learning models to training data. Previously, sensitivity was often studied through empirical investigations which was costly and do not always generalize across models. In this paper, the authors present the memory perturbation equation (MPE), based on the Bayesian...	Rebuttal 1: Rebuttal: >Q1: I don’t see much discussions on related work or empirical comparisons to state-of-the-art methods A1: Thanks, we will improve this point, but could the reviewer mention some related works they would like us to add? For now, we considered influence function as state-of-the-art and we have new...	Rebuttal 1: Rebuttal: We appreciate the feedback provided by all reviewers. The main weakness seems to be that \ (1) the relevant background and related work can be improved; and\ (2) the experimental results are weak at times and not on large models. For issue 1, we believe it can easily be fixed and we have given r...	NeurIPS_2023_submissions_huggingface	2,023	Summary: The paper presents a generic method, relying on some Bayesian analogy, to perform instance-wise sensitivity analysis of learned models. A generic updating rule and derivative is presented, as well as its different variations on popular models (Neural networks and Gaussian models). One notable use of the propos...	Rebuttal 1: Rebuttal: >Q1: Bayesian approximation validity: In L92, I found it very strange that one can assume that there is always a proportional link between a Bayesian posterior and the exponential of a Loss function… is it guaranteed that any loss function can be related to a legitimate (proper) prior/likelihood p...	null	null	null	null	null	null
Large-Scale Distributed Learning via Private On-Device LSH	Accept (poster)	Summary: This paper presents a novel framework for on-device locality-sensitive hashing (LSH) that addresses the limitations of existing algorithms in terms of computational efficiency, memory constraints, and privacy concerns. The authors introduce a new family of hash functions that enables each device to generate ha...	Rebuttal 1: Rebuttal: We thank the reviewer for their valuable feedback. We address all questions and concerns below. --- > Is there any intuition on the choosing between PGHash version of DWTA and SimHash in practice? DWTA and SimHash approximate similarity in fundamentally different ways: * DWTA computes similar...	Summary: The paper proposes a family of hashing functions that reduces the memory cost. Let W be a weight matrix. It is observed that in many applications, one needs to draw a Gaussian matrix T and compute T * W when constructing hash tables. Also, given a query x, one needs to compute T * x which requires the storage ...	Rebuttal 1: Rebuttal: We thank the reviewer for their valuable feedback. We address all questions and concerns below. >My primary concern is the novelty of the approach. To the best of our knowledge, our LSH approach is novel due to the following reasons: - Existing literature is not focused on the serial gener...	Summary: Main Idea : If you are using LSH based sparsity, then why not project the weight matrices down using random / even structured projections and then use the LSH. Potential applications: device-based training / inference (resource-constrained) / federated learning. The authors provide theoretical anal...	Rebuttal 1: Rebuttal: We thank the reviewer for their valuable feedback. Below, we address all questions and concerns. ## Theory > [Q1]: Misleading impression of Theorem 1 being the main result Theorem 1 and its corollary are warmup results intended to illustrate a simple connection between traditional SimHash an...	null	null	Rebuttal 1: Rebuttal: ### Summary of Rebuttal: General Comments: We thank the reviewers for their valuable feedback and questions on theory and experiments. Reviewer highlights: Reviewer 7yAa believes that our "approach allows for personalized and private LSH analysis while conserving memory and...	NeurIPS_2023_submissions_huggingface	2,023	null	null	null	null	null	null	null	null
Revisit Weakly-Supervised Audio-Visual Video Parsing from the Language Perspective	Accept (poster)	Summary: The goal of this paper is to create a model that can solve the task of audio-visual video parsing in a weakly-supervised way. For this, the authors propose to use the language modality which proves to provide benefits and improve the performance of the proposed model over the previous baseline. The language mo...	Rebuttal 1: Rebuttal: We thank you for your reviews and address your concerns as follows. Q1: Do the authors also train the Resnet and VGGish? Or are they pre-trained and then frozen? A1: Yes, they are pre-trained and then frozen. The Resnet is pre-trained on ImageNet and VGGish is pre-trained on Audioset. *...	Summary: This paper tackles the audio-visual parsing task by dividing the video-level label into segment-level labels with the help of language based on CLIP. Training with the fine-grained segment-level labels, instead of the coarse video-level labels, makes the model perform better. In the process, the processing for...	Rebuttal 1: Rebuttal: We thank you for your reviews and address your concerns as follows. Q1: The main task is the “audio-visual” video parsing task, then how would the relationship between the two modalities influence the task? A1: Both modalities contribute to the recognition of events in audio-visual pars...	Summary: This paper proposes LSLD, which leverages CLIP model to denoise unreliable segment-level video labels for audio-visual video parsing. Strengths: $+$ The proposed method can achieve state-of-the-art results on LLP datasets in several metrics. $+$ The improvement of video and audio-visual events is significan...	Rebuttal 1: Rebuttal: We thank you for your reviews and address your concerns as follows. And LSLD means only visual denoise and LSLD* means both audio and visual label denoise in the following answers. Q1: The proposed method can only denoise labels for visual domains. Now we have extended our denoising method ...	Summary: The authors have observed that weakly-supervised labels in audio-visual tasks are noisy at the segment-level label. To solve the problem, the authors introduce language as an additional source of information to assign soft labels to each of the 1-second video segments. More specifically, a similarity score is ...	Rebuttal 1: Rebuttal: We thank you for your reviews and address your concerns as follows. Q1: Section 3.4, especially Equation (4), needs to be clarified. According to Equation (4), $\mathbf{\widetilde{y}}_t^v[c]=1$is already assigned with 1 or 0. Where does the value come from? A1: Sorry for the confusion. ...	null	NeurIPS_2023_submissions_huggingface	2,023	null	null	null	null	null	null	null	null
Fully Dynamic $k$-Clustering in $\tilde O(k)$ Update Time	Accept (poster)	Summary: This paper studies dynamic algorithms for k-median (and k-means) in general metrics. The main result is an \tilde{O}(k) update time, poly(k) query time algorithm that achieves O(1)-approximation. Here, the update model is point insertion/deletion, and the algorithm has access to the distance oracle. A query op...	Rebuttal 1: Rebuttal: We thank the reviewer for their effort and the review. Weakness 1: We would like to point out here that in the HK20 result, if one wishes to ensure that $\epsilon$ is small, then the resulting coreset has to be very large. For example, for the HK20 guarantee to hold with probability $½$ and $\eps...	Summary: This paper studies the $k$-median/means problems in fully dynamic settings. Clustering in fully dynamic setting is a recent hot topic, where fully dynamic $k$-center has been well studied in literature. However, little was known for fully dynamic $k$-median/means problems. Inspired by the static framework prop...	Rebuttal 1: Rebuttal: We thank the reviewer for their effort and the review. Regarding the challenges behind getting our result: Typically, the first challenge towards designing a dynamic algorithm for a given problem is to identify a suitable static algorithm which is flexible enough that it can be adapted to the dy...	Summary: Fully dynamic k-clustering in O(k) update time This paper studies fully dynamic k-clustering. It gives a fully dynamic algorithm that maintains O(1)-approximate solutions to k-median and k-means with \tilde O(k) amortized update time and \tilde O(k^2) worst-case query time. On the negative side, the authors s...	Rebuttal 1: Rebuttal: We thank the reviewer for their effort and the review. Q1: The approximation ratio of our algorithm is (very close to) a factor $4$ off the approximation ratio of the static algorithm by Mettu-Plaxton. This is about $83 + 168\alpha$, where $\alpha$ is the approximation ratio of the static algori...	Summary: This paper considers the dynamic version of the $k$-median and the $k$-means clustering problem in an arbitrary metric space. The authors provide an $O(k)$ amortized time (which is near optimal based on a lower bound that the authors provide) for insertions and deletions of points and a query time of $O(k^2)$....	Rebuttal 1: Rebuttal: We thank the reviewer for their effort and the review. We agree that we extend the Mettu-Plaxton algorithm to the dynamic setting. We believe this to be an important contribution, since our algorithm is simple and practical to implement, and improves upon the prior bound for this fundamental clu...	null	NeurIPS_2023_submissions_huggingface	2,023	null	null	null	null	null	null	null	null
Efficient Multi-Task Reinforcement Learning via Selective Behavior Sharing	Reject	Summary: This work presents Q-switch Mixture of Policies (QMP) that identifies shareable behaviors and incorporates shareable behaviors. The authors propose to utilize each task’s learned Q-function to evaluate shareable behaviors, and incorporate helpful behaviors from other tasks to aid the exploration of the current...	Rebuttal 1: Rebuttal: Thank you for your thoughtful review and feedback that helped us improve the paper. We address each of your concerns below, including new experiments. ### More Complex Environment added: Metaworld MT10 As advised, we have added an MT10 experiment to test QMP’s contribution to parameter-shared and...	Summary: The paper studies sharing behaviors between tasks in multi-task reinforcement learning. In the proposed method, each task maintains an independent policy network. During online exploration, it selects the action maximizing the task's Q function among actions proposed by all the tasks' policies. Experimental re...	Rebuttal 1: Rebuttal: Thank you for your thoughtful and constructive feedback and suggestions for additional parameter sharing experiments which we have added. We address each of your concerns below. ### Shared state space and transition function As we address in more detail in the general response, shared state spac...	Summary: The paper introduced a new exploration mechanism for MTRL. They suggested training a different policy for each task and “sharing the behaviors” between them. In order to do so, each policy, in a certain state, chooses the most suitable action using its own Q function. The author evaluates several MTRL benchmar...	Rebuttal 1: Rebuttal: Thank you for your rigorous, thoughtful and constructive feedback on the technical soundness of the Q-switch and suggestions for additional baselines. We address the main concerns below and the questions after. ### Technical Soundness of Q-switch While theoretical guarantees are outside of the sc...	Summary: This paper proposes Q-switch Mixture (QMP) for identifying shareable behaviors over tasks and incorporating them to guide exploration. QMP identifies shareable behaviors from other tasks and incorporates them to make exploration efficient. The proposed framework is tested on three different multi-agent tasks a...	Rebuttal 1: Rebuttal: Thank you for your helpful feedback on the multi-task problem setting and the complexity of the environments. We refer to the general response where we address these concerns in detail. To summarize again: - our multi-task setup with shared observation and action space, i.e., single-agent RL, is a...	Rebuttal 1: Rebuttal: We thank the reviewers for the constructive feedback. We appreciate the positive notes on the motivation for selective behavior sharing, the novelty and simplicity of Q-switch as a policy selection metric, comprehensive analysis of identified behaviors and combination with parameter sharing, and t...	NeurIPS_2023_submissions_huggingface	2,023	null	null	null	null	null	null	null	null
The Rank-Reduced Kalman Filter: Approximate Dynamical-Low-Rank Filtering In High Dimensions	Accept (poster)	Summary: This paper develops a method for performing approximate Kalman filtering and smoothing using a low-rank approximation. Fully low-rank predict, update and smoothing equations are derived. Under certain reasonable assumptions, these equations reduce the cubic complexity to linear complexity (and quadratic if t...	Rebuttal 1: Rebuttal: Many thanks for your strong review and your detailed, positive assessment of our work! We greatly appreciate your appraisal that "there is definitely an audience at NeurIPS that will enjoy and really benefit from this work". In the following we will address your remaining concerns, which we hope f...	Summary: The paper proposes an approximation method for Kalman filtering of intractable, high-dimensional Gauss-Markov settings with linear observations. It adopts existing DLRA methods to compute a discrete-time representation of the sampled process and suggests a recursive filtering and smoothing methods in a low-ran...	Rebuttal 1: Rebuttal: Many thanks for your detailed and positive assessment of our work. In the following we will address your remaining concerns, which we hope further improves and solidifies your valuation of the quality of our work. **Q: The method introduces the low-rank dimension $r$ as a tunable parameter of the...	Summary: The paper presents a deterministic low-rank approximation of the Kalman filter for estimation in high-dimensional linear Gaussian setting. This setting is important for many problems that arise in meteorology and solved by Ensemble Kalman Filter (EnKF). The paper presents formula for low rank approximation of ...	Rebuttal 1: Rebuttal: Many thanks for your detailed and positive assessment! We hope that we can additionally resolve your main point of criticism in the following and thereby convince you to raise your score. **Q: It is not clear how to apply this method in case one has to forward the dynamics with a numerical simula...	Summary: The paper proposes a deterministic low rank approximation to the Kalman filter in high dimensional settings where the computational complexity of the full Kalman filter is very expensive. They compare to the ensemble Kalman filter and ensemble transform Kalman filter. Strengths: The paper is well written. W...	Rebuttal 1: Rebuttal: Many thanks for your critical assessment of our work! We identified your main points of contention to be (a) that the ensemble Kalman filter is not a low-rank method and (b) we are not comparing our method to a low-rank filter. Both of these points are based on a misconception, which we will expla...	Rebuttal 1: Rebuttal: We are very grateful to all reviewers for their high-quality reviews and for providing constructive feedback. We are inspired by the reviewers' confirmation that our work addresses a "critical issue" (e5WD) and an "important problem" (U69V) to which our "proposed solution is appealing" (e5WD). We ...	NeurIPS_2023_submissions_huggingface	2,023	Summary: Filtering (estimation) and smoothing of large-dimensional state-space models are computationally challenging. Exact Kalman filtering, at least insofar as LTI SDEs are considered, is a mature solution for such tasks but the cubic complexity of such methods makes it computationally intractable. This paper propos...	Rebuttal 1: Rebuttal: Many thanks for your careful reading of our paper as well as your thoughtful critique. In our response, we would like to divide the major points of contention into two categories: "Address literature on reduced order methods" and "model order reduction versus our approach". These will be respond...	null	null	null	null	null	null
Trust Your $\nabla$: Gradient-based Intervention Targeting for Causal Discovery	Accept (poster)	Summary: The paper proposes to exploit the gradient information to efficiently select intervention targets. The novel approach is called Gradient-based Information Targeting (GIT), it is claimed that it is the first gradient-based intervention targeting method. A number of realistic benchmarks are used to validate the ...	Rebuttal 1: Rebuttal: We thank the Reviewer for the feedback. We are glad that the Reviewer appreciates that GIT is sound and experiments are well tailored for the problem. Below we provide answers to the Reviewer's questions and concerns: 1. [Lack of novelty]- It is true that the idea of using gradient signal in guid...	Summary: This paper proposes a method for efficient acquisition of interventional data. Specifically, given observational data and causal discovery algorithm A, the authors propose a method (GIT: Gradient-based Intervention Targeting) that chooses particular intervention(s) to enrich the data and enhance the causal dis...	Rebuttal 1: Rebuttal: We appreciate the Reviewer's feedback. It's gratifying to know that the Reviewer recognizes GIT as a robust solution to the crucial problem of intervention targeting, and acknowledges its compatibility with various causal discovery frameworks. We're also pleased that the Reviewer acknowledges the ...	Summary: This paper addresses the problem of targeting interventions to learn the causal graph. The ideal goal is to have the minimum number of interventions that is necessary to have identifiability. The authors propose a Gradient-based Intervention Targeting method (GIT) that utilize the gradient estimator of a grad...	Rebuttal 1: Rebuttal: We thank the Reviewer for the time and effort put into providing valuable feedback on our work. We are pleased to hear positive comments on the novelty of our work, the clarity of the manuscript, and the extensiveness of our experiments. We appreciate the constructive critique and questions which ...	Summary: The paper studies the problem of efficiently inferring causal structure from data. Observational data often falls short in providing a full picture of the causal structure, while obtaining interventional data tends to be costly. Therefore, optimizing the collection of interventional data to reduce the number o...	Rebuttal 1: Rebuttal: We thank the Reviewer for positive feedback on our work. We are glad that the Reviewer considers the studied problem to be important and that they appreciate a very good empirical performance of our method. Is it good to hear that we managed to clearly describe our work and present the results nic...	Rebuttal 1: Rebuttal: We would like to thank all the Reviewers for taking the time to review our work and providing us with insightful feedback. We are glad that the Reviewers appreciate the strengths of our paper, including clear writing and good presentation (Reviewers 6LvC, qbh5, 8Vao, E5qP), extensive experimental ...	NeurIPS_2023_submissions_huggingface	2,023	Summary: The authors present an approach for inferring the most informative intervention targets for the task of causal discovery; where they propose an alternative to bayesian experimental design approaches. They use the gradients of the casual discovery algorithm's loss function to find the intervention targets, whic...	Rebuttal 1: Rebuttal: We thank the Reviewer for a very encouraging review. We are glad that the Reviewer appreciated the importance of the problem, novelty, soundness of the approach and experiments, and clear presentation. Below we answer the specific questions asked by the Reviewer: - [Scalability of the method]: W...	null	null	null	null	null	null
Sample Complexity of Forecast Aggregation	Accept (spotlight)	Summary: The authors study the problem of forecast aggregation in a Bayesian setting. Here n experts observe an individual signal that is correlated with the truth, and each expert reports their posterior belief to the principal. The principal aggregates these reports and outputs a prediction, whose quality is assessed...	Rebuttal 1: Rebuttal: Alternative loss functions: Our response to Reviewer YcHV mentions several reasons why we focus on the squared loss $E[ \| f(r) - \omega \|^2 ]$ in this work. We did think about other loss functions like the logarithmic loss $E[(\omega \log(f(r)) + (1-\omega)\log(1-f(r)))]$ and the absolute los...	Summary: The paper initiates the study of sample complexity of forecast aggregation under a Bayesian forecasting model. In this problem, n experts, each observe a private signal about an unknown binary event, and then report their posterior beliefs about the event to a principal, who then aggregates the reports into a ...	Rebuttal 1: Rebuttal: Response to Questions: Maximum a posterior (MAP), or posterior mode, is the point estimate $f(r)$ that minimizes the absolute loss $E[ \| f(r) - \omega \| ]$. We choose to use the squared loss $E[ \| f(r) - \omega \|^2 ]$ instead in our work for a few reasons: (1) Squared loss is a very popular ...	Summary: The paper studies the aggregation of expert opinions in a Bayesian setting for discrete (binary) distributions. The setting further limits each expert to have at most m different opinions, making the problem fully discrete. This reduction enables the analysis of the sampling complexity of the problem, i.e., th...	Rebuttal 1: Rebuttal: 1. Experts decide on Bayesian arguments? This is an interesting observation! Indeed, our results for the case of general distributions do not need the experts to be Bayesian. It is OK if an expert reports 0.1 even if its posterior belief given signal is 0.8, as long as the expert reports differ...	Summary: This paper studies the problem of forecast aggregation: There are n experts who are each given signals about some unknown binary event, and each output a posterior probability based on these signals. We have access to k such reports (corresponding to k events). The task is to aggregate these reports to produce...	Rebuttal 1: Rebuttal: Question 1: Signals are conditionally independent if they are independent draws from some distributions but the distributions are determined by the unknown state of the world. For example, the unknown state of the world can be the quality of a school, which is either high or low. If the school q...	null	NeurIPS_2023_submissions_huggingface	2,023	null	null	null	null	null	null	null	null
Scalable Transformer for PDE Surrogate Modeling	Accept (poster)	Summary: The paper presents a high-dimensional function decomposition technique for mitigating the complexity encountered by Transformers when dealing with high-dimensional data. The authors have developed a clear logical argument and a comprehensive review of the relevant literature. It conducts extensive experiments ...	Rebuttal 1: Rebuttal: We thank the reviewer TtUN for the very detailed comments and suggestions. We also appreciate the recognition on our work. Here we would like to address your concerns and questions as below. --- > *The experimental section is somewhat weak......I suggest the authors incorporate additional represe...	Summary: This paper uses separable axis attention to reduce the complexity from being exponential in spatial dimension to linear. Strengths: NA Weaknesses: The standard Transformer has two components: attention and MLP. This paper only addresses the attention. But the MLP complexity O(Nd^2) would still make the overa...	Rebuttal 1: Rebuttal: We thank reviewer EAHP for the efforts spend on reviewing our work. However, we believe there are some misunderstanding and misinterpretation about our work and we would like to provide some clarifications below. --- > *The standard Transformer has two components: attention and MLP. This paper on...	Summary: This paper proposes a new method to scale Transformer-based models for PDE surrogate modeling to higher dimensions. The new method combines several "tricks" from previous works, and most importantly, uses factorization to further reduce the computational cost of training and evaluating Transformer-based operat...	Rebuttal 1: Rebuttal: We thank reviewer y1WA for the helpful comments and greatly appreciate your recognition on our work. Here we would like to address your concerns and questions as below. --- > *The saving of computation in the presentation of formula (7) may not be that obvious to the community of Transformer res...	Summary: The proposed work presents a scalable transformer architecture for modeling partial differential equations (PDEs). The input function is linearly projected to multiple functions with one-dimensional domains using a learnable integral projection operator. The attention mechanism is then applied to these project...	Rebuttal 1: Rebuttal: We thank the reviewer UCqx for the helpful comments and suggestions on improving the paper. Here we would like to address your concerns and questions as below. --- > *The experiments in this work use fluids with very high viscosity, resulting in simulations of flows with low Reynolds numbers. As ...	Rebuttal 1: Rebuttal: We would like to thank all the reviewers for their efforts spent reviewing our work and the suggestions on helping us improve our work. Below we summarize the newly added results that are included in the one-page pdf. #### Table 1: We tested our proposed model and Dilated-ResNet on two be...	NeurIPS_2023_submissions_huggingface	2,023	Summary: This paper proposes Factorized Transformers (FactFormer) for surrogate modeling of partial differential equations (PDEs). FactFormer first projects the input function to multiple one-dimensional projected functions, which are then used for multi-dimensional factorized attention. The factorized attention greatl...	Rebuttal 1: Rebuttal: We thank the reviewer hvcj for the insightful comments and suggestions on the paper. Here we would like to address your concerns and questions as below. --- >*The design of learnable projection and factorized kernel integral may limit the model capacity, especially for problems that do not have s...	Summary: The authors propose FactFormer, a neural PDE solver that is based upon the transformer, but scales much better with the dimensionality of the PDE problem. The authors compare FactFormer to several other neural PDE solvers on three different problems and find that Factformer achieves similar performance to st...	Rebuttal 1: Rebuttal: We thank the reviewer YpQV for the detailed comments and suggestions on improving the paper. Here we would like to address your concerns and questions as below. --- > The motivation for creating new datasets. One of the major goals of our work is to improve the scalability of attention-based P...	null	null	null	null
Improving Adversarial Transferability via Intermediate-level Perturbation Decay	Accept (poster)	Summary: The paper addresses the limitations of intermediate-level attacks, highlighting how their two-stage training scheme leads to perturbations that deviate from directional guides, significantly impairing attack performance. To overcome this limitation, the authors propose a single-stage optimization method that e...	Rebuttal 1: Rebuttal: Thanks for your positive feedback. Our response to the comments is provided as follows.   > The method seems sensitive to $\gamma$ and the intermediate layer to split $h$ and $g$. Response: Just like all other intermediate-level attacks, although our ILPD needs to tune the position ...	Summary: In this paper, the authors propose an intermediate-level attack method with a single stage of optimization to improve attack transferability. When computing gradients, it adopts intermediate-level perturbation decay to process a larger magnitude. The experiment validations on the ImageNet and CIFAR-10 datasets...	Rebuttal 1: Rebuttal: Thanks for your feedback. Our responses to the comments are provided as follows.   > From Figure 1, there is no obvious difference between the victim model and the substitute model. The reviewer is confused that these two figures are less related to the statement in Section 3.1. **Respo...	Summary: This paper presents an approach to enhancing the transferability of adversarial examples in a black-box scenario. Previous studies have typically followed a two-stage process involving the derivation of effective guiding directions and subsequently maximizing the perturbation magnitudes accordingly. Unfortunat...	Rebuttal 1: Rebuttal: Thanks for your feedback. Our responses to the comments are provided as follows.   > Despite the empirical effectiveness, the proposed method is not well-explained. Or, at least, it is unclear to me. The authors may like to elaborate more on how the "dual" equation (Eq. 3) is derived and...	Summary: The authors first introduce the observation that traditional Intermediate-Level Attack (ILA) and some subsequent works deviate from the directional guides, resulting in sub-optimal transferability. With such observation, the authors propose Intermediate-Level Perutbration Decay (ILPD), a one-staged optimizatio...	Rebuttal 1: Rebuttal: Thanks for your positive feedback. Our responses to the comments are provided as follows.   > Although the authors perform evaluations on robust models, the choice of models does not align with what used to be evaluated in the previous works (i.e. the NeurIPS 2017 Competition, some model...	null	NeurIPS_2023_submissions_huggingface	2,023	Summary: This paper notes that ILA (intermediate level attack), which maximizes projection onto a guide attack, suffers when the direction of the attack ends up differing from the direction of the guide, even when the projection is large. Thus, they propose that instead of allowing the direction to vary from a guide, w...	Rebuttal 1: Rebuttal: Thanks for your positive feedback. Our response to comments about our presentation is provided as follows. Response: The common belief in intermediate-level attacks, _e.g._, ILA, ILA++, and NAA, is that a larger magnitude of intermediate-level perturbations along the guide leads to improved ...	null	null	null	null	null	null
Fast Exact Leverage Score Sampling from Khatri-Rao Products with Applications to Tensor Decomposition	Accept (poster)	Summary: This work studies fast, exact leverage score sampling for Khatri-Rao product matrices (given access to its factor matrices) and an application to CP tensor decomposition via alternating least squares with row sampling. The main theoretical contribution is a data structure for sampling from the leverage score d...	Rebuttal 1: Rebuttal: We thank the reviewer, and we have already updated our draft to address the typos and suggestions. The table below gives the requested comparisons to out-of the-box methods. All times reported are average seconds per single ALS iteration; randomized algorithms were benchmarked with $2^{16}$ samp...	Summary: The paper studies algorithms to efficiently sample a row from a matrix $A = U_1 \odot U_2 \odot ... \odot U_N$ with probability proportional to the leverage scores of $A$. Here, $B \odot C$ denotes the _Khatri-Rao_ product of $B$ and $C$, or the column-wise kronecker-product of the columns of $B$ and $C$. In p...	Rebuttal 1: Rebuttal: We appreciate your comments, which have strengthened the quality of our draft. The global rebuttal response to all reviewers includes an extra PDF that contains updated figures (2, 3, 4, 5) with error bars / axes swaps mentioned in this review. It also includes a frequency spectrum analysis (figur...	Summary: This work proposes a new algorithm to efficiently perform exact leverage score sampling on Khatri-Rao products. This work is built on top of the TNS-CP algorithm (Malik et al., 2022), and a more efficient data structure is used to achieve better sampling computational cost. This algorithm can be used to accele...	Rebuttal 1: Rebuttal: We appreciate your review of our paper. Per the instructions, the PDF attached to our global response at the top of these reviews has a three-part figure that illustrates the sampling process, including the matrices involved, the significant operations, and the two-stage sampling procedure. These ...	Summary: This paper designs a method for efficiently minimizing least squares regression, when the design matrix is equal to the Khatri-Rao product of multiple matrices. Because this Khatri-Rao product can have a huge dimension, the paper is motivated to study sampling techniques to reduce the dimension of the product ...	Rebuttal 1: Rebuttal: We thank the reviewer for their comments and useful feedback, which have strengthened the quality of our draft. Regarding the third weakness [W3] about the novelty of our contribution: our approach offers a significant advance in the best theoretical complexity for sketched ALS CP Decomposition. ...	Rebuttal 1: Rebuttal: We thank all the reviewers for their constructive comments. The PDF attached to this rebuttal contains figures that were requested by some reviewers. These include diagrams that illustrate our sampling procedure and updated graphs with error bars. Below, you can find a summary of some revisions to...	NeurIPS_2023_submissions_huggingface	2,023	null	null	null	null	null	null	null	null
Functional Equivalence and Path Connectivity of Reducible Hyperbolic Tangent Networks	Accept (poster)	Summary: The paper investigates the functional equivalence class for reducible neural network parameters and its connectivity properties. The authors focus on single-hidden-layer hyperbolic tangent networks, but the findings can be generalized to other feed-forward network components. Strengths: The paper provides a c...	Rebuttal 1: Rebuttal: We thank reviewer Wcpf for their thorough review of our submission. We think the strengths section provides a fair summary of the main contributions of the work and that the weaknesses section provides a fair summary of the main scope limitations of the work. We are pleased to reply to the reviewe...	Summary: The paper introduces an algorithm to calculate the canonical equivalent parameter for a tanh feed-forward network. Further, the full equivalent class of each parameter is given by a union of subsets constructed out of (inverse-)reductions and permutations of the canonical parameter. Finally, the paper shows th...	Rebuttal 1: Rebuttal: We thank reviewer QPQA for the thorough review of our work. We appreciate in particular the kind comments about the clarity, novelty, and potential importance of the paper. We hope to be able to sway the reviewer's recommendation to reject the paper by defending the relevance of the results in an...	Summary: This paper focuses on the study of functional equivalence classes in neural networks, specifically in fully-connected, single-layer networks with hyperbolic tangent activation. It presents a Canonicalisation algorithm for reducible networks to determine canonical representative parameters for different functi...	Rebuttal 1: Rebuttal: We thank Reviewer HCLr for taking the time to thoroughly review our work. We aim to respond here to the several weaknesses and questions raised. Our question for the reviewer is, if the discussion of limitations and practical considerations in the paper were expanded along these lines, would the r...	Summary: This paper deals with functional equivalence problems in neural networks, i.e., the characterization of all neural networks that lead to the same given output function. This is a problem that has been studied since the early 1990s, including work by the fields medalist Charles Fefferman. The present paper con...	Rebuttal 1: Rebuttal: We thank reviewer NhDx for their review of our work and for their generous praise of our contributions. We appreciate the reviewer's acknowledgement of our novel perspective on the functional equivalence problem in particular. We do acknowledge that our work definitely has important limitations i...	Rebuttal 1: Rebuttal: We thank the reviewers for their time spent thoroughly reviewing our submission. We have responded individually, and we wanted to take the opportunity here in the author rebuttal to expand on the discussion of limitations in the paper. Limitations of architecture generality: All reviewers rai...	NeurIPS_2023_submissions_huggingface	2,023	Summary: The paper addresses fully connected feed-forward neural networks (NNs) with a single hidden layer and the hyperbolic tangent activation function. A parameter is thus a vector of weights and biases. It considers reducible parameters, which means that a NN with strictly less neurons can implement the same functi...	Rebuttal 1: Rebuttal: We thank reviewer virJ for their thorough review and for their kind words about the clarity of the paper. We appreciate the reviewer's feedback that some of the more technical elements of the proofs were difficult to follow. We are interested in broadening the accessibility of the paper to the ext...	null	null	null	null	null	null
Differentially Private Decoupled Graph Convolutions for Multigranular Topology Protection	Accept (poster)	Summary: Graph neural networks have privacy leakage in both their topology information and node attribute information. This paper proposes a differential privacy framework to protect both graph topology and node attributes. A model that decouples graph convolution and node attribute embedding is proposed. Strengths: ...	Rebuttal 1: Rebuttal: We thank reviewer fPNi for their thoughtful feedback, comments, and positive assessments of our work. All questions are addressed below. - W1: ``A weakness of using the differential privacy (DP) metric is the significant deterioration of utility … for even a very generous privacy budget. … One s...	Summary: This paper introduces a new framework called Differentially Private Decoupled Graph Convolutions (DPDGC) for graph learning settings that ensures both provably private model parameters and predictions. The framework is designed to protect sensitive user information and interactions in graph-structured data. Th...	Rebuttal 1: Rebuttal: We thank reviewer ssFv for their comments. We addressed all questions below. - W1: `` The paper's presentation is difficult to follow.`` We appreciate the reviewer’s comments regarding readability. We spend significant efforts to find a good order of exposition and lengths of explanations unde...	Summary: The paper presents a well-written and easily understandable framework called Graph Differential Privacy (GDP) tailored for graph learning methods. The proposed framework aims to address the privacy challenges associated with GNNs by ensuring both model parameter and prediction privacy. The authors introduce th...	Rebuttal 1: Rebuttal: We thank reviewer tLab for their thoughtful feedback and comments. We addressed all questions below. - W1: `` Lack of comparison with other methods: ... include a comparison with SOTA non-private graph learning methods to better understand the tradeoffs between privacy and utility.`` We apprec...	Summary: The paper introduces a differentially private GNN model that allows for different privacy requirements for node attributes and graph structure. The model decouples graph convolutions from node attributes and graph topology and provides provable privacy guarantees. Experimental results are provided to show the ...	Rebuttal 1: Rebuttal: We thank reviewer XfoH for their thoughtful feedback, comments, and positive assessment of our work. We addressed all questions below. - Q1: `` From line 310: "...". However, it is not clear what MLP refers to in Table 3.`` We apologize for the confusion. MLP in Table 3 refers to both non-priv...	Rebuttal 1: Rebuttal: We appreciate the time and effort of all reviewers and the AC. The reviews indeed provided helpful feedback. In this general response, we elaborate on some common questions raised by the reviewers (G1, G2). We also highlight some thoughtful questions and comments that initiate valuable discussions...	NeurIPS_2023_submissions_huggingface	2,023	Summary: The authors introduce a new model for Graph Differential Privacy (GDP) that ensures a parametric level of topological privacy through decoupling of the graph convolution mechanism (i.e., preventing direct neighborhood aggregation of features = standard $AXW$ aggregation). The key definition is that up to $k$ i...	Rebuttal 1: Rebuttal: We thank reviewer WwFs for their thoughtful feedback and comments. We addressed all questions below. - W1: `` A big portion of the paper is spent defining the problem and setting up conventions for future Differential Privacy studies to be more suitable to the GNN field/setting.`` We believe t...	null	null	null	null	null	null
Deep Gaussian Markov Random Fields for Graph-Structured Dynamical Systems	Accept (poster)	Summary: This paper proposes a Gaussian model for graph-temporal data, with two sorts of sparseness in the precision matrix: blockwise due to an assumed Markov property in time, and within blocks due to assumed sparse graph structures. To achieve inference and parameter learning the paper proposes a two-stage approach ...	Rebuttal 1: Rebuttal: # Official Rebuttal to Upf6 We would like to thank Reviewer Upf6 for the positive assessment, the thorough feedback and interesting questions. In the following, we will respond to each point separately, and refer to our general response to all reviewers whenever there is overlap with other review...	Summary: The paper extends Gaussian Random Fields with precision-based spatiotemporal structure. Strengths: Clarity: The paper is professionally written, with flawless math. The method is technical and dense, and there are many parts to it. It would have been useful to include illustrations, perhaps in the supplement...	Rebuttal 1: Rebuttal: # Official Rebuttal to k13T We would like to thank Reviewer LAkB for praising the quality and clarity of our method and for providing constructive feedback. In the following, we will respond to each point separately, and refer to our general response to all reviewers whenever there is overlap wit...	Summary: The author extend work on deep Gaussian Markov random field to a spatio-temporal setting. Earlier it has been explored in spatial and graphical setting. The authors show how to retain the sparsity structure of the previous work, which is need to be computationally feasible. They also show the various types of...	Rebuttal 1: Rebuttal: # Official Rebuttal to LAkB We would like to thank Reviewer LAkB for the positive assessment and good suggestions. In the following, we will respond to each point separately, and refer to our general response to all reviewers whenever there is overlap with other reviewers. ## Responses to Weakne...	Summary: This work extends DGMRF to account for temporal dependency in data with a SSM reformulation. With necessary assumptions, the proposed method produces accurate posterior, competitive performance to DGMRF and faster inference to Kalman smoother. Strengths: - The method extends DGMRF to account for temporal depe...	Rebuttal 1: Rebuttal: # Official Rebuttal to 9F96 We would like to thank Reviewer 9F96 for the positive feedback and the good suggestions. In the following, we will respond to each point separately, and refer to our general response to all reviewers whenever there is overlap with other reviewers. ## Responses to Weak...	Rebuttal 1: Rebuttal: # General rebuttal to all reviewers First of all, we would like to thank all the reviewers for their thoughtful and constructive feedback. We are delighted to hear that our paper is "clear" (LAkB), "easy to read" (Upf6) and "professionally written" (k13T), and that the reviewers found our propose...	NeurIPS_2023_submissions_huggingface	2,023	null	null	null	null	null	null	null	null
Finding Order in Chaos: A Novel Data Augmentation Method for Time Series in Contrastive Learning	Accept (poster)	Summary: This paper presents a new time series augmentation method which adopts Mixup on the amplitude and phase, respectively, after transforming the time series in the frequency domain, thus avoiding the destructive extrapolation from linear Mixup. Strengths: 1. The authors propose a novel mixup method in the freque...	Rebuttal 1: Rebuttal: We thank the reviewer for the thorough review. We appreciate the feedback and the recognition of the importance of our work. We replied to each concern as follows. 1)Regarding the ablation experiments focusing on amplitude or phase alone for the mixup. We'd like to clarify that we conducted a par...	Summary: This paper proposes a novel mixup method for non-stationary time-series data to generate positive samples for the contrastive learning formulation. The proposed method mixes the magnitude and phase of each frequency component of two samples that are close in the latent space of a variational autoencoder. The a...	Rebuttal 1: Rebuttal: We thank the reviewer for the thorough review of our work. We appreciate the feedback and are glad to hear that our work has been received positively. Regarding the weaknesses, thank you for bringing hyperlinks to our attention. We will fix this issue in the revised version of the manuscript. Y...	Summary: This paper is about a novel data augmentation method that can be used in contrastive learning for time-series tasks and aims to connect intra-class samples together, and find order in the latent space. The proposed method builds upon the mixup data augmentation technique by controlling the degree of chaos crea...	Rebuttal 1: Rebuttal: We thank you for your comments. And, we would like to clarify some misunderstandings and points arising from your review of our paper. Regarding the claim of your incremental novelty, it's important to note that there currently exists no prior study that has specifically addressed the destructive...	Summary: This paper explores data augmentation for contrastive learning that better aligns with the nature of time series. Specifically, they tailor the mix-up method separately to amplitude and phase in the frequency domain. A theoretical analysis illustrates how the method enhances task-specific information, contrast...	Rebuttal 1: Rebuttal: We appreciate your feedback about our paper on data augmentation for contrastive learning in the context of quasi-period time series data. We also thank you for valuing the importance of the problem. We would like to address certain points raised in your review. We are well aware of all the work...	Rebuttal 1: Rebuttal: We would like to thank all the reviewers for taking the time to read our paper and provide feedback. In response to the reviewers' comments, we included two more data augmentations for comparison. We also applied our proposed data augmentation method to one more contrastive learning framework, wh...	NeurIPS_2023_submissions_huggingface	2,023	null	null	null	null	null	null	null	null
Epistemic Neural Networks	Accept (spotlight)	Summary: This paper introduces Epistemic Neural Networks (ENN) as a general approach for uncertainty estimation for deep learning, and then the epinet, which is a particular ENN instantiation. The paper argues that joint predictions are essential for sequential decision making problem, and shows epinets offer s...	Rebuttal 1: Rebuttal: We thank the reviewer for carefully reviewing our paper and writing detailed reviews. Our responses are as follows. Due to the length limit of the rebuttal, we will omit parts of the questions and provide concise answers. Regarding your questions: **1. L42-43 "evaluating each model on out-of...	Summary: In this paper, the authors present a new approach for uncertainty estimation utilizing joint predictions. They introduce the concept of an Epistemic Neural Network (ENN). Within this framework, they propose an innovative architecture called 'epinet,' which supplements any conventional neural network. This addi...	Rebuttal 1: Rebuttal: We thank the reviewer for their time and effort in reviewing this paper. Our responses are: > I think that there is a missing opportunity to show the performance of the proposed approach on direct uncertainty quantification... It would be helpful to see the authors run standard experiments, like ...	Summary: The paper introduces epinets as part of Epistemic Neural Networks (ENNs), a novel approach to uncertainty estimation in deep learning models. Epinets extend neural networks to create ENNs, which can be used to estimate uncertainty. The paper then proceeds to present experiments on image classification tasks, d...	Rebuttal 1: Rebuttal: We thank the reviewer for reviewing our paper. Our point-to-point responses are as follows: **Even though the framework is rather general, the focus of the experiments seems rather narrow and confined to image classification tasks. It would be useful to see how well the epinet generalizes to othe...	Summary: In this work, the authors introduce a new framework "Epistemic Networks" to better capture uncertainty by integrating over epistemic indices when computing the joint distribution over multiple inputs. This approach is flexible and can be added to any existing neural network architecture with some differences i...	Rebuttal 1: Rebuttal: We would like to thank Reviewer TShA for reviewing our paper. The following are our point-to-point responses: **I think the authors need to do a better job to motivate the case for joint log-loss. Why should one care for joint log-loss performance as opposed to marginal? Can you describe any appl...	null	NeurIPS_2023_submissions_huggingface	2,023	Summary: This paper proposes epistemic neural networks (ENNs) as a novel framework for uncertainty estimation in neural network predictions. ENNs introduce an epistemic index that expresses uncertainty and correlations across multiple inputs via joint predictive distributions. The paper argues that joint predictions ar...	Rebuttal 1: Rebuttal: We would like to thank the reviewer for their time and effort in reviewing the paper. We address the comments below: **> Theoretical analysis regarding regret bound: Could you provide a more thorough theoretical analysis quantifying the advantages of joint modeling for decision making? Any regret...	null	null	null	null	null	null
Causal Discovery in Semi-Stationary Time Series	Accept (poster)	Summary: The problem of causal discovery is tackled for time series data in those cases where stationarity of multivariate time series data can not be assumed. A problem formulation based on structural causal model is given by considering the case of semi-stationary time series. Under this assumptin, the paper consider...	Rebuttal 1: Rebuttal: We really appreciate the feedback. Regarding your valuable questions, please see below for our responses. ## Q1. Computational Complexity Executing the PCMCI algorithm on the entire time series constitutes the initial phase of the proposed approach (designated as "line 1" in Algorithm 1). The...	Summary: The paper describes a constraint-based causal discovery method for semi-stationary SCMs, where the causal structure is periodically changing over time. This type of structure is relevant in real-life scenarios where seasonal or diurnal variation is present for example. The problem is rigorously formalized and...	Rebuttal 1: Rebuttal: We really appreciate the feedback. Regarding your questions, please see below for our responses. ## Q1. Sampling issue Thank you for this great question. Our proposed method can handle the "unchanged" causal mechanism as long as there exists a periodicity in how the causal mechanisms change ov...	Summary: ## Summary The paper addresses the problem of discovering causal relations in semi-stationary time series data, where a finite number of causal mechanisms occur sequentially and periodically across time. The authors propose a constraint-based, non-parametric algorithm called $PCMCI_\Omega$ to handle this type ...	Rebuttal 1: Rebuttal: Your insights are greatly appreciated as we work towards enhancing the robustness of our method. Regarding your questions, ## Q1. Scope of experiments To our best knowledge, there is no such non-parametric algorithm that can handle the periodicity in causal mechanisms and discover the time-lag e...	Summary: The authors propose a causal discovery algorithm for a class of periodic random processes called semi-stationary process. The algorithm is built upon an existing algorithm assuming stationarity. Consistency results of the algorithm is provided. Strengths: 1. Consistency results of the algorithm was shown eve...	Rebuttal 1: Rebuttal: We really appreciate the feedback. Regarding your questions, ## Q1. Description of PCMCI Thank you for the suggestion. In the final version of the paper, we will provide a concise overview of how PCMCI functions, utilizing the additional page available in the camera-ready version. Here is a bri...	Rebuttal 1: Rebuttal: Dear reviewers, We extend our sincere gratitude for dedicating your time to meticulously assess our submission and offer invaluable insights. With the initial rebuttal phase comes to an end, we believe that our response effectively addresses the raised points of concern. We have attached a one-...	NeurIPS_2023_submissions_huggingface	2,023	null	null	null	null	null	null	null	null
Self-Correcting Bayesian Optimization through Bayesian Active Learning	Accept (poster)	Summary: The paper proposed a novel function SAL used for active learning based on distribution difference. Then the author further incorporated the SAL function into Bayesian optimization (BO) to improve BO performance. Most numerical experiments are conducted on synehtic benchmakr functions. Better performances are a...	Rebuttal 1: Rebuttal: We thank the reviewer for their thorough work reviewing the paper. We have added additional experiments, and hope to clarify the notation-induced confusion. __0)__ _Distinguish among three different things: $f(x), f(x) + ε, GP(x)$_ Using the notation $f(x)$ to denote both the true black-box f...	Summary: The paper introduces a technique called statistical distance-based active learning (SAL) for the purpose of learning the hyperparameters of a Gaussian process. Additionally, SAL is integrated with information-theoretic Bayesian optimization (BO) in a framework referred to as self-correcting BO (SCoreBO). This ...	Rebuttal 1: Rebuttal: We thank the reviewer for their through review. There are, however, crucial misunderstandings regarding our work: __all baselines use fully Bayesian hyperparameter treatment__ and the active learning presented in this paper __goes beyond conventional hyperparameter marginalization__. The feedback ...	Summary: The paper presents two algorithms/methods, SAL and SCoreBO, for active learning and Bayesian optimization using GPs. Both build on the premise that hyperparameters are critical to effective AL or BO; hence it is important to learn both the function and the hyperparameters through an acquisition function that a...	Rebuttal 1: Rebuttal: We thank the reviewer for their feedback and are glad to see that the reviewer appreciates both the importance of the problem setting as well as the proposed approach. --- > __1)__ _Lack of details regarding a fully Bayesian treatment_ --- We agree with the reviewer that the paper would benefit ...	Summary: Standard acquisition functions in Bayesian Optimization (BO) only aim to finding the optimum, but do not directly consider the problem of hyperparameter learning in Gaussian Processes (GP) which has considerable impact on the optimization performance. This paper introduces a new acquisition function for Activ...	Rebuttal 1: Rebuttal: We thank the reviewer for their feedback. We are pleased to see that the reviewer recognizes the significance of the joint BO/hyperparameter learning problem setting and our proposed approach. --- > __1.__ _This paper would strongly benefit from showing ScoreBO at work on other real-world tasks,...	Rebuttal 1: Rebuttal: The authors would like to thank all reviewers for their effort. As per the reviewers' requests, we have added 4 additional plots to the rebuttal PDF. __Fig. 1:__ Hyperparameter convergence of fully Bayesian JES, MES and SCoreBO on the SAASBO task, the noisy 25-D Ackley function. We see that the f...	NeurIPS_2023_submissions_huggingface	2,023	null	null	null	null	null	null	null	null
Train Faster, Perform Better: Modular Adaptive Training in Over-Parameterized Models	Accept (poster)	Summary: The authors introduce a novel modification of the neural tangent kernel (NTK) called modular NTK (mNTK). It is essentially the decomposition of the NTK as a sum of tangent kernels for the disjoint modules that make up the network. This allows for module-level analysis of training dynamics. They find that princ...	Rebuttal 1: Rebuttal: > `Q1` For the language models, the paper only looks at perplexity scores of the pretrained model for text. I'm not the most up to date on this, but I think typically practitioners are more concerned about the performance of the model when fine-tuned on down stream tasks. I think comparing perform...	Summary: This paper introduces modular adaptive training based on the largest eigenvalue of the NTK. In particular, based on the assumption that the principal eigendirection matters the most for generalization, the authors propose to dynamically turn off gradient updating on certain modules if they have small principa...	Rebuttal 1: Rebuttal: > `Q1` In Figure 2(a), can we also see a picture showing the distribution of all eigenvalues? Now even the 16th eigenvalue seems pretty large, is there very small eigenvalue? `A1` Thanks for your question. The eigenvalues can be refered to Figure A of PDF in General Response. The computation of...	Summary: In this work the authors propose a method for analyzing modules of a neural network, in particular their utility in test-time generalization, via a modular neural tangent kernel (mNTK). The mNTK is eigenspectrum the NTK of the network modules, e.g. different attention heads. In this work it was found that the ...	Rebuttal 1: Rebuttal: > `Q1` The mNTK isn't super novel, as it is very similar to spectral analysis work which has been performed before, although I do think it is sufficiently novel enough for to ML venues. `A1` Thanks for your comments. Indeed, the spectrum of NTK has been applied to two-layer networks to analyze t...	Summary: Inspired by the Neural Tangent Kernel literature, this paper analyzes the training dynamics of over-parameterized neural networks from the perspective of modularity. In particular, the paper proposes to decompose the Jacobians of the NTK into those of modules. Here, the concept of "module" is a bit ill-defined...	Rebuttal 1: Rebuttal: > `Q1` The biggest limitation ... the speed gain. > I would really appreciate a deeper comment on the entire computation time when accounting for the algorithm itself. `A1` Thanks for your comment. The proposed method MAT will introduce additional computational overheads as it involves calculati...	Rebuttal 1: Rebuttal: # General Response Thank you very much for reviewing our manuscript and providing detailed and constructive comments, which have been very helpful for us to improve the quality of our work. Please see our answers to address the comments of individual reviewers. Additionally, enclosed pdf with fig...	NeurIPS_2023_submissions_huggingface	2,023	null	null	null	null	null	null	null	null
Human-like Few-Shot Learning via Bayesian Reasoning over Natural Language	Accept (oral)	Summary: This paper presents a model for how humans learn abstract symbolic concepts from induction. The model uses an off-the-shelf language model as a meta-prior, which is then tuned to form a task-specific prior over hypotheses using a small number of human samples. This prior is then incorporated into a Bayesian in...	Rebuttal 1: Rebuttal: Thank you for the thoughtful review and kind words. Below we respond to some of your main points, but can discuss further during the discussion period: > Do you have thoughts on exactly how "strong" you should make the prior and likelihood functions? Our thoughts are that the proposal distributi...	Summary: This paper proposes a model for human learning of concepts from few examples (aka "few shot" setting) that leverages natural language as an internal concept representation. This is a key trick because it means that * an LLM can be used to act as a proposal distribution for (efficiently) generating data-depend...	Rebuttal 1: Rebuttal: Thank you for the thoughtful review and enthusiastic support. Please find below our clarifications: > …mention of some task or problem setting that successfully "red teams" this approach - ie, some problem which is adversarially chosen to be challenging or ill-suited for this framework (while sti...	Summary: The authors introduce a computational approach for explaining how humans perform few-shot concept learning. The proposal is that humans use Bayesian inference over natural-language definitions of concepts. In more detail, a bottom-up proposer generates a set of candidate concept definitions, and Bayesian infer...	Rebuttal 1: Rebuttal: Thank you for the thoughtful input and enthusiastic support. We are happy to correspond more during the discussion period, but here we mainly just address the specific question you raised: > …unclear to me what the paper’s main contribution is: Is it (i) a new hypothesis about how humans perform ...	Summary: The paper proposes an approach to scaling up intractable Bayesian models of few-shot concept learning. The key idea is to (1) train an amortized posterior distribution q(C\|X_{1:K}) over concepts (represented as natural-language expressions) and then (2) make predictions about membership by marginalizing over a...	Rebuttal 1: Rebuttal: Thank you for the thoughtful input, kind words, and suggested improvements. We address the main issues, with new results bolded, but can answer all of your questions during the discussion period. > My primary concerns are around how to do credit assignment to the many potentially brittle comp...	Rebuttal 1: Rebuttal: Thank you all for your reviews, and especially for the encouragement and constructive criticism. Below we summarize some important strengths and weaknesses identified across the reviews, and overview the new results in the attached PDF that address those weaknesses. Strengths: qc38 describe...	NeurIPS_2023_submissions_huggingface	2,023	null	null	null	null	null	null	null	null
RECKONING: Reasoning through Dynamic Knowledge Encoding	Accept (poster)	Summary: The paper introduces a two-level learning algorithm called RECKONING, which enhances the in-context reasoning performance by addressing the issue of distractor facts. The algorithm consists of two learning steps, an inner and an outer loop, where the inner loop trains the model to encode contextual knowledge i...	Rebuttal 1: Rebuttal: ## Responses to Reviewer 5k3Z (R5) We thank the reviewer for their constructive comments, as well as describing our idea as interesting and important, and recognizing the benefits of our approach. We address the reviewer’s concerns and questions below: **W1: The reviewer argues that if there...	Summary: This work aims to solve reasoning tasks where models need to rely on knowledge provided as part of the task input. Motivated by the fact that pre-trained models encode a lot of irrelevant facts and they are good at retrieving the relevant ones for solving a downstream task, this work proposes an algorithm, REC...	Rebuttal 1: Rebuttal: ## Responses to Reviewer W7M8 (R4) We thank the reviewer for their constructive comments and viewing our idea as interesting and sound, our experiments as sufficient and well-motivated toward the study, and our algorithm as well explained. We address the reviewer’s comments and questions below: ...	Summary: This paper introduces a novel method for addressing logical reasoning in natural language question answering, specifically when the required knowledge is part of the context. One major challenge highlighted in this paper is that the included knowledge often contains irrelevant information, which can mislead th...	Rebuttal 1: Rebuttal: ## Responses to Reviewer ZqBF (R3) Dear reviewer, thank you for the constructive comments and we appreciate your time and effort. We thank the reviewer for recognizing our approach as interesting and sound, and our paper as well-written. We are encouraged to see the reviewer acknowledging our con...	Summary: Summary: The paper introduces RECKONING, a bi-level learning algorithm designed to improve reasoning in transformer-based language models. RECKONING encodes contextual knowledge into the model's parameters using gradient updates, allowing the model to answer questions based on its updated parameters. The autho...	Rebuttal 1: Rebuttal: ## Responses to Reviewer UhBM (R2) We thank the reviewer for their helpful comments and are encouraged that the reviewer agrees that our algorithm is “an interesting and novel idea in the field of natural language reasoning” and recognizes its benefits. We also thank the reviewer for complimentin...	Rebuttal 1: Rebuttal: ## General responses to all reviewers We would like to thank the reviewers for providing us with thoughtful comments and constructive feedback. We appreciate that the reviewers recognize our proposed bi-level learning algorithm for language reasoning as novel/interesting (R2, R3, R4, R5), us...	NeurIPS_2023_submissions_huggingface	2,023	Summary: The paper presents a two-phase learning algorithm that 1) encodes background knowledge in the parameters of an LM through fine-tuning (instead of providing it in context), and 2) learns how to use and reason the encoded knowledge for a given question (?). The paper shows better performance than in-context lea...	Rebuttal 1: Rebuttal: ## Responses to Reviewer ymTg (R1) We thank the reviewer for viewing our proposed bi-level learning algorithm as interesting and novel and for recognizing its benefits and contribution. Below, we address the questions: Q1: The reviewer asks if the order of the questions affect the results N...	null	null	null	null	null	null
Does a sparse ReLU network training problem always admit an optimum ?	Accept (poster)	Summary: The authors are interested in the following question : « Given a deep learning architecture possibly with sparsity constraints, does its corresponding optimisation problem actually admit an optimal $\theta^*$? » While it is mostly taken for granted that the answer is « yes », the author proves it to be true in...	Rebuttal 1: Rebuttal: We thank the referee for his/her comments and questions. Comments on the weakness 1.1. While regularization with a coercive function indeed ensures the existence of a minimizer, this however implies tradeoffs between the data fitting term and the penalty, usually tuned via a parameter $\lamb...	Summary: This work studies the existence of global minimum in a sparse deep learning setting Strengths: I think this work touches on a very important problem that is understudied in the field: the existence of the global minimum for deep nonlinear networks. It is easy to image toy examples for which the global minimum...	Rebuttal 1: Rebuttal: We thank the referee for his/her comments. Comments on the weakness For the first point, our Example 3.1 about the "LU architecture" is meant to be pedagogical, and its ultimate goal is to show that even in pretty simple cases there may be actual issues related to non-closedness. In our re...	Summary: This paper provides necessary and sufficient conditions for whether a neural network training problem admits an optimal solution, that is, whether weights and biases exist such that the infimum of the loss function is actually attained. This is done for the classical case of empirical risk minimization on a fi...	Rebuttal 1: Rebuttal: We thank the referee for his/her comments and questions. Comments on weakness We really appreciate the comments on the overall organization and presentation of the paper and will implement them correspondingly in the final version. Regarding the limitation to 2-layers only, all our *negat...	Summary: The paper studies the existence of an optimal solution to the objective of training a ReLU neural network under certain sparsity patterns. They consider two topological properties of the input function space (the best approximation property, and closedness). They provided series of results on scalar-valued neu...	Rebuttal 1: Rebuttal: We thank the reviewer for his/her comments on our work. Comment on the weakness Our work allows to answer the following question: if the supports of the weight matrices are randomly sampled from a distribution, what is the probability that the corresponding training problem potentially admits...	Rebuttal 1: Rebuttal: Dear reviewers, In this global response, we attach a pdf file containing figures for our rebuttal. Pdf: /pdf/9440f8eade73540011d67f0011e14a52020b0564.pdf	NeurIPS_2023_submissions_huggingface	2,023	null	null	null	null	null	null	null	null
Slow and Weak Attractor Computation Embedded in Fast and Strong E-I Balanced Neural Dynamics	Accept (spotlight)	Summary: The paper combines two established models in theoretical neuroscience, the continuous attractor neural network (CANN), and the excitation-inhibition balanced neural networks (E-INN). Though those models have been applied to explain various phenomena in the brain, they possess seemingly contradictory characteri...	Rebuttal 1: Rebuttal: Thanks for the encouraging and valuable comments of the reviewer, which are very helpful for us to improve the paper. Below, we would like to address reviewer’s concerns in weakness and questions in detail. On weakness: Thank the reviewer for pointing out that there exist inconsistencies of...	Summary: The authors present a model of a continuous attractor network in a balanced-state spiking network. The model combines fast and slow connections. The fast connections are in charge of the balanced state and are thus strong 1/sqrt(K). The slow connections are in charge of the continuous attractor, and are weak 1...	Rebuttal 1: Rebuttal: Thanks for the encouraging and valuable comments of the reviewer, which are very helpful for us to improve the paper. Below, we would like to address reviewer’s concerns in weakness and questions in detail. 1. Localized inputs disrupt the balance condition, if the EI balance dynamics is mediated...	Summary: The paper proposes a new spiking neural network where structured connectivity, consistent with a CANN, is combined with random connectivity, consistent with a excitation and inhibition network, with two inhibitory subpopulations. The network depends onj two different sets of weights: weak weights for CANN dyna...	Rebuttal 1: Rebuttal: Thanks for the encouraging and valuable comments of the reviewer, which are very helpful for us to improve the paper. Below, we would like to address reviewer’s concerns in weakness and questions in detail. On weakness: 1. Whether the model can maintain a persistent bump is mainly dependent...	Summary: This study explores the compatibility of attractor networks and excitation-inhibition balanced networks (E-INNs) in neural circuits. It proposes that a neural circuit can exhibit traits of both by utilizing two sets of synapses: one set for strong and fast irregular firing and another set for weak and slow att...	Rebuttal 1: Rebuttal: Thanks for the encouraging and valuable comments of the reviewer. Below are our replies to reviewer’s concerns. 1. On the biological plausibility of our model. We actually just found an experimental study [1] (which we did not know when writing this paper), which has already provided strong evide...	Rebuttal 1: Rebuttal: On the biological plausibility of our model We are delighted to find that an experimental study [1], which we did not know when writing this NeurIPS paper, has already provided strong evidence for the key assumption of our model, namely, the neural circuit consists of two sets of synapses: a ...	NeurIPS_2023_submissions_huggingface	2,023	null	null	null	null	null	null	null	null
NeuralGF: Unsupervised Point Normal Estimation by Learning Neural Gradient Function	Accept (poster)	Summary: The paper proposes a method to learn normals for point clouds using a neural representation. The idea is to perform a multi-step point moving and compute a series of losses to constrain the gradient of the field to describe shape locality with consistency. The paper proposes a large series of comparisons with ...	Rebuttal 1: Rebuttal: ### 1. Discussion of contributions, and improvements to previous works. As we can see from Table 2, existing supervised methods can achieve high-precision unoriented normal. However, from Table 1, we can observe that higher-precision unoriented normal do not result in more accurate oriented normal...	Summary: The paper introduces a method for estimating oriented normals from a given point cloud by utilizing a neural networkto model a sign distance field (SDF). The proposed approach involves training the SDF representation, which allows for easy querying of gradients at the positions of the point cloud. The method e...	Rebuttal 1: Rebuttal: ### 1. Some definitions, notations, and wordings need to be revised. $f_i$ in Eq.(2) represents the signed distance of each point during the i-th position movement process. $f_i^G$ in Eq.(3) represents the signed distance of each point in point set G during the i-th position movement process. As...	Summary: In this work the authors present an unsupervised framework for predicting globally consistent, accurate normals given a point cloud as input. The crux of the method is to predict an implicit surface representation (a signed distance function to be exact) from the input point cloud by leveraging the fact that t...	Rebuttal 1: Rebuttal: ### 1. Comparison with other implicit representations. We use some implicit representation methods to estimate oriented normals. The comparison of normal RMSE on point clouds in the datasets PCPNet and FamousShape is reported in the following table and the visual comparison is shown in *Rebuttal ...	Summary: This work proposes to learn neural gradient functions from point clouds to estimate oriented normal in an unsupervised manner. Specifically, this method introduces several loss functions to constrain query points to iteratively fit the underlying surface, which is defined by the sampled points. Meanwhile, the ...	Rebuttal 1: Rebuttal: ### 1. Why need a distribution D? The optimization of our method is formulated as an iterative moving operation of the points to the target positions. We use this distribution to generate the training data, i.e., query points, which are pulled onto the surface during optimization. As in Line 117,...	Rebuttal 1: Rebuttal: We thank the reviewers for their valuable comments, and our responses to all reviewers are as follows. More visual comparison results are shown in the provided PDF. Pdf: /pdf/3d5b39f9f5d7af2d4f1bb75b4cf71cb3a8d76e29.pdf	NeurIPS_2023_submissions_huggingface	2,023	Summary: The paper proposes a method that predicts oriented surface normals given a 3D point cloud. The method is based on SDF surface reconstruction. Given a potentially noisy point cloud, a SDF is fitted using an approach similar to Neural-Pull [45]. The surface normal at each point can then be computed based on the ...	Rebuttal 1: Rebuttal: ### 1. Comparing with other implicit representation methods and repurposing surface reconstruction methods for normal estimation. We use some implicit representation methods to estimate oriented normals. The comparison of normal RMSE on point clouds in the datasets PCPNet and FamousShape is repor...	null	null	null	null	null	null
Safe Exploration in Reinforcement Learning: A Generalized Formulation and Algorithms	Accept (poster)	Summary: This paper studies the safe RL problem with a generalized stepwise safety chance (probability) constraint, essential for many safety-critical systems with RL. The authors propose a meta-algorithm to solve this problem (MASE), by combining unconstrained RL with an uncertainty quantifier to guarantee safety with...	Rebuttal 1: Rebuttal: We deeply appreciate the reviewer’s encouraging comments. Your feedback and suggestions are valuable and help us improve the quality of the manuscript. We will first answer the Question and then address the comments regarding Weakness. Lines 96 - 98 [Question]. Thank you for the important qu...	Summary: The authors present a novel safe RL algorithm for safety constraints with probability one. First, the authors present a problem formulation that can be used to derive different common safe RL formulation (state constraints, accumulated constraints etc). Then the authors leverage a sophisticated technique to re...	Rebuttal 1: Rebuttal: We deeply appreciate the reviewer’s encouraging comments. Your thoughtful comments are valuable and help us to improve the quality of the manuscript. We will answer the Questions first and then address the comments in Weakness. Safety-Gym [Question 1]. Thank you for the very important questi...	Summary: Paper considers the generalized safe exploration problem (problem 4) and compares it to the other safe exploration problems in the literature, leading to Thm 3.1 that concludes that it is more general than the others. The authors then introduce MASE, a meta-algorithm for safe exploration, that attempts to solv...	Rebuttal 1: Rebuttal: We deeply appreciate the reviewer’s valuable comments and feedback. We will answer the Questions first and then address the comments provided in Weakness. Chance constraint [Question and Weakness 3]. Thank you for the good question! In general, it is known that joint chance constraints repr...	Summary: This paper addresses the problem of safe reinforcement learning, in particular safe exploration. In a nutshell, the authors provide an algorithm that is supposed to go beyond standard 'safety' measures in safe RL, such as the constrained MDP setting. There, an agent is bound to satisfy an additional (expected)...	Rebuttal 1: Rebuttal: We deeply appreciate the reviewer for helpful and thoughtful comments and questions. First of all, please let us emphasize that our main contributions include the formulation of the GSE problem and its theoretical result (i.e., Theorem 3.1) as well as the proposal of the MASE algorithm. Our c...	Rebuttal 1: Rebuttal: Dear reviewers and AC, We deeply thank all the reviewers for their insightful comments and constructive suggestions. - We have conducted new experiments based on the reviewers' comments. Additional experimental results are provided in a one-page PDF containing new figures attached in this "globa...	NeurIPS_2023_submissions_huggingface	2,023	null	null	null	null	null	null	null	null
Locally Invariant Explanations: Towards Stable and Unidirectional Explanations through Local Invariant Learning	Accept (poster)	Summary: This work extends [LIME](https://arxiv.org/abs/1602.04938) to generate stable and unidirectional feature attributions through the [invariant risk minimization framework](https://arxiv.org/abs/1907.02893). Their method, LINEX, ensures that it follows the desiderata of faithfulness, robustness to neighborhood sa...	Rebuttal 1: Rebuttal: Thank you for highlighting the strengths of our approach. We now address your concerns. > qualitative insights into how models work We provide several qualitative insights in last paragraph of Section 5 and in appendices I, J, . Please also look at Figure 25. 1. Figures 2 and 3 show LINEX expla...	Summary: The paper proposes Locally Invariant Explanations (LINEX) as a variation of the well-known LIME explanation method. While LIME is widely used in interpretability, it has been demonstrated to be dependent on the sampling procedure to create neighbouring points as well as the choice of the explanation model. The...	Rebuttal 1: Rebuttal: Thank you for your suggestions in improving the structure of our paper. We are also glad that you liked the approach and our experimental evaluations. > Scope: modifying the sampling scheme of LIME LIME is one of the most widely used explanation approaches and being model agnostic is applicable ...	Summary: Explanation methods are known to be unstable with flickering attributions on inputs with mild perturbations. This work addresses the explanation instability problem by re-posing it as computing attributions that are invariant across different environments (that are obtained through input perturbations). Their ...	Rebuttal 1: Rebuttal: We are glad that you found our idea to be interesting. Below are our responses to your concerns. > Connection to IRM Besides the aspect of environments which you mentioned being similar to IRM, citation [4] mentioned in the paper is a game theoretic algorithm to solve IRM. IRMv1 which was mentio...	Summary: In this paper, the authors put forward a local, attribution-based explanation method that fits a local, explainable surrogate (a linear model in this case) to perturbation output of the original model. In essence, this is exact scenario as the immensely popular LIME explanation method; however, this work puts...	Rebuttal 1: Rebuttal: Thank you for realizing the potential of our approach. We now address your concerns. > Reliance of LINEX on the “diversity” of environments Typically when using IRM environments are given. One of our contributions here is proposing simple alternatives to create them for our problem. In our case,...	Rebuttal 1: Rebuttal: We thank all the reviewers for their constructive comments. We are glad that you found our work to be an interesting idea, that could be widely adopted, having clear motivation, with straightforward development, strong experimental evidence, **easily adaptable into post-hoc e...	NeurIPS_2023_submissions_huggingface	2,023	null	null	null	null	null	null	null	null
From Discrete Tokens to High-Fidelity Audio Using Multi-Band Diffusion	Accept (poster)	Summary: The paper proposes to use diffusion model to recover high-fidelity audio from compressed audio tokens. To improve the performance of diffusion model, the paper proposes multi-band diffusion with 1) Frequency Eq. Processor; 2) Scheduler Tuning; and 3) Band-Specific Training. The experiments show that the propos...	Rebuttal 1: Rebuttal: Thank you for your review and suggestions. Clarification about the setting: We will clarify the model's input-output relationship in the paper, as it currently appears to be unclear. All models take the full Encoder compressed representation as input. The diffusion process occurs in the wavef...	Summary: This paper presents Multi-Band Diffusion, a diffusion-based decoder for neural audio codecs. The proposed method demonstrates superior generation quality across various audio domains compared to a publicly available state-of-the-art neural audio codec method. Based on the analysis of time-frequency representat...	Rebuttal 1: Rebuttal: High bit rate: The Neural audio codec literature has mainly focussed on low bit rates. Moreover, these low bitrate setups align with the configurations employed by language modelling approaches that our method seeks to enhance, such as Musicgen, or Vall-E.. While our method is applicable to hi...	Summary: This paper proposes a novel multi-band diffusion (MBD) model that generates high-fidelity audio of multiple modalities, e.g., speech, music, and environmental sounds, from low-bitrate discrete representations. The authors show that MBD outperforms EnCodec and Opus in terms of perceptual quality. Strengths: Th...	Rebuttal 1: Rebuttal: Thank you for your review and comments. Comparison to SoundStream: Despite our willingness to perform a direct comparison with SoundStream, it's important to note that the authors have not released a public repository, and there is no public reimplementation on GitHub that is reproducing Soun...	Summary: This is an interesting submission proposing the use of a diffusion model in a band-by-band manner to generate high-fidelity audio from (potentially) a variety of low bit rate inputs. The literature so far has only applied diffusion models for audio generation given spectrogram inputs. A specific diffusion nois...	Rebuttal 1: Rebuttal: First thank you very much for the review and taking the time to point at typos and grammar issues. We have addressed all of your comments and will conduct a thorough proofreading of the paper to ensure that no error appears in the camera-ready version. Regarding the computational cost: We wi...	Rebuttal 1: Rebuttal: We want to make an ethical statement, the AudioGen and MusicGen team recently released the new public implementations with pre-trained models. We trained MultiBand Diffusion with those new versions of the compression models. We found improvement on MusicGen using MultiBand Diffusion as a decoder. ...	NeurIPS_2023_submissions_huggingface	2,023	Summary: This paper presents a novel approach to processing audio data by developing and implementing a band-specific diffusion model, a frequency equalizer processor, and a power noise scheduler. The band-specific diffusion model processes various frequency bands independently, thereby reducing the accumulation of ent...	Rebuttal 1: Rebuttal: Thank you very much for your review. Baselines: We conducted a new subjective studies (MUSHRA) including new baselines such as a discrete version of HifiGan used in https://arxiv.org/abs/2104.00355 and Priorgrad (https://arxiv.org/abs/2106.06406) both conditioned on EnCodec units at 6 kbps. ...	null	null	null	null	null	null
Constrained Policy Optimization with Explicit Behavior Density For Offline Reinforcement Learning	Accept (poster)	Summary: The paper proposes a novel algorithm for offline model-free reinforcement learning by constraining the actor to lie within a safe area defined by an explicit behavioral density. This behavioral density is obtained by thresholding a FlowGAN model. The resulting algorithm is competitive across various D4RL datas...	Rebuttal 1: Rebuttal: Thank you for your review. We appreciate your kind suggestions for improving our manuscript. Response to Weakness 1: We agree that MORel is a useful baseline for CPED, and we will certainly include the MORel[1] in our manuscript as baseline method. From our understanding, MoReL provides us ...	Summary: This paper proposes a novel approach named Constrained Policy optimization with Explicit Behavior density (CPED) to mitigate the existing policy control methods, e.g., overly conservative or failing to identify OOD areas accurately. CPED use a flow-GAN model to explicitly estimate behavior policy density. The ...	Rebuttal 1: Rebuttal: We thank the reviewer for their efforts and time reviewing our paper. Response to Weakness 1: Thank you for your comment. We want to emphasize that the original intention of CPED is not to directly apply the Flow-GAN idea to offline RL problem for the following reasons. 1. **Advantage o...	Summary: This paper proposes a new offline RL algorithm called Constrained Policy optimization with Explicit Behavior density (CPED). The main idea is utilizing a flow-GAN model to estimate the probability density explicitly, which limits the exploration in the safe region. To demonstrate the effectiveness, the author ...	Rebuttal 1: Rebuttal: Thank you for your review. Your comments are very useful for refining our manuscript. Response to Weakness 1:Thank you very much for this important question. **We want to emphasize that the original intention of CPED is not to directly apply the Flow-GAN idea to offline RL problem or to simpl...	Summary: This work proposes to use flow-GAN model to explicitly model the distribution of the behaviour policy and use that to alleviate the OOD actions problem in offline RL. A theoretical analysis of convergence is presented. Empirical evaluation on d4rl gym locomotion and antmaze show that the proposed CPED algorith...	Rebuttal 1: Rebuttal: Thank you so much for your review. We appreciate your valuable questions and suggestions. We have replied your comments in detail, and please see the following for your response. Response to Weakness 1: We agree that the proposed CPED introduces an additional generative modeling module in t...	Rebuttal 1: Rebuttal: We are grateful for the valuable questions and suggestions were given by all four reviewers that help us to revise our manuscript. After reading all the reviews, we have answered each reviewer's questions in detail, and also added relevant experiments based on each reviewer's suggestions to suppor...	NeurIPS_2023_submissions_huggingface	2,023	null	null	null	null	null	null	null	null
Double and Single Descent in Causal Inference with an Application to High-Dimensional Synthetic Control	Accept (poster)	Summary: This paper investigates over-parametrized models in causal inference, specifically focusing on high-dimensional linear regression models and high-dimensional synthetic control estimators with a large number of control units. The authors examine the prediction risk behaviors associated with these two estimators...	Rebuttal 1: Rebuttal: Thank you for your comments and suggestions. 1. For the concerns you lay out: - We believe that important next steps on this agenda include (i) developing low-level sufficient conditions under which more complex synthetic control performs better than less complex synthetic control, (i...	Summary: The paper studies the issue of double and single descent for linear regression as well as synthetic control. For the former, empirical motivation is given in the context of predicting wages and theoretical results are given from the perspective of model averaging. For the latter, empirical motivation is given ...	Rebuttal 1: Rebuttal: Thank you for your thoughtful comments and suggestions. 1. For the point about bias, we agree that our analysis does not separately provide results for the bias of the estimator, which will generally depend on the relationship of the estimator to the data-generating process. We note, however,...	Summary: This paper examines single and double descent phenomena in two causal inference estimators: high-dimensional linear regression and synthetic control estimators with many controls. To begin, the paper starts with a high-dimensional linear regression problem and illustrates the double descent phenomena using the...	Rebuttal 1: Rebuttal: Thank you for your detailed and thoughtful comments and questions. 1. In response to your question about reasonable comparison models, we are reporting additional comparisons: - For linear regression, we report a comparison to low-dimensional linear regression using the original covari...	Summary: The paper studies overparameterized linear regression in the context of imputing data for estimation of causal effects. This involves both unconstrained regression for imputing wages in the CPS dataset and regression under a simplex constraint on the weights for synthetic controls in the California smoking rat...	Rebuttal 1: Rebuttal: Thank you for your thoughtful comments. 1. The question under which conditions we obtain better generalization goes to the heart of the issue. While the model-averaging property holds mechanically for these estimators, the resulting generalization error bounds (like the one in _Proposition 5_...	Rebuttal 1: Rebuttal: We are attaching a one-page PDF with two additional figures: _Figure R1_ presents the performance of linear regression in estimating average treatment effects in samples of varying size, and _Figure R2_ presents the performance of high-dimensional synthetic control with ten pre-treatment periods. ...	NeurIPS_2023_submissions_huggingface	2,023	null	null	null	null	null	null	null	null
SwiFT: Swin 4D fMRI Transformer	Accept (poster)	Summary: This paper extends Swin Transformer to 4D brain functional MRI data. Unlike existing deep learning methods in the neuroscience field that either use an ROI-based method (taking as inputs functional connectivity) or a two-step method (3D spatial encoding followed by temporal encoding), the proposed model SwiFT ...	Rebuttal 1: Rebuttal: Thank you for the review and giving us some helpful pointers. Here is our response to the questions and concerns: >*Due to memory constraint, the fMRI sequence needs to be divided into shorter sub-sequences and the model predictions for sub-sequences are aggregated through averaging. In my opinio...	Summary: This paper presents SwiFT (Swin 4D fMRI Transformer), a novel Swin Transformer architecture designed for modeling spatiotemporal brain dynamics from high-dimensional 4D functional MRI data. The architecture incorporates a 4D window multi-head self-attention mechanism and absolute positional embeddings, making ...	Rebuttal 1: Rebuttal: Thank you for the review and raising some important questions. Here is our response: >The paper lacks a proper discussion of its limitations... - Although our model has demonstrated high performance and efficiency compared to existing models, it still presents certain limitations for neuros...	Summary: The paper focuses on the expansion of the Swin Transformer to a 4D version, enabling it to be trained on 4D fMRI data in an end-to-end manner. Specifically, the proposed method commences by constructing an absolute positional embedding layer across spatially neighboring patches, prior to implementing deep 4D S...	Rebuttal 1: Rebuttal: Thank you for the review and providing us with helpful suggestions. Here is our response to some of the questions and concerns raised: > Swin Transformers primarily restrict self-attention computations to specific sub-windows... - We agree that it is important to address the issue. Please refe...	Summary: * The authors seek to develop an approach for modeling spatiotemporal brain dynamics, as measured with resting state functional MRI. For this, they have extended a transformer-based architecture to handle time-varying 3D scan data. * They seek to predict subject traits and attributes like age, sex, and cognit...	Rebuttal 1: Rebuttal: Thank you for raising some important questions and providing us with helpful feedback. Here is our response to the questions and concerns raised: >The novelty of the architecture ... - The novelty of our architecture may seem to be limited, however there are several points that we claim to be ...	Rebuttal 1: Rebuttal: We thank the reviewers for taking their time to review our work and provide constructive feedback that would help us improve our work. Here are our responses to common questions and concerns raised by multiple reviewers. ### Our computational resources (for reviewer bLJJ, pTkQ) - We have used f...	NeurIPS_2023_submissions_huggingface	2,023	Summary: This paper introduces SwiFT, a novel Swin Transformer-based architecture for analyzing high-dimensional functional brain MRI data. SwiFT learns end-to-end to process spatiotemporal brain functional data, and it achieves state-of-the-art performance on several tasks, including sex classification, age prediction...	Rebuttal 1: Rebuttal: Thank you for the detailed analysis of our work and for the insightful suggestions. Here we address some of the questions and concerns raised: > ...It would be interesting to see how SwiFT compares with more types of models... - Thank you for providing us with some pointers. To faithfully eva...	null	null	null	null	null	null
Keypoint-Augmented Self-Supervised Learning for Medical Image Segmentation with Limited Annotation	Accept (poster)	Summary: This paper tackles self-supervised learning for medical image segmentation. The main contribution is the keypoint identification in input images in self-supervised pre-training to consider long-range spatial dependencies. Each UNet encoder layer is followed by a keypoint-augmented fusion (KAF) layer, where key...	Rebuttal 1: Rebuttal: > While contributions are novel, literature review is missing a lot of works on self-supervised medical image segmentation via global/local similarity/contrastive learning. > Experimental comparisons against Wang et al. 2022 would further strengthen the novelty. Thank you for offering additiona...	Summary: This paper focuses on improving medical image segmentation performance with limited labeled data. They (1) propose a modification to the standard UNet convolutional architecture by incorporating local features (at locations identified by SIFT) into the backbone, (2) show how their modification can be used to c...	Rebuttal 1: Rebuttal: > The GLCL baseline reported in Table 1 is much lower than the numbers reported in the GLCL original paper; the original GLCL paper actually reports performance better than the proposed method on the ACDC dataset. Please comment on why the reported GLCL performance in 14-18 points lower than the G...	Summary: SIFT keypoints are used in conjunction with a UNET to construct a transformer-like model for medical image segmentation. Global and local self supervised learning (SSL) losses are proposed. Correspondences identified using the SuperGlue method. Various experiments on two cardiac image datasets, including ablat...	Rebuttal 1: Rebuttal: > It would be interesting to know if 3D keypoints would improve results. Thank you for the insightful point! As discussed in our limitations section, a promising future avenue is expanding our model to 3D. This direction, however, poses unique challenges. Firstly, 3D keypoint detection lacks esta...	Summary: The authors introduce an approach to improve the performance of the UNet model, specifically in the task of segmentation. They propose a keypoint augmented fusion layer to enhance the CNN-based encoder of UNet. This layer leverages self-attention to capture long-range spatial dependencies among keypoints extra...	Rebuttal 1: Rebuttal: > Comparison with transformer-based UNets. Thank you for your suggestions! We now added two recent SoTA transformer-based unets including Swin-Unet[1] and Swin-UNETR[2]. To illustrate the effects of architectural changes, we train all models with random weight initialization, under limited annota...	Rebuttal 1: Rebuttal: We thank the reviewers for their time and expert feedback. As a summary of the reviews, we were happy to see that the submission was found to be well-written [`b18J,PaeV`], and it presents a sound and novel [`b18J,9qHo,m84B,8g4p,PaeV`] method for local self-supervised learning, with well form...	NeurIPS_2023_submissions_huggingface	2,023	Summary: The paper presents a self-supervised learning approach for medical image segmentation that uses matching keypoints to contrast local features that may be spatially distant from one another. The proposed model pretrains a segmentation network on unlabelled images, using three contrastive losses: a global loss o...	Rebuttal 1: Rebuttal: > Comparison with other local SSL methods. Thank you for offering additional literature. We will add them in our paper. To clarify the difference between our method and `Peng et al., Zhong et al.`, `Chaitanya et al.`: while they incorporate local losses, their frameworks are designed for *semi...	null	null	null	null	null	null
Data Augmentations for Improved (Large) Language Model Generalization	Accept (poster)	Summary: This paper focuses on counterfactual data augmentation for training robust machine learning models on text data. The concerned applications such as healthcare are safety-critical, highlighting the importance of this work. To generate data, the authors propose to employ a Large Language Model (LLM) to rewrite t...	Rebuttal 1: Rebuttal: Thank you for reviewing our paper and recognizing positive points on the importance of the topic we study, the achieved empirical performance and writing quality. In what follows we aim to address each question and concern in your review. Limited novelty: The novelty of our work, besides the ...	Summary: The paper addresses the issue of text classifiers relying on spurious correlations, leading to poor generalization vis-a-vis out of domain, particularly in critical areas like healthcare. The authors propose using counterfactual data augmentation, guided by knowledge of causal data structure, to create text cl...	Rebuttal 1: Rebuttal: Thank you for the in-depth review of the paper and for raising important questions that help us improve it. Please find our responses to the concerns and questions below. Knowledge of Causal Structure: While our method assumes some knowledge of causal structure (i.e. Figure 1, an anti-causal ...	Summary: This paper develops and analyzes the counterfactual data augmentation strategies for the setting where the data-generating process is anti-causal. First, they first show the Bayes optimal predictor in the setting they considered and then how can we obtain that predictor using counterfactual data augmentation. ...	Rebuttal 1: Rebuttal: Thank you for taking the time to review our paper. We greatly value your feedback and would like to address each comment you’ve provided while highlighting the essential aspects crucial for evaluating our work. Distinction from previous work: Our contribution stands out first and foremost by ...	Summary: The authors propose to use knowledge from the causal structure of the data to counterfactually simulate interventions on spurious features and to learn more robust classifiers. They focus on text classification tasks and argue that this approach is appropriate in prediction problems for which the label is spur...	Rebuttal 1: Rebuttal: We appreciate your review and questions, and are eager to clarify all points regarding our data, baselines, and assumptions. In response to questions: Causal vs. anti-causal prediction problem: As you correctly suggest in the review, we solve an anti-causal prediction problem. A true condit...	Rebuttal 1: Rebuttal: We thank the reviewers for their efforts and valuable comments, which led us to make important revisions to the paper, particularly by adding two additional baselines to our experiments (IRMv1 and GroupDRO), clarifying the connections of our work to previous papers on invariant learning and counte...	NeurIPS_2023_submissions_huggingface	2,023	Summary: In this work, the authors develop causally-driven data augmentation methods to improve model robustness. Strengths: In general, I find the paper to be very, very well written and clear. The authors do a really great job of explicitly stating their assumptions, and acknowledging when certain assumptions are st...	Rebuttal 1: Rebuttal: Thank you for the insightful and positive review. We are happy to see that you think the paper is a very interesting first step in utilizing recent advances in LLMs for OOD generalization, which is exactly our goal in this work. The points for improvement that you lay out are very helpful, and we...	null	null	null	null	null	null
On the Convergence of Black-Box Variational Inference	Accept (poster)	Summary: The paper proves convergence results for black-box variational inference (BBVI) with ordinary stochastic gradient descent (SGD) and proximal SGD under several assumptions: it considers only the reparameterization gradient setting with the location-scale variational family (in particular, mean-field and Cholesk...	Rebuttal 1: Rebuttal: Thank you for review. > The main weakness of the paper is how it deals with the limitations: while it is good that the assumptions are highlighted in the text, there are more limitations that should be pointed out. For example, the paper only deals with reparameterization gradient BBVI, which is ...	Summary: The authors analyze the smoothness and convexity of the ELBO under different parameterizations (linear vs nonlinear) of the scale for location-scale families, building on the work of Domke (2020). This enables convergence analysis for BBVI with standard and proximal stochastic gradient descent. Their main find...	Rebuttal 1: Rebuttal: Thank you for your review. > The story is a bit disjointed. Assumption 4 and Theorem 3 might be better placed directly after Example 4 to complete a story about convergence of BBVI for nonlinear and general nonconvex $f$ that satisfies the smoothness conditions (among others). This might have e...	Summary: This paper tries to prove convergence of BBVI algorithm in a more general setting compared to previous work: the main being that the target needing to be only log-smooth and not needing to be log-concave, which means that the objective can be non-convex. The paper claims that non-linear parameterizations of th...	Rebuttal 1: Rebuttal: Thank you for your review. > I think, the paper is trying to unify many different parameterizations, and so at times the paper has become a bit harder to read We agree that the paper is quite dry in the submitted form. Due to the page limit, we had to remove many of our results' commentaries. We...	Summary: The paper establishes the first complete convergence result of BBVI as it is used in practice. It is shown that a linear parametrization of the covariance leads to better convergence rates, which is confirmed in some numerical experiments. Finally, a proximal version for BBVI due to Domke et al 2020 is analyze...	Rebuttal 1: Rebuttal: Thank you for your review. > The paper missed to cite the recent line of works Thank you for the suggestions. We will add them to the next version. > It is known that BBVI is inferior to natural-gradient algorithms which exploit the KL / Fisher-geometry (https://arxiv.org/abs/2107.04562). This...	null	NeurIPS_2023_submissions_huggingface	2,023	null	null	null	null	null	null	null	null
LMC: Large Model Collaboration with Cross-assessment for Training-Free Open-Set Object Recognition	Accept (poster)	Summary: In this paper, the authors tackle the problem of open-set object recognition. The key challenge in open-set recognition problem they tackle is the model's reliance on spurious-discriminative features. To tackle this challenge, they propose an open-set object recognition system consisting of multiple pre-traine...	Rebuttal 1: Rebuttal: >Q1: Technical contribution. A1: In this work, we propose a novel framework that collaborates large models and performs training-free open-set object recognition. Moreover, to effectively extract implicit knowledge from large models, we also incorporate several novel designs, such as iterat...	Summary: This work demonstrates a very interesting and sophisticated usage of various large models, including ChatGPT, CLIP, DINO and DALL-E., to tackle the open-set recognition problem. The main challenge in open-set recognition is that the chosen classifier may get confused by certain spurious-discriminative features...	Rebuttal 1: Rebuttal: >Q1: Pave a new way for follow-up research. A1: Thank you for pointing out our method is an interesting way. Below, we also discuss potential follow-up research in solving open-set recognition (OSR) that could be inspired by our work. In our method, we identify suitable large models, ...	Summary: This paper proposed to incorporate several large models to solve the spurious-discriminative problem in open-set recognition. It first prompts ChatGPT to describe the known classes and generate a list of new classes that share the spurious-discriminative features with the known classes. Then it prompts DALL-E ...	Rebuttal 1: Rebuttal: >Q1: May consider citing the related papers [1,2]. A1: We will cite [1,2] in the paper. Note that different from [1,2] which fuse the uncertainty score from two aspects (e.g., from FS-KNN and Softmax), we are the first to collaborate large models and leverage their rich and distinct knowled...	Summary: This paper proposes an open-set recognition framework, Large Model Collaboration (LMC), which collaborates several large models (ChatGPT, DALL-E, CLIP, and DINO) to make use of the rich implicit knowledge to reduce the reliance on spurious-discriminative features. The proposed framework consists of the followi...	Rebuttal 1: Rebuttal: >Q1: Comparisons under the cross-dataset setup. A1: Below, we evaluate our method under a common cross-dataset setup following [15,36,45,55,63]. In this setup, closed-set-class images are from CIFAR-10, and open-set-class images are from ImageNet-crop, ImageNet-resize, LSUN-crop, and LSUN-r...	Rebuttal 1: Rebuttal: We thank all reviewers for recognition of our contributions (Reviewer Kpbj:"novel and interesting"; Reviewer Qabu:"the overall idea is new and interesting"; Reviewer BEEJ:"demonstrates a very clever usage of several off-the-shelf large models of different modalities", "possesses many advantages, s...	NeurIPS_2023_submissions_huggingface	2,023	null	null	null	null	null	null	null	null
Cookie Consent Has Disparate Impact on Estimation Accuracy	Accept (poster)	Summary: Cookies enable accurate identification and tracking of user behavior, leading to personalized ads and better ad campaign performance. However, this raises privacy and fairness concerns. This study investigates the impact of user consent on a recommender system's ability to learn about users' demographics and p...	Rebuttal 1: Rebuttal: Thanks for the review. - Regarding your comment that our paper does not contain a solution, our view is that there is value in drawing attention to an issue without necessary having a fix. We’ve proposed changes to industry practice that should be made, along the lines of ensuring that how much a...	Summary: * This paper aims to investigate the effect of cookie consent (as mandated by GDPR, ePD), on the accuracy of recommendation algorithms using a simulation study. * Model: * Each advertisement $a$ is characterized by a topic $\tau \in \mathcal{T} \cong [m]$. * Each user $i\in[n]$ described by a cookie $\phi_...	Rebuttal 1: Rebuttal: We appreciate your careful review. We hope that our responses below resolve some of your concerns. Impact of recommendation algorithm on accuracy disparity: - The impact of different recommendation algorithms on the observed disparity differs depending on how the algorithm deals with missing inf...	Summary: This paper simulates a recommendation system that is aware of and responds to users decision to share cookies, along with data from users' clicks. In this simulation, the paper finds that the recommendation accuracy could be higher for users who do not consent to provide cookies under certain conditions. The p...	Rebuttal 1: Rebuttal: We appreciate your detailed review of our paper. We’ve responded to each of your queries (either in this response or in the global rebuttal). Response to: "Complex simulation setup, results possibly driven by unrealistic recommendation model": - Practically, the set of recommendable ads may be...	Summary: This paper investigates the influence of varying consent rates on the accuracy of estimation. The paper utilizes a simulation-based approach, in which the decision to share cookies plays a significant role. The findings reveal interesting patterns that can contribute to the development of more refined fairness...	Rebuttal 1: Rebuttal: Thank you for your review. We've provided some answers to your questions below. The need for simulations: - While the priors do indeed dictate the relative accuracy disparities across cohorts and user consent decisions, the evolution of these errors as a consequence of the recommendation process...	Rebuttal 1: Rebuttal: Real-world relevance of the studied model: - We appreciate the reviewers’ concerns regarding the relevance of the model to real-world settings. We’d like the emphasize that the goal of this investigation is to present an intentionally simplified model in order to isolate and understand the effect...	NeurIPS_2023_submissions_huggingface	2,023	null	null	null	null	null	null	null	null
CP-SLAM: Collaborative Neural Point-based SLAM System	Accept (poster)	Summary: The paper presents a collaborative implicit neural simultaneous localization and mapping (SLAM) system for RGB-D image sequences, which is able to merge sub-maps into a global map and also handle loop-closure for neural mapper for the first time. To facilitate the capability of doing loop-closure and sub-map m...	Rebuttal 1: Rebuttal: Thank you for your careful review and valuable suggestions. Here are our responses to your comments. 1. To demonstrate advantages from both collaboration and loop-closure capability, evaluation in this paper becomes more challenging. (W1) In CP-SLAM, loop detection and pose graph optimizatio...	Summary: This paper proposed a collaborative RGB-D neural implicit SLAM method named CP-SLAM. CP-SLAM is based on neural point representation similar to PointNeRF, which is natually suitable for collaborative SLAM and loop closing. Based on this neural point representation, CP-SLAM is able to perform full SLAM with a s...	Rebuttal 1: Rebuttal: Thanks for your constructive comments and recognition of our strengths. Our responses to your concerns are as follows: 1. Some components in CP-SLAM are incrementally modified based on existing works. (W1) We aim to build a complete collaborative NERF-based SLAM system, focusing more on syst...	Summary: The authors propose a novel collaborative RGB-D SLAM pipeline, where the map is represented by an implicit field based on Point-NERF that are transformed together with their keyframes allowing pose graph optimization. They use a loop closure detection based on learned features (NetVLAD) also for sub-map alignm...	Rebuttal 1: Rebuttal: Thank you for your valuable comments. Here are our responses to your concerns. 1. What does it mean to incrementally project a pixel to a 3d location in line 101? (W1) In the SLAM task, both the map and trajectory are incrementally constructed over time. In Line 101, when a new mapping frame...	Summary: The paper introduces a neural implicit representation based collaborative SLAM technique. They introduce a learnable feature encoding for each point in the 3d scene based on neural rendering subsequently using it for depth and color map estimation. They introduce a centralized learning strategy to share MLPs a...	Rebuttal 1: Rebuttal: Thank you for your valuable comments! Here are our responses to your concern. 1. It would be nice to see evaluations on more datasets. (W1) We extended experiments on the TUM-RGBD real-world dataset, and the results are shown in the Tab.G3 in global response. Following recently proposed ...	Rebuttal 1: Rebuttal: We thank all the reviewers for their valuable comments and appreciate the positive comments like “the contributions to the pipeline are significant” (Reviewer Zpdr), "CP-SLAM introduces a novel neural point-based representation, enabling dense collaborative neural SLAM." (Reviewer 1EKn), "...	NeurIPS_2023_submissions_huggingface	2,023	Summary: The paper presents CP-SLAM, a collaborative neural implicit SLAM system developed by the authors. CP-SLAM incorporates neural point-based odometry, loop detection, sub-map fusion, and global refinement. Additionally, the authors propose a novel neural point 3D scene representation with keyframes to enhance map...	Rebuttal 1: Rebuttal: Thank you for your careful reading and constructive comments on our paper, and noting the novelty of our CP-SLAM system and its superiority. Here are our responses to your comments. 1. Large viewpoint change is challenging for NERF-based SLAM methods. (W1) As pointed out in Section.5 in our ...	Summary: There have been some papers on getting dense Visual SLAM with neural fields published in the last couple of years. In this work, the authors proposed a set of additions to prior pipelines that get the NN SLAM closer to the traditional approaches concerning the overall framework. Specifically, the authors addre...	Rebuttal 1: Rebuttal: Thank you for your valuable comments and for appreciating our contributions and writing. Here are our responses to your comments. 1. The paper focuses more on adding new steps to prior Visual SLAM pipelines. (W1) We aim to build a complete collaborative NERF-based SLAM system, focusing mo...	null	null	null	null
Separable Physics-Informed Neural Networks	Accept (spotlight)	Summary: This paper introduces a novel approach to using separable neural networks to represent solutions of partial differential equations (PDEs). In contrast to the conventional form used in physics-informed neural network (PINN) methods, the specialized form of neural networks proposed in this work takes advantage o...	Rebuttal 1: Rebuttal: R: It is unfair to test SPINN on solutions that align with the variable separation form. A: We tested on manufactured solutions because this is the most conventional way of making an exact reference solution. As you can see in the diffusion and (2+1)-d Navier-Stokes equation, SPINN is still ...	Summary: This paper presents a novel variant of PINN for learning multi-dimensional PDEs. Unlike traditional approaches considering point-wise inputs, this method incorporates each axis information separately during the learning process and leverages tensor decomposition to interpret the predicted solution variables. A...	Rebuttal 1: Rebuttal: R: The primary concern of PINNs is the optimization issue. Therefore, the contribution is moderate. A: We believe that scalability is also an important issue to tackle in PINNs’ literature. In our Navier-Stokes experiment, where the solution shows complex and turbulent behavior, previous stu...	Summary: The paper proposes a methodology for significant savings in computation and memory in learning physics informed networks for approximately solving PDEs. This enables learning with a much larger number of points, resulting in better accuracy. The idea is to use a specific function class, where features are cons...	Rebuttal 1: Rebuttal: R: Discussion of why SPINN delivers more improvements A: Our performance enhancement mainly stems from the adoption of large-scale collocation points. This is particularly effective when dealing with PDEs where the solution is complex, such as a solution that contains high-frequency componen...	Summary: The paper proposes a new class of architecture called Separable PINN (SPINN) that operates on a per-dimension basis instead of the standard coordinate based MLP architecture. This architecture significantly reduces the computational complexity and allows the use of a large number of collocation points to enfor...	Rebuttal 1: Rebuttal: R: Computational stability and gradient vanishing/exploding problems A: We have tested on the (5+1)-d system to check whether our model can handle higher dimension PDEs (please see section G.1 in the supplementary material). Additionally, we acknowledge the potential numerical instability in...	Rebuttal 1: Rebuttal: ## For all reviewers We thank the reviewers for taking their time to suggest all the invaluable comments and constructive feedback. We attached a PDF file to show tables and visualizations of additional experiments conducted during the rebuttal period. Please see our responses below. Pdf: /...	NeurIPS_2023_submissions_huggingface	2,023	Summary: In this paper, the authors propose two main contributions: (i) forward mode AD for PDEs (ii) separating the contributions from each dimension (just like in separable conv) and performing computation using tensor multiplication. The results naturally show significant improvement over PINNs. Strengths: + Signi...	Rebuttal 1: Rebuttal: R: Expressions when stating our contribution A: In the related work section, we mentioned that we utilized forward-mode AD in training PINNs (lines 108-109), but regrettably, this detail was inadvertently omitted in the conclusion section (line 348). Thank you for pointing this out. We wil...	null	null	null	null	null	null
Fed-CO$_{2}$: Cooperation of Online and Offline Models for Severe Data Heterogeneity in Federated Learning	Accept (poster)	Summary: In this paper, the authors focus on tackling the feature skew issue, and introduces a new approach that utilizes knowledge distillation. Specifically, the proposed method transfers the knowledge between the FL online model and the local offline model to enable them to have both domain invariant and domain spec...	Rebuttal 1: Rebuttal: >Q1: Lack of important baselines: a. Ditto: Since Ditto is a very classic pFL method, it should be included as the baseline. b. FedMask [1]: There is no hypernetwork based method that directly personalizing parts of the model parameters. FedMask is one of the SOTA method, which learns distinct...	Summary: The paper proposes a personalized federated learning (PFL) method to handle both label and feature distribution skew.Specifically, each clients holds a partially perosnalized model (personalization happens at the level of batch-normaizliation layers), and a fully personalized local model. The paper conducts a...	Rebuttal 1: Rebuttal: >Q1: The paper does not provide stong explantion to motivate the proposed method. In particular, it is unclear where (6) and (7) come from, and why they take this specific form. R1: The motivation of our proposed method is answered together with Q3. Formulas (6) and (7) describe the gener...	Summary: The paper presents Fed-CO2, a novel framework for Federated Learning (FL), a distributed learning paradigm that allows multiple clients to collectively learn a global model without sharing their private data. FL's effectiveness depends heavily on the quality of the data used for training. Data heterogeneity is...	Rebuttal 1: Rebuttal: >Q1: The model may be complex to implement, especially in environments where the number of clients is very large, due to the dual model structure (online and offline) and the knowledge transfer mechanisms. R1: We partially disagree with this opinion. Our method will be easy to implement i...	Summary: This paper introduces an algorithm to address both label and feature heterogeneity in federated learning. The algorithm entails that each client learns two models: one that has personalized batch normalization parameters and otherwise global parameters, and another that is fully personalized. To learn the para...	Rebuttal 1: Rebuttal: >Q1: The writing can be improved. There are numerous sentences with improper grammar/semantics, e.g. “Federated learning for Data Heterogeneity of Feature Skew”. R1: We will improve improper grammar/semantics in the final version. >Q2: The model parameter notation is not consistent...	Rebuttal 1: Rebuttal: We thank the reviewers for their careful reading of our paper and help with improving our manuscript. We sincerely appreciate that you find our work proposes 'a novel and effective approach' (Reviewer 5E9k, Reviewer NQAC), addresses 'both label distribution skew and feature skew within a unified f...	NeurIPS_2023_submissions_huggingface	2,023	null	null	null	null	null	null	null	null
Online Inventory Problems: Beyond the i.i.d. Setting with Online Convex Optimization	Accept (poster)	Summary: In this paper, the authors consider the classic online inventory control problem. While previous works mainly consider the case where the demand is stochastically drawn from a stationary distribution, this paper considers the case where the demand can be adversarial. In addition, the authors consider a more ge...	Rebuttal 1: Rebuttal: 1. *One major concern is the novelty of the proposed algorithm. From the technical perspective, I think the analysis generally follows the classic analysis of online gradient descent with a self-confidence tuning of the learning rate in order to handle the issue raised by the scale of the gradient...	Summary: The paper studies a general online decision making problem for inventory management. The problem setting extends the generic online convex optimization by introducing structures such as states and demands. Compared to settings studied by related works in inventory management, the setting of this paper relaxes ...	Rebuttal 1: Rebuttal: 1. One aspect I'm not sure is the novelty. ; The backbone of the proposed algorithm is a batch-update version of online gradient descent ; the novelty of the paper is mainly on the combination of OGD with inventory structures We address this in our global response. 2. *Has the geometric cy...	Summary: The paper introduces the online inventory optimisation problem. In the problem, we have $n$ products to stock and every day choose the levels to stock up the products to. We then suffer loss according to a convex loss function, which may vary from day to day (e.g., it may depend on the demand). We also get to ...	Rebuttal 1: Rebuttal: 1. The algorithm and the pound in the paper are an extension of the online subgradient descent (OSD). One can argue that the result is of the incremental nature. Our answer : we address this in the global response. 2. *A constant strategy will only work if the constant is above the remaind...	Summary: The paper considers a general family of online inventory management problems under convex losses and stateful inventory dynamics. They propose a general algorithm based on an adaptation of ideas from online convex optimization. The method is essentially AdaGrad enhanced with an additional delayed update mechan...	Rebuttal 1: Rebuttal: 1. *I worry that this comparator strategy is problematic in the stateful setting. First off, some fixed constant strategies may be infeasible when some of the carryover inventory levels x_t are large. Thus, the best feasible constant strategy may be a very poor one, ordering very high levels to ma...	Rebuttal 1: Rebuttal: We will use this general response to address issues which were raised simultaneously by several reviewers. We responded to all the other issues in the individual responses to reviewers. 1) *Reviewers 5kk9, UtoR, ZTux raised concerns on the fact that the constant competitor in the regret definitio...	NeurIPS_2023_submissions_huggingface	2,023	null	null	null	null	null	null	null	null
Supervised Pretraining Can Learn In-Context Reinforcement Learning	Accept (spotlight)	Summary: This paper introduces a new pretraining objective for decision-making tasks. Leveraging the in-context learning capabilities of transformers, the authors propose to pretrain a model that predicts the optimal action given the current state and a set of history interactions as context. Empirically, the pretraine...	Rebuttal 1: Rebuttal: Thanks for detailed feedback and questions! Optimal actions at pretraining.: Indeed, the theoretical formulation of DPT requires that task-specific optimal actions can be obtained for pretraining. While restrictive in some cases, we actually view this as being potentially quite flexible. This...	Summary: This paper proposes an in-context learning based algorithm for meta-learning decision-making algorithms, combined with a novel analysis which shows that unlike prior works, the proposed algorithm is guaranteed to work. Moreover, the authors also show empirically and theoretically that their algorithm is able t...	Rebuttal 1: Rebuttal: Thank you for the positive comments and detailed feedback! Can you relax the consistency assumption on the learned model for the theory?: Thank you for the great question. In Appendix C, we provide some justification for this via a generic MLE concentration inequality (cf. Zhang, 2006). This...	Summary: This paper studies in-context learning for sequential decision-making problems. It proposes an approach, Decision Pretrained Transformer (DPT), which takes as input a (state, context) pair and produces an action. It is pretrained on a large number of tasks drawn from some task distribution and evaluated on hel...	Rebuttal 1: Rebuttal: Thank you for your positive comments and detailed feedback! How does the computational cost of the proposed method compare to that of the baselines? Due to the need to perform attention… this might be expensive: AD also uses a GPT-like architecture, so its computational cost is approximately ...	Summary: This work leverages transformers to solve decision making problems in a few-shot manner. In particular, they train transformers to predict optimal actions from a task given a query state and a few-shot dataset of state, action, reward transitions. The approach is tested in both multi-armed-bandit problems and...	Rebuttal 1: Rebuttal: Thank you for your detailed feedback! Novelty: We apologize this was not clear. DPT indeed offers significant novel insights, distinct from prior work. 1. It is not necessary to start with a good RL algorithm to learn a good RL algorithm. Supervised pretraining can produce one. Prior work ha...	Rebuttal 1: Rebuttal: Dear Reviewers and AC, Thank you for your detailed and thoughtful reviews. We are delighted to hear that the reviewers found the problem and results interesting and novel (R3a7, pHmJ), and the analyses well-designed and thorough (pHmJ, CX5S). We share excitement about the insights such as the con...	NeurIPS_2023_submissions_huggingface	2,023	null	null	null	null	null	null	null	null
This Looks Like Those: Illuminating Prototypical Concepts Using Multiple Visualizations	Accept (poster)	Summary: This paper proposes a new interpretable prototype-based classifier, called ProtoConcepts. Unlike existing prototype-based classifiers that use one-to-one comparisons, ProtoConcepts learns prototypical concepts using multiple image patches. This approach aims to make it easier to identify the underlying concept...	Rebuttal 1: Rebuttal: Thank you for your review comments. We are happy to make some clarifications as follows: >It is definitely not a new idea in the XAI community to use multiple images/patches to visualize a concept. This is a common practice in the works of concept-level explanations, like TCAV, ACE, and so on. How...	Summary: The paper proposes a modification of the prototype layer of the existing prototype-based networks. Existing methods rely on a single training image patch as a prototype. The main drawback of this is that it is difficult for the user to understand the meaning of the prototype from a single visualization. The pa...	Rebuttal 1: Rebuttal: Thank you for your review comments. We are happy to make some clarifications as follows: >However, the experiments do not report one or more metrics that show this reduction of ambiguity with respect to the plain prototypical networks. A: As per your suggestion, we evaluated the added interp...	Summary: This paper extends the patch comparison of prototypical part-based classifiers from one-to-one patch to one-to-multiple (e.g. from this looks like that to this looks like those). They compare their proposed network with Proto-PNet variants and show no classification accuracy improvements and little evidence of...	Rebuttal 1: Rebuttal: Thank you for the review. However, we would like to make some clarifications as following: >I believe maybe this work also fails to help humans [1] as its siblings [2,3]. The authors could improve the interpretability analysis by running more extensive study (both automatic and human studies) rath...	Summary: This paper improves interpretability for prototypical learning. Most previous methods use a single prototype as an interpretation -- and this is not as informative as we don't know what portion of the prototype corresponds to our target image. The authors introduce a prototypical ball to interpret what pro...	Rebuttal 1: Rebuttal: Thank you for your review comments. We are happy to make some clarifications as follows: > Why are these selected as prototypes? A: These prototypes are learned. The data chooses these prototypes. The interpretation of the prototypes as a concept is done by the human. The algorithm doesn'...	Rebuttal 1: Rebuttal: User Study We thank all the reviewers for their comments. To show the reduction of ambiguity (and resulting improvement in user interpretability), we created a distinction user study similar to HIVE[1] to compare our ProtoPConcepts method with ProtoPNet. We randomly picked ten samples from...	NeurIPS_2023_submissions_huggingface	2,023	null	null	null	null	null	null	null	null
SmooSeg: Smoothness Prior for Unsupervised Semantic Segmentation	Accept (poster)	Summary: The paper tackles unsupervised semantic segmentation which aims to group pixels into semantic clusters without manual annotation. This paper proposes the smoothness prior by enforcing that adjacent features in metric space share the same semantics. The approach relies on features from a self-supervised represe...	Rebuttal 1: Rebuttal: Thank you for taking the time to review our work. ### Why SmooSeg is superior to Kmeans used in STEGO The performance of Kmeans heavily relies on the quality and discriminative power of the features it operates on. In the case of DINO features, as shown in the 2nd column of Figure 5, the Kmeans c...	Summary: The paper presents an unsupervised training method for semantic segmentation models. It trains the network on a loss with two terms: one encouraging smoothness in the segmentation labels, and a "data" term based on self-training. It assumes a neural net architecture with a) an extractor that produces per-pixel...	Rebuttal 1: Rebuttal: Thank you for taking the time to review our work. ### Why the method works What the network is learning from. Please refer to the general response. ### The relative role/contribution of the within-image and between-image comparisons. We apologize for any confusion caused. We clarify that ou...	Summary: This work addresses the problem of unsupervised semantic segmentation. In contrast to STEGO (baseline), the approach learns semantic embeddings directly in a student-teacher regime without the need for the K-means. The objective is entirely unsupervised and is reminiscent of a CRF energy formulation with data ...	Rebuttal 1: Rebuttal: ### Technical similarities to STEGO We acknowledge some technical similarities between our smoothness loss and the correlation loss in STEGO. These loss formulations are not entirely novel and have been seen in various dimensionality reduction methods (e.g., [a]). Essentially, STEGO boils down to ...	Summary: The paper introduces a new approach called SmooSeg for unsupervised semantic segmentation, which aims to segment images into semantic groups without manual annotation. SmooSeg is based on the idea of smoothness: adjacent features in a metric space should share the same semantics. Specifically, it formulates un...	Rebuttal 1: Rebuttal: Thank you for taking the time to review our work. Some of the concerns are responded in the "General Response" thread. Please kindly refer to the response regarding "why this method works" and "CRF". ### Roles of the smoothness term and the student-teacher predictor We apologize for any ...	Rebuttal 1: Rebuttal: # General Response We appreciate all reviewers for the highly constructive comments that help improve the paper quality. We also thank the reviewers for the recognition of our work being well structured (1NSo, gyDm) and well presented (iJZU, yftt), the proposed solution being simpl...	NeurIPS_2023_submissions_huggingface	2,023	Summary: The paper tackles the difficult task of unsupervised semantic segmentation at the level of the scene (dense prediction). The authors exploit the piecewise coherence regarding the semantics, texture, or color that similar objects naturally have (termed as smoothness prior). The problem formulation raises some c...	Rebuttal 1: Rebuttal: Thank you for taking the time to review our work. ### Architecture description is tough to follow, the dedicated figure is overcomplicated，and the term "prototypes" is confusing. 1\) We apologize for the confusion on the architecture description and the figure (we guess you are referring to Fig...	null	null	null	null	null	null
Contrastive Sampling Chains in Diffusion Models	Accept (poster)	Summary: This paper analysis why diffusion models need addtional contrastive loss. The main target is the reduce $D_{KL}(p_t\|\| p_t^{SDE})$. This paper provides a detailed theoretical analysis and solid examples to compare its method with existing ones. Strengths: 1. This paper analysis the the gap $D_{KL}(p_0\|\| p_0^{S...	Rebuttal 1: Rebuttal: ## Weaknesses ### 1. Provide further analysis on the role of the second item and clarify how consistent loss and contrastive loss have different influences. This question is quite valuable and may provide us further research direction! The contrastive loss is designed for reducing the distance bet...	Summary: This paper proposes to fine-tune pre-trained diffusion models using contrastive losses to reduce discretization errors. The positive pair is formed by the same image at different steps, while the negative pair is formed by different images. To better optimize the contrastive loss, dynamic weighting schedules a...	Rebuttal 1: Rebuttal: ## Weaknesses ### 1. The term "Contrastive Sampling Chain" is a bit confusing to me, since the proposed method only fine-tunes the diffusion model without changing the sampling process. I would recommend using something like "Contrastive Diffusion Chain". Likewise, the authors should avoid the say...	Summary: This paper focuses on diffusion models (DMs). Current DMs suffer from discretization error since it leverage the numerical solvers to solve SDEs. To overcome this issue, authors propose a a contrastive loss when optimizing DMs. Authors present a theoretical analysis to demonstrate that combining both the gen...	Rebuttal 1: Rebuttal: ## Weaknesses ### 1.Why it use the pretrained model? how about the performance when optimizing it from scratch? I am not convincing about the generalization capability. # It is traditional [1-6] to improve generative performance via fine-tuning pre-trained diffusion models (DMs). For instance, [1,...	Summary: This paper employs the contrastive loss to construct a contrastive sampling chain, which optimizes the KL divergence between the true sampling chain and the simulated chain at each time step to reduce the discretization error associated with numerical solvers used for solving SDEs. Experimental results demonst...	Rebuttal 1: Rebuttal: ## Weaknesses ### 1. It is preferable to provide evaluation metrics and visual results for the outputs of generative models, rather than solely focusing on the performance of enhancing pre-trained diffusion models. To qualitatively evaluate generated images, we visualize some generated images in t...	Rebuttal 1: Rebuttal: We incorporate more additional details into the Figure 1 of our paper and visualize some generated images for qualitative analysis. Pdf: /pdf/e50dbab67e3f0e0554e40b5d5efbc06f5f057d24.pdf	NeurIPS_2023_submissions_huggingface	2,023	Summary: The paper targets the discretization error of the diffusion SDE sampling, especially when the number of sampling steps is small. The authors propose a contrastive loss for diffusion sampling where instances on the same sampling chain are deemed positive pairs. The InfoNCE loss provides an upper bound on the K...	Rebuttal 1: Rebuttal: ## Weaknesses ### 1. MoCo is pre-trained on ImageNet larger than CIFAR10. The MoCo V2 encoder utilized in this paper is pre-trained on ImageNet, and we use this encoder to fine-tune DMs in all our experiments. To further analysis the performance of our method, we conduct an ablation experiment on ...	null	null	null	null	null	null
A*Net: A Scalable Path-based Reasoning Approach for Knowledge Graphs	Accept (poster)	Summary: The paper introduces ANet, a scalable path-based approach for reasoning on extensive knowledge graphs (KGs). In contrast to embedding techniques, path-based methods exhibit inductive capabilities but encounter challenges in terms of scalability due to the exponential growth of paths. ANet addresses this issue ...	Rebuttal 1: Rebuttal: Thanks for your comments. Here is our response to your concerns. W1: Although the method proposed in this article has better scalability, the contributions from theoretical perspectives are incremental compared to NBFNet. A1: The major contribution of this paper is scalability, not the theor...	Summary: This paper presents a scalable path-based method for knowledge graph reasoning, which is inspired by the A* algorithm for shortest path problems. Strengths: 1. The intriguing approach of applying the A$^*$ algorithm's principle to path reasoning in KG is proposed in this paper, along with the introduction of ...	Rebuttal 1: Rebuttal: Thanks for your recognition and constructive comments. Here is our response to your concerns. W1: The proposed method performs slightly worse than NBFNet as shown in Table 1, and no results of NBFNet are reported on tail prediction in Table 2. A1: By design, A\*Net cannot be better than NBFN...	Summary: The main contribution of this paper is presenting a scalable path-based method ANet, for link prediction on large-scale knowledge graphs. ANet is inspired by the A* algorithm for solving shortest path problems, where it learns a priority function to select important nodes and edges at each iteration. This al...	Rebuttal 1: Rebuttal: Thanks for your constructive comments. Here is our response to your concerns. *W1: In KG reasoning, the target point is unknown, rendering the core principle of A\ unfeasible. In fact, the A\* algorithm's priority rule, involving the distance to the target node, is not pertinent to the priority...	Summary: This paper proposes a scalable path-based knowledge graph reasoning approach. The idea is to extend only important paths from the exponentially growing set of all possible paths. A heuristic priority function is parametrized by a feed-forward network and is trained to predict the priority of nodes to expand. E...	Rebuttal 1: Rebuttal: Thanks for your comments. We notice that you raised only one concern for the rating of 3. Feel free to bring up other concerns for discussion. Here is our response to your concern. **W1: The first part $h^{(t)}_q(u, x)$ is already conditioned on q, u, and x, so in principle the second part $g([h^...	Rebuttal 1: Rebuttal: # Summary of Responses We would like to thank all reviewers for your time and patience on our submission. Here is a summary of reviewers’ points and our responses. *We attached a PDF file to illustrate the correspondance between A\Net and the A\* algorithm. Contributions - Scalability...	NeurIPS_2023_submissions_huggingface	2,023	null	null	null	null	null	null	null	null
ZipLM: Inference-Aware Structured Pruning of Language Models	Accept (poster)	Summary: This paper proposed a novel structured compression approach for the inference of LLM, which reduces the model size by removing entire sub-components, like rows or columns from the model’s weight matrices. Based on their algorithm, they manage to drop out some attention heads and shrink the size of FFN layers. ...	Rebuttal 1: Rebuttal: Question 1: The performance of the proposed method seems strongly dependent on the data distribution of inference, which may make it not that practical in real-world inference settings where the incoming inputs can vary a lot. We would like to note that encoder-only models, such as BERT, are ...	Summary: This paper looks into structure pruning (e.g. neurons, or attention heads) of Transformer based models. Authors propose to formulate the pruning objective by requiring the output of each pruned layer to be as close as possible to that of an unpruned layer. Then they adopt optimal brain surgeon algorithm to com...	Rebuttal 1: Rebuttal: Question 1: Computational Cost Please note that in the section “5 Discussion and Extensions”, paragraph “Computational efficiency”, we report two additional metrics: end to end runtime and efficiency with respect to the number of epochs, comparing also against the strongest competitor, CoFi: ...	Summary: The paper proposes ZipLM, a structured pruning method that can achieve desired target runtime speedups. The idea of ZipLM is to solve a layer reconstruction problem with a structural constraint which minimizes the output changes on a set of calibration examples if the layer is reconstructed as it. The propose...	Rebuttal 1: Rebuttal: **Question 1: I believe the paper (especially the methodology part) can be presented better and a lot more details should be added (even in the appendix). For example, how do you solve the reconstruction problem and obtain the optimal mask and weight update (equation (2) & (3))? How do you conside...	Summary: This paper proposes ZipLM, a structured pruning and reconstructing + layer-wise distillation + inference-aware pruning algorithm. The authors first extend the pruning formula of OBS to structured pruning and utilize estimation of inference for each structure and structured SPDY search to achieve more accurate ...	Rebuttal 1: Rebuttal: Question 1: Ablations In our submission, “Appendix B: Ablation Studies” presents results of an ablation study on the impact of distillation during the fine-tuning stage. As can be seen, our distillation technique can improve results up to _2 points_ in accuracy. Additionally, in Figure 2 of t...	Rebuttal 1: Rebuttal: We wish to sincerely thank the reviewers and the AC for their work, and for the valuable feedback. We have provided individual responses for each reviewer question. We outline answers to some common topics below: - Regarding the discussion of limitations and potential negative social impact in o...	NeurIPS_2023_submissions_huggingface	2,023	Summary: This paper proposes a structured compression method to optimize inference efficiency for language models. The proposed method uses the accuracy-efficiency trade-off under specific inference objectives as the importance measure of the model component( attention head, MLP neurons, entire FC/attention blocks.) Ex...	Rebuttal 1: Rebuttal: Question 1: Discussion of alternative approaches (quantization, Deja Vu) Please note that quantization is complementary to both distillation and pruning; in practice, it is applied in conjunction with these two. Thus, rather than comparing pruned and quantized models, one can _combine_ these ...	null	null	null	null	null	null
3D Copy-Paste: Physically Plausible Object Insertion for Monocular 3D Detection	Accept (poster)	Summary: This paper proposes a data augmentation approach to assisting the training of monocular 3D detectors by inserting 3D objects into indoor scenes in a physically plausible manner. Specifically, it addresses two main challenges in the entire pipeline: 1) where and how to insert those objects; 2) which illuminatio...	Rebuttal 1: Rebuttal: Thank you for your time and comments! Please see our response below. - Related work: Thanks for your suggestion and comments! We will modify the statement and add all the provided related papers. - Methodology: To conduct 3D object insertion for data augmentation, traditional methods [1...	Summary: This work introduces 3D Copy-Paste, a physically plausible indoor object insertion technique for automatically generating large-scale annotated 3D objects. This approach ensures the plausibility of the objects’ physical location, size, pose, and illumination within the scene Using this 3D copy-paste as augmen...	Rebuttal 1: Rebuttal: Thank you for your time and comments! Please see our response below. - Indoor/outdoor difference and novelty: Different from outdoor scenes, the challenges in indoor scenes include (1) complex spatial layouts (in particular, cluttered background and limited object-placeable space) that necess...	Summary: This paper presents a novel 3D augmentation method to augment the variety of 3D scenes with the corresponding 2D images. This augmentation method focuses on addressing the data-hungry problem when doing monocular 3D detection. The strategy here is intuitive: to insert 3D synthetic object models into 3D real sc...	Rebuttal 1: Rebuttal: Thank you for your time and comments! Please see our response below. - Methodology contribution: To conduct 3D object insertion for data augmentation, traditional methods[1] require manually locating support planes, designing poses, estimating lighting, etc, which are hard to scale up. Out me...	Summary: This paper addresses the scarcity of large-scale annotated datasets which is challenging for rapid deployment of monocular 3D object detection models. A physically plausible indoor 3D object insertion approach is proposed to automatically copy and paste virtual objects into real scenes. The resulting objects h...	Rebuttal 1: Rebuttal: Thank you for your time and comments! Please see our response below. - Comparison to Common 3D Corruption: 3D Common corruptions (3DCC) use 3D information to generate real-world corruptions, which can evaluate the model robustness and be used as a data augmentation for model training. Our met...	Rebuttal 1: Rebuttal: ## Global response (common questions) to all reviewers We would like to thank our reviewers, which put considerable time and thoughts for helping improve our paper. We are pleased that the reviewers find our paper "very well-written and easy to follow"(R-sM4b, R-mHBD, R-DWmW); our method being d...	NeurIPS_2023_submissions_huggingface	2,023	Summary: The manuscript introduces system to create RGBD datasets augmented with additional 3d objects (one per rgb frame). The inserted objects are placed such that they dont intersect other objects, stand on the ground floor and the objects are rendered into the original RGB frame such that the lighting of the scene ...	Rebuttal 1: Rebuttal: Thank you for your time and comments! Please see our response below. - Experiments on ScanNet: We use the SUN RGB-D dataset as this dataset offers 10,000+ monocular RGB-D images (scenes), and is densely annotated with 146,617 2D polygons and 58,657 3D bounding boxes. Many 3D object detection ...	null	null	null	null	null	null
Meta-AdaM: An Meta-Learned Adaptive Optimizer with Momentum for Few-Shot Learning	Accept (poster)	Summary: The paper presents Meta-AdaM, a meta-learned adaptive optimizer that also includes momentum. The parameter of the optimizer in the inner loop, i.e. inner learning rate and momentum, are predicted by a LSTM network, that takes into account the previous gradient and momentum steps. The optimization process also ...	Rebuttal 1: Rebuttal: Table 1: Ablation Study for class dynamic weight using Convnet4 on Mini-ImageNet, TieredIma- geNet, and Cifar100 datasets \| Dataset\|Method\|5-way-1-shot\| 5-way-5-shot \| \| :----: \| :-----:\|:--------------:\|:-------------:\| \|Mini-ImageNet\|ALFA \|50.58 ± 0.51\|69.12 ± 0.47\| \|Mini-ImageNet\|ALFA+DW\| 5...	Summary: The authors introduce an adaptive optimizer for meta-learning that uses double lookahead to incorporate momentum in scenarios where ‘conventional’ approaches of using momentum usually fall short (few-shot inner-loop, due to high momentum/gradient fluctuations during initial steps). The use of an LSTM to predic...	Rebuttal 1: Rebuttal: Table 1: Ablation Study for class dynamic weight using Convnet4. \| Dataset\|Method\|5-way-1-shot\| 5-way-5-shot \| \| :----: \| :-----:\|:--------------:\|:------------:\| \|Mini-ImageNet\|ALFA \|50.58 ± 0.51\|69.12 ± 0.47\| \|Mini-ImageNet\|ALFA+DW\| 50.65 ± 0.49\|70.02 ± 0.45\| \|Mini-ImageNet\|Meta-AdaM w/o DW\|...	Summary: The paper proposes Meta-Adam, a meta-learned adaptive optimizer that employs momentum to rapidly adapt a meta-learned initial model to a few-shot task. The proposed framework consists of four components. First, an LSTM that predicts a learning rate for each weight inside the model using the weight update histo...	Rebuttal 1: Rebuttal: Table 1: Additional Comparison results using Convnet4 on Mini-ImageNet. We report performances in terms of accuracy (%) with standard deviation. \|Dataset\|Method\| 5-way-1-shot\|5-way-5-shot \| \|:-----------------:\|:-------------------------:\|:------------------------:\|:------------------------:\| ...	Summary: This paper introduces a novel meta-learned optimizer with momentum for few-shot learning. The proposed approach incorporates the parameter-update history by leveraging LSTMs to adaptively adjust the learning rate. To enhance convergence speed, the momentum is integrated into the meta-optimizer by using a doub...	Rebuttal 1: Rebuttal: Table 1: Additional Comparison results using Convnet4 on Mini-ImageNet, TieredImageNet, and Cifar100 datasets. We report performances in terms of accuracy (%) with standard deviation. \| Dataset \| Method \| 5-way-1-shot \| 5-way-5-shot \| \|:----------------...	Rebuttal 1: Rebuttal: Dear Reviewers, Firstly, we'd like to express our deep gratitude for the comprehensive review and insightful feedback on our paper. We have carefully addressed each of your comments, and our individual responses to each query can be found in the subsequent sections. Incorporating your feedback, ...	NeurIPS_2023_submissions_huggingface	2,023	Summary: In this paper, the authors aim to solve the few-shot learning problem by proposing a meta-learned learning rate learner. This learner utilizes weight-update history as input to predict more appropriate learning rates for rapid convergence. Furthermore, they incorporate momentum into the optimization process of...	Rebuttal 1: Rebuttal: Table 1: Additional Comparison results using Convnet4 on Mini-ImageNet, TieredImageNet, and Cifar100 datasets. We report performances in terms of accuracy (%) with standard deviation. \|Dataset\|Method\| 5-way-1-shot\|5-way-5-shot \| \|:-----------------:\|:-------------------------:\|:--------------...	null	null	null	null	null	null
How to Fine-tune the Model: Unified Model Shift and Model Bias Policy Optimization	Accept (poster)	Summary: This paper investigates the problem of model shift in model-based Reinforcement Learning (RL). The author proposes a two-stage model learning method called USB-PO (Unified model Shift and model Bias Policy Optimization). The first stage is the same as previous work, updating based on the previous model using m...	Rebuttal 1: Rebuttal: We would like to appreciate the reviewer for the recognition and valuable comments! Our specific responses to the questions raised by the reviewer are as follows: ___ ## 1. The basis for lines 133-135 According to the previous work MBPO, the return discrepancy scheme refers to the difference bet...	Summary: This paper proposes a new MBRL algorithm which adaptively adjusts the model update by considering the model shift (distance towards previous step update) and the model bias (towards true env model) simultaneously, which avoids model overfitting and shows better empirical performance compared to only considerin...	Rebuttal 1: Rebuttal: We thank the reviewer for identifying our technical contributions. The valuable comments help us improve our submission! The reviewer’s primary concern seems to be how the paper is written and organized. Our specific responses to the concerns raised by the reviewer are as follows: ___ ## 1. Sect...	Summary: The paper considers model based RL (MBRL), focusing on the question of fine-tuning the model while learning the policy (the policy optimization). While prior works treat these two aspects somewhat separately with tuning thresholds, the authors derive a single cost function that incorporates both. The paper p...	Rebuttal 1: Rebuttal: We sincerely thank the reviewer for the high recognition of our work! Our specific responses to the questions raised by the reviewer are as follows: ___ ## 1. Policy shift wrt the model bias · According to the experiment in MBPO, when the amount of data provided by the policy for training the mo...	Summary: This paper studies the learning of the dynamics model in model-based reinforcement learning. The authors propose a novel method USB-PO that has provable properties. Strengths: 1. The paper studies an important problem in model-based RL, namely the model shift and model bias balance during model update without...	Rebuttal 1: Rebuttal: We greatly appreciate the reviewer’s insightful comments on our paper. The relevant literature mentioned in your comment helps us better polish the paper! Our specific responses to the questions raised by the reviewer are as follows. ___ ## 1. The difference between $M$ and $M_2$ $M$ and $M_2$...	Rebuttal 1: Rebuttal: We are very grateful to all the reviewers for the valuable feedback, helping us improve the paper further! We revise the paper and add suggested experiments according to the reviewer’s comments. The detailed revisions are as follows. The additional figures and a table are in the pdf file. ___ ...	NeurIPS_2023_submissions_huggingface	2,023	null	null	null	null	null	null	null	null
Unsupervised Image Denoising with Score Function	Accept (poster)	Summary: The paper introduces an extension of Noise2Score for unsupervised image denoising and demonstrates its efficiency on several noise models, including non-exponential family distributions. Strengths: The extension of Noise2Score allows to deal with different kinds of noise, as opposed to prior works that only f...	Rebuttal 1: Rebuttal: We thank the reviewer for the positive feedback and constructive comments. We answer the raised questions below. We hope that our answers clarify the doubts and address the concern of the reviewer. Q1: The contribution is relatively incremental compared to Noise2Score, and is based on an approxim...	Summary: This paper presents a self-supervised learning algorithm for image denoising which only requires the score of the noisy image distribution to perform denoising. The algorithm builds on ideas of a recent paper Noise2Score, extending the family of noise distributions that can be handled via an application of Fis...	Rebuttal 1: Rebuttal: We thank the reviewer for the positive feedback and constructive comments. We answer the raised questions below. We hope that our answers clarify the doubts and address the concern of the reviewer. Q1: It would be good if the authors could specify some distributions where SURE or Noise2Score woul...	Summary: The following work attempts to solve a single image denoising task in an unsupervised fashion. The authors propose to predict the score function and then denoise the input noisy image by solving the system of equations. Moreover, the proposed approach is a generalization of the Noise2Score method, which can wo...	Rebuttal 1: Rebuttal: We thank the reviewer for the positive feedback and constructive comments. We answer the raised questions below. We hope that our answers clarify the doubts and address the concern of the reviewer. Q1: Proposed method is highly susceptible to noise parameter estimations according to Figure 2. A1...	Summary: This paper proposes a general unsupervised image denoising approach by solving system with an estimated score function. The proposed system can be applied on multiple noise models by changing the equation system rather than retraining the model. Strengths: 1) Compared to the Noise2Score limited to the noise m...	Rebuttal 1: Rebuttal: We thank the reviewer for the positive feedback and constructive comments. We answer the raised questions below. We hope that our answers clarify the doubts and address the concern of the reviewer. Q1: Qualitative comparison should be provided in the supplementary materials. A1: We will offer a ...	Rebuttal 1: Rebuttal: We thank the reviewer for the positive feedback and constructive comments. The qualitative comparison is in the attached PDF. Pdf: /pdf/b41145f4e7840d27aa86b17e6f61582b42fbb129.pdf	NeurIPS_2023_submissions_huggingface	2,023	Summary: This paper proposes a unsupervised image denoising method based on score function. Compared with Noise2Score, it discards the noise distribution on exponential family, and derives the solutions for additive Gaussion noise, multiplicative noise, and mixture noise. The experiments verify the effectiveness of the...	Rebuttal 1: Rebuttal: We thank the reviewer for the positive feedback and constructive comments. We answer the raised questions below. We hope that our answers clarify the doubts and address the concern of the reviewer. Q1: The detailed mathematical derivation of this equation (Eq 7) is not clearly presented. A1: Equ...	null	null	null	null	null	null
Structural Pruning for Diffusion Models	Accept (poster)	Summary: This paper presents Diff-Pruning, an efficient compression method for learning lightweight diffusion models from pre-existing ones. DPMs have shown impressive capabilities in generative modeling, but they often come with significant computational overhead during training and inference. Diff-Pruning addresses ...	Rebuttal 1: Rebuttal: > Q1: Provide additional details about the pruning process like step-by-step explanations or pseudocode algorithms. A1: Thanks for the advice. Section 4 of the appendix provides a brief discussion about the pruning pipeline and training configurations. We will make it more detailed follow...	Summary: This work introduces a structural pruning method for diffusion models, called Diff-pruning. The authors leverage Taylor expansion on each term of the ELBO loss as the creteria to decide the pruned weights. By calculating the error of the final image induced by the pruning at timestep t and discussing about the...	Rebuttal 1: Rebuttal: > Q1: The motivation of the method is not clearly explained. A1: In essence, we harness the informative gradients stemming from early timesteps ($t\rightarrow 0$) to identify unimportant parameters for the pruning process. As demonstrated in our experiments, we disregard larger timestep...	Summary: The paper introduces a pruning-based method aimed at addressing the efficiency challenges of diffusion models. Unlike existing approaches that focus on accelerating sampling or enhancing architectures, it specifically focus on the time cost of compression and the consistency in generated images. The authors pr...	Rebuttal 1: Rebuttal: > Q1: the generated images may still exhibit some distortions or changes, such as the watermarks shown in Figure 2. A discussion about controllable preservation of such information. it would be beneficial to explore methods that allow for controllable preservation of such information while pru...	Summary: This paper presents Diff-Pruning, a novel structural compression technique for learning efficient diffusion models from pre-trained ones. The fundamental idea behind Diff-Pruning is the utilization of Taylor expansion on pruned timesteps, which effectively combines informative and clean gradients to estimate t...	Rebuttal 1: Rebuttal: > Q1: Extra comparison of the memory requirement. A1: Thanks for the suggestion. Methods like Random Pruning and Magnitude Pruning do not require additional memory during pruning. Taylor pruning and the proposed diff-pruning require $O(N)$ space to store the gradient, where $N$ is th...	Rebuttal 1: Rebuttal: We thank all reviewers for their constructive comments. Following the advice of Reviewer S8k8, we report our results on conditional LDMs on ImageNet-1K 256. Generated images are available in the attached PDF file. \| Method \| #Params $\downarrow$ \| MACs $\downarrow$ \| FID $\downarrow...	NeurIPS_2023_submissions_huggingface	2,023	Summary: The paper proposed a method to prune diffusion models to achieve 50% FLOPS reduction at 10% to 20% of original training budget. The authors show that prune and fine-tune strategy based on the random selection, magnitude based selection, Taylor expansion based selection all lead to non-ideal / suboptimal perfor...	Rebuttal 1: Rebuttal: > Q1: The lack of objective quantitative metrics, and the effectiveness on conditional diffusion models like LDMs. A1: Thanks for your valuable comments. In the table below, we provide the results of pruning Conditional Latent Diffusion Models on ImageNet-1K (256$\times$ 256). We follow t...	null	null	null	null	null	null
Towards Test-Time Refusals via Concept Negation	Accept (poster)	Summary: The paper proposes a method to remove a negative concept in a text-to-image diffusion model during inference time; the approach is different from prior work, which assumed that the benign concepts and the removed concepts are independent and added a penalty score; instead, this paper proposes to remove the con...	Rebuttal 1: Rebuttal: Thank you for your valuable feedback! Below we address your questions and concerns. Please feel free to post additional comments if you have further questions. 1. Deep learning architectures can be susceptible to adversarial perturbations. The introduction of deliberate perturbations can disrupt t...	Summary: Generative models require refusal techniques to limit their output and uphold ethical and copyright standards. 'Concept negation' is a promising solution, yet current methods have limitations, such as only accommodating independent concepts without considering their interconnected nature. This paper proposes P...	Rebuttal 1: Rebuttal: Thank you for your valuable feedback! Below we address your questions and concerns. Please feel free to post additional comments if you have further questions. 1. Thank you for bringing this to our attention. It is imperative to acknowledge that our current approach may exhibit diminished effecti...	Summary: This paper tackles the problem of concept negation in the context of image editing (i.e., excluding user specified concepts in the generated image). To this end, this paper proposes ProtoRe, which consists of three steps: 1) first a collection of negative prompts are encoded using CLIP and eventually aggregate...	Rebuttal 1: Rebuttal: Thank you for your valuable feedback! The paper "Imagen Editor and EditBench: Advancing and Evaluating Text-Guided Image Inpainting" centers around a distinct task known as text-guided image editing (TGIE). Unlike complete image generation, TGIE involves the editing of pre-existing or captured v...	Summary: Paper is concerned with stable diffusion of images guided by a prompt but with stipulated concepts not present. In particular it is concerned with situations where it is unclear how the prompt may be visualized without visualizing the banned concepts. Method works by incrementally pushing the generated image ...	Rebuttal 1: Rebuttal: Thank you for your valuable feedback! Below we address your questions and concerns. Please feel free to post additional comments if you have further questions. 1. Thank you for your attentive reading. Positive and negative concepts can only overlap at the level of the textual prompt, occurring sim...	null	NeurIPS_2023_submissions_huggingface	2,023	Summary: This paper proposes ProtoRe as a method for getting generative image models to refuse to show a certain concept. It incorporates the CLIP model’s knowledge, developing a prototype, and then refusing to generate (in effect) that prototype. The paper then presents a number of benchmarks showing ProtoRe’s perform...	Rebuttal 1: Rebuttal: Thank you for your valuable feedback! Below we address your questions and concerns. Please feel free to post additional comments if you have further questions. 1. In response to less easily recognizable concepts, we divide them into three groups: small objects, large objects (including backgrounds...	Summary: This paper introduces an innovative concept negation approach designed to effectively eliminate violent, unethical, or copyright-infringing content from generated images in diffusion-based models. The method operates during inference and stands apart from existing techniques by grounding concept negation in na...	Rebuttal 1: Rebuttal: Thank you for your valuable feedback! Below we address your questions and concerns. Please feel free to post additional comments if you have further questions. (1) In this context, the $\epsilon_\theta^i$ signifies the output of a noise predictor implemented by a neural network. The plus or minus...	null	null	null	null
On the Gini-impurity Preservation For Privacy Random Forests	Accept (spotlight)	Summary: This paper presents a novel encryption mechanism for isolation forest algorithms that preserves the Gini impurity metric. The authors provide theoretical evidence demonstrating that the proposed mechanism effectively preserves the Gini impurity and withstands attacks. Additionally, the authors have conducted c...	Rebuttal 1: Rebuttal: [Q1] … regarding the solidity of the proof for the theoretical guarantee in the security analysis (Theorem 3) … the proof relies on experimental analysis utilizing different random seeds rather than a traditional mathematical proof. This raises questions about the rigorousness of the proof and its...	Summary: This work develops a new encryption to maintain the data's Gini impurity and combines it with CKKS to perform random forests training and predicting, which outperforms other SOTAs in terms of the trade-off between latency and security level. Strengths: 1. This paper gets solid technique details and correspond...	Rebuttal 1: Rebuttal: [Q1] One weakness would be the limited application scenario. It would be better to have some variants for other applications. [A1] We will clarify that our work takes the first step on encryption by incorporating some learning ingredient, i.e., Gini impurity, and we can present some variants such...	Summary: This work investigates the privacy-preserving random forest, and the main contributions include an interesting property-preserving encryption method that can preserve data’s Gini impurity. Based on the proposed encryption method, the authors present an innovative privacy-preserving training and prediction algo...	Rebuttal 1: Rebuttal: [Q1] Homomorphic encryption is used to encrypt labels … high computation overhead … does this impact the efficacy of the proposed method? [A1] We will clarify that the encrypt labels (by homomorphic encryption) are only used for majority voting in random forests, and it does not require bootstrap...	Summary: This work focuses on security and privacy for machine learning, especially random forest and its splitting criterion, Gini impurity. In order to do that, this work proposes a new encryption scheme for features of data, which is against chosen plaintext attack. Moreover, this work utilizes a homomorphic encrypt...	Rebuttal 1: Rebuttal: [Q1] In section 4, authors mention [50, 72, 76, 77]. Since they relate to this work, do authors consider to update asymptotical analysis in table 1? [A1] We will clarify that our asymptotical analysis includes [50] in table 1, but does not include [72,76,77], since [76,77] only focuses on predict...	null	NeurIPS_2023_submissions_huggingface	2,023	Summary: This work focuses on the privacy of random forests. The authors develop a new encryption to preserve the Gini impurity of data, which plays an important role on the construction of random forests. Based on this new encryption, the authors introduce an effective algorithm for privacy random forests under the cl...	Rebuttal 1: Rebuttal: [Q1] Is the proposed method able to handle multiclass data? Does the number of classes impact the efficiency of the encryption method? If the number of classes is large, could this potentially lead to a sparsity problem? [A1] We will clarify that our method could deal with multiclass data efficie...	null	null	null	null	null	null
Unbalanced Low-rank Optimal Transport Solvers	Accept (poster)	Summary: The authors focus on the problem of efficiently compute discrete Optimal Transport (OT) problems, which are known to have a cubic complexity w.r.t. the number of input samples. They propose to approximate it by assuming that the transport plan is a low-rank matrix and propagate this property in the computation...	Rebuttal 1: Rebuttal: We thank you for your encouraging review. > **The low-rank approximation is not sufficiently discussed in this paper. I mentioned that it is considered to accelerate the computations. However, one might use this variant because it makes sense to have a prior of low-rank transport plan in their ap...	Summary: In this paper, the authors combine two variants of optimal transport (OT) known as unbalanced OT and low-rank OT. While the former relaxes the marginal constraints to ease the modelling rigidities and discard possible outliers from input measures, the latter helps reduce the computational cost of $\mathcal{O}(...	Rebuttal 1: Rebuttal: Many thanks for providing so much feedback, and giving 3 good grades to the soundness / presentation / contribution of our paper. We hope we can convince you to raise your final score with the answers below. > **Originality: Although the combination of unbalanced OT and low-rank OT is n...	Summary: The authors proposed unbalanced low-rank OT (ULOT) solver, which is an extension of the balanced counterpart. They show how to adapt this solver to the other unbalanced low-rank settings, namely translation-invariant and GW. The experiments compare the performances of multiple low-rank solvers, and of the Unba...	Rebuttal 1: Rebuttal: Many thanks for the wealth of comments you have provided. > **The content in the main paper and appendix is not adequately partitioned, namely the section 3 is somewhat too long that there is few space left for the experiment section, and some experiment details in the Appendix should be put in t...	Summary: In its original exact formulation, optimal transport (OT) suffers from an $\mathcal{O}(n^3)$ cost when applied to clouds of $n$ points. In addition, the exact constraint on the marginals makes it sensitive to outliers. One solution for each of these problems exist: - unbalanced OT [Schiebinger et al., 2019] is...	Rebuttal 1: Rebuttal: Many thanks for your time and for your comments. > Although the paper heavily sells low rank optimal transport, isn't the formulation in eqs 5/6 non convex? If so, doesn't it make the exact solution intractable? How to ensure convergence towards a good critical point? Yes, the formulation is...	Rebuttal 1: Rebuttal: We thank the reviewers, AC and SAC assigned to this paper for their time and work looking into our submission. We thank them in advance for reading our rebuttal and interacting with us for a few more days during the discussion period. We were happy to see that the paper was overall well rec...	NeurIPS_2023_submissions_huggingface	2,023	null	null	null	null	null	null	null	null
Zero-Shot Anomaly Detection via Batch Normalization	Accept (poster)	Summary: The paper presents an approach to address the challenging task of zero-shot anomaly detection. The authors base their methodology on three crucial assumptions: the presence of a training set that shares relevant structures with the target task, performing batch-level analysis during inference, and having the m...	Rebuttal 1: Rebuttal: > The assumption of a meta-training set is strong This assumption is common in the meta-training literature, both for general methods, such as MAML as well as meta-training AD work [8, 21, 31, 39]. There are multiple ways to mitigate this assumption. If one does not have a meta-training set at ha...	Summary: This paper proposes a method called Adaptive Centered Representations (ACR) for zero-shot batch-level anomaly detection. The method utilizes off-the-shelf deep anomaly detectors trained on a meta-training set, combined with batch normalization layers, to achieve automatic zero-shot generalization for unseen an...	Rebuttal 1: Rebuttal: > Why ACR-DSVDD with the addition of Batch Normalization Layers performs better overall than ACR-BCE is not further explained here. DSVDD has a specific inductive bias for anomaly detection to learn a compact normal data boundary, while a binary classifier doesn’t have such an inductive bias. Li ...	Summary: This paper proposes Adaptive Centered Representations (ACR) for zero-shot batch-level AD. The simple recipe, batch normalization plus meta-training, seems highly effective and versatile. The experiments are well conducted on multiple datasets. Strengths: 1. This paper is well-written, and very easy to follow....	Rebuttal 1: Rebuttal: > relation between one-class classification and anomaly detection. We wonder whether our understanding of the question aligns with your intention, so let us give our understanding of the anomaly detection (AD) problem. AD is a broad field that can be formulated in many ways (Table 3 in Ruff et a...	Summary: I think the paper has an interesting motivation but fails to deliver a convincing contribution. The method is incremental and lacks novelty and analysis. Strengths: The paper proposes a method for zero-shot batch-level anomaly detection using batch normalization and meta-training. The main idea is to train a ...	Rebuttal 1: Rebuttal: > The method is incremental and lacks novelty. We respectfully disagree. Before our work, zero-shot anomaly detection relied on large, pre-trained foundation models specifically on image or language data. We enrich the toolbox and provide a surprisingly simple solution. That is, one can use any...	Rebuttal 1: Rebuttal: We thank the reviewers for all the valuable suggestions on additional ablations. Based on their inputs, we have run additional experiments to study different normalization layers (Rebuttal Tables 1 and 2), studying different values for the hyperparameter $\pi$ (Rebuttal Table 3), meta-OOD performa...	NeurIPS_2023_submissions_huggingface	2,023	Summary: This paper introduces Adaptive Centered Representations (ACR) to tackle the zero-shot anomaly detection task. To this end, the ACR enables off-the-shelf deep anomaly detectors to generalize to unseen anomaly detection via training data distribution adaption concerning batch normalization and meta-learning. The...	Rebuttal 1: Rebuttal: > Effects of other normalizations, such as LayerNorm, InstanceNorm, and GroupNorm We thank the reviewer for mentioning other normalization techniques. As follows, we report new experiments involving LayerNorm, InstanceNorm, and GroupNorm for zero-shot AD. We will add these discussions to the revi...	null	null	null	null	null	null
Optimality of Message-Passing Architectures for Sparse Graphs	Accept (poster)	Summary: This paper studies a specific graph convolutional architecture (and the resultant estimators from an architecture with appropriate learned values) based in the setting of a contextual stochastic block model: one where a graph with latent clustering also has node features available that are informative for di...	Rebuttal 1: Rebuttal: We thank the reviewer for their concise and thought-provoking questions! They really help us develop a better perspective of our contributions and improve our presentation. We address the comments below. > It is far from clear that GCN architectures uninformatively initialized would converge to t...	Summary: The paper considers the inference problem of node classification in feature-decorated locally tree-like sparse graphs. The authors introduce and motivate a notion of asymptotic local Bayes optimality for local estimators. The first main result, Theorem 1, concerns the characterisation of the Bayes optimal (in ...	Rebuttal 1: Rebuttal: We thank the reviewer for their positive review and encouraging comments about our work! We address the questions below. > Maybe the most relevant weakness of the work is, in my opinion, that it solely focuses in locally tree-like graphs. Although this kind of models are very extended in the lite...	Summary: The authors propose a local analysis of the CSBM. They define and derive a local Bayes-optimal classifier and show it can be achieved by a message-passing architecture. They give an interpretation of it in the limiting cases when the graph caries no or all information and support it with numerical experiments....	Rebuttal 1: Rebuttal: We thank the reviewer for their encouraging review and insightful comments that helped us improve the clarity and credibility of our work. We address these below. > In the CSBM usually one considers the high-dimensional limit. the authors consider d=O(1) which seems limited. We agree that the hi...	Summary: The focus of this research paper was the investigation of node classification problem on sparsely populated graphs that exhibit local tree-like structures. Thy introduced asymptotic local Bayes optimality to define the ideal performance standard for node classification tasks. By utilizing this criterion, the p...	Rebuttal 1: Rebuttal: We thank the reviewer for their helpful comments and questions. We address the comments below. > The analysis in the paper seems interesting but more experimental results are required to support the claims. For example, comaprison with the baselines on existing benchmarks in GCN literature as wel...	Rebuttal 1: Rebuttal: In this comment, we respond to the general comments and questions of the reviewers about the convergence of the architecture's parameters, referred to as $(\theta_1, \theta_2)$ in the attached pdf, to the optimal values associated with $\rho$ and $Q$. We emphasize that figures 1 and 2 in the paper...	NeurIPS_2023_submissions_huggingface	2,023	null	null	null	null	null	null	null	null
VRA: Variational Rectified Activation for Out-of-distribution Detection	Accept (poster)	Summary: This paper conducts a fine-grained analysis of ReAct, a simple but effective OOD method by rectifying high activation. The authors perform a variational method to analyze the optimal activation function by clipping and amplifying features. Strong empirical results are presented to support the effectiveness of ...	Rebuttal 1: Rebuttal: # Response to Reviewer JjUp We thank the reviewer's appreciation of the writing and sufficient experimental results. We try to address your concerns as follows: Q1: My main concern is that the methodology seems to need to calculate $p_{out}$ in advance. However, the statistics of OOD data ar...	Summary: The authors propose a technique called "Variational Rectified Activation (VRA)", which simulates these suppression and amplification operations using piecewise functions. Theoretical analysis is provided to illustrated VRA. Extensive experiments demonstrate the effectiveness and generalization for the proposed...	Rebuttal 1: Rebuttal: # Response to Reviewer J9PQ We sincerely thank the reviewer for your detailed comments and suggestions. We try to address each comment as satisfactorily as possible: Q1: The proposed method of simultaneously truncating high and low activations is very similar to the existing method in BATS, ...	Summary: This work leverages the variational method to find optimal operation and proposes a new technique variational rectified activation (VRA) for out-of-distribution (OOD) detection. This paper finds the best operation for OOD detection and verifies the necessity of suppressing abnormally low and high activations a...	Rebuttal 1: Rebuttal: # Response to Reviewer Tdya We thank the reviewer's appreciation of our clear writing and sufficient experimental results. We try to address your concerns as follows: Q1: Can the authors report the mean and std for the main results in the experiment section? A1: Thanks for your valuable...	Summary: In this paper, the authors explore the out-of-distribution detection ability. They propose Variational Rectified Activation based on ReAct. Specifically, they suppress both the high and low values of the penultimate layer rather than only focusing on high values. Their performance on several benchmark dataset ...	Rebuttal 1: Rebuttal: # Response to Reviewer GN4S We thank the reviewer's appreciation of our theoretical analysis and rich experimental results. We try to address your concerns as follows: Q1: The theory part should be more clear about the relationship between the objective you minimize and the final score. **A1...	Rebuttal 1: Rebuttal: # Global Response Dear Reviewers, Area Chairs, and Program Chairs: We would like to express our gratitude to all reviewers for taking their valuable time to review our paper. We sincerely appreciate all reviewers for their positive comments on our theoretical analysis and technique soundness. For...	NeurIPS_2023_submissions_huggingface	2,023	null	null	null	null	null	null	null	null
On the Role of Randomization in Adversarially Robust Classification	Accept (spotlight)	Summary: The paper investigates randomized classifiers and their robustness against adversarial examples. The authors' contributions can be found along three different axes. First, the authors give the necessary and sufficient conditions so that from a set of deterministic classifiers one can build a randomized classif...	Rebuttal 1: Rebuttal: Thank you for your thoughtful review and pointing out relevant references. We are very pleased to know you found our work "interesting and well-written for the largest part". We also share the expectation that our contribution will "influence to some extent the work that will be done on the robust...	Summary: Several proposed adversarial defenses involve randomized classifiers. Prior theoretical work analyzes both the randomized and non-randomized cases. This paper answers: 1) When does randomization 'help' in adversarial classification? 2) When does it suffice to consider deterministic classifiers? The authors st...	Rebuttal 1: Rebuttal: Thank you for all the effort you put in reviewing our work and pointing out relevant references. We appreciate that you find that our paper "answers an important question in the field of adversarial learning". We will put all out effort in clarifying the exposition of the paper to make it clearer ...	Summary: This paper provides an analysis of the role of randomization in developing adversarially robust classifiers. It explores the conditions under which randomized ensembles outperform deterministic classifiers regarding adversarial risk. The study also highlights the existence of deterministic classifiers that is ...	Rebuttal 1: Rebuttal: We would like to thank you for your review. We are very happy to know that you find that our work offers a foundational understanding of the significance of randomization in building robust models. Furthermore, we also agree with you in that this theoretical contribution should be followed by prac...	Summary: The paper addresses the theoretical problem of how randomization affects adversarial robustness. It establishes a condition that is both necessary and sufficient for enhancing the robustness of probabilistic models. Furthermore, it proves that in binary classification, a deterministic model can always achieve ...	Rebuttal 1: Rebuttal: Thank you for your review. We agree with you in that a very interesting path for future work is the practical implementation of this theory. We are very interested in exploring how we can actually boost adversarial robustness, either by considering the right probabilistic classifiers (that maximiz...	Rebuttal 1: Rebuttal: We thank all the reviewers for their very valuable comments and questions. Many of them also pointed out interesting references that sparked interesting discussions within the team. We are glad that multiple reviewers found that our work tackles an interesting question related to adversarial robus...	NeurIPS_2023_submissions_huggingface	2,023	null	null	null	null	null	null	null	null
Exploring Diverse In-Context Configurations for Image Captioning	Accept (poster)	Summary: This paper presents a case study that examines the design of prompts and its effectiveness in generative visual language models for image captioning tasks. The authors delve into four methods for image selection and four methods for assigning captions to create contextually relevant learning samples. To evalua...	Rebuttal 1: Rebuttal: 1. DIIR in Lines 141-150. In Diversity-based Image-Image Retrieval (DIIR), we first extract the discrete semantic tags including object categories, attributes, and relations from an image. Suppose there are total $M$ kinds of tags, we divide them into $N$ clusters and there are $M/N$ kinds of ...	Summary: The paper studies how to choose in-context examples (images and text) for open flamingo models on the CoCo captioning task. Extensive experiments are run with many baselines. Based on these experiments, the authors conclude (1) given sufficient descriptiveness, simpler captions are better, (2) similarity of in...	Rebuttal 1: Rebuttal: 1. Presentation Issues. We will add key findings in "Abstract" (W1), enlarge Figure 1 (W2), divide the good and bad captions in Figure 1 (W2), distill the contents in Lines 51-59 (W3), and use the full names of different sampling strategies in "Introduction" and at the beginning of each secti...	Summary: This paper examines the effects of varying configurations on Vision-Language (VL) in-context learning, specifically in the field of image captioning. The authors developed four strategies each for image selection and caption assignment to determine how different methods of configuring image-text pairs impact t...	Rebuttal 1: Rebuttal: 1. Quality of Image Captions. Image captioning quality hinges on the ability of the caption to accurately and grammatically describe the main events of an image. Inadequate grammar or descriptions that either neglect or misinterpret key events indicate poor quality. Human-labeled ground truth...	Summary: This work investigates the configuration of in-context image-caption pairs for few-shot image captioning in a VL model: OpenFlamingo. Specifically, authors compare four different image selection methods: (1) random selection, (2) image-image similarity based selection (SIIR), (3) image-caption similarity based...	Rebuttal 1: Rebuttal: 1. Consistent Format. When we say that the model-generated captions (MGC) have "simple patterns", part of the meaning we wish to convey is that they often have a consistent format for captions, e.g., in Lines 236-237, we mention that ground-truth captions (GTC) contain more diverse patterns th...	null	NeurIPS_2023_submissions_huggingface	2,023	Summary: This paper explores the in-context configurations for few-shot ability in Vision-Language Models (VLMs). Therefore, they design four strategies for image selection and four for caption assignment to explore the influences of in-context pairs in image captioning. As a result, extensive experiments uncover two v...	Rebuttal 1: Rebuttal: 1. CLIPScore Assessment. Upon your recommendation, we employ the CLIPScore to re-assess the quality of our generated captions. Due to space constraints, we present only the key findings here which validate two key conclusions shown in Lines 227-229 and 302-306. We first show the results about ...	null	null	null	null	null	null
Efficient Neural Music Generation	Accept (poster)	Summary: This paper introduces an LM-guided diffusion model designed for the generation of music audios. The proposed system follows a three-step process, starting with the extraction of audio-text embeddings (MuLan tokens) using a pre-trained MuLan model. Next, semantic tokens (wav2vec tokens) are generated with the a...	Rebuttal 1: Rebuttal: We are grateful to the the reviewer's comments and suggestions. In the following, we address each of the concerns raised by the reviewer. ### Response to concern #1 Optimizing overall structure of this paper: We are thankful for the reviewer's constructive comment, which is taken seriously t...	Summary: The authors present a text to music model, with the aim of achieveing faster than real time generation on a V100 gpu. The authors introduce a latent space diffusion model. The latent space is obtained with an adverdarial VAE (replacing the VQ VAE from MusicLM), and a first LM is trained on the semantic tokens ...	Rebuttal 1: Rebuttal: We appreciate the reviewer's careful reading of our paper. We hope the following fully address all concerns mentioned by the reviewer: ### Response to concern #1 ``large number of models to train'': The reviewer is correct that MeLoDy comprises 5 components for a successful training, i.e...	Summary: The authors introduce a framework (MeLoDy) for music generation. This new framework significantly improves the inference speed of MusicLM while producing high-quality results. in terms of musicality, audio quality, and text correlation. The main reason for this speedup is that the authors replaced the coarse ...	Rebuttal 1: Rebuttal: We appreciate the thorough review regarding our study. We provide detailed responses to all reviewer's concerns, as summarized below. ### Response to concern #1 The reviewer concerned about the complex system design of the proposed MeLoDy with **``many components, moving parts, and many hyper...	Summary: This paper proposes MeLoDy, a diffusion-based text-to-music generation system. The system comprises four parts: 1) the MuLan contrastive model, 2) semantic language modeling using Wav2Vec2-Conformer, 3) a dual-path diffusion model to generate the latent, and 4) a 1D VAE-GAN autoencoder for learning the latent....	Rebuttal 1: Rebuttal: We are grateful for the reviewer's overall positive response. The concerns and questions raised by the reviewer is addressed below. ### Response to concern #1 Justification of the choice of the dual-path diffusion model: Thanks for the insightful comments. We would like to first clarify one ...	null	NeurIPS_2023_submissions_huggingface	2,023	Summary: The paper introduces MeLoDy, a music generation method combining Language Models (LMs) and Diffusion Probabilistic Models (DPMs). It addresses the challenge of generating music from diverse free-form text descriptions while mitigating high computational costs associated with current state-of-the-art models lik...	Rebuttal 1: Rebuttal: We are grateful for the reviewer's an overall positive response. The question raised by the reviewer is addressed below. Question: What's the rule of thumbs for selecting the number of chunks when training? Answer: If we understand correctly, by "the number of chunks", the reviewer refer...	null	null	null	null	null	null
Privacy Assessment on Reconstructed Images: Are Existing Evaluation Metrics Faithful to Human Perception?	Accept (spotlight)	Summary: The paper points out that current metrics, such as PSNR, MSE, SSIM, FID, and LPIPS, do not always reflect human's judgements on evaluation of model privacy risk under reconstruction attacks. Therefore, it proposes a learning-based measure called SemSim which is more compatible with human opinions. SemSim was t...	Rebuttal 1: Rebuttal: Thank you for your thoughtful review and valuable feedback. Below we address your concerns. Please let us know if further clarification is needed. **Q1: Inconsistency between lines 63-65 and the experimental setting mentioned from lines 230 to 231 and 256 to 257. I think the latter is correct, ri...	Summary: This paper evaluates the limitations of existing hand-crafted image quality metrics in indicating privacy leakage in reconstructed images. The findings reveal weak correlation and contradictions among these metrics and with human perception, highlighting the risks of solely relying on them. To address this iss...	Rebuttal 1: Rebuttal: Thank you for your thoughtful review and valuable feedback. Below we address your concerns. Please let us know if further clarification is needed. Q1: Details of FID. In lines 118-120 of the submission, we referred to [8] where FID is utilized to assess the similarity between generated and r...	Summary: The paper establish a system for human to annotate whether a reconstruction (from privacy attacks) is recognizable (thus breaches privacy). Based on human annotation results, a new learning-based metric is proposed, which shows significantly better alignment and faithfulness with human perception. Strengths: ...	Rebuttal 1: Rebuttal: We would like to thank the reviewer for the constructive comments. We address the concerns raised by the reviewer below. Please let us know if further clarification is needed. Q1: Can the proposed method generalize well enough when the distribution changed a lot, such as medical images? Gre...	Summary: This paper studies the faithfulness of hand-crafted metrics to human perception of privacy information from the reconstructed images, and discovers hand-crafted metrics only have a weak correlation with the human evaluation of privacy leakage. A learning-based measure called SemSim is proposed to evaluate the ...	Rebuttal 1: Rebuttal: We thank the reviewer for many insightful comments. We answer the questions in what follows. Please let us know if further clarification is needed. Q1: Can we just apply the original classifier on the reconstructed images as a metric for privacy leakage? Yes, and we have discussed this possi...	Rebuttal 1: Rebuttal: We thank all reviewers for the thoughtful feedback. We are greatly encouraged that the reviewers appreciate this manuscript for its originality and novelty, its implications for the machine learning privacy community (reviewer oRwe), clear motivation (reviewers oRwe and Lywc), compreh...	NeurIPS_2023_submissions_huggingface	2,023	null	null	null	null	null	null	null	null
ITEM3D: Illumination-Aware Directional Texture Editing for 3D Models	Reject	Summary: The research introduces ITEM3D, a model that enhances texture editing in 3D modeling, addressing challenges like complexity and text ambiguity. Leveraging a diffusion model, ITEM3D uses images to bridge text and 3D representations, optimizing texture and environment map. It employs a relative editing direction...	Rebuttal 1: Rebuttal: We would like to express our sincere gratitude for your valuable feedback on our paper. Q1: [Experiments on real-world dataset and experiments about efficiency] A1: We have incorporated a qualitative experiment on a real-world 3D dataset, as shown in Fig.1 of the rebuttal PDF. The da...	Summary: Given a set of multi-view calibrated images, the authors aim to first reconstruct the 3D objects based on the input images and then edit them via a pre-trained diffusion model. To accomplish this, they developed ITEM3D, which utilizes a differentiable marching tetrahedron (DMTet) as its 3D implicit representat...	Rebuttal 1: Rebuttal: Thank you sincerely for your positive and encouraging feedback on our paper. Q1: [Lack of related works] A1: We appreciate your suggestion to include these references and provide a more comprehensive discussion of the related works. Based on recent state-of-the-art studies, we have inve...	Summary: This paper proposes ITEM3D, a model for automatic 3D object editing according to text prompts. ITEM3D bridges the gap between 3D representation and natural images using rendered images. It optimizes disentangled texture and environment maps using a relative editing direction to bypass ambiguous text descriptio...	Rebuttal 1: Rebuttal: Thank you for providing additional references and addressing the lack of comparison with other baselines in our paper. Q1-Q2: [Lack of related works and comparison experiments] A1-A2: CLIP-NeRF, ARF, and SINE are CLIP-based editing models that, compared to diffusion-based models, lack th...	Summary: This paper tackles the task of texture editing for 3D models with the guidance of text. The task aims to manipulate the surface properties based on the text guidance to create corresponding visually realistic appearance. The prior works that this work targets to improve upon is the SDS loss related methods. ...	Rebuttal 1: Rebuttal: Thank you for bringing up this insightful comment. Q1: [Quantitative evaluation] A1: When evaluating generative models, quantitative metrics are often limited. In our study, we utilize CLIP-scores, which are commonly employed in evaluations of Instruct-NeRF2NeRF and other related researc...	Rebuttal 1: Rebuttal: General Response: Dear Reviewers, We would like to express our gratitude for your thoughtful feedback on our submitted academic paper. Your comments and suggestions have been invaluable in refining and strengthening our work. In this general response, we will address the three important part...	NeurIPS_2023_submissions_huggingface	2,023	Summary: The paper introduces ITEM3D, an illumination-aware model for automatic 3D texture editing based on text prompts. The authors address the challenges of complex 3D models and ambiguous text descriptions in texture editing. They propose leveraging the power of the diffusion model and optimizing disentangled textu...	Rebuttal 1: Rebuttal: Thank you for your kind words. We appreciate your advice. Q1: [Experiment on real-world dataset] A1: We have conducted additional experiments on a real-world dataset, which comprises several complex objects reconstruced from multi-view images captured by professional cameras. As depicted...	null	null	null	null	null	null
GPT4Tools: Teaching Large Language Model to Use Tools via Self-instruction	Accept (poster)	Summary: The goal of this paper is to utilize locally available models, such as vicuna and opt, to learn the tool utilization capability of black-box models like GPT-3.5. They use the large models to generate instructions and samples using image context and apply instruction tuning on these data with the fine-tunable m...	Rebuttal 1: Rebuttal: Q1: *The method proposed in this paper, which utilizes self-instruction techniques of large models to enable tool using capabilities for primitive models, is not novel. Similar techniques have been proposed in several works, such as tool-llama (https://arxiv.org/abs/2304.08354), tool-alpaca (h...	Summary: This work propose GPT4Tools, a self-instruction approach to teach large language models to use tools to solve vision tasks. The authors construct an instruction-following dataset by ChatGPT from the combinations of a set of images and 23 image-related tools. In addition, they augment the dataset to enable to c...	Rebuttal 1: Rebuttal: Q1: How effectively can the proposed instruction data collection procedure generalize to other domains? \ A1: Different from VisProg which utilizes few-shot prompt engineering, our method uses a smaller open-source language model and zero-shot manner. As shown in Table 2, our method demo...	Summary: This paper utilizes ChatGPT to generate instruction data for training moderately-sized open-source models to use tools. The authors fine-tune LLMs using LoRA and demonstrate their effectiveness on various visual problems. Additionally, the paper establishes a benchmark for evaluating the LLMs' tool usage abili...	Rebuttal 1: Rebuttal: Q1: It would be intriguing to understand the reasons behind the significant discrepancy in generalization ability between OPT and LLaMA in Table 2. A1: Compared to Vicuna and LLaMA, one possible reason is that OPT has weaker capabilities for multi-turn interaction. There are many sampl...	Summary: The authors propose GPT4Tools, a new approach to data collection for tool use, along with a novel evaluation method. In this approach, ChatGPT is utilized as a teacher to generate instruction-following data for vision-language tasks related to tools. In addition to the single-turn samples, the authors suggest ...	Rebuttal 1: Rebuttal: Q1: *The process of creating synthetic data from ChatGPT and using it to fine-tune smaller models can be seen as a distillation process. Previous studies have demonstrated that, for many tasks, distilled student language models can outperform their teacher language models across various tasks....	null	NeurIPS_2023_submissions_huggingface	2,023	null	null	null	null	null	null	null	null
On the Learnability of Multilabel Ranking	Accept (spotlight)	Summary: This paper is mainly about the learnability of multilabel ranking. The authors show that a ranking hypothesis class embeds $K^2$ different binary hypothesis classes and defines two families of ranking loss functions that capture most, if not all, ranking losses used in practice. The paper addresses the fundame...	Rebuttal 1: Rebuttal: We thank the reviewer for noting that our work provides a significant contribution to the learnability of multilabel ranking problems. Q: "This paper does not contain any experiments for their theories and algorithms" A: We thank the reviewer for noting this concern. We would like to emph...	Summary: This work study the learnability of multilabel ranking problems in both batch and online settings for different ranking losses family. They show that the multilabel ranking learnability is equivalent to the learnablities of a family of binary classification problems. Strengths: * The paper tackles a well-moti...	Rebuttal 1: Rebuttal: We thank the reviewer for finding that our paper tackles a well-motivated problem and presents a novel and original contribution to the theoretical foundations of multilabel ranking.	Summary: The authors in their work focus on a single theoretical contribution - showing that the multilabel ranking hypothesis class is PAC learnable for a wide range of ranking losses in both batch and online settings. The authors first define two general families of ranking losses that are equivalent to popularly use...	Rebuttal 1: Rebuttal: We thank the reviewer for finding that our paper contributes to the fundamental understanding of statistical learning theory for multilabel ranking. Q1: "The main proofs are only slightly outlined in the main paper; " A1: We thank the reviewer for pointing this out. We will make sure to ad...	Summary: The authors study the problem of learnability of label ranking for a large family of ranking losses. The paper is purely theoretical. The main results have been built on the equivalence of realizable and agnostic learnability of the PAC model for binary classification. Using the techniques from Hopkins et al. ...	Rebuttal 1: Rebuttal: We thank the reviewer for noting that this work is an important theoretical contribution to a practical learning problem. Q1: *"The problem setting seems to be a bit confusing. I would call it rather label ranking (potentially with ties) instead of multi-label ranking. The latter suggests th...	null	NeurIPS_2023_submissions_huggingface	2,023	null	null	null	null	null	null	null	null
Demystifying Softmax Gating Function in Gaussian Mixture of Experts	Accept (spotlight)	Summary: This paper addresses the long-standing challenge of parameter estimation in softmax gating Gaussian mixture of experts. The authors identify three theoretical challenges related to softmax gating: (i) the non-uniqueness of parameter estimation, (ii) the interaction between softmax gating and expert functions, ...	Rebuttal 1: Rebuttal: Q1: The absence of a comparison or discussion of popular techniques such as Gumbel-softmax and max-propagate raises concerns. These techniques are widely used in probabilistic MOE methods, and their exclusion from the paper may limit its completeness and applicability. Thanks for your suggest...	Summary: The authors explore the theoretical complexity of standard Gaussian Mixture of Experts and propose a Vononoi loss which estimates different components of the MLE using different Voronoi cells. Update: I have read the rebuttal and the other reviewers' responses, and will keep the score at 7 (accept). Strengt...	Rebuttal 1: Rebuttal: Q1: Unfortunately, there does not seem to be any practical exploration of the Voronoi loss in the real world setting to verify the theoretical findings, that should be fine due to the posit of the paper as a more theoretical work. Thanks for your comment. We would like to refer the reviewer t...	Summary: This paper establishes convergence rates for parameters (Gaussian mean / variance and expert weight) for softmax Gaussian mixture of experts. The rates depend on the setting: exact-fitted vs. over-fitted. The solutions are in terms of Voronoi losses and Hellinger distances. Strengths: Originality To the ...	Rebuttal 1: Rebuttal: Q1: The paper targets the convergence of the softmax mixture of expert parameters, however, in practical settings, I would imagine that these parameters are the outputs of deep networks, which may complicate the convergence picture. Thanks for your question. Regarding the softmax gating Gauss...	Summary: This paper studies convergence rate of Maximum Likelihood Estimate (MLE) of parameters of Gaussian mixture of experts with softmax gating. Unlike some of the previous papers, the authors allow for the gating probabilities to be input dependent. Convergence rate (upper bound) is provided for both the setting wh...	Rebuttal 1: Rebuttal: Q1: As someone with only cursory understanding of the area, the paper was hard to follow and full of technical detail that to me felt like obfuscating rather than facilitating understanding. Thanks for your comment. We would like to refer you to Common Question 1 in the General Response to un...	Rebuttal 1: Rebuttal: General Response Dear AC and reviewers, Thanks for your thoughtful reviews and valuable comments, which have helped us enhance our paper substantially. In this section, we will address some common concerns from the reviewers, and then include these changes to the revision of our paper. **CQ...	NeurIPS_2023_submissions_huggingface	2,023	Summary: This work considers the problem of parameter estimation for the softmax-gated mixture of Gaussian experts. For simplicity, first consider the case in which the mixture weights do not depend on the covariates X. Then, this model can be seen as a mixture of linear regressions where each “regression component...	Rebuttal 1: Rebuttal: Q1: The full details of the system of polynomial equations in the introduction are unnecessary and should be omitted. Instead, the authors should briefly introduce the quantity $r(m)$ and its properties. Additionally, how do the authors arrive at these polynomial equations? Thanks for your su...	null	null	null	null	null	null
Counterfactual Generation with Identifiability Guarantees	Accept (poster)	Summary: This paper proposes a framework for unsupervised style transfer by disentangling the content and style representations. Unlike previous research, they do not assume independence between the content and style variables in the generation process but rather only a lower influence of the style variable compared to...	Rebuttal 1: Rebuttal: Thank you so much for the valuable comments and suggestions. We also hope that our method could provide insights for the following research. We respond below to address your concerns. ### Title Thank you for letting us know this. The title should be “Counterfactual generation with identifiabilit...	Summary: In the paper, the authors take varying dependence between content and style into account in the counterfactual generation process. The identification problem is addressed when using a VAE for the task. It is proved that the subspaces of latent variables for contents and style are identifiable. Then based on th...	Rebuttal 1: Rebuttal: Thank you so much for the insightful comments and suggestions. We hope that generation tasks, including advanced LLMs would benefit from principles for representation learning, as developed in our work. ### Human Evaluation Thank you for providing us with valuable suggestions to enhance our wor...	Summary: This paper wants to do a controllable text style transfer by tackling the dependence between the content and the style variables. They adopt the concept of influence sparsity requiring the influence of the style variable to be sparser than the content variable. They evaluate their method on several NLP dataset...	Rebuttal 1: Rebuttal: Thank you so much for the valuable comments and suggestions. We address each point as follows. ### Advanced baselines and evaluation on other styles Thanks for providing the comprehensive list! It should be noted that our proposed method does not rely on attribute/style labels in the training, ...	Summary: The work "Controllable Text Generation with Identifiability Guarantee" presents a model for unsupervised counterfactual generation based on a model with disentangled style and content. The model is based on a variational autoencoder augmented with two flow-based models operating in the latent space. The flow-b...	Rebuttal 1: Rebuttal: Thank you so much for the valuable comments and suggestions. We have revised and included some results and we hope the revision could address your concerns. ### Task is limited Thank you for the suggestion! We will make it explicit that our empirical study is limited to sentiment transfer. At th...	Rebuttal 1: Rebuttal: We all thank all reviewers for their valuable feedback and dedicated time! We address the individual comments in separate response and will incorporate the reviews' suggestions in our revision. Pdf: /pdf/d6b520c328c048451d9ef12b9feb9e47344c4bec.pdf	NeurIPS_2023_submissions_huggingface	2,023	Summary: This paper discusses controllable text generation and tackles the dependence between content and style in the counterfactual generation task. Identification guarantees are proven and used to enhance disentangling of the variables. Strengths: Style and content disentanglement, especially for scenarios with spa...	Rebuttal 1: Rebuttal: Thank you for your valuable time and your detailed feedback! We address each point as follows. ### Writing (1) Key Concepts. Identifiability guarantee refers to the standard notion of identifiability in statistics (https://en.wikipedia.org/wiki/Identifiability), which describes the poss...	null	null	null	null	null	null
Efficient RL with Impaired Observability: Learning to Act with Delayed and Missing State Observations	Accept (poster)	Summary: This paper studies online learning in tabular MDPs with impaired observability, which means one of two models: states are observed in stochastic delay or some observations are missing (sampled independently). The authors present algorithms based on optimism and prove regret bounds that scale optimally with $S,...	Rebuttal 1: Rebuttal: Thank you for your comments! >Q1: The novelty of this paper is limited to the definition of the new models. It appears to be a factored MDP with known solutions. A1: We disagree respectfully. Our augmented MDP is not a factored MDP. Factored MDPs encode rewards and transitions exhibiting...	Summary: The paper considers reinforcement learning with delayed and missing observations. It is shown that a MDP with delayed observations is equivalent to an augmented MDP with perfect observations. Based on the augmented MDP, an optimistic algorithm is proposed for delayed MDPs and it is shown to achieve a near-opti...	Rebuttal 1: Rebuttal: Thank you for your comments! >Q1: In the definition of the transition below line 164, the collections of actions for time $h$ and $h+1$ may have different sizes. Is that just some typo or some additional things need to be handled? A1: Varying sizes of actions is a consequence of our gene...	Summary: This paper aims to provide a theoretical analysis of reinforcement learning with delayed and missing state observations, and establish near-optimal regret bounds, for RL in both the delayed and missing observation settings. Despite impaired observability posing significant challenges to the policy class and pl...	Rebuttal 1: Rebuttal: Thank you for your comments! >Q1: The paper does not include preliminary experimental results, which may hinder the empirical understanding of the methods. A1: Thanks for the suggestion. In the RL theory literature, developing theory for tabular MDP is usually the first step, and most pa...	Summary: This paper studies reinforcement learning in the setting where MDP is with delayed state information, but delayed observations still arrive in order, and in the setting where the state information of MDPs could be missed and never be observed. For both settings, this paper provides provably efficient RL algori...	Rebuttal 1: Rebuttal: Thank you for your comments! >Q1: Novelty of the proposed algorithms. A1: A naive application of tabular RL algorithm to our problem would yield exponential $\tilde{\mathcal{O}}({\rm poly}(H) \sqrt{SA^HK})$ regret. We are able to achieve regret that scales with $\sqrt{S A}$ by utilizing ...	Rebuttal 1: Rebuttal: We would like to thank all the reviewers for a thoughtful review and valuable comments, which help to improve the quality of the paper. We have provided our response to your individual questions and concerns. In addition to clarify our contributions and technical novelties, we conduct numerical e...	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.