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 |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
Testing the General Deductive Reasoning Capacity of Large Language Models Using OOD Examples | Accept (poster) | Summary: The paper presents a programmable natural language reasoning dataset, PrOntoQA-OOD, specifically designed for testing the out-of-distribution generalization abilities of machine learning models. 4 LLMs, FLAN-T5, LLaMA, GPT-3.5, and PaLM, are tested on this dataset for their ability to generalize to different d... | Rebuttal 1:
Rebuttal: Thank you for the feedback and helpful comments! We would greatly appreciate any additional comments or questions during the discussion period.
1. “[1] presents a benchmark for theorem proving that is programmable and tests generalization across six dimensions. These dimensions include different ... | Summary: This submission proposes a new dataset PRONTOQA-OOD to systematically study the general deductive reasoning capabilities of large language models (LLMs). Some findings are obtained, e.g., LLMs exhibit mixed generalization to unseen deduction rules and more robust generalization to longer, wider, and compositio... | Rebuttal 1:
Rebuttal: Thank you for the feedback and helpful comments! We would greatly appreciate any additional comments or questions during the discussion period.
1. “Similar to existing datasets for evaluating the deductive reasoning capabilities of LLMs, the proposed dataset is built on simple and short templates... | Summary: In this paper, the authors explore the general deductive reasoning ability of LLMs. They test LLMs on a broad set of deduction rules and measure the ability to generalize to more complex proofs from simpler demonstrations from multiple angles. They construct a programmable dataset that enables control over ded... | Rebuttal 1:
Rebuttal: Thank you for the feedback and helpful comments! We would greatly appreciate any additional comments or questions during the discussion period.
1. “The main issue with the writing is that the outcomes of the experiments are not clearly communicated, likely due to a lack of correlation between fig... | Summary: The authors propose a new dataset ProntoQA-OOD, to evaluate LLMs' reasoning abilities across more complex proofs across multiple aspects such as depth, compositionality, etc. The authors find that LLMs are able to generalize to longer and compositional proofs but usually require demonstrations to get these goo... | Rebuttal 1:
Rebuttal: Thank you for the feedback and helpful comments! We would greatly appreciate any additional comments.
1. “The main weakness is the lack of meaningful insights from the results, which are mostly contradictory to previous findings…”
In addition to the summary of contributions in the common rebutta... | Rebuttal 1:
Rebuttal: Thank you to all reviewers for the very helpful feedback and comments! We will address common concerns below and also respond to each reviewer individually.
1. Some reviewers have raised the concern that there is a lack of meaningful insights from the results, especially those that are contradict... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Fast Rank-1 Lattice Targeted Sampling for Black-box Optimization | Accept (poster) | Summary: The paper presents the Rank-1 Lattice Targeted Sampling (RLTS) technique as a solution to the challenge of efficiently scaling up black-box optimization in high-dimensional problems. The primary objective of this technique is to enhance the query efficiency while ensuring fast Gaussian process (GP) training an... | Rebuttal 1:
Rebuttal: We sincerely appreciate the reviewer's constructive advice and detailed comments. However, we respectfully disagree that some of the comments may stem from a misunderstanding. We will kindly clarify the reviewer's misunderstanding first. Afterward, we will provide thorough and detailed responses ... | Summary: This paper studies high-dimensional black-box optimization. In order to achieve high efficiency, this paper proposes a Rank-1 Lattice Targeted Sampling (RLTS) technique, using random rank-1 lattice Quasi-Monte Carlo, to perform fast local exact Gaussian processes (GP) training and inference with O(n log n) com... | Rebuttal 1:
Rebuttal: We thank the reviewer for dedicating time and effort to review our paper. Your support is greatly appreciated.
Q1. Can the authors explain if RLTS still has a faster convergence than other compared methods in terms of physical time?
A1. We further compare our RLTS with the high-dimensional... | Summary: For the setting of black-box optimization where only function queries are available, this paper propose a rank-1 lattice targeted sampling technique to scale up to high-dimensional problems with good query efficiency. Based on the special structure, this paper use FFT/IFFFT to do fast coordinate search with O(... | Rebuttal 1:
Rebuttal:
We sincerely appreciate the reviewer's constructive advice and valuable comments. Detailed responses to the reviewer's questions are given below.
Q1. Some concept appears without explanation such as "dual lattice".
A1. Thanks for the reviewer’s suggestion. We will revise the paper according... | Summary: This paper proposes a novel Rank-1lattice targeted sampling (RLTS) strategy to significantly reduce the GP training & inference time complexity and also enable $O(n\log(n))$ time complexity in order to find a targeting reference point at each iteration. Combined with the evolution strategies, the method propos... | Rebuttal 1:
Rebuttal:
We sincerely appreciate the reviewer's constructive advice and valuable comments. Our detailed responses to the reviewer's questions are presented as follows:
Q1. “ In GP training, in addition to considering the complexity of computing the value of likelihood, isn't it also important to consid... | Rebuttal 1:
Rebuttal: # Overall Response
We sincerely appreciate all reviewers’ constructive advice and valuable comments. We first present an overall response regarding the common concerns raised. Subsequently, we provide detailed responses to each reviewer’s specific questions.
### Clarification
In this work, our... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Geometric Algebra Transformer | Accept (poster) | Summary: This paper is about geometric algebra, which is interesting but not widely known. In this paper, the authors propose a transformer-based model for geometric data, which can represent general geometric data and transformation. The authors also apply this method in two cases, which shows the effectiveness of the... | Rebuttal 1:
Rebuttal: Thank you for the constructive review and helpful questions.
> There is no ablation study.
Thank you for the suggestion. We ran an ablation study on the main architectural decisions in GATr.
We performed this experiment on a new real-world problem: estimating the wall shear stress from artery ... | Summary: This paper introduces the Geometric Algebra Transformer, a novel architecture that leverages the projective geometric algebra to embed inputs, outputs, and hidden features. The transformer incorporates specific linear layers, nonlinear layers, normalization layers, and attention mechanisms designed for the pro... | Rebuttal 1:
Rebuttal: Thank you for the in-depth review, kind praise, and on-point questions.
> The proposed model's requirement of 16-dimensional vectors for feature representation may present scalability challenges compared to traditional transformers.
Good point. To address it, we added a study of the computation... | Summary: The paper proposes a (transformer-based) network architecture that operates on a projective geometric algebra that can encode simple 3D transformations and 3D objects (such as points, planes, etc) in a unified description. The network performs operations on these mathematical objects that are (also in composit... | Rebuttal 1:
Rebuttal: Thank you for the detailed and constructive review as well as the kind words.
> In how far could permutation invariance (or the lack thereof) become an issue?
Our architecture is fully **permutation equivariant**: a permutation of the input tokens leads to the same permutation of the output tok... | Summary: This paper presents an adaptation of the transformer architecture to enforce equivariance to E(3), which can be helpful to a variety of practical applications necessitating such geometric inductive bias.
This is achieved partly by leveraging geometric algebra rather than vector algebra, which allows for a sim... | Rebuttal 1:
Rebuttal: Thank you for the thorough review, helpful suggestions, and kind words.
> A number of practical applications would require equivariance to SE(3) only. [...] Equivariance in 2D would also be rather useful.
Good point, we agree. There are two strategies to address this: either designing separate ... | Rebuttal 1:
Rebuttal: Thanks to the reviewers for their insightful feedback. We are delighted that they find the key ideas behind GATr - geometric algebra representations, equivariance, and a transformer architecture - "sound and useful" (reviewer **wqKZ**) and "very methodological" (**8oHP**). They appreciated both th... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Feature Dropout: Revisiting the Role of Augmentations in Contrastive Learning | Accept (poster) | Summary: This paper questions a common assumption on data augmentations in contrastive unsupervised representation learning, specifically label-preserving, where data augmentations do not change the label of the original sample. Indeed, this assumption does not always holds in data augmentations used in contrastive uns... | Rebuttal 1:
Rebuttal: We thank the reviewer for their review! We are glad they found our research question novel and significant, and to give a better view of the problem.
**W1a: Details of transfer experiments.**
To perform the transfer experiments in Tables 1 and 2 we use the same setup as in the Viewmaker Paper ... | Summary: **Summary**. The authors question the common assumption in contrastive learning that good augmentations should be label-preserving, and suggest that dropping features out (for example, with data augmentation) actually is helpful in practice because it helps learning of "weaker" features: features that may be h... | Rebuttal 1:
Rebuttal: We thank the reviewer for their review! We are glad they found the results interesting, and appreciated our “questioning common assumptions in the field.”
**[Q1 and Q2] Do our synthetic experiments generalize to the the real world?**
This is an important question, a big reason we conducted exper... | Summary: The authors of this work challenge the belief that label-preserving augmentations are necessary for contrastive learning by demonstrating that label-destroying augmentations can be useful for learning representations useful for a variety of downstream tasks. The authors treat augmentations as feature dropout,... | Rebuttal 1:
Rebuttal: We thank the reviewer for their thoughtful review! We are glad that they found our experiments and theoretical results to be thorough. From our understanding, the reviewer’s main point of concern is clarifying the paper's additional contribution on top of the two papers cited.
We are happy to sta... | null | null | Rebuttal 1:
Rebuttal: We thank the reviewers for their reviews! We briefly highlight two points mentioned across more than one reviewer.
**Results on naturalistic datasets**
Our experiments on the naturalistic CelebA dataset are a significant contribution of our work, but we realize from the reviewers’ feedback that ... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Enhancing Minority Classes by Mixing: An Adaptative Optimal Transport Approach for Long-tailed Classification | Accept (poster) | Summary: This paper aim to solve long-tailed problem in the perspective of data augmentation. Specifically, the backgrounds of majority classes are mix with the foreground of minority classes. This work found out that randomly mixup images might face semantic misalignment, which impedes the effectiveness of this scheme... | Rebuttal 1:
Rebuttal: **(1) Response to performance gain:** We would like to emphasize that in the highly competitive area of imbalanced classification, our performance gain over CMO is consistent and significant. Specifically, on CIFAR-LT-10 and CIFAR-LT-100 (imbalance factor=100,50,10), and ImageNet-LT (imbalance fac... | Summary: This paper addresses the interesting long-tailed classification problem, which is usually confronted by real-world machine learning tasks. In particular, this work extends the existing idea of mixing the background images from the majority class and foreground images from the minority class for imbalanced clas... | Rebuttal 1:
Rebuttal: **(1) Response to contribution:** Please see global response.
**(2) Response to the comparison with well-known baselines:** Thanks for your suggestion. We provide more comprehensive evaluation experiments with Focal loss and Class-Balanced (CB) loss on long-tailed datasets in the pdf uploaded on ... | Summary: The authors tackle the notable challenge of learning from imbalanced data, proposing an adaptation to mixup-based or copy-paste-based data augmentation for rare categories that takes sample-level semantic similarity into account. Their proposed method uses optimal transport to underpin an adaptive mixup strate... | Rebuttal 1:
Rebuttal: **(1) Response to unreasonable samples:** Thanks for your insightful comments. We agree that for any imperfect method, there can be cases where foreground patches are mixed unreasonably with background patches, like the example you pointed out. We would like to point out that CMO mixes patches com... | Summary: To tackle the data imbalance problem, the authors suggest to manipulate the data distribution by introducing a new data augmentation mechanism. This mechanism is a guided mix-up operation that uses backgrounds mostly from major classes and foregrounds from low frequency classes. To make the augmented images ... | Rebuttal 1:
Rebuttal: **(1) Response to typos:** Thanks for your reminder and we will carefully address them in the revised version.
**(2) Response to the fraction about the mixed images:** According to your suggestion, we conduct an additional experiment with the fraction of the images unchanged in every batch in the... | Rebuttal 1:
Rebuttal: We thank all reviewers for their constructive comments. We upload a global pdf about the figures and tables. Below we first address common concerns and then respond to each reviewer.
**Response to Reviewers HKTJ and mXEg about contribution and novelty:** Learning relatively balanced classes from ... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Multiply Robust Federated Estimation of Targeted Average Treatment Effects | Accept (poster) | Summary: In this paper, the authors propose a multiply robust targeted average treatment effect estimator when the data is distributed across multiple sites. The proposed method only requires transmitting statistics among the target site and the source site, without directly sharing their own data.
Strengths: 1. The p... | Rebuttal 1:
Rebuttal: * Thank you very much for raising this important issue. The DP perspective is very interesting and we will consider adopting DP methods in future work in order to claim formal privacy guarantees. The current work is privacy-preserving in the sense that the only information that is shared is the fo... | Summary: This paper introduces a new federated approach for conducting multi-source studies, allowing valid average treatment effect inferences to be derived for a target population while preserving privacy and addressing heterogeneity in covariate distributions. The proposed methodology incorporates transfer learning ... | Rebuttal 1:
Rebuttal: Weaknesses
1. When the identification assumptions do not hold, we cannot identify the TATE from the observed data. However, we can still derive bounds under relaxed identification requirements. These bounds are useful because they provide a range of plausible values for the TATE. In practice, we c... | Summary: Authors address estimating a target average treatment effect (TATE) in the setting of multi-site studies that include a target site and several source sites. For the target and each source site, authors propose estimating the TATE; then, TATE estimates are ensembled. Authors allow for source sites to have addi... | Rebuttal 1:
Rebuttal: Weaknesses
1. As one concrete example, suppose the ATE is 0 in both the target and source. However, the CATE across racial groups is not the same. In the target, the treatment effect is positive for Caucasians but negative for African Americans; in the source, it is null for both. Exchangeability ... | Summary: This paper focuses on federated ATE estimation. The main challenges lie in deriving valid causal inferences from decentralized data with heterogeneity, different data structures, and privacy constraints. To address these issues, the authors propose a privacy-preserving estimator that allows incorporating site-... | Rebuttal 1:
Rebuttal: * Thank you for raising this question. The Reviewer is correct in noting that some sites can have extreme covariate shifts against the target site in practice. Under this scenario, correctly specifying the density ratio model can be challenging. To increase the flexibility of our density ratio mod... | Rebuttal 1:
Rebuttal: * We have responded to each reviewer's comments.
* We also attach a PDF containing results from an updated experiment, in response to Reviewer 27aW's inquiry regarding non-comparable sample sizes in the target site and source sites.
Pdf: /pdf/3607ef17afa7bf68437b15d32c6479d074a17bf7.pdf | NeurIPS_2023_submissions_huggingface | 2,023 | Summary: This paper considers federated learning for causal inferences using multi-site data. The authors propose an estimator which is multiply-robust (to model misspecification and covariate shift) and privacy-preserving with additional nuisance function estimation. The authors then provide theoretical guarantees for... | Rebuttal 1:
Rebuttal: * We wish to clarify two aspects that are fundamental to the contribution of our paper:
1. Covariate mismatch can occur regardless of whether assumption (A6) is met.
2. We do not require the outcome models to be the same across different sites.
Based on the availability of covariates and invest... | null | null | null | null | null | null |
Language Models can Solve Computer Tasks | Accept (poster) | Summary: The paper introduces a novel approach, Recursively Criticizes and Improves (RCI), for guiding large language models (LLMs) to execute computer tasks using natural language commands. The RCI method outperforms existing methods and sets a new state-of-the-art on the MiniWoB++ benchmark when combined with the Ins... | Rebuttal 1:
Rebuttal: Thank you for your insightful feedback on our submission. We appreciate the time you took to review our work. We have carefully considered your comments and have made improvements to address the weaknesses and address your questions and concerns.
1. More comparison with other methods (e.g. self-d... | Summary: The paper presents a new method to enable LLMs to execute computer tasks guided by natural language. The method uses a simple prompting scheme called Recursive Criticism and Improvement (RCI), which prompts the LLMs to find and fix problems in their output. Without large amounts of expert demonstrations and ta... | Rebuttal 1:
Rebuttal: Thank you for the review and valuable feedback. We appreciate the reviewer's careful examination of our submission. Below is our rebuttal addressing the raised concerns:
1. Impact of RCI iteration steps
We agree that the impact of RCI iteration steps should have been explicitly presented in the p... | Summary: This paper proposes a prompting approach that utilizes a critic to improve the web navigation performance. Within the same context, the model reviews a candidate answer and improves it based on reasoning steps. For web navigation, there are two phases: explicit and implicit RCI. Explicit RCI takes a task and g... | Rebuttal 1:
Rebuttal: We express our gratitude to the reviewer for their valuable feedback, which has significantly contributed to identifying areas for improvement in our paper. Below, we address the points raised to enhance the clarity of our content.
1. Comparison with ReACT and InnerMonologue on computer tasks
Reg... | Summary: This paper proposes an algorithm for executing tasks using LLMs. While this can in principle be applied to any sequential task with language input and output spaces, they tested primarily on computer control tasks.
The method, RCI, works as follows:
- The environment state, action space, and task are represen... | Rebuttal 1:
Rebuttal: We appreciate the reviewers' valuable feedback on our paper. We have carefully considered the comments and suggestions and would like to address each of the concerns raised.
1. Task selection:
We understand the concern about task selection for our evaluation. We followed the WebN-T5-3B's [1] task... | null | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
H-nobs: Achieving Certified Fairness and Robustness in Distributed Learning on Heterogeneous Datasets | Accept (poster) | Summary: This paper introduces a fair and robust algorithm in distributed learning setup, which combines q-FFL [8] (or the one by [6]) with the norm-based screening robust method. Three convergence results w.r.t. the algorithm are developed for non-convex, convex and strongly convex models. Experimental results are pre... | Rebuttal 1:
Rebuttal: **Comment:** *The developed algorithm looks a simple combination of the prior works: ([6],[8]) for fairness objectives and [4] for gaining insights for NBS.*
**Response:**
Thank you for your comment. While it is true that the key elements of our framework, NBS and fairness objectives, both come f... | Summary: This paper proposes a protocol to address both fairness and Byzantine robustness in federated learning, and specifically i) the objective function is designed to take fairness into account, and ii) the algorithm takes an average of gradients with large norms and reduces the influences of large gradients. In ad... | Rebuttal 1:
Rebuttal: **Comment:** *I’m not sure why having a large norm (at least possibly) implies maliciousness? I’m not an expert in Byzantine robustness, but if this is a result in this area, I’d certainly prefer to have a reference or at least a brief discussion.*
**Response:**
Thank you for your question. Let's... | Summary: The paper explores robustness to byzantine attacks/failures as well as fairness in a federated learning setting.
The main focus of the paper is on a heterogenous user setting, i.e. data is non-iid between users. Particularly, the paper considers egalitarian fairness, i.e. uniform performance for all users.
To ... | Rebuttal 1:
Rebuttal: **Comment:** *With regards to novelty and contribution: Both NBS and fairness-promoting losses are not novel. Can you clarify the contributions of the paper and comment on the novelty of the different aspects of the proposed approach?*
**Response:**
Thank you for your insightful question. Althoug... | Summary: This paper acknowledges and tackles the challenges associated with fair and robust distributed learning. It addresses these challenges by introducing the H-nobs framework and provides convergence guarantees under different convexity assumptions. The proposed algorithm is then evaluated through experiments unde... | Rebuttal 1:
Rebuttal: **Comment:** *My main concern is the experimental evaluation. In particular, the paper considers just one dataset, examines a single value of $\alpha$, one data heterogeneity setting. A single setting cannot provide enough evidence to thoroughly evaluate the applicability and utility of the propos... | Rebuttal 1:
Rebuttal: We thank all the reviewers for their valuable time and helpful comments. Here we first want to highlight two key aspects of our work and then address two major concerns raised by the reviewers. In particular, we note that a reject recommendation was made mainly based on the extensiveness of our si... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
The RL Perceptron: Generalisation Dynamics of Policy Learning in High Dimensions | Reject | Summary: This paper focuses on developing theories for the learning dynamics of policy gradient reinforcement learning (RL) algorithms with a particular focus on high-dimensional latent feature space. As an early work along this direction, the authors study a binary-action environment setup for simplicity. The authors ... | Rebuttal 1:
Rebuttal: We thank you for the time you took to review our paper, and appreciate the comments raised, which will help to improve the quality of the paper. We address these comments below:
* **Data space dimensions**: The accuracy of the ODE description of the learning dynamics hinges on the concentration o... | Summary: This paper proposes a model for solving high-dimensional problems in reinforcement learning (RL) referred to as the RL perceptron. The model is used as a framework for analyzing generalization dynamics of simple neural networks for RL tasks. More precisely, the model employs a student teacher design in which t... | Rebuttal 1:
Rebuttal: We thank you for the time spent reviewing the paper and for the positive comments, and we're glad you enjoyed reading the paper. We really appreciate the comprehensive feedback.
Below, we go through the points made and address the questions after.
* **Connection to policy gradient**: We agree wi... | Summary: The authors develop a set of differential equations that describe the learning dynamics in high-dimensional settings, allowing for a quantitative analysis of learning behaviors. This framework enables the computation of optimal hyper-parameter schedules and the visualization of phase diagrams for learnability.... | Rebuttal 1:
Rebuttal: We thank the reviewer for providing a review of our paper. However, we have to disagree on the two points raised by the review:
We thank the reviewer for providing a review of our paper. However, we have to disagree on the two points raised by the review:
* We want to stress that our paper does ... | Summary: The work proposes a theoretical framework to study the average case learning behavior of deep RL policy gradient methods. The framework is based on ODEs that can describe the typical learning dynamics of PG RL agents. The framework is used in various settings to describe learning behaviors in these and a final... | Rebuttal 1:
Rebuttal: We thank you for the time spent reviewing, and appreciate the feedback given. We understand clarity is something we can work on, and will strive to order definitions and assumptions in a way that is easier to follow in the revised version. Below we respond to the points you raised, and we would be... | Rebuttal 1:
Rebuttal: We thank the reviewers for their time spent and comments on our paper. Here is a list of references referred to in other responses, and attached is the PDF of new figures.
[1] Andreas Engel and Christian Van den Broeck. Statistical mechanics of learning Cambridge University Press, 2001
[2] Maryl... | NeurIPS_2023_submissions_huggingface | 2,023 | Summary: The paper discusses the application and theoretical understanding of Reinforcement Learning (RL) algorithms in high-dimensional settings. The authors propose a high-dimensional model of RL that can capture a variety of learning protocols and derive its dynamics as a set of closed-form ordinary differential equ... | Rebuttal 1:
Rebuttal: We thank you for the time spent reviewing and the positive comments on our work, we really appreciate it. We’re glad you acknowledged the value of analysing scenarios in average-case instead of worst-case.
We thank you for your questions, they are indeed ones which we have been considering oursel... | Summary: This paper introduces a new problem definition and class of solution methods for decision making.
The motivation put forward by the authors is that the current pool of solution methods do not have theoretical results that capture the neural functional classes used by practical applications in RL, leaving a ... | Rebuttal 1:
Rebuttal: We thank the reviewer for their time spent reading our paper, and for the thoroughness of the review. Due to character limits, we only briefly respond to some points, but we are happy to elaborate during the discussion period.
Clarity is paramount and we clearly can improve thanks to the feedback... | null | null | null | null |
Learning Complete Protein Representation by Deep Coupling of Sequence and Structure | Reject | Summary: The paper addresses the challenge of integrating protein sequence and structure information to improve protein representations. The authors propose CoupleNet, a network that utilizes graph convolutions to model the relationships between protein sequences and structures. The approach involves constructing two t... | Rebuttal 1:
Rebuttal: **Q1** Differences with GearNet.
**A1** Thanks for the valuable feedback! (1) CoupleNet uses two graph types (radius and sequential) rather than GearNet's [1] three (adding k-nearest), the sequential edges in the sequential graph of GearNet are further divided into different edge types according... | Summary: This paper proposes a representation learning framework using GNN for protein datasets. A key contribution is combining sequence and structure information using the proposed GNN. The learned representations achieve better performance on downstream tasks of protein fold classification and function prediction.
... | Rebuttal 1:
Rebuttal: **Q1** Limitations of existing approaches
**A1** (1) Sequences and structures provide complementary evolutionary and geometric information, but integrating them is challenging due to different data types, as shown in Figure 1 in the manuscript. This may make the methods complex and computational... | Summary: The authors proposed CoupleNet to co-model the protein sequences and structures. CoupleNet separately builds a sequence-based graph and a radius graph for message passing. It achieved state-of-the-art performance on several datasets compared with recent baselines.
Strengths: 1. The proposed CoupleNet utilized... | Rebuttal 1:
Rebuttal: **Q1** The major differences from GearNet in terms of graph construction.
**A1** Thank you for your valuable feedback! (1) There are mainly three different types of graphs in GearNet [1], including the K-nearest graph, radius graph, and sequential graph, where the sequential edges in the sequent... | Summary: This paper introduces a network called CoupleNet that combines protein sequence and structure information to generate informative protein representations. The network utilizes multiple levels of features, including residue identities and positions for sequences, as well as geometric representations for tertiar... | Rebuttal 1:
Rebuttal: **Q1** The relationships between protein sequence, structure, and function.
**A1** Thank you for your valuable feedback! Please allow us to elaborate on the relationships. (1) As shown in Figure 1 in the manuscript, protein sequence and structure provide different aspects of information, e.g., t... | Rebuttal 1:
Rebuttal: We sincerely thank all the reviewers for their insightful and constructive reviews of our manuscript. We are encouraged to hear that the reviewers find **our paper is clearly written, figures ares concisely defined** (Reviewer S326, 7zKz); **Our work/idea is effective/novel**(Reviewer S326, 4x6Z, ... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Transformers learn to implement preconditioned gradient descent for in-context learning | Accept (poster) | Summary: The paper tries to tackle an important and relevant problem in LLM training: whether in-context learning properties can be explained by a form of gradient descent?
For the setting, the authors have chosen to focus specifically on the linear transformer with a single or multiple layers with an input that is gi... | Rebuttal 1:
Rebuttal:
**Regarding the setting** We agree with the reviewer that our setting is theoretical. However, we emphasize that despite our assumptions, the setting is expressive enough for ICL, and has been studied in a number of other papers.
- [1,2,3] use linear attentions as linear attentions are expressiv... | Summary: This paper studies linear transformer and validates its ability in solving linear problems via applying conditioned gradient descent. The authors firstly state that there exists a unique global minimum for single attention layer. Then they explore multi-layer linear transformer and demonstrate that stationary ... | Rebuttal 1:
Rebuttal:
Thank you for your constructive comments! We address your comments one by one as below.
> Compared to the previous work (von Oswald et al., 2022), the novelty is insufficient except the introduction of non-isotropic data distribution.
Our results in Theorems 1, 2, and 3 are novel even for is... | Summary: LLMs trained with transformers have shown remarkable abilities that have led to a lot of recent theoretical research towards understanding the origins of these abilities. In context learning is one such ability that refers to a model's ability to adapt in-context to new tasks without weight updates. In the wor... | Rebuttal 1:
Rebuttal:
For the linear attention vs softmax attention comment, we provide an extended response in the [joint response](https://openreview.net/forum?id=LziniAXEI9¬eId=p362P0UrGI) above. Thank you for pointing it out!
> What would happen if we were to consider the standard parametrization. Do we stil... | Summary: The paper presents an in-depth analysis of in-context learning (ICL) in transformers, particularly focusing on linear transformers. The authors propose a novel interpretation of ICL as an implicit form of '**preconditioned**' gradient descent. They further extend this analysis to multi-layer transformers and p... | Rebuttal 1:
Rebuttal: > I did not find the ICL experiments --- x-axis represents the number of ICL samples, and the y-axis represents the test loss (similar to the experiments in Figure 2 of Garg et al. (2022)). It would be interesting to check those results besides the current ones on training dynamics.
Thank you for... | Rebuttal 1:
Rebuttal: We thank all the reviewers for very constructive comments.
- We thank **Reviewer m6sU** recognizing the novelity of our work: _"The paper provides a novel perspective -- the preconditioning/adaptive perspective -- on in-context learning in transformers... Given the growing interest in understand... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Diffused Redundancy in Pre-trained Representations | Accept (poster) | Summary: Authors argue that pre-trained representations exhibit a degree of "diffused redundancy", i.e. the accuracy on downstream tasks with linear probing is similar when we use the whole layer compared to a randomly subset of neurons. This claim is validated using representations pre-trained using ImageNet-1k and Im... | Rebuttal 1:
Rebuttal: We thank the reviewer for their positive feedback and excellent suggestions! We answer their concerns below:
### [W#1 comparison to last $k\%$ principal components]
Thank you for this suggestion! We’ve added these results in the response PDF (Fig 4). We see that projections onto the lower PCA co... | Summary: The paper empirically investigates the various factors that affect how the features of a pretrained representations are encoded. They find that the features exhibit a degree of diffuse redundancy and randomly chosen subsets of the features in the penultimate layer have similar accuracy on downstream tasks and ... | Rebuttal 1:
Rebuttal: We thank the reviewer for their valuable comments! We address their concerns below:
### [W#1 decreased downstream training time]
We would like to emphasize that our goal is not model optimization. There has been long-standing research on structured pruning (He et al.[1]; Li et al.,[2]) that achi... | Summary: The authors introduce the concept of `diffused redundancy', which is the idea that a random sample of neurons reproduces the performance of the whole layer in many downstream tasks. The authors support the existence of this phenomenon with experiments on a variety of architectures pre-trained on ImageNet and e... | Rebuttal 1:
Rebuttal: We thank the reviewer for their positive feedback and great suggestions! We address their concerns below:
### [W#1 neural collapse]
Thank you for pointing out this interesting connection! We agree that diffused redundancy and neural collapse go hand-in-hand, and it would be especially interestin... | Summary: The paper proposes investigating the emergence of redundant representations in pre-trained models. The setup is pretraining of various architectures (VIT, vgg, resnet) on Imagenet1k. The study pertains to how much of the representation is redundant and this is estimated by dropping features and finetuning on s... | Rebuttal 1:
Rebuttal: We thank the reviewer for their comments. We address their concerns below:
### [W#3 confusion in PCA vs Diffused Redundancy results]
We believe the reviewer is referring to L225-235, and associated figures 2c and 2d. The goal of this analysis is to compare diffused redundancy (our work) with PCA... | Rebuttal 1:
Rebuttal: We thank the reviewers for their time, positive comments and constructive feedback. There were many excellent points raised by the reviewers. We ran the following experiments to address their concerns:
- Fig 1: One common comment raised by reviewers was that all our experiments were only on the ... | NeurIPS_2023_submissions_huggingface | 2,023 | Summary: Focusing on the (penultimate layer) representation learned by the pre-training model, this paper proposes a "diffused redundancy hypothesis" that includes two main concepts:
- Redundancy: A small subnet can achieve comparable performance with the full model.
- Diffused redundancy: The selection of the small s... | Rebuttal 1:
Rebuttal: Thank you for your careful review! We address your concerns below:
### [W#1 obvious in light of model compression]
We draw differences between our work and model compression literature in our paper in Section 1.1. To expand on that, model compression works aim to reduce the number of parameters ... | null | null | null | null | null | null |
Sharp Spectral Rates for Koopman Operator Learning | Accept (spotlight) | Summary: This paper develops bounds on how close the approximated Koopman modes and eigenvalues are to the true eigenvalues and modes, for two important classes of methods used to compute the Koopman mode decomposition. The authors find that one class, Principal Component Regression, which included extended dynamic mod... | Rebuttal 1:
Rebuttal: We appreciate the reviewer's insightful evaluation and valuable comments. In what follows, we aim to address the highlighted weaknesses and respond to the reviewer's questions.
__Weakness 1 & Question 1:__
First, we would like to stress out that [Korda and Mezic 2018] prove the consistency (infi... | Summary: This paper studies the approximation and learning of the Koopman operator. Koopman operator is helpful in modeling a broad class of Markovian dynamical systems. In this paper, two types of approximation strategies are studied, namely, "Principal Component Regression (PCR)" and "Reduced Rank Regression (RRR)." ... | Rebuttal 1:
Rebuttal: We appreciate the reviewer's insightful evaluation and the positive comments. In what follows, we aim to respond to the reviewer's questions.
__Questions:__
1. We agree, the notation is not the best. We will follow the advice and denote by ${\rm B}_r(\mathcal{H})$ finite rank operators $\mathca... | Summary: The authors provide bounds for the spectral decomposition of the estimated Koopman operators. The bounds are given for self-adjoint time-reversible operators in terms of two new metrics. Compared to estimation guarantees given in terms of the Hilbert Space distance norm, the proposed bounds require less restri... | Rebuttal 1:
Rebuttal: We appreciate the reviewer's insightful evaluation and valuable comments. In what follows, we first aim to address the weaknesses and respond to the reviewer's questions.
__Weaknesses:__
>_It is unclear..._
We are not sure to understand the reviewer's remark. In the appendix all the bounds are... | Summary: This paper analyses two approximation techniques for self-adjoint compact Koopman operators. Both estimators are constructed based on regularized least-squares with a certain rank condition in mind. The Principal Component regression (PCR) performs unconstrained Tikhonov estimation of data projected on a low... | Rebuttal 1:
Rebuttal: We appreciate the reviewer's insightful evaluation and valuable comments. In what follows, we aim to address the highlighted weaknesses and respond to the reviewer's questions.
__Weaknesses:__
- Regarding the variance bounds' constants, we carefully derived the proofs in the Appendix so that they... | Rebuttal 1:
Rebuttal: We wish to thank all reviewers for their insightful evaluation of our paper. We appreciate all their comments and remarks, which we will incorporate in our revision. Before addressing each review in detail, we would like to point out some general remarks that apply to all of them.
__Assumptions &... | NeurIPS_2023_submissions_huggingface | 2,023 | Summary: This paper proves sharp upper bounds for eigenvalue estimation of a Koopman operator for a time-homogeneous Markovian dynamical system using either reduced rank regression or principal component regression. The bounds include operator norm error and metric distortion. These results are illustrated on simple mo... | Rebuttal 1:
Rebuttal: We appreciate the reviewer's insightful evaluation and valuable comments. In what follows, we aim to address the highlighted weaknesses and respond to the reviewer's questions.
__Weaknesses:__
Indeed, the tools used to establish variance bounds in Section 5 are not particularly novel, and their ... | null | null | null | null | null | null |
Prompt-augmented Temporal Point Process for Streaming Event Sequence | Accept (poster) | Summary: The author introduces PromptTPP, a method that integrates classical neural temporal point processes with a continuous-time retrieval prompt pool. This approach enables the model to learn event streams sequentially, eliminating the need to buffer previous examples or task-specific features in the context of con... | Rebuttal 1:
Rebuttal: > The method description is not clear.... the buffer of continual learning is dynamically changing. You can show the case study of the Taobao dataset and the Amazon dataset.
In our experiment we found the performance w.r.t. hyperparameters is relatively stable over time. See below table. For each... | Summary: This paper addresses the problem of Temporal Point Processes (TPP), which involves modeling continuous-time events. Specifically, the authors focus on developing a model for streaming data, where the model needs to be improved in a Continual Learning (CL) setting. To tackle this challenge, the authors propose ... | Rebuttal 1:
Rebuttal: > Can the authors provide justification on using only two datasets in the experiment section?
Firstly, using two benchmark datasets for a certain task is common in TPP papers. For example, [1] use MIMIC and StackOverflow to test their transformer structure on event sequences; [2] use Retweet and ... | Summary: The paper proposes a novel prompt-augmented continual learning (CL) framework for modeling streaming event sequences, called PromptTPP. The framework incorporates a continuous-time retrieval prompt pool and a retrieval mechanism to dynamically lookup a subset of task-relevant prompts based on the instance-wise... | Rebuttal 1:
Rebuttal: > The paper does not provide a detailed analysis of the computational complexity of the proposed framework.
The computational complexity comes from two parts: the TPP model and prompt pool.
Take ANHP as the base model. Assume the input sequence length is $K$, event embedding size $d_e$. Recall ... | Summary: This paper aims to enhance the Neural Temporal Point Processes (TPPs) used in modeling continuous-time event sequences. Particularly in the real-world, the distribution of event data shifts over time, which existing TPPs fail to reflect. To address this, the authors propose PromptTPP, a continual learning fram... | Rebuttal 1:
Rebuttal: > The authors refer to CLTPP (Dubey et al., 2022) as the only neural TPP with continual learning (CL) capabilities. However, they did not conduct experiments on the Yelp and Meme datasets used in this paper (Dubey et al., 2022). Is there any reason for it?
The problem formulation and dataset proc... | Rebuttal 1:
Rebuttal: We thank all the reviewers for their constructive feedback. In this response, we will first use a few reply-to-all messages to clarify our dataset settings and present some new experiment results. Then we will provide a point-to-point response to each individual review. We can surely consolidate a... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Bridging RL Theory and Practice with the Effective Horizon | Accept (oral) | Summary: The authors aim to improve the theoretical understanding of RL. The contributions of the paper are as follows:
1. The authors curated a dataset called BRIDGE consisting of 155 deterministic MDPs along with their tabular representations
1. The authors found that, for a given deterministic MDP, deep RL is more l... | Rebuttal 1:
Rebuttal: We thank the reviewer for their careful and comprehensive review. We are glad that they found our results on acting greedily with respect to the random policy’s Q-function “extraordinary and quantitatively verified.” Below are responses to questions and comments.
**Our PPO results vs. benchmarks:... | Summary: This paper proposes a new complexity metric, denoted as effective horizon, for deterministic MDPs. It is defined as the number of full value iteration updates starting from an initial policy, plus a logarithmic factor for concentration bounds, needed to reach an optimal policy. A corresponding algorithm, GORP,... | Rebuttal 1:
Rebuttal: We thank the reviewer for their comprehensive review. We appreciate that they found our results “extensive” and that they think the tabular representations in our BRIDGE dataset “can be very useful for future research on tighter theoretical bounds.” Below are responses to individual comments and q... | Summary: The paper investigates the question of what finding an explanation for the performance of deep RL algorithms. To accomplish this, the authors propose the notion of effective horizon. It is found that this notion is better able to answer questions about sample complexity and effect of reward shaping or policy i... | Rebuttal 1:
Rebuttal: We thank the reviewer for their thorough and thoughtful review. Below are responses to questions and comments raised in the review.
**Effective horizon vs $k$:** First, we noticed that there may have been some confusion between the value of $k$ for which an MDP is $k$-QVI-solvable versus the effe... | Summary: The paper aims at addressing the gap between RL theory and deep RL practice.
The central finding of the paper is a property that correlates well with the performance of PPO and DQN algorithms on a set of 155 environments (such as Atari, Procgen, and MiniGrid).
The property looks at whether the argmax action o... | Rebuttal 1:
Rebuttal: We thank the reviewer for their careful and thoughtful review. We appreciate that the reviewer found that our work offered “an interesting and non-trivial way to think about RL.”
**Difference between random policy’s Q-function and randomly initialized Q network:** In this work, we only focus on t... | null | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
StEik: Stabilizing the Optimization of Neural Signed Distance Functions and Finer Shape Representation | Accept (poster) | Summary: The authors demonstrate that the eikonal equation which is often used for representing 3D shapes leads to training instabilities. Based on this finding the authors propose a new regularizer for representing 3D shapes based on the eikonal equation. This regularizer is based on the hessian of the SDF function. M... | Rebuttal 1:
Rebuttal: **Q1** *In table 3, why does the runtime decrease when adding the new regularizer which is based on the hessian? I would assume that this impacts the runtime negatively. Can the authors elaborate on this?*
**->** The major savings comes from the number of back-propagation calls. In our approach,... | Summary: The paper first utilizes Von Neumann analysis to put forward a plausible explanation for instability of neural SDFs as the function space enlarges (as network parameters increase). Along the way, they show that some of the commonly used regularizers can combat this instability if they are weighted heavily enou... | Rebuttal 1:
Rebuttal: **Weakness1** *The paper could have provided a bit more context for the Von Neumann analysis. They provide a reference, but it still might be helpful to give a bit of an introduction to the analysis (or give a more specific chapter/section/theorem reference).*
**->** We will add more related refe... | Summary: The paper provides a theoretical framework to study the optimization of neural SDFs. In this framework, the eikonal loss - used as regularization to obtain a SDF - results unstable. Existing regularizers are proven to stabilize the eikonal loss, but they empirically lead to oversmoothing. To overcome this issu... | Rebuttal 1:
Rebuttal: **Weakness1** *Increased Training Time: The major weakness is the increased training time associated with the proposed method due to the use of quadratic layers, as noticed by the authors. The ablation study suggests that the regularizer alone does not improve performance, necessitating the use of... | Summary: The paper gives a theoretical framework for analysis the stability of neural SDF optimisation, by considering the gradient descent optimisation as a PDE. They can then analysis the gradient descent on Eikonal loss PDE, and show that it is inherently unstable, and further show that normal loss or divergence los... | Rebuttal 1:
Rebuttal: **Weakness1** *The ablation results on SRB seem to show that quadratic layers could be more important than the Laplacian normal regularisation, but it is hard to tell as the overall improvements are minor. It would be better to run the ablation on ShapeNet, where there is a much larger improvement... | Rebuttal 1:
Rebuttal: We thank the reviewers for a thorough review. There are some concerns mentioned by several reviewers so we will address them here.
**[Slight Correction on ShapeNet Results]**: We apologize that there are some minor errors in the evaluation of ShapeNet in the initial submission. We mistakenly put ... | NeurIPS_2023_submissions_huggingface | 2,023 | Summary: This paper presents a novel approach to learning implicit neural representations (INR) of shapes. The paper tries to address the issue of instability in optimizing neural signed distance functions (SDFs) using the eikonal loss, which can lead to irregularities in the reconstructed surfaces and convergence to s... | Rebuttal 1:
Rebuttal: **Weakness 1** *The paper states that their method improves the representation of shapes compared to state-of-the-art methods. However, the evaluation on the surface reconstruction benchmark dataset does not demonstrate significant performance improvements over other methods.*
**->** Since SRB is... | null | null | null | null | null | null |
Compositional Generalization from First Principles | Accept (poster) | Summary: This paper theoretically investigates mathematical conditions for compositional generalization.
It defines compositional representation and compositional support, then derives a set of sufficient conditions on training distribution support and model architecture.
It also conducts empirical validation and ablat... | Rebuttal 1:
Rebuttal: We thank the reviewer for reading our work and for commending the clarity of work. We encourage the reviewer to check our global response where we have addressed some of their concerns, nonetheless we address some specific points here:
1. (Weakness 0: “The theorem is weak and only covers limited ... | Summary: Compositional generalization, which is easy for humans but hard for deep neural networks, gradually attracts more attention. Different from most existing works focusing on how to train a more to compositionally generalize well, this paper aims at the characteristics of the datasets. It tries to find sufficient... | Rebuttal 1:
Rebuttal: We thank reviewer 5Ysh for reading our paper, for commending the novelty and relevance of our approach, and for asking intriguing questions. We have clarified some concerns in the global response to all reviewers. Nonetheless, we would like to address some specific points below:
1. We concur with... | Summary: This paper proposes formal conditions on a data generating procedure such that "compositional generalization", i.e. generalization to unseen combinations of elements observed during training, can be shown to be possible. The authors construct a toy task and show empirically that when the proposed conditions ar... | Rebuttal 1:
Rebuttal: We thank reviewer SW1j for reviewing our paper and for lauding our effort to formalize notions of compositional generalization. While we hope to have already captured most of your concerns in our general reply above, we would like to also address your individual concerns here.
1. (Weakness 1): Ou... | Summary: In this work, compositionality is examined as a fundamental attribute of the data-generating process, rather than being solely associated with the data. By establishing a precise mathematical framework, it aims to comprehensively analyze and understand compositional generalization. Notably, the work identifies... | Rebuttal 1:
Rebuttal: We express our gratitude to reviewer zapL for their thorough review of our paper, for highlighting the paper’s strengths, and for providing valuable suggestions. While we have already tackled a number of concerns outlined in our global response and will proceed to address specific questions below.... | Rebuttal 1:
Rebuttal: We would like to thank all 5 reviewers for their time and valuable feedback!
We appreciate their assessment of our work as **“valuable for future work related to compositional generalization”** (yWdt). The reviewers concur that **“the lack of compositional generalization ability has been a longst... | NeurIPS_2023_submissions_huggingface | 2,023 | Summary: This work combines the compositional generalization problem with the data generation process and mathematically defines and formalizes the relevant framework. In this process, two main conditions are proposed: the first is the condition for a given problem to be considered a compositional generalization proble... | Rebuttal 1:
Rebuttal: We thank reviewer yWdt for reviewing our paper and providing helpful suggestions. While we hope to have already captured most of your concerns in our global reply above, we would like to also address your individual concerns here.
1. (Weakness 1.) We agree that the explanation of *broad* and *nar... | null | null | null | null | null | null |
Flow Factorized Representation Learning | Accept (poster) | Summary: The paper introduces a new generative model called Flow Factorized Representation Learning that learns a distribution over sequences of observed variables using which one can learn both disentangled and equivariant representations. The model is based on the idea of representing transformations in the input spa... | Rebuttal 1:
Rebuttal: We thank $\textcolor{pink}{Reviewer\ wDok}$ for the positive feedback and constructive suggestions. Below we respond to the concerns point by point.
---
**Q1. Pseudocode of the Algorithm**
Thanks for the constructive suggestion. We provide the following pseudo-codes for training our flow-factor... | Summary: This paper proposes a flow factorized representation learning approach for disentanglement and equivariant representation learning in a learned latent space. This work introduces a generative model that defines a set of latent probability paths to specify different input transformations and each latent flow is... | Rebuttal 1:
Rebuttal: We thank $\textcolor{green}{Reviewer\ znKv}$ for the constructive suggestions. Below we respond to the concerns point by point.
---
**Q1. Application on Real-world Datasets**
Thanks for the great advice! As suggested by $\textcolor{magenta}{Reviewer\ ww47}$, we take a step further to apply our ... | Summary: This paper proposes the Flow Factorized Representation to learn structured representations. It introduces a generative model that defines different input transformations with a set of latent flow paths. The latent evolution is generated by the gradient flow of learned potentials, which are parameterized by a s... | Rebuttal 1:
Rebuttal: We thank $\textcolor{blue}{Reviewer\ xnNs}$ for the careful review and valuable advice. Below we address the concerns in detail.
---
**Q1. Disentanglement Metrics**
Thanks for the constructive suggestions! We strongly agree that the VP metric [a] indeed suits our method as it does not pose any ... | Summary: The authors study disentangled and equivariant representation learning as the authors claim existing approaches are either ill-specified or can not effectively separate realistic factors of interest in the learned latent space. The key idea is to define a distinct set of tangent directions that associate diffe... | Rebuttal 1:
Rebuttal: We thank $\textcolor{magenta}{Reviewer\ ww47}$ for the encouraging feedback and valuable suggestions. Below we respond to the concerns in detail.
---
**Q1. Disentanglement Metrics**
As stated in line29 of the supplementary (and also indicated by $\textcolor{blue}{Reviewer\ XnNs}$), most traditi... | Rebuttal 1:
Rebuttal: We sincerely thank all the reviewers for their constructive suggestions and valuable feedback. We appreciate the reviewers' common sentiment that our work is **intuitive and novel** ($\textcolor{red}{R1}$, $\textcolor{magenta}{R2}$, $\textcolor{green}{R4}$), **mathematically sound and technically ... | NeurIPS_2023_submissions_huggingface | 2,023 | Summary: This paper proposes a novel approach to unsupervised representation learning; More precisely, the authors focus on the learning of latent dynamics with the aim to design an equivariant mapping from the observational to the representational space; Authors leverage variational posterior approximation and dynamic... | Rebuttal 1:
Rebuttal: We thank $\textcolor{red}{Reviewer\ tD2D}$ for the constructive suggestions and positive feedback. Below we respond to the questions point by point.
---
**Q1. Choosing the Value of $k$**
For MNIST, we choose the value of $k$ in line with TVAE [a] and PoFlow [b]. For Shapes3D, we use the self-co... | null | null | null | null | null | null |
From Tempered to Benign Overfitting in ReLU Neural Networks | Accept (spotlight) | Summary: This work examines asymptotic overfitting behavior on a simple class of toy classification problem for two-hidden-layer ReLU nets, finding that increasing input dimension tends to push the model from tempered towards benign overfitting.
Strengths: Originality: this is, to my knowledge, the first analysis of t... | Rebuttal 1:
Rebuttal: We thank the reviewer for the thorough review and constructive comments.
- "Comparison to the experiment of Figure 9 of Mallinar et al.": This is indeed an interesting comparison, although we point out that we analyze a classification setting, while the aforementioned experiment deals with regres... | Summary: This paper considers the various settings of overfitting, including benign overfitting where an interpolating estimator is optimal in spite of perfectly fitting noisy labels, tempered overfitting where an interpolating estimator is inconsistent but error is a function of the label noise, and catastrophic overf... | Rebuttal 1:
Rebuttal: We thank the reviewer for the very thorough review, constructive comments and positive feedback.
- Bound for Theorem 3.1: We acknowledge that this bound is not tight (even for very large values of $m$), while the bound in Theorem 3.2 is tight yet requires stronger assumptions. We believe that the... | Summary: This work studies the phenomenon of overparameterized neural networks (NNs) generalizing well even when trained on noisy data. Previous research focused on "benign overfitting," where interpolating predictors achieve near-optimal performance. However, recent empirical observations suggest that NN behavior is b... | Rebuttal 1:
Rebuttal: We thank the reviewer for the constructive review and insightful comments.
We used this data distribution as it was already studied in previous works (e.g. Mallinar et al. 2022), while making the analysis tractable. Our results for one-dimensional data can be readily extended for other distributi... | Summary: This paper explores the spectrum of benign to catastrophic overfitting in a specific instance of a 2-layered neural network with ReLU activation when the training data has noisy labels with proportion $p$. The authors when the input dimension is , the test error is lower bounded proportionally to $poly(p)$ und... | Rebuttal 1:
Rebuttal: We thank the reviewer for the thorough review, helpful comments and positive feedback.
Weaknesses:
- 1. If space permits, we will consider adding a short discussion on practical implications of our results. We note that such discussions also appear in several previous papers on benign and temper... | null | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Individualized Dosing Dynamics via Neural Eigen Decomposition | Accept (poster) | Summary: This develops a robust SDE algorithm for medical forecasting. This is demonstrated on both synthetic and real data.
Strengths: The problem solved is clearly compelling and the performance is evaluated on a wide range of datasets, where it exhibits superior performance to reasonable alternatives. It would appe... | Rebuttal 1:
Rebuttal: We thank the reviewer for their comments. If our response is satisfactory, we kindly ask the reviewer to update the score.
**Components Description**
In the current version, we briefly describe the components in Section 4. The observation filter is further elaborated in Appendix A, the ESDE is ... | Summary: The paper presents a piece-wise linear neural SDE with control for biomedical forecasting. The key contribution is combining rigorous linear SDE solutions with hypernet parameterisation.
Strengths: The paper is fantastically written and very clear.
The idea of linearising SDEs is relatively novel, and the o... | Rebuttal 1:
Rebuttal: We thank the reviewer for their positive feedback.
**Regarding the value of the method**
NESDE aims to solve a hard problem – forecasting and filtering a sparsely-observed SDE with an independent control signal. To address the accompanying challenges, NESDE combines several components. First, in... | Summary: This paper proposes a novel network approach, named Neural Eigen Stochastic Differential Equations (NESDE), to solve sequential prediction problems mainly in medical dosing control field of applications.
Strengths: $\mathbf{1}.$ The proposal of NESDE is novel, along with the hypernet that determines the subs... | Rebuttal 1:
Rebuttal: We thank the reviewer for their positive and detailed feedback.
**General applicability of NESDE and Neural-ODE**
NESDE constitutes a framework for forecasting individualized dosing dynamics. As described in Section 3, we consider the general problem of forecasting the dynamics of a controlled S... | Summary: This paper presents a method for predicting time series at irregular time points. Its approach systematically integrates control signals, a convenient attribute for modelling dosing regimes in medicine. The latent process follows a piecewise linear stochastic differential equation (SDE) with control. The SDE i... | Rebuttal 1:
Rebuttal: We thank the reviewer for their helpful feedback. We will modify the manuscript as discussed below to address the reviewer’s comments. In particular, we will clarify the selection of baselines, which seems to be the main concern in the review. If our responses are satisfactory, we will appreciate ... | Rebuttal 1:
Rebuttal: We are excited that all the reviewers recommended accepting the paper, and thank them for their helpful comments. The reviewers found that the “problem solved is clearly compelling” (a6Bv), and the paper is “fantastically written and very clear” (hUXX). They describe our proposed method by “the pr... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
What functions can Graph Neural Networks compute on random graphs? The role of Positional Encoding | Accept (poster) | Summary: This paper theoretically studies node classification on latent position random graphs, a generalization of graphons. More precisely, the family of functions that can be approximated using GNNs with positional encodings are determined in general and for the examples, sign-net (eigenvector PEs) and distance-base... | Rebuttal 1:
Rebuttal: Thank you for your very detailed review.
**Questions:**
1. In our analysis (Example 1) the adjacency matrix is normalized: $A/(\alpha_n n)$, which guarantees that its eigenvalues are bounded with high probability. It requires however the knowledge (or approximate knowledge) of the ``sparsity lev... | Summary: The paper provides a complete description of the function space generated by equivariant GNNs on latent position random graphs in order to study the expressive power of GNNs on large graphs for node level tasks. Then the paper investigates how node features such as positional encodings would affect the functio... | Rebuttal 1:
Rebuttal: Thank you for your detailed and insightful review. We agree with the reviewer's opinion that many questions remain on the interpretability of the function spaces defined in this work, beyond the results presented in this paper. We will complete the outlook section if space permits. We also agree t... | Summary: The paper investigates the theoretical expressiveness of Graph Neural Networks (GNNs) on large random graphs, and the role of node features and Positional Encodings (PEs) in enhancing their approximation power. The paper makes several contributions, such as:
- Providing a complete and intuitive description of... | Rebuttal 1:
Rebuttal: Thank you for your review. We are delighted that you found our paper to be well-written and organized, with detailed technical proofs in the appendix. | Summary: The paper analyzes the approximation capability of Graph Neural Networks (GNNs) in node-level tasks. The work presents a novel concept of expressivity that examines whether GNNs can approximate $L^2$-functions on measure spaces as the underlying graph grows and approximates operators on $L^2$ itself. The paper... | Rebuttal 1:
Rebuttal: We thank the reviewer for their insightful and review.
- **On related work and overstatement**: thank you for pointing this out, we will add more related work on universality (and other places where we feel it might be necessary). You are correct in pointing out that the normalization ``good prac... | Rebuttal 1:
Rebuttal: Dear reviewers,
We thank you for your valuable comments. We are glad to see that see that our paper is well-received and considered as *well-written* (gCL6, mk3D, rtyh, 2T44). The novelty is highlighted by several of you (gCL6, mk3D, 2T44, Jsrn).
A common concern of the referees is the practica... | NeurIPS_2023_submissions_huggingface | 2,023 | Summary: This paper aims to enhance the theoretical understanding of Graph Neural Networks (GNNs) concerning their expressive power, focusing on large graphs. The paper provides a comprehensive description of the function space generated by equivariant GNNs for node tasks, incorporating notions of convergence from pre... | Rebuttal 1:
Rebuttal: We thank the reviewer for their detailed and insightful review.
- **Practical implementation**: the main goal of this paper is to better understand the theoretical properties of existing GNNs on large graphs, and we focus on classical implementations that use the adjacency matrix or normalized La... | Summary: In this paper the authors provide characterisations of the class of functions that can be approximated by GNNs, possibly in the presence of positional encodings (PE). Here the notion of convergence is stated with high probability and for increasing graph size. So, in large graph limit.
Strengths: - There is ... | Rebuttal 1:
Rebuttal: We thank the reviewer for their detailed and insightful review.
- **Practical insight**: as mentioned by the reviewer, on large (random) graphs, in the infinite-node limit, node-task GNNs approach functions. Approximation results indicate *which labelling functions the GNN is able to approximate*... | null | null | null | null |
Non-autoregressive Machine Translation with Probabilistic Context-free Grammar | Accept (poster) | Summary: This work attempts to learn hierarchical structure in form of a PCFG as part of a non-autoregressive translation (NAT) system. The intuition is that the PCFG relaxes the conditional independence assumption between target tokens in vanilla NAT. The authors argue that related approaches with the same objective f... | Rebuttal 1:
Rebuttal: Thank you for your valuable feedback and comments on our paper. We appreciate your recognition of the novelty of our proposed PCFG-NAT method and its potential to learn hierarchical dependencies in NAT. We would like to address your concerns and questions in the following rebuttal.
#### **PCFG-st... | Summary: This paper proposes PCFG-NAT, which alleviates the issue of conditional independent modeling through probabilistic context-free grammar. PCFG-NAT leverages syntax information to capture bidirectional dependencies among predicted tokens. In this way, this method also provides a better explanation of the transla... | Rebuttal 1:
Rebuttal: We appreciate your valuable feedback on our paper and would like to respond to your comments and suggestions.
#### **Comparison with earlier methods and reported results:**
Regarding the discrepancy in the reported results for the CTC+GLAT method, we would like to clarify that the difference is ... | Summary: This paper proposes to rely on PCFG to incorporate the most likely parsing tree into the non-autoregressive transformer model as a way to mitigate multi-modality problem. Since considering all possible parsing trees is computationally expensive, the authors propose a right-heavy PCFG based on the assumption th... | Rebuttal 1:
Rebuttal: We appreciate your valuable feedback on our paper and would like to respond to your comments and suggestions.
#### **Balancing between expressive capabilities and computational complexity:**
We acknowledge that our paper could have provided more experimental evidence to support our claim that RH... | Summary: The paper proposes a new approach called PCFG-NAT to address the limitations of conventional Non-autoregressive Transformer (NAT) models. PCFG-NAT uses a specially designed Probabilistic Context-Free Grammar (PCFG) to capture complex dependencies among output tokens and enhance the expression power of NAT mode... | Rebuttal 1:
Rebuttal: We appreciate your valuable feedback on our paper and would like to respond to your comments and suggestions.
#### **Training complexity analysis:**
In our paper, we provided the algorithm and complexity analysis for both training and inference of PCFG-NAT, along with detailed pseudocode in the ... | Rebuttal 1:
Rebuttal: In response to the reviewers' suggestions, we have incorporated two sets of analysis experiments during the rebuttal stage.
First, we conducted ablation studies to examine the impact of Glancing Training and Knowledge Distillation on our model's performance.GLAT consistently improves the transla... | NeurIPS_2023_submissions_huggingface | 2,023 | Summary: This article proposes a new NAT training method based on PCFG. To address the limitations of critical and unidirectional dependencies, the authors provide a new way of establishing dependencies from the perspective of CFG, hoping to capture both grammatical and semantic information. Therefore, this article sim... | Rebuttal 1:
Rebuttal: We appreciate your constructive feedback on our paper and would like to respond to your comments and concerns.
#### **Capturing complex dependencies:**
We acknowledge that our experimental section could be more explicit in comparing and analyzing the ability of our proposed PCFG-NAT method to cap... | null | null | null | null | null | null |
Training shallow ReLU networks on noisy data using hinge loss: when do we overfit and is it benign? | Accept (spotlight) | Summary: The paper studies overfitting when training shallow ReLU networks with the hinge loss using noisy data (i.e., with label flipping noise). They consider a setting where the data points consist of a direction $\mathbf{v}$ that determines the (clean) labels, and noise from a sphere orthogonal to $\mathbf{v}$. Thi... | Rebuttal 1:
Rebuttal: We thank the reviewer for their positive review and thoughtful questions which we address below.
- *"Assumption 1 in the main text is different from Assumption 1 in the appendix. Specifically, condition 5 of Assumption 1 in the appendix includes a parameter that does not appear in the assumption ... | Summary: The paper studies training a noisy dataset with label noise using GD on a 2-layer network, without biases, and where the second layer is fixed. The main results are, that under the data distribution defined in the paper (namely, linearly separable with orthogonal noise), different types of overfitting (or no f... | Rebuttal 1:
Rebuttal: We thank the reviewer for their positive review and insightful questions which we address below.
- *"How the results would change if the network had bias terms? or the second layer would also be trained?"* We believe our arguments can readily be extended to cover the case of bias terms and sketch... | Summary: In this work a simple model of linear classification is analyzed to study the phenomenon of benign overfitting and generalization of neural networks with noisy data. A theoretical analysis of the trajectory of GD shows that it can be in three regimes (1) benign overfitting (2) non-benign overfitting and (3) no... | Rebuttal 1:
Rebuttal: We thank the reviewer for their positive review and for appreciating the novel contributions of our work. We are happy to answer the questions raised.
- *"The theoretical result holds for a bounded number of samples (i.e., cannot go to infinity)."* As long as $d$ scales appropriately we do not r... | Summary: This paper studies the benign overfitting (or not) of two-layer ReLU neural networks under the hinge loss. Under a new noisy (almost orthogonal) data setting without separable data setting,
- Theorem 3.1 shows that benign overfitting exists occurs the signal-level parameter $\gamma$ is small in $\Theta(1/n)$;... | Rebuttal 1:
Rebuttal: We would like to thank the reviewer for their time and comments and hope we can address each of them below. We hope that if you are satisfied with our responses you might consider increasing your score.
- *"...emphasize the differentiation between ... the logistic loss and the hinge loss"* A comp... | null | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
CS-Isolate: Extracting Hard Confident Examples by Content and Style Isolation | Accept (poster) | Summary: Summary: This paper presents a solution to infer content factors for classification when learning with the noisy labels especially with hard samples that are close to decision boundary. In order to achieve the goal, they tend to separate the style information from the semantic information with some techniques ... | Rebuttal 1:
Rebuttal: >**Q1. The authors used confusing words "random 1, 2, 3, worst" to call noise types on CIFAR-10N.**
**A1**. The terms "Random 1", "worst," and "aggregate" are directly from the original paper [1]. They correspond to specific noise types on the real-world noise dataset CIFAR-10N. For example, “Wor... | Summary: This paper provides insights into the extraction of hard and confident examples from noisy datasets, which is an important problem when learning with noisy labels. The authors discover that by removing irrelevant style information from the data, hard examples can become more discernible. Building on this findi... | Rebuttal 1:
Rebuttal: >**Q1. It remains unclear how concretely the different augmentations are sampled for each dataset.**
**A1**. In our paper, we utilize 6 types of data augmentation techniques that follow [1].
**Detailed types of data augmentation.** These techniques include shift scale rotation, random crop and ... | Summary: This paper focuses on improving the ability to extract hard confident examples (close to the decision boundary), a critical aspect of noisy label learning. The authors find that a key reason for some hard examples being close to the decision boundary is due to the entanglement of style factors with content fa... | Rebuttal 1:
Rebuttal: >**Q1. Support that hard examples are close to the decision boundary is due to the entanglement of style and content factors.**
**A1**.
- **Visualization of Examples.** In Fig. 1 of Rebuttal PDF, We demonstrate examples that failed to be identified as confident by DivideMix but were successfully... | Summary: The problem of training deep neural networks with label noise is considered. Unlike traditional methods proposing to extract confident examples far from the classification boundary, this work proposes an approach named CS-Isolate to extract hard examples close to the classification boundary by disentangling co... | Rebuttal 1:
Rebuttal: We sincerely thank you for your time and effort to review our manuscript and please find the response to your comments below.
>**Q1. The effectiveness of CS-Isolate on large-scale datasets and real-world label noise (e.g., Clothing1M) is worth investigating.**
**A1**. The results on the Clothing... | Rebuttal 1:
Rebuttal: Dear Reviewers,
**We sincerely thank you all for reviewing our paper and providing insightful feedback to help us improve our work.**
We are glad that the reviewers collectively acknowledge the paper's deep **insight into the extraction of hard confident examples** (Reviewer fCED, Reviewer NWTZ... | NeurIPS_2023_submissions_huggingface | 2,023 | Summary: The paper proposed a novel method to extract the hard confidence examples under the noisy labels. By splitting the images into content factors and style factors, the hard confidence examples can be more easily extracted by solely considering the content factors. By training with the hard confidence examples, t... | Rebuttal 1:
Rebuttal: >**Q1. Visualize the hard confident examples and the content/style factors.**
**A1**. We visualize some examples that failed to be identified as confident examples by DivideMix, but are successfully identified by our method in Fig. 1 of Rebuttal PDF.
We also visualize some examples that are succe... | null | null | null | null | null | null |
Optimizing protein fitness using Bi-level Gibbs sampling with Graph-based Smoothing | Reject | Summary: This paper proposes a sampling method for directed protein evolution. The method includes three parts: Bi-level Gibbs sampling (BiG), graph-based smoothing (GS), and iterative extrapolation (IE). BiG improves sampling efficiency for multi-point mutant sequences, GS helps to sample toward high fitness sequences... | Rebuttal 1:
Rebuttal: We thank the reviewer for their time and insightful feedback. Below we address the reviewer’s concerns and questions.
> The proposed approach is somewhat heuristic and lacks theoretical assurance.
Our approach is well-grounded in the theoretical guarantees of Gibbs With Gradients (GWG) [1]. Theo... | Summary: This work applies gibbs sampling methods for generating random sequences of proteins to find mutants with high fitness. The model is compared against a set of sampling methods and the results are evaluated on GFP by fitness, diversity, and novelty.
Strengths: - The problem formulation and model construction ... | Rebuttal 1:
Rebuttal: We thank the reviewer for their time and insightful feedback. Below we address the reviewer’s concerns and questions.
> The motivation for random sampling the whole sequence for a protein in the context of is unclear.
Could the reviewer clarify which component of our method they refer to? Our met... | Summary: This paper proposes a novel method to optimize protein fitness by using bi-level Gibbs sampling with graph-based smoothing. Experiments on real datasets are used to verify the effectiveness of the proposed method.
Strengths: 1. The studied problem about designing novel protein is interesting and challenging. ... | Rebuttal 1:
Rebuttal: We thank the reviewer for their time and insightful feedback. Below we address the reviewer’s concerns and questions.
> The writing can be improved. There exist many typos and grammatical errors. Line 106, where is Algorithm 2? Furthermore, the organization can also be improved. For example, Sect... | Summary: This work considers the problem of optimization over a noisy fitness landscape of sequences. There are two main contributions: 1) Graph-based smoothing (GS) --- considering the landscape as a noisy landscape and applying L1 graph Laplacian regularization. 2) Bi-level Gibbs (BiG) --- using bi-level sampling to ... | Rebuttal 1:
Rebuttal: We thank the reviewer for their time and insightful feedback. Below we address the reviewer’s concerns and questions.
> The main weakness is in the empirical evaluation. The evaluation is based on machine learning models as oracles…. it is notoriously challenging to evaluate protein sequence desi... | Rebuttal 1:
Rebuttal: We sincerely thank the reviewers for their constructive feedback. Our manuscript has improved as a result. We are glad reviewers noted the novelty of graph-based smoothing and Markov Chain Monte Carlo (MCMC) to protein fitness optimization. Here we highlight important improvements and responses to... | NeurIPS_2023_submissions_huggingface | 2,023 | Summary: In this paper, the authors proposed a Gibbs sampling algorithm with graph signal smoothing that can generate high fitness protein sequence. The proposed method, BiGGS, constructs a smoothed space between protein sequence nodes, and use the gradient of a trained fitness predictor to guide the sampling of both n... | Rebuttal 1:
Rebuttal: We thank the reviewer for their time and insightful feedback. Below we address the reviewer’s concerns and questions.
> The complexity of method is quite high, especially the retraining step after smoothing, which could be a major bottleneck of optimizing the methods in other use cases.
We addre... | Summary: The paper introduces BiGGS, a new sequence-based protein fitness optimization algorithm. BiGGS employs bi-level Gibbs sampling for efficient mutation sampling, graph-based smoothing to regularize the fitness landscape, and iterative extrapolation for progressive mutation towards higher fitness. The algorithm's... | Rebuttal 1:
Rebuttal: We thank the reviewer for their time and insightful feedback. Below we address the reviewer’s concerns and questions.
> While the paper presents compelling results, it could be further enhanced by incorporating a case study that applies the proposed method in a real-world scenario…
We agree a ca... | null | null | null | null |
When Can We Track Significant Preference Shifts in Dueling Bandits? | Accept (poster) | Summary: This paper studies the non-stationary multi-armed dueling bandit under the SST and STI conditions. The paper first shows a hardness result that SST+STI are necessary for sub-linear regret, while merely SST or STI cannot guarantee sublinear regrets. Then, the paper studies under SST+STI conditions and propose a... | Rebuttal 1:
Rebuttal: * _Lower Bound and $L$ vs. $\tilde{L}$_: the lower bound is in fact $\sqrt{\tilde{L} KT}$ since the previous work’s lower bound construction in terms of $L$ switches in fact uses $\tilde{L}=L$ (Theorem 5.1; Saha and Gupta, 2022). This is also stated in Remark 3.1 of the cited paper Saha and Buenin... | Summary: The paper tackles a regret minimizing problem in a dueling bandits scenario, where the underlying preference probabilities are non-stationary but instead shifting over time. In case these preference probabilities fulfill SST (strong stochastic transitivity) and STI (stochastic triangle inequality) at any time,... | Rebuttal 1:
Rebuttal: Thanks for the close reading, pointing out typos, and useful writing suggestions.
* _On experiments_: please see the general author rebuttal.
* _$\sqrt{K}$ Regret in Other Subclasses_: indeed there may be carefully constructed subclasses/instances outside of $SST\cap STI$ where $\sqrt{\tilde{L} KT... | Summary: This work studies the problem of a non-stationary bandits problem. First of all, the hardness of dynamic regret minimization under significant shifts problem without SST+STI condition is discussed. Then, with such additional conditions, an algorithm is proposed with a regret of $O(\sqrt{KLT})$. An innovative a... | Rebuttal 1:
Rebuttal: Thanks for your positive comments. Please see simulation results in the general author rebuttal.
---
Rebuttal Comment 1.1:
Comment: Thanks for the response. You have addressed my concerns. | Summary: This paper addresses the problem of dueling bandits with distribution shifts in preferences. The authors investigate the possibility of designing an adaptive algorithm with dynamic regret that depends on the number of significant shifts in preferences. They provide an impossibility result for well-studied pref... | Rebuttal 1:
Rebuttal: Thanks for encouraging comments!
* _On experiments_: please see preliminary experimental results in the general author rebuttal.
* _Innovation in analysis_: the key innovation is explained in Section 5.1 of the paper. At first glance, the analysis appears challenging as optimal dueling algorithms ... | Rebuttal 1:
Rebuttal: We thank reviewers for their time and careful reading.
A common concern raised by reviewers was the lack of empirical experiments. We'd like to emphasize that the main contribution of the paper is a theoretical understanding of achievability: whether $\sqrt{K}$ dependence in regret is possible in... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
AlberDICE: Addressing Out-Of-Distribution Joint Actions in Offline Multi-Agent RL via Alternating Stationary Distribution Correction Estimation | Accept (poster) | Summary: the paper presents a novel approach, AlberDICE, an offline MARL algorithm that performs centralized training of individual agents based on stationary distribution optimization, to tackle the distribution shift challenge in offline MARL. The proposed method avoids OOD joint action selection by computing the bes... | Rebuttal 1:
Rebuttal: # Response to Reviewer $\text{\textcolor{blue}{5Yz6}}$
Thank you for taking the time to provide detailed feedback on our paper. Please feel free to let us know if any of our responses below need further clarification. We’d be more than happy to address them during the reviewer-author discussion ... | Summary: This paper proposes a new algorithm called AlberDICE for addressing the challenge of out-of-distribution joint actions in offline Multi-Agent Reinforcement Learning (MARL). The proposed algorithm performs centralized training of individual agents based on stationary distribution optimization, which circumvents... | Rebuttal 1:
Rebuttal: # Response to Reviewer $\text{\textcolor{red}{aFwQ}}$
Thank you very much for the detailed review. Please feel free to let us know if any of our responses below need further clarification. We’d be more than happy to address them during the reviewer-author discussion period.
> **W1. Since $\pi_{... | Summary: In this paper, the authors propose an alternative multi-agent offline reinforcement learning algorithm, where each agent is optimized for the best response policy and avoiding the out-of- distribution joint actions. The theoretical part proves the nash policies convergence after the presented training paradigm... | Rebuttal 1:
Rebuttal: # Response to Reviewer $\text{\textcolor{green}{82R1}}$
Thank you for taking the time to review our paper. Please feel free to let us know if any of our responses below need further clarification. We’d be more than happy to address them during the reviewer-author discussion period.
> **W1. The s... | Summary: The authors consider the offline multi-agent reinforcement learning setting. They provide an example to demonstrate that decentralisation can lead to out of distribution actions at test time. They derive AlberDICE, an offline RL method based on computing alternating best responses to the other agents' policies... | Rebuttal 1:
Rebuttal: # Response to Reviewer $\text{\textcolor{magenta}{gucm}}$
Thank you for taking the time to provide detailed feedback on our work. We are happy to hear that you have enjoyed reading our paper. Please feel free to let us know if any of our responses below need further clarification. We’d be more th... | Rebuttal 1:
Rebuttal: # General Response
We thank all reviewers for taking the time to provide constructive feedback on our paper. We are encouraged that they find the following strengths in our work:
- Clear presentation of the problem with OOD joint actions and the failures of previous approaches (Reviewers $\text... | NeurIPS_2023_submissions_huggingface | 2,023 | Summary: This paper introduces the AlberDICE framework that targets to address the out-of-distribution challenge (or distributional shift) in off-policy MARL. AlberDICE addresses this challenge by the alternating optimization procedure along with the regularization term. Also, AlberDICE addresses the curse of dimension... | Rebuttal 1:
Rebuttal: # Response to Reviewer $\text{\textcolor{purple}{s9Ry}}$
Thank you for taking the time to provide feedback on our paper. Please feel free to let us know if any of our responses below need further clarification. We’d be more than happy to address them during the reviewer-author discussion period.
... | null | null | null | null | null | null |
Deep Stochastic Processes via Functional Markov Transition Operators | Accept (poster) | Summary: The paper introduces a new class of Bayesian nonparametric models by extending neural processes. The fundamental idea is to create a Markov chain of stochastic processes, culminating in a flexible enough stochastic process.
Strengths: The paper introduces a simple but a clever modification of vanilla neural ... | Rebuttal 1:
Rebuttal: We are grateful for the reviewer’s constructive feedback and have provided our responses below:
> *“The paper uses an unclear notation for probabilistic notions. It’s not just that this choice is unconventional, forcing the reader to keep a mental map of how things are redefined, but also that it... | Summary: This paper proposes Markov neural process, a new class of stochastic processes. The work is based on prior art on neural process, which was published in 2018. The paper is reasonably written. The effectiveness of the new method is very clear from the experimentations.
Strengths: 1. The proposed markov neural... | Rebuttal 1:
Rebuttal: We thank the reviewer for dedicating their time to evaluating our manuscript. Below are our responses:
> *“The explanation of the overall MNP is hard to follow. Fig 2 is crucial to explain the overall structure, plus the training and inference flow. But it's unclear. For example, the three type o... | Summary: The paper attempts to improve the flexibility of Neural Processes without sacrificing exchangeability and consistency properties of classical stochastic processes. The paper proposes a diffusion-like iterative refinement of the predicted values, however, ensuring exchangeability and consistency.
Strengths: - ... | Rebuttal 1:
Rebuttal: We are grateful to the reviewer for their detailed comments. Below are our responses:
> *“I feel Eq. (11) is somehow incomplete considering that y_i^{(t)} (for t=0,..., T-1) are also random variables that do not appear here. Are they absorbed in the reconstruction term (i.e., the first term.”*
T... | Summary: The paper presents Markov Neural Processes (MNPs), a novel class of Stochastic Processes (SPs) that leverage neural networks for data modeling. MNPs enhance the flexibility and expressivity of the Neural Processes (NPs) framework without sacrificing consistency or imposing restrictions. The proposed iterative ... | Rebuttal 1:
Rebuttal: We deeply appreciate the time and effort the reviewer has dedicated to evaluating our paper. Please see below for our responses to the questions raised.
> *“How is the scalability of this method? Were any tricks used to make the method more scalable? Can you provide an analysis on the time/space ... | Rebuttal 1:
Rebuttal: Here we updated Figure 2 in the uploaded PDF, which illustrates the generative and inference models of MNPs. In the reviewer rebuttals below, we address each reviewer's questions and comments separately.
Pdf: /pdf/5a1dd47e421568890fe0760e2101b7f36d783471.pdf | NeurIPS_2023_submissions_huggingface | 2,023 | Summary: The paper introduces Markov Neural Processes (MNPs), a new variant of Stochastic Processes (SP). The paper points out that predictive SP models that construct context-to-target mappings no longer satisfy conditional consistency property and also are not fully generative. An expressive NP variant called MNP is ... | Rebuttal 1:
Rebuttal: We sincerely thank the reviewer for their feedback on our manuscript. Below are our responses to the questions raised.
> *“Figure 2: I could not understand clearly why the auxiliary latent variables should be inferred in the reverse order. Can you expand further?”*
The auxiliary latent variables... | null | null | null | null | null | null |
Hybrid Policy Optimization from Imperfect Demonstrations | Accept (poster) | Summary: This paper addresses the exploration challenge in RL by introducing imperfect demonstrations. These demonstrations are used to train an offline guider policy through imitation learning. Subsequently, the guider policy is employed to provide instructions during online training, thereby significantly enhancing t... | Rebuttal 1:
Rebuttal: We thank the reviewer for the insightful and valuable feedback. We explain the concerns point by point below.
**Q1: About the superiority of HYPO and the ablation study.**
A: We further elaborate on the superiority of our algorithm compared to other relevant methods and conduct more ablation exp... | Summary: This work proposes a reinforcement learning algorithm to learn in a sparse reward environment with the help of data from a sub-optimal expert. The proposed algorithm, HYbrid Policy Optimization (HYPO) consists of three main parts, a discriminator, an offline guider, and an online agent. The primary idea is to ... | Rebuttal 1:
Rebuttal: We thank the reviewer for the insightful and valuable feedback. We explain the concerns point by point below.
**Q1: The ablation study of the input of discriminator.**
A: We conducted additional ablation experiments to emphasize the significance of utilizing the offline guider as an additional s... | Summary: The paper introduces a method (HYPO) for RL using LfD, specifically in the setting where available demonstrations are suboptimal and/or incomplete. The method maintains a discriminator trained to differentiate agent-generated $(s,a)$ pairs from those available as demonstrations. The method further trains two p... | Rebuttal 1:
Rebuttal: We thank the reviewer for the insightful and valuable feedback. We explain the concerns point by point below.
**Q1: About the incomplete demonstrations.**
A: In our method, "incomplete" indicates that expert trajectories lack reward signals. Specifically, for offline learning, complete trajecto... | Summary: This paper presents HYPO, a Learning from demonstration (LfD) method to effectively learn policy from a limited amount of imperfect demonstrations. At the core of the method are a hybrid discriminator to tell positive/negative state-action pairs through positive-unlabeled reward learning, an offline imitation ... | Rebuttal 1:
Rebuttal: We thank the reviewer for the insightful and valuable feedback. We explain the concerns point by point below.
**Q1: About the presentation and confusion of notations.**
A: In the paper, the **guider** is trained using **offline imitation learning**, and thus, we use the term offline imitation po... | Rebuttal 1:
Rebuttal: We would like to express our sincere gratitude for the thorough review and valuable feedback on our paper. The reviewers' insights and suggestions are quite helpful to improve the quality and clarity of our work. In this author rebuttal, we conduct additional experiments about the **ablation study... | NeurIPS_2023_submissions_huggingface | 2,023 | Summary: The goal of this work is to develop an approach for RL in the sparse reward setting, wherein the RL agent is directed by a guidance policy that provides direction to the online learner. While this approach has been used before, the main contribution of this work is on developing a guidance policy that utilize... | Rebuttal 1:
Rebuttal: We thank the reviewer for the insightful and valuable feedback. We explain the concerns point by point below.
**Q1: About the decaying of the KL regularization coefficient.**
A: The primary purpose of the regularization term here is to facilitate knowledge distillation from the guidance policy t... | null | null | null | null | null | null |
On Dynamic Programming Decompositions of Static Risk Measures in Markov Decision Processes | Accept (poster) | Summary: The paper shows that the dynamic programming decomposition methods used in risk-averse MDP are suboptimal for CVaR and EVaR, but optimal for VaR. These findings seem important as such decomposition methods become increasingly popular recently.
Strengths: The findings that the decomposition fails for CVaR and... | Rebuttal 1:
Rebuttal: We thank the reviewer for the constructive comments.
First, we need to clarify a fact that the reviewer might have missed while reading the paper. Namely, both CVaR and EVaR do possess valid dynamic programming-based solution methods, as presented in Bäuerle and Ott. (2011) and Hau et al. (2023) ... | Summary:
The paper tackles the risk-level decomposition of risk measures in the finite Markov Decision Process (MDP). The work firstly points out the suboptimality of the decomposition proposed in prior works (Conditional-Value-at-Risk), by mathematically proving that the decomposition in CVaR does not maintain the eq... | Rebuttal 1:
Rebuttal: Thank you for your review and suggestions. Adding a numerical experiment is a good suggestion. In the final version of the paper, we will add a simple example that demonstrates how the duality gap translates into a sub-optimal policy. Please see the attached pdf with the global response, where we ... | Summary: The main contribution of the paper is to demonstrate that a popular dynamic programming decomposition for well-known risk metrics only provides suboptimal policies. More precisely, the underlying idea of such decompositions is to operate on an extended state space that incorporates a continuous state variable ... | Rebuttal 1:
Rebuttal: Thank you for the detailed review and the encouraging words. We answer the questions and clarify some of the points below.
**Lack of discussion of possible research avenues to address this issue.**
This is a great point. Given how widespread this error is in the extant literature, we believe that... | Summary: The authors demonstrate that dynamic programming decompositions for CVaR and EVaR can be suboptimal, and show that VaR may not be subject to the same problems in MDPs of horizon 1.
Strengths: The paper provides some interesting observations conflicting with previous papers on CVaR and EVaR optimization in RL... | Rebuttal 1:
Rebuttal: We thank the reviewer for the constructive comments, which encourage us to suggest adding to the appendix of the paper a natural extension of the VaR decomposition presented as Theorem 5.2 to the general case of $T>1$ (more on this below). Indeed, we had not noticed that a reader might consider Se... | Rebuttal 1:
Rebuttal: We would like to take this opportunity to provide a holistic response to the reviewers' main questions and concerns below.
**Concern 1: Paper does not present a complete algorithm or numerical results**
This paper aims to correct the literature on risk-averse RL and retard the propagation of inc... | NeurIPS_2023_submissions_huggingface | 2,023 | Summary: This paper investigates the effectiveness of decomposition approaches for solving risk-averse Markov Decision Processes (MDPs) with Conditional Value-at-Risk (CVaR), Expected Value-at-Risk (EVaR), and Value-at-Risk (VaR) objectives. The goal is to assess the validity and accuracy of these common decomposition ... | Rebuttal 1:
Rebuttal: We appreciate the detailed review and the helpful suggestions. We provide some clarification and answer your questions below.
**It would be valuable if the authors could mention other state-of-the-art techniques in a related works section.**
That is a good suggestion; thank you. In fact, the intr... | null | null | null | null | null | null |
ODE-based Recurrent Model-free Reinforcement Learning for POMDPs | Accept (poster) | Summary: This manuscript presents GRU-ODE, a new ODE-based recurrent model designed to tackle POMDP control tasks. By utilizing a numerical ODE solver, GRU-ODE effectively incorporates past observations, actions, and rewards into its context. In comparison to previous RNN-based methods, GRU-ODE offers several advantage... | Rebuttal 1:
Rebuttal: We thank the reviewer for recognizing the contributions of our paper and giving us constructive suggestions. We will address the reviewer’s concerns in the following parts. We address all of your concerns and show distinctions between our methods with previous works [1]. We hope that our work will... | Summary: The paper proposes GRU-ODE, which encodes historical observations, actions, and rewards into a latent variable. This model is based on recurrent networks and modifies the GRU topology by separating its memory cell into time-continuous states. The integration of variables is accomplished through neural ODEs. Gi... | Rebuttal 1:
Rebuttal: We thank the reviewer for recognizing the contributions of our paper and giving us constructive suggestions. We have endeavored to address all of your inquiries and provide explanations for the issues raised in the weaknesses section. We aspire to resolve your concerns about assumptions and ablati... | Summary: The authors propose a neural ODE-based GRU for use in partially observable RL. They effectively "filter" GRU recurrent states using an ODE gate, before feeding the filtered state back into the GRU. They use a variational approach, where their model $g$ spits out a mean and variance. They find that training wit... | Rebuttal 1:
Rebuttal: We thank the reviewer for recognizing the positive aspects of our paper, and we will address the reviewer’s concerns in the following parts.
**W1**: computation
We agree with your statement that calculating ODE at each step requires additional computational resources. However, we would like to ... | Summary: The author present a method for doing deep reinforcement learning for POMDPs with uneven time between observations. Their method uses TD3 with Neural-ODE+GRU encoders for the actor and the critic. They test on various partially observed control tasks, as well as tasks with uneven observation times.
Strengths:... | Rebuttal 1:
Rebuttal: We thank the reviewer for recognizing the positive aspects of our paper, and we will address the reviewer’s concerns in the following parts.
**W1**: POMDP setting
We agree with your opinion about the POMDP framework, during our experiments, we found almost all the baselines are proposed on posit... | Rebuttal 1:
Rebuttal: We are incredibly grateful for the positive assessment and highly detailed and constructive feedback on our paper. In this rebuttal, we have addressed the comments of the review team. This has led to multiple improvements including a better exposition of the proposition, more complete ablation stu... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Residual Q-Learning: Offline and Online Policy Customization without Value | Accept (poster) | Summary: This paper proposes a problem setting called policy customization, which involves training a policy that inherits the characteristics of a prior policy while satisfying additional requirements imposed by a target downstream task. They present a novel approach to interpreting and determining the two task object... | Rebuttal 1:
Rebuttal: > The derivation process in Equation 5 lacks clarity, particularly the transition from eq. 5b to eq. 5c and eq. 5d. It is necessary to provide a more detailed and step-by-step derivation to enhance the understanding of the readers.
Thank you for your suggestion on including a step-by-step deriva... | Summary: This paper proposes a new problem setting called pulicy customization, where the goal is to train a policy that inherits the characteristics of a prior policy while satisfying some additional requirements imposed by a downstream task. To solve this problem, this paper uses a novel and principled framework by f... | Rebuttal 1:
Rebuttal: Thank you for your valuable feedback. We are very excited that you find policy customization a novel problem setting and find the proposed residual Q-learning framework a principled and flexible solution to the policy customization problem.
> Does the performance of residual-Q highly rely on the... | Summary: << I have read the authors' rebuttal and have raised my score based on the discussion >>
This paper introduces a novel problem setting, named 'policy customization,' which seeks to train a new policy that inherits the properties of an existing pre-trained policy and meets additional requirements from a given ... | Rebuttal 1:
Rebuttal: Thank you for your thoughtful feedback and suggestions. We are very excited that you find our paper commendably articulated, novel, and clearly motivated. We are also glad to have your recognition of the technical soundness of the problem formation and derivation. Please find below our responses t... | Summary: This paper proposes a new problem setting called policy customization, where the goal is to train a new policy that inherits the characteristics of a prior policy while satisfying additional requirements imposed by a downstream task. The authors propose a novel Residual Q-Learning framework to solve this probl... | Rebuttal 1:
Rebuttal: Thank you for your thoughtful feedback and valuable suggestions. We are glad you find the policy customization problem and residual Q-learning framework novel. Please find below our responses to address your concerns regarding our work:
> The evaluation for the proposed method is limited. It wou... | Rebuttal 1:
Rebuttal: We thank the reviewers for their thoughtful feedback and suggestions. We are delighted that the reviewers found policy customization novel (R1, R2, R3, R4, and R5) and essential (R1, R5). We are glad that they found residual-Q novel (R1, R2, R3, R4, and R5), principled (R2, R4), and flexible (R4).... | NeurIPS_2023_submissions_huggingface | 2,023 | Summary: This paper introduces a novel concept policy customization, which trains a new policy based on a pre-trained policy so as to solve the downstream tasks. Compared to RL fine-tuning algorithms, policy customization directly optimize the imitative objectives and downstream RL objectives jointly. The authors propo... | Rebuttal 1:
Rebuttal: Thank you for your thoughtful feedback and valuable suggestions. We are glad you find the policy customization problem novel and our theoretical formulation sound. Please find below our responses to address your concerns regarding our work:
> The author should clarify the significance of policy ... | null | null | null | null | null | null |
Echoes Beyond Points: Unleashing the Power of Raw Radar Data in Multi-modality Fusion | Accept (poster) | Summary: This paper presents a method to perform image and raw radar fusion for object detection from a Mobile platform. The authors explain that the commonly used radar data processing pipelines discard too much data, resulting in very sparse point clouds. In contrast, the proposed method operates on raw data and perf... | Rebuttal 1:
Rebuttal: Thank you for your thoughtful comments and advice. The responses to the main concerns are as follows:
## Q1: Clumsy formulations and jargons should be avoided
Thanks for reminding! Our motivation involves some concepts of radar signal processing. We are trying our best to make them intuitive and... | Summary: This paper proposes a Polar-Aligned Attention module to guide the fusion of radar data and images. The module projects the image and radar features onto a polar coordinate system to calculate the similarity of the associated information. This effectively enhances the direct fusion of radar data and images. Add... | Rebuttal 1:
Rebuttal: We are grateful for your valuable comments and constructive advice, which helps us a lot to make the paper better. The responses to the main concerns are as follows:
## Q1: Whether the novelty and contribution of this work is significant
Indeed, our work is based on PolarFormer, but the fusion m... | Summary: This work presents a method for camera-radar detection and fusion. The authors propose an approach that follows:
Yanqin Jiang, Li Zhang, Zhenwei Miao, Xiatian Zhu, Jin Gao, Weiming Hu, and Yu-Gang Jiang. Polarformer: Multi-camera 3d object detection with polar transformers. AAAI, 2023.
and
Julien Rebut, Ar... | Rebuttal 1:
Rebuttal: Thank you very much for your kind comments on our manuscript. The responses to the main concerns are as follows:
## Q1: Whether the novelty and contribution of this work is significant
Indeed, our work is based on PolarFormer, but the fusion mechanism is newly designed because raw radar data has... | Summary: The paper proposes a method for fusing raw radar data with images in bird’s eye view (BEV) space. To this end, the method extract multilevel features from images and radar and uses polar-aligned attention mechanism with polar-BEV decoder for bounding box prediction. Evaluation is done on re-annotated RADIal da... | Rebuttal 1:
Rebuttal: We sincerely thank you for the very professional comments. We are inspired and hope our discussion brings more insights. The responses to the main concerns are as follows:
## Q1: How robust is the proposed model to spatio-temporal misalignment?
Both spatial and temporal misalignment can be allev... | Rebuttal 1:
Rebuttal: Dear all,
Thanks again for the very constructive comments and for spending your valuable time reading our responses! We are open to any further discussions during the author-reviewer discussion period (Aug 9-16). PDF material contains Figure R1-R4 used in personal replies.
## More Experiments on... | NeurIPS_2023_submissions_huggingface | 2,023 | Summary: The paper proposes a new method to fuse radar and imagery data for autonomous driving. Instead of using processed radar data, which discards most of the information in the original signal while attempting to filter out noise, this method leverages the raw data directly as input to the network. The method is b... | Rebuttal 1:
Rebuttal: We really appreciate your affirmation of our work and constructive advice! The responses to the main concerns are as follows:
## Q1: Possibility of expanded annotation and tasks beyond vehicle detection
In the original dataset there was only one class. In our refined annotation, we provide fine ... | null | null | null | null | null | null |
High-dimensional Contextual Bandit Problem without Sparsity | Accept (poster) | Summary: The paper addresses the problem of high-dimensional linear contextual bandits where the number of features is greater than the budget or even infinite. Unlike previous works, the paper does not impose sparsity on the regression coefficients but instead leverages over parameterized models. The authors propose a... | Rebuttal 1:
Rebuttal: We thank your careful reading and insightful comments. The score is a bit low even though no critical issue was raised. If the main concerns are the unclarity of the problem setting, we would be happy to address any concerns. Regarding the clarification on the benign covariance, we will revise the... | Summary: This paper considers a high-dimensional contextual linear bandit setting and uses recent results on
the risk of minumum
norm linear interpolators to evaluate the regret of two variants of an explore-then-commit algorithm.
The first variant makes use of the risk results to balance regret in the "explore" vs "... | Rebuttal 1:
Rebuttal: Thank you for your time and many comments.
> It is unclear to me why the paper is focused on covariance structures of the type in Proposition 1. Presumably this is because these are the covariances for which the minimum norm interpolators have non-trivial risk guarantees, but that seems like a w... | Summary: The authors studied the high-dimensional linear contextual bandit problem where the number of features 𝑝 is greater than the budget 𝑇, or it may even be infinite.
Problem setting: They assumed each arm has an individual hidden parameter, and the contexts are drawn from nature for each round.
They argue t... | Rebuttal 1:
Rebuttal: We thank your careful reading and insightful comments. The score is a bit low compared with the tone of the review. The main questions raised by the reviewers are about the problem setting. To address this, we will revise the paper as outlined in the global reply. Below, we answer each of the ques... | Summary: This paper addresses the linear contextual bandit problem in high dimensions where the number of features is greater than the budget. While the recent works impose an assumption about sparsity on the regression coefficients, this paper uses the minimum-norm interpolating estimator which is inspired by recent ... | Rebuttal 1:
Rebuttal: Thank you for your insightful review. These questions are fundamental to the benign assumption, and we will revise the paper based on your comments. Below, we answer the questions raised.
> Although the paper introduces a new direction for the high-dimensional linear contextual bandits, the autho... | Rebuttal 1:
Rebuttal: Thank you all for your thoughtful and important comments.
We are making the following major changes based on your comments.
> Add more introduction to benign covariance and related notion
To enhance our readers' comprehension of overparameterized models, we propose to provide a more comprehensi... | NeurIPS_2023_submissions_huggingface | 2,023 | Summary: This paper analyzes the Explore-then-commit algorithm for the linear contextual bandit problem in high dimensions, when the number of features is higher than the number of rounds T. The authors use recent studies on overparametrized models for linear regression to analyze the algorithm. Since this algorithm re... | Rebuttal 1:
Rebuttal: Thank you for your insightful review. We would improve the introduction of benign overfitting theory as suggested.
> The introduction of the term benign covariance is a bit abrupt and very technical very quickly. The authors could provide more intuition about what it means before defining the bi... | null | null | null | null | null | null |
Large Language Models can Implement Policy Iteration | Accept (poster) | Summary: This paper examines the possibility of taking advantage of the ability of in-context learning to use large language models to implement policy iteration so that large language models can be used to solve reinforcement learning problems without querying the gradients. Experiments have been conducted on 5 artifi... | Rebuttal 1:
Rebuttal: > The applications of ICPI are limited. Right now it can only be applied to handcrafted small environments like chain and mini catch. It is unclear if large language models can do complex planning in more practical tasks such as Atari and mujoco tasks. Additionally, the current formulation of ICPI... | Summary: This paper formalizes a framework where language model can use its in-context learning capability to predict actions and Q-values, and the training occurs by updating the prompt with the latest history. As a result, the LM can perform reinforcement learning without gradient. The paper explored a few language m... | Rebuttal 1:
Rebuttal: > While the formalization seems nice, it is not particularly novel; the empirical results/datasets presented also seem toy and low-hanging fruits.
The Policy Iteration algorithm itself is, of course, many decades old. What is novel is the in-context implementation of policy iteration using LLMs. ... | Summary: The authors show how LMs can solve simple MDPs by implementing policy iteration in-context. They do that by constructing prompts that simulate policy rollouts using partial trajectories from an experience buffer as few-shot examples. They empirically validate their approach on six toy MDPs that test for specif... | Rebuttal 1:
Rebuttal: > The authors attribute the relatively good performance of ICPI its “ability to generalize from its context to unseen states and state/action pairs (unlike “Tabular Q” and “Matching Model”)” (lines 296-298). How is that different from function approximation in more traditional RL?
Yes, this is a ... | Summary: The paper demonstrates a way to implement policy iteration by means of in-context learning with a large language model. Authors suggest ICPI, a reinforcement learning algorithm that performs policy evaluation and policy improvement solely by encoding tokens and inferencing the model, without updating model par... | Rebuttal 1:
Rebuttal: > To the best of my knowledge, there is nothing in the paper that evaluates if ICPI is "faster" than equivalently tuned gradient-based methods.
Thank you for pointing out this omission. [This link](https://www.overleaf.com/read/ncjhbrxkvyrz) provides a demonstration of the performance of Proxima... | null | NeurIPS_2023_submissions_huggingface | 2,023 | Summary: The problem of using large-pretrained foundation models (LLMs) like codex as backbones for learning RL policies without gradient updates or expert demonstrations is addressed by the authors. They propose an algorithm called In-Context Policy Iteration (ICPI) to perform model-based RL policy iterations using a ... | Rebuttal 1:
Rebuttal: > To perform Q-value computation, the algorithm needs to unroll action sequences for all actions. Unfortunately, this greatly restricts the algorithm's usefulness to finite action-spaces.
You are correct that we only demonstrated the method using discrete action domains. However we suggest two co... | null | null | null | null | null | null |
Diffusion Models and Semi-Supervised Learners Benefit Mutually with Few Labels | Accept (spotlight) | Summary: In this work, the authors propose a novel training strategy, dual pseudo training (DPT), that involves three steps: First, DPT uses state-of-the-art semi-supervised learning methods to predict pseudo-labels on a partially-labeled dataset and diffusion models. Second, DPT trains a conditional generative model w... | Rebuttal 1:
Rebuttal: We thank reviewer JKn2 for the interest and acknowledgement of our contributions and the valuable comments. We respond below to your questions and concerns.
## Weakness1: claim may mislead to the readers.
We appreciate your feedback and understand your concern. We will follow your suggestion an... | Summary: To advance semi-supervised generative and classification tasks, this paper proposes a simple yet effective training strategy termed dual pseudo training (DPT). The proposed DPT operates in three stages: training a strong semi-supervised learner to predict pseudo labels, training a strong diffusion model using ... | Rebuttal 1:
Rebuttal: We thank reviewer hJXD for the acknowledgement of our contributions and the valuable comments. We respond below to your questions and concerns.
## Weakness1: Regarding technical novelty.
Thank you for your feedback. Our paper's novelty lies in answering the two key questions and in exploring ho... | Summary: This work presents an interesting interaction between a generative model and a semi-classifier. The pipeline can boost both the generative sample quality and discriminative performance of the semi-classifier in the few label settings. A three-stage pipeline is designed. First, the semi-classifier is trained on... | Rebuttal 1:
Rebuttal: We thank the reviewer wQJF for the valuable and constructive comments. We respond below to your questions and concerns.
## Weakness1: Claim about DPT with four labels per class may be incorrect.
We appreciate your pointing out the work [c] and your valuable suggestion. DPT with four labels per ... | Summary: This paper proposes to exploit generative diffusion models to benefit semi-supervised learning and vice versa.
Specifically, the proposed training strategy called dual pseudo training (DPT) consists of 3 stages:
(1) Train a semi-supervised learning (SSL) model on labeled and unlabeled data
(2) Train a condi... | Rebuttal 1:
Rebuttal: We thank reviewer 2c3F for the interest and acknowledgement of our contributions and the valuable comments. We respond below to your questions and concerns.
## Weakness1: the novelty is limited.
Thank you for your feedback. Our method is simple but effective. The novel contributions of our work ... | Rebuttal 1:
Rebuttal: # Summary of the revision
We sincerely thank the reviewers for their valuable comments, which help to further improve the quality of our work. We have addressed the detailed comments, and summarize the main revision which will be updated in our final version as follows:
## New results
- We **a... | NeurIPS_2023_submissions_huggingface | 2,023 | Summary: This paper empirically studies the effectiveness of combining diffusion models and semi-supervised classification. First, an image encoder is trained using a self-supervised method, then finetuned with a small fraction of the ground-truth labels. Second, the image classifier assigns pseudo-labels to all unlabe... | Rebuttal 1:
Rebuttal: We thank reviewer kGQr for the acknowledgement of our contributions and the valuable comments. We respond below to your questions and concerns.
## Weakness1: Lower FID translates to image quality improvement or better Top-N alignment?
Thank you for your insightful suggestion and pointing out the... | null | null | null | null | null | null |
Exploring Question Decomposition for Zero-Shot VQA | Accept (poster) | Summary: This paper studies the question decomposition of VQA.
To begin with, the authors demonstrate that large VLMs equipped with LLMs can effectively leverage question decomposition to improve answering performance.
Thereafter, some experiments are conducted to prove that instruction-tuned language models are capabl... | Rebuttal 1:
Rebuttal: Great questions, and thank you for the positive appraisal of our work. We have added an experiment for you, as well as a detailed discussion of the of $E_{IC}$ and $E_{CR}$ which should clear up any confusion: we do believe large models can decompose well.
### Why does selective decomposition wo... | Summary: To answer the questions that require reasoning or outside knowledge, this paper proposes a question decomposition method which uses LLM to infer subquestions
Strengths: 1. The proposed method is intuitive and interesting.
2. The methods conduct comprehensive experiments on BLIP2 and FlanT5, and the results c... | Rebuttal 1:
Rebuttal: We appreciate the positive review, and have added the requested experiments and will add the requested references.
### We add experiments with 2 new VLMs and 2 new LLMs.
Thank you for the great suggestion! We have added two additional VLMs (Flamingo and InstructBLIP), and two additional LLMs (Fa... | Summary: The paper examine whether the current instruction-tuned VLM (e.g. BLIP2 ) can (1) learn to generate simple perception questions that can help to answer a more complex reasoning visual question; (2) learn to actually benefit from the perception questions.
The conclusion is that
for (1): language models (with... | Rebuttal 1:
Rebuttal: Thank you for the positive review! These are great questions and we have considered them ourselves.
### Why doesn't visual input help generate better perception questions?
We believe the answer to this is related to the textual biases of the dataset. For example, in the second half of Table 1, we... | Summary: In this paper, the authors focus on question difficulty in visual question answering (VQA). In particular, they propose to give subquestions that improve the percentage of correct answers to answer inferential questions more correctly. Using a large-scale vision-language model (VLM), the authors show that givi... | Rebuttal 1:
Rebuttal: We believe there is a misunderstanding in your summary of our work. We do not propose any method for properly generating subquestions with a specific prompt, we propose a method for *selecting which questions to decompose* based on the model's likelihood of the answer being correct.
This is an em... | Rebuttal 1:
Rebuttal: We thank all reviewers for the detailed feedback and engagement with the paper, and the positive appraisals of our work!
Our rebuttal includes a number of experiments requested by reviewers, and a human study of the decompositions. **Please see the PDF for detailed tables.**
### Additional Exper... | NeurIPS_2023_submissions_huggingface | 2,023 | Summary: This paper presents an investigation of a 1-step question decomposition strategy for VQA.
An example of a decomposition is the sub-question "What is the name of the plane company?" for the visual question "What country headquarters this plane company?" (Figure 1).
This paper finds:
- Including human-annotate... | Rebuttal 1:
Rebuttal: Thank you for the detailed review! We have done our best to respond to everything. Please note the user study.
- We do measure the decomposer method on VQAv2 in the supplement (Table 6). This table also includes $E_{IC}$ and $E_{CR}$ for the human-annotated decompositions.
- Thank you for bringi... | null | null | null | null | null | null |
Stochastic Optimal Control for Collective Variable Free Sampling of Molecular Transition Paths | Accept (poster) | Summary: The paper proposes an interesting method called PIPS for sampling transition paths between metastable states in molecular systems. These transitions are difficult to sample using standard Molecular Dynamics (MD) simulations due to high energy barriers. Traditional methods augment MD with a bias potential based... | Rebuttal 1:
Rebuttal: We thank the reviewer for their detailed comments. We appreciate the reviewers comments regarding the clarity and originality of the work.
We have addressed the reviewer's concerns in detail below. Where possible, we have clustered related questions.
We are looking forward to discussing furthe... | Summary: This paper addresses the task of sampling transition paths between metastable states in molecular systems. This problem is particularly pronounced when dealing with complex systems, where traditional methods often fall short due to the need for chemical intuition and the limitations in scalability. To overcome... | Rebuttal 1:
Rebuttal: We thank the reviewer for taking the time to review our work. We greatly appreciate the reviewers comments regarding the work's novelty and creativeness.
We have the reviewer's concerns below in more detail. We look forward to further discussing our work.
---
> The experiments are conducted on... | Summary: This manuscript addresses the molecular transition pathway sampling problem, which is a fundamental and significant topic in computational chemistry/biology research. Specifically, the authors first made connections between transition path (TP) sampling, the Schrödinger Bridge Problem (SBP), and Stochastic Opt... | Rebuttal 1:
Rebuttal: We thank the reviewer for their time and feedback. We appreciate that the reviewer has found our work to be significant, clearly written, and inspirational.
We will provide detailed comments to the reviewer's major remarks below. Where appropriate, we have grouped comments together. Minor remark... | Summary: Molecular dynamics simulations yield trajectories which show how a system develops from a starting configuration to a final configuration over time.
Given a system with more than one (meta-)stable state, sampling trajectories which include transitions from one (meta-)stable state to another one might occur spa... | Rebuttal 1:
Rebuttal: We thank the reviewer for their detailed review of our work. We are happy to hear that the reviewer finds our work to be addressing a significant problem and in general well written and understandable.
Below, we have included some detailed answers to your questions and comments. We have tried to... | null | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Learning Visual Prior via Generative Pre-Training | Accept (poster) | Summary: The paper proposes VisorGPT to model the propabilistic visual prior via generative pretraining. This is achieved by transforming the continuous values (locations, bounding boxes) into discrete tokens. Prompt engineering are employed to arrange these tokens into a sequence. The author conducts experiments to pr... | Rebuttal 1:
Rebuttal: **The Figures and Tables are provided in the global pdf**
1. The rationale and motivation for utilizing GPT to model visual priors.
There are three main considerations:
i) **Visual priors are not just location**. Our visual priors scope extends beyond the modeling of pure continuous numbers (e... | Summary: This paper proposes to translate image layout information such as bounding box layouts, masks and keypoints into discrete sequences that can be modelled by NLP architectures, such as GPT. By doing so, the image layout distribution of a dataset can be learned to later sample from it and obtain plausible in-dist... | Rebuttal 1:
Rebuttal: **The Figures and Tables are provided in the global pdf**
1. Generalize to novel layouts.
The training set to train VisorGPT contains approximately 4 million images with around 40 million bounding boxes, providing **an extensive and diverse range of layouts to enable novel layout generation**. W... | Summary: This is impressive work, striving to explicitly learn visual priors and enabling customization in the sampling process for conditional image synthesis. Such a task normally occurs inaccurate synthetic results due to improper conditional priors, such as spatial location, shape, and interaction of objects.
Tak... | Rebuttal 1:
Rebuttal: **The Figures and Tables are provided in the global pdf.**
1. Add guidance and precaution to address the closed-set bias.
Thanks for your valuable suggestion. Actually, 1,000 categories learned by VisorGPT can cover most of the common objects in the world. To include more potential categorie... | Summary: The paper proposes a method to learn a high-level generative model for images based on language modeling, which allows the generation of object masks, bounding boxes, and human pose keypoints. The paper proposes a novel tokenization scheme to turn annotations into token sequences that can be fed into an off-th... | Rebuttal 1:
Rebuttal: **The Figures and Tables are provided in the global pdf.**
1. The multimodal training and inference.
Thanks for pointing out. We train VisorGPT using a mix of boxes and keypoints and present some examples in Fig. 1. One can observe that VisorGPT exhibits the capability to infer bounding ... | Rebuttal 1:
Rebuttal: **The Figures and Tables are provided in the global pdf.**
Pdf: /pdf/e9cd4981e6de4f9b674604af694826bc7a2fca8a.pdf | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Feature-Learning Networks Are Consistent Across Widths At Realistic Scales | Accept (poster) | Summary: This paper studies the convergence of key quantities (like test predictions, feature kernels) in Neural Networks as width increases, in feature-learning regimes. The paper is largely an empirical study, and presents width-convergence results on both image and language tasks. The paper also uses ensembling to c... | Rebuttal 1:
Rebuttal: We thank the reviewer for these thoughtful comments. We'd like to address them in sequence.
* *for the qualitiative results on ImageNet experiments on Figure 2b and 3b, it isn't clear to me that these curves have converged in width (say comparing N=256 or 512).*
This is a good point. In the mai... | Summary: The authors study the so-called “feature learning” parameterization of neural networks that is given by the maximal update parameterization (muP). This is in contrast to the NTK parameterization for infinite-widths, which leads to the “lazy learning” regime which is claimed to not learn features, and also in c... | Rebuttal 1:
Rebuttal: We would like to thank the reviewer for their detailed review. We address the concerns and questions below.
1. **Usage of 'bias' does not correspond to usual statistical notion of bias'**: We agree that the way we use bias is not the statistical notion of bias for exactly the reason pointed out b... | Summary: The paper investigates the impact of width on the dynamics of feature-learning neural networks with both vision and language tasks. The empirical results demonstrate that during the early stage of training or during the entire training process for easy tasks, wide neural networks trained on online data not onl... | Rebuttal 1:
Rebuttal:
We thank the reviewer for their comments. We are glad that the paper was received as well-organized with well-motivated results.
We would like to address the main concern: the broader applicability of our findings.
* *"In the conclusion section, the authors assert that the findings 'motivate t... | Summary: This is a primarily empirical paper studying the agreement between real, finite-width neural networks and their limiting behavior under infinite width. Rather than using the neural tangent kernel as the descriptor of infinite-width behavior, the paper focuses on a different parameterization called feature-lear... | Rebuttal 1:
Rebuttal: We thank the reviewer for their useful comments and suggestions. Below we address the various weaknesses and questions which arose in the review.
### Weaknesses
1. We will include a sentence in the abstract specifying that we mean networks in $\mu$P. We agree that we should define what we mean b... | Rebuttal 1:
Rebuttal: We thank the reviewers for their detailed reading of our paper and useful suggestions and comments. In response to many repeated questions and concerns, we have performed the following new experiments (see attached PDF)
#### New Figures and Experiments
1. We have added a comparison of network dyn... | NeurIPS_2023_submissions_huggingface | 2,023 | Summary: In this work, the authors conducted an empirical investigation into the consistency across widths of the training dynamics of neural network models in the maximum-update scaling. Specifically, they demonstrated through experiments on both vision and language tasks that many behaviors in the online learning dyn... | Rebuttal 1:
Rebuttal: We thank the reviewer for suggesting these important related works. We will be sure to include a proper comparison of our study to these previous papers. Below, we provide a brief comparison of our results to each prior work.
#### Comparison to prior works
1. First [1] Chen et al study two layer ... | null | null | null | null | null | null |
Sample Efficient Reinforcement Learning in Mixed Systems through Augmented Samples and Its Applications to Queueing Networks | Accept (spotlight) | Summary: This paper focuses on a specific class of Markov Decision Processes (MDPs) in which a state can be decomposed into a stochastic component and a pseudo-stochastic component. The transition dynamics of stochastic states depend solely on the current stochastic state and the action taken, while the transition dyna... | Rebuttal 1:
Rebuttal: We would like to sincerely thank the reviewer for the constructive comments. We will incorporate these great suggestions in our revision and we'd like to address the major concern in the following.
> (1) Building upon point 1, Assumption 1 poses challenges in designing the function class F for th... | Summary: The paper introduces a novel setting called mixed systems, in which environments comprise both stochastic and deterministic transitions. This leads to two types of states: stochastic states and pseudo-stochastic states. Stochastic states follow a stochastic transition kernel, while pseudo-stochastic states hav... | Rebuttal 1:
Rebuttal: We first would like to sincerely thank the reviewer for the willingness to reevaluate the rating. We recap your comments and present our detailed response as follows.
> (1) Compared with MDPs with exogenous inputs.
- Our discussion from lines 132-137 explains the fundamental difference. For the ... | Summary: This paper consider the sample efficiency of RL in a setting involving "mixed systems" where the state space is factored into a stochastic component and another much larger component that is deterministic when conditioned on the stochastic component. The authors develop an algorithm for this setting relying on... | Rebuttal 1:
Rebuttal: We appreciate your time and thoughtful evaluation of our paper. We recap your comments and present our detailed response as follows.
> (1) How realistic is it that the factorization of stochastic states and pseudo-stochastic states is provided to you? Can the cost of identifying these be bounded?
... | null | null | Rebuttal 1:
Rebuttal: We appreciate the reviewers' insightful comments and suggestions. Your feedback has significantly contributed to enhancing the quality of our work. We would now like to address each reviewer's questions individually to ensure we provide comprehensive responses to your specific concerns. We include... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Three Towers: Flexible Contrastive Learning with Pretrained Image Models | Accept (poster) | Summary: CLIP is trained using contrastive loss. Previous works have shown that freezing the image tower, and only training the text encoder on pretraining datasets yields better performance compared to training both the encoders. This paper proposes that completely freezing the image encoder is drastic and will underp... | Rebuttal 1:
Rebuttal: Dear reviewer Y95u,
Thank you for your hard work and helpful feedback. We look forward to actively engaging with you during the discussion period to clarify any remaining points. Please read our comment above, addressed to all reviewers, first.
> Section 4.3 Scaling pretraining data [...]
Th... | Summary: This paper presents Three Towers (3T), a method to enhance the contrastive learning of vision-language models by incorporating pretrained image classifiers. Unlike the previously proposed Locked-Image Text Tuning (LiT) method, which directly uses frozen pretrained classifier embeddings, 3T introduces a third t... | Rebuttal 1:
Rebuttal: Dear reviewer Wqsw,
Thank you for your hard work and helpful feedback. We will very gladly incorporate some of your excellent suggestions and look forward to engaging with you during the discussion period to clarify any remaining points. Please read our comment above, addressed to all reviewers,... | Summary: In this paper, a technique called Three Towers (3T) is presented as a means to enhance the contrastive learning of vision-language models. The approach involves integrating pretrained image classifiers to achieve flexible and effective transfer in contrastive vision-language models. The authors provide a forma... | Rebuttal 1:
Rebuttal: Dear reviewer ni71,
Thank you for your hard work and helpful feedback. We look forward to actively engaging with you during the discussion period to clarify any remaining points. Please read our comment above, addressed to all reviewers, first.
> One weakness of the Three Towers (3T) method is... | Summary: This paper focuses on improving LiT. A concern with LiT is that it may be overly reliant on the pretrained model, completely missing out on any potential benefits the image tower might get from contrastive training. The proposed 3T method is formalized for flexible and effective transfer of pretrained classifiers... | Rebuttal 1:
Rebuttal: Dear reviewer GeUR,
Thank you for your hard work and helpful feedback. We will gladly incorporate some of your suggestions and look forward to engaging with you during the discussion period to clarify any remaining points. Please read our comment above, addressed to all reviewers, first.
> 1 [... | Rebuttal 1:
Rebuttal: We thank all reviewers for their careful consideration, insightful comments, and helpful suggestions. We are glad to see the paper was generally well received and hope that our responses alleviate the concerns that were raised. We believe your input has already helped improve the paper and look fo... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Asymptotically Optimal Quantile Pure Exploration for Infinite-Armed Bandits | Accept (poster) | Summary: This paper addresses a problem of pure exploration in the infinite-armed bandit setting : in this problem, each arm has mean reward $p_i$ drawn i.i.d. from a distribution with quantile function $G^{-1}$, and the aim is to output an arm $a_{i*}$ after sampling $N$ times an action, such that $\mathbb{P}\left(p_{... | Rebuttal 1:
Rebuttal: We thank the reviewer for their time and insightful reviews. Here is our detailed response.
Thank you for all of the helpful suggestions. We have taken them carefully into account while revising the paper.
Typos:
1. Thank you! We have corrected the typo.
2. We apologize. C is a universal cons... | Summary: This paper considers pure exploration problem in multi-armed bandits with infinite number of arms. Its goal is to return an arm with sufficiently high expected reward (within the top $\eta$ region) with high probability (at least $1-\delta$). A naive method could lead to a complexity upper bound of $O({\log(1/... | Rebuttal 1:
Rebuttal: We thank the reviewer for their time and insightful reviews. Here is our detailed response.
We would like to emphasize that we not only give an algorithm achieving failure probability $\exp(-N/\log^2 N)$, but we also show this is unimprovable in a natural asymptotic parameter regime. This lower b... | Summary: This paper studies pure exploration problem in bandits with infinite arms, where the mean reward of each arm is i.i.d. according to some unknown distribution. In fixed-confidence setting, this paper proposes an algorithm that nearly matches the optimal sample complexity. In fixed-budget setting, this paper pro... | Rebuttal 1:
Rebuttal: We thank the reviewer for their time and insightful reviews. Here is our detailed response.
Suggestions:
1. For this suggestion, we actually intended to write “known” because we are discussing the lower bound. This “known” makes pure exploration easier, hence the lower bound is stronger. We added... | Summary: **[Setting]** This paper studies best arm identification among countably infinite arms in both fixed confidence and fixed budget setting. The mean rewards for these arms are drawn iid from a distribution $\mu$. Let $\alpha$ be the $1 - \eta$ percentile of $\mu$. The goal is to find an arm with mean reward at l... | Rebuttal 1:
Rebuttal: We thank the reviewer for their time and insightful reviews. Here is our detailed response.
Quality Questions:
1. L212-215 are comments for the fixed budget setting (as opposed to the fixed confidence setting). Thank you for your suggestion! We have added some intuition in the paper following tho... | Rebuttal 1:
Rebuttal: We thank the reviewers for insightful comments and suggestions. We have adopted most of your suggestions and added clarifications in the paper.
Please see the one-page pdf where we highlight some of the more significant modifications so far.
Pdf: /pdf/77c21ce28704c8820bd02d4465ffd07151af1058.pdf | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
An Inverse Scaling Law for CLIP Training | Accept (poster) | Summary: This work explores how to increase the training speed of CLIP in order to make it possible to train on academic resources. They examine how to best compress both the visual and text information given a limited token budget. They do this across multiple CLIP sizes and find that as the size of CLIP increases, it... | Rebuttal 1:
Rebuttal: We first thank the reviewer for the detailed comments and the appreciation of our work. We address the concerns below:
### Q1: Relation to RECLIP
Thanks for bringing up this concurrent work! We totally agree that a discussion of RECLIP [1] will make this work more complete, and we will do so in ... | Summary: Different methods to reduce the number of tokens entering the image or text encoders when training a CLIP model are compared.
In doing this the authors discover that larger models are able to learn faster from fewer tokens per example than smaller models; this is the inverse scaling law in the title.
Models a... | Rebuttal 1:
Rebuttal: We first thank the reviewer for the detailed comments and the appreciation of our work. We address the concerns below:
### Q1: Figure 1 difficult to parse
Thanks for raising the concern! The number of tokens in the top row indeed indicates the number of input tokens. We will update Figure 1 to ma... | Summary: This paper is about increasing the training efficiency of vision-language models such as CLIP. These models can serve as a foundation for many vision applications and are therefore very useful. However, they are typically (pre-)trained on large datasets for many steps, which makes research involving training s... | Rebuttal 1:
Rebuttal: We first thank the reviewer for the detailed comments. We address the concerns below:
### Q1: Incomplete evaluation
Thanks for raising this concern!
Firstly, it is essential to note that our training pipeline contains two stages — the reduced-length input is only applied during pre-training to ... | Summary: This work studies the impact of token length (image and text) on the quality of image-language pretraining in CLIP-style models. Crucially, the authors find that as the model size (both image and language encoders) are scaled up, the drop in accuracy owing to a smaller input token length keeps on decreasing. T... | Rebuttal 1:
Rebuttal: We first thank the reviewer for the detailed comments. We address the concerns below:
### Q1: Figure 1 confusing & Discussion about patchifying
Sorry for the confusion about Figure 1. There is a typo in the figure; the dimension of the encoder output solely depends on the encoder size (i.e., a la... | Rebuttal 1:
Rebuttal: ### Global Response:
We would like to first thank all the reviewers for their invaluable comments and recognition of the interest and value in our discovery of the inverse scaling law in CLIP training (Reviewer Depv, TeGF, Mhr1), enabling the potential to improve accessibility to large models and... | NeurIPS_2023_submissions_huggingface | 2,023 | Summary: The key finding in this paper is that, the larger the model is, the shorter the token sequence is required to maintain a certain level of performance. Based on this finding, this paper proposes to train larger model with shorter sequences, which is more computationally favorable than training smaller models wi... | Rebuttal 1:
Rebuttal: We first thank the reviewer for the detailed comments and the appreciation of our work. We address the concerns below
### Q1: Missed related works of scaling law in language models
Thanks for this suggestion! We completely agree that discussing these related works will make this paper more comple... | null | null | null | null | null | null |
Disentanglement via Latent Quantization | Accept (poster) | Summary: The paper proposes learning a disentangled latent representation by quantizing the latent space into learnable discrete values, with a separate scalar codebook for each dimension. The main assumption that the paper relies on is that "many datasets of interest are generated from their sources in a compositional... | Rebuttal 1:
Rebuttal: Thank you for your review. We believe we can address the main concerns you had about our submission, in particular the similarity to VQ-VAE and only evaluating with InfoMEC.
> The proposed method appears to be very similar to VQ-VAE.
Indeed, latent quantization can be summarized as three modific... | Summary: The paper presents an alternative approach to the standard disentanglement representations learning framework whereby the continuous latent space is substituted by a quantized one. Based on this new quantized latent space, the authors propose novel ways to measuring several properties related to disentanglemen... | Rebuttal 1:
Rebuttal: Thank you for your extensive review. We agree with the main criticisms you have about our manuscript at the time of submission, and we have taken action to address them. We appreciate your willingness to re-assess our work.
> While the authors present numerical results for five datasets, they onl... | Summary: The paper discusses the problem of disentangled representation learning (finding latent variables only from observed data that ideally correspond to the true underlying sources of variation in the data). The authors suggest both a method for disentangled representation learning and a metric for evaluating it. ... | Rebuttal 1:
Rebuttal: Thank you for your review and appreciation of the strengths of our work. We have striven to address your main concerns with our submission.
> The proposed method is only evaluated using InfoMEC. This is a significant drawback in my opinion. [...] Another limitation, which I'm not sure is discusse... | Summary: The authors propose to achieve disentanglement via latent quantization which constructs an inductive bias that matches the compositional nature of data. The authors also proposed three metrics to evaluate the disentanglement and achieved better performance according to the metrics they proposed. Ablation studi... | Rebuttal 1:
Rebuttal: Thank you for your positive evaluation of our work! We have taken your constructive criticism very seriously, and we believe we have addressed the most pressing points.
> The authors propose a few new metrics, and the metrics are sound. The authors claim that the metrics addressed a few issues of... | Rebuttal 1:
Rebuttal: We thank the reviewers for their detailed and insightful feedback. Our manuscript has greatly benefited as a result of addressing the reviewers’ most critical comments. A brief summary of the major changes:
- **Quantitative results based on DCI evaluation [all reviewers].** We have repeated our ev... | NeurIPS_2023_submissions_huggingface | 2,023 | Summary: This paper tackles the topic of learning disentangled representations from observations in a fully unsupervised manner. In particular, the authors propose a novel inductive bias for disentangled representation learning as well as an associated evaluation metric; The inductive bias relies on latent quantization... | Rebuttal 1:
Rebuttal: Thank you for your thorough evaluation of our submission and thought-provoking questions.
> Have authors tried to incorporate the proposed inductive bias into a Variational form of an AE?
Our proposed quantized latent autoencoder (QLAE) is variational to the same extent that VQ-VAE is, specifica... | null | null | null | null | null | null |
Spatially-Guided Temporal Attention (SGuTA) and Shifted-Cube Attention (SCubA) for Video Frame Interpolation | Reject | Summary: This work introduces two attention modules for video frame interpolation that achieve promising results in terms of quantitative and qualitative evaluation. The researchers conducted numerous experiments to provide a comprehensive evaluation of the proposed approach.
Strengths: The proposed attention mechanis... | Rebuttal 1:
Rebuttal: We sincerely appreciate your valuable insights and thoughtful evaluation of our work. Here we present our explanation to the specific points raised during the review process:
1. The reported results in Tab. 1/2 derived through pretraining the previous state-of-the-art models on the [Vimeo90k Se... | Summary: In video tasks, the computational complexity and memory requirements of Transformer is challenging. This paper employ two different Transformers in VFI task resulting in new state-of-the-art results. Furthermore, this paper conduct an analysis of existing embedding strategies, and put forth a novel half-overla... | Rebuttal 1:
Rebuttal: We are sincerely grateful for your time and thoughtful evaluation of our work. Here we present our explanations to the specific points raised during the review process:
1. First of all, We actually compared our approach with recent relevant works, including CVPR20-AdaCoF, AAAI20-CAIN, WACV23-FLAV... | Summary: This paper aims to address the computational complexity and memory requirements of Transformers to enhance their suitability for the video frame interpolation (VFI) task. The authors propose two novel methods, SCuTA and SGubA, which are integrated into a multi-stage multi-scale framework. SCuTA leverages the c... | Rebuttal 1:
Rebuttal:
1. Thank you for your reminder. We will add explanations for G-MSA and SC-MSA in the revised manuscript.
2. In terms of our model's inference speed we compared our proposed method against the top two existing multi-frame interpolation methods, FLAVR and VFIT. We initially followed the AMT appr... | Summary: This paper tackles video frame interpolation (VFI). It particularly aims to deploy a transformer architecture for VFI tasks. To address the computational complexity and memory requirements in transformers, it proposes two transformer networks, SGuTa and SCubA. While SGuTa uses the spatial (global) information ... | Rebuttal 1:
Rebuttal:
Thank you for your thoughtful and constructive feedback. Here, we present our explanations addressing the specific points raised during the review process:
1. **Concerning the motivation and novelty of the work:**
- While it may appear that we did not explicitly reference the progress made i... | null | NeurIPS_2023_submissions_huggingface | 2,023 | Summary: This paper proposed two types of Transformer for video frame interpolation: Spatially-Guided Temporal Attention (SGuTA) and
Shifted-Cube Attention (SCubA). SGuTa merges the temporal dimension and the embedding dimension during the self-attention process to explore the inherent correlations between the spatial ... | Rebuttal 1:
Rebuttal: Thank you for your thoughtful and constructive feedback.
1. As you mentioned, SGuTA's performance lags behind SCubA in various aspects. We still chose to include SGuTA in our work for a specific reason.
- On one hand, SGuTA demonstrates better performance on various datasets compared to prev... | null | null | null | null | null | null |
Variational Weighting for Kernel Density Ratios | Accept (poster) | Summary: The paper presents an optimized weight function for kernel density estimations, which reduces bias and improves estimates of prediction posteriors and information-theoretic measures. The weight function is derived using tools from multidimensional calculus of variations and utilizes information from the first ... | Rebuttal 1:
Rebuttal:
Thank you for taking the time to review our paper. We deeply value your feedback.
> l. 23-27: Mention the cause of underestimation and overestimation in convex and concave regions
First, we will include Larry Wasserman (2005) *All of nonparametric statistics*, Springer, in our revised version ... | Summary: The paper studies the weighted kernel density estimation (wKDE) and proposes a method to learn the weights.
Strengths: 1. The authors illustrate biases in the standard KDEs and propose a mitigation by introducing weights in the KDE.
2. An estimator for the weight function is provided based on minimizing a qua... | Rebuttal 1:
Rebuttal:
Thank you for taking the time to review our paper. Below, we rebut the claims that the reviewer provided. We hope the concerns of the reviewer to be resolved.
> The bias occurs only when the data x is not sampled from the true distribution p(x).
This is not true. KDE should produce asymptotic... | Summary: The authors propose an optimal weighting scheme to reduce bias in kernel density ratio estimation that alleviates the problem of under (or over) estimation due to the geometry of the distribution. They show that employing this weighting scheme leads to better estimates of prediction posteriors.
Strengths: - ... | Rebuttal 1:
Rebuttal: Thank you for your constructive questions.
> The proposed method is computationally expensive. Fair comparison
The following is a table for the computation time with the 10D non-isotropic experiment in Fig. 4 for 4.2k samples per density:
| Algorithm | Time (s) |
|------------|----------|
| KD... | Summary: The paper aims at improving the KDE estimator for the estimation of density ratios or KL divergences. To do so, they introduce a weighting function $\alpha$ whose role is to reduce the bias of the estimates. A variational problem is obtained using standard derivations, allowing to learn $\alpha$ in some model ... | Rebuttal 1:
Rebuttal:
We acknowledge the relevance of all the comments and have outlined our responses below, explaining how we intend to revise the manuscript in accordance with the feedback.
> Major comment on the structure and the illustration of the main equations
We will highlight the main equations and make su... | null | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Dynamic Tensor Decomposition via Neural Diffusion-Reaction Processes | Accept (spotlight) | Summary: The paper presents the framework to generate dynamic embeddings for multi-entity (e.g., traffic type, severity, latitude, and longitude) interactions with timestamps. The temporal high-order interactions are modeled via neural diffusion-reaction processes, and to tack the temporal drift of underlying node, aut... | Rebuttal 1:
Rebuttal: Thanks for your valuable comments and suggestions. Here are our responses. C: comments; R: responses
> C1: "It seems that the text can be further improved to present better. Some phrases are repeated several times unnecessarily, and some points, on the contrary, are covered too briefly. I would r... | Summary: In this work, the authors introduce DEMOTE which can both consider both structure information and temporal information via dynamic tensor decomposition. Structure information is captured by multi-partite graph diffusion process which in substance is a special ODE, whereas temporal information is analyzed using... | Rebuttal 1:
Rebuttal: Thanks for your insightful comments and the great reference TANGO. We will add discussion and comparison with this nice, related work.
>C1: why can learning dynamics produce more reliable and long-term trajectory predictions?
R1: Great question. To better illustrate it, we can consider an alte... | Summary: In this paper, to capture higher-order interaction among different modes of tensor data and their evolution process over time, the authors proposed an ODE-based method called DEMOTE to capture the dynamics of latent embeddings. In detail, the authors propose to use a diffusion-reaction process to model the e... | Rebuttal 1:
Rebuttal: Thanks for your valuable comments. Here are our responses. C: comments; R: responses
> C1: "The main concern lies in its novelty since there are some efforts studying the interaction results of higher order continuous time data, as clarified in the paper (that is [1]) ... Although these two paper... | Summary: This paper proposes a dynamic tensor decomposition (TD) model, in which the latent factors are functions of time stamps described by neural ODEs. Moreover, they encode an adjacent matrix into the diffusion process to better model the correlations among different tensor modes. Finally, they use an additional ne... | Rebuttal 1:
Rebuttal: We thank the reviewer for their valuable comments. Here are our responses. C: comments; R: response.
>C1: "the authors construct a complex TD with ODEs and NN to parameterize the latent relation. Besides, the model parameters are learnt from MAP estimates (Eq. 3). I am wondering if such a model s... | Rebuttal 1:
Rebuttal: We thank all the reviewers for their careful, insightful and constructive comments! Here we post a few global responses. C: comments; R: responses
>C1: provide the code
R1: Yes, we are happy to provide the code. We have created an anonymous repository that includes the implementation of our mode... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Handling Data Heterogeneity via Architectural Design for Federated Visual Recognition | Accept (poster) | Summary: This paper conducted extensive experiments to show the effect of the neural architecture design on non-iid federated learning (FL) for visual recognition. More specifically, experimental results demonstrated that applying the SOTA neural architecture design to FL can significantly mitigate the performance degr... | Rebuttal 1:
Rebuttal: We thank the reviewer for the encouraging and insightful comments. **Please find our responses to specific queries below.**
>**1. On technical novelty**
Rather than introducing an entirely novel approach to federated learning (FL), our paper focuses on conducting thorough experiments using sta... | Summary: This paper presents an experimental analysis comparing several state-of-the-art neural networks architectures on four different federated vision datasets in highly heterogeneous settings using multiple different optimization approaches.
Strengths: Thorough experimental study: Four datasets, four optimizers, ... | Rebuttal 1:
Rebuttal: We thank the reviewer for the encouraging and insightful comments. **Please find our responses to specific queries below.**
> **1.1 *(Weakness)* Various existing works have already emphasized the problem BatchNorm causes in FL settings, and proposed to use GroupNorm instead of BatchNorm, for ex... | Summary: This paper focuses on exploring the effects of architectural design choices on the performance of federated learning (FL) systems in visual recognition tasks, particularly when faced with data heterogeneity. To accomplish this, the paper conducts a comprehensive experimental analysis involving 19 state-of-the-... | Rebuttal 1:
Rebuttal: We thank the reviewer for the encouraging and insightful comments. **Please find our responses to specific queries below.**
>**1. While this paper provides a comprehensive study of architectural design for FL, no novel methods are developed.**
We acknowledge that our work does not propose ne... | Summary: In this comprehensive research study, the authors dive into the realm of federated learning (FL) and its application in visual recognition. They address a crucial aspect often overlooked in FL literature - the impact of architectural design choices on achieving optimal performance. While FL offers collaborativ... | Rebuttal 1:
Rebuttal: We thank the reviewer for the encouraging and insightful comments. **Please find our responses to specific queries below.**
>**1. Smaller figure captions?....**
We thank the reviewer for the feedback. We will address this in the final draft.
>**2. Reposition Figure 2 adjacent to Table 2...... | Rebuttal 1:
Rebuttal: We thank the reviewers for the positive comments. In this rebuttal, we have exerted our best efforts to address nearly all of the reviewers' comments, resulting in numerous enhancements to our final draft. *These improvements include*: (i) additional experiments on using Group normalization, (ii) ... | NeurIPS_2023_submissions_huggingface | 2,023 | Summary: This work conducted extensive empirical evaluation to study the influence of model structure on federated learning performance. The paper evaluated the performance of 19 models from 5 model families on 4 different federated learning datasets. The results showed that architecture design has a significant impact... | Rebuttal 1:
Rebuttal: We thank the reviewer for the encouraging and insightful comments. **Please find our responses to specific queries below.**
> **1. I appreciate that the author conducted such an extensive empirical evaluation. It takes a lot of effort. ,..., I suggest the authors conclude more insights and guidel... | null | null | null | null | null | null |
Certified Robustness via Dynamic Margin Maximization and Improved Lipschitz Regularization | Accept (poster) | Summary: This paper proposes a new bound for margin maximization, as well as tighter estimation of Lipschitz constant, to improve the adversarial robustness of neural networks. Experiments on three benchmark datasets are presented to show the proposed method can improve the certified radius.
Strengths: - This paper i... | Rebuttal 1:
Rebuttal: We would like to thank the reviewer for the comments.
>Regarding the improved bound for input margin, I think the contribution is limited since the similar strategy has be used frequently in recently works.
> The novelty of margin regularization is limited and the experimental results are not s... | Summary: The authors explore how a model’s robustness to local pertubations in the input space can be enhanced through regularization at a low cost on the clean accuracy.
The goal is to have large margins on the desicion boundaries in input space.
Since adding a regularization directly on the size of these margins is ... | Rebuttal 1:
Rebuttal: We thank the reviewer for their perceptive remarks and suggestions. We will add your useful experiment and ablation study suggestions to the final manuscript. We'll try to address your concerns in the following:
### Addressing weaknesses
- In the limited rebuttal period we were able to incorporat... | Summary: This paper developed a robust training algorithm whose objective is to increase the margin while regularizing the Lipschitz constant of the model along vulnerable directions. The authors also developed an improved Lipschitz estimation for a multi-layer networks with residual connections. Experiments on MNIST a... | Rebuttal 1:
Rebuttal: We thank the reviewer for their thoughtful suggestions and constructive criticism. In the following, we respond to the weaknesses raised by the reviewer.
### Addressing weaknesses
- Our hope was to present a unifying loss function in Eq. (6) that can recover several existing loss functions. For e... | Summary:
This paper proposes a loss to induce robustness w.r.t. adversarial attacks. This loss is composed of a term for enlarging the logit difference (such as classical Cross-Entropy Loss), and a surrogate term to maximize the robust radius. The latter imply to compute an upper bound of the Lipschitz constant (cal... | Rebuttal 1:
Rebuttal: We would like to thank the reviewer for their constructive criticism and detailed comments. We have provided a global response to address the comments that are common to all the reviewers. Here we address the comments specific to the reviewer.
> Batchnorm layers and dropout are not supported.
A... | Rebuttal 1:
Rebuttal: We would like to sincerely thank all the reviewers for their useful comments and feedback. We have improved the paper in terms of presentation, related work, notation, and technical aspects. Below, we report a list of global responses that would be useful for all the reviewers.
**A new regulariz... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Hardness of Low Rank Approximation of Entrywise Transformed Matrix Products | Accept (poster) | Summary: This paper studies the problem of computing a low-rank approximation (LRA) for $f(UV)$, where $U\in \mathbb{R}^{n\times r}$, $V\in \mathbb{R}^{r\times n}$, and $f:\mathbb{R}\rightarrow \mathbb{R}$ applies to each entry of $UV$. This problem has very important applications in deep learning, in particular for Tr... | Rebuttal 1:
Rebuttal: Thank you for your positive review and careful reading of our paper.
We address your questions below:
1. $B_i$ is the $i$-th row of matrix $B$.
2. Yes, the lower bounds do hold for randomized algorithms (and OVP is stated with constant probability). In general, by Yao’s minimax principle, deter... | Summary: The authors discuss a setting where one observes a scalar transformation of an r-rank matrix, f(UV) and wants to find the best k-rank approximation of it.
It is known that under various assumptions, if f(UU^t) is PSD then the problem is solvable in nearly linear time. It is also known that for various kernels... | Rebuttal 1:
Rebuttal: Thank you for your positive comments.
We address your comment regarding the “importance of choosing the $x^p$ activation function”: Note that we extend our base case results over $x^p$ to a broader class of functions (via a Taylor expansion) in Theorem 3.2. Specifically, our results also hold fo... | Summary: This paper studies the problem of low rank approximation of the form f(U.V) where the matrices U,V $\in \mathbb{R}^{n \times r}$ and $r=O(\log n)$ and $f(.)$ is applied entrywise. The main contributions of the paper are the following:
1) The paper presents a $\Omega(n^{2-o(1)})$ time lower bound for getting r... | Rebuttal 1:
Rebuttal: Thank you for your careful review of our paper - the following are our responses to the main concerns raised:
Addressing Weaknesses:
1. We emphasize that we presented our results in a way that highlights the novelty and difficulty of our lower bounds, which could be of general interest even in a... | Summary: For two matrices U,V one is often interested in computing f(UV) for some entrywise function f. Several papers in the past decade considered the problem of approximating f(UV) by a product AB with small inner-dimension (low-rank approximation). While a number of past papers presented clever positive results, th... | Rebuttal 1:
Rebuttal: Thank you for your positive feedback. We would like to address your main comment concerning “irrelevant complexity results to most attendees” and “[NeurIPS] is mostly interested in doing things rather than in enumerating all things that cannot be done”:
* First, while our submission mainly focus... | null | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Normalization-Equivariant Neural Networks with Application to Image Denoising | Accept (poster) | Summary: The paper proposes normalization-equivariant CNNs, in which the convolutional kernels are forced to sum to 1 and the activation sorts pairs of neurons. This is shown to perform similarly to other models on image denoising when trained with the same noise levels, but generalizes much better to unseen noise leve... | Rebuttal 1:
Rebuttal: W2. They are “blind” indeed since the models were trained on a single noise level (and the noise level is not known at inference time). For the sake of clarity, we will explicitly mention it in the legend of Figures 4 and 5.
W3. Thank you for pointing this out. The transition from a dense layer (... | Summary: The authors of this paper propose an architectural change in neural networks, in a way that enforces the normalization equivariance property. The authors demonstrate how networks that satisfy this property can be used in the setting of image denoising.
Strengths: - The paper is clearly written overall. The in... | Rebuttal 1:
Rebuttal: Regarding your reservations about paper limitations, please refer to the global response to reviewers.
- “the authors compare against techniques which use regular feedforward neural networks. While the comparisons support the points made by the authors, I believe that it would be interesting to c... | Summary: In this paper, the authors propose an architectural design which guarantees shift and scale equivariance. The proposed method is called normalization equivariance in which scale equivariance is inherited from prior work by removing layers biases and shift equivariance is achieved through affine-constrained con... | Rebuttal 1:
Rebuttal: Minor issues:
- “I believe rotation needs to generally be an invariance (as opposed to equivariance) with respect to labels in image classification (line 52)”.
Thank you for pointing this out. We agree with you, this was an inaccuracy and we removed it from the manuscript.
- “Proof for proposi... | Summary: This work proposes a normalization equivariant deep convolutional neural network for image denoising. The authors provide a mathematical construction of a normalization equivariant neural network and propose sorting nonlinearity. The proposed nonlinearity shows comparable performance with the baseline and show... | Rebuttal 1:
Rebuttal: W1. First of all, we would like to stress that our aim is not to introduce new layers that might perform better than traditional ones, but only to observe that it is possible to use the proposed substitutes as a viable alternative that allows normalization-equivariance, without degrading performan... | Rebuttal 1:
Rebuttal: First of all, we want to thank the reviewers for dedicating their valuable time to the evaluation of our paper and for providing us with constructive feedback, which will improve the quality and clarity of the manuscript. We have revised the manuscript, taking into account all the comments receive... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Cascading Bandits: Optimizing Recommendation Frequency in Delayed Feedback Environments | Accept (poster) | Summary: This paper introduces several modifications to the cascading bandit setting to make it more suitable for settings where users can provide delayed feedback and abandon the system in response to receiving messages too frequently, such as online ad retargeting. It then provides an algorithm for solving the messag... | Rebuttal 1:
Rebuttal: We appreciate your commenting that our writing is generally good. Please find our point-by-point responses below.
> Should Algorithms 2 and 3 call Algorithm 1 with $q^{LCB}$ instead of $q^{UCB}$? The optimism in the face of uncertainty heuristic is based on always using upper bounds on how good s... | Summary: This work studied a generalization setting of the cascading bandit model which has been explored in existing literature. In detail, beyond the existing cascade model, it allows delayed feedback from the user, different number of pulled arms at different time steps, and takes the action frequency into considera... | Rebuttal 1:
Rebuttal: We would like to sincerely thank you for your effort in reviewing our paper.
> The regret upper bound for Algorithm 2 is provided in Theorem 3.3, I think a detailed comparison to existing bounds should be provided here:
I guess Algorithm 2 also works for the traditional cascading bandit setting, ... | Summary: This work introduces a new variant of cascading bandit that differs with the classical cascading bandit in the following ways: (1). the user might abandon the system with certain probability depending on the list length; (2). the list length is some variable to be optimized (even could be contextualized), inst... | Rebuttal 1:
Rebuttal: Thank you very much for your great comments! Please see our point-by-point response below.
>One thing I am concerned about is the computational feasibility of the method. It seems for each user, the exact DP is performed, which might be super expensive when the length list $m$ is large. Could the... | Summary: The paper considers a bandit setting where the agent selects a list of messages to be sent one at a time in a time window of length $D$. For each message, the user may be attracted (and sent a reward) or not (in which case she may leave the system with a probability that depends on the frequency of the message... | Rebuttal 1:
Rebuttal: Thank you very much for commenting that our paper is technically solid. We answer your questions below:
>Apart from this weakness, I have a comment: The literature on click models includes a click model with probability of leaving the system: the dependent click model.
Thanks for your comments ... | Rebuttal 1:
Rebuttal: We thank all reviewers for their time and insightful comments to improve our paper. We appreciate the positive feedback. We provide clarifications below.
Pdf: /pdf/2ef77bc33eb3e15afb1cddb2f50ea79a53182c8c.pdf | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
SNAP: Self-Supervised Neural Maps for Visual Positioning and Semantic Understanding | Accept (poster) | Summary: Currently，2D semantic maps have limitations: they lack detail, often contain inaccuracies, and are difficult to create and maintain， in order to solve these problem, the author proposes fusing features from different input modalities to obtain a 2D neural map that is robust over time and retains sufficient ... | Rebuttal 1:
Rebuttal: **[...] there is no relevant experimental evidence to prove whether the fusion of multi-mode features and the proposed supervision method are reasonable.
It is necessary to supplement experimental evidence for the effectiveness of fuse and contrastive learning supervision in the generation process... | Summary: This paper proposes a self-supervised learning approach to create a neural representation of the 3D environment from aerial-view and ground-view images. It utilizes contrastive learning to align neural maps based on the same pose and distinguish maps with inconsistent poses. Large-scale pretraining enhances th... | Rebuttal 1:
Rebuttal: **Exploring the impact of pose accuracy [of the training data] on the model's performance would provide valuable insights.**
We did not try polluting the poses while training. While we can try this, it was not possible to regenerate the data and retrain the model within the rebuttal window.
**It... | Summary: This paper presents a method for self-supervised learning of BEV features, by using a contrastive loss that links features computed from perspective-view images to features computed from aerial images, where BEV correspondences are given by poses. Interesting method details include depth-weighted splatting in ... | Rebuttal 1:
Rebuttal: **I am particularly interested to see if the pre-trained model outperforms some SOTA on some BEV segmentation task (e.g., on nuScenes). The danger is something like: as you add more supervised data, the impact of the pre-training goes to zero.**
Obtaining training labels for 3D semantic tasks is ... | Summary: This work develops neural 2D maps in an self-supervised manner by matching multi-view ground level imagery with aerial images.
To construct a neural map (2D map of features), they encode an image into a feature frustum (similar to lift-splat-shoot), then pool vertically - for multi-view images, they perform a ... | Rebuttal 1:
Rebuttal: **Quantitative evaluation is not as fair as I would like. Especially when comparing streetview inputs only, the performance improvement is not particularly strong.**
We note that with StreetView inputs only, our model still outperforms SfM+SuperGlue (a strong and popular baseline) on the medium (... | Rebuttal 1:
Rebuttal: We thank the reviewers for their time and well-thought comments. Reviewers noted very clearly that our method is **novel, well-motivated, and exciting**:
* _"the large-scale nature of training, self-supervised learning, and use of top-down + street-view images is exciting"_ -- R.sYmT
* _"the appro... | NeurIPS_2023_submissions_huggingface | 2,023 | Summary: This paper proposes SNAP, a novel architecture to obtain 2D neural maps from street-view and top-down images. The model works by encoding street-view and top-down images separately using corresponding encoders, projecting features to the top-down 2D space, and fusing them via max-pooling to obtain the final fe... | Rebuttal 1:
Rebuttal: **Will the "50M street view images" dataset be shared publicly for reproducibility?**
As stated in the global response, we would really like to, but we cannot, because we do not own the data. Additionally, please note that a large-scale dataset of ground-level images must comply with regulations ... | null | null | null | null | null | null |
Trial matching: capturing variability with data-constrained spiking neural networks | Accept (poster) | Summary: This paper considers fitting high-fidelity recurrent spiking neural networks to multiple populations of neurons across sessions. The networks include LIF neurons for each fitted neuron and model within & across region interactions at fine timescales, and include a soft-thresholding mechanism for neural spiking... | Rebuttal 1:
Rebuttal: We would like to thank the reviewer for reading our paper and providing useful feedback.
Given the second question of the reviewer, we wonder whether there might be a misunderstanding on the nature of the trial-matching loss function. It appears that the reviewer understands that we only model "... | Summary: The authors present a method to fit recurrent spiking neural network models to experimental data, extending prior work to multi-session recordings that involve different subpopulations in each session. To capture single trial variability, they introduce a trial matching loss function. The technique is applied ... | Rebuttal 1:
Rebuttal: We would like to thank the reviewer for evaluating our work. We reply to each weakness and question below.
Weaknesses:
- We are not sure that we understand correctly what the reviewer meant by "individual trial averages". In case the reviewer meant neuron specific individual trials, we tried t... | Summary: This paper presents a framework for training a large recurrent spiking neural network on multi-session recordings by leveraging
an optimal transport-based trial matching between the real data and generated data. This model is use to model the cortical sensory-motor pathway during a tactile detection task.
S... | Rebuttal 1:
Rebuttal: We would like to thank the reviewer for the detailed review, and the positive feedback. Regarding the weaknesses of the paper raised by the reviewer:
1. About the comparison to other baselines, in case this was overlooked by the reviewer, we would first like to highlight that we had compared tria... | Summary: This paper extends previous works on generative model of neural data using a recurrent spiking neural network.
The main contributions are design of new loss function based on optimal transport to match the trial-to-trial variability in the data, and the approach to extend RSNN generative model to leverage reco... | Rebuttal 1:
Rebuttal: We would like to thank the reviewer for reading our work in depth. We address below the main weakness and questions raised by the reviewer.
- *"Although the authors mention in the discussion section the difference in the approach between the proposed method and existing methods, the (potential)... | Rebuttal 1:
Rebuttal: We would like to thank all reviewers for reading our paper in depth and providing insightful feedback. All reviewers understood the paper and wrote that sections were "clear and novel" and "appropriately discussed and clearly stated". Three of the reviewers commented that the submitted manuscript ... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Explaining the Uncertain: Stochastic Shapley Values for Gaussian Process Models | Accept (spotlight) | Summary: The work proposes a new kind of Shapley value (SV) based explanation method. The new methods GP-SHAP and BayesGP-SHAP are like TreeSHAP, LinearSHAP or RKHS-SHAP model-specific SV methods for Gaussian Process (GP) models. The authors further propose another approach of utilizing GP models to amortize (i.e learn... | Rebuttal 1:
Rebuttal: We thank the reviewer for their constructive feedbacks and we will answer their questions below:
## 1. How does your approach compare to FastSHAP in terms of predictive performance? What's the tradeoff in terms of predictive performance and explainability?
- To answer this, we have ran a predicti... | Summary: The authors consider Shapley values computed from games $\nu$ modelled with a Gaussian process. Since the game is stochastic, the Shapley values are now stochastic, and the stochasticity reflects epistemic uncertainty in the explanations. This is unlike previous methods like BayesSHAP, where the uncertainty is... | Rebuttal 1:
Rebuttal: We thank the reviewer for their in-depth suggestions. We would like to address their concerns below:
### Proof of Theorem 4:
We agree with the reviewer that using the d-game to establish stochastic Shapley values as a push forward measure is easier to understand. In fact, this approach was the on... | Summary: This works introduces the stochastic Shapley values to the explainable AI community. It formulates the stochastic Shapley values for GP models with a closed-form expression for the covariance matrix. Then, it proposes BayesGP-SHAP, which incorporates both sources (of uncertainty from the GP posterior and of th... | Rebuttal 1:
Rebuttal: We thank the reviewer for their positive feedback and pointing out the typos in the references, we will amend them accordingly in the camera ready version. | Summary: The paper introduces a novel approach for explaining Gaussian processes (GPs) by utilizing the analytical covariance structure present in GPs based on extending the concept of Shapley values to stochastic cooperative games, resulting in explanations that are random variables. The GP explanations generated usin... | Rebuttal 1:
Rebuttal: We thank the reviewer for sharing literature pointing out the limitations of Shapley-value based explanations. We have included a discussion and included the references in the discussion section in the camera ready version. We will also enlarge figure 2 in the camera ready version. | Rebuttal 1:
Rebuttal: We include further experiments to compare FastSHAP with GP regression using the Shapley prior in light of reviewer uH3p's comments.
Pdf: /pdf/9ebcb9fcd9c00071904b748fab7d8d8c7d3154b2.pdf | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
GradOrth: A Simple yet Efficient Out-of-Distribution Detection with Orthogonal Projection of Gradients | Accept (poster) | Summary: The paper proposes to project the per-example parameter-gradients of the prediction head onto the orthogonal subspace of the learned features and uses the projection norm as the OODness score. The key motivation for doing so is that the parameter gradients lie in the subspace of the corresponding input feature... | Rebuttal 1:
Rebuttal:
Dear Reviewer MCXS,
Thanks for your thorough evaluation of our paper. Your suggestions will significantly enhance the comprehensiveness and clarity of our work. We would like to address your comments and queries as follows:
- Experiments on near-OOD and semantic outliers:
Thank you for your va... | Summary: The work proposes GradOrth, a simple post-hoc OOD detection method that seeks to improve on existing gradient-based OOD detection approaches by leveraging a selected lower-rank subspace of the gradient most relevant for OOD detection. The resulting GradOrth method shows superior OOD detection performance on a ... | Rebuttal 1:
Rebuttal:
Dear Reviewer 6mNB,
We appreciate your thoughtful evaluation of our work. Your feedback provides valuable insights that will undoubtedly contribute to the refinement and enhancement of our research. We would like to address your comments and concerns as follows:
We have compared our method aga... | Summary: The paper proposes an OOD detection method by exploiting the orthogonal projections of gradients to a subspace. Once a model is trained on an in-distribution (ID) dataset, the k most significant features in the last layer of a network are found using SVD using a small subset of n ID samples. Then, the gradient... | Rebuttal 1:
Rebuttal:
Dear Reviewer 81LC,
We greatly appreciate your insightful review of our paper. Your comprehensive assessment of our methodology and experimental results provides valuable feedback that will undoubtedly contribute to the refinement and advancement of our work. Please find the responses to your qu... | Summary: This paper focuses on OOD detection based on gradient information. It seems based on the framework of GradNorm (the paper is confusing on this) to use the gradient response as the score for detecting OOD data. Specifically, it uses SVD based subspace modeling and projection of the testing sample’s gradient as ... | Rebuttal 1:
Rebuttal:
Dear Reviewer fiAZ,
We appreciate your time and effort in reviewing our paper. Your feedback provides us with valuable insights into the perception of our work, its strengths, and aspects that require improvement. In the following, we provide the response to your questions:
- How is gradient ca... | Rebuttal 1:
Rebuttal:
Dear Reviewers,
We would like to express our sincere gratitude for your diligent and insightful reviews of our paper. Your thoughtful evaluation and constructive feedback are invaluable to us as we strive to enhance the quality and impact of our work.
We definitely consider your insightful sugg... | NeurIPS_2023_submissions_huggingface | 2,023 | Summary: A method is proposed to perform out-of-distribution detection on a pre-trained classifier. The approach used by the authors is to utilize a low-rank approximation of the parameter space for in-distribution data. Gradients for in-distribution samples are more likely to be projected directly onto the parameter s... | Rebuttal 1:
Rebuttal:
Dear Reviewer ne6G,
Thank you for taking the time to evaluate our proposed method. We appreciate your feedback and insights. Your assessment provides us with valuable insights into the strengths and areas for improvement of our approach.
we want to assure you that we address your questions and ... | null | null | null | null | null | null |
Latent Field Discovery in Interacting Dynamical Systems with Neural Fields | Accept (poster) | Summary: In this paper, the authors address the often-overlooked influence of underlying field effects on the evolution of interacting systems, postulating the existence of latent force fields. They suggest utilizing neural fields for learning these dynamics. Equivariant networks, although commonly used, fail to captur... | Rebuttal 1:
Rebuttal: We would like to thank the reviewer for their insightful comments. We appreciate that they find our method novel and insightful with broad applicability. Please find our answers below.
```
I found it interesting that there's no explicit mention of the training objective. This omission can potenti... | Summary: In this paper, the authors have introduced a novel method, Aether, for discovering the latent field within dynamical systems in order to better model their dynamical evolutions. This is achieved by disentangle the interactions between objects and the effect of the external global field, conditioned on the stat... | Rebuttal 1:
Rebuttal: We would like to thank the reviewer for their insightful comments. We appreciate that they find our problem statement original and our method well-designed. Please find our answers below.
```
Q1. Most of the datasets used in the paper are artificial/simulation data. Are there any practical applic... | Summary: The main assumption of this paper is that local object interactions generally happen within external global fields. Based on this assumption, the paper then proposes to disentangle the local from the global interactions, respectively, by learning both an equivariant graph network, capturing the local interacti... | Rebuttal 1:
Rebuttal: We would like to thank the reviewer for acknowledging the novelty of our idea and approach. Please find our answers below.
```
Is it possible to generalize the ideas in this paper to a structuring mechanism relevant to a more general class of ML applications?
```
The notion of the field is tightly... | Summary: This paper proposes to learn the latent fields governing the dynamics of interacting objects in a system without observations of the field. The authors propose unconditional and conditional latent field models to learn the static and dynamic fields respectively. The learned latent field signals are represented... | Rebuttal 1:
Rebuttal: We would like to thank the reviewer for their insightful comments. We appreciate that they find the idea of field discovery very interesting, and our method simple and effective. Please find our answers below.
```
Unclear Notation
```
- We describe most shapes in detail in appendix A. We will inc... | Rebuttal 1:
Rebuttal: ### Reponse to question by Reviewer hUPP
```
Figure 3 shows the conditional neural field setup for dynamic fields, while most of the experiments are using static fields. To the reviewer, this is a bit misleading and does not showcase the proposed model properly. Please add both static and dynamic ... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Efficient Batched Algorithm for Contextual Linear Bandits with Large Action Space via Soft Elimination | Accept (poster) | Summary: This paper proposes the first efficient batch algorithm for (contextual) linear bandits. Instead of relying on the classical elimination-based algorithmic design, this paper adopts a linear optimization oracle over the large action set. The proposed algorithm achieves $\tilde{O}(\sqrt{}T)$ regret with the numb... | Rebuttal 1:
Rebuttal: We thank the reviewer for their input and for appreciating our contribution.
* **[Numerical simulations]** Indeed, as the reviewer notes, in this work we focused on the theoretical understanding of efficient batch learning in contextual bandits, where we show a batched bandit algorithm that prova... | Summary: This paper proposes an algorithm for the batched (contextual) linear bandit problem with a large action space. The proposed algorithm has low time and space complexity. The proposed algorithm for the batched linear bandit problem selects arms in each batch according to a G-optimal design based on a modified ac... | Rebuttal 1:
Rebuttal: We thank the reviewer for their input and for appreciating our contribution.
* **[Regret dependency on $d$]** We would like to highlight that our regret bounds match the best known regret bounds for state-of-the-art (unbatched) efficient contextual linear bandit algorithms [Agrawal et al., 2013; Z... | Summary: This paper proposes a new batched contextual linear bandits algorithm that achieves $\widehat{O}(\sqrt{T})$ regret while using $O(\log\log T)$ batches. The algorithm combines various techniques from previous work including approximate barycentric spanner computation and reduction from contextual linear bandits... | Rebuttal 1:
Rebuttal: We thank the reviewer for their input and for appreciating our contribution.
* **[Empirical results]** Indeed, as the reviewer notes, in this work we focused on the theoretical understanding of efficient batch learning in contextual bandits, where we show a batched bandit algorithm that provably ... | Summary: This paper studies batched online learning for linear bandits and linear contextual bandits in large action space. It shows that the learner can use $O(\log \log T)$ batches (in each of which the sampling distribution over actions remain the same) to achieve near-optimal regret. Compared to previous work, the ... | Rebuttal 1:
Rebuttal: We thank the reviewer for their review and valuable comments.
* **[Distinction from existing algorithms]** We respectfully disagree that the contributions in our work are incremental with respect to previous work. Our work provides the first efficient batched algorithms, both for linear and conte... | Rebuttal 1:
Rebuttal: We thank all the reviewer for their input and for appreciating our contribution. This common part is devoted to summarizing the novelty of our proposed schemes. Detailed responses on the reviewers questions are provided as a response to each reviewer. We also attach a figure with preliminary sim... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
PrObeD: Proactive Object Detection Wrapper | Accept (poster) | Summary: The authors propose a proactive object wrapper for the object detection method. The warper is similar to an input transformation that makes the object detection model focus more on the salient regions. The experiments show that a consistent improvement after the proposed method is applied on top of existing ob... | Rebuttal 1:
Rebuttal: Thank you for your feedback. We provide a global response containing positives comments and major concerns of all reviewers. We also provide a global pdf file containing tables and figures with reference to added evaluation by some reviewers. We address all of your concerns below:
$\textbf{Compar... | Summary: The paper presents a wrapper based on proactive object detection which can be utilized for both general object detection and camouflaged object detection problems. The proposed algorithm is flexible to different detection algorithms like FasterRCNN, Yolo and DeTR. Reasonable performance have been reported in b... | Rebuttal 1:
Rebuttal: Thank you for your feedback. We provide a global response containing positives comments and major concerns of all reviewers. We also provide a global pdf file containing tables and figures with reference to added evaluation by some reviewers. We address all of your concerns below:
$\textbf{Motiv... | Summary: Authors propose a new object detection wrapper, PrObeD, based on proactive scheme, which tries to enhance the performance of existing object detectors. PrObeD generate a custom template for each input images which makes the weights of object detector closer to the ideal weights. Besides, a decoder who recovers... | Rebuttal 1:
Rebuttal: Thank you for your feedback. We provide a global response containing positives comments and major concerns of all reviewers. We also provide a global pdf file containing tables and figures with reference to added evaluation by some reviewers. We address all of your concerns below:
$\textbf{Benef... | Summary: In this paper, the authors study the problem of proactive object detection. To be specific, the authors introduce a wrapper around an existing object detector that can produce a template image and improve the performance of the detector if the image is multiplied with this template. Prior works have proposed l... | Rebuttal 1:
Rebuttal: Thank you for your feedback. We provide a global response containing positives comments and major concerns of all reviewers. We also provide a global pdf file containing tables and figures with reference to added evaluation by reviewers. We address all of your concerns below:
$\textbf{Clarificati... | Rebuttal 1:
Rebuttal: $\textbf{Positives}$
We thank all the reviewers for their feedback regarding our work. All the reviewers have greatly appreciated our proposed work. The positives comments by the reviewers are mentioned below:
$1.$ Novelty of PrObeD (Reviewers rzBs, 93HY, NMUy).
$2.$ A new aspect to deve... | NeurIPS_2023_submissions_huggingface | 2,023 | Summary: The paper proposes a proactive approach to enhance the performance of object detection through adding a learnable image-dependent template into the image. The optimal template can effectively bridge the gap between the learned and the ideal weights for object detectors. As demonstrated by the results, the temp... | Rebuttal 1:
Rebuttal: Thank you for your feedback. We provide a global response containing positives comments and major concerns of all reviewers. We also provide a global pdf file containing tables and figures with reference to added evaluation by some reviewers. We address all of your concerns below:
$\textbf{Perfor... | null | null | null | null | null | null |
Truly Scale-Equivariant Deep Nets with Fourier Layers | Accept (poster) | Summary: The authors of this study introduce architectural modifications to the conventional CNN framework by incorporating Fourier layers. Their aim is to achieve scale-equivariance, addressing the anti-aliasing problem that occurs when images are downscaled in the continuous domain. To tackle this issue, the authors ... | Rebuttal 1:
Rebuttal: > Q1. Computational/time complexity.
Please see the general response for a detailed discussion.
> Q2. Discussion on limitation
Please see the general response.
> Q3. Typos
Thanks for the careful review. We will address the typos.
> Q4. Caption clarification for Fig. 1
Fig. 1 illustrates th... | Summary: This paper introduces Truly Scale-Equivariant Deep Nets with Fourier Layers, which directly formulate down-scaling in the discrete domain to address the anti-aliasing problem. The proposed methods are validated on the MNIST-scale and STL-10 datasets, demonstrating good classification performance while preservi... | Rebuttal 1:
Rebuttal: > Q1. Evaluation of large-scale datasets and dense prediction tasks.
Thanks for the suggestion. For a consistent evaluation, we followed prior works on scale-equivariance to evaluate using the MNIST and STL10 datasets, e.g., [36].
We agree that large-scale datasets and dense prediction tasks wo... | Summary: The submitted paper explores the problem of scale-equivariant neural networks.
To get the desired property by design, the authors propose new building blocks: a spatially local Fourier layer, a novel scale-equivariant pooling layer, and scale-equivariant way of applying activation functions.
All the introduced... | Rebuttal 1:
Rebuttal: > Q1. Computational or memory overhead of the approach
Please see the general response for a detailed discussion.
> Q2. Evaluation of dense prediction tasks, e.g., semantic segmentation?
For a consistent evaluation, we followed prior works on scale-equivariance to evaluate using the MNIST and S... | Summary: This paper proposes a formulation of scale equivariant networks, which takes anti-aliasing into account. This is in contrast to previous works, in which the issue of aliasing is often disregarded. The proposed method achieves competitive classification accuracy on MNIST-scale and STL-10, while maintaining zero... | Rebuttal 1:
Rebuttal: > Q1. Group equivariance vs. resolution agnostic.
We believe our work covers both aspects, where we construct equivariant features to build a classifier that is resolution agnostic.
First, we view our proposed method (Sec. 4.1) to be a method for learning equivariant features. Following the def... | Rebuttal 1:
Rebuttal: We thank the reviewers for their constructive feedback and appreciation of our contribution. We first address the common concerns. Individual questions are answered separately in each reviewer's thread.
> Q1. Computational and memory complexity of the method
We now provide further analysis and e... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Learning Descriptive Image Captioning via Semipermeable Maximum Likelihood Estimation | Accept (poster) | Summary: ## Summary
- In the paper, the authors introduce a new method for generating descriptive captions for images. They call it SMILE, which stands for "Semipermeable Maximum Likelihood Estimation." Unlike traditional image captioning models that often produce generic descriptions, SMILE allows for more detailed a... | Rebuttal 1:
Rebuttal: Thank you for the insightful understanding and valuable feedback. We address each of your concerns and questions as follows:
> **Q1: (a)** All metrics are automatic metrics. Have you considered RefCLIPScore instead of CLIPScore? **(b)** Perhaps another metric could be running an object detector.
... | Summary: This paper proposes a novel further training approach called SMILE (Submodular Mixup Language Evaluation) for descriptive image captioning, which aims to improve the diversity and quality of generated captions. SMILE introduce to suppress conciseness optimization while encourge richness optimization to incenti... | Rebuttal 1:
Rebuttal: Thank you very much for the thoughtful comments and valuable feedback. We address the specific concerns as follows:
> **Q1:** The criterion used to assess richness and conciseness appears to be relatively simple.
**A1:** We understand your concern, but we have to clarify that what we refer to as... | Summary: The authors proposed the SMILE optimization method to enrich the detailed description of images in Captioning models. The authors analyzed the optimization process in the Image Captioning task and divided the optimization process into conciseness optimization and richness optimization. By suppressing the conci... | Rebuttal 1:
Rebuttal: Thank you very much for the thorough understanding and insightful comments. We address the specific concerns and questions as follows:
> **Q1: (a)** SMILE shows a significant decline in the CLIPScore metric across multiple experiments, and there is also a significant drop in the Acc metric for se... | Summary: This paper argues that captioning models with maximum likelihood estimation as the training objective led to conciseness optimization, which limits the model to generate descriptive captions. To this end, they introduce semipermeable maximum likelihood estimation (SMILE), which allows richness optimization for... | Rebuttal 1:
Rebuttal: Thank you very much for the thoughtful comments and valuable suggestions. Our responses to the specific questions are as follows:
> **Q1: The SMILE is not generalized to the large language models.** It doesn’t observe significant improvement in response length or overall quality.
**A1:** The ini... | Rebuttal 1:
Rebuttal: We sincerely thank all the reviewers for their valuable feedback, which is crucial for further refining our work. After the initial submission, we dedicated considerable time to improving the paper's structure and writing. Consequently, some of the potentially confusing and ambiguous areas highlig... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Spatial-frequency channels, shape bias, and adversarial robustness | Accept (oral) | Summary: This study applies a well-established procedure from psychophysics, critical band masking, to an object recognition task on which human and a wide range of neural networks are compared.
The core finding is that while humans use roughly the same spatial frequency channels for object recognition as they do for e... | Rebuttal 1:
Rebuttal: **Strengths:**
>Overall, I really liked this study...
Thank you!
**Weaknesses:**
>Significance testing:
>Figure 6 does not correct for multiple comparisons...
Done. We applied the correction to our results and do not see any change in any of the results discussed in the paper. There were a fe... | Summary: The authors propose to measure the differences in how humans and models rely on spectral information to recognize objects by using critical band masking to ablate. They conduct a human study revealing that humans detect objects with the same channels they use to detect letters and grating. On the other hand, m... | Rebuttal 1:
Rebuttal: **Strengths:**
>Originality
The paper explores a less common yet highly important research direction. Some "model" findings were partially reported in literature, but for the most part the findings are highly original.
>Quality
The submission is of high quality (minor weaknesses see below). Th... | Summary: This paper presents an analysis of neural network response to object classification from the perspective of frequency components. The authors use an analytic tool based on critical band masking, which is a technique borrowed from neuroscience.
The focus of the network to use certain frequency bands for object... | Rebuttal 1:
Rebuttal: **Strengths:**
>Use of a novel approach to analyze neural networks: the critical band masking…
>Extent of models considered in the analysis …
Thank you!
**Weaknesses:**
>Usefulness and outlook of the results: …
Thank you for raising this important point. We agree that the significance of the... | Summary: This work presents an empirical analysis of the sensitivity of humans and neural networks to critical band masking. This is the process of perturbing specific spatial frequencies, and determining which frequencies and at what magnitude, will lead to recognition errors. This in turns identifies the frequencies/... | Rebuttal 1:
Rebuttal: **Strengths:**
>Originality
>Critical band masking-based …
>Quality
>The design of the study is mostly sound: 1100 images from …
>The evidence for the claims of bigger channel bandwidth than humans, inverse correlation of bandwidth and shape bias, and …
Thank you!
>Clarity
>The paper is we... | Rebuttal 1:
Rebuttal: Dear Reviewers kLx7, QrRK, 9K3N, vBv1 and TCTa
Thank you for your thoughtful reviews, with scores ranging from 4 to 10. The paper seems much improved by our efforts to respond to your questions and comments. We are particularly grateful for the explicit questions asking us how our work can be use... | NeurIPS_2023_submissions_huggingface | 2,023 | Summary: I have read the authors' rebuttal and will maintain my already very high rating.
This paper discovers a new fact (which should be replicated by others) about human vision: That we use the same frequency band for objects as we do for words and gratings. It seems remarkable that this has not been tested before,... | Rebuttal 1:
Rebuttal: **Strengths:**
>-This is one of those papers that one wishes one had written. The result about the human bandwidth is Nature-worthy!
>-The sequence of experiments: First human psychophysics, then network “psychophysics”, then showing that these measurements of noise-sensitivity (mean, std dev., ... | null | null | null | null | null | null |
Tester-Learners for Halfspaces: Universal Algorithms | Accept (oral) | Summary: This work studies the problem of designing tester-learners for halfspaces in the agnostic setting and under Massart noise. In the setting of tester-learners, we do not make distributional assumptions on our samples, but rather require that our algorithm produces an accurate output whenever the input satisfies ... | Rebuttal 1:
Rebuttal: We wish to thank the anonymous reviewer for their suggestions and comments!
Prior work in the area has demonstrated that, in the distribution specific setting, assuming that the marginals are “nice” (i.e., that $\lambda$ is some constant) is sufficient and there is no dependence of the error boun... | Summary: The paper develops a tester-learner for halfspaces in the recently proposed testable learning framework of [RV23]. A recent work [GKSV23] had studied the same problem and proposed tester-learners for halfspaces with Massart noise and in the agnostic setting, for any particular strongly-log concave distribution... | Rebuttal 1:
Rebuttal: We thank the reviewer for their constructive feedback. Here are our responses to the questions, which we will aim to incorporate in a future revision:
Theorem 4.1 states that in the Massart noise case, we can achieve the optimal $\mathrm{opt}+\epsilon$ guarantee, while in the agnostic setting, we... | Summary: The paper proposes the first tester-learner for learning the class of halfspaces universally over a class of strongly log-concave distributions (or the class of all log-concave distributions under the KLS assumption). Unlike prior works that crucially rely on testing a specific given marginal distribution whil... | Rebuttal 1:
Rebuttal: We wish to thank the anonymous reviewer for their feedback and for appreciating our work!
The reviewer is right that it is an interesting open question whether our techniques can be applied to achieve optimal guarantees (i.e., $\mathrm{opt}+\epsilon$) for testably learning halfspaces when the noi... | Summary: Learning halfspaces is a very well studied problem in machine learning. In the agnostic (adversarial label noise) and distribution free setting, this problem has been known to be computationally intractable. As a result, there have been several works of agnostic learning in distribution specific settings (wher... | null | null | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Describe, Explain, Plan and Select: Interactive Planning with LLMs Enables Open-World Multi-Task Agents | Accept (poster) | Summary: This paper focuses on building an LLM-based planner for sequential decision-making tasks in open-world environments such as Minecraft. It proposes a method called "Describe, Explain, Plan and Select" (DEPS), which incorporates multiple stages including LLM describe, LLM explain, planning, and goal selection. T... | Rebuttal 1:
Rebuttal: > Concerns on Explainer
>
Regarding the example mentioned by the reviewer (L523-L549), the explanation generated by the Explainer contains the potential reason for the current failure, not just summarizing the current state. Specifically, the LLM is prompted to explain the reason for the failure... | Summary: The paper discusses the problem of planning in Minecraft and proposes an interactive planning approach based on Large Language Models (LLMs) called DEPS. The LLM based planner has the ability to re-plan upon failure and can take distance into consideration.
Strengths: 1. To my knowledge, this is one of the fi... | Rebuttal 1:
Rebuttal: Thanks for your comments. We provide clarifications below on the major comments, which will be included in the final version.
> It is quite confusing that the long standing ObtainDiamond challenge has been re-labeled as MT8 in a lot of result tables
>
Thanks for your comments. We will further cl... | Summary: The paper proposes a new planning algorithm for solving tasks in open-ended environments, and tests its performance against baselines in Minecraft. The proposed model uses large language models (LLMs) to split the large task into subgoals (for which there are controllers that can complete the subgoals). Howeve... | Rebuttal 1:
Rebuttal: Thanks for your comments. We provide clarifications below on the major comments, which will be included in the final version.
> Concern on the set of controllers and their abilities.
>
DEPS primarily focuses on generating more feasible plans for long-horizon tasks in an open-world scenario, usin... | Summary: The paper presents a novel approach called "Describe, Explain, Plan and Select" (DEPS) for task planning of multi-task embodied agents within open-world environments. The authors identify two main challenges in this area: the need for accurate and multi-step reasoning due to the long-term nature of tasks, and ... | Rebuttal 1:
Rebuttal: Thanks for your comments. We provide clarifications below on the major comments, which will be included in the final version.
> Similar works in zero-shot task planning.
>
Please refer to Q2 and A2 in the global rebuttal.
> The explanation of the selector could be improved. This seems like a w... | Rebuttal 1:
Rebuttal: Appreciate the reviewers’ constructive comments. We provide clarifications on the common comments as follows, which will be amended in the final version.
> Q1: Concerns on using OpenAI closed Language Models.
A1: We noticed that there are some concerns about DEPS using OpenAI's language model, a... | NeurIPS_2023_submissions_huggingface | 2,023 | Summary: In order to solve the long-term planning problem in the open world, this paper designed an interactive planning framework based on LLMs, and trained the first planning-based agent that successfully implements the ObtainDiamond task in the Minecraft scene.
Strengths: 1. The article is clear and the idea propo... | Rebuttal 1:
Rebuttal: Thanks for your comments. We provide clarifications below on the major comments, which will be included in the final version.
> The failure criteria of the controller that triggers the re-planning mechanism is not clear enough. Can the author explain in detail how to determine whether the failure ... | null | null | null | null | null | null |
IDRNet: Intervention-Driven Relation Network for Semantic Segmentation | Accept (poster) | Summary: This study focuses on the utilization of deletion diagnostics for modeling pixel relationships. The main contribution of this work can be divided into two parts. Firstly, the simplification of pixel relation modeling as object relation modeling. Secondly, the introduction of deletion diagnostics to enable netw... | Rebuttal 1:
Rebuttal: We highly appreciate your insightful reviews and positive comments on our novel methodology, the extensive evaluation, and clear writting. Our responses are shown below.
> **Q1: Missing results for Mask2Former+IDRNet on LIP/COCO-Stuff/Pascal Context.**
Following your suggestion, we also conduct... | Summary: This work presents IDRNet, a context aggregation scheme for semantic segmentation networks which is driven by intervention in the object query stage. The work is a follow up on OCR (object-contextual representations), but adds to the previous work by an intervention-based strategy to update, augment, or delete... | Rebuttal 1:
Rebuttal: We sincerely thank you for the comments. We hope that our responses could thoroughly address your concerns.
> **Q1: How this method is different from other works performing token pruning.**
We are glad to clarify the differences between our work and those token pruning works.
- The objectives... | Summary: This paper presents a context relation modeling method, which can be incorporated with existing semantic segmentation model. Instead of modeling pixel-level relation, it models class-level semantic relation by weighted averaging features at each location and deletion diagnostics. The proposed module is flexibl... | Rebuttal 1:
Rebuttal: We highly appreciate your insightful reviews and positive comments on our methodology design, extensive evaluation, and framework generalizability. Our responses are shown below.
> **Q1: Explanation of symbols in Section 3.**
We are glad to explain the meaning of these symbols and will add thes... | Summary: To further enhance the segmentation performance, the authors utilize deletion diagnostics to model pixel relationships. Specifically, to address the computational cost of pixel relation modelling, they simplify it as object relation modelling. Then, they introduce deletion diagnostics to facilitate the network... | Rebuttal 1:
Rebuttal: We sincerely thank you for the comments. We hope that our responses could thoroughly address your concerns.
> **Q1: Confusion of IDRNet.**
We are sorry for the confused representation of IDRNet. In the revision, we will utilize ***IDRNet*** to indicate the whole segmentation framework and ***I... | null | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Circuit as Set of Points | Accept (poster) | Summary: This paper introduces a novel perspective on circuit design by treating circuit components as point clouds and utilizing point cloud perception methods for feature extraction. Previous approaches either relied on hand-crafted methods to transform circuit designs into images for CNN feature extraction or employ... | Rebuttal 1:
Rebuttal: We would like to thank you for your detailed comments to help us improve our work, and we will improve our manuscript correspondingly. The responses to the main concerns are as follows.
**Question 1**：In figure 5 shows, 3x model size than unet. Which creates the question of scaling model. If a sm... | Summary: This paper proposes to use point cloud perception methods to extract features from the circuit by treating the circuit components as point clouds. The proposed method solves the infeasibility of end-to-end training in the CNN-based method and the limitation of time-consuming preprocessing in the GNN-based meth... | Rebuttal 1:
Rebuttal: Thanks for the valuable comments and suggestions which help us improve this work. Following are our detailed responses to your concerns.
**Question 1**：In the decoder part, motivations for using ResNet18 and UNet++ are missing. Is the selection of network architecture in the decoder part based on... | Summary: This paper proposes a new perspective on circuit design by viewing circuit components as point clouds and using point cloud perception methods to extract features from circuits. This achieves the goals of directly extracting features from raw data, end-to-end training, and high performance. Experimental result... | Rebuttal 1:
Rebuttal: Dear Reviewer rfmi
Thank you for your recognition of our work. If you have any new questions or concerns after reviewing our discussions with other reviewers, we welcome further discussion and are available to address any additional inquiries you may have. Please feel free to reach out to us.
Be... | Summary: The manuscript presents several claims. Firstly, it introduces a novel perspective on "point cloud perception methods" for performing EDA tasks. By treating circuit components as point clouds and utilizing point cloud perception networks, the paper proposes an innovative approach to solving these tasks. Second... | Rebuttal 1:
Rebuttal: We would like to thank you for your detailed comments to help us improve our work, and we will improve our manuscript correspondingly. The responses to the main concerns are as follows.
Firstly, I would like to apologize for the inconvenience caused. Due to the word limit imposed on rebuttal, to ... | Rebuttal 1:
Rebuttal: ##
| Method | pearson | spearman | kendall |
| -------------- | ---------- | ---------- | ---------- |
| Gpdl [22] | 0.5202 | 0.1880 | 0.1423 |
| UNet [31] | 0.5253 | 0.1481 | 0.1124 |
| UNet++ [45] | 0.4593 | 0.1862 | 0.1423 |
... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Time-Independent Information-Theoretic Generalization Bounds for SGLD | Accept (poster) | Summary: The paper provides novel information-theoretic generalization bounds for the SGLD. They derive the generalization error bounds by focusing on the time evolution of the Kullback–Leibler divergence. The bound is time-independent and removes the step size dependence in previous works.
Strengths: 1. This paper... | Rebuttal 1:
Rebuttal: We would like to express our deepest appreciation for your thoughtful reviews. We sincerely summarize our responses to you as follows.
---
## For Weaknesses (Questions)
### For weakness 1 (about dissipativity):
The dissipative assumption plays an important role in guaranteeing the geometrical con... | Summary: This paper provides novel generalization bounds for SGLD under smooth and dissipative loss conditions via the information-theoretic approach. The proposed bounds are time-independent and decay to zero as the sample size increases, regardless of the number of iterations and whether the step size is fixed. In ad... | Rebuttal 1:
Rebuttal: We would like to express our deepest appreciation for your insightful reviews and for giving us very convincing papers.
We sincerely summarize our responses to you as follows.
---
## For Questions
Thank you for your insightful feedback.
As you pointed out, our derived bounds are based on the mu... | Summary: The paper develops "time-independent" information-theoretic generalization bounds for SGLD. The key ingredients are: assuming the smoothness and dissipativity assumptions of the loss function , tracking evolution of the weight densities for two independent training sets, and their KL divergence. Such an approa... | Rebuttal 1:
Rebuttal: We would like to express our deepest appreciation for your time and careful check.
---
## For Weaknesses
### For weakness 1 (carelessness for dissipativity):
First and foremost, we would like to extend our sincerest apologies for our carelessness.
As you pointed out, there was a typo in the dissi... | Summary: In the literature, there are two approaches to obtain generalization performance bounds. One approach is to study the algorithmic stability and there are well-known results that can bound the generalization error using the algorithmic stability bounds. Another approach is the information-theoretic analysis, wh... | Rebuttal 1:
Rebuttal: We would like to express our deepest appreciation for your insightful reviews.
We sincerely summarize our responses to you as follows.
---
## For Weaknesses
We appreciate your careful check. As you pointed out, the term $V_{\Delta}:=\max_{t\in\[0, T\]}\beta \mathrm{Var}[\nabla F(W_{t}, B_{t})]$ c... | Rebuttal 1:
Rebuttal: Dear all PCs, SACs, ACs, and reviewers,
We sincerely appreciate your consideration of our paper.
We also would like to extend our deepest appreciation to the reviewers for providing insightful reviews.
Before summarizing our responses to each of the reviewers, due to word limitation, we organize ... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
StyleDrop: Text-to-Image Synthesis of Any Style | Accept (poster) | Summary: This paper proposes a new method that conducts the image synthesis following a new style using a text-to-image model. The method is called StyleDrop, which works by efficiently learning a new style by fine-tuning few parameters, and improving the quality via iterative training with either human or automated fe... | Rebuttal 1:
Rebuttal: * Fine-tuning efficiency
* Training an adapter takes only 3 mins on TPU v3 with the batch size of 8 for 1000 steps for base transformer and 17 mins for super-res transformer for both rounds. Note that the super-res transformer does not need to be trained at every iterative training as it mostly ... | Summary: The paper introduces a method to fine-tune a pre-trained large text-to-image model using a few images that provide an artistic style, so that the fine-tuned model can generate images that follow the provided artistic style. The proposed method is a combination of Muse [1], a transformer-based text-to-image gen... | Rebuttal 1:
Rebuttal: * DreamBooth is not bad at all, so I guess the performance improvement is marginal?
* We observe significant style score improvement with StyleDrop over DreamBooth in Tab 2 (0.644 → 0.705). Moreover, the user preference study clearly shows that the StyleDrop (86%) generates much more style consi... | Summary: To synthesize arbitrary image styles, this paper proposed a StyleDrop method that enables an image synthesis with the style of a given image, using text-to-image models, e.g., Muse and Imagen. StyleDrop consists of three crucial components: 1) a transformer-based text-to-image generation model using Muse, 2) a... | Rebuttal 1:
Rebuttal: * Visual comparison to Muse baseline
* Please find a visual comparison to the Muse baseline in Figure R1 of the attached rebuttal PDF in the global response. Generated images by Muse and StyleDrop share the same text prompt (e.g., "A baby penguin in watercolor painting style"), but we see that t... | Summary: This paper works on personalized text-to-image synthesis. Given reference style images (even a single one), the model can adapt the style to the synthesized images. There are several key components in the papers including: using a transformer-based t2i model rather than a diffusion-based one; adding an adaptor... | Rebuttal 1:
Rebuttal: * Reproducibility
* We'll provide more complete pseudo-code on the training and inference of the StyleDrop, in addition to what we've already provided in the paper (e.g., dataset, hyperparameters, pseudo-code on generating adapters). In addition, we'll look into options to release the StyleDrop ... | Rebuttal 1:
Rebuttal: We thank reviewers for reviewing our submission. We are encouraged with positive feedbacks on the **quality of image generation results** (9biG, bhPK, xBBj, oHZZ, vFSY), **extensive experiments** (bhPK, xBBj, oHZZ, vFSY), **well-motivated method** (bhPK, xBBj), and **well-written manuscript and pr... | NeurIPS_2023_submissions_huggingface | 2,023 | Summary: This paper proposes StyleDrop to perform text-to-image generation following the style of a reference image. StyleDrop finetunes Muse [1] via parameter-efficient fine-tuning (PEFT) using the style reference images and their corresponding text prompts. To better finetune Muse, this paper proposed (1) a templated... | Rebuttal 1:
Rebuttal: * Technical contribution
* We note that the most previous works (e.g., [DreamBooth](https://dreambooth.github.io/), [Custom Diffusion](https://github.com/adobe-research/custom-diffusion)) on t2i adaptation are based on diffusion models, and no framework has been proposed for t2i generative visio... | null | null | null | null | null | null |
Distributed Personalized Empirical Risk Minimization | Accept (poster) | Summary: This paper proposes a new paradigm for learning from multiple heterogeneous data sources to achieve optimal statistical accuracy across all data distributions without imposing stringent constraints on computational resources shared by participating devices. The proposed Personalized Empirical Risk Minimization... | Rebuttal 1:
Rebuttal: Thank you for your constructive comments! We will try to address your concerns as follows.
**Inadequate evaluation and lack of verification on real data**
We discuss this issue in Global comment. We have already reported results on real data sets in Appendix C of main paper and reported new resul... | Summary: This paper studies a distributed training strategy to obtained a personalized model that shares a bases for all clients while an adaptor for each clients.
Strengths: The overall derivation of the algorithm is theory-driven and reasonable.
Good experiment result in the simulation dataset.
Weaknesses: This pa... | Rebuttal 1:
Rebuttal: Thank you for your valuable comments! We will try to address your concerns as follows.
**Limitation of weight sum based mixing method**
In multi-source learning, weighted summing source domains and learning on it is the most popular and natural method [1,2], due to its simple form, easy implement... | Summary: This paper proposes an optimization strategy for federated learning (FL). A key challenge of FL is data heterogeneity, where each client device has different data distributions. Unlike a naive FL, which minimizes the empirical risk simply weighted over clients (called WERM), the proposed PERM first minimizes t... | Rebuttal 1:
Rebuttal: Thank you for your compliments and valuable comments! We will try to address your concerns as follows.
**The scalability of the proposed method**
We discuss the scalability of PERM in comparison to existing methods in Global answer. To summarize, the compute burden on clients and server is roug... | Summary: The paper proposes a novel learning framework for personalizing empirical risk minimization in a given federated learning problem where the data is heterogeneous (non-IID). The core idea of the paper is to use optimization relaxations and data shuffling to achieve near-optimal solutions for the general NP-har... | Rebuttal 1:
Rebuttal: Thank you for your valuable comments! We will try to address your concerns as follows.
**The reliance on the strong concavity and smoothness is a serious limitation**
Strongly convexity and smoothness are standard assumptions in the proof of shuffling-based optimization algorithm. Our objective i... | Rebuttal 1:
Rebuttal: We would like to thank all reviewers for your careful and constructive comments. We appreciate the compliments from reviewers on our algorithmic novelty and theory and will gladly incorporate the suggestions.
**Empirical evaluation on real data** We understand the main concern raised is that ou... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Disentangling Voice and Content with Self-Supervision for Speaker Recognition | Accept (poster) | Summary: In this paper, the authors proposed RecXi, a framework that disentangles the voice and content for speaker recognition. Also, the authors used self-supervised learning methods to improve performance. The authors focused on the fact that speaker information is static and the content information is dynamic, so t... | Rebuttal 1:
Rebuttal: Dear Reviewer,
Thank you for taking the time to review our work and providing us with your valuable suggestions and generous comments! **More experiments are presented in the PDF file in the ‘Global’ rebuttal following this year’s NeurIPS guideline.** Please kindly let us know during discussion if... | Summary: This paper describes a novel speaker representation based on a three layer Gaussian inference network for speaker recognition. The model approach uses a static representation for layers 1 and 3, but a time-variant representation for layer 2. The authors position this network structure as disentangling speake... | Rebuttal 1:
Rebuttal: Dear reviewer:
We sincerely appreciate your effort in reviewing our work and providing valuable and insightful suggestions. In response to your feedback, **we have conducted additional experiments, and the results are included in the PDF file attached to the 'Global' rebuttal, following this year... | Summary: In this paper, authors propose a disentanglement framework for speaker recognition that extract speaker representations from the mixture of speaker traits and content. The proposed framework utilizes three Gaussian inference layers to decompose the static and dynamic speech components. Additionally, a self-sup... | Rebuttal 1:
Rebuttal: Dear Reviewer,
Thank you for taking the time to review our work and providing us with your valuable suggestions and insightful comments. We have carefully considered your feedback, and we would like to address your concerns as follows:
**Weakness:** Thanks for pointing this out. We use the best ... | Summary: This paper proposes an improvement over Xi-vector speaker recognition algorithm. They propose an iterative 3 layer Gaussian inference mechanism to disentangle the content information from speaker information. The first layer gets precursor speaker representation similar to Xi-vector framework. The second layer... | Rebuttal 1:
Rebuttal: Dear Reviewer:
Thanks for reviewing our work and providing valuable and insightful suggestions. **More experiments are reported in the PDF file in the ‘Global’ rebuttal, as per NeurIPS guideline.** Please let us know if you have any difficulty accessing the file. **Due to the word limit of the ... | Rebuttal 1:
Rebuttal: Dear all reviewers,
We greatly appreciate the insightful suggestions and valuable comments from each reviewer. These have been immensely helpful and enlightening for refining this paper.
We have included a figure and several experimental results to aid our rebuttal. **Following the NeurIPS guide... | NeurIPS_2023_submissions_huggingface | 2,023 | Summary: The authors propose RecXi, a Gaussian inference based disentanglement method for speech that itself is based on Xi-Vector. Overall, the methods makes sense with sufficient motivation and extensive experiments. The results are promising, and have further mileage in ASR, voice conversion and TTS.
The paper's ma... | Rebuttal 1:
Rebuttal: Dear reviewer,
Thank you for reviewing our work and sharing the valuable suggestions and generous comments! Following is our rebuttal to address the concerns.
**Weakness 1**: Thanks for pointing this out. It is true that we do not consider explicitly prosody features in our approach. As observe... | null | null | null | null | null | null |
Generalizing Importance Weighting to A Universal Solver for Distribution Shift Problems | Accept (spotlight) | Summary: The paper studied importance weighting that is a fundamental technique in machine learning and data mining. It pointed out a serious issue of existing importance weighting methods that has not been realized or considered carefully --- the support of the training distribution may not fully enclose the support o... | Rebuttal 1:
Rebuttal: Thanks for spending a lot of time reviewing our manuscript!
***1. Experiment scale***
Please find our reply *"1. On the common concern about the experiment design"* in the top-level rebuttal above as well as the reply to Reviewer uQtu (https://openreview.net/forum?id=KmdlUP23qh¬eId=m5Q81fgcQF... | Summary: This paper focuses on an important technique in the field of learning under distribution shift. Compared to previous importance weighting techniques, this paper consider more general scenarios: support shift. This paper first investigates how the previous techniques fail on general support-shift scenarios, the... | Rebuttal 1:
Rebuttal: Thanks for your constructive comments!
For *Questions 2 to 4*, please find our replies in the top-level rebuttal above as well as the reply to Reviewer uQtu (https://openreview.net/forum?id=KmdlUP23qh¬eId=m5Q81fgcQF). In addition, do you have any suggestions for the type of support shift that ... | Summary: The paper introduces a Generalized Importance Weighting (GIW) method designed to address distribution shift problems, where the support of the probability density function varies between training and test data. The authors demonstrate that GIW outperforms existing methods in scenarios where the test support is... | Rebuttal 1:
Rebuttal: Thanks for your constructive comments!
The experiment scale is concerned by 3 reviewers (i.e., 3 reviewers commented on the size of datasets and 2 commented on the size of models), so please find our reply *"1. On the common concern about the experiment design"* in the top-level rebuttal above. I... | Summary: The paper considers a distribution shift problem where the test distribution $D_\text{te}$ is different from the training distribution $D_\text{tr}$.
Specifically, the learner is given training data sampled according to $D_\text{tr}$ and (typically a much smaller size of) validation data sampled according to $... | Rebuttal 1:
Rebuttal: Thank you for reviewing but there seem to be some serious factual misunderstandings. We will clarify the issues one by one.
***1. "generally, the notion of consistency is defined for estimators"***
Here, we guess estimators meant the empirical objectives from the previous sentence. This claim is... | Rebuttal 1:
Rebuttal: ***1. On the common concern about the experiment design***
We sincerely thank all the reviewers. First of all, since 3 reviewers asked for additional experimental comparisons of the proposed method against the baseline methods on ImageNet, we'd like to explain why it is difficult to do so. We wil... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
An Information Theory Perspective on Variance-Invariance-Covariance Regularization | Accept (poster) | Summary: This paper presents an information-theoretic analysis of the self supervised learning algorithm VICReg. The authors first introduce an approach to study neural networks using an information-theoretic perspective by shifting the stochasticity from the networks to the input data. Therefore they derive an informa... | Rebuttal 1:
Rebuttal: Dear Reviewer 2fL7,
We appreciate the comprehensive review and your valuable insights. Your insights will significantly enhance our paper. We want to address your comments in detail:
## Extending the Empirical Evaluations
Our paper emphasizes VICReg's theoretical aspects and advantages over ot... | Summary: This paper analyzes the SSL methods from an information-theoretic perspective. Unlike previous theoretical analyses, the paper focuses on deterministic networks more suitable for real-world scenarios. Assuming a data distribution as a mixture of Gaussians (MoG) and a deep neural network (DNN) as a conditional ... | Rebuttal 1:
Rebuttal:
Dear Reviewer sVT5,
Thank you for your comprehensive review and the constructive feedback you've provided. Your insights will significantly enhance our paper. We want to address your comments in detail:
## Weaknesses:
### Limitations of assuming randomness from DNNs
Numerous methods propose u... | Summary: This paper presents two theoretical results on the VICReg self-supervised learning technique. Taking an information theoretic perspective, it shows that the VICReg objective can be seen as maximizing a lower bound on the mutual information between Z and X' where X' is an augmented version of data point X and Z... | Rebuttal 1:
Rebuttal: We appreciate your detailed review and the constructive feedback you've provided. Your insights are invaluable and will significantly contribute to the refinement of our paper. Allow us to address each of your concerns in detail:
## Weaknesses:
### Empirical Evaluations
Our primary goal was to ... | Summary: This paper explained the effect of VICReg learning method using information theoretic perspective under some assumptions on the distribution of the input data. Based on the assumption, it connected the VICReg to mutual information minimization, and proposed new self-supervised method that performed better than... | Rebuttal 1:
Rebuttal: Dear reviewer tvXs,
We appreciate your constructive comments and suggestions. We believe your feedback will significantly help us improve our paper's quality and readability. Here, we address each of your concerns.
## Weaknesses:
1. **Presentation of the Paper** - We value your feedback on impr... | Rebuttal 1:
Rebuttal: Dear Reviewers,
Thank you for your thoughtful and detailed reviews of our work. We appreciate your time and the constructive feedback you have provided. We have carefully considered your comments and concerns and would like to address them in a unified response:
While our empirical explorations ... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Transferable Adversarial Robustness for Categorical Data via Universal Robust Embeddings | Accept (poster) | Summary: This paper proposes to use the input embeddings from neural networks as universal robust embeddings. They can be used to enhance the robustness of robustness of machine learning models.
Strengths: This paper is well written and well formulated.
Weaknesses: - It is better to add more theoretical analysis abou... | Rebuttal 1:
Rebuttal: We thank you for your comments.
> *It is better to add more theoretical analysis about why embeddings from neural networks still work on other machine learning models.*
We agree with this suggestion. We mentioned the potential reasons behind this in Lines 197-200: *“By using first layer embeddin... | Summary: This paper investigates adversarial robustness of tabular data. By transferring embeddings achieved by bilevel alternating minimization framework, tree-based models can then learn robustness without the need of adversarial training.
Strengths: The proposed approach is novel and seems to be general for wide ap... | Rebuttal 1:
Rebuttal: We thank you for pointing out the novelty and practicality of our approach. We omitted the comparison with Chen et al. (2021) in our work, since we consider its threat model to be sufficiently different. The two key differences are:
1. Chen et al. operate within per-feature constraints, while we ... | Summary: This paper presents a method to train adversarially robust deep networks for tabular data and to transfer this robustness to other classifiers via universal robust embeddings tailored to categorical data. These embeddings created using a bilevel alternating minimization framework can be transferred to boosted ... | Rebuttal 1:
Rebuttal: We thank the reviewer for the positive assessment of our work.
> *Yet, it would be easier to understand the proposed approach if this paper could provide more explanations about the difference of using gradient-based adversarial training on the tabular data rather than the image or text data.*
T... | Summary: This study presents novel approaches aimed at enhancing the robustness of models trained on categorical data on the deep networks and tree-based models. It addresses the challenges posed by real-world adversaries within the constraints of financial resources. Through extensive empirical analysis, they demonstr... | Rebuttal 1:
Rebuttal: We thank you for the positive feedback.
> *1. The assumption in line 138 is that the cost of modifying features is additive. If the two features have the inverse function for the final decision, is it reasonable to use this assumption?*
Yes, we are confident that it is a reasonable assumption si... | Rebuttal 1:
Rebuttal: We thank the reviewers for the feedback. We are glad that our work received positive comments such as:
- *“the potential for impact is high (at least insofar as adversarial robustness is impactful)”* (**Reviewer saBk**).
- *“The paper is good-written and easy to follow”* (**Reviewer a4Uk**)
- *”T... | NeurIPS_2023_submissions_huggingface | 2,023 | Summary: The authors propose an adversarial training algorithm for tabular data with categorical features which can produce robust embeddings for use with downstream non-differentiable models like decision trees. The authors evaluate on three datasets and with several downstream models on the robust embeddings.
Stren... | Rebuttal 1:
Rebuttal: We thank you for the detailed response. We discuss the raised concerns below.
> *Gradient boosted decision tree methods like XGBoost and CatBoost often achieve superior performance compared to neural networks ... Do they maintain their advantage over neural networks after adversarial training (bot... | null | null | null | null | null | null |
Online Label Shift: Optimal Dynamic Regret meets Practical Algorithms | Accept (spotlight) | Summary: The general scope is online classifier learning under online label shift where the key assumption is that the class conditionals $Q(x|y)$ are invariant. The paper focuses on both unsupervised and supervised problems, which they coined as Unsupervised Online Label Shift (UOLS) and Supervised Online Label Shift ... | Rebuttal 1:
Rebuttal: We thank the reviewer for your generous appraisal of the main theoretical idea. We are glad that you find the paper detailed and well-written.
>**Although well acknowledged in the paper, the fact that the approach, like [8], requires memory that scales linearly with time is quite a concern from ... | Summary: The authors study the important problem of learning in an online setting.
In particular, they focus on two important problems:
- Online Unsupervised learning with Label Shift (UOLS)
- Online Supervised Learning with Label shift (SOLS)
In both cases, they assume, as in previous work, that the distribution of th... | Rebuttal 1:
Rebuttal: We thank the reviewer for your thoughtful feedback. We are glad that you find the paper well-written, and are grateful that you appreciate the experimental and theoretical results.
>**The results are limited to changes in the probability of obtaining an element of a certain class (label shift). ... | Summary: this paper applies techniques from online regression oracles to tackle the problems of unsupervised and supervised label shift, where the marginal probabilities of the labels can be changed over time by an adversary.
Strengths: the main contributions include:
C1 an algorithm that adapts to the changing label... | Rebuttal 1:
Rebuttal: We want to thank the reviewer for your positive feedback and for championing our paper. We are glad that you find the paper well-written and appreciate the experiments.
>**1 figure 2 does not explain what the uncertainty bands are. I assume these are std's or are they standard errors? similairly... | Summary: This paper delves into the supervised and unsupervised online label shift problem, which focuses on adapting to changing class marginals using online data. The authors propose new algorithms that convert the adaptation problem into an online regression task, guaranteeing the minimax optimal dynamic regret. The... | Rebuttal 1:
Rebuttal: We are thankful for Reviewer Cpae’s constructive feedback and positive perspective on the work. We have incorporated your feedback, and believe they have greatly enhanced the manuscript.
>**... "sample efficient"...**
We apologize for the confusion. The term “sample efficient” in the theoretica... | Rebuttal 1:
Rebuttal: We are thankful to the reviewers for their careful consideration of our work and thoughtful feedback. We are glad to find that the reviewers appreciate our work and gave positive scores (8, 8, 7, 4). We also appreciate that all reviewers find the paper well-written. While reviewer Cpae has some cl... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
VoxDet: Voxel Learning for Novel Instance Detection | Accept (spotlight) | Summary: The paper introduces a new 3D geometry-aware framework, VoxDet, to address the challenge of detecting unseen instances based on multi-view templates. VoxDet includes three main modules: a Template Voxel Aggregation (TVA) module, an open-world detection module, and a Query Voxel Matching (QVM) module. TVA trans... | Rebuttal 1:
Rebuttal: Dear Reviewer NFHY,
Thanks for your insightful questions and feedback. We provide specific answers below:
1. **Is this a novel experiment setup?**
Thanks for this question.
Yes, this is a novel setup, we have developed new training data (OWID) and test benchmarks (authorita... | Summary: This paper tackles the task of novel instance detection when given multi-view reference images.
The key is the utilization of a 3D voxel representation, contrary to 2D representations used by prior work that fair under occlusion and pose variation.
The proposed detector, VoxDet, combines reference features int... | Rebuttal 1:
Rebuttal: Dear Reviewer dtGe,
Thank you very much for the questions and comments. We address them as follows:
1. **VoxDet requires known rotations of the references during training.**
Thanks for raising this concern.
* It is true that most 2D methods didn't consider the geometry information du... | Summary: The manuscript introduces VoxDet, a system for novel instance detection. The method takes in a set of query images of an object instance and aims to detect this instance in a target image. The core novelty of the approach lies in the lifting of the query instance and all potential target instances (detected vi... | Rebuttal 1:
Rebuttal: Dear Reviewer W6QG,
We sincerely appreciate your feedback and concerns.
We provided detailed explanations and responses as follows:
1. **Concerns to reimplement this system due to complex framework figures**
We sincerely appreciate this concern.
- To ensure reproducibility ... | Summary: The paper introduces a method to address the problem of novel instance detection. The key challenge of this problem over object detection is that the network needs to detect novel types of objects that were not in the training data. In contrast to prior art, this paper presents a 3D geometry-aware framework t... | Rebuttal 1:
Rebuttal: Dear Reviewer 6hBZ,
Thanks a lot for providing valuable suggestions and raising insightful questions about our work.
We address and answer your questions below:
1. **Why not consider both translation and rotation in the alignment?**
Thanks for raising up this issue.
***In the TVA modu... | Rebuttal 1:
Rebuttal: Dear AC, Reviewers,
We would like to express our sincere gratitude for the careful inspection and constructive feedback from all the reviewers and the area chair. We are glad to see that **all of the reviewers in general hold a positive attitude** towards our paper in the pre-rebuttal period.
Fo... | NeurIPS_2023_submissions_huggingface | 2,023 | Summary: The paper presents a new 3D geometry-aware instance detection framework, VoxDet, which effectively transforms multi-view 2D images into 3D voxel features using a Template Voxel Aggregation (TVA) module. This representation demonstrates improved resilience to occlusions and pose variations. VoxDet also uses a Q... | Rebuttal 1:
Rebuttal: Dear Reviewer 7Hv8,
We would like to express our sincere gratitude for your careful inspection and constructive feedback on our paper.
We address your specific concerns as follows:
1. **Synthetic training data raises concerns about a potential domain gap**
Thanks for this concern.
... | null | null | null | null | null | null |
Demystifying Oversmoothing in Attention-Based Graph Neural Networks | Accept (spotlight) | Summary: The authors provide theoretical proofs on the oversmoothing problem in GCNs and GATs through Ergodicity analysis of infinite product of matrices. The authors provide a bound in convergence of oversmoothing problem and provides suggestions on choices in activation function to relief oversmoothing problem.
Stre... | Rebuttal 1:
Rebuttal:
Thank you for your constructive feedback and positive assessment of our work. Below, we provide individual responses to your questions.
**Q1: It would be better if the authors could add a experiment with a plot directly showing JSR.**
Thank you for the comment. As mentioned in the paper (line 3... | Summary: The objective of this paper is to examine the occurrence of over-smoothing phenomena in attention-based Graph Neural Networks (GNN). Over-smoothing refers to a situation where deepening the network results in similar node representations. The researchers demonstrate the presence of over-smoothing in attention-... | Rebuttal 1:
Rebuttal: We greatly appreciate your positive assessment and insightful comments, which have helped strengthen our work. Below, we provide individual responses to your questions.
**Q1: The main shortcoming of the paper is the lack of suggestions for solving the over-smoothing problem. [1] suggests scaling ... | Summary: This paper provides a definitive answer to the question that whether the graph attention mechanism can mitigate over-smoothing through a rigorous mathematical analysis. The theoretical results establish that the graph attention mechanism cannot prevent over-smoothing and loses expressive power exponentially.
... | Rebuttal 1:
Rebuttal: We thank the reviewer for taking the time to review our paper and providing constructive feedback. In line with the reviewer’s suggestion, we have performed additional numerical experiments, validating our theoretical findings on three heterophily datasets (Cornell, Texas and Wisconsin [1]) and a ... | Summary: The paper addresses the question of whether graph attention mechanism could prevent oversmoothing in Graph Neural Networks (GNN) and provides a negative result. Existing results on graph attention are available for 1-layer GNN. The current work analyzes the general case of multi-layered graph attention in the ... | Rebuttal 1:
Rebuttal: We appreciate your thoughtful comments and positive assessment of our work. After carefully reviewing your feedback, below we provide answers to the comments you raised.
**Q1: Some of the prior work on reducing oversmoothing [1] and impact of the current work for understanding them could be inclu... | Rebuttal 1:
Rebuttal:
We would like to thank the reviewers for carefully reading our paper and giving insightful comments. We have provided detailed responses to each of the reviews separately.
In the rebuttal pdf file, we also provide additional numerical results under the same experimental setup as in our paper. S... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Towards Optimal Effective Resistance Estimation | Accept (poster) | Summary: The authors consider the problem of estimating effective resistance both for all edges of a graph and between any pair of nodes. This quantity is incredibly useful for sampling from a graph to produce a sparser graph with similar spectral properties. Furthermore, this quantity can provide a useful metric for t... | Rebuttal 1:
Rebuttal: Thank you for your thoughtful comments and feedback! We have addressed the limitations of the expander case and high-level motivation for considering the general response (see Comment 3 therein). In addition, we have described the key bottleneck in the general response as well. Thank you again for... | Summary: The paper studies the (approximate) estimation of effective resistance on expander graphs, i.e., graphs with edge expansion $\Omega(1/\text{polylog}\ n)$. The main upper-bound result is a data structure (sketch) for effective resistance that takes $\tilde{O}(m/\epsilon)$ time to build, answers each query in $\... | Rebuttal 1:
Rebuttal: Thank you for your thoughtful comments and questions! In the general response, we have discussed how we plan to incorporate some of the changes to the preliminary/introduction of the paper in response to your helpful suggestions for content to include therein.
**Question 1: Clarifying question o... | Summary: In the paper, the authors study the problem of estimating the effective resistance between pairs of vertices in weighted, undirected expander graphs. The authors present a novel algorithm for effective resistance estimation and effective resistance sketching. Moreover, they give a conditional lower bound based... | Rebuttal 1:
Rebuttal: Thank you for your thoughtful comments and questions!
**Question 1: It would be helpful if the authors could provide a more close comparison between their work and that of [3].**
**Response:** Thanks for the suggestion. We would be happy to incorporate more of this comparison in the introducti... | Summary: The paper theoretically analyzes the time complexity of estimating effective resistances in $n$-node $m$-edge undirected, expander graphs. Specifically, the paper present an algorithm which can achieve $\tilde{O}(m \varepsilon^{-1})$ time complexity for estimating the effective resistance of all edges in such ... | Rebuttal 1:
Rebuttal: Thank you for your thoughtful comments and questions!
We have discussed interesting directions for future empirical experiments in the general response, as it was a common theme brought up by multiple reviewers. In addition, we have discussed the primary bottleneck to extending results to genera... | Rebuttal 1:
Rebuttal: We thank the reviewers for their helpful comments and questions. We are encouraged that the reviewers appreciated novelty in our technical contributions (R2, R3, R5) and cleverness in our techniques for obtaining our improved upper and lower bounds (R4, R6). We are glad that they found our paper t... | NeurIPS_2023_submissions_huggingface | 2,023 | Summary: Given a graph $G$, the effective resistance approximation problem is one where you are asked to approximate the effective resistances of some subset $S$ of vertex pairs. The authors have two contributions:
Firstly, the authors demonstrate an $O(m\epsilon^{-1} + |S|)$ algorithm for computing effective resistan... | Rebuttal 1:
Rebuttal: Thanks for your thoughtful comments and questions! We discuss more motivation for considering the expander case in the general response.
**(1) Why consider conditional lower bounds on algorithms using the subcubic non-FMM triangle detection hardness assumption? This is standard, but I can’t find... | Summary: The paper studies the problem of estimating the effective resistances in $n$-node $m$-edge undirected, expander graphs, and provides a $O(m/\epsilon)$ time algorithm which produces an effective resistances sketch in $O(n/\epsilon)$ size of space. This improves the runtime of the previous algorithm that estimat... | Rebuttal 1:
Rebuttal: Thank you for your thoughtful comments and questions! We have discussed interesting directions for future empirical experiments, as well as the modifications to the preliminaries section which you suggested in the general response. Please let us know if you have any additional feedback on those po... | null | null | null | null |
Wasserstein distributional robustness of neural networks | Accept (poster) | Summary: This paper presents a novel adversarial attack methodology premised on Wasserstein distributionally robust optimization. This approach serves as a generalized technique that encompasses many established adversarial attack methods such as FGSM and TRADES. To ensure computational tractability, the authors introd... | Rebuttal 1:
Rebuttal: Thank you for your comments and praise. This is much appreciated. Thank you in particular, for emphasising our contributions in introducing the ReDLR loss function and results in the "Bounds on Out-of-Sample Performance" section which, while important, we were not able to stress appropriately ours... | Summary: The paper focuses on the issue of distributional robustness in deep neural networks and introduces a novel approach to address it. The authors present a tractable distributional attack within a Wasserstein ball, which allows attackers to perturb inputs in a non-uniform manner. They propose a first-order approx... | Rebuttal 1:
Rebuttal: Thank you for your praise of our contribution. In particular, your comments that our results offer "key advancements to the current understanding and applications of deep learning models under adversarial conditions" as well as demonstrating "a deep and accurate understanding of the state-of-the-a... | Summary: The paper proposes a new method combining Wasserstein distance, DRO, and adversarial attack. The novelty is that the method computes the adversarial attack bounded by the Wasserstein distance. The paper provides a bound on the performance of out-of-domain samples. It applies the new adversarial attack on CIFAR... | Rebuttal 1:
Rebuttal: Thank you for your comments and questions which raise important points. While you thought our writing and presentation were clear, your questions allowed us to understand where we could have explained the key insights and contributions better, for which we are particularly grateful. We stride in t... | Summary: This paper links more general attacks with question of out-of-sample performance and Knightian uncertainty. To evaluate the distributional robustness of neural networks, this paper proposes a first-order AA algorithm and its multistep version. The proposed attack algorithms include Fast Gradient Sign Method (F... | Rebuttal 1:
Rebuttal: Re **Weakness 1** - We trust your first point was addressed by our new experiments showcasing the performance on CIFAR100 and ImageNet. In fact, while we agree that the complexity of Wasserstein distance in higher dim can cause troubles, this does not happen for our methods at all due to our formu... | Rebuttal 1:
Rebuttal: We are deeply grateful to all the reviewers for their comments and questions. We were pleased to hear the reviewers felt our paper was well structured and clearly presented. It was rewarding to see all the reviewers list numerous strengths of our work and very interesting that their emphasis was d... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Learning Depth-regularized Radiance Fields from Asynchronous RGB-D Sequences | Reject | Summary: This paper explained and analysed the issue of unsynchronous RGB and depth measurement under the UAV city modelling scenario, which makes it challenge to incorporate depth supervision for NeRF optimisation under such scenerio. The authors proposed a novel solution to it by modelling the continuous RGB camera t... | Rebuttal 1:
Rebuttal: We thank 9aLD for recognition. The detailed responses to the weakness section are as follows:
- **W1**: We present additional experiments and ablation studies in the author rebuttal section B, which demonstrates that our solution out-performs the baseline of directly using RGB poses as initializa... | Summary: The paper proposes a way to train NeRF with asynchronous RGBD videos. Specifically, three technical contributions have been made:
1. New problem formulation for NeRF training from async RGBD video.
2. Propose a time-pose function to use async RGB and Depth stream, resulting in better pose estimation and NeRF t... | Rebuttal 1:
Rebuttal: We thank JhZu for recognition and pointing out the depth distortion issue in real-world depth data. Addressing the depth distortion problem will be a promising future work but is not currently considered in this project. Our responses are three-fold:
(1) During the capture of our real-world data... | Summary: The authors propose a method to reconstruct aerial scenes with Neural Radiance Fields that are supervised with RGB images and asynchronous depth images. To address the asynchronicity they propose a novel time-pose function that provides a prior to optimise the poses of the depth images. To validate their metho... | Rebuttal 1:
Rebuttal: - **W1**: We provide additional experiment results that use the misaligned RGB poses (with calibrated sensor transformation $\mathcal{T}_{RGB\rightarrow D}$) as the initial value of depth sensor poses and perform BARF-style joint optimization. The quantitative results can be found in the Author Re... | Summary: This paper proposes to include depth supervision in NeRF in the UAV city modeling scenario. The key problem is that the images and depth maps are asynchronous. This paper exploits a prior that RGB-D frames are sampled from the same physical trajectory. It fits a time-pose function to the available RGB cameras ... | Rebuttal 1:
Rebuttal: We thank 2eof for recognition and professional feedbacks. Here are detailed responses:
- **Weakness #1 & #2**: We provide another ablation experiment in the Author Rebuttal section A. As demonstrated by the experiment results, our time-pose function out-performs linear interpolation.
- **Weakness... | Rebuttal 1:
Rebuttal: We thank all the reviewers for their professional and in-depth comments. We respond to two shared concerns here.
## A. Comparisons with linear interpolation and non-hash implicit network architectures (along with wall-clock runtime)
> The weakness 1&2 of reviewer 2eof, weakness 4 of reviewer GDq... | NeurIPS_2023_submissions_huggingface | 2,023 | Summary: This paper proposes a pipeline to build depth-supervised neural radiance fields (NeRF) using asynchronous RGB-D sequences. Since the task is novel, the paper also contributes a synthetic dataset and demonstrates that it outperforms certain baselines in the experiments.
Strengths: ## Originality
The task is n... | Rebuttal 1:
Rebuttal: ## Weakness 1
> The most obvious baseline is missing. What if authors simply initialize the depth frames with their (misaligned) poses from the sensor and perform BARF [1] style joint optimization on the depth frames’ poses (note that depth frames can have different poses from the RGB images after... | null | null | null | null | null | null |
Layer-Neighbor Sampling --- Defusing Neighborhood Explosion in GNNs | Accept (poster) | Summary: The paper proposes the LABOR sampling algorithm to address the NEP problem in GNN. LABOR combines layer sampling (LADIES), neighbor sampling (NS), and Poisson sampling. In contrast to prior works, LABOR calculates the sampling probabilities for importance sampling “iteratively” for each layer. Moreover, the Po... | Rebuttal 1:
Rebuttal: We would like to thank the reviewer for the valuable feedback.
For the question about why LADIES and PLADIES vertex efficiencies are higher than LABOR on reddit: Table 2 shows how many vertices and edges are
sampled in each layer. We are interested in the $|V^3|$ and $|E^2|$ columns as they give ... | Summary: This paper proposes a new sampling algorithm called LAyer-neighBOR sampling (LABOR). The authors show the variance of the estimator is controlled. Experiments show this method is much more efficient in sampling and faster in convergence.
Strengths: 1. The paper is well-written and clear.
2. Some analysis is ... | Rebuttal 1:
Rebuttal: We would like to thank the reviewer for the valuable feedback.
For the first point, since LABOR is a combination of NS and LADIES methods, it wouldn't be hard to extend any proof for their
convergence to LABOR. We think that such a theoretical analysis wouldn't be the most interesting part of th... | Summary: This paper improves the previous layer-wise importance sampling algorithm in GNN training by: 1) introduce poison sampling for debiasing; 2) it modifies the sampling importance by incorporating the probability of node-wise sampling, to make sure all nodes have similar neighbor nodes.
It shows that the varianc... | Rebuttal 1:
Rebuttal: We would like to thank the reviewer for the valuable feedback. We will be responding to the weaknesses the reviewer pointed out below:
1. FastGCN is superseded by LADIES because FastGCN doesn't even consider the recursive structure of sampling and the previous layer's sampled
vertices. A prelimi... | Summary: This paper proposes a novel neighbor sampling (NS) method, called layer-neighbor sampling (LABOR), for training GNNs on large-scale graphs. Unlike the representative NS method, which samples neighbors for each node from a uniform distribution, LABOR samples neighbors for each node using Poisson sampling from a... | Rebuttal 1:
Rebuttal: We would like to thank the reviewer for the valuable feedback.
When it comes to computational complexity of sampling, let us note that we implemented all parts of LABOR using CUDA, fully parallelized. LABOR-0
sampling runtime is very similar to or even faster than NS. When importance sampling is ... | Rebuttal 1:
Rebuttal: We would like to thank all the reviewers for their valuable feedback.
Some reviewers question why we kept our main experimental evaluation to NS, LADIES and LABOR. Firstly, Appendix A.8 has
some comparison against GraphSAINT in terms of time taken to converge to a given
accuracy target, and resu... | NeurIPS_2023_submissions_huggingface | 2,023 | Summary: This paper studies the problem of improving the efficiency of GNN training. It first points out that the node-based sampling methods that sample independently ad recursively for each node suffer the neighbor explosion problem (NEP), while the layer-based sampling methods that sample for the whole layer collect... | Rebuttal 1:
Rebuttal: We would like to thank the reviewer for the valuable feedback. For the camera-ready version, we will have one more page available. We can move the
main algorithm to the main paper as you suggest.
Appendix A.4 has a figure where the x-axis is the number of training iterations. Appendix A.8 has a s... | null | null | null | null | null | null |
Blockwise Parallel Transformers for Large Context Models | Accept (spotlight) | Summary: The paper extends the work in Flash Attention to apply block wise attention to not just the self attention block, but also to the feedback forward part of the transformer block. While doing so, the output of the self attention block is not stored for backprop, and recalculated. This leads to very efficient gai... | Rebuttal 1:
Rebuttal: We thank the reviewer for their feedback and for a positive assessment of the work. We found the reviewer’s questions and suggestions insightful, and we list our plans to incorporate them below. Please let us know if our answers address your questions.
> Q: In the RL task, the AT baseline is fo... | Summary: This paper proposed to merge the computation of the FFN layers in Transformer into the block-wise computation of self-attention. This eliminates the need to wait for the self-attention computation to finish before performing the feedforward step on the entire sequence, because FFN is entirely position-wise and... | Rebuttal 1:
Rebuttal: Thank you for the review. We found the reviewer’s questions and suggestions insightful, and we list our plans to incorporate them below. You raised a good question about how exactly does BPT help save memory. Please let us know if our answers address your questions.
> Q: If I understand correct... | Summary: The paper presents a novel variant of the Transformer model, the Blockwise Feedforward Transformer (BFT), which is designed to address the memory inefficiency issue in standard Transformer models. The authors propose a blockwise computation method that significantly reduces the memory footprint without comprom... | Rebuttal 1:
Rebuttal: We thank the reviewer for their feedback and for a positive assessment of the work. We found the reviewer’s questions and suggestions insightful, and we list our plans to incorporate them below. Please let us know if our answers address your questions.
> Q: The paper lacks a comprehensive analy... | Summary: "Blockwise Parallel Transformer for Large Models" presents a novel approach to handle the memory demands of Transformer models, mainly when dealing with long sequences or tasks involving multiple sequences or long-term dependencies. The authors propose a method called Blockwise Parallel Transformer (BPT) that ... | Rebuttal 1:
Rebuttal: We thank the reviewer for their feedback and for a positive assessment of the work. We found the reviewer’s questions and suggestions insightful, and we list our plans to incorporate them below. Please let us know if our answers address your questions.
> Q: The use of RL as a benchmark for test... | null | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Prediction and Control in Continual Reinforcement Learning | Accept (poster) | Summary: In this work, the author proposes a new continual reinforcement learning method, the PT-TD, which consists of two value networks, one learns the transient memories and the other learns the permanent memories. The author proves that in the semi-continual setting, the PT-TD algorithm will get better performance ... | Rebuttal 1:
Rebuttal: Thank you for your review and positive comments on our paper. Please find our answers to your questions below:
> 1. I think the formulation of the Theorem 8 is a little confusing. Maybe you can give out explicitly which network it is using (P or T).
We will update the statement in the main paper... | Summary: This paper proposes new algorithms for prediction and control in continual reinforcement learning. The new algorithms use permanent and temporary memories. The permanent memory captures a general value function for all the tasks, while the transient memory is used to adapt to the current tasks continually. The... | Rebuttal 1:
Rebuttal: Thank you for your thorough review and positive comments on our paper. Please find our answers to your questions below:
> I'm confused about Equation 5, specifically line 112. Does it mean the "gradient" or the "semi-gradient" of the transient memory?
This is a semi-gradient update rule, in the ... | Summary: The paper proposes a method to improve continual learning in reinforcement learning with non-stationary dynamics. It proposes to maintain two value functions: one that tracks information about the current task (transient) and one that maintains a general knowledge of the task distribution (permanent) in order ... | Rebuttal 1:
Rebuttal: Thank you for your review. We answer your questions below, please let us know if further clarification is needed.
> What do the authors mean when they say prior work “reflected the nature of that task” in line 41?
We meant that the method presented in Silver et al [1] leverages domain knowledge ... | Summary: This work proposes a reinforcement learning method for prediction and control. The value function in reinforcement learning is decomposed into two terms: a permanent value function, and a transient value function. The permanent function contains the long term knowledge of the interplay between the agent and th... | Rebuttal 1:
Rebuttal: Thank you for the review. We address the main points below.
> The settings for the experiments are very simple.
We designed the experiments in the paper to cover several aspects that we deemed important: environment non-stationarity, different function approximators (tabular, linear, non-linear)... | Rebuttal 1:
Rebuttal: Thank you for your time and useful suggestions for improving our paper. We appreciate the positive feedback on the ideas, their connections to complementary learning systems and the theoretical analysis. We address specific questions below and would be happy to provide any further clarifications. | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
ReContrast: Domain-Specific Anomaly Detection via Contrastive Reconstruction | Accept (poster) | Summary: The paper presents an approach, called ReContrast, for domain-specific anomaly detection and segmentation. The authors aim to overcome the transfer limitations of existing methods by addressing and mitigating the challenges associated with the use of pre-trained encoders from natural image domains. They propos... | Rebuttal 1:
Rebuttal: Thank you for your insightful comments and suggestions.
__1. The authors handle training instability by selectively changing the batch normalization (BN) mode of the encoder based on performance on specific datasets or categories. This raises concerns about the proposed method's reliability and g... | Summary: This paper proposes an epistemic method, namely ReContrast, for unsupervised anomaly detection (UAD). The proposed method is deeply based on UniAD and RD4AD, with some technical improvement to adapt to the data from different domains. Experiments are conducted on two popular industrial defect detection benchma... | Rebuttal 1:
Rebuttal: Thank you for all your valuable suggestion and concerns.
__1. (Weakness 1 & Question 1) The paper lacks novelty. The proposed method is deeply based on UniAD and RD4AD, with some technical improvement to adapt to the data from different domains. Unclear contributions compared to existing methods ... | Summary: This paper proposes a novel unsupervised anomaly detection (UAD) method called ReContrast. The method addresses the issue of poor transferability of pre-trained encoders from natural image domains to target UAD domains such as industrial inspection and medical imaging. ReContrast optimizes the entire network t... | Rebuttal 1:
Rebuttal: First and foremost, we are thrilled that you like our writing logic. We made a lot of considerations to use this exploration-like narrative instead of a conventional paper structure.
__1. In the ablation experiments of MVTec AD and APTOS, for example, the experiment on Global Cosine Similarity, t... | Summary: This work proposes an unsupervised anomaly detection method using contrastive learning method to reduce the biases between the pre-trained image domain and the target domain. In the paper, different designs of reconstruction networks are introduced and compared in terms of detection performance, model coverage... | Rebuttal 1:
Rebuttal: We would like to thank you for your constructive comments with great insights.
__1. During both training and inference, the feature loss is calculated using six feature pairs from two encoders. The utilization of two encoders in the methods leads to an increase in parameters and inference time. ... | Rebuttal 1:
Rebuttal: __Note: *Revised Figures are in the one-page PDF. Please check "pdf" below.*__
__*The following area is only the extension of response to Reviewer 4 (e2VQ) due to page limitation.*__
__Reply to Review 4 (e2VQ). 4. (Question 4) They are suggested to follow the unified setting of UniAD [9] (all... | NeurIPS_2023_submissions_huggingface | 2,023 | Summary: This paper introduces a feature reconstruction model, namely ReContrast, for anomaly detection. By incorporating a set of techniques including feature global similarity, stop gradient, contrastive pair optimizations, the authors shows good performance of ReContrast on both industrial and medical datasets.
Str... | Rebuttal 1:
Rebuttal: We would like to thank you for your constructive comments with great insights. The replies to specific remarks are as followed:
__1. The paper presentation should be improved. The current version is in a lab report format, outlining the observed issues and employing trial-and-error strategies to ... | null | null | null | null | null | null |
Faster Relative Entropy Coding with Greedy Rejection Coding | Accept (poster) | Summary: The paper presents a coding scheme for greedy rejection sampling by Harsha, which builds upon the existing A* coding framework. The trick here is to add a partition operation between each reject/accept steps, which allows a more efficient scheme under certain assumptions. Overall, the work contributes meaningf... | Rebuttal 1:
Rebuttal: We thank TPKF for their time reviewing our paper and their feedback. We were pleased to hear TPFK thought our work was a meaningful contribution and our paper well-written and relatively easy to follow. TPKF raised a few points which we would like to address below.
__Sketch proof:__ We agree the ... | Summary: The authors propose a greedy rejection coding based algorithm for performing compression. The authors claim that GRC an ERC based compression method is the first method which does not make unrealistic assumptions, is guaranteed to finish in fixed amount of time. They generalize the algorithm to arbitrary parti... | Rebuttal 1:
Rebuttal: We would like to thank reviewer BHNG for their time to review our work and their feedback. We were happy to hear that the reviewer thought our paper was well-written, discussed relevant challenges well, and was satisfied by the correctness of our proofs. The only point of weakness they mentioned w... | Summary: This paper presents a new algorithm, Greedy Rejection Coding (GRC), for Relative Entropy Coding (REC). REC is an alternative to entropy coding and quantization, which naturally interfaces with continuous variables and has applications in data compression. Current REC algorithms have faced challenges like slow ... | Rebuttal 1:
Rebuttal: We thank reviewer ULAV for their time reviewing our paper. We are happy they found our theoretical and experimental results as strong points. Reviewer ULAV also raised some concerns, which we address below.
__Dependency on CDF:__ ULAV commented that since GRC requires access to the CDF of the tar... | Summary: This paper proposes new Greedy Rejection Coding methods that generalize to continuous probability space and arbitrary partitioning schemes. The new methods produce correct samples with the expected code length of $O(D_{KL}[Q||P])$ where $Q$ is the target distribution and $P$ is the proposal distribution. The p... | Rebuttal 1:
Rebuttal: We thank reviewer T3AN for their time reviewing our paper and valuable feedback. We were happy to hear they found our theoretical and experimental results as strong points of this work. We were also pleased that they thought our exposition, discussion of related work and limitations, and vizualisa... | Rebuttal 1:
Rebuttal: Dear Reviewers and ACs,
We would like to thank all four reviewers for the time they took reviewing our paper and for their valuable feedback. We found that the feedback from all four reviewers was generally positive. Most reviewers listed our theoretical contributions and the experimental validat... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Trading-off price for data quality to achieve fair online allocation | Accept (poster) | Summary: Existing works in the fair online allocation space assume that the decision-maker has access to protected attributes. In contrast, this paper studies the choice of a data source as a multi-armed bandit problem in conjunction with the primary allocation problem. This paper introduces a computationally efficient... | Rebuttal 1:
Rebuttal: Thank you for this review!
Regarding code and experimental results, we have presented a simple numerical toy example in the appendix F.2. Additionally, we made additional numerical simulations in a simple case showing how the optimal solution and value vary as we vary some parameters. We refer th... | Summary: The authors formulate the online fair allocation problem, where a decision maker may choose an information source $k_t$ based on public context $z_t$ and generate an allocation decision $x_t$ with a reward $1$ if $x_t=1$. The fairness penalty is some convex function that is minimized when the decision $x_t$ is... | Rebuttal 1:
Rebuttal: Thank you for this review!
**Concerning the novelty and the technical difficulties**:
As discussed in the abstract (line 7-8) and in the introduction (line 57-58), the main novelty of this work is to consider jointly the problem of source selection and of long term fairness penalty. Indeed, those... | Summary: This paper studies an online allocation problem with long-term fairness constraints where, unlike prior works, the protected attributes of online users (e.g., gender, ethnicity) are not observed, but they can be purchased from $K$ different sources for some additional cost. The goal is to maximize the overall ... | Rebuttal 1:
Rebuttal: Thank you for this review!
**Relationship between various quantities and usefulness of $z_t$**: The quantities $(z_t,a_t,u_t,c_{t1},...,c_{tK})$ are random variables non necessarily independent, which are simultaneously drawn from some distribution, but which are only partially observed by the ad... | Summary: This paper studies the problem of online resource allocation with long-term fairness constraints. Different from mainstream literature, the authors assume that the fairness context is not known to the player in advance. However, the player does have the option of paying to buy fairness context information from... | Rebuttal 1:
Rebuttal: Thank you for this review!
– **Fairness Penalty**: We do mention that we consider a fairness penalty and not a hard constraint in the abstract and in the model (e.g., lines 1,5,10 and 119), but we will make it clearer by also mentioning it in the introduction. We feel that the confusion may have ... | Rebuttal 1:
Rebuttal: In response to reviewer 2Yz9 we have conducted an additional numerical experiment for a simple example on the impact of varying fairness and information cost over the allocations and other meaningful quantities. The results are presented in the figure of the one-page PDF attached to this global re... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Reinforcement Learning with Fast and Forgetful Memory | Accept (poster) | Summary: The paper proposes a new memory model for reinforcement learning agents, targeting POMDP setting. The memory (FFM) aggerates past data via an exponentially weighted summation operation where the weights are learnable parameters. The memory is used to compute the Markov state through sophisticated gating mechan... | Rebuttal 1:
Rebuttal: Thank you for the detailed review. Our primary concern in this paper was to design a __more efficient RNN for model-free RL__, with the idea that anyone with an existing recurrent RL implementation (e.g. CleanRL's recurrent PPO) could easily replace their LSTM/GRU with FFM, and train a comparable ... | Summary: This work introduces a new recurrent neural network layer called Fast and Forgetful Memory (FFM), which, differently from traditional architectures such as GRUs and LSTMs, FFM is explicitly designed for deep reinforcement learning
Strengths: The authors present a novel and promising memory layer for reinforce... | Rebuttal 1:
Rebuttal: Thank you for taking time to review, we hope you enjoyed the paper.
Thank you for the corrections, we will fix these at once. The footnote was originally in the main text, but it took up a lot of space and we were worried it might put some readers to sleep. We will see where we can fit it back in... | Summary: The paper introduces Fast and Forgetful Memory (FFM), a memory model designed for reinforcement learning. FFM is based on theories of composite memory and contextual drift from computational psychology. It consists of two main components: an aggregator and a cell. The aggregator computes complex-valued summari... | Rebuttal 1:
Rebuttal: Thank you for taking the time to review our paper, and thank you for your interest. We hope we are understanding your first question correctly: we only ever store the __sum__ of time-contextualized traces. We assume psychologists like using distinct traces like Eq. 6 because they are conceptually ... | Summary: The paper aims to propose an algorithm-agnostic memory model specifically designed for reinforcement learning (RL). The proposed neural network (NN) model is claimed to offer advantages over existing sequence-based memory models like Recurrent Neural Networks (RNNs) and can be seamlessly substituted for RNNs. ... | Rebuttal 1:
Rebuttal: Thank you for taking the time to review our paper. We believe a reject rating for the issues you raised is unduly harsh, and we believe there might also be some misunderstanding regarding the scope of our paper. The goal of this paper is to design an efficient replacement for RNNs in model-free RL... | Rebuttal 1:
Rebuttal: Thank you to all the reviewers and ACs who spent time reviewing this paper. We are happy to hear that our proposed method is "novel and promising", as well as "inuitive...well explained, and easy to follow", that "the empirical analysis is extensive" and "The contributions of this paper are signif... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
On the Constrained Time-Series Generation Problem | Accept (poster) | Summary: The authors propose a set of methods for generating synthetic time series, with or without constraints. They defined the different types of constraints possible, including soft and hard constraints. They developed a COP approach for generating time series from a seed, then presented ways of adding constraints... | Rebuttal 1:
Rebuttal:
We thank the reviewer for the positive and constructive feedback. We are happy to add more discussion on the type of constraints and time-series we currently support, and to extend our work in the future, following the reviewer suggestion.
**Q.1** Currently our work cannot cover constraints th... | Summary: The paper proposes a model for generating synthetic time series data given user-specified constraints. The constraints can be either soft (e.g. following trend) or hard ((in)equalility). The first proposed method is framed as a Constrained Optimization Problem (COP), it inputs a seed time series which is eithe... | Rebuttal 1:
Rebuttal: We thank the reviewer for the positive and constructive feedback. We are happy to revise and improve the paper according to the reviewer's comments, and to better motivate our work and its novelty.
## Weaknesses
**Q.1**
Our paper presents a set of methods for incorporating constraints during t... | Summary: This paper proposes multiple approaches, including a Constrained Optimization Problem (COP) method and a conditional diffusion model called DiffTime, to cope with the problem of generating constrained time series data that adheres to statistical properties and specific user-defined constraints. It also provide... | Rebuttal 1:
Rebuttal: We appreciate the constructive suggestions provided by the reviewer, and we are happy to revise and improve our paper accordingly, incorporating the writing and presentation suggestions pointed out.
## Weaknesses
**Q1.**
In particular, we thank the reviewer for pointing out unclear terms and d... | Summary: The authors propose multiple methods to solve the constrained time-series generation problem: Given a time-series and a set of (hard or soft, local or global) constraints, the goal is to generate a time-series which is realistic, adheres to the input constraints, yet sufficiently different from the original in... | Rebuttal 1:
Rebuttal: We thank the reviewer for the positive feedback, and for pointing out an aspect that was not sufficiently discussed. We agree that discussing the limitations of our work will enhance the overall quality of our paper and its impact on the academic community. We will add a more extensive discussion ... | Rebuttal 1:
Rebuttal: We would like to thank all the Reviewers for their constructive and valuable comments. We revised our work according to the reviewer's comments, and new results are included in this rebuttal, and attached PDF. In particular:
$\bullet$ We included 2 new state-of-art benchmarks for time-series gen... | NeurIPS_2023_submissions_huggingface | 2,023 | Summary: In this paper, the authors investigate the problem of time-series generation under different types of conditions. Since the existing methods usually require re-training when the conditions should be changed and the rejection sampling might require heavy computation, the authors use the conditional diffusion mo... | Rebuttal 1:
Rebuttal: We thank the reviewer for pointing out aspects that were not clearly explained in the paper.
**Q1/3** Our diffusion models use a Markov chain to convert a simple known distribution (i.e., a Gaussian) into the target (data) distribution using a diffusion process[1], without any strong assumptions ... | null | null | null | null | null | null |
Conformal Prediction for Time Series with Modern Hopfield Networks | Accept (poster) | Summary: This paper proposes a novel method for estimating uncertainty in prediction of time series. Modern Hopfield Networks is used as a method of measuring uncertainty based on the existence of similar historical data. This paper presents its construction method and shows that it works for multiple data and forecast... | Rebuttal 1:
Rebuttal: > This paper consists only of a method proposal and an experimental evaluation, and does not include theoretical effects. In this structure, is an evaluation on four data sets sufficient to demonstrate the effectiveness of the proposed method?
Besides a brief theoretical motivation in the main pa... | Summary: The work presents HopCPT, a conformal prediction algorithm for Modern Hopfield Networks in the setting of time series forecasting. The work uses a novel (attention-like) weighted quantile approach to leverage similar time steps from the past, improving both coverage and sharpness compared to SOTA CP methods. T... | Rebuttal 1:
Rebuttal: > From the paper alone, it's difficult to tell what exactly are the theoretical guarantees of your algorithm. While the authors did refer to Appendix B, I think it is important to briefly describe the assumptions and state the main bound (B.8 and B.9) of your theoretical analysis in the paper.
Th... | Summary: This paper deals with providing conformal prediction (CP) sets for the time series prediction problem or, more generally, the cases where the exchangeability assumption is violated. Similar to the prior work of Foygel Barber et al. (2022), they also try to assign weights to the points in the calibration set wi... | Rebuttal 1:
Rebuttal: > The contribution of this paper seems to be relatively marginal. Indeed, the authors introduced a new way of weighting the non-conformity scores (i.e., reforming the distribution of non-conformity scores) on top of the NexSCP framework proposed by Foygel Barber et al. (2022) while adopting the st... | Summary: The authors propose a conformal prediction approach for black box time series forecasting models. The method is based on learning to predict the model’s errors using a small neural network structured similarly to the softmax computation in an attention mechanism. The prediction interval of the proposed method ... | Rebuttal 1:
Rebuttal: > Without a background on Hopfield networks, the method seems an application of the attention idea to conformal prediction. It would be helpful to lay out the similarities and differences between MHNs and the attention mechanism.
We agree that the connection between MHN and attention was not clea... | Rebuttal 1:
Rebuttal: We would like to thank all reviewers for their thorough reviews, helpful suggestions, and constructive comments. To summarize, we made the following main changes to the manuscript to follow the suggestions and comments of the reviewers:
- We added a discussion paragraph that elaborates on the main... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Aligning Optimization Trajectories with Diffusion Models for Constrained Design Generation | Accept (poster) | Summary: This work studies the use of diffusion models for constrained topology design, a task where physics-based optimization methods still outperform deep learning approaches. The authors introduce diffusion optimization models (DOMs), a diffusion-based approach to generate topology candidates rapidly and efficientl... | Rebuttal 1:
Rebuttal: ## **Authors**: *We would like to clarify that DOM is a well-defined conditional diffusion model. TA is used only during training. At inference time, we sample DOM using standard ancestral sampling and the model does not have access to any intermediate optimization step. See the attached PDF for e... | Summary: The paper presents a diffusion-based method for topology optimization. The method seems to base on TopoDiff, but uses trajectory alignment instead of guidance to produce plausible results even under few-step sampling.
Strengths: Diffusion models seem like a natural fit for solving inverse problems, but it's s... | Rebuttal 1:
Rebuttal: ## **Authors**: *Topology Optimization is a hard problem without analogy in the machine learning domain. Our goal is not to replace classical solvers but to augment them.*
>Strengths:
Diffusion models seem like a natural fit for solving inverse problems, but it's still open to what is the most ef... | Summary: The paper proposes a model that combines engineering optimization with conditional diffusion generative models for constrained design generation. The main idea is to match the sampling trajectory of the diffusion model with the trajectory derived from the traditional physics-based model.
Strengths: Strength:
... | Rebuttal 1:
Rebuttal: ## **Authors**: *We emphasize that trajectory alignment is performed only during training and that DOM does not require extra steps of optimization*.
> Strength: The idea of using a physics-based trajectory to guide the sampling process of the diffusion model is very interesting.
**Authors**: th... | Summary: The paper introduces Diffusion Optimization Models (DOM) to generate feasible and high-performance designs in structural topology optimization. DOM is based on diffusion model, and a key design component is trajectory alignment method, in which trajectory in diffusion-based optimization is biased towards given... | Rebuttal 1:
Rebuttal: ## **Authors**: *We emphasize that we do not use post-processing for any results or visualizations in the paper except to reduce floating material in Table 4*.
>Strengths:
The exposition for the motivation is written crisply and is generally easy to follow. Additional details on backgrounds for D... | Rebuttal 1:
Rebuttal: We would like to thank the reviewers for the feedback. After reading the reviews, we feel to clarify three raised issues:
- Q1) Is DOM a generative model?
#### **Yes, DOM is a conditional diffusion model trained with the standard epsilon parameterization and a regularization term (TA). The atta... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
DAW: Exploring the Better Weighting Function for Semi-supervised Semantic Segmentation | Accept (poster) | Summary: This paper studies semi-supervised semantic segmentation, where the key challenge is how to select high-quality pseudo-labels. This paper systematically analyzes the trade-off between correct unreliable predictions and wrong confident predictions and theoretically finds the optimal weighting function. Interest... | Rebuttal 1:
Rebuttal: Thanks for taking the time to share your comments and suggestions in the review assessment, as well as for acknowledging the good writing, sufficient supplementary material, novel idea and convincing experiment result.
We see that your main concern is how to better estimate the distribution.
All... | Summary: In this paper, the authors thoroughly analyze the pixel-level pseudo label selection criteria and the trade-off associated with weighting functions in semi-supervised semantic segmentation. The authors propose a novel method called distribution-aware weighting (DAW) based on their analysis and confidence distr... | Rebuttal 1:
Rebuttal: Thanks for taking the time to share your comments and suggestions in the review assessment, as well as for acknowledging the well-thought-out design, comprehensive analysis, sophisticated mathematical derivation and valuable insight.
All your comments are addressed point by point in the following... | Summary: This paper discusses the challenge of utilizing unlabeled data to improve the generalization performance of semantic segmentation models. Existing methods use weighting functions to select pseudo-labels, but face a trade-off between accuracy and utilization. The proposed method, DAW, models the confidence dist... | Rebuttal 1:
Rebuttal: Thanks for taking the time to share your comments and suggestions in the review assessment, as well as for acknowledging the novelty and well-done presentation.
We see that your main concern is about the effectiveness of DAW.
All your comments are addressed point by point in the following.
**Q... | Summary: This paper presents a systematic analysis of the trade-off involved in previous methods that deal with pseudo-labels. The authors define the trade-off between inaccurate yet utilized pseudo-labels and correct yet discarded pseudo-labels by modeling the confidence distribution of both types of pseudo-labels. Th... | Rebuttal 1:
Rebuttal: Thanks for taking the time to share your comments and suggestions in the review assessment, as well as for acknowledging the thorough analysis, rigorous experiments, and well-structured presentation.
We see that your main concern is about the imprecise statement.
All your comments are addressed ... | Rebuttal 1:
Rebuttal: **To Reviewer SG4E -- Rebuttal Part2:**
**Q3:** From the results of Tables 2 and 3, it can be seen that the improvement in the final performance in most scenarios is relatively minimal compared to previous works.
**A3:**
Semi-supervised semantic segmentation is a very challenging task, and DAW ... | NeurIPS_2023_submissions_huggingface | 2,023 | Summary: This paper introduces DAW, a method for semi-supervised semantic segmentation that addresses the trade-off between false-positive and false-negative pseudo-labels. DAW utilizes Gaussian functions to model confidence distributions and derives an optimal weighting function based on their inter-section. It also i... | Rebuttal 1:
Rebuttal: Thanks for taking the time to share your comments and suggestions in the review assessment, as well as for acknowledging the novelty and adequate technical approach.
We see that your main concerns are more experiments on a larger dataset and further explanation in Table 1.
All your comments are ... | null | null | null | null | null | null |
SOC: Semantic-Assisted Object Cluster for Referring Video Object Segmentation | Accept (poster) | Summary: This paper proposes a novel framework that can effectively exploit video-level visual-linguistic alignment to improve the performance of RVOS, filling a gap in existing methods. The extensive experiments on three datasets demonstrate the effectiveness of SOC.
Strengths: The proposed method outperforms existi... | Rebuttal 1:
Rebuttal: We sincerely thank the Reviewer 9DMq for his/her comments and the appreciation of our work. Our response to the Reviewer 9DMq is as follows:
*Questions:* Thanks for your question. Sorry we are not sure whether you are referring to the training stage or the inference. Therefore, we explain both of... | Summary: This work studies the referring video object segmentation problem and proposes the semantic-assisted object cluster approach. It aggregates video content and textual guidance for unified temporal modeling and cross-modal alignment.
Strengths: 1) It performs object aggregation and promotes visual-linguistic a... | Rebuttal 1:
Rebuttal: We sincerely thank the Reviewer d9RC for his/her comments and the appreciation of our work. To the concerns expressed in Weaknesses and Questions, our response is as follows:
*Weakness-1:* Thanks for the question. MMF is just a component of our framework. It is not our contribution, and it is not... | Summary: This research paper focuses on referring video object segmentation. It combines temporal frame features and language features using a cross attention design. Additionally, it extracts object features from each frame and applies contrastive learning with language representation. The paper presents various desig... | Rebuttal 1:
Rebuttal: We sincerely thank the Reviewer ihRG for his/her review time and valuable comments. To the concerns expressed in Weaknesses and Questions , our response is as follows:
*Weakness-1*: Thanks for your suggestion. The main contribution and motivation of this paper is to understand video content from ... | Summary: This article introduces the Semantic-assisted Object Cluster model, which is used to solve the problem of visual language alignment between videos and texts. The model uses a semantic integration module and visual language contrastive learning to improve the understanding of video content and has achieved sign... | Rebuttal 1:
Rebuttal: We sincerely thank the Reviewer NrNc for his/her comments and the appreciation of our work. To the concerns expressed in Weaknesses and Questions, our response is as follows:
*Weakness-1:* Good question! Inaccurate or incomplete descriptions do affect the effectiveness of Visual-Linguistic Contra... | Rebuttal 1:
Rebuttal: Dear Reviewers and ACs:
We sincerely appreciate all the reviewers. They give positive and high-quality comments on our paper with a lot of constructive feedback. We are glad to be recognized that our motivation is clear and straightforward (R2, R3, R6), the paper is well-written and easy-to-follo... | NeurIPS_2023_submissions_huggingface | 2,023 | Summary: This paper mainly studies referring video object segmentation (RVOS) which is a fundermental problem in vision-and-language applications. Compared with previous works which pay little attention to temporally modeling relationships between different frames in the entire video clip, the authors proposed a framew... | Rebuttal 1:
Rebuttal:
We sincerely thank the Reviewer 8Kjg for his/her comments and the appreciation of our work. To the concerns expressed in Weaknesses and Questions, our response is as follows:
*Weakness-1:* (1) Thanks for your valuable question. Table 7 in the manuscript shows the ablation of the structure of th... | Summary: In this paper, the author presents a framework called SOC, and a visual-linguistic contrastive loss for RVOS to unify temporal modeling and cross-modal alignment.
The performance is good according to the leaderboard (https://paperswithcode.com/sota/referring-video-object-segmentation-on-refer)
The extra module... | Rebuttal 1:
Rebuttal: We sincerely thank the Reviewer iFku for his/her comments and the appreciation of our work. To the concerns expressed by Reviewer iFku in Weaknesses and Questions , our response is as follows:
*Weakness-1:* Thanks for reminding us that we should make a clear demonstration of model parameters. In ... | null | null | null | null |
Fairness Continual Learning Approach to Semantic Scene Understanding in Open-World Environments | Accept (poster) | Summary: This paper proposes a fairness continual learning approach in semantic segmentation. Derived from the therotical analysis of the effect of class distribution, a new learning paradigm that is consisted of three main objectives: the task-specific objective to handle the catastrophic forgetting problem, the fairn... | Rebuttal 1:
Rebuttal: ## Response to Reviewer 9fXu
Dear Reviewer 9fXu,
We thank the reviewer for your careful read and valuable feedback.
We are pleased you encourage that our theoretical analysis is novel and our approach achieves strong performance. We appreciate your constructive comments and would like to a... | Summary: This paper proposes a Fairness Continual Learning approach to address the continual semantic segmentation problem. Specifically, the fairness continual learning framework is proposed for fairness objective, while a Prototypical Contrastive Clustering loss for reducing catastrophic forgetting. When compared to ... | Rebuttal 1:
Rebuttal: ## Response to Reviewer 2Nqo
Dear Reviewer 2Nqo,
We would like to thank the reviewer for your careful read and valuable feedback.
We are very happy you encourage that our proposed losses are well-formulated and the ablation studies have illustrated the effectiveness of our approach. We apprec... | Summary: The paper explores the continual semantic segmentation problem and highlights the lack of fairness in existing methods. It proposes a novel fairness objective based on class distributions to address the bias on the majority group classes. The paper proposes a novel prototype contrastive clustering loss to addr... | Rebuttal 1:
Rebuttal: ## Response to Reviewer ReY5
Dear Reviewer ReY5,
We would like to thank the reviewer for your careful read and valuable feedback.
We appreciate that you think our paper is well-motivated and well-written, the proposed approach is well-formulated and novel, and our theoretical analysis of t... | Summary: The paper focuses on the semantic segmentation problem in the context of continual learning. In the past, the emphasis was mainly on addressing catastrophic forgetting and background shift issues. However, this paper goes beyond that and investigates the balance problem between different object categories. It ... | Rebuttal 1:
Rebuttal: ## Response to Reviewer jVAT
Dear Reviewer jVAT,
We thank the reviewer for your careful read and valuable feedback.
We are pleased you think of our proposed problem and approach as our strengths and significant contributions. We appreciate your constructive comments and would like to address the... | Rebuttal 1:
Rebuttal: ## Global Response
We thank all the reviewers for their careful read and valuable feedback. Reviewer 9fXu and Reviewer jVAT ***weakly accept*** our paper; Reviewer ReY5 considers our work a ***borderline accept***; and Reviewer 2Nqo considers ***a borderline reject***. Our implementation will be... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Fast Attention Requires Bounded Entries | Accept (poster) | Summary: This paper shows that whether or not there exists a fast algorithm for approximately computing the attention matrix ($A \in \mathbb{R}^{n \times d}$, where $d = O(\log n)$ depends on the bound of the entry value $B$
1. Particularly, the authors show that when $B = o(\sqrt{\log n)}$, their proposed algorithm c... | Rebuttal 1:
Rebuttal: We thank the reviewer for the thoughtful review.
Minor typo p5: Yes, it is a typo and the max operator is missing. Thank you for pointing it out. We will fix it in the updated version.
Question: The construction of U_1 and U_2 in Lemma 3.2 is actually very simple and practical. At a high level:... | Summary: The paper is a purely theoretical paper discussing improving the computational time required to solve the attention computation problem which is known to be quadratic in time. The idea hinges upon the use of the boundness of the entries of the matrices involved in such a problem.
The authors discuss mainly tw... | Rebuttal 1:
Rebuttal: We thank the reviewer for the thoughtful review. We really appreciate the reviewer for pointing out these typos, we will fix them in the next version. (We remark that none of the typos impact the correctness of the paper.) | Summary: In this work, the author studied the problem of how to efficiently compute or approximate the attention matrix. The major contribution can be summarized as:
1. they found for the case when $d = O(\log n)$ and entries of the input matrices are uniformly bounded by $o(\sqrt{\log n})$, then an algorithm that ap... | Rebuttal 1:
Rebuttal: We thank the reviewer for the thoughtful review.
1. We discuss our technical contributions and how they compare with prior work in sections 1.1 and 1.2 of our paper. Perhaps the biggest technical insight is that a number of tools from the literature on KDE, which were not previously known to be a... | Summary: The attention mechanism, whose runtime complexity is approximately quadratic in the sequence length $n$, is the main bottleneck to LLM inference. It is thus quite reasonable to study whether this can be *approximated* with sufficient accuracy in subquadratic time. In particular, for this paper the authors defi... | Rebuttal 1:
Rebuttal: We thank the reviewer for the thoughtful review. We really appreciate the reviewer for pointing out these typos, we will fix them in the next version. (We remark that none of the typos impact the correctness of the paper.)
---
Rebuttal Comment 1.1:
Title: Thanks for the responses
Comment: I have... | Rebuttal 1:
Rebuttal: We want to thank the reviewers for their thoughtful and detailed reviews.
We want to briefly expand on the connection between our result and practical papers [ZBIW19] and [KVPF20] which multiple reviewers asked about. At a high level, these two papers showed that techniques for enforcing and mani... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
One-Pass Distribution Sketch for Measuring Data Heterogeneity in Federated Learning | Accept (poster) | Summary: The paper proposes to use a locally sensitive hashing based sketch to measure the similarity between two datasets in a privacy-preserving manner. This sketch can be computed efficiently and differential privacy can be guaranteed via the Laplace mechanism. The paper concludes with presenting two potential appli... | Rebuttal 1:
Rebuttal:
We thank the reviewer for the comments and suggestions to improve our paper! Please see the following clarifications.
**Q: Would the definition 3.1 of TVD not be more appropriately placed at the beginning of Sec 3.2.2?**
We introduce the TVD early to motivate the theory and provide a high-level... | Summary: The presence of **data heterogeneity** poses a critical challenge in **federated learning** (FL), as it leads to slower and noisier collaborative learning. To address this issue and improve the learning process with faster convergence, it is here proposed a more efficient client selection strategy based on dis... | Rebuttal 1:
Rebuttal: We thank the reviewer for the comments and suggestions to improve our paper! Please see the following clarifications.
**Reference**
We thank the reviewer for a great summary of the literature. We have updated the paper by including the mentioned reference in the Introduction and Related Work sec... | Summary: This paper proposed to use sketching for measuring heterogeneity in user data distributions in federated learning settings. The sketching is based on locality sensitive hashing and the authors have shown that in theory the measurement from the count sketch converges to the true kernel density estimation. The a... | Rebuttal 1:
Rebuttal: We thank the reviewer for the comments and suggestions to improve our paper! Please see the following clarifications.
**Choices of parameters**
We suggest choosing $B$ and $R$ as follows: $R$ should be chosen proportional to $O(\sqrt{n})$, where $n$ where
is the maximum possible client data si... | Summary: The authors propose a private strategy for measuring data heterogeneity using a RACE matrix. The RACE matrix essentially measures the normalized frequency of data points being assigned to particular LSH codes. These LSH count matrices are sent up to a server where it is applied into two different regimes:\\
1.... | Rebuttal 1:
Rebuttal: We are glad that you find the method simple and promising, and we have done our best to address your questions carefully. We hope that you might consider raising the score in light of our response.
**Q1: Our contribution with respect to the RACE matrix**
The RACE matrix, indeed, proves to be an... | Rebuttal 1:
Rebuttal: We thank reviewers [R1(Jfur), R2(QJNz), R3(iXQ3), R4(c2Su)] for their thoughtful and highly supportive feedback! We were glad that the reviewers found the problem significant and interesting [R2 R3], the observations and theoretical analysis insightful and valuable [R3, R4], the methods novel and ... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
First Order Methods with Markovian Noise: from Acceleration to Variational Inequalities | Accept (poster) | Summary: The authors study first order stochastic algorithms under a Markovian noise: the noise is no longer i.i.d. as for vanilla SGD, but follows a Markov chain, of finite mixing time $\tau$. The proposed algorithms are SGD-like algorithms, with batchsizes that can vary. In particular, they use a randomized batchsize... | Rebuttal 1:
Rebuttal: We thank Reviewer __8XQi__ for the work and for the positive feedback! Next, we answer the issues raised.
__W1__ and __Q1__(noise):
Thank you! The noise assumptions are indeed an important part that needs to be covered. We analyze in detail the assumptions on noise from [28]. In Section 5.1, th... | Summary: The authors design and analyze first-order methods for stochastic minimization and stochastic variational inequalities (VIs), with Markovian noise. Typically in stochastic minimization, the target function is only accessible at point x, defined as $f(x)$, and its gradient $\nabla f(x)$, through the sampling of... | Rebuttal 1:
Rebuttal: We thank the Reviewer __RVr5__ for the work and for the positive feedback! Next, we answer the issues raised.
__Q1__(examples of Markovian noise):
This is a very interesting and important question. Next, we give a number of examples which we will try to fit into the revised paper:
_Coordinate.... | Summary: This paper presents an optimization study focusing on stochastic optimization with Markovian noise as opposed to i.i.d. noise. The authors explore three distinct scenarios: strongly convex, non-convex, and Variational Inequalities (VI). In comparison to previous research, they relax the boundedness assumptions... | Rebuttal 1:
Rebuttal: We thank the Reviewer __9UP4__ for the work and for the positive feedback! Next, we answer the issues raised.
__W1__(mixing time):
We agree that the demand of knowing mixing time in advance is a significant limitation of the method. Estimating mixing time $\tau$ from a single trajectory of the r... | Summary: This paper studies the convergence of first-order methods with Markovian noise for solving stochastic optimization problems and variational inequalities. The authors establish finite-time complexity bounds which scale linearly with the mixing time of the underlying Markov chain.
Strengths: This paper analyze... | Rebuttal 1:
Rebuttal: We thank Reviewer __11kd__ for the work and for the positive feedback! Next, we answer the issues raised.
__W1__ and __Q1__(experiments):
Unfortunately in a week to respond, we couldn't make the final version of the experiments, but we promise to add a set of experiments to vary the theory in th... | Rebuttal 1:
Rebuttal: Reviewer __9UP4__ requested to make the analysis with stepsize, which is decreasing and independent of a large set of unknown parameters. In part because of the size limitations of a personal reply to Reviewer, we put this in a general reply.
Here we give a sketch of how to make a decreasing ste... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
SE(3) Equivariant Augmented Coupling Flows | Accept (spotlight) | Summary: The manuscript proposes a normalizing flow architecture that is i.) equivariant wrt to the euclidian group SE(3) and permutation group S_n and ii.) is based on coupling blocks. Existing work on this specific equivariance has relied on continuous normalizing flows which rely on a Neural ODE instead of a couplin... | Rebuttal 1:
Rebuttal: Thank you for your review - we appreciated that you find our work relevant to your research! Below we have responded to your feedback.
## Similarity to Klein et al
We assume by Klein et al you are referring to https://arxiv.org/pdf/2302.01170.pdf.
We are aware of this work and indeed it was an i... | Summary: The paper proposes a method of producing a special Euclidean group - SE(3) - equivariant method of training coupling flows by augmenting the input space then conducting standard normalizing flow transformations in a rotationally invariant space by means of an invertible transformation. This is applicable parti... | Rebuttal 1:
Rebuttal: Thanks for your feedback! Below we have responded to your comments.
## Diffusion on Alanine Dipeptide
We agree that this would be a good additional baseline. Hence, we will add it to the camera ready version of the paper.
## Speedup of CNF
The reviewer is correct that the large relative speedup... | Summary: This paper proposes a novel method for developing an SE(3) equivariant coupling flow that works on the Cartesian coordinates of atoms, crucially enabling quick sampling and density evaluation. The standard coupling architecture is unable to operate on the Cartesian coordinates with the Special Euclidean group ... | Rebuttal 1:
Rebuttal: Thanks for your review! Below we have responded to your feedback and questions.
## Proof of equivarance
Indeed in our code we do test equivariance of our flow by testing that $g \cdot F(x, a) = F(g \cdot x, g \cdot a)$ (and that the log det is invariant).
We do not think that our notation for ... | Summary: This paper tackles the problem of sampling molecule configurations using deep generative models. The learned distribution should be permutation invariant (in terms of permutations of the particles in $\mathbb{R}^3$) and SE(3) invariant (i.e., invariant to rotations and translations applied to the particles' ca... | Rebuttal 1:
Rebuttal: Thank you for your review! We have addressed the issues you highlighted below.
## Clarity of methods section
We have updated the description of our method to make it more clear. Additionally [Reviewer goAi](https://openreview.net/forum?id=KKxO6wwx8p¬eId=fraAiN4VoH) found a typo in one of the ... | Rebuttal 1:
Rebuttal: # General response
Thank you for the reviews - these have been very helpful and we have incorporated the feedback into the paper. Below we highlight key changes to the paper: **(1)** Added new results studying the data-efficiency of the equivariant vs non-equivariant variants of the flow. **(2)** ... | NeurIPS_2023_submissions_huggingface | 2,023 | Summary: This paper constructs equivariant normalizing flows that are discrete to sidestep speed issues with equivariant continuous normalizing flows. The method does this by augmenting the input with some an additional variable to avoid representation issues.
Strengths: * The method is a reasonable extension of stand... | Rebuttal 1:
Rebuttal: Thank you for your feedback! We have addressed each of the weaknesses you highlighted below. We hope that this helps clarify the paper.
## Usage of $\gamma$:
You are correct; there is a typo in the text - thanks for pointing this out. The property of interest is $\gamma_{g \cdot r} = \gamma_{ r} ... | null | null | null | null | null | null |
SEGA: Instructing Text-to-Image Models using Semantic Guidance | Accept (poster) | Summary: The paper introduces Semantic Guidance, a method that enables user control and interaction with text-to-image diffusion models to generate images that align with their intended semantics. The authors highlight the challenge that small changes in the input prompt can result in vastly different images. To addres... | Rebuttal 1:
Rebuttal: We thank the reviewer for their valuable feedback and suggestions. We will address each of the stated weaknesses and questions individually.
### W1: Similarity to previous work
We respectfully disagree with the reviewer that our work is almost identical to Composable Diffusion. We agree that the... | Summary: This paper proposes a method for attribute editing of the images generated with Text-to-Image Diffusion models (T2I-DM). This is done in a manner similar to CFG, where extra edit prompts are passed alongside the text prompt to generate the image. Noise predicted from the edit prompt in the denoising process gu... | Rebuttal 1:
Rebuttal: We thank reviewer ZrN7 for their feedback and subsequently address each stated weakness and question.
### W1: Additional results for negative attributes
While we believe the presented attributes for negative guidance to be representative of SEGA’s capabilities, we agree that the results on nega... | Summary: The paper proses a novel approach called Semantic Guidance (SEGA) to edit an image using text only and without fine-tuning the T2I model. It does this by manipulating the noise estimate in the T2I diffusion model. The space of sparse noise-estimate vectors is referred to as the semantic space. The paper justif... | Rebuttal 1:
Rebuttal: We thank the reviewer for their great feedback and address each weakness separately.
### W1: Large-scale evaluation w/ face attribute detection
In line with the reviewer’s suggestion, we initially attempted to use face attribute detection for large-scale evaluation. As outlined in the Appendix, h... | Summary: This work proposes SEGA (semantic guidance) for text-to-image diffusion models. It is a architecture-agnostic technique and is only applied during the sampling process. In addition to classifier-free guidance, a semantic guidance is applied that pushes noise prediction along the direction of a specific semanti... | Rebuttal 1:
Rebuttal: We thank reviewer i1VN for their input and questions. Below we address each weakness and question individually.
### W1: Image quality measurement.
We agree with the reviewer that a measure of image quality comparison with unaltered images will improve the paper substantially and conducted a res... | Rebuttal 1:
Rebuttal: We thank all reviewers for their valuable feedback and suggestions, which we have taken under careful consideration to improve the paper.
## Comparison
As suggested by the reviewers, we conducted randomized user studies comparing the capabilities of SEGA with related methods. Namely: Prompt2Prom... | NeurIPS_2023_submissions_huggingface | 2,023 | Summary: This paper analyzes classifier guidance in a more detailed way. The main idea is to tie the direction of guidance and the semantic, so that multiple guidance can be used. This idea is simple and effective. The paper also presents many small tricks to improve the performance.
SEGA is evaluated on the face man... | Rebuttal 1:
Rebuttal: We thank the reviewer for their valuable feedback and will address each weakness individually.
### W1: Analysis of classifier-free guidance
1. **Linearity of Semantics.**
We would like to clarify that SEGA concepts scale monotonically and not necessarily linearly, as the space of noise estimates... | null | null | null | null | null | null |
MeCo: Zero-Shot NAS with One Data and Single Forward Pass via Minimum Eigenvalue of Correlation | Accept (spotlight) | Summary: This paper proposes a novel zero-cost proxy called MeCo. The proxy is data-independent, which requires only one random data with a single forward pass. It analyzed that the minimum eigenvalue of P(X) is correlated to the training convergence rate of the network. The experiments on several NAS benchmarks evalua... | Rebuttal 1:
Rebuttal: Thanks a lot for your time and valuable comments and suggestions! We would like to respond to your reviews point by point.
**Q1: How can the theory be connected with network generalization? Can the MeCo metric be applicable to the self-attention layer?**
Answer: 1) MeCo can also reflect the gene... | Summary: This paper proposes MeCo, a new zero-shot NAS method that can rank model architectures without training, without any backward pass, and just using 1 data sample (instead of a batch of data like existing methods). The method is evaluated on multiple NAS Benches and the authors have also run NAS experiments on C... | Rebuttal 1:
Rebuttal: Thanks a lot for your time and valuable comments and suggestions! We would like to respond to your reviews point by point.
**Q1: Results on ImageNet.**
Answer: We adopt MeCo-based NAS to ImageNet-1k dataset on MobileNet-v2 search space. The results are shown in Table 4 in the attached PDF. We tr... | Summary: This paper proposes a new proxy ie MeCo for zero-cost neural architecture search. The proxy design aims at overcoming two important disadvantages of zero-cost proxies proposed earlier. The first one is that these proxies invoke at least one back prop and hence are still expensive. The second disadvantage is th... | Rebuttal 1:
Rebuttal: Thanks a lot for your time and valuable comments and suggestions! We would like to respond to your reviews point by point.
**Q1: Could you evaluate the proxy on all tasks from NAS-Bench suite zero and transformer spaces?**
Answer: Thanks for your suggestions! We conduct the experiments on all 28... | Summary: This paper introduce a new zero-cost proxy for NAS, called MeCo, which is fully independent of data and labels and requires only one data sample for a single forward pass. The paper provides a theoretical analysis of the relationship between the Pearson correlation matrix of the feature maps and the training c... | Rebuttal 1:
Rebuttal: Thanks a lot for your time and valuable comments and suggestions! We would like to respond to your reviews point by point.
**Q1: What are the advantages of Meco over #Param, besides the marginal improvements in ranking correlation? Is there a high correlation between Meco and #Param?**
Answer: W... | Rebuttal 1:
Rebuttal: General Response to the Reviewers.
We thank all the reviewers for their time and valuable comments and suggestions, which are indeed promising and can improve the quality of our paper! In this rebuttal period, we made several improvements and additions to the current version in terms of both the... | NeurIPS_2023_submissions_huggingface | 2,023 | Summary: The paper proposes a new zero-cost proxy (called MeCo) for neural architecture search motivated by theoretical insights about the relationship of the statistics of Pearson Matrix and a network's convergence rate. The proposed proxy is evaluated by investigating ranking correlation on NATSBench-TSS, NATSBench-S... | Rebuttal 1:
Rebuttal: Thanks a lot for your time and valuable comments and suggestions! We would like to respond to your reviews point by point.
**Q1: MeCo and generalization capacity:**
Answer: MeCo can also reflect the network generalization capacity. Specifically, we introduce the following theorem:
**Theorem 3.... | null | null | null | null | null | null |
Multi-Step Generalized Policy Improvement by Leveraging Approximate Models | Accept (poster) | Summary: The authors present an to the Generalized Policy Improvement (GPI) framework, where the agent is able to leverage approximate models to derive a policy in a zero-shot fashion that adapts to unseen tasks related to others it has previously solved.
The paper derives an h-step planning horizon style action select... | Rebuttal 1:
Rebuttal: We thank the reviewer for the positive assessment of our work, and for highlighting the theoretical and experimental contributions of the paper.
### How to identify the best possible value for $h$ ?
This is a great question relating to our formal bounds and the information needed to deploy our me... | Summary: The paper presents a hybrid model-free and model-based policy improvement method for zero-shot policy generalization with successor features. The generalized value function is calculated using a generalized version of the multi-step Bellman operator, where a learned model is used in the rollout to predict futu... | Rebuttal 1:
Rebuttal: We thank the reviewer for the constructive feedback and for the suggestion of new experiments. Below we discuss how our newly added experiments clarify your raised concern.
### Comparing $h$-GPI with SF-MPC with longer horizons
The reviewer asked about the meaning of the experimental results pres... | Summary: This paper studies the combination of model-based planning/rl and model free generalized policy improvement and proposes h-GPI. The authors provide both theoretical and experimental results to support the method.
Strengths: 1. This paper is clearly written and easy to follow.
2. The motivation to study the i... | Rebuttal 1:
Rebuttal: We thank the reviewer for acknowledging that the paper is clearly written, that the theoretical results are thorough and correct, and that both theoretical and experimental results support the method. Below, we carefully answer all concerns raised in your review:
### On the definition of GPI
Than... | Summary: This paper proposes a new hybrid solution of model-free and fully model-based reinforcement learning solutions for planning. It introduces a multi-step extension of generalized policy improvement by having certain planning steps to approximate model behavior, increasing the efficiency of the model. It is follo... | Rebuttal 1:
Rebuttal: We thank the reviewer for the positive assessment of our work. Below, we address their remaining questions:
### On $h$-GPI's underlying motivations and advantages
Thank you for your feedback that further intuitive discussion on the motivation and advantages of $h$-GPI may be provided. The goal of ... | Rebuttal 1:
Rebuttal: We would like to thank the reviewers for their positive feedback and for the constructive suggestions for improving the paper.
In summary, one of the main concerns raised by reviewers was the need to compare GPI with SF-MPC under longer horizons. To address this concern, we have performed additio... | 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.