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 |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
ResoNet: Noise-Trained Physics-Informed MRI Off-Resonance Correction | Accept (poster) | Summary: The Authors describe a physics informed neural network based method to improve the quality and or reduction of MR image acquisition time. The raw signal from magnetic resonance imaging is in Fourier domain, hence normally an inverse Fourier transformation is performed to obtain an image in spatial domain that ... | Rebuttal 1:
Rebuttal: We thank the reviewer for their positive evaluation and valuable feedback on our work. We are pleased that the presentation, figures, and supplementary material have been well-received.
Regarding the discussion on the uniqueness and potential generative hallucinations in the de-blurring operation... | Summary: This work proposes a physics-informed unrolled DL framework for off-resonance correction in MRI trained on synthetic data and they demonstrate the effectiveness of their approach on phantom and in-vivo data. Overall, the approach is sound and has a potential big clinical impact. Empirical evaluations are convi... | Rebuttal 1:
Rebuttal: We appreciate the positive evaluation of our work and the reviewer's confidence in its impact and soundness.
Regarding open-sourcing our datasets, we are committed to making our work reproducible and accessible to the community. We will release our code (which includes the code for generating th... | Summary: The paper presents an innovative approach to off-resonance correction in Magnetic Resonance Imaging (MRI) using a physics-inspired unrolled-deep-learning framework. This framework is designed to address the challenges of off-resonance effects that significantly degrade MRI quality, especially in non-Cartesian ... | Rebuttal 1:
Rebuttal: Yes, we understand the importance of clear and consistent notation to ensure the ease of understanding for all readers, and will clarify raised confusion and improve the writing throughout the paper.
Regarding the notational inconsistency, we acknowledge that this is an oversight on our part. We ... | Summary: This work introduces a physics-informed unrolled-deep-learning framework for off-resonance correction in MRI. The proposed approach maintains data consistency by incorporating a forward model with coil sensitivities, multi-frequency bins, and non-uniform Fourier transforms.
The proposed model is first traine... | Rebuttal 1:
Rebuttal: Our short answer is that recent deep learning methods on off-resonance correction are not directly comparable. See details below.
We chose Autofocus as a fair baseline comparison since it is an analytical method for direct off-resonance correction. Compared to other methods and recent deep-learn... | Rebuttal 1:
Rebuttal: A figure including a comparison to a deep learning baseline is attached. The deep learning baseline performs decently in regions of water-only content but struggles in regions with fat content (e.g., knee scan (d) and brain outer region (c)).
Pdf: /pdf/78e971aa7c61122b8eb0bb126e9797dfeb50d9a6.pdf | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Efficient Test-Time Adaptation for Super-Resolution with Second-Order Degradation and Reconstruction | Accept (poster) | Summary: This paper proposes a fast test-time adaptation SR framework named SRTTA, which is capable of super-resolving images with various degradation types while maintaining efficiency. Specifically, the authors first exploit a pre-trained degradation classifier to predict the degradation type of the test image and co... | Rebuttal 1:
Rebuttal: > Q1. Lack of evaluation on domains with mixed multiple degradations.
A1: In our paper, we seek to adapt pre-trained SR models to target domains when real-world images contain different degradations, which can severely limit the application of most SR models. To this end, we propose a novel SRTTA... | Summary: The paper proposes a test-time adaptation framework called SRTTA for image super-resolution (SR) to address the degradation shift issue between training and test images. The SRTTA framework adapts a pre-trained SR model to different degradation types observed in real-world test images. It introduces a second-o... | Rebuttal 1:
Rebuttal: > Q1. Evaluation merely on a real-world dataset limits the generalizability and applicability of the proposed framework.
A1: Our SRTTA is able to remove the degradation of real-world images and generate images with fewer artifacts, as shown in Figure 3 of the paper and Figure 5 of the supplementa... | Summary: This paper proposed a fast test time adaption method for image SR. The authors point out that the existing image SR method use synthesized paired LR-HR data for training, which degradation type has a gap with the real-world sceneries. The authors also proposed a second-order based data synthesis circuit with a... | Rebuttal 1:
Rebuttal: > Q1. In the abstract, it's better to say 'conventional SR' or 'PSNR/fidelity oriented SR' use paired data for training.
A1: We will revise the sentence of the abstract in the revised paper to avoid misleading.
> Q2. Differences from Real-ESRGAN[1].
A2: We would like to highlight the **differen... | Summary: The submission presents Test-time adaptation Super Resolution (SRTTA) via learning from the second-order degradation and reconstruction. This method super-resolve images with various degradation types while maintaining high efficiency. In addition, the paper also introduce a new dataset DIV2K-C, C stands for c... | Rebuttal 1:
Rebuttal: > Q1. **Significance of the problem setting**: I'm really doubt about the problem setting. Modern ISP always removes noises like Poisson noise, most images are clean besides some compression artifacts.
A1: In this paper, we seek to adapt a trained super-resolution (SR) model to new domains with u... | Rebuttal 1:
Rebuttal: # **General Response**
> G1. Significance and novelty of our SRTTA.
In this paper, we seek to adapt a trained super-resolution (SR) model to domains with unseen degradations. This problem is practical and challenging, as the real-world environment may dynamically change and the collected images ... | NeurIPS_2023_submissions_huggingface | 2,023 | Summary: The paper proposes to solve the domain shift problem for a pretrained super-resolution model. A pretrained classifier is utilized to predict the type of degradation of the test image. After the degradation type is detected, the degradation is used as an augmentation scheme for adaptation training for the new S... | Rebuttal 1:
Rebuttal: > Q1. What are the consequences of domain shift when applying pretrained SR models?
A1: When the domain shift issue occurs, pre-trained SR models often generate HR images with unplausible artifacts, such as over-sharp edges or amplified noise [1, 2].
[1] Blind super-resolution with iterative ker... | null | null | null | null | null | null |
Det-CGD: Compressed Gradient Descent with Matrix Stepsizes for Non-Convex Optimization | Reject | Summary: This paper studies optimization problem under non-convex and smooth setting, where the smoothness is parametrized by a PSD matrix instead of a scalar. It considers using matrix instead of scalar as stepsize, which can trace back to the Newton's method. In addition, it considers the distributed/federated optimi... | Rebuttal 1:
Rebuttal: $\newcommand{\bbS}{\mathbb{S}}
\newcommand{\mS}{\mathbf{S}}
\newcommand{\mD}{\mathbf{D}}
\newcommand{\mT}{\mathbf{T}}
\newcommand{\mW}{\mathbf{W}}
\newcommand{\mL}{\mathbf{L}}
\newcommand{\Exp}[1]{{\mathbb E}\left[#1\right]}$Dear Reviewer WhEe,
Thank you very much for your careful reviews and tho... | Summary: This paper analyzed two different non-convex optimization algorithms with matrix step sizes with sketching.
Strengths: 1. The paper coupled sketching with matrix stepsize, and detailed a number of conditions that need to be satisfied in order for the algorithm to work.
Weaknesses: 1. (9) and (10) do not clea... | Rebuttal 1:
Rebuttal: Dear Reviewer DJA3,
Thank you very much for your careful reviews and thoughtful questions for the paper. Below is a point-to-point response to your questions and concerns.
**Weaknesses**
> a) `(9) and (10) do not clearly communicate why using a matrix stepsize D is better than a constant stepsi... | Summary: The paper studies matrix step-sized and compressed gradient descent method for non-convex optimization. Two types of update are given (det-CGD1 and det-CGD2) . They are extensions of the vanilla gradient descent from scalar step-sizes to matrix cases, and can be considered as Newton-Type methods with a constan... | Rebuttal 1:
Rebuttal: $\newcommand{\mD}{\mathbf{D}}
\newcommand{\bbS}{\mathbb{S}}
\newcommand{\mT}{\mathbf{T}}
\newcommand{\mW}{\mathbf{W}}
\newcommand{\mL}{\mathbf{L}}
$Dear Reviewer DW9L,
Thank you for your reviews and thoughtful questions for the paper. Below is a point-to-point response to your questions and conce... | Summary: This paper proposes two novel single-node and distributed algorithms for non-convex optimization via SGD with matrix stepsizes and compressed gradients via random sketches. The authors also provide theoretical guarantees for convergence of their methods to a stationary point under a matrix smoothness assumptio... | Rebuttal 1:
Rebuttal: $\newcommand{\norm}[1]{\lVert #1 \Vert}
\newcommand{\mD}{\mathbf{D}}
\newcommand{\R}{\mathbb{R}}
\newcommand{\Exp}[1]{{\mathbb E}\left[#1\right]}
$Dear Reviewer 72cD,
Thank you very much for your careful reviews and thoughtful questions for the paper. Below is a point-to-point response to your qu... | Rebuttal 1:
Rebuttal: We thank the reviewers for their time and effort spent on reading our submission and their constructive comments. The reviewers highlighted the following strengths of the paper:
* The novel design of the first matrix stepsized CGD in combination with sketch matrices.
* The extension of the settin... | NeurIPS_2023_submissions_huggingface | 2,023 | Summary: 1. This work analyzes two compressed gradient descent methods with matrix step sizes for non-convex optimization.
2. This work designed the sketches and step sizes to take advantage of the layer-wise structure.
3. The distributed version of the proposed method is also discussed under the Federated Learning f... | Rebuttal 1:
Rebuttal: $\newcommand{\mD}{\mathbf{D}}
\DeclareMathOperator{\Diag}{Diag}
\newcommand{\mI}{\mathbf{I}}
\newcommand{\R}{\mathbb{R}}
\newcommand{\mS}{\mathbf{S}}
$Dear Reviewer i86q,
Thank you for your review and suggestions for the paper. What follows is a point-to-point response to your questions and conc... | null | null | null | null | null | null |
Intriguing Properties of Quantization at Scale | Accept (poster) | Summary:
The paper empirically studies the effect of optimization/regularization hyperparameters -- weight-decay, dropout, gradient-clipping -- on post-training quantization (PTQ) performance of large language models (LLMs). Specifically, it performs PTQ to int8 by simple affine scaling, adopts a GPT-style architectu... | Rebuttal 1:
Rebuttal: We thank **R TcCw** for their valuable feedback and noting that the problem of quantizing both activations and weights is “well motivated “ and that our paper is “clearly written” and “easy to follow”.
> While paper focuses on the effect of regularization choices using a GPT style architecture ... | Summary: Recent research suggests that the trade-off incurred by quantization is an emergent property, resulting in sharp performance drops in models with over 6 billion parameters. In this study, the authors investigate whether quantization cliffs in performance are solely determined by model scale. Considering the in... | Rebuttal 1:
Rebuttal: We thank **R xK34** for the positive feedback, including their observations that 1) We conducted a series of “controlled experiments, maintaining consistent architecture while varying key optimization choices” which allowed us “systematically investigate the factors influencing sensitivity quantiz... | Summary: There is a quantization tradeoff in LLMs, where quantization methods drop in particular for some larger models. This problem has been attributed to large activation magnitude "outliers". The authors note that activation outliers are not an emergent property of LLMs and that instead a byproduct of how the netwo... | Rebuttal 1:
Rebuttal: We thank **R vYyu** for their helpful feedback and their observations noting that our paper is “easy to follow” and “clear” in describing the motivation, problem, and related work.
> As noted by the authors, other work has also noted that quantization techniques are more stable if the network is... | Summary:
The authors study the question of whether quantization (of weights and
activations in decoder-only language transformer architectures) fundamentally
reduces performance at large parameter counts, or whether this degradation of
performance, observed in prior works on this topic, is simply a consequence of
cert... | Rebuttal 1:
Rebuttal: We thank **R egL3** for noting that our findings “lead to a significant improvement” and our results “highlight useful general takeaways about how regularization can help mitigate PTQ performance drop”. We are glad they found that our experimental setup is presented “very clearly”, aiding “the ab... | Rebuttal 1:
Rebuttal: We greatly appreciate the thoughtful and positive feedback from the reviewers. We are encouraged that reviewers found the results notable [xK34, egL3] “insights lead to a significant improvement over OPT-66B PTQ accuracy”. Reviewers also positively reviewed our experimental setup as rigorous, clea... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Optimization or Architecture: How to Hack Kalman Filtering | Accept (poster) | Summary: The paper presents a new Kalman Filter (KF) algorithm named Optimized Kalman Filter, which applies an optimization to the supervised learning loss as usually used for training non-linear models (e.g. NKF). For linear KF, a more straightforward way is to estimate the noise covariance, which is directly plugged ... | Rebuttal 1:
Rebuttal: We thank the reviewer for their helpful feedback. We will modify our manuscript accordingly. Please see our detailed responses below.
**In section 4, it would be interesting to compare with KF with known covariance matrices.**
We conducted this experiment in Appendix B1. The oracle performed ver... | Summary: In the domain of nonlinear filtering, it's a usual practice to draw comparisons between nonlinear architectures like neural networks and traditional linear Kalman Filters (KF). This approach blends two distinct aspects: the nonlinear structure and the numerical optimization technique, thereby introducing poten... | Rebuttal 1:
Rebuttal: We thank the reviewer for their helpful feedback. Please see our responses below.
**It would be nice to suggest various optimization methods.**
Our key insight is that noise estimation (Algorithm 1) is not a reliable proxy to MSE optimization, and instead the MSE should be optimized directly. Th... | Summary: The paper is considering a linear Gaussian filtering problem where the noise covariance are unknown. It assumes it has access to state and observation data. It aims to find the noise covariances by minimizing the estimation error of a Kalman Filter using those noise covariance matrices. This is in contrast to... | Rebuttal 1:
Rebuttal: We thank the reviewer for reading our work and providing honest feedback. We respond to the reviewer’s comments below. More importantly, the review specifies that the main message was not clear to the reviewer. **Since it is important for us to make the presentation as clear as possible, we recap ... | Summary: This paper uses stochastic optimization to estimate the covariance of Kalman filters in supervised tracking problems. They find that online stochastic optimization in this manner tends to generalize better than either neural Kalman filters with LSTM base models or the older plug-in estimators for Kalman filter... | Rebuttal 1:
Rebuttal: We thank the reviewer for their detailed feedback. **Before responding in detail and presenting the new experiments as requested, we wish to focus on the comments about the significance of our contribution. This may clarify some of the reviewer’s questions, and we will appreciate it if the reviewe... | Rebuttal 1:
Rebuttal: We thank all the reviewers for their helpful comments. We are pleased that the reviewers found our work “**original**” (JSXK), **”honest”** (mJXo), **”meaningful”** and “**very novel**” (GLsN), and acknowledged that we point to a **“prevalent methodological error”** (GLsN) and a **“profound neglec... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Fine-grained Late-interaction Multi-modal Retrieval for Retrieval Augmented Visual Question Answering | Accept (poster) | Summary: This paper studies two problems in knowledge-based VQA: 1) the image input is often transformed into the text; 2) the relevance score between query and document is estimated based on one-dimensional dot product.
To address these problems, the authors leverage several image features, including ROIs features fro... | Rebuttal 1:
Rebuttal: Thank you very much for raising these concerns.
- **Is there also ground-truth knowledge for the questions in the test set? Did the authors use ground-truth knowledge for testing?**
- No and no. There is no ground-truth knowledge annotation of the document collection for the question in the t... | Summary: The paper proposed a fine-grained Late-interaction multimodal knowledge retriever for knowledge-based VQA. The core idea to the to both visual tokens and question tokens as token-level queries to retrieve document in late interaction fashion. The paper suggests that (1) squeezing query representation and docum... | Rebuttal 1:
Rebuttal: Thank you very much for your very positive review. Thank you for your careful reading as well. We respond to your constructive suggestions as below:
1. **Hope to know if it requires very good SSD or RAM though it is cheaper than the GPUs.**
Thank you for raising this point. Our cluster e... | Summary: This paper proposes a multi-vector (late interaction) multi-modal retrieval-based approach for Knowledge-based Visual Question Answering, inspired by the ColBERT multi-vector representations for text. The retriever improves the recall substantially over prior multi-modal methods that use single-vector encoding... | Rebuttal 1:
Rebuttal: Thank you for your very careful reading and the accurate summarization of our strengths. Thank you very much!
1. **What are the connections to FILIP (Yao et al., 2021), which is also a late interaction retriever for some multi-modal settings. FILIP seems to tackle zero-shot image classification a... | Summary: This paper introduces a new method, Fine-grained Late-interaction Multi-modal Retrieval, to improve knowledge retrieval in RA-VQA. The method addresses two major limitations in RA-VQA’s retriever: incomplete image understanding and lossy compression of visual scenes and questions.
Strengths: 1) The paper pro... | Rebuttal 1:
Rebuttal: Thank you for your careful reading and kind suggestions. We will answer your questions point by point below:
1. **This work used the VinVL approach, which leverages regions of interest for visual features. I could not pinpoint the main contributions of this work, which DPR does not make. For a be... | Rebuttal 1:
Rebuttal: We appreciate the time and efforts of all seven reviewers in assessing our paper. The positive feedback on our work is highly encouraging, and we have taken great care to respond to each reviewer individually. Here, we hope to provide direct and concise clarifications addressing the common concern... | NeurIPS_2023_submissions_huggingface | 2,023 | Summary: This paper proposed an updated version of RA-VQA by incorporating a vision model aligned with text-based retriever in a late interaction way. Extensive experiments show the effectiveness of the proposed method. Thorough ablation studies and analysis were conducted.
Strengths: 1. The paper is well-written and ... | Rebuttal 1:
Rebuttal: Thank you very much for your very positive review of our work. We are excited to learn that you enjoyed the presentation and the experiments.
We would like to respond to your concerns as below:
**1. Though the performance shows the superiority of the proposed method, the method of adding visual ... | Summary: In this paper, the authors focus on RA-VQA framework to solve the KB-VQA task. Specifically, this paper proposed Fine-grained Late-interaction Multi-modal Retrieval (FLMR) to address the limitation of inaccurate image representations and one-dimensional embeddings. FLMR leveraged late interaction and multi-dim... | Rebuttal 1:
Rebuttal: Thank you for your positive review of the strength of our work. We hope that you enjoyed reading our paper. Below we would like to respond to your major concern and question:
1. **Have the authors performed experiments on other KB-VQA datasets or is that planned? If not, do authors have any parti... | Summary: This paper focuses on retrieval augmented visual question answering and addresses two major limitations in previous RA-VQA. Specifically, they propose FLMR and overcome the limitations by obtaining image representations using a vision model. FLMR also encodes images and questions using multi-dimensional embedd... | Rebuttal 1:
Rebuttal: **Weakness and Question 1**:
Please kindly refer to the general response where we provide a list of our technical novelties. If accepted, we will highlight these points more clearly in the camera-ready version following your suggestions.
**Question 2: In line 163-164, please explain the improve... | null | null |
Are Emergent Abilities of Large Language Models a Mirage? | Accept (oral) | Summary: The authors argue that purported emergent abilities of language models are mostly explained by the use of specific metrics, including non-linear or discontinuous scoring, and evaporate when linear or continuous scoring metrics are applied. The authors (1) present a toy model to explain emergence, (2) demonstra... | Rebuttal 1:
Rebuttal: We thank Reviewer ccWP for their time and feedback! Replying to each comment, suggestion and/or question in turn:
> As the authors point out in their related work, the idea that emergence appears due to the metric is not novel, and was proposed by [1], so the novelty of this paper lies in the qua... | Summary: This paper aims to debunk a possibly myth about the emergent properties of language models with regard to scale. The emergent behaviors are defined as abilities that are not demonstrated with smaller models, but emerge suddenly with an increase in the model capacity, and cannot be explained by smoothly extrapo... | Rebuttal 1:
Rebuttal: We thank Reviewer eK8N for their time and feedback! Replying to each comment, suggestion and/or question in turn:
> The authors cite the following definition early on -- emergent abilities are "abilities that are not present in smaller-scale models but are present in large-scale models; thus the... | Summary: This paper shows that the widely-discussed phenomenon of "emergent abilities" in large language models is largely due to the choice of (discontinuous/nonlinear) metrics and underpowered analyses.
Strengths: - Well-written paper studying an important/commonly discussed phenomenon
- Comprehensive, well-designe... | Rebuttal 1:
Rebuttal: We thank Reviewer GXCv for their time and feedback, and we’re overjoyed to hear you write that the paper was a pleasure to read! Replying to each comment, suggestion and/or question in turn:
> l227: Not sure what is meant by "multiple controls" - is the implication that the community should have ... | Summary: The paper shows that the apparent suddenness with which capabilities of language models seem to emerge with scale may be an artifact of the metrics by which they are measured. Theoretically, emergence should be more gradual if the metric scales linearly with log loss on some test set. Three sources of empirica... | Rebuttal 1:
Rebuttal: Thanks for time & feedback! You made clear effort to think critically about how to improve our paper! 6k space limited -> dense response; try our best.
> Why few-shot int arithmetic
1. Few-shot int arithmetic featured prominently in multiple prev papers L102-104.
2. To make point about sufficien... | null | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Implicit Bias of Gradient Descent for Logistic Regression at the Edge of Stability | Accept (spotlight) | Summary: The authors study gradient descent for logistic regression on linearly separable data, with a focus on large stepsizes, and in particular stepsizes in the edge of stability (EoS) regime.
With the logistic loss, they show that for *any* positive stepsize, the empirical loss is minimized as the number of GD step... | Rebuttal 1:
Rebuttal: We appreciate your strong support! Thank you for pointing out additional related papers. We will make sure to cite and discuss them properly in the revision. We address your question below.
---
**Q1**. “...since the implicit bias is the same whatever the stepsize is (it is thus the same for infin... | Summary: This paper analyzes the convergence and implicit bias of logistic regression on linearly separable datasets with arbitrarily large step sizes. The paper proves that gradient descent converges _for any_ constant step size, even those which result in non-decreasing loss curves. Similar to the regime of small s... | Rebuttal 1:
Rebuttal: We appreciate your affirmative comments! Thank you for pointing out the paper by [Liu, Huang, and Xu, 2023]. We will properly cite and discuss the paper in the revision. We will address your other questions as follows.
---
**Q1**. “Minor: would Assumption 3 hold if the dataset were the last-layer... | Summary: The paper studies the convergence of gradient descent with large step sizes beyond the classical stepsize suggest by smoothness-based convergence. Across the iterates, the risk here exhibits non-monotonic behaviour and yet converges in the long range. To show these phenomenon, the paper rigorously studies the... | Rebuttal 1:
Rebuttal: Thank you for supporting our paper! We address your comments in the following.
---
**Q1**. “The proof is tailor made for linear regression and does not work beyond that for any convex functions.”
**A1**. We respectfully point out that our analysis is for logistic regression rather than linear r... | Summary: This research investigates the convergence and implicit bias of constant-stepsize GD for logistic regression on linearly separable data in the EoS regime. The authors assert that despite local oscillations, the logistic loss can be minimized by GD with constant stepsize over a long time scale. Experiments on ... | Rebuttal 1:
Rebuttal: Thank you for appreciating our contributions including a novel theory, new techniques, and practical insights. We address your concerns below.
---
**Q1**. “The analysis is limited to logistic regression on linearly separable data. Extending it to more complex models like neural networks is not c... | Rebuttal 1:
Rebuttal: The attached rebuttal PDF contains the results of suggested experiments for logistic regression with extremely large stepsizes and homogenous neural networks.
Pdf: /pdf/ab1557eb68f4c8908fc20ec73b499641974dbbfa.pdf | NeurIPS_2023_submissions_huggingface | 2,023 | Summary: This work investigates the implicit bias in linear regression with logistic loss with large learning rates. It shows the convergence result of logistic loss for any learning rates, as long as an counter-example of exponential loss to show the gap between these two losses.
Strengths: 1. The motivation and sett... | Rebuttal 1:
Rebuttal: Thank you for your comments, suggestions, and positive evaluation. Please find our answers to your questions below.
---
**Q1**. “The covered setting is somehow simple in lack of detailed discussions of what impact logistic loss + large learning rates have on more complicated settings possibly. F... | null | null | null | null | null | null |
Coordinating Distributed Example Orders for Provably Accelerated Training | Accept (poster) | Summary: This paper is devoted to the development of the Gradient Balance for the centralized distributed setting.
The proof is highly non-trivial because the extension to distributed ML is not natural. To this end, the authors use kernel thinning to translate the benefits of provably faster permutation-based example ... | Rebuttal 1:
Rebuttal: Thank you for your review. We respond below.
> 1. The presentation may need to be improved. For example, if we turn to Table 1 (Algorithm 1), the algorithm is not self-explanatory because some notation are not defined.
**Algorithm 1**: Thank you for this feedback. Can you please indicate where t... | Summary: Propose an online method, CD-GraB for shuffling data samples during distributed model training (parameter-server setting) in a fashion that accelerates convergence in both theory and practice.
Theoretical assumptions are reasonable and quite standard in the literature. Empirically, this work mostly builds on ... | Rebuttal 1:
Rebuttal: Thank you for observing the challenging nature of our work and the value of our contributions (both in terms of novel results, and in clarifying them / communicating them effectively). We respond below.
> Significance While the overall contribution of the work is impressive, it remains the case t... | Summary: The authors propose a distributed version of Grab. The authors both theoretically and experimentally show that there exists a pattern which can improve training convergence of such a method in a distributed setup.
The paper is very well and tightly written with clear limitations. The authors verified their c... | Rebuttal 1:
Rebuttal: Thank you for your review and questions. Please see our answers below.
> As the authors have stated that there is additional memory overhead keeping older gradients. As memory becomes more scarce due to larger models, do authors think there is a way of approximating the previous epoch gradients?
... | Summary: The paper introduces Coordinated Distributed GraB, a new method that leverages stale gradients from prior epochs to order examples and achieve a provably faster convergence rate than random reshuffling. CD-GraB demonstrates a linear improvement in convergence rate compared to centralized GraB and showcases sup... | Rebuttal 1:
Rebuttal: Thank you for your thoughtful review and questions,
> 1. This method is designed specifically for the paradigm involving a parameter server and workers. However, it is important to note that as the number of workers increases, there will be a linear increase in overhead, potentially affecting pe... | Rebuttal 1:
Rebuttal:
# Top-level response to all reviewers
Thank you for your comments. We appreciate the detailed feedback, and think that there are ideas in the reviews that would be interesting to pursue in future work (but unfortunately don’t fall in scope for the present project). We clarify below.
## A. Prio... | NeurIPS_2023_submissions_huggingface | 2,023 | Summary: This paper proposes a method to extend the optimal data order finding method GraB to centralized distributed setting. This is achieved by using an alternative balancing method that does not require using stale averages which would cause a problem given higher learning rates in the distributed settings. Further... | Rebuttal 1:
Rebuttal: Thank you for your review, and for appreciating the writing quality of our submission. We respond directly to your comments about weaknesses below.
**Scale of experiments**
* We note in the paper/Appendix that we ran we could afford to. CD-GraB converges faster than D-RR, which the existing resu... | null | null | null | null | null | null |
Mutual-Information Regularized Multi-Agent Policy Iteration | Accept (poster) | Summary: This paper propose a new method: LIMI, which utilizes mutual information as a regularizer to help MARL algorithms better face to the variation of team decomposition and generalize on unseen tasks. Both detailed theory proof and practical details are provided.
Strengths: - The paper is well-written, both detai... | Rebuttal 1:
Rebuttal: > More environments needed: SMAC is very sensitive to hyper-parameters. Therefore more results on other environments are needed to make the claims in the paper persuasive (e.g. smacv2, MAMuJuCo).
Thanks for your valuable suggestion. We have carried out additional experiments on more scenarios to ... | Summary: The paper presents a MARL algorithm that utilizes an SAC-based auxiliary mutual information reward regularizer to generalize learned decentralized multi-agent policies to different team compositions. The method has been evaluated against related baselines and in two simple and complex domains, demonstrating a ... | Rebuttal 1:
Rebuttal: > There’s no clear definition of “team composition”...
As we described in Line[136-138], we assume each state $s$ can be decomposed into the state of entities $s_e$, and entities include agent entities and non-agent entities. So the term "team composition" here simply means the composition of ag... | Summary: This paper proposes a method which use the mutual information as intrinsic reward to prevent the agents over-rely on its collaborators. Meanwhile, the proposed method can encourage the agents to learn robust policies in different team compositions. Moreover, this paper utilizes Blahut–Arimoto algorithm and an ... | Rebuttal 1:
Rebuttal: > Some important baselines are missed. Since the authors have mentioned them in the related work, the new published and common used baselines (e.g. MAPPO, HATRPO) should not be neglected.
> Performance on more challenging scenarios of different benchmarks deserve to be verified.
> How the model ... | Summary: To address the issues of overfitting in Multi-Agent Reinforcement Learning (MARL) and dynamic team formation, this paper proposes a new approach of using mutual information regularization as intrinsic rewards to prevent individual policies from excessively relying on team-related information. Experimental resu... | Rebuttal 1:
Rebuttal: > This paper aims to reduce the reliance of agents on team-related information in decision-making. However, one potential issue may exist with this approach. It may result in fewer cooperative behaviors among agents, leading to a greater emphasis on individual performance than team collaboration. ... | Rebuttal 1:
Rebuttal: We would like to thank all reviewers for their time and efforts to review our paper and give insightful comments that could further improve our work.
As several reviewers suggest including more experiments, we are now working on this. We plan to include experiments on Resource Collection (an MPE... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Gradient strikes back: How filtering out high frequencies improves explanations | Reject | Summary: This paper investigates the issue of noisy gradient-based explanations caused by high-frequency content. The authors empirically identify that several down-sampling operations, such as MaxPooling or stride convolution, could be the primary source of these high frequencies. To address this, they present FORGrad... | Rebuttal 1:
Rebuttal: The reviewer's dedicated time and their kind appraisal of our work are greatly appreciated by us. Each of the weaknesses noted by the reviewer has been addressed in the following content, as detailed in our responses.
**- 1:** Thank you for this insightful comment and for carefully checking the d... | Summary: The paper analyzes the two types of attribution methods: gradient-based/ white-box methods and prediction-based/ black-box methods. The authors observe that the faithfulness of black-box methods surpasses the white-box methods despite white-box methods accessing the internals of the classifier being explained.... | Rebuttal 1:
Rebuttal: We extend our appreciation to the reviewer for sparing their time and for the thoughtful remarks they've provided regarding our work. We have taken the opportunity to respond to the weaknesses raised by the reviewer in the content that follows.
**- Possible misrepresentation of the related litera... | Summary: The paper explores the differences between prediction-based and gradient-based attribution methods for explaining the decisions of deep neural networks. The authors observe that these two approaches produce attribution maps with distinct power spectra, with gradient-based methods exhibiting more high-frequency... | Rebuttal 1:
Rebuttal: We express our thanks to the reviewer for allocating their time and for the encouraging feedback they've shared about our work.
The ensuing content contains our responses to the weaknesses indicated by the reviewer.
**- Flaw in Observation:** You are absolutely right, and it is indeed common know... | Summary: The paper investigates why gradient-based attributions contain more high-frequencies compared to prediction-based
attribution. The authors argue that these high-frequency components are not relevant for the model and can
therefore seen as noise. Different saliency methods are analyzed on three different image ... | Rebuttal 1:
Rebuttal: Our sincere gratitude goes to the reviewer for their valuable time invested and for their generous appreciation of our efforts in this endeavor. In the subsequent content, we address each of the weaknesses highlighted by the reviewer.
**- Missing relevant work:** Thank you for bringing this pap... | Rebuttal 1:
Rebuttal: We express our heartfelt appreciation to the reviewers for devoting their time to meticulously read and evaluate our paper. Their critiques have been perceptive and thought-provoking. We believe that we were able to address most of their comment in a satisfactory manner and the resulting manuscrip... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Towards Characterizing the First-order Query Complexity of Learning (Approximate) Nash Equilibria in Zero-sum Matrix Games | Accept (poster) | Summary: This paper investigates the first-order query complexity of two-player zero-sum games. More precisely, given a bi-matrix zero-sum game $(A,-A^\top)$ (with $A \in \mathbb{R}^{k\times k}$) a first-order query is a set of strategies $(x,y)$. The answer to such a query is the set of vectors $(A y, -A^\top x)$ resp... | Rebuttal 1:
Rebuttal: We thank Reviewer Wvc4 for the nice comments and for the reference, we will add it to the final version.
---
Rebuttal Comment 1.1:
Comment: Thank you for response. I have carefully read the other reviews as well as your responses. I believe that your result are interesting for the *learning in g... | Summary: This paper studies the query complexity lower bound of first-order methods for finding exact and approximate Nash equilibria in two-player zero-sum games. It is first established that if the matrix entries come from a fixed countable set, one query suffices to reconstruct the full matrix and hence to compute t... | Rebuttal 1:
Rebuttal: We thank Reviewer EExk for the helpful remarks and kind comments! We will fix the typos!
**The current lower bound rate is still quite far from the existing upper bound** See our answer to 'Reviewer neqq':
We agree that our results are quite technical and do not fully
characterize the rates of s... | Summary: Define the gap of a two player zero sum game with payoff matrix M and mixed strategies p and q as:
gap(M,p,q):=max_j (M^Tp)_j - min_i(Mq)_i
An equilibrium is called \eps approximate if the gap is at most 2\eps. An important question is to compute the approximate Nash equilibrium by observing the gradients of ... | Rebuttal 1:
Rebuttal: We thank Reviewer nqee for the helpful remarks and the interesting question!
**'the actual result is fairly technical and may not be of broad interest since the improvement is marginal'** We agree with the reviewer that our results are quite technical and do not fully characterize the rates of s... | Summary: This paper delves into the complexities associated with identifying a mixed Nash equilibrium within the context of a zero-sum game. Specifically, it scrutinizes situations where the algorithm does not directly observe the payoff matrix. Instead, it learns about it by querying its product with either a column v... | Rebuttal 1:
Rebuttal:
We thank Reviewer 3esb for the remarks and comments!
We kindly ask Reviewer 3esb to read the other reviewers' feedback regarding the concerns about the clarity of our paper. In particular, please note that all other reviewers did specifically mention the clarity and soundness of our paper as a st... | null | NeurIPS_2023_submissions_huggingface | 2,023 | Summary: This paper focuses on the query complexity of learning (approximate) Nash equilibria in zero-sum matrix games. The authors provide lower bounds on the query complexity for exact and approximate equilibria. They also show that if the game matrix has entries in a known countable set, the learner can recover the ... | Rebuttal 1:
Rebuttal: We thank Reviewer VSPH for the helpful remarks and comments!
**Limited novelty due to heavily building on existing results:**
We did our best to thoroughly present the state of the problem and the related literature, and to properly credit the ideas we build upon. This, we believe, highlights the... | null | null | null | null | null | null |
HQA-Attack: Toward High Quality Black-Box Hard-Label Adversarial Attack on Text | Accept (poster) | Summary: This paper proposed a new approach HQA-Attack using iterative substituting normal words and replacing with synonyms for creating high quality adversarial text samples in the setting of black box attack + hard label. The experimental results on several NLP datasets and two commercial APIs validate the effective... | Rebuttal 1:
Rebuttal: **For Reviewer BWcn.**
Thanks very much for your valuable and positive comments.
**1. Theoretical analysis of computational complexity and the number of queries.**
We have attempted to analyze the computational complexity and the number of queries for the main procedures of HQA-Attack, includin... | Summary: This paper presents a black-box adversarial attack method that can create high-quality adversarial samples without accessing the predictive probability/gradients of the victim model. The approach is based on synonyms substitution. In the attack framework, this paper also proposes some tricks to reduce the quer... | Rebuttal 1:
Rebuttal: **For Reviewer pVxP.**
Thanks very much for your valuable and helpful comments.
**1. About the contributions.**
We would like to clarify our contributions.
(1) We propose a novel method for substituting original words back (Sec. 4.2), which is different from previous methods. Specifically, pre... | Summary: - The authors consider a problem of black-box hard-label adversarial attack, which adversarially perturbs the input text while having access only to the output label of the victim model.
- They propose HQA-Attack method which can successfully perturb the input text into the adversarial text with higher semanti... | Rebuttal 1:
Rebuttal: **For Reviewer RbiV.**
Thanks very much for your valuable and positive comments.
**1. About the experiments with the sentence-dependent word substitution set.**
(1) For fair comparison, we follow the previous works HLGA, TextHoaxer and LeapAttack to generate 50 synonyms for each word by using t... | Summary: This paper proposes a new hard-label black-box attack on textual data that aims to improve the performance on semantic similarity and perturbation rate. The main novel techniques are word back-substitution and a new strategy for optimizing the adversarial example. Evaluations on 5 classification tasks, 3 infer... | Rebuttal 1:
Rebuttal: **For Reviewer 27hb.**
Thanks very much for your valuable and positive comments.
**1. About the word back-substitution component.**
(1) For the word back-substitution component, HQA-Attack is different from previous methods. Specifically, as shown in Alg. 1 (Appendix A), after replacing one ori... | Rebuttal 1:
Rebuttal: **For everyone.**
We would like to thank everyone for devoting the valuable time and effort to assist us in enhancing the quality of our manuscript.
We have provided the detailed responses for each review, and we really appreciate your consideration. | NeurIPS_2023_submissions_huggingface | 2,023 | Summary: This paper proposes a novel and effective methods for black-box textual adversarial attacks. The proposed framework focuses on maximizing the similarity of the adversarial texts and the original texts through shrinking the perturbation. In the experiments, the authors prove that the proposed methods can signif... | Rebuttal 1:
Rebuttal: **For Reviewer gEht.**
Thanks very much for your insightful and positive comments.
**1. Add human evaluation experiments.**
We have conducted the human evaluation experiments on the Bert model using HLGA, TextHoaxer, LeapAttack and HQA-Attack for the MR and IMDB datasets. Specifically, for each... | null | null | null | null | null | null |
DSR: Dynamical Surface Representation as Implicit Neural Networks for Protein | Accept (poster) | Summary: The work proposes to model the surface of proteins using signed distance field (SDF) expressed by implicit neural representations. Previously, protein surfaces were studied to characterize protein interactions and ligand binding. In this work, surface representation learning is for analyzing MD simulation data... | Rebuttal 1:
Rebuttal: **Q1**: According to the provided visualizations (e.g. Figure 3), the proposed method could add some non-existent tiny areas close to the structure but disconnected. I believe these artifacts are due to implicitly representing surfaces.
**A1**: We acknowledge the complexity of protein surfaces, w... | Summary: This paper proposes to learn an implicit neural network to model protein surface and dynamics. A protein surface is represented as a signed distance function (SDF), which can be modeled by implicit neural networks based on MLP. The training objective follows standard practice: the difference between true and p... | Rebuttal 1:
Rebuttal: Thanks for your review comments.
**Q1**: Currently, DSR cannot generalize to a new protein.
**A1**: We appreciate your valuable feedback regarding the generalizability of our DSR method to new proteins. We acknowledge that the current set of results demonstrates a limitation in the ability of D... | Summary: The author puts forwards the DSR model, which is an implicit neural network, uses this model to accurately capture the protein dynamic trajectories and interpolate and extrapolate in 3D and time. The DSR is a scalable and efficient method to modeling protein dynamics, overcomes the limitations of first-princip... | Rebuttal 1:
Rebuttal: Thanks for your review comments.
**Q1**: In the reconstruction task, there are little comparison between the DSR and other methods.
**A1**: Thank you for your feedback regarding the comparison between our DSR method and other methods in the reconstruction task. As we are the first to propose and... | Summary: The paper proposes a framework for predicting protein dynamics through learning the time-dependent SDF. The authors showcase that the learned SDF can reconstruct/interpolate protein surfaces and predict the dynamics, thus bypassing the computationally expensive molecular dynamics simulation.
Strengths: * The ... | Rebuttal 1:
Rebuttal: Thanks for your review comments.
**Q1**: What is the advantage of the proposed method over a point-cloud GNN that learns to predict atom-wise displacement?
**A1**: Thank you for your question regarding the advantages of our proposed method over a point-cloud GNN that learns to predict atom-wise ... | Rebuttal 1:
Rebuttal: I would like to extend my sincere appreciation for the reviewers’ efforts in coordinating the review process for our submission. In response to the reviewers' comments, we have carefully analyzed and compiled their concerns and suggestions. I would like to summarize the key points raised by the re... | NeurIPS_2023_submissions_huggingface | 2,023 | Summary: The paper proposes to learn an implicit neural representation to model protein surface across continuous spatial and temporal space.
Strengths: This is a great application of implicit neural networks for modeling protein surfaces, and the results seem promising.
Weaknesses: The evaluation of the method can b... | Rebuttal 1:
Rebuttal: Thanks for your review comments.
**Q1**: Comparing the learned model with simple methods, such as linearly interpolate?
**A1**: Following the reviewer's suggestion, we conduct additional experiments comparing our methods to linear interpolation that generates intermediate structures between two... | null | null | null | null | null | null |
TopoSRL: Topology preserving self-supervised Simplicial Representation Learning | Accept (poster) | Summary: The paper introduces a new self-supervised learning (SSL) method called TopoSRL for simplicial complexes. It claims to overcome the limitations of existing graph-based SSL methods by capturing higher-order interactions and preserving topology in learned representations. The authors propose a simplicial augment... | Rebuttal 1:
Rebuttal: We thank the reviewer for the valuable feedback. We address your concerns and questions as follows:
### Improving readability
To improve the readability, we will place the important definitions in the main text, reduce the overflow of equations, and add an image to explain intuitively the working... | Summary: This paper proposed a novel self-supervised learning (SSL) method for simplicial complexes. Advanced results are achieved by introducing appropriate augumentation methods and contrastive loss, taking into account the topological properties of the simplicial complex. It also compares the proposed method with th... | Rebuttal 1:
Rebuttal: We appreciate the reviewer's insightful comments.
### On comparison with graph analysis methods
While it is true that simplicial complexes cannot be directly processed as graphs, it is important to note that every simplicial complex inherently includes an underlying graph that captures the 0-simp... | Summary: The paper introduces a new algorithm for self-supervised learning for simplicial complex-based networks.
The paper proposes a new simplicial contrastive loss function that contrasts the generated simplices to preserve local and global information present in the simplicial complexes.
Experiments show the utili... | Rebuttal 1:
Rebuttal: We thank the reviewer for the valuable feedback. We address your concerns and questions as follows:
### On similarity with graph augmentation and contrastive learning
We acknowledge that the concept of augmentation for graphs has gained popularity in graph SSL methods. However, as highlighted in ... | Summary: This work designs an SSL pipeline coined as TopoSRL for simplicial complex data, and proposes a new contrastive loss function to preserve topology information to learn more expressive representations.
Strengths: 1. The motivation is meaningful.
2. The experiments are fairly comprehensive.
Weaknesses: 1. In... | Rebuttal 1:
Rebuttal: We thank the reviewer for the valuable feedback. We address your concerns and questions as follows:
### Lack of SSL baselines
To compare with more graph SSL baselines, we have added BGRL and GCA as graph SSL baselines and GIN [1] as graph supervised baseline. The results are available in Table G1 ... | Rebuttal 1:
Rebuttal: We thank all the reviewers for their insightful comments and thoughtful questions. As suggested by reviewers, we have conducted the following experiments and presented them in the global response PDF. The new results are in red color.
### **Graph Classification on TU Datasets [GC]**
- As suggeste... | NeurIPS_2023_submissions_huggingface | 2,023 | Summary: The work proposes a Self-Supervised Learning (SSL) approach for simplicial complexes, extending graph SSL to higher-order structures. While recent works exist that extend and adapt standard graph representation learning techniques to simplicial complexes, none consider SSL. In particular, the authors propose t... | Rebuttal 1:
Rebuttal: We thank the reviewer for the valuable feedback. We address your concerns and questions as follows:
### Graph classification on TUDatasets using clique lifting
We have incorporated your suggestion and included Table G2 in the global response PDF, presenting the performance comparison between the p... | null | null | null | null | null | null |
Task-aware world model learning with meta weighting via bi-level optimization | Accept (poster) | Summary: The authors propose a novel task-aware model learning scheme for model-based RL, which combines the idea of value aware losses with bilevel optimization to obtain a stable training objective from SOTA pixel-based RL approaches such as Dreamer. The method uses the VAML loss as a surrogate objective for learning... | Rebuttal 1:
Rebuttal: Dear Reviewer YURA,
We sincerely appreciate your detailed feedback and constructive suggestions! Here is our response to your concerns:
---
**Q1: The authors (correctly) note that VAML (as is) is inapplicable in a POMDP setting, but propose a method that relies on inferring value functions from ... | Summary: The authors present an elegant approach to combine meta-learning and world models. The authors cover the components of their Task-aware Environment Modeling Pipeline with Bi-level Optimization (TEMPO) approach. They then cover experiments across continuous and discrete domains, finding that TEMPO meets or exce... | Rebuttal 1:
Rebuttal: Dear Reviewer RGcg,
We sincerely thank you for giving us encouraging comments and recognizing the contribution of our work. Here are our answers to your questions:
---
**Q1: While DreamerV2 is a strong baseline, there have been other world model approaches released after it with improved perform... | Summary: This paper proposes a novel model-based algorithm termed Task-aware Environment Modeling Pipeline with Bi-level Optimization (TEMPO). To achieve a proper trade-off between MLE and Value-Equivalence models, TEMPO utilizes a bi-level optimization framework that trains an upper-level weight network generating non... | Rebuttal 1:
Rebuttal: Dear Reviewer kk1X,
Thank you for your detailed and constructive feedback. Here is our response to your concerns:
---
**Q1: The learned representations of DreamerV2 not only rely on the reconstruction loss but the reward prediction and transition prediction that indeed provides task-specific inf... | Summary: The authors propose to combine the benefits of value-based model-based methods (like MuZero) with reconstruction-based model-based methods (like Dreamer). They do so through TEMPO, which is a bi-level model-learning algorithm that includes a meta weighter network that weights each training sample with the goal... | Rebuttal 1:
Rebuttal: Dear Reviewer 9zSq,
We sincerely appreciate your positive comments and feedback, as well as the suggestions for improvement. In response to your questions, we have prepared the answers below:
---
**Q1: I encourage the authors to make more clear what is their contribution and what is contributed ... | Rebuttal 1:
Rebuttal: Dear Reviewers,
Thank you for your comments and suggestions. We hope our response is sufficient to address your questions.
Please kindly let us know if you have any additional concerns.
Paper 3253 Authors
Pdf: /pdf/b5d37febc08afda83f2cfdeee4b1e1e655ae0202.pdf | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
“Why Not Looking backward?” A Robust Two-Step Method to Automatically Terminate Bayesian Optimization | Accept (poster) | Summary: The authors of this paper propose a new stopping criterion for Bayesian optimization. The idea is to consider the last designs visited during optimization to compute a local regret involving another optimization problem. A theoretical analysis is provided as well as an empirical comparison with some alternativ... | Rebuttal 1:
Rebuttal: **W#1:** Thank you for your suggestion, we will add the two papers you mentioned to Section 5 of the main paper. However, both the reviewer mentioned works are not applicable in our context. As for the Jones et al.'s work, it terminates BO when EI is less than 1% of the current optimal value. Howe... | Summary: The paper proposes a new termination criteria for Bayesian Optimization (BO). Early stopping is an important, and understudied problem in BO, as function evaluations tend to be expensive and therefore terminating the optimization (once a good enough solution has been reached) can save a large amount of resourc... | Rebuttal 1:
Rebuttal: **W#1:** We will amend the proofs to improve their readability in the final version. The proof of Lemma 2 is elaborated in the response to **Q#3**, while the intuition of $\mathrm{I}_\mathrm{cdf}$ is explained in the response to **Q#4**.
**W#2:** The choice of the termination thresholds are non-t... | Summary: This paper proposes a two-step termination method for Bayesian optimization. The proposed termination method first identifies if the optimization has converged within convex region, then it would stop the optimization procedure when estimated local regret is below predefined threshold. The authors compute esti... | Rebuttal 1:
Rebuttal: **W\#1:** We do not assume the existence of convexity across the entire search space $\Omega$. Instead, as claimed in the **Proposition 2** of our manuscript, we suppose $\underset{\mathbf{x}\in\tilde{\Omega}}{\operatorname{minimize}}-\sigma^2(\mathbf{x})$ exhibits convexity in its local optimal r... | null | null | Rebuttal 1:
Rebuttal: To facilitate the understanding of reviewers, we provide three figures in the submitted PDF file. Specifically, the Figure 1 corresponds to the response to **pzre Q\#1**, the Figure 2 corresponds to the response to **TmDT Q\#3**, the Figure 3 corresponds to the response to **TmDT Q\#4** and **ffw... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
CellPLM: Pre-training of Cell Language Model Beyond Single Cells | Reject | Summary: This paper introduces a new pre-training strategy that takes cells as tokens and tissues as sentences. It also encodes cell-cell relations by leveraging the spatial information of cells acquired from spatially-resolved transcriptomic data. The proposed method achieves state-of-the-art performance in various do... | Rebuttal 1:
Rebuttal: > W3. Why not evaluate the most common and important task like cell-type classification?
A1. Thanks for the suggestion. We did not introduce cell-type classification as a benchmark task because the ground-truth labels of cell-type annotation depend on the specific dataset and we didn't know which... | Summary: This paper proposed a pre-trained model based on single-cell data motivated by the special characteristics (i.e., bag of genes structure, cell-cell relation, noisy) in cell-data, which viewed cells as tokens and tissues as sentences. This is different from the most existing pre-trained models treat genes as to... | Rebuttal 1:
Rebuttal: > W1. The technical details of pre-trained data process are not clear enough.
A1. We have indeed presented a reference of pre-training dataset in Appendix D.2. Specifically, the scRNA-seq data consist of 4.7 million cells from human tumor cell atlas, 1.4 million cells from human cell atlas, and ... | Summary: The paper proposes a scRNA-seq model that takes into account correlations between gene expression values in nearby cells in a tissue. The proposed model represents cells as tokens with a very simple embedding model. The tissue representation is modelled using self-attention between cells. The model is trained ... | Rebuttal 1:
Rebuttal: Thank you for providing detailed feedback and helpful suggestions. To address your concerns, we have provided explanations as follows.
> W1 & Q1-a. For the denoising experiments are you using the transformer attention with noisy cells? Because that seems hardly fair to the other methods.
A1-a. Y... | Summary: The paper presents a study on the problem of enhancing cell representation learning through pretraining. Inspired by the recent success of pretrained language models, the researchers explore the use of Transformer architecture to achieve improved cell representations.
In contrast to previous approaches that o... | Rebuttal 1:
Rebuttal: Thank you for your constructive suggestions! Below are responses to address your key concerns.
> W1& Q1 It would be valuable to observe how the proposed model performs on tasks that involve multi-cell information. Are there other datasets requiring multi-cell modeling?
A1. Thanks for your suggest... | Rebuttal 1:
Rebuttal: General Response
We thank five reviewers for their detailed and thoughtful comments. We have improve our manuscript on the following aspects.
**i) Comparison with Other Pre-trained Models on Cell-type Annotation Task**
We now add a new experiment on cell type annotation tasks on Zheng68k[1] be... | NeurIPS_2023_submissions_huggingface | 2,023 | Summary: The CellPLM paper presents a novel pre-trained model that encodes cell-cell relationships in spatially-resolved transcriptomic data. The authors demonstrate that their model outperforms existing state-of-the-art methods in various downstream tasks, including cell type classification, cell-cell interaction pred... | Rebuttal 1:
Rebuttal: Thank you for your detailed comments and constructive suggestions. Below are responses to address your key concerns.
> W1-a. It is unclear to me why to use a Gaussian Mixture Latent Space. The distinct functional groups of cells have not been explained clearly.
A1. We would like to clarify that ... | null | null | null | null | null | null |
EmbodiedGPT: Vision-Language Pre-Training via Embodied Chain of Thought | Accept (spotlight) | Summary: The given text discusses the field of embodied AI and its applications in robotics. It highlights the potential of large language models (LLMs) in enabling robots to process natural language instructions, perform multi-modal tasks, and plan actions in physical environments. However, the collection of high-qual... | Rebuttal 1:
Rebuttal: We are immensely grateful for your invaluable efforts and insightful suggestions, which have been of great significance and played a crucial role in the improvement of our paper.
**Q1.[Potential overfitting to specific domains]**
Great question! Your insight highlights the very reason we constru... | Summary: This paper proposes an end-to-end multi-modal foundation model for embodied AI. Specifically, the authors create a dataset termed EgoCOT from the Ego4D dataset and introduce a paradigm for both the vision and embodied tasks using large language models. However, I have some concerns about this paper. My detaile... | Rebuttal 1:
Rebuttal: Thanks for your efforts to improve the quality of our paper. We appreciate the flaws you identified in our paper's presentation and we believe they have been thoroughly clarified and addressed as follows.
**Q1.[Limited novelty on architecture.]**
We would like to clarify that our approach's nove... | Summary: In this work, the authors propose an end-to-end multi-modal model for Embodied AI field. Based on the Ego4D dataset, the authors build a large-scale Embodied planning dataset, i.e., EgoCOT, which contains machine-generated language instruction with chain-of-thoughts mode. Besides, an EmbodiedGPT based on LLM i... | Rebuttal 1:
Rebuttal: We sincerely appreciate the time and effort you have invested in reviewing our paper. Your meticulous evaluation and insightful recommendations have significantly contributed to improving the quality of our research.
**Q1.[Evaluating on vision-language navigation (VLN) tasks to test the navigatio... | Summary: This paper introduces EmbodiedGPT, a new multi-modal foundation model designed for application to embodied (robotic) tasks. EmbodiedGPT combines existing large pretrained (and frozen) vision and language models (namely the ViT-G/14 EVA-CLIP vision model and LLaMA-7B) through the use of an "Embodied-Former" whi... | Rebuttal 1:
Rebuttal: We are deeply grateful for your careful review and meticulous line-by-line suggestions!
**Q1. [It is hard to know how representative the 20 tasks are of general captioning/planning tasks, how were these chosen?]**
To ensure comprehensive representation, the tasks are selected across real-world a... | Rebuttal 1:
Rebuttal: # General Response
We sincerely appreciate all reviewers’ time and efforts in reviewing our paper. We are glad to find that reviewers generally recognized our contributions:
* **Model Development.** Taking a pioneering step in developing a foundational model for embodied AI [BD8Q, P2YP], integrati... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Initialization Matters: Privacy-Utility Analysis of Overparameterized Neural Networks | Accept (poster) | Summary: This work aims to connect the structure of a neural network and its initialization scheme to the privacy guarantees it can offer. KL divergence is used as privacy metric, under the settings of Langevine diffusion.
Strengths: The paper aims to solve a very important problem, i.e. the effects of overparameteriz... | Rebuttal 1:
Rebuttal: We thank the reviewer for pointing out the typos and repeated symbols. Below we clarify the questions and concerns.
> same line: $m = \Omega(n^{14}L^{16}\phi^{-4})$. To me this assumption looks severely irrealistic for any deep learning model.
This condition comes as is from the prior analysis [... | Summary: This paper presents the KL privacy bound for training fully connected networks and its linearized variant using the Langevin diffusion algorithm. The study delves into the impact of network width, depth, and initialization on privacy. Notably, the paper reveals that, with specific parameter initialization, the... | Rebuttal 1:
Rebuttal: > The paper lacks numerical experiments to validate the disparity between empirical KL privacy leakage and the privacy bound proposed in this paper. This would shed light on the tightness of the bound. Given that the arguments are based on the upper bound, providing numerical experiments demonstra... | Summary: This paper theoretically analyzes how the width, depth, and initialization of deep neural networks affect the privacy utility trade-off. For the special setting of linearized networks, the paper shows that gradient norm and privacy loss hinge on the per-layer variance of the initialization distributions. Thus,... | Rebuttal 1:
Rebuttal:
We thank the reviewer for the detailed comments and pointers to typos, which we clarify below.
---
> [Definition for KL privacy]
A randomized algorithm $\mathcal{A}$ satisfies $\varepsilon$-KL privacy if for any neighboring datasets $D$ and $D'$, we have that the KL divergence $KL(\mathcal{A}... | Summary: The authors propose a KL privacy bound for DP-SGD that depend on the width, depth, and initialisation of the neural network. They show that the privacy bound improves with increasing depth for the LeCun and Xavier initialisation while it degrades with increasing depth.
Strengths: 1. The authors came up with ... | Rebuttal 1:
Rebuttal: We thank the reviewer for spotting the typos, we will correct them in revisions of the paper. Below we clarify the questions and concerns.
---
> Often there are mistakes in privacy proofs. Having some experiments with privacy auditing to back up the surprising results would make the results mor... | Rebuttal 1:
Rebuttal: # General response
We’d like to thank the reviewers for their insightful comments. Below we address the shared questions and concerns.
---
**1. Intuition for improving privacy bound**
On the one hand, when depth is fixed, increasing width $m$ always increases the gradient norm and thus harms o... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Towards Label Position Bias in Graph Neural Networks | Accept (poster) | Summary: This paper presents a study on Label Distance Bias in Graph Neural Networks (GNNs). The authors investigate the performance disparity across various GNN models on three datasets and analyze three metrics related to Label Distance Bias. The study reveals that Label Position Bias is a significant factor affectin... | Rebuttal 1:
Rebuttal: Dear reviewer,
We appreciate your constructive feedback. We are pleased to provide detailed responses to address your concerns.
**Q1: I cannot determine the actual performance of your work as you have not provided any code.**
**A1:** Thank you for highlighting your concern. We apologize for any... | Summary: This paper demonstrates a new bias in GNN models called label position bias. A metric called the Label Proximity Score is introduced to quantify the bias. The paper proposes a framework to learn an unbiased graph structure, which can be applied to existing GNN models.
Strengths: Label position bias is an inte... | Rebuttal 1:
Rebuttal: Dear reviewer,
We appreciate your constructive feedback. We are pleased to provide detailed responses to address your concerns.
**Q1: Why the label position bias phenomenon occurs?**
**A1:** Thanks for your thorough consideration. We attribute the root cause of label position bias to the inhere... | Summary: This paper pointed out that message-pass based GNNs training usually suffers from label position bias problem for semi-supervised node classification especially for quite few-shot settings. That is, the unlabeled nodes which far away from the labeled ones might suffer from performance degradation.
The paper ... | Rebuttal 1:
Rebuttal: Dear reviewer,
We appreciate your constructive feedback. We are pleased to provide detailed responses to address your concerns.
**Q1: What is the root to cause the label position bias in GNNs? The performance degradation might due to shallow GNNs or deep GNNs suffering from over-smoothing or ove... | Summary: This paper presents a new bias problem on graph neural networks: label position bias, which implies that nodes perform better closer to labeled nodes. A new metric - the Label Proximity Score(LPS) is proposed to quantify this bias, and a new optimization framework is proposed to learn a new unbiased graph stru... | Rebuttal 1:
Rebuttal: Dear reviewer,
We appreciate your constructive feedback. We are pleased to provide detailed responses to address your concerns.
**Q1: What is the difference between the Topology-Imbalance Node Representation Learning (TINL) and our proposed Label Position Bias (LPB) problem?**
**A1:** **TINL*... | Rebuttal 1:
Rebuttal: # Global Response
We thank the reviewers for the valuable comments and suggestions. We carefully revise our paper and provide responses to address the concerns. In this global response, we are willing to provide information about tables and figures in the rebuttal pdf file and some common concern... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
QuACK: Accelerating Gradient-Based Quantum Optimization with Koopman Operator Learning | Accept (spotlight) | Summary: This paper utilizes Koopman operator theory to accelerate the optimization of quantum circuits. Stability analysis is performed and the benefit of alternating between standard and Koopman optimization is discussed. A number of experiments are performed, across several different areas of interest to the quantum... | Rebuttal 1:
Rebuttal: This paper shows, via a number of experiments, ...
The Appendix contains many details on quantum computing and the architectures used (the availability of this information should be mentioned in the main text).
**We thank the reviewer’s appreciation of our contributions.**
Weaknesses:
While The... | Summary: This paper proposes QuACK, which leverages Koopman operator learning theory to accelerate gradient-based quantum optimization and machine learning. QuACK utilizes the sliding window DMD to enrich the embedding, and neural networks to learn and predict the nonlinear dynamics. Experiments on quantum Ising model,... | Rebuttal 1:
Rebuttal: Weaknesses:
Θ(td+1)≈KΦ(Θ(t0)) in page 5 seems to lack theoretical support, since the koopman operator is Kg(x(t))=g(x(t+1), which is different from the given formula.
**Thanks for the question. This is a good point to note. As the reviewer points out, this is different than the conventional Koopm... | Summary: The paper presents an innovative framework known as Quantum-circuit Alternating Controlled Koopman learning (QuACK), which effectively combines Koopman operator theory and natural gradient methods in quantum optimization. Through the utilization of an alternating algorithm, QuACK allows for the efficient predi... | Rebuttal 1:
Rebuttal: Strengths:
The paper's strength lies in its contribution to accelerating the optimization of variational quantum algorithms (VQAs), which is crucial for leveraging the capabilities of modern quantum machines in practical problem-solving. The proposed QuACK addresses this challenge and presents int... | Summary: The paper presents an algorithm that effectively "extends" $k$ gradients updates to $k'$ gradient updates by fitting a neural koopman operator. The additional cost is made worth it because the algorithm is derived in a quantum machine learning setting where quantum gradients are expensive and limited and clas... | Rebuttal 1:
Rebuttal: Strengths:
Originality: people don't use koopman operators enough, using them as a framework to perform a -step extension of expensive gradient steps is very clever and I might steal the idea for some of my own work post review period Quality: Argument is clear and sufficient analysis is done to... | Rebuttal 1:
Rebuttal: **We thank the reviewers’ comments and suggestions. Reviewer M6jg acknowledges originality, clarity of our argument, “full marks” on our experiments, experiments on real hardware, and significance of the results. Reviewer kkUu acknowledges “intriguing findings” in approaching the limitations of cr... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Goal-Conditioned Predictive Coding for Offline Reinforcement Learning | Accept (poster) | Summary: Update 8/24: As discussed with the authors, I'm updating my score to "weak accept."
-----------------------
This paper studies the role of representation learning for goal-conditioned behavioral cloning methods, focusing on those representations produced by sequence models. The paper fines that a statewise a... | Rebuttal 1:
Rebuttal: Thank you for your constructive feedback. We address your concerns in detail below and will update our paper accordingly:
>***Some of the confusion is from the different masking strategies, which I didn't see explicitly defined anywhere.***
We discussed the masking strategies in section 3.3 and a... | Summary: This work introduces Goal-Conditioned Predictive Coding (GCPC) – a goal conditioned policy learning method that incorporates trajectory representations learned by a bi-direction transformer. The authors introduce a two-phase framework for goal-conditioned learning, consisting of a trajectory encoding phase, an... | Rebuttal 1:
Rebuttal: Thank you for your constructive feedback. We address your concerns in detail below and will update our paper accordingly:
>***Make it clear when numbers are quoted from published results***
We completely agree! We will incorporate the reviewer’s suggestion and explicitly explain if a baseline’s p... | Summary: The paper presents a perspective on RL via Supervised Learning (RvS) algorithms that identifies two steps—trajectory representation and policy learning—and proposes a specific algorithm for offline RL called Goal-Conditioned Predictive Coding (GCPC) based on this view. The trajectory encoder is modelled as Bid... | Rebuttal 1:
Rebuttal: Thank you for your constructive feedback! We address your concerns in detail below and will update our paper accordingly:
>***More seeds are needed for ablation experiments.***
We reported ablation experiments on FrankaKitchen in the supplementary materials Table A3 and A4. In Table R1, R2, and ... | Summary: The manuscript presents a novel framework for formulating decision making as a supervised learning problem on offline-collected trajectories. Contrary to previous methods, the authors propose to decompose the process into two stages: a) a trajectory representation learning stage (using sequential modeling tech... | Rebuttal 1:
Rebuttal: Thank you for your constructive feedback! We address your concerns in detail below and will update our paper accordingly:
>***The numbers of trajectory rollouts***
In our experiment, we regard IQL[a] as a robust value-based baseline, which has demonstrated its effectiveness across a diverse arra... | Rebuttal 1:
Rebuttal: # General Response to All Reviewers
We thank all reviewers for their insightful suggestions and constructive feedback. We are glad that the reviewers found our work to be “interesting”, “novel” (EZxU, tnif, 3wNA), offers a “nice” and “fresh” perspective on existing RL methods (tnif, EzxU), and th... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Leveraging sparse and shared feature activations for disentangled representation learning | Accept (spotlight) | Summary: The paper proposes a novel regularization technique for disentangled representation learning. This regularizer is based on two principles formalized in the paper: 1) sparsity and 2) minimality. It assumes access to multiple tasks during training where each task uses only a subset of features, and different tas... | Rebuttal 1:
Rebuttal: We thank the reviewer for all their suggestions, questions and comments. We will address them in the following. We remain available during the discussion period for any further question or concern.
### **Clarity**
We thank the reviewer for pointing out unclear sections and typos.
Concerning P... | Summary: The paper studies the problem of learning disentangled representations in a setting with access to multiple supervised tasks, each of which is assumed to depend only on a subset of the underlying factors of variation. While the factors of variation are never directly observed, two principles are incorporated t... | Rebuttal 1:
Rebuttal: We thank the reviewer for all their suggestions, comments and questions. We will address them in the following. We remain available during the discussion phase to further clarify any question or concern.
### **Metalearning and baselines**
For comparison with the ERM baseline we detach the origi... | Summary: The foundation of the proposed work is a form of regularization motivated by the problem of disentanglement. A supervised task is deconstructed into randomly generated sub-tasks, and the same trained representations are used for all sub-tasks (followed by a linear head specific to each sub-task). Two regular... | Rebuttal 1:
Rebuttal: We thank the reviewer for all their suggestions, comments and questions. We will address them in the following. We remain available during the discussion phase to further clarify any question or concern.
### **Disentanglement**
The reviewer is correct in that each task does not depend on all t... | Summary: The paper proposes to learn disentangled representations by leveraging knowledge from a diverse set of supervised task objectives. Specifically, the paper shows that disentangled representations can emerge without ground-truth annotations of the factors of variation under two assumptions: 1) sparse sufficiency... | Rebuttal 1:
Title: Response to Aoq2
Comment: We thank the Reviewer for all the positive comments and suggestions. We will include the suggested modifications in the main paper, in particular, we will make sure to polish all captions to make them self-contained. | Rebuttal 1:
Rebuttal: # General Response
We thank all reviewers for their valuable insights, constructive feedback, and inquiries. In particular, we're glad that all reviewers agree on the value of the contribution, the soundness of the idea, the usefulness of the proposed properties, clarity (reviewers Aoq2,BjS9) and... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Connecting Multi-modal Contrastive Representations | Accept (poster) | Summary: The SOTA approach in multi-modal learning employs contrastive loss to project embeddings from different modalities into a unified embedding space. However, models trained on different modal pairs result in disparate embedding spaces. This paper introduces a novel method to align distinct multi-modal embedding ... | Rebuttal 1:
Rebuttal: ## Clarification of our core innovation and the real reason our method works:
The really effective part of our method is not constructing pseudo audio-image pairs using CLIP and CLAP. Instead, our methods focus on learning the alignment between text representations from CLIP and CLAP in a new spa... | Summary: This paper introduced a novel technique to learn multimodal contrastive representations (MCR) without paired data from the two modalities by using a third modality to bridge between the two existing modalities. The technique assumes the existence of a learned MCR between each existing modality and the third mo... | Rebuttal 1:
Rebuttal: **W1 Apply the proposed methods to other modality settings:**
Please refer to “response to all reviewers” for a detailed comparison and analysis of the experiment on the 3D-text field. Briefly, our method also achieves state-of-the-art performance on the 3D-text field.
**W2 Performance of the le... | Summary: This paper presents a Connecting Multi-modal Contrastive Representations (C-MCR) scheme that aligns the spaces of two already pre-trained MCR models by utilizing the overlapping modality between two, in a data- and training-efficient way. One interesting property of C-MCR is that aligning those two MCR spaces ... | Rebuttal 1:
Rebuttal: **W1 Whether trained projectors can improve the alignment between image and language/audio and language:**
As depicted in the visualization provided in the newly submitted PDF files, the original CLIP and CLAP space (Figure (a) ) exhibits a noticeable modality gap. After remapping by the trained ... | Summary: - This paper proposes C-MCR, a multimodal contrastive framework that aligns multimodal representations of overlapping modality pairs to learn effective representations of the non-overlapping modality pairs. Specifically, the authors concentrate on learning the inherent semantics of visual-audio data solely thr... | Rebuttal 1:
Rebuttal: **W1: Too informative Figure 1**
Thank you for your constructive suggestion. We will add subsection numbers to the subfigures in Figure 1 and make the necessary adjustments to the reference statements in the main text accordingly.
**W2 & Q1: Generalize our approach to different and more modaliti... | Rebuttal 1:
Rebuttal: # Response to all reviewers
**1 Experiment of 3D-text**
1.1 Performance comparison:
By connecting text-image representations of CLIP and image-3D representation of ULIP 2 (Objaverse), we successfully learned a contrastive 3D-text representation without explicitly using any 3D-text data. We eval... | NeurIPS_2023_submissions_huggingface | 2,023 | Summary: This paper focuses on extending existing Multimodal Contrastive Representation (MCR) to more modalities without massive high-quality data pairs. To this end, they propose a training-efficient method for learning a new MCR without paired data. Specifical, they connect one existing MCR pre-trained on modality $(... | Rebuttal 1:
Rebuttal: **W1: The main results are not sufficient:**
Please refer to “response to all reviewers” for the experiment of learning 3D-text representation via image-text and image-3D pre-trained model.
**W2: Unfair comparison experiments:**
Thank you for your suggestion. We will add a column in Table 2 and... | null | null | null | null | null | null |
SHAP-IQ: Unified Approximation of any-order Shapley Interactions | Accept (poster) | Summary: This paper provided an unified perspective to several existing Shapley-based interaction attribution methods as Cardinal Interaction Index (CII), in addition provide a more efficient sampling based approximation method. Connection between their formulation with unbiased KernelSHAP is established and they showe... | Rebuttal 1:
Rebuttal: We sincerely thank the anonymous reviewer for carefully reading and appreciation of our submitted work.
We will follow your suggestions and add the reference for CII [11] already in line 33-34 and refer to the definition of $\mathcal S_{s_0}$ (given in line 84) in line 129 to improve readability. | Summary: This paper presents a unified approximation technique for Shapley interaction indices. It reformulates the general form of various Shapley interaction methods (Shapley interaction index, Shapley-Taylor, Faith-Shap) from a sum involving the set in question to a sum not involving the set, but instead a computati... | Rebuttal 1:
Rebuttal: We sincerely thank the anonymous reviewer for carefully reading and appreciation of our submitted work as highly relevant in terms of theoretical and empirical contributions.
- Could you clarify, is this method an application of a technique used to approximate Shaley Values, but applied to intera... | Summary: This paper presents a novel method, SHAPley Interaction Quantification (SHAP-IQ), for efficiently computing Shapley interaction indices in the context of explainable artificial intelligence (XAI). Shapley interaction indices extend the concept of Shapley values, a method from cooperative game theory, to measur... | Rebuttal 1:
Rebuttal: We would like to sincerely thank the anonymous reviewer for their thoughtful and thorough assessment of our submitted work. We greatly appreciate the detailed list of strengths and contributions of our submitted manuscript.
We strongly believe that the utilization of feature interaction scores wi... | Summary: The paper proposes a new sampling-based approximation method called Shapley interaction quantification(SHAP-IQ) for the cardinal interaction index(CII), which is a generalization of the Shapley interaction index. To this end, the authors give a novel representation, and showed that SHAP-IQ is an unbiased and p... | Rebuttal 1:
Rebuttal: We sincerely thank the anonymous reviewer for carefully reading and providing helpful feedback.
- Shapley interaction index implies that the number of features to be observed exponentially increases, which in turn may make interpretation more difficult. In some cases, multiple sets of features ma... | Rebuttal 1:
Rebuttal: We would like to thank the anonymous reviewers for carefully reading and evaluating our manuscript. As a response to reviewer wrD8, we have uploaded a run-time plot showing the computational effort of SHAP-IQ compared to baseline methods on varying budgets.
Pdf: /pdf/de1d9aa37767cc1a233d8a43746b9b... | NeurIPS_2023_submissions_huggingface | 2,023 | Summary: **Context.** This paper aims at quantifying the contribution of a set of features to the prediction of a machine learning model by leveraging the Shapley value (SV), a principled tool coming from cooperative game theory. Previous works either focused on (i) estimating the contribution of a single feature (via ... | Rebuttal 1:
Rebuttal: We would like to thank the anonymous reviewer for carefully reading and evaluating our submitted manuscript.
- Theorem 4.3 is quite classical since SHAP-IQ stands for a Monte Carlo approximation and proofs techniques to obtain high-probability deviation bounds via Chebyshev inequalities have been... | null | null | null | null | null | null |
Estimating Causal Effects Identifiable from a Combination of Observations and Experiments | Accept (poster) | Summary: The paper presents a method for causal effect estimation through a combination of experimental and observational data.
Strengths: - The paper addresses a very important problem
- The solution appears to be mathematically sound
- Experimental results appear to show improvement over other methods
Weaknesses:... | Rebuttal 1:
Rebuttal: We appreciate your feedback, thank you.
---
> W1. The paper is incredibly hard to parse. Pages 4-8 are just a wall of mathematical notation with little to no explanation.
> W2. The paper would improve significantly if after each definition / lemma / proof a plain word intuition was given.
Give... | Summary: This paper presents a new framework for estimating the causal effect when multiple observational and experimental datasets are available. The authors proved that any g-identifiable causal effects can be written as a functional of "g-mSBD" operators, based on which they further developed a doubly robust estimat... | Rebuttal 1:
Rebuttal: We thank the reviewer for the valuable feedback provided and hope that the answers below help to address the raised issues. Still, we would be happy to provide further clarification if you find that this was not the case.
---
> W1. The first regards the assumption imposed in their Proposition 2 ... | Summary: The authors study a class of causal effect that are called “g-identifiable” (gID) following Lee et al. (UAI 2019). There are two main results.
1. Expressing these causal effects as transformations of generalized multi outcome sequential back door adjustments (g-mSBDs).
2. Proposing doubly robust estimators for... | Rebuttal 1:
Rebuttal: Thank you for your feedback and for the opportunity to provide further elaboration. We respectfully ask the reviewer to reconsider our paper based on the clarifications provided below. If there are any points that remain unclear, we would be glad to offer additional explanations.
---
> Weakness ... | Summary: The paper aims to fuse the recent literature on causal effect identification and estimation. It expresses any identification (in the general sense) of a causal effect in terms of generalized mSBD adjustments, and then develops a corresponding estimator. For the latter robustness properties and empirical studie... | Rebuttal 1:
Rebuttal: We thank the reviewer for the time and valuable feedback, and appreciate the positive assessment of our work.
---
> W1. I believe the contribution to be valuable but the methods are not novel; they seem to be iterative applications of existing ones. In many places it's too dense: for example, Ex... | Rebuttal 1:
Rebuttal: Thank you for your valuable feedback. We responded to each review individually. Here, we uploaded an enlarged plot for our simulation result (Figure 2).
Pdf: /pdf/270c164cc1d11213f6e544d6e127f54a1c07416f.pdf | NeurIPS_2023_submissions_huggingface | 2,023 | Summary: This paper builds on prior work in causal effect estimation, from observational as well as experimental data, by providing a more general framework with desirable properties. This more generalized framework is shown to be robust and consistent in its estimation and overcomes the limitations of some of the prio... | Rebuttal 1:
Rebuttal: Thank you for sharing your thoughts and feedback!
---
> W1. It is not clear what challenges remain for g-identification in practice with the proposed new estimator. Some guidance for practitioners and other researchers as to the context of problems and types of data they can consider using the n... | null | null | null | null | null | null |
Evaluating and Inducing Personality in Pre-trained Language Models | Accept (spotlight) | Summary: The authors study how personality manifests in large language models (LLMs). The authors first construct a machine personality inventory (MPI), which is a set of questions designed to measure the personality of a language model on the five OCEAN axes. The authors then construct a set of personality prompts int... | Rebuttal 1:
Rebuttal: > it is still a bit of a stretch -- again, in my opinion -- to make a claim that a model indeed has an inherent personality (specifically for the zero-shot experiments).
We believe that back to the first time the concept of personality was introduced as a measurement of human behaviors, psycholog... | Summary: This paper investigates approaches to use psychometric tests in order to elicit personality traits from large language models (LLMs). To do so, the authors introduce the Machine Personality Inventory (MPI), a collection of multiple-choice questions, to assess an LLM’s personality. The MPI is based on the Big F... | Rebuttal 1:
Rebuttal: > To fully assess whether these statistics are equal between groups (i.e., humans vs LLMs), rigorous statistical testing (incl. measuring effect sizes) is essential, also because the two compared groups have different sample sizes (there are 620,000 human responses).
Yes, unlike other cognitive/d... | Summary: This paper focuses on investigating human-like personality in Language Model Models (LLMs) and assessing the feasibility of personality control in these models. To establish the presence of human-like personality, the study introduces the Machine Personality Inventory (MPI) dataset and evaluate LLMs from 5 asp... | Rebuttal 1:
Rebuttal: Thank you immensely for your comprehensive and constructive review of our paper. Your insightful feedback and recognition of our work's novelty are highly appreciated!
> Conducting additional experiments to demonstrate LLMs' ability to resist personality induction to some extent could strengthen ... | Summary: The authors introduce the topic of Machine Personality using the Big Five Personality Traits. They propose a new evaluation dataset called Machine Personality Inventory (MPI) to evaluate the existence of personality that are shown by modern LLMs. Big Five categorizes human personality using five key traits: Op... | Rebuttal 1:
Rebuttal: > Controlling the personalities with prompts are quite naive and not sure if it works in reality. Use LoRA to few-shot fine-tune on certain personality traits to see better performance.
Thanks for your suggestion! We use prompting since it is one of the most universal methods for tuning large lan... | Rebuttal 1:
Rebuttal: # Overall Response
We thank all reviewers for their time and valuable comments. The feedback is substantial and helpful for improving our paper. In this work, we are inspired by the theoretical propositions and the behavioral observations of human personality; we explore whether pre-trained langua... | NeurIPS_2023_submissions_huggingface | 2,023 | Summary: This paper dives into an interesting territory: the personalities of Large Language Models (LLMs). The researchers inquire whether LLMs possess discernible personalities and how these can be evaluated or even influenced. They introduce the Machine Personality Inventory (MPI), based on well-established personal... | Rebuttal 1:
Rebuttal: > W (1) Human personality is complex and multifaceted, and this model may not fully capture the nuances of personality.
We agree with the statement, as factorizing and disentangling human personality is still an open problem in the field of personality/social psychology. However, we'd also like ... | null | null | null | null | null | null |
Formulating Discrete Probability Flow Through Optimal Transport | Accept (poster) | Summary: The paper answers two important questions for the diffusion modelling community:
(1) How to interpret the encoding given by the ODE transport map of a diffusion model - it is proven to be an optimal transport map.
(2) How can this ODE type integration be translated to discrete diffusion models - a low stoch... | Rebuttal 1:
Rebuttal: > **The paper has not provided any experiments to demonstrate the results on continuous state spaces however I believe these would mostly be similar to Khrulkov et al.**
Thank you for your comments. In [1], Khrulkov et al. have demonstrated the results on continuous state spaces through numerical... | Summary: Building upon the concept of probability flows, the paper establishes a clear link between diffusion processes and optimal transport. It first shows that the continuous probability flow corresponds to a Monge map for every finite-length time interval. It then states and proves analogous statements in the conte... | Rebuttal 1:
Rebuttal: > **Notation is hard to parse ...**
We apologize for not describing the notations. We use $X_t$ and $Y_t$ to represent the OU process and Brownian motion. For consistency, we adopt the notation '^' and '~' to signify the corresponding probability flow and infinitesimal transport. The same convent... | Summary: In this paper, the authors present a theoretical way to define the probability flow for continuous-time diffusion models with discrete state-space, a notion which is largely used for continuous state-space diffusion models, but does not generalize to the discrete setting. They bring several contributions:
- I... | Rebuttal 1:
Rebuttal: > **Scalability of the experiments.**
Please refer to our general response.
> **In Proposition 6, the existence of a Kantorovitch plan in the discrete setting is not provided for the whole time interval. Do you observe empirically any change in the transport of the particles on the whole time in... | Summary: The paper presents a formulation of discrete probability flow using tools of Optimal Transport. First establishing conditions under which continuous probability flows are OT maps, the authors establish a version for discrete case. The authors then define a discrete probability flow using the connection to OT. ... | Rebuttal 1:
Rebuttal: > **Unclear expression "This holds under the 2..." in Line 102.**
As the gradient flow of the same functional is different under different metrics, we would like to emphasize here that Eq. (13) is the gradient flow of Eq. (14) under the $W_2$ metric. Thanks for your advice. We will refine our exp... | Rebuttal 1:
Rebuttal: # General Response to All Reviewers
We thank all the reviewers for their time, insightful suggestions, and valuable comments. We are delighted that they find our approach interesting and theoretically sound (**Reviewer Rnpe, wEPe**), establishes a clear link between diffusion processes and optim... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
State Regularized Policy Optimization on Data with Dynamics Shift | Accept (poster) | Summary: The paper proposes a context-aware method called SRPO (State Regularized Policy Optimization) to address the issue of training RL algorithms on data with dynamics shift, i.e., with different underlying environment dynamics. The authors reveal and exploit the property of similar stationary state distributions i... | Rebuttal 1:
Rebuttal: Thank you for your valuable feedback! We provide discussions on your concerns as follows.
Q: **The paper would benefit from additional theoretical analysis on the intuition of similar optimal state distributions.**
A: Thm. 4.3 (lines 258-259) of our paper provides evidence for the existence of s... | Summary: The paper considers leveraging existing data in RL but with dynamics shifts. The proposed method augments the reward with a density ratio estimation term that encourage the policy to recover the optimal state distribution under the distribution shift. The theory motivates the algorithm design by showing that u... | Rebuttal 1:
Rebuttal: Thank you for your valuable feedback! We provide discussions on your concerns as follows.
Q: **The theory section (section 4) only provides some motivation for the algorithm design, instead of actually analyzing the algorithm.**
A: The SRPO algorithm works by augmenting the reward function. Such... | Summary: This paper presents an approach to RL/policy learning in the case where the dynamics parameters are potentially time-varying. The approach relies on the view that the optimal state distribution is similar across (small) variations in the underlying parameters, subject to assumptions. Thus, the authors propose ... | Rebuttal 1:
Rebuttal: Thank you for your valuable feedback! We provide discussions on your concerns as follows.
Q: **The similarity of similar state distributions is said to arise when the dynamics are similar and when the reward functions and dynamics functions are Lipschitz. These conditions are quite strong and the... | Summary: The paper aims to improve sample efficiency for RL algorithms trained in environments with dynamics shift. The core motivation is that, for different environments and their corresponding optimal policies, the state visitation distribution could be similar. Building upon this motivation, the authors propose a c... | Rebuttal 1:
Rebuttal: Thank you for your valuable feedback! We provide discussions on your concerns as follows.
Q: **In equation 5, we use Rmax directly. What if Rmax is unknown or different across states?**
A: In Eq. 5, $R_\max$ serve as a normalization factor to guarantee the probability $p(\mathcal{O}\_t|s\_t)$ is... | Rebuttal 1:
Rebuttal: We extend our sincere gratitude to all the reviewers for providing insightful comments. Both reviewer PXH4 and ZrWb mentioned the absence of discussions on the Lipstchitz coefficients in the assumptions. In this common response, we will address this issue and discuss how Lipstchitz coefficients ca... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Robust Concept Erasure via Kernelized Rate-Distortion Maximization | Accept (poster) | Summary: This paper considers the concept erasure task removing an attribute from distributed representations while retaining other information from the original representation space as much as possible. Specifically, it proposes a Kernelized Rate-Distortion Maximizer (KRaM) framework and provides a theoretical analysi... | Rebuttal 1:
Rebuttal: Thanks for your review and questions. Please find our responses below.
(W1)
We believe there may be a misunderstanding here. And so we would like to clarify things in detail because we firmly believe that our work is not just a simple introduction of kernels to [1]. In fact, **our task and fin... | Summary: The KRaM framework is proposed in this paper to perform concept erasure. The approach aims to differentiate similar concept labels in the learned representation space while preserving other information. The paper demonstrates the effectiveness of KRaM in erasing various concepts across different forms of repre... | Rebuttal 1:
Rebuttal: Thank you for your review and questions. Please find our responses below.
Weaknesses:
(W1) We would like to clarify the claim made in the paper.
We emphasize that the purpose of concept erasure is to make it difficult to predict the deleted concept from the representation set. And indeed, our ... | Summary: The paper proposes a novel method to erase concepts from a distributed/latent data representation. The method is based on a kernelized rate-distortion function that can be maximized to achieve concept erasure. The kernel encodes the "concept similarity" between two samples, and it is therefore supervised. That... | Rebuttal 1:
Rebuttal: Thank you for your review and questions. Please find our responses below.
Questions:
(Q1) We are unfamiliar with and did not get to try alternative formulations of the rate-distortion function. However, even in the non-parametric setting where the representations are available as finite samples... | Summary: This paper studies concept erasure, a task proposed by prior work [1,2] which aims to remove information from a chosen attribute while retaining other information from the original input as much as possible. This paper achieves this by maximizing the kernelized rate-distortion maximizer (KRaM), an objective fr... | Rebuttal 1:
Rebuttal: Thank you for your detailed review. Please find our responses below:
Weaknesses:
(W1) Thank you for pointing out these works. We will remove the specific claim and cite these works. We have compared with the suggested works and the results on categorical concept erasure are reported below:
*DIA... | Rebuttal 1:
Rebuttal: We thank the reviewers for their valuable feedback and suggestions. We believe that we have carefully considered all the comments raised by the reviewers. We have conducted several additional experiments and theoretical analysis as summarized below:
# New Experiments & Results
**New Baselines**:... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Deductive Verification of Chain-of-Thought Reasoning | Accept (poster) | Summary: Despite demonstrating appealing reasoning capabilities, Language Models (LLMs) still suffer from erroneous and hallucinated reasoning. In order to enhance the trustworthiness of LLMs' reasoning process, this paper introduces a deductive reasoning framework based on natural language. The framework extracts prem... | Rebuttal 1:
Rebuttal: We sincerely thank you for your constructive feedback and your appreciation of our work! We address the comments and questions below.
> Q1: Add experiments over more benchmarks like CLUTRR and StrategyQA.
Thanks for your suggestion. For StrategyQA, we sample 1000 instances from the dataset. ... | Summary: This paper propose natural program, which a novel method to make verifiable Chain-of-Thoughts reasoning as well as verify each step. The idea is to first generate reasoning steps in the format of '(# by evidences), caculate...'. And then they prompt another LLM to verify each relatively independent step. The p... | Rebuttal 1:
Rebuttal: We sincerely thank you for your constructive feedback! We address the comments and questions below.
We'd like to first clarify that our primary goal is *NOT* to improve the final answer accuracy of reasoning tasks.
When LLMs leverage CoT to carry out problem solving, hallucinations and accu... | Summary: This paper introduces a format for chain-of-thought reasoning in language models called Natural Programs. In a Natural Program, the model explicitly breaks down the given problem into a numbered list of assumptions. Then, each reasoning step explicitly lists out the premises (starting assumptions or previous s... | Rebuttal 1:
Rebuttal: We sincerely thank you for your constructive feedback! We address the comments and questions below.
> Differences between NLProofS and our work
Thanks for your insightful question and we'll cite the "NLProofS" paper. We'd like to highlight several perspectives that distinguish our work from NLPr... | Summary: The authors propose Natural Program, a suite of prompt-based steps to verify the reasoning process generated by LLM by reusing the LLM itself. The key observation used is that LLMs are good verifiers natively if they are presented with just the necessary set of premises to judge the statement. Any distractor c... | Rebuttal 1:
Rebuttal: We sincerely thank you for your constructive feedback! We address the comments and questions below.
> Circular nature of using LLM for verification
We'd like to clarify that "the main hypothesis here is that we don't trust the generations sampled from LLM" is actually NOT our hypothesis. Not onl... | Rebuttal 1:
Rebuttal: # Common Response
We sincerely thank all reviewers and ACs for their constructive feedback! We're glad that reviewers found our paper well-motivated (`LzaJ, BLD6, TPbw`), our verification framework and ideas plausible (`bmW6`) and good (`kmTd`), our observations intersting (`BLD6`), and our paper... | NeurIPS_2023_submissions_huggingface | 2,023 | Summary: The authors present an approach to build easily verifiable prompted systems by proposing a new reasoning format - Natural Program. In this proposed form, LLMs are expected to generate a rationale but additionally also indicate previous steps that led to each step of reasoning. This allows a verifier to only lo... | Rebuttal 1:
Rebuttal: We sincerely thank you for your constructive feedback! We address the comments and questions below.
> Q1: Include stronger baselines for comparision with "Natural Program".
We'd like to note that our primary goal is to improve the *truthworthiness and reliability* of the reasoning via step-by-s... | null | null | null | null | null | null |
On Robust Streaming for Learning with Experts: Algorithms and Lower Bounds | Accept (poster) | Summary: This paper studied the memory complexity of the expert learning problem against an *adaptive* adversary. It proved a lower memory bound for deterministic algorithm and gave an optimal deterministic algorithm matching this bound. It also proposed a randomized algorithm which uses less memory if the regret and b... | Rebuttal 1:
Rebuttal: > I believe your theorems are correct. However, some parts of your proof doesn't seem rigorous to me or could just be wrong. For example, in line 272, you claim "at least $m/3$ of the instance". I don't know where the constant $1/3$ comes from because you seem to be quite sloppy with the constants... | Summary: The paper studies the experts problem in online learning under an additional memory constraint. Recent development on this setup gave algorithms with a sublinear memory usage (in $n$, the number of experts) and also a sublinear regret (in $T$, the time horizon). However, many of these results require the obser... | Rebuttal 1:
Rebuttal: > The results are restricted to the discrete prediction setup (i.e., with binary outcomes, binary forecasts, and 0/1 loss), whereas some previous work along this line applies to the more general bounded-loss case. (Lines 25-26 mentions a setup with losses bounded in $[0,\rho]$, but all the results... | Summary: This paper considers the problem of individual sequence prediction with a finite pool of experts, under memory constraints. While this problem was considered in previous works when the learner faces oblivious adversaries. The authors consider the case of adaptive adversaries that choose the experts costs in e... | Rebuttal 1:
Rebuttal: > The algorithm requires prior knowledge of the number of mistakes by the best expert and a pre-specified target regret.
We remark that the natural deterministic algorithm only requires a loose upper bound on the number of mistakes by the best expert. Moreover, it is not the focus of our paper. ... | Summary: The paper studies the learning with experts problem with a focus on: 1) the trade-off between the memory usage and the regret; 2) robustness against an adaptive adversary who decides the experts’ predictions and the correct prediction on each time step, based on all previous stream updates and the player’s out... | Rebuttal 1:
Rebuttal: > Most techniques in the proofs are already used in previous works. For example, the proof of the space lower bound is by a reduction to the communication problem of $\epsilon$-DIFFDIST. Similar reduction is already used in the closely related work [49] to show memory lower bound for learning with... | Rebuttal 1:
Rebuttal: We thank the reviewers for their thoughtful comments and valuable feedback. In particular, our experiments show the importance of algorithms that are robust to adversarial input, demonstrating that the existing algorithms perform poorly against a simple attack while deterministic algorithms are a ... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Provably Bounding Neural Network Preimages | Accept (spotlight) | Summary: This paper presents a novel approach to bounding the preimages of neural networks using a sequence of relaxations. Results are provided for a toy controls problem and a robustness verification task.
Strengths: 1. The paper addresses an important recent developement in the literature.
2. The paper is well pres... | Rebuttal 1:
Rebuttal: Thank you for the feedback! Thanks to your review, we have demonstrated our scalability through much more difficult robustness verification (including **8112-dimensional inputs** and **167k neurons**) and tested another reachability analysis problem of a **6-dimensional state quadrotor**.
> Give... | Summary: The authors propose INVPROP, an algorithm for constructing certified bounds on the pre-image of neural networks with ReLU activations. The method constructs a convex over-approximation of the pre-image by implicitly optimizing the dual relaxation of a MILP problem. Empirical evaluation shows considerable impro... | Rebuttal 1:
Rebuttal: Thank you for the helpful feedback! Thanks to your review, we have run more expansive evaluations (including **8112-dimensional inputs** and **167k neurons**), improved our discussion of limitations, and better situated ourselves within prior work.
> I think the authors' contribution would shine ... | Summary: The paper considers verified neural network inversion. Given a set of output values, the goal is to compute the pre-image of these, i.e., the inputs that lie within this output set (given as a linear constraint over the output space).
To address this, the authors propose INVPROP, an algorithm that computes
an... | Rebuttal 1:
Rebuttal: Thank you for the helpful feedback! Thanks to your review, we have run more expansive evaluations (including **8112-dimensional inputs** and **167k neurons**), extended our reachability analysis to a **6-dimensional state quadrotor**, and improved our discussion of limitations.
> Can you discuss ... | Summary: The paper considers the problem of computing approximations of the set of
inputs for which a given neural network satisfies certain output constraints.
To solve the problem, the paper discusses adaptations of a number of approaches
from standard neural network verification, including MILP and LP formulations,
... | Rebuttal 1:
Rebuttal: Thank you for the helpful feedback! Thanks to your review, we have run more expansive evaluations (including **8112-dimensional inputs** and **167k neurons**), **fully implemented branching** from abCROWN with INVPROP, and improved our discussion of limitations.
> At its present stage the method ... | Rebuttal 1:
Rebuttal: We thank the reviewers for their helpful and constructive comments about our paper.
Thanks to the reviews, we have run a large suite of new experiments where we have evaluated our method on robustness verification for **inputs of dimension 8112** and neural networks of **167k neurons**, includin... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Randomized and Deterministic Maximin-share Approximations for Fractionally Subadditive Valuations | Accept (poster) | Summary: This paper considers several classical fair division problems. This paper considers both the deterministic and randomized allocations under the XOS valuations. The XOS valuation is a subclass of the subadditive functions and it generalizes the submodular function. The deterministic allocation is an assignment ... | null | Summary: The paper studies fair allocation under fractionally subadditive (XOS) valuations, where a valuation function is XOS if it can be represented by a maximization among a set of additive valuation functions. Both deterministic allocation and randomized allocation settings are considered in the paper. For the dete... | Rebuttal 1:
Rebuttal: That is an interesting special case to consider. After receiving your review, we found a negative result that exact $MMS$ allocations cannot be guaranteed in this case. For the positive results, to the best of our knowledge, no better approximation factor is know for this particular setting.
---
... | Summary: This paper focuses on the problem of fair allocation of indivisible items to agents with fractionally subadditive valuations. The authors propose both deterministic and randomized allocation algorithms with improved approximation guarantees for the maximin-share objective.
In the deterministic setting, the a... | Rebuttal 1:
Rebuttal: The size of the support is $2$. Due to the space limit, we could not discuss how we convert the fractional allocation into a randomized one. The idea is to construct two allocations such that for all items like $g$ which are fractionally allocated to two agents like $i$ and $j$, $g$ is allocated t... | Summary: This paper studies the problem of fair division of indivisible goods for XOS valuations. The notion of fairness considered here is the well-studied maximin-share (MMS). While it is known that an MMS allocation may fail to exist, even for additive valuations, obtaining positive results for increasingly better a... | Rebuttal 1:
Rebuttal: We appreciate your feedback. We will enhance the writing accordingly for the final version.
Regarding the first question,
- for Additive valuation, the state-of-the-art gives a constructive algorithm which is not polynomial but yields a PTAS.
- In Submodular valuation class, a 10/27-MMS alloca... | Rebuttal 1:
Rebuttal: We found all the comments fair and valid. Therefore, we would like to use this opportunity to thank all the reviewers for the time and effort they dedicated to our paper. We will apply all the comments in the final version. | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
GeoCLIP: Clip-Inspired Alignment between Locations and Images for Effective Worldwide Geo-localization | Accept (poster) | Summary: This paper tackles the image geolocalization problem. Instead of formulating the task as image retrieval or classification over a set of geospatial bins as is typical, the authors frame it as image-to-GPS retrieval by matching image features with location features. The proposed training strategy leverages Rand... | Rebuttal 1:
Rebuttal: We thank the reviewer for acknowledging our problem as “important and interesting”, for finding our approach and our problem formulation to be “novel”, for finding our evaluation to be “extensive and compelling”, and for mentioning our demonstrated applications as “intriguing”. We reply to the que... | Summary: The paper addresses the problem of geo-locating an image, i.e. estimating the latitude and longitude of the where the image was taken.
Previous works treat the problem as a classification problem (of cells distributed across the globe). In contrast the this work proposes to retrieve the GPS coordinates from ... | Rebuttal 1:
Rebuttal: We thank the reviewer for acknowledging our GPS encoding method (“encoding GPS coordinates via random fourier features”) to be “novel”, for mentioning our representation of 2D GPS coordinates using equal earth projection to be “beneficial”, for finding our ablation studies to be “useful”, and for ... | Summary: This paper proposes an image-to-GPS retrieval method, with an image encoder from the pre-trained CLIP and a tailored GPS encoder.
The GPS encoder is implemented by using: i) the Equal Earth Projection to mitigate the GPS distortion; ii) Random Fourier transform with multiple MLPs to obtain high-frequency `po... | Rebuttal 1:
Rebuttal: We thank the reviewer for the encouraging feedback on our section 4.4 and for finding our work to be “well-written and easy to follow”, and our idea to be “interesting and new”. Also, we are thankful to the reviewer for mentioning - “experimental evaluations are strong”. We reply to the queries be... | Summary: This work learns an image-to-GPS retrieval approach where the image is a query, and the gallery
is a collection of GPS coordinates. The authors propose GeoCLIP, which consists of a location encoder, and an image encoder. The retrieval then is based given a new query image against a collection of gps embeddings... | Rebuttal 1:
Rebuttal: We thank the reviewer for the positive feedback on our work: “well-written paper, with good motivation”, “hierarchical learning experiments are solid”, and “good justification for design choices”. We are also thankful to the reviewer for the encouraging remarks about our section ‘Utility of our lo... | Rebuttal 1:
Rebuttal: We sincerely thank all the reviewers for their valuable and encouraging feedback on our work. We are happy to see that all the reviewers have appreciated our work. The reviewers vH7S, AAT4, and SMLq even explicitly mentioned our approach to be novel. Moreover, we are very thankful that our demonst... | NeurIPS_2023_submissions_huggingface | 2,023 | Summary: For global localization task for a query photo, the authors propose a CLIP-inspired image-to-GPS retrieval approach (named ‘GeoCLIP’) where we retrieve the GPS coordinates of an unseen query image by matching it with the gallery of GPS coordinates. Regarding location encoder, some new ideas such as Random Fou... | Rebuttal 1:
Rebuttal: We thank the reviewer for finding our method GeoCLIP (“matching of GPS location embeddings and image embeddings”) as “novel”, for appreciating our location encoder having “new ideas such as Random Fourier Features and Hierarchical Representation”, and also for mentioning our method’s ability to “m... | null | null | null | null | null | null |
Future-Dependent Value-Based Off-Policy Evaluation in POMDPs | Accept (spotlight) | Summary: - This paper proposes a novel future-dependent value function to study the problem of OPE in POMDPs with function approximations.
- A new Bellman-style conditional moment equation is derived for the future-dependent value function, based on which the value function can be learned via finite sample using minim... | Rebuttal 1:
Rebuttal: **Regarding the completeness assumption (2b), how can it be justified in the examples provided in the paper, e.g., HSE-POMDPs? I would appreciate it if the authors could provide more detailed clarifications on that.**
Thank you for the question. Lemma 5 in Appendix D.2 establishes conditions for ... | Summary: The paper introduces a novel model-free approach for off-policy evaluation (OPE) in partially observable Markov decision processes (POMDPs). The authors propose the use of future-dependent value functions (FDVFs) to capture the latent states by leveraging future observations. They provide a Bellman-like equati... | Rebuttal 1:
Rebuttal: ### Rebuttal for the Weakness
**The readability of the paper can be enhanced, particularly by providing a more intuitive motivation and explanation of FDVFs.....**
Thank you for your suggestion. The primary motivation is FDVFs give embeddings of the latent states such that the expectation is t... | Summary: The paper considers off-policy evaluation for POMDPs, where the data is collected using a behavior policy that is only dependent on the observations available to the evaluation policy as well. The work defines a new "future-dependent" value function and its associated Bellman equations that are used to define ... | Rebuttal 1:
Rebuttal: **Staring at Definition 1, I am wondering if the importance-weighted return $\rho G$ extracted from $F$, is a valid instance of $g_{V}(F)$?**
You are right: this would work if $F$ were the infinite future including _all_ future observations, actions, and rewards. Then, this choice of $g_{V}(F)$ e... | Summary: The paper addresses the problem of off-policy evaluation in POMDPs. The main idea is to modify the Bellman equation by incorporating future and historical observations instead of relying on unobserved states. This new future-dependent value function allows for policy evaluation without the need for unobserved ... | Rebuttal 1:
Rebuttal: **My only concern is the practicality of the proposed algorithm. The experimental results are not sufficiently convincing, and conducting more extensive studies, including (1) more baselines and (2) more complex environments, particularly real-world POMDP applications, would greatly enhance the pa... | null | NeurIPS_2023_submissions_huggingface | 2,023 | Summary: The paper proposes a new OPE method for in POMDPs. The paper first introduces a new future-dependent value function and derive a new Bellman equation for the future-dependent value functions. Then, the paper proposes an estimator to learn such future-dependent value function. The paper provides theoretical res... | Rebuttal 1:
Rebuttal: Thank you for your incisive comments. We will review the draft again, and improve the clariy.
**Are all expectations in the paper, when not explicitly specified, taken with respect to the offline data sampled from the data collection policy?**
Yes. We will make it more explicit and reiterate r... | null | null | null | null | null | null |
Contextual Bandits with Knapsacks beyond Worst Cases via Re-Solving | Reject | Summary: This work studies the CBwK problem beyond the worst case scenario, presenting two results regarding the worst-case locations and log rate respectively. The proposed algorithm utilizes a re-solving heuristic that achieves $O(1)$ under full-information and $O(\log(T))$ regret under partial-information with some ... | Rebuttal 1:
Rebuttal: We thank the reviewer for the detailed comments. We will answer your doubts in the following. Please also refer to **CQ1 in the global comment**.
**Q1: "Hard" budget and "soft" budget.**
We believe there lies some misunderstanding here. **This paper does consider "hard" budget constraints**, in... | Summary: This works considers a general setup of Contextual Bandits with Knapsacks (CBwK). The authors identify sufficient conditions under which constant of logarithmic regrets are possible (while previous authors were mainly dealing with $\sqrt{T}$ worst-case regrets). The studied algorithm is rather intuitive and is... | Rebuttal 1:
Rebuttal: We thank the reviewer for the detailed and constructive comments. We will answer your questions according to the logical order. Please also refer to **CQ1-3 in the global comment**.
**Q1: Stopping time of the algorithm.**
Under Assumption 3.1 and full information feedback, $\mathbb{E}[T - T_0]... | Summary: The paper studies stochastic contextual Bandits with Knapsacks. The authors provide an algorithm that guarantees regret smaller than the worst-case $\tilde O(\sqrt{T})$ in non-degenerate instances. In particular, they provides an algorithm that achieves $\tilde O(1)$ regret when the fluid LP has a unique and n... | Rebuttal 1:
Rebuttal: We thank the reviewer for the meaningful comments. The following are our responses to your questions. Please also refer to **CQ1 in the global comment**.
**Q1: The definition of Assumption 3.1.**
Assumption 3.1 supposes that the optimal solution of $J(\mathbf{\rho}_1)$ is unique and non-degener... | Summary: This paper considers the problem of contextual bandits with knapsacks and provides an algorithm that goes beyond worst-case (i.e., provide logarithmic regret) under the mild condition of unique optimal solution and non-degenerate solution. This also simultaneously enjoys an optimal worst-case regret bound when... | Rebuttal 1:
Rebuttal: We thank the reviewer for the constructive comments. The following are our responses to your questions. Please also refer to **CQ1-3 in the global comment**.
**Q1: Rewording of comparison with prior work.**
Thanks for the kind notice. We will adopt your suggestions, correct inaccurate expressio... | Rebuttal 1:
Rebuttal: Here, we will answer some common questions raised by the reviewers.
**CQ1: Technical novelty.**
Our work has the following technical contributions:
1. In proving the worst-case regret (Theorem 2.1), we consider a hybrid benchmark to benefit the analysis, which first appears in literature to o... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Synthetic-to-Real Pose Estimation with Geometric Reconstruction | Accept (poster) | Summary: The authors proposed a geometric reconstruction-based domain adaptation method for synthetic to real transfer of hand/human pose estimation models. Based on the geometric transfer and image generation, the reconstruction loss can help the model generalize to the target domain. The proposed method can benefit f... | Rebuttal 1:
Rebuttal: **Q1: Reconstruction from pseudo-labels for domain adaptation is not entirely new. However, the model in this work used optical flow and occlusion maps for inpainting-based reconstruction, which seems to be quite effective without a complex reconstruction network. However, some implementation of t... | Summary: This paper proposed a domain adaptation method from synthetic to real for 3D hand/human pose estimation. Specifically, two self-supervised losses are designed, one is a geometric reconstruction loss which measures a multi-scale perceptual loss between an image in target domain and a generated one with pairwise... | Rebuttal 1:
Rebuttal: **Q1: Does the occlusion map in Sec 3.3.2, represent the occlusion between hand/human and background? or the self-occlusions between fingers/palm?**
A1: The occlusion does not represent either. Instead, it marks out regions that should be inpainted by the base image, and is applied as a mask to d... | Summary: This paper introduces a self-supervised synthetic-to-real domain adaptive pose estimation method. The approach utilizes a reconstruction-based strategy. It involves geometric transformations with an image generation network to reconstruct a driving image from a base image and uses pixel-wise reconstruction los... | Rebuttal 1:
Rebuttal: We greatly appreciate you highlighting the parts of our paper which could benefit from more explanation and discussion. We will definitely include the suggested points in the revision and hope it will help you view our work more positively.
**Q1: Explanation of StyleNet.**
A1: Thanks for poi... | Summary: This paper addresses the task of learning human body and hand pose estimation from labeled synthetic data and unlabeled data of the target domain. The main idea is to improve the key point estimation pre-trained on synthetic data by reconstructing a video frame using another frame with key point estimation, af... | Rebuttal 1:
Rebuttal: **Q1: Experimental comparisons to reconstruction-based methods [38,11,29,17].**
A1: Thanks for your suggestions. We added experiments on [29] and opted to exclude others due to the inapplicable rendering [38] or additional requirements like 2D labels [11, 17].
Specifically, the rendering in [38... | Rebuttal 1:
Rebuttal: Dear reviewers and ACs,
Thank you for your careful and thoughtful reviews. R1 and R4 commented that the proposed geometric reconstruction for domain adaptive pose estimation is new and novel. All reviewers appreciated the empirical verifications to validate the effectiveness of the proposed metho... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Generalized Balancing Weights via Deep Neural Networks | Reject | Summary: This paper proposes a method for estimating the balancing weights in causal effect estimation. The balancing weights are given by the density ratio of the counterfactual distribution and the observed distribution, and the density ratio can be estimated by solving the optimization of the variational expression ... | Rebuttal 1:
Rebuttal: We thank the reviewer for a detailed and thoughtful review. We apologize for the delay in responding to your questions.
We will incorporate other suggestions raised by the reviewer regarding the presentation and other experimental details in the final version. Below we address the main concerns ra... | Summary: This paper proposes an approach for estimating balancing weights. The ultimate target is to use these weights for the estimation of the casual effects of interventions. The search of balancing weights are formulated as density ratio estimation. To tackle this problem, the authors employ a variational represent... | Rebuttal 1:
Rebuttal: We thank the reviewer for a detailed and thoughtful review. We apologize for the delay in responding to your questions.
We will incorporate other suggestions raised by the reviewer regarding the presentation and other experimental details in the final version. Below we address the main concerns ra... | Summary: This paper derives a novel method of reweighting covariates for the purpose of causal inference. This is done through learning optimal change of measure weights modeled via neural networks and trained using alpha-divergence measures between the true joint distribution and a mutually independent one. Further tw... | Rebuttal 1:
Rebuttal: We thank the reviewer for a detailed and thoughtful review. We apologize for the delay in responding to your questions.
We will incorporate other suggestions raised by the reviewer regarding the presentation and other experimental details in the final version. Below we address the main concerns ra... | Summary: This work presents the way to estimate the balance weight, that represents the causal effects of arbitrary mixture of intervention, by using the neural network. Specifically, authors recognize this balance weight as the density ratio between the source and balanced distributions, and estimates this ratio by op... | Rebuttal 1:
Rebuttal: We thank the reviewer for a detailed and thoughtful review. We apologize for the delay in responding to your questions.
We will incorporate other suggestions raised by the reviewer regarding the presentation and other experimental details in the final version. Below we address the main concerns ra... | null | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Chameleon: Plug-and-Play Compositional Reasoning with Large Language Models | Accept (poster) | Summary: This work presents Chameleon, a system that uses a library of tools or modules to assist a LLM to answer questions. First, a LLM is prompted with a description of available modules, constraints on how the modules should be used, few-shot demonstration examples, and finally the question to answer. The LLM then ... | Rebuttal 1:
Rebuttal: Dear Reviewer,
Thank you so much for your thoughtful review. We are encouraged by your recognition of our paper's originality, extensive experiments, and significant research topic. We appreciate your concerns and your insightful comments to lead to a more impactful paper.
We are committed to a... | Summary: The paper presents Chameleon, an AI system that enhances large language models (LLMs) with plug-and-play modules for compositional reasoning. Chameleon uses various tools, including LLMs, vision models, web search engines, and Python functions, to perform complex reasoning tasks. The system, powered by GPT-4, ... | Rebuttal 1:
Rebuttal: Dear Reviewer,
Thank you for your comprehensive and positive remarks about our work. Your recognition of Chameleon's novel approach, addressing the inherent limitations of LLMs reaffirms the value of our work. Your commendation of the achievements emphasizes the effectiveness of our framework. Y... | Summary: This work proposes Chameleon, an approach to connect large language models to different tools for compositional reasoning. Chameleon is powered by a GPT-4 planner that chains different tools to accomplish the described goal. Chameleon achieve improved performance in comparison with strong baseline models.
S... | Rebuttal 1:
Rebuttal: Dear Reviewer,
Thank you for your thoughtful analysis and encouraging words. We are pleased to know that you appreciated the great presentation of our framework. Your recognition of our solid experiments on two challenging benchmarks. Moreover, we are glad that you perceive our study on how to m... | Summary: The paper introduces Chameleon, a novel AI system that enhances large language models (LLMs) with plug-and-play modules for compositional reasoning. The authors argue that while LLMs have achieved remarkable progress in solving various natural language processing tasks, they have inherent limitations in access... | Rebuttal 1:
Rebuttal: Dear Reviewer,
Thank you for your insightful comments and constructive suggestions. We're grateful for your acknowledgment of our work is among the first works in the field and the distinctiveness by clear evaluation metrics. Your insights on our significant improvements and qualitative analysis... | Rebuttal 1:
Rebuttal: We appreciate the efforts of all reviewers in reviewing our paper. Thank you for the insightful comments that recognize the novelty and contributions of our work, and for the valuable suggestions to enhance our paper. We would like to address the common concerns, emphasize the novelty and contribu... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Understanding and Mitigating Copying in Diffusion Models | Accept (poster) | Summary: This paper aligns with a range of recent works which investigate the tendency of diffusion models to memorize training data and emit this during sampling, when prompted to, and presents strategies to mitigate this for data-processing, training and inference. Other than previous works, which mainly focused the... | Rebuttal 1:
Rebuttal: We sincerely thank the reviewer for your insightful comments and positive feedback on our paper. Your recognition of our rigorous empirical experiments in dissecting training factors, coupled with the acknowledgment of the novelty and significance of our main takeaways, is truly encouraging.
### ... | Summary: This paper discusses the issue of data replication in text-to-image diffusion models, specifically focusing on the memorization problem. It mentions that these models, such as Stable Diffusion, are known for unintentionally replicating their training data, which has led to recent lawsuits. The commonly held be... | Rebuttal 1:
Rebuttal: We thank the reviewer for recognizing the merit of our work. We are pleased to know that you found our experiment setup clear and sound.
### [Toy cases]
We thank the reviewer for this suggestion. In our camera-ready version, we will visualize a toy problem in 2 dimensions where we know the clos... | Summary: The paper tries to explain the memorization and copying of training data in diffusion models. The paper establishes a connection between caption duplications and memorization on top of data duplications analyzed in previous works. In their main experiments, the authors fine-tune various SD models on ImageNette... | Rebuttal 1:
Rebuttal: We thank the reviewer for their thorough feedback. We greatly appreciate your acknowledgment of our experiment design, improved comprehension of the memorization phenomenon, and writing style, and that you found our approach to controlling captions and image complexity analysis insightful.
### [D... | Summary: This paper is a deep dive into the phenomenon of memorization in Diffusion models. I find that the paper is very well position given the current state of generative ML and the paper does a great job at asking the right questions about various factors concerning the memorization of images. Some of the key takea... | Rebuttal 1:
Rebuttal: We are happy to learn that you perceive our contributions as both innovative and crucial in initiating a deeper understanding of memorization.
### [Generation quality]
We provide FID scores in Appendix Table 3 for different mitigation strategies. Irrespective of data duplication, Multiple Captio... | Rebuttal 1:
Rebuttal: We would like to thank **all six reviewers** for their highly constructive feedback! We very much appreciate their assessment that we are "studying an important problem" (ciAp, 4RBY, R9Bm), with "thorough experimentation" (ciAp, qP9c, rsTQ, 4RBY, MEzn, R9Bm), "first to introduce text conditionin... | NeurIPS_2023_submissions_huggingface | 2,023 | Summary: This paper systematically investigates the memorizing phenomenon of stable diffusion. As they empirically observed, most of the memorizing phenomenon originates from the duplication of images and captions in the training set. Owing to this, the authors propose to deduplicate the captions during the training st... | Rebuttal 1:
Rebuttal: We thank the reviewer for recognizing the merit of our work and acknowledging the thoroughness of the experiments.
### [Is SSCD score $> 0.5$ a reasonable proxy for memorization?]
Somepalli et al [1] found if the similarity of SSCD embeddings of 2 images is greater than 0.5, they are most likel... | Summary: The authors conduct a large-scale empirical study about the memorization problem in diffusion models. Their experimental results indicate that text conditioning plays an important role in this problem. Based on these observations, they propose to randomize the text prompts during training to mitigate memorizat... | Rebuttal 1:
Rebuttal: We thank the reviewer for recognizing the significance of the memorization problem in the AI safety domain. We sincerely appreciate your acknowledgment of our extensive experimental design.
### [Clarifications on caption diversity]
We apologize that our language at L177 is slightly confusing. W... | null | null | null | null |
Addressing the speed-accuracy simulation trade-off for adaptive spiking neurons | Accept (poster) | Summary: This paper focus on addressing the speed-accuracy tradeoff in adaptive spiking neurons. They utilize the existence of the absolute refractory period (ARP), and the fact that a neuron can spike at most once within this period, which could reduce the complexity to O(1), and develop an algorithmic reformulation o... | Rebuttal 1:
Rebuttal: > Motivation: The paper is well-motivated, and addressing the speed-accuracy tradeoff in the biological simulation is an important question. And using the fact that neuron can spike at most once within the absolute refractory period (ARP) to speed up the computation is novel.
Thank you for acknow... | Summary: The submitted manuscript proposed a method to speedup simulations of adaptive leaky integrate-and-fire (ALIF) neurons on a GPU by using the fact that during the absolutely refractory period ARP, neurons don't exchange any information and thus "blocks" on dt/ARP can be parallelized on GPU. This approach is then... | Rebuttal 1:
Rebuttal: > Once the code is made available, I am willing to improve my score and recommend the paper for publication at NeurIPS.
Thank you! We have uploaded all the code, including instructions to replicate all the experimental findings for the machine learning and neural-fitting benchmarks. In addition, ... | Summary: The authors implement an adaptive leaky integrate-and-fire model in a parallelized way on GPUs by incorporating a fixed refractory period that enables a window of time (‘block’) to be processed in one go. This accelerates the inference and training of spiking neuron models, while retaining accuracy, as shown b... | Rebuttal 1:
Rebuttal: We are grateful for your positive comments about the implications of our work.
> It is unclear why the accuracy does not go down in the empirical results presented.
The accuracy does go down slightly (not quite visible in original Figure 4a, but definitely visible in original Figure 5e.
> If t... | Summary: Spiking neural networks can be computationally expensive to implement because of their history dependence. However, spikes can't occur at any time whatsoever; instead, after spiking, there is a refractory period during which they cannot spike again. While this period is ~1-2ms, to attain precise dynamical inf... | Rebuttal 1:
Rebuttal: We thank you for your positive comments about the strengths of our paper and suggested edits to improve our work.
> Fig 5b): A timescale bar is needed, as is the ARP used. Given that there is very little performance change with DT=0.1ms, ARP=8ms (e), it would be helpful to visualize this here as ... | Rebuttal 1:
Rebuttal: Dear reviewers,
Thank you for taking the time to review our paper. As requested, we have run additional experiments, and have appropriately updated Figures 4 and 5, which we have uploaded in the attached single-page PDF. Furthermore, we have included a zoomed-in Supplementary plot of Figure 5b o... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Causal Component Analysis | Accept (poster) | Summary: The authors propose CauCA as a middle ground between ICA and CRL. They introduce two new assumptions, which is (i) knowledge of the causal variables and how they are related, and (ii) knowledge of the causal variables that were intervened upon to generate each dataset. With these two assumptions, CauCA identif... | Rebuttal 1:
Rebuttal: Thank you for the detailed review and for valuing _“the effort the authors made to intuitively explain their definitions, assumptions, and results, and to contextualize their work in relation to the broader landscape of identifiable representation learning”_.
Let us emphasize that learning causal... | Summary: The paper provides identifiability results for causal component analysis, where the causal structure of the latent variables is assumed to be known. The goal is to identify the causal mechanisms and mixing functions up to a certain degree of uncertainty. Compared to the previous work with unknown graphs and li... | Rebuttal 1:
Rebuttal: > _“it seems that the considered setting could possibly be transferred to the traditional one where there is no dependence among latent variables.”_
Thank you for this suggestion. As we explain below, CauCA cannot be reduced to nonlinear ICA when the graph is non-trivial. The intuition is that th... | Summary: The authors consider a class of settings in the realm of causal representation learning (CRL) and study their identifiability. More specifically, they consider Causal Component Analysis where the latent causal graph of the generative model is known and interventions are performed. As the authors show, the prob... | Rebuttal 1:
Rebuttal: Thank you for your positive evaluation of our work.
Comments on theory:
> _“The latent graph is known and interventions are performed on almost all latent variables. [...] the assumptions are too strong to have practical applications.”_
We agree these are strong assumptions. As you said, part ... | Summary: This paper introduces the concept of Causal Component Analysis (CauCA), which lies between Independent Component Analysis (ICA) and Causal Representation Learning (CRL). CauCA generalizes ICA by considering the causal dependence among latent components while assuming knowledge of the causal graph, focusing on ... | Rebuttal 1:
Rebuttal: Thank you for your positive feedback on our work.
> _“The authors did not discuss the limitations, so one paragraph on this may need to be added.”_
Based on this comment and on the discussion with all reviewers, we will add a paragraph discussing limitations, including: (i) Known intervention ta... | Rebuttal 1:
Rebuttal: Thank you for the valuable reviews, acknowledging that Causal Component Analysis is “important for the development of causal analysis in the context of deep learning” (xr1H), and that our work will have “an impact on [Causal Representation Learning] research” (nsXw). Moreover, our results introduc... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Assessor360: Multi-sequence Network for Blind Omnidirectional Image Quality Assessment | Accept (poster) | Summary: This paper introduces a BOIQA model called Assessor360, which takes into consideration the observer's browsing behavior. Specifically, RPS method is proposed to combine content and detailed information to generate multiple pseudo-viewport sequences from a given starting point. It also uses an MFA module with D... | Rebuttal 1:
Rebuttal: We thank Reviewer #4 for his/her comments and the appreciation of our work. To the concerns expressed by Reviewer #4 in `Weaknesses` and `Questions`, our response is as follows:
**`R4-Weakness 1`**: Thanks for your suggestion. We replace the GRU of our method with Transformer, LSTM and Temporal P... | Summary: This manuscript proposes a novel multi-sequence network for BOIQA, which is inspired by the actual multi-assessor omnidirectional image quality assessment and introduces a generalized recursive probability sampling method for BOIQA tasks. In addition, a multi-scale feature aggregation module with distortion-aw... | Rebuttal 1:
Rebuttal: We thank Reviewer #3 for his/her comments and the appreciation of our work. To the concerns expressed by Reviewer #3 in `Weaknesses` and `Questions`, our response is as follows:
**`R3-Weakness 1`**: Thanks for your question. We utilize the Swin Transformer to process each viewport within the samp... | Summary: A blind quality assessment model is proposed for 360-degree images.
Strengths: 1. The problem is interesting and practical.
2. The attempt to model the observer’s browsing process is a plus.
3. The results look promising.
Weaknesses: 1. The recursive probability sampling is a little ad hoc to the reviewer. I... | Rebuttal 1:
Rebuttal: We thank Reviewer #2 for his/her comments. Our responses are as follows:
**`R2-Weakness 1`**: Thanks for your comment. We respectfully **DISAGREE** with your opinion that the handcrafted RPS weakens our motivation.
Firstly, it is true that previous OIQA methods have relied on handcrafted viewpo... | Summary: This paper proposed a novel blind omnidirectional image quality assessment framework based on the modeling of the observer’s browsing process, which is called Assessor360. In detail, Assessor360 designs a novel viewpoint sampling methods named “Recursive probability sampling (RPS)” based on content-guided-samp... | Rebuttal 1:
Rebuttal: We thank Reviewer #1 for his/her insightful comments. Our responses are as follows:
**`R1-Q1`:** Thanks for the questions. For your `first question`, we agree that the name "content-guided" may be commonly used when the method is guided by the object information and semantic information in the tr... | Rebuttal 1:
Rebuttal: Dear Reviewers and ACs:
We sincerely appreciate all the reviews. They give positive and high-quality comments on our paper with a lot of constructive feedback.
We would like to emphasize that our work proposes a novel, simple and effective new pipeline called Assessor360. We appreciate your recog... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Construction of Hierarchical Neural Architecture Search Spaces based on Context-free Grammars | Accept (poster) | Summary: In this paper, the authors focus on the search space issue in neural architecture search and propose a unifying search space design framework that models all candidate neural structures with context-free grammars. This approach enlarges the search space, which is more likely to find high-performance architectu... | Rebuttal 1:
Rebuttal: We want to thank the reviewer for the encouraging comments and thoughtful questions. We are very happy that the reviewer finds our work “precious” and “easy for readers, even researchers not in this area, to understand the main idea of this paper.”
By following the reviewer’s suggestion, we were ... | Summary: The authors propose improvements to several parts of the neural architecture search (NAS) pipeline. First, they introduce a methodology for specifying search spaces based on a combinator grammar that is both more flexible than previous structured search spaces and encourages more regularity and hierarchy than ... | Rebuttal 1:
Rebuttal: We want to thank the reviewer for the thorough review and thoughtful raised questions. We are glad that the reviewer finds that our “paper stands out as both careful and ‘obvious in retrospect’”. We are very happy that the reviewer acknowledges our “strong results”, while using “relatively little ... | Summary: This paper uses context-free grammar to improve the efficiency of neural architecture search. Experiments show that the proposed approach outperforms existing methods.
Strengths: - Simple, well-motivated and efficient approach to improve neural architecture search.
- Clear presentation.
Weaknesses: No maj... | Rebuttal 1:
Rebuttal: We thank the reviewer for taking the time to review our work. We are glad that the reviewer finds our approach “simple, well-motivated and efficient”. Below, we address an interesting point raised by the reviewer in more detail.
> Exclusion of the optimal architecture from the search space (as di... | Summary: The authors propose an approach to unify the search space for Neural Architecture Search (NAS) using context-free grammars. They further develop a Bayesian Optimization based search strategy to effectively explore a huge search space. They experiment over (proposed) hierarchical extension of NAS-Bench 201 benc... | Rebuttal 1:
Rebuttal: We thank the reviewer for the review. We are happy that the reviewer views our approach as a “principled solution for macro to micro search” and that it “allows exploration of more controlled and viable architectures”.
By following the reviewer’s suggestion, we additionally conducted experiments... | Rebuttal 1:
Rebuttal: We thank all the reviewers for their thorough review, thoughtful questions, and constructive criticisms. We are very glad to hear that reviewers find that our work is “precious” (mjYa), “principled” (sktw), “simple, well-motivated and efficient” (gqtA), as well as “stands out as both careful and ‘... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Ensemble-based Deep Reinforcement Learning for Vehicle Routing Problems under Distribution Shift | Accept (poster) | Summary: This paper suggests a novel method to mitigate the distributional shift of combinatorial optimization problems. The idea is to bootstrap and aggregate multiple diverse policies which are similar to a classical ensemble method called bagging. The idea is simple and effective having promising empirical performan... | Rebuttal 1:
Rebuttal: Thanks for reviewing our paper with positive feedback. We appreciate that the reviewer acknowledges the motivation and significance of our work. Our responses to the comments are as follows.
Q1: One of the other major topics of the distributional shift of combinatorial optimization is scale shift... | Summary: This paper investigated the classical vehicle routing problems and proposed an ensemble-based deep reinforcement learning algorithm. Moreover, two regularization terms (Theil index and L2-dist) were designed to improve the generalization ability of learned policies. The motivation is convincing and the propose... | Rebuttal 1:
Rebuttal: We thank the reviewer for the comments and support. We also appreciate that the reviewer acknowledges the quality and contribution of the paper.
Q1: Why not just use multi-agent reinforcement learning to solve the problem? What is the relationship between ensemble learning and multi-agent policy ... | Summary: In this paper, an ensemble-based deep reinforcement learning algorithm is proposed for vehicle routing problems. The goal of the method introduced in this work is to improve handling of distribution shifts at test time. The approach, EL-DRL, combines multiple ideas: i) and ensemble of sub-policies, ii) distinc... | Rebuttal 1:
Rebuttal: We appreciate your insightful and helpful feedback on our paper. Please find below our detailed responses to the concerns/comments raised by you.
Q1: Verify whether the improvements due to the diversity-enhancing mechanisms are significant or marginal.
As suggested, we used Google rliable to com... | Summary: The paper proposes a deep reinforcement learning approach for vehicle routing problems under distribution shift that trains an ensemble of models. During solution construction, all ensemble models are evaluated at each decision step and the actions are selected based on the combined output probabilities. The e... | Rebuttal 1:
Rebuttal: Thanks for reviewing our paper with constructive and valuable feedback. Our detailed responses to the concerns/comments are as follows.
Q1: How many solutions do you generate per test instance? In the considered neural combinatorial setting there is no direct need to combine model outputs, inste... | Rebuttal 1:
Rebuttal: We appreciate all reviewers for their thoughtful reviews and constructive comments. In this general response, we provide the reference list, as well as a figure in the PDF file to support our response. Our detailed point-to-point responses can be found below.
## STATISTICALLY SIGNIFICANT
We use... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Learning in the Presence of Low-dimensional Structure: A Spiked Random Matrix Perspective | Accept (poster) | Summary: This manuscript presents an exploration of high-dimensional regression using a theoretical approach. It focuses on a model characterized by:
- A single-index model where the output is dependent on a one-dimensional projection of the input in a specific direction, $\mu$.
- Spiked covariance data with Gaussian i... | Rebuttal 1:
Rebuttal: Thank you for the detailed and thoughtful feedback. We address the technical points below.
**Q: Motivation of single-index $f_*$**
The single-index model is a canonical setting of learning low-dimensional signals from high-dimensional data; it covers many statistical learning problems such as p... | Summary: This paper considers the problem of learning a single-index target function under the spiked covariance data in the proportional asymptotic limit.
The results show the spike magnitude required for both kernel ridge regression (KRR) and two-layer neural networks (NN) trained with gradient descent and reveal tha... | Rebuttal 1:
Rebuttal: Thank you for the comments and questions. We address the technical points below.
**Q: Single-index assumption**
The single-index model is a canonical setting of learning low-dimensional signals from high-dimensional data; it covers many statistical learning problems such as phase retrieval and... | Summary: The paper studies the learning capacity of Kernel Ridge Regression and 2-Layer Neural Network to approximate a target function $f^\star$ (denoted teacher) under a low dimensional structure. In particular, starting from a generative process where the covariance of the data has a low dimensional structure ($\mat... | Rebuttal 1:
Rebuttal: Thank you for the thoughtful feedback. We address the technical points below.
**Q: low dimensional structure both for the data and the target function $f_*$**
As remarked in the Introduction, the spiked covariance model is a classical statistical model that captures low-dimensional structure in... | Summary: This paper studies a problem of parameter recovery under a generalized linear model-like setting. The novelty of this problem is revisiting this setting in the proportional limit setting: The sample size divided by the feature dimension converges to a fixed constant as both increases to infinity. This is a hig... | Rebuttal 1:
Rebuttal: Thank you for the comments and questions. We address the technical points below.
**Q: Lack of insight**
Here we briefly summarize the takeaways of our analysis.
1. By introducing a low-dimensional structure to the input data in the form of a spiked covariance (which is a common model in high... | null | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Diffusion-Based Probabilistic Uncertainty Estimation for Active Domain Adaptation | Accept (poster) | Summary: This paper tackles active domain adaptation based on the diffusion probabilistic model to estimate the prediction uncertainty. The authors apply the variational inference to the joint distributions of input data, latent codes in feature and prediction spaces, and labels. To realize adaptation, KL loss, and adv... | Rebuttal 1:
Rebuttal: Dear Reviewer 1fQe:
Thanks for your comments. Please kindly find below our responses and answers. Due to limited length, Table R6 and R7 are in the pdf file attached in the Common Response.
```
Weakness 1: The combination nature.
```
Thanks for the comment. We hope to clarify that the contributio... | Summary: This paper introduces a probabilistic framework for ADA, which incorporates both data-level and prediction-level uncertainties using variational inference. It employs a VAE and a diffusion-based classifier to approximate the joint posterior distribution. Adversarial learning is used to ensure an invariant late... | Rebuttal 1:
Rebuttal: Dear Reviewer rDAp:
Thanks for your positive words and constructive comments. Below are our responses to the weaknesses and questions.
```
Weakness 1: Increased complexity and computational cost, limited scalability and practicality.
```
Response: Thanks for your comments. The computational compl... | Summary: This paper aims to tackle the uncertainty estimation problem in ADA with a probabilistic framework, which captures two levels of uncertainties by a variational autoencoder and a diffusion-based classifier, respectively. An adversarial adaptation strategy and a statistical test-based criterion are further intro... | Rebuttal 1:
Rebuttal: Dear Reviewer Grmw,
We appreciate the comments and questions you post. Here are our point-to-point responses.
```
Weakness 1: Suboptimal domain adaptation component.
```
Response: Thank you for the thoughtful feedback on improving the domain adaptation component of our approach. We totally agree ... | Summary: Authors propose a new model for active domain adaptation. They train a VAE with domain-invariant latents (using adversarial alignment) and a supervised+teacher/student classifier head on top of learned embeddings to obtain point estimates of class probabilities, and then train a denoising diffusion model on pr... | Rebuttal 1:
Rebuttal: Dear Reviewer fVMQ,
Thank you very much for your feedback and questions. We provide our responses below.
```
Weakness 1: Marginal improvements over a very strong deterministic backbone method.
```
Thanks for recognizing this. While the gain may seem marginal, we believe formulating ADA in a proba... | Rebuttal 1:
Rebuttal: **Common Response:**
We sincerely appreciate the time and effort the reviewers have dedicated to reviewing our paper and providing thoughtful and constructive feedback. Here we provide responses to some common concerns raised by multiple reviewers. The attached pdf file contains the revised Figur... | NeurIPS_2023_submissions_huggingface | 2,023 | Summary: This paper concerns Active Domain Adaption leveraging a diffusion-based adversarial probabilistic model with T-statistics for uncertainty quantification. The paper combines many existing ideas such as variational auto-encoder, diffusion model, knowledge distillation, and T-statistics. Then it proposes a new fr... | Rebuttal 1:
Rebuttal: Dear Reviewer 3siL:
Thank you very much for your detailed comments and insightful questions. We provide point-to-point responses and answers below.
```
Q1: Generalization to common domain adaptation scenarios.
```
Response: Thanks for the comment. It is natural to employ our method to common DA s... | null | null | null | null | null | null |
ID and OOD Performance Are Sometimes Inversely Correlated on Real-world Datasets | Accept (spotlight) | Summary: This work performs an empirical study on the correlation between ID and OOD performances of models in vision and NLP. While some previous work claimed that there is usually a positive correlation between ID and OOD, this work argues that there are real datasets on which ID and OOD are negatively correlated, so... | Rebuttal 1:
Rebuttal: Thanks for the thorough review. We are grateful for your appreciation of the importance of this paper for the OOD community.
> Q1. Section 6 is not very useful
We agree that the empirical results can stand on their own. But we are also fearful of readers dismissing the claims as "hand-wavy" unle... | Summary: The manuscript observe a phenomenon that ID and OOD performance can be inversely correlated because spurious correlations under the partially informative invariant feature (PIIF) assumption. The authors analyze the reasons of this missing phenomenon in previous studies, and conduct dataset-wise and shift-wise ... | Rebuttal 1:
Rebuttal: Thanks for your time and suggestions.
> The most obvious weakness of this paper is that the analyses are mostly experiment-centric
On the contrary, this is actually the key strength of the paper. The theoretical possibility of an inverse ID/OOD correlation is almost trivial (especially from a ca... | Summary: The paper presents an empirical examination of ID vs. OOD performance on several real-world datasets showing inverse correlation patterns can indeed arise in many datasets and experimental settings. This conflicts with past papers mainly observing a positive dependence structure. The paper further discusses wh... | Rebuttal 1:
Rebuttal: Thanks for your time and suggestions.
> I would make sure not to oversell the "discovery of inverse ID-OOD tradeoffs" but not lose focus on explaining how previous papers could have overlooked this possibility (which the paper already does for the most part but not everywhere).
This is a very va... | Summary: Through extensive experiments, the authors showed that ID and OOD performances can inversely be correlated. The key ingredient is to track the performances throughout the training dynamics, instead of looking only at the end point.
Strengths: By tracking the training process, instead of the final point, the ... | Rebuttal 1:
Rebuttal: Thanks for your time and suggestions.
> Q1. Theorem 1 demonstrates this diverse performance phenomenon can happen even to simple linear models, how NN does so was not analyzed. I'd like to see the authors analyze a shallow neural net.
The claims of this paper are much more fundamental: they are ... | null | NeurIPS_2023_submissions_huggingface | 2,023 | Summary: The paper examines the previously proposed idea that in-distribution (ID) and out-of-distribution (OOD) performance of models are positively correlated. Through various experiments, the authors demonstrate that ID and OOD performance can in fact be inversely correlated. They also discuss how the previous work ... | Rebuttal 1:
Rebuttal: Thanks for your time and suggestions.
> I need to see the relevance of the theoretical analysis (section 6) to the experiments. In the previous section, I elaborated my recommendations to make this connection happen, and I am asking the authors to address my concern.
The theoretical possibility ... | null | null | null | null | null | null |
SLIBO-Net: Floorplan Reconstruction via Slicing Box Representation with Local Geometry Regularization | Accept (poster) | Summary: The paper presents a learning-based framework on floorplan reconstruction from point cloud. To tackle with the complex polygons emerge widely in floorplan, the polygon is decomposed to non-overlapping, vertical-slicing boxes as regression target of the networks. A bbox transformer and a room transformer are de... | Rebuttal 1:
Rebuttal: **W1**: *Bbox assignment: the paper assigns the slicing bboxes to rooms by nearest distance search. However, this may not always be the case when some irregular polygons (the center living room shown in Figure 7 left). Some bboxes may be incorrectly matched with other rooms when there exists some ... | Summary: The paper introduces SLIBO-Net, a novel approach for reconstructing 2D floorplans from unstructured 3D point clouds. By employing a transformer-based architecture and an efficient floorplan representation, the method improves room shape supervision and manages token numbers effectively. The proposed approach a... | Rebuttal 1:
Rebuttal: **W1**: *It is notable that the significant improvement in performance can be largely attributed to the Box IoU function, as acknowledged by the authors themselves. From a practical standpoint, this suggests that the additional technical contributions and associated complexity may have limited val... | Summary: This paper focuses on 2D floorplan reconstruction from 3D point clouds. The authors represent 2D floorplans as sliced boxes and apply a transformer network to predict box characteristics, as well as post-processing steps to build the final floorplan. The authors conduct experiments on Structure3D and show be... | Rebuttal 1:
Rebuttal: **W1**: *I still don't understand why two parallel transformers are used instead of a single transformer for room center and slice box prediction. Perhaps it is better to show quantitatively that joint prediction is not a viable solution. Otherwise I find the proposed method a bit clumsy.*
**Repl... | Summary: The authors propose a new method for floorplan reconstruction from point clouds. Specifically, the method formulates the problem from predicting polygons for each room to predicting sliced boxes and their corresponding room centers by imposing the Manhattan grid assumption on the floorplans. The authors also u... | Rebuttal 1:
Rebuttal: **W2**: *Extending the local planes across the whole space might be unnecessary to form the cell complex. The extended planes will intersect at many empty locations.*
**Reply**: Extending local planes across the whole space has several advantages. Firstly, such unlimited expansion can propagate ... | Rebuttal 1:
Rebuttal: We are very grateful to all the reviewers for their valuable and insightful comments. To address all the questions, we have prepared a pdf document that contains figures to better illustrate our responses. Due to the format limitation of this pdf document, we will also explain all the figures belo... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Alignment with human representations supports robust few-shot learning | Accept (spotlight) | Summary: This paper explores the relationship between the alignment of deep learning based representations with humans and the downstream performance of the NN models. Using triplets as a non-metric measure of similarity, the authors use information theory to show that alignment and downstream performance would have a ... | Rebuttal 1:
Rebuttal: Thank you for your valuable feedback! We are really glad you appreciated the scale of our experiments and the architectural insights. We are especially grateful for your question about “outlier” examples as this led to an exciting new insight detailed below.
**1: Contribution seems limited to U-... | Summary: The work provides theoretical and empirical analysis for the relationship between human alignment, few-shot performance, and robustness to distribution shift. The theoretical analysis predicts models that have high and low alignment with human annotations will have better performance on few-shot and distributi... | Rebuttal 1:
Rebuttal: Thank you for your valuable feedback! We’re very glad to hear that the key result and information-theoretic framework will be useful for the community!
**Comment 1: Other works [1][2] have shown that the validation accuracy on an in-distribution (ID) dataset is highly correlated with accuracy on ... | Summary: In this paper, the authors present a theory for a u-shaped effect in terms of model representation alignment, where models that are more and less aligned with human representations should have better performance over a number of tasks. They then perform a large number of experiments, comparing 491 models with ... | Rebuttal 1:
Rebuttal: Thank you for your valuable feedback! We’re grateful that you enjoyed the paper and expressed such enthusiasm for the theory and experiments!
**Comment 1: The largest of these weaknesses is that while the authors repeatedly refer to a U-shaped effect, the results actually suggest a J-shaped effec... | Summary: - The authors predict that there should be an U-shaped relationship between the degree of representational alignment and performance on few-shot tasks based on information theoretic analysis, and empirically show that is the case using hundreds of vision models. They further show that the degree of alignment p... | Rebuttal 1:
Rebuttal: Thank you for your valuable feedback! We’re really glad that you appreciated the U-shaped results and both the theory and experiments that support them!
**Comment 1: I believe ImageNet-A is just a subset of ImageNet-test set that happens to be difficult, but the term "adversarial robustness" ofte... | Rebuttal 1:
Rebuttal: **We thank all the reviewers for their valuable feedback! We are encouraged that the reviewers found the paper to be “well-written”, “well-presented”, and “well-organized”; the theory to be “novel”, “interesting”, “valuable”, and “useful” for the community; and the experiments to be “detailed”, “e... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Video Prediction Models as Rewards for Reinforcement Learning | Accept (poster) | Summary: This paper proposes Video Prediction Rewards (VIPER), a general architecture that extracts reward functions from action-free expert demonstrations. To match the agent's trajectory distribution with the source distribution, VIPER optimizes the agent to maximize a log-likelihood estimated by an auto-regressive v... | Rebuttal 1:
Rebuttal: We would like to thank the reviewer for their insightful comments. We provide detailed answers to your questions below:
**[Weakness 1-1] The data efficiency of VIPER seems to be low as it requires nearly 10M data to converge in DMC.**
While agents trained with VIPER may require millions of train... | Summary: The authors present Video Prediction Rewards, an algorithm that leverages transformer-based video prediction models as action-free reward signals for reinforcement learning. The reward induced by this video prediction model incentivizes the agent to find the most likely trajectory under the expert video distri... | Rebuttal 1:
Rebuttal: We are happy that the reviewer recognized the importance of learning reward functions for RL and found our method concise and effective across a wide range of tasks. The PDF in the global rebuttal contains:
- Additional experiments with more baselines [A, B] (Figure A)
- Further analysis of OOD g... | Summary: This paper proposes a learning-from-observation algorithm that builds a reward based on a video predictor trained from action-free expert videos. Experimental results show that online reinforcement learning algorithms can learn a working policy from their reward only effectively. This paper contains rich and i... | Rebuttal 1:
Rebuttal: We thank reviewer d6Ah for their thoughtful comments, and provide detailed answers to their questions below.
**[Q1] How good is the generalization ability of VIPER? We definitely need evaluations on more OOD combinations to support the statement in the third contribution.**
Thank you for pointin... | Summary: This paper proposes to use prediction likelihoods from autoregressive video models as reward functions to train reinforcement learning agents. Specifically, the conditional likelihood $\log p(x_{t+1}|x_{1:t})$ is augmented with an exploration reward to avoid suboptimal behavior. The authors conduct extensive e... | Rebuttal 1:
Rebuttal: We would like to thank the reviewer for their thoughtful comments. We provide detailed answers to your questions below:
**[Weaknesses] one fundamental problem unresolved in this work is why the autoregressive likelihood, which is inherently non-Markov, can work with general RL algorithms, in whi... | Rebuttal 1:
Rebuttal: We thank all reviewers for their constructive feedback and for helping us make our paper a stronger submission!
- First of all, we highlight that we have included **two new imitation learning from observation baselines (Figure A)**, BCO and GAIfO. In short, VIPER consistently shows significant i... | NeurIPS_2023_submissions_huggingface | 2,023 | Summary: The paper proposes to use a large video-prediction model for learning a reward model for RL. The agent's performance is evaluated on a total of 8 envs from 3 benchmarks: DMC, Atari, and RLBench. The paper argues that the proposed model also generalizes to different environments for which no training data was p... | Rebuttal 1:
Rebuttal: We would like to thank the reviewer for their helpful comments. We provide detailed answers to their questions below.
**[Weaknesses] Overall, I think the paper is interesting but I want to see performance improvement over a bunch of baselines and some ablations.**
We have included additional bas... | Summary: This paper proposes a simple method that uses a pre-trained video prediction model to provide rewards for online RL. The design includes using VQ-GAN to encode discrete embeddings and incorporating an exploration bonus (opt for Plan2Explore and RND). In experiments, the authors also show the learned rewards pr... | Rebuttal 1:
Rebuttal: Thank you for your constructive feedback. We provide detailed answers to the questions below.
**[Weakness 1 & 2] The effectiveness of the proposed method may be limited if the expert data is scarce. … this paper undoubtedly requires a lot of expert data (to train the video prediction model).**
F... | null | null | null | null |
On the Convergence and Sample Complexity Analysis of Deep Q-Networks with $\epsilon$-Greedy Exploration | Accept (poster) | Summary: This paper studies DQN (one of the most popular RL algorithms in the literature) with $\epsilon$-greedy behavior policy. Specifically, the authors provide the convergence and sample complexity guarantees and numerically verify the algorithm's performance.
Strengths: This paper is well-written, and the appendi... | Rebuttal 1:
Rebuttal: *Q1: Regarding Corollary 1.*
**A1**: Thanks for your insightful question! $e_t$ in Corollary 1 can be replaced by its upper bound, which is independent of $\epsilon_t$. See GR for details.
*Q2: Regarding “W^T depends on the iteration number.” and “almost sure convergence of the weights”.*
**A2... | Summary: The paper studies non-asymptotic behavior of deep Q-network where the inner loop uses momentum and the behavior policy is time-varying epsilon greedy policy. With nice choice of initialization and exploring strategy $\epsilon_t$, the authors prove that the DQN algorithm enjoys $1/\sqrt{N}$ sample complexity.
... | Rebuttal 1:
Rebuttal: ***Q1**: Regarding $\epsilon_t$.*
**A1**: Thanks for your insightful question! $e_t$ in Corollary 1 can be replaced by its upper bound, which is independent of $\epsilon_t$. See GR for details.
Moreover, we would like to clarify the insights from our upper and lower bound for $\epsilon_t$. The u... | Summary: This paper provides a theoretical analysis of the converge of DQN in deep reinforcement learning. Its contributions are mostly theoretical and concern the important setting of epsilon-greedy exploration DQN with general function approximation. The paper includes convergence and sample complexity analysis for D... | Rebuttal 1:
Rebuttal: We thank the reviewer's comprehensive reading of this paper and valuable suggestions. The point-to-point responses will be posted based on the reviewer’s questions.
***Q1**: As far as weaknesses go, there are minor orthographic typos that remain in the main text, and should be addressed in the ... | Summary: This paper analyzes the sample complexity of using $\epsilon$-greedy for exploration in the setting of general function approximation.
Strengths: This paper provides an convergence analysis of using $\epsilon$-greedy in Q-learning.
Weaknesses: In my opinion, this paper is studying optimization rather than sa... | Rebuttal 1:
Rebuttal: ***Q1**: In my opinion, this paper is studying optimization rather than sample complexity.*
**A1**: Thank you for sharing your thoughts. Your input is greatly appreciated. We find that the concept of "sample complexity" lacks a precise definition within the RL literature. To make a fair compariso... | Rebuttal 1:
Rebuttal: We thank all the reviewers' comprehensive reading of this paper and valuable suggestions, and the paper will be revised according to the reviewers’ comments. We will first provide our general response (GR) to capture the major modifications made, focusing particularly on our revision of Theorem 1.... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Online RL in Linearly $q^\pi$-Realizable MDPs Is as Easy as in Linear MDPs If You Learn What to Ignore | Accept (oral) | Summary: This submission provides the first algorithm learning an $\epsilon$-optimal policy with poly many samples obtained through online interactions with a $q^\pi$-realizable MDP. This is an important contribution to an open question in the community. Indeed previous works either adopted a generative model or they r... | Rebuttal 1:
Rebuttal: We thank the reviewer for their helpful comments. We answer their questions below:
Computational complexity and Q4: Regrettably, we eliminated our discussion on computational complexity when shortening the initial submission -- thank you for pointing this out. We will include it in the final subm... | Summary: This paper proves that the standard online episodic RL with the $q^\pi$-realizability assumption can be learned with polynomial samples, without generative models. Technically, this paper proves that the states of any MDPs with $q^\pi$-realizability assumption can be partitioned into two disjoint sets, where t... | Rebuttal 1:
Rebuttal: We thank the reviewers for their helpful comments. We address their concern mentioned in “weaknesses” in the [“global” response](https://openreview.net/forum?id=HV85SiyrsV¬eId=26AjXdraOA) and answer their questions below:
Q1. We do not show a formal reduction of $q^\pi$-realizable MDPs to line... | Summary: This paper studies online RL under linear $q^\pi$ realizable MDPs, and proposes an algorithm that learns an epsilon-optimal policy in polynomial time and polynomial sample complexity. Previous online RL in linear RL either assumes that transitions and rewards are linear individually (e.g., linear MDP), or ass... | Rebuttal 1:
Rebuttal: We thank the reviewer for bringing these presentational issues to our attention, and a special thanks for taking the time to provide useful concrete suggestions for improvement! The proof of our result uses many novel ideas and techniques, and we strived hard to present it well despite its inheren... | Summary: The paper addresses online RL in episodic MDPs under linear $q^\pi$-realizability, where policy action values are expressed as linear functions of state-action features. This class is shown to be more general than linear MDPs, where only the transition kernel and reward function are linear functions of feature... | Rebuttal 1:
Rebuttal: We thank the reviewer for their helpful comments, based on which we have added a table to the revised version of our paper containing the summary of how our work fits in the wider literature. In short, (i) for linear MDPs, Jin et al. [2020] provide algorithms with both polynomial sample and comput... | Rebuttal 1:
Rebuttal: Please see our individual responses to the reviewers. We use this space to address a comment made by multiple reviewers that the exposition of the results could be improved (thank you for this feedback). We provide below a high-level explanation of our method and proof compared to ELEANOR (such a ... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Detecting hidden confounding in observational data using multiple environments | Accept (poster) | Summary: The authors develop a test for hidden confounding using only observational data from multiple environments with the same DAG under the independent causal mechanisms assumption.
Strengths: The paper is well-written and covers the related work well.
Their setting, assumptions, and contributions are clear and th... | Rebuttal 1:
Rebuttal: We thank the reviewer for their insightful questions, as well as pointing out the novelty of designing the test in our paper. We further agree that it is important to keep in mind that our test relies on new strong assumptions that should not be taken for granted.
## The real-world practical util... | Summary: The authors focus on a setting where we observe treatments, outcomes, and a set of confounders across multiple environments. While it's generally not possible to identify unobserved confounders in a single data set, when we have multiple data sets and assume the mechanisms operate in the same way across those... | Rebuttal 1:
Rebuttal: We thank reviewer 7Ey5 for their valuable feedback on our work. As suggested by them, in the final version of the paper, we intend to bring the algorithm back to the main paper if space allows in the camera-ready version. Similarly, we will also extend the introduction with two additional examples... | Summary: The authors proposed a detection method for hidden confounders, with only observational data in multiple environments. They, in particular, designed a testable independence condition for it, under the assumptions of (i) Faithfulness & Causal Markov Property, (ii)Shared Causal Graph, (iii) Independent Causal Me... | Rebuttal 1:
Rebuttal: We thank reviewer ZLeH for their helpful feedback. We have corrected the typo on line 113 pointed out by them. For clarity, we will address their questions in a Q&A format below.
**Q: Since the $E^{(k)}$ consists of four causal mechanisms shown in Fig.1(a), how could it serve as an indicator vari... | Summary: This paper presents a method for detecting hidden confounding in observational data. Without further assumptions, detecting hidden confounding is known to be impossible. This paper takes advantage of a common scenario where one has access to observational data collected from multiple environments. Given par... | Rebuttal 1:
Rebuttal: We thank reviewer 1FZJ for their insightful feedback and questions. Based on their recommendation, to help future readers of our paper, we have added an earlier reference to our discussion on the violation of our assumptions when first presenting the assumptions in the paper.
## The influence of ... | Rebuttal 1:
Rebuttal: This is a joint reply to reviewers 7Ey5 and g1qg for their questions on the real-world practical utility of our proposed method. For the camera-ready version, if space allows, we will include the below examples in the introduction of our paper to help future readers understand our method’s practic... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Guiding The Last Layer in Federated Learning with Pre-Trained Models | Accept (poster) | Summary: The work proposes to address the training of the classifier (last layer of the network) separately in the setup of federated learning with a pretrained weights initialization. In particular, the suggestion is to use a nearest class means (NCM) approach, which saves a lot of computation time. The proposed appro... | Rebuttal 1:
Rebuttal: We thank the reviewer for acknowledging benefits of our work and their helpful comments which help us improve the clarity of the paper. We address the comments in turn:
**W1: is the final layer updated during fine-tuning**
We apologize for the confusion, the final phase also updates the final la... | Summary: This paper studies strategies to fine-tune models in Federated Learning, starting from a pre-trained model. It has been shown in the literature that pre-training is beneficial in FL, that it improves convergence speed and robustness to heterogeneity.
This paper shows that existent fine-tuning pipeline are not... | Rebuttal 1:
Rebuttal: We thank the reviewer for their overall positive view of the work and their helpful comments which will improve the paper. We address the comments in turn:
**Weakness 1: Comparisons to first stage LP**
We have added a comparison below with the suggested baseline of running LP for several rounds. ... | Summary: This paper proposes to use nearest class means (NCM) with pre-trained models in federated learning, coined as FedNCM. Experiments show that FedNCM is effective in terms of convergence and communication cost, due to the application of pre-trained models and light-weighted last layers.
Strengths: 1. Using pre-t... | Rebuttal 1:
Rebuttal: We thank the reviewer for their helpful comments which help us to improve the paper, we address each comment in turn:
**Clarification:** Our results show substantial improvement over prior work, not just the random baseline.
**W1: Novelty**
We emphasize that the simplicity of the proposed approa... | Summary: This paper studies the effect of tuning the linear head on federated feature learning. They show that well-tuning the linear head will make the base feature extractor converge closer to the optimal solution while reducing the total communication cost in FL. They evaluate their method on multiple FL datasets an... | Rebuttal 1:
Rebuttal: We thank the reviewer for their helpful comments and their interest in our research direction. We address the comments in turn:
**Weakness 1: Closed form solution**.
Our experiments largely consider the cross entropy loss for which there is no closed form solution. Indeed, it is an interesting i... | Rebuttal 1:
Rebuttal: We thank the reviewers for their helpful comments and suggestions that help us improve the paper. We have responded to the reviewers in the individual comments.
We additionally include a pdf with NLP experiments suggested by reviewer tHeD. These show additional results on NLP Datasets using the ... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Direct Preference Optimization: Your Language Model is Secretly a Reward Model | Accept (oral) | Summary: This paper tackles a very important problem in LLM training, i.e., how to simplify the complex process of reinforcement learning from human feedback? RLHF is different to use, it is a complex and often unstable procedure, first fitting a reward model that reflects the human preferences, and then fine-tuning th... | Rebuttal 1:
Rebuttal: Thank you for your detailed summary of the paper, your feedback, and believing that our work is highly significant for researchers!
---
Rebuttal Comment 1.1:
Comment: I have read the author's response, and I maintain my score and recommend accepting this paper. | Summary: This paper presents a new policy optimization algorithm with human preferences, called Direct Preference Optimization (DPO), which eliminates the need for explicitly fitting a reward model. The authors formulate the process of RLHF into a single stage policy learning by leveraging a mapping between reward func... | Rebuttal 1:
Rebuttal: Thanks for your helpful feedback and questions!
### Generalization of DPO vs existing RLHF
First, we note that all of the evaluations in the paper already compute test win rates on unseen test prompts.
We performed a follow-up experiment to assess the performance of PPO and DPO under distributi... | Summary: The authors propose DPO, a method to fine-tune a language model to human preferences that does not use reinforcement learning or learn a reward model, but rather, directly optimizes the language model with preference data.
Through a change of variables, the authors express the reward function in terms of the ... | Rebuttal 1:
Rebuttal: Thanks for assessing our paper to be a significant contribution, and for your feedback!
### Single-turn dialogue experimental results did not compare DPO with a RLHF baseline…
Please, see Fig. 3 in the rebuttal pdf for an additional baseline to our dialogue experiments, the oracle best-of-N meth... | Summary: RLHF is generally done by training a reward model on a dataset of preferences and then using on-policy RL to finetune a language model with the reward model. This work proposes to bypass explicit learning of a reward model and finetune the language model directly on the preference dataset. This is conceptually... | Rebuttal 1:
Rebuttal: Thanks for appreciating our work and providing a great summary for it!
### DPO vs PPO generalization.
We performed a follow-up experiment to assess the performance of PPO and DPO under distribution shifts. We evaluate the summarization policies trained on Reddit TL;DR data on the CNN/DailyMail d... | Rebuttal 1:
Rebuttal: We appreciate the detailed feedback provided by the reviewers. We address a few common points in this response. All other questions are addressed in reviewer specific responses.
## Re: Generalization of PPO and DPO
First, we note that all of the evaluations in the paper compute test win rates on... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Learning Dense Flow Field for Highly-accurate Cross-view Camera Localization | Accept (poster) | Summary: The paper addresses the problem of cross-view camera localization, i.e. of estimating a 3D camera pose for a ground-level image with respect to a satellite image of its local surroundings (assuming a coarse localization is already given). For this, it proposes a deep learning pipeline that aims at estimating a... | Rebuttal 1:
Rebuttal: We thank Reviewer-NkYA for the valuable comments. Please find our response below.
**1:Minor technical contribution, limited impact.**
Instead of a trivial ensemble of existing techniques, our contribution is mainly on how to solve unique challenges in a new scenario faced by existing technique... | Summary: This paper deals with the 3-DoF camera pose estimation method for a ground-level image with respect to a satellite image. The proposed end-to-end approach that leverages the learning of dense pixel-wise flow fields between the projected ground-level image and the satellite image is built upon RAFT architectur... | Rebuttal 1:
Rebuttal: We thank Reviewer-63Cb for the valuable comments. Please find our response below.
**1:Differences compared to Shi and Li[29] and Zamir, Amir R., et al. [\*] are minor.**
From a high-level perspective, Shi and Li [29], Amir R., et al. [\*], and our method all estimate the relative pose from the ... | Summary: In this paper, the authors try to address the cross-view geo-localization by learning dense pixel-wise flow fields between ground and satellite images. To achieve this, they utilize a ground feature projection module to align the spatial feature and propose a refinement block to enhance the context feature for... | Rebuttal 1:
Rebuttal:
We thank Reviewer-1HCJ for the valuable comments. Please find our response below.
**1: Details of satellite images in the datasets, including collection, preprocessing, ground sample distance, data distribution, and so on.**
Please refer to our response to all the reviewers in the overall resp... | Summary: The paper presents a method that estimates a 3-DoF pose that aligns correspondences between ground-level and satellite/aerial images. The proposed solution uses a CNN-based model to extract image features, map image features from the ground-level camera to a BEV so that it can be matched with features from the... | Rebuttal 1:
Rebuttal:
We thank Reviewer-Bgje for the valuable comments. Please find our response below.
**1:Reasons for using CNN.**
There are currently two lines of work for estimating correspondences between two images (or two point clouds, etc.) The first line of work first constructs a noisy cost volume between... | Rebuttal 1:
Rebuttal: We appreciate all reviewers' valuable feedback and constructive suggestions. We are glad that two ethics reviewers confirmed that the datasets used in this paper are all licensed for academic use and not original to this work. We are committed to adhering to the guidelines set forth by the origina... | NeurIPS_2023_submissions_huggingface | 2,023 | Summary: The paper introduces a method for estimating 3DOF camera pose of satellite images given a ground-level image and an initial pose estimate as input. The ground-level image is mapped to a satellite view using the camera parameters, confidence-weighted 2D matches are estimated between the warped image and the sat... | Rebuttal 1:
Rebuttal: We thank Reviewer-pNph for the valuable comments. Please find our response below.
**1:Missing resource requirements.**
Our method is trained and evaluated with an NVIDIA TITAN V GPU with 12GB memory. The average evaluation time for each ground image is 0.25s. Using the same GPU device, the eval... | null | null | null | null | null | null |
A Heat Diffusion Perspective on Geodesic Preserving Dimensionality Reduction | Accept (poster) | Summary: In the paper "A heat diffusion perspective on geodesic preserving dimensionality reduction", the authors develop a new embedding method (HeatGeo), conceptually close to PHATE (Moon et al 2019) in that it performs MDS on some kind of diffusion distances. The authors make some strong claims about HeatGeo: (1) it... | Rebuttal 1:
Rebuttal: We would like to thank the reviewer for their thorough analysis of our manuscript. In the following we will address the main points raised by the reviewers. Typos or undefined notation were fixed.
**Relationship to PHATE is not very clear.** Our method follows the same main steps as Isomap and PH... | Summary: The authors propose heat geodesic embeddings, a novel approach to perform graph embeddings that is motivated by Riemannian geometry (Varadhan's formula and the parabolic Harnack inequality), which allows for directly linking heat diffusion on graphs to manifold distances. They performed experiments to demonstr... | Rebuttal 1:
Rebuttal: We are pleased to read such an encouraging perspective on our work and thank the reviewer for their feedback. We remain at their disposal for any further clarification.
---
Rebuttal Comment 1.1:
Title: reply to authors
Comment: This is to acknowledge that I have read the author's response to my ... | Summary: The authors develop specific heat diffusion kernels for manifold embedding.
They define a heat kernel based on the Varadha formula.
The authors try to weave some connections with existing methods of DR such as PHATE, t-SNE, etc.
They run some experiments
Strengths: The paper is well written except for the org... | Rebuttal 1:
Rebuttal: We would like to thank the reviewers for their constructive comments on our manuscript. Below we will address the main concerns of the reviewer.
>The writing style is too dense, poorly structured, with a succession of (too) short sections, referring to each other too many times, and often missing... | Summary: The paper introduces a dimensionality reduction technique based on heat kernels and called HeatGeo. The main idea is to learn a mapping that projects points to a low-dimensional space (ideally 2 or 3-dimensional for visualization) so that relative distances between pairs of points satisfy the same order as dis... | Rebuttal 1:
Rebuttal: We thank the reviewer for their encouraging review. Below we respond to the main weakness raised by the reviewer.
>The main weakness of the paper is that it mainly focuses on dimensionality reduction and visualization which already have widely used techniques (e.g. t-SNE and UMAP) that are used o... | Rebuttal 1:
Rebuttal: We would like to thank the reviewers for their time and valuable feedback when reviewing our paper. These suggestions led us to make a few modifications, which we believe greatly improved the quality of the manuscript. We now respond below to the main clarification points grouped by theme.
- **On... | NeurIPS_2023_submissions_huggingface | 2,023 | Summary: This article deals with representation learning and dimensionality reduction of high dimensional noisy datasets. The proposed method relies on heat diffusion. More precisely, assuming their data can be viewed as Riemannian manifolds and based on Varadhan’s formula, the authors define the notion of heat-geodesi... | Rebuttal 1:
Rebuttal: We would like to thank the reviewer for their suggestions, and we agree that the quality of the writing could be improved. We did a thorough pass on the manuscript, and corrected some typos including the ones mentioned. We also added the missing definitions. We believe that these minor corrections... | null | null | null | null | null | null |
A Simple Yet Effective Strategy to Robustify the Meta Learning Paradigm | Accept (poster) | Summary: The authors address an existing gap in fast adaptation of meta-learning by developing a robustification framework that draws from the concept of expected tail risk and recasts robustification as a distributional optimization problem. They adopt a heuristic two-stage optimization strategy that is broadly useful... | Rebuttal 1:
Rebuttal: We sincerely thank \# Reviewer Wn6T for these insightful comments. The remainders focus on concerns to address.
***
**1. Misconception of the tail risk**
Thanks for the comment.
(1) In our paper, ***the concept of the tail risk is from the literature work [1-2] and the standard definition in W... | Summary: The paper proposes a novel expected tail risk objective for meta-learning, which provides an intermediate of the worst-case and average risk minimization algorithms. The proposed algorithm can be adapted to existing meta-learning algorithms with improved robustness.
Strengths: * The paper is well-written an... | Rebuttal 1:
Rebuttal: We sincerely thank \# Reviewer vUMr for these constructive comments. The remainders focus on questions to answer.
***
**1. Consider and discuss generalization related work.**
Thanks for this suggestion.
(1) Most previous work studies the generalization in the context of the complete task space... | Summary: This paper considered a robust meta-learning framework by addressing the worst-tasks performance. Specifically, this paper adopted conditional value at risk (CVR) to filter poorly performed tasks. In order to effectively learn the CVR, this paper adopted a surrogate loss, a variant of hinge loss with a learnab... | Rebuttal 1:
Rebuttal: We sincerely thank \# Reviewer NA6N for these insightful comments. The remainders focus on concerns to address.
***
**1. Group DRO baseline comparison, research motivation and data augmentation**
Thanks for this precious question.
(1) Sorry for confusing you. ***The Group DRO baseline can be f... | Summary: In this paper, the authors present a method by which to improve the expected tail risk during meta learning via a two-stage optimisation strategy. In addition, the authors provide theoretical and empirical analysis of their proposed approach, demonstrating improvements in the average, worst-case and conditiona... | Rebuttal 1:
Rebuttal: We sincerely thank \# Reviewer ANpj for these helpful comments. The remainders focus on questions to answer.
***
**1. Potential improvement on crude Monte Carlo**
This is a crucial question.
(1) We choose the crude Monte Carlo in quantile estimate due to its simplicity. However, accurate estim... | Rebuttal 1:
Rebuttal: We sincerely thank all reviewers and area chairs for wonderful work. In this response, I will summarize reviews, address confusion/concerns, and report future changes in the manuscript.
***
### **I. Positive Comments**
These include: (1) *a novel meta learning pipeline/objective* **[\# Rev... | NeurIPS_2023_submissions_huggingface | 2,023 | Summary: This paper proposes a novel meta-learning pipeline from the perspective of distributionally robust perspective to robustify fast adaptation. Through a two-stage strategy heuristically, the model could control the worst fast adaptation cases at a certain probabilistic level. And experiments show the effectivene... | Rebuttal 1:
Rebuttal: We sincerely thank \# Reviewer YmWX for insightful comments. The remainders focus on questions to answer.
***
**1. Goal to introduce $(R^+,B)$ and the slack variable in Line157**
Thanks for this question.
(1) The goal of introducing $(R^+,B)$ is to describe the risk value distribution in the ta... | null | null | null | null | null | null |
Unifying GANs and Score-Based Diffusion as Generative Particle Models | Accept (poster) | Summary: In this paper, the authors consider the task of generative modeling as one of particle flow, and provide a unifying theory for score- and diffusion based generator-free models, and generator-based GANs. They propose the particle-flow models that encompassed diffusion and score-based methods such as NCSN or EDM... | Rebuttal 1:
Rebuttal: We would like to thank the reviewer for their detailed feedback. Due to rebuttal size constraints, we answer the reviewer's main questions below. We will follow-up with the rest of the response during the discussion period.
## Training of Discriminator Flows
Discriminator Flows do not attempt at... | Summary: The paper points out commonalities between diffusion and GAN models by examining the time evolution of the sample particles. For diffusion, this evolution happens at inference time by following the theory-given flow that transports the initial noise distribution to match the data distribution (in practice this... | Rebuttal 1:
Rebuttal: We would like to thank the reviewer for their feedback.
We concur with the reviewer regarding the weaknesses raised, which are acknowledged in the paper. Notably, we mention that Score GAN can serve as a distillation method in the conclusion. We address the reviewer's questions below.
## Latent ... | Summary: The authors interpret GANs through the lens of particle optimization and show the links to score-based diffusion models. Based on these theoretical results, they propose to either 1) do particle learning (without a generator) based on the flow of a discriminator, or 2) learn a generator by minimizing the scor... | Rebuttal 1:
Rebuttal: We would like to thank the reviewer for their feedback. We address the weaknesses they raised below.
## Novelty
> Particle GANs are not very novel; [...] this is not really new, but as far as I know, its rarely done
> line 21: 'challenge the conventional view that particle and adversarial gener... | Summary: The paper proposes a framework that unifies particle-based deep generative models, such as gradient flows and score-based diffusion models, and adversarial generative models, such as GANs, by framing generator training as a generalization of particle models. The authors suggest that a generator is an optional ... | Rebuttal 1:
Rebuttal: We thank the reviewer for their feedback. We address the weaknesses and questions they raised below.
## Wasserstein Notation
> Some mathematical symbols (e.g., $\nabla_W$) require further explanation to enhance the readability of this paper for a wider range of readers.
The notation $\nabla_W$ ... | Rebuttal 1:
Rebuttal: ## Response to Reviews
We would like to thank the reviewers for their insightful, constructive, and timely reviews. We are glad that most reviewers appreciated the paper and the perspective we propose, generally agreeing on its relevance and potential for the community.
Below, we respond to each... | NeurIPS_2023_submissions_huggingface | 2,023 | Summary: This paper proposes a framework, i.e., a generalization of particle models, that unifies particle-based generative models like score-based diffusion and gradient flows with adversarial generative models like GANs, where generator training is framed as interacting Particle Models. Based on this perspective, the... | Rebuttal 1:
Rebuttal: We thank the reviewer for their feedback. We address the questions they raised below.
> The claim that "Discriminator Flows have the expected advantage of converging faster to the target distribution as compared to the state-of-the-art diffusion model EDM, and thus have higher efficiency during i... | Summary: This work presents a unified generative particle model framework that explains generative adversarial networks (GANs) and score-based diffusion models.
This model is characterized by an evolution equation with a functional ${h_{\rho}}_t$ that captures the transitioning over the time index $t$.
This work expl... | Rebuttal 1:
Rebuttal: We thank the reviewer for their feedback. We address the weaknesses they raised by answering their questions below.
## (1). Wording Clarification
> (1). The following phrase is not clear: Line 51 on Page 2: "In prior work, ...". It is not clear if this is referring to a specific work or prior wo... | null | null | null | null |
Embedding Space Interpolation Beyond Mini-Batch, Beyond Pairs and Beyond Examples | Accept (poster) | Summary: The paper introduces a novel data augmentation method called MultiMix, which aims to go beyond the limitations of existing mixup methods. The authors propose interpolating an arbitrarily large number of examples beyond the mini-batch size in the embedding space, rather than along linear segments between pairs ... | Rebuttal 1:
Rebuttal: We appreciate the valuable feedback of R-SS7F. We address the concerns as follows:
### 1. Importance of using dense $\lambda$ in embedding space
Using the description of our contributions presented in global thread **Author Rebuttal by Authors**, we elaborate on motivation here.
``R-SS7F - one ... | Summary: This paper considers a generalized form of Manifold mixup by expanding the concept of a pair-wise interpolation method to the multi-sample-based interpolation on the entire convex hull. The proposed methods called MultiMix and Dense MultiMix show the best performance over prior mixup-based methods for the vari... | Rebuttal 1:
Rebuttal: We appreciate the valuable feedback of R-6vc1. We address the concerns as follows:
### 1. Why multi-sample-based mixup is effective
Using the description of our contributions presented in global thread **Author Rebuttal by Authors**, we elaborate on motivation here.
``R-6vc1 - why the multi-sampl... | Summary: The paper aims to increase the performance and robustness of deep learning model by increasing sample diversity. The authors propose data augmentation techniques which interpolate samples from a convex hull. The experimental evaluations show that the algorithm outperforms existing techniques.
**I have read th... | Rebuttal 1:
Rebuttal: We appreciate the valuable feedback of Reviewer Hzfq. We address the concerns as follows:
### 1. Presentation
We would like to point out that the other reviewers appreciate the writing and quality of the presentation. R-Xd45 and R-SS7F rate presentation as *excellent*. R-6vc1 acknowledges that Th... | Summary: The authors introduce MultiMix, an approach aimed at enhancing the Mixup class of techniques, which involve interpolating between input examples and their corresponding targets. This enhancement is achieved through the interpolation of a substantially more training examples, increased number of loss terms for ... | Rebuttal 1:
Rebuttal: We appreciate the valuable feedback from R-Xd45. We address the concerns as follows:
### 1. Motivation
Please refer to our comment on *Author Rebuttal by Authors*.
### 2. Comparison with AugMix
Thanks for sharing this work.
**AugMix vs. MultiMix.** AugMix applies a convex combination of ... | Rebuttal 1:
Rebuttal: As summarized in our first contribution (L81-83), MultiMix consists of the following three elements:
- Increase the number $n$ of generated mixed examples beyond the mini-batch size $b$ (L46-54, L71-78, Table 1/column *terms*, Fig. 4(b)).
- Increase the number $m$ of examples being interpolated ... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Aligning Language Models with Human Preferences via a Bayesian Approach | Accept (poster) | Summary: This paper proposes a method for modeling human preference with a Bayesian approach that considers the disagreement among human annotators. The authors also propose a contrastive learning method to replace RL in the alignment process. The authors conduct experiments on two text generation tasks and show some i... | Rebuttal 1:
Rebuttal: $\textbf{Mistakes in W1, Q1:}$ The second dataset's human preference data is not actual collected from human annotation, which does not makes sense. \& Can it be validated on some held out data that have disagreement annotations?
To clarify, the human preference data from the MIC dataset is genui... | Summary: The authors propose a novel approach for training preference models (PMs) for aligning language models (LMs) that accounts for human preference diversity. Their algorithm, d-PM, trains a PM to match a distribution over human preference in human feedback data. The authors also propose a contrastive learning alg... | Rebuttal 1:
Rebuttal: Thank you for your comprehensive feedback. We'd like to underscore the primary contribution of our work: considering the disagreement of human preference for LM alignment. Our work is focused on scenarios where human preferences are highly subjective. In such contexts, it is imperative to factor i... | Summary: The paper proposes a novel alignment method, which takes a Bayesian approach, termed Preference Modeling with Dis-
295 agreement (short as d-PM); the method is designed to capture disagreement in human labels, rather than assuming a gold human label. The Bayesian approach is basically a new way to train reward... | Rebuttal 1:
Rebuttal: Thank you for the detailed comments.
$\textbf{W2:}$ For the experiment result, looking at the Rouge-L score, the differences between majority voting and d-PM isn't huge for both Blender and MIC datasets. In some cases, like Table 4, Row-BART-Greedy, Rouge-L and other metrics are giving conflictin... | Summary: This paper proposed to train a reward model to capture the disagreement among human raters and use it to calibrate NLG model's generation. The contributions includes: align text generation modes with human preference considering disagreement, a d-PM (preference modeling with disagreement) approach to train re... | Rebuttal 1:
Rebuttal: We are extremely grateful for the positive assessment and highly detailed and constructive feedback on our paper.
$\textbf{W1:}$ a simple baseline for reward model should be included: taking each human label as independent (without aggregation), with a cross entropy loss.
We appreciate the valu... | Rebuttal 1:
Rebuttal: We thank all the reviewers for their comments and valuable feedback. We have made the following major updates to improve our work further:
1. Following the suggestion of Reviewer $\color{blue}{\textbf{Ru2e}}$, we adopt GPT-4 as a reward model-based automatic evaluation metric to evaluate models. ... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
ResMem: Learn what you can and memorize the rest | Accept (poster) | Summary: Several recent works investigated the role of memorization in Deep Neural Networks as long as its interplay with generalization capabilities. In this context, this work proposes a two-stage procedure where a standard prediction model is augmented via a k-NN based mechanism that explicitly memorizes the model... | Rebuttal 1:
Rebuttal: We thank the reviewer for seeing the originality of the ResMem algorithm. We address your questions below.
> Joint training of kNN and DeepNet. Role of the duration of the first training phase? What happens when the second stage is applied too early or too late?
This is a very insightful quest... | Summary: The paper proposes ResMem (residual memorization) to improve neural network predictions by explicitly memorizing training residual patterns. The final prediction is computed by summing the neural network's output and the fitted residual obtained by KNN. The paper provides theories proving that ResMem yields fa... | Rebuttal 1:
Rebuttal: > The assumption of the proof (linear function, k=1) is not practical and not aligned with the experiment
We agree that the toy experiment with linear function and 1-nearest neighbor is very different from the actual empirical. The main purpose of the theory is to design an abstract setting that ... | Summary: This paper proposes a new learning algorithm, "ResMem," that combines deep neural networks (DNNs) with k-Nearest Neighbors (kNNs). The algorithm is a form of two-stage boosting: we first train a neural network. Then we construct a kNN algorithm that is trained to predict the residuals. The final output combine... | Rebuttal 1:
Rebuttal: Thanks for seeing the strength of the proposal and providing insightful comments! We will answer your questions below:
> A separate algorithm section.
We agree this is a good idea. The dilemma was that ResMem is a very general algorithm and can be applied to any supervised learning problem (reg... | Summary: The paper proposes the residual-memorization (ResMem) algorithm.
This algorithm aims to improve the generalization performance by explicitly memorizing the residual between the neural network's outputs and the training labels with a kNN regressor.
The paper demonstrates the validity of the algorithm from both ... | Rebuttal 1:
Rebuttal: We thank the reviewer for seeing the novelty of ResMem algorithm. We will fix the minor issues and adopt the suggestions.
We address Q1-Q4 below.
> ReQ1: equation under line 163.
We use the linearity of expectation and use the fact that
$\mathbb{E}\_{\mathbf{x}\sim\mathbb{P}\_{\mathbf{x}}} [\ma... | null | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
On quantum backpropagation, information reuse, and cheating measurement collapse | Accept (spotlight) | Summary: This paper explores the efficiency of training parameterized quantum models, from the perspective of backpropagation scaling. By leveraging some recent developments in shadow tomography and accessing multiple copies of a quantum state, the authors propose an algorithm that matches backpropagation scaling in qu... | Rebuttal 1:
Rebuttal: We are grateful for the positive feedback from the reviewer and thank them for taking the time to provide an insightful review. We seek to address the reviewer’s concerns and questions below and would be happy to have any further discussions on any points that may remain unclear:
Practical implic... | Summary: The paper explores whether parameterized quantum models can achieve comparable training efficiency to classical neural networks. From the perspective of reusing quantum information, the paper demonstrates that achieving backpropagation scaling in quantum models is not feasible without access to multiple copies... | Rebuttal 1:
Rebuttal: We would like to thank the reviewer for taking the time to go through the manuscript and provide useful feedback and suggestions. We are glad that the reviewer sees our work as relevant for the quantum machine learning community and we have drafted a response to the reviewer’s questions below:
Re... | Summary: The paper studies the scaling of computing the gradient of a quantum neural network. While in the classical case we can use backpropagation, which gives the same linear scaling for computing the gradient and the forward pass, in the quantum case, we would naively have to run a circuit for each component of the... | Rebuttal 1:
Rebuttal: We highly appreciate that the reviewer regards our paper as well-written and technically sound, and we are grateful for their comments and insightful feedback. We aim to answer the questions raised below and would be happy to provide any further details as needed:
Related work section: We would b... | Summary: The authors go over the backpropagation scheme for both classical and quantum machine learning methods. They also propose a novel quantum backpropagation algorithm based on quantum shadow tomography to reuse information and reduce the time complexity.
Strengths: 1. This paper provides a link between quantum ... | Rebuttal 1:
Rebuttal: We thank the reviewer for their helpful comments on our work and are grateful for the positive feedback. Regarding the reviewer’s questions, we have the following responses:
The proposed algorithm: While the proof that our algorithm works leans heavily on the existing machinery of shadow tomograp... | null | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Training Transformers with 4-bit Integers | Accept (poster) | Summary: This paper presents a 4-bit training method mainly to speed up the training process. In the forward propagation, Hadamard quantizer is proposed to suppress the outliers whereas structure sparsity and bit splitting are used for quantization of gradients. It has been shown that the proposed training method can s... | Rebuttal 1:
Rebuttal: We thank you for your comments and feedback. Below we provide a point-to-point response to all comments.
**W1**: limited speedup of this paper
**A1**: We would like to emphasize that our main contribution is the first INT4 training **algorithm**. The implementation is only a small fraction of ou... | Summary: The authors propose 4-bit training methods for Transformers. It first proposes a Hadamard quantizer with activation outliers issue being solved by Hadamard Transformation. Optimization for backpropagation is proposed by leveraging gradient sparsity with bit splitting and score sampling. The proposed 4-bit trai... | Rebuttal 1:
Rebuttal: We thank you for your comments and feedback. Below we provide a point-to-point response to all comments.
**Q1**: The comparison with SmoothQuant
**A1**: There might be some misunderstandings.
It's important to note that SmoothQuant is primarily designed for **INT8 inference** and not designed f... | Summary: The research paper presents a framework for training transformer-based neural networks using 4-bit integers, offering enhanced computational and memory efficiency.
The proposed method includes a dedicated Hadamard quantizer for forward propagation, designed to suppress outliers which typically cause degradati... | Rebuttal 1:
Rebuttal: We thank the reviewer for the acknowledgment of the potential and effectiveness of our work and the detailed constructive comments. Below we provide a point-to-point response to all comments.
**W1**: compare the performance and computation complexity with FP16 rather than examining its efficiency... | Summary: This paper deals with 4-bit training (fine-tuning) of transformers. For forward propagation, a Hadamard quantizer is proposed to solve the problem of outliers. For backpropagation, the authors leverage the structural sparsity of gradients by proposing bit splitting and leverage score sampling
techniques to qua... | Rebuttal 1:
Rebuttal: We thank the reviewer for the acknowledgment of our contributions and the careful constructive comments. Below we provide a point-to-point response to the reviewer's questions.
**Q1**: The proposed method introduces extra FP16 computations. The BS with selection needs data rearrangement, which c... | Rebuttal 1:
Rebuttal: Due to the length limit, we have included the references we utilized in the general rebuttal section.
[1] Zhu, Feng, et al. "Towards unified int8 training for convolutional neural network." Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2020.
[2] Zhao, Kang, e... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Failure-Aware Gaussian Process Optimization with Regret Bounds | Accept (poster) | Summary: 1. A failure-aware gaussian process optimization algorithm is proposed. A key difference from existing GP constrained optimization is that here we cannot obtain a direct observation from the latent failure function.
2. The algorithm's regret bound is presented.
3. Some practical issues are discussed.
4. Numeri... | Rebuttal 1:
Rebuttal: We thank the reviewer for the comments.
> Can we have a more "rigorous" theoretical analysis of the "real" algorithm(replacing set covering part by its heuristic approximation) and still derive the regret bound?
We consider the input domain of bounded hypercube, on which the NP-complete set-cove... | Summary: This paper studies Bayesian optimization in the setting where observing the unknown function might fail, for instance due to numerical simulator returning NaN. The authors propose a UCB-type acquisition function for this setting, which is based on first restricting the domain to a feasible regions where failur... | Rebuttal 1:
Rebuttal: We thank the reviewer for the detailed comments.
In the revised version of our paper, we will carefully fix the typos, grammar mistakes, and clarity issues, which the reviewer kindly pointed out.
We describe the answers to the questions of the reviewer below.
> Then, at certain points in time w... | Summary: The paper extends the gaussian process UCB algorithm to the case where some observations are failures. i.e. They solve for the case where x^* = argmax_{x in X} {f(x) such that c(x) =0}. Here c(x) = 1 may be thought of as an indicator of failure. The crucial part is to adaptively ignore the neighbourhood of the... | Rebuttal 1:
Rebuttal: We thank the reviewer for the comments.
However, we believe that there are some misunderstandings in the "Weaknesses" section that the reviewer pointed out.
> The definition of r_t says that when \hat{x}t is not observable, r_t is f(x^*) - min{x in X} f(x). This is just 0.
Since our setup consid... | Summary: This paper investigates GPBO under a robust setting, where the failure of observations is taken into account in the overall budget. An adaptive algorithm based on GP-UCB has been proposed to address this problem, and both theoretical and practical behavior have been analyzed and verified.
Strengths: Overall,... | Rebuttal 1:
Rebuttal: We thank the reviewer for the comments.
> To my knowledge, greedy elimination algorithms based on maximizing posterior variance usually achieve better theoratical convergence rates than UCB algorithms when it comes to robutsness considerations.
Do you have any idea that in the failure setting, a... | null | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Facilitating Graph Neural Networks with Random Walk on Simplicial Complexes | Accept (poster) | Summary: The authors generalize node-level random walk positional and structural encoding to edge-level (1-simplex) and higher-order simplex random walks. They focus on the edge-level for both theoretical and empirical results. Their approach is motivated primarily through theory, and then expanded empirically to the a... | Rebuttal 1:
Rebuttal: Thank you for your review points. We address your questions as follows.
Weakness
1. Applying to SOTA baselines: We have applied our method to SOTA baselines, including GRIT[5] and Exphormer[6]. We also improve the performance by improving the implementation and hyper-paremeters of Hodge1Lap and ... | Summary: A major issue with graphs is the absence of positional information on nodes, which decreases the representation power of GNNs to distinguish e.g., isomorphic nodes. An approach to tackle this issue is to introduce Positional Encoding (PE) and Structural Encoding (SE) of nodes and inject it with node features.... | Rebuttal 1:
Rebuttal: Thank you for your valuable reviews. We address the weakness and your questions as follows.
1. Test on more GNN architectures: In our new experiments, we apply our Hodge1Lap and EdgeRWSE to a variety of different GNN models, see Table 1 and general response for more details. For all base models, ... | Summary: The paper studies random walks over simplicial complexes to construct novel positional and structure encoding methods for improving graph neural networks.
In particular, the authors propose two strategies to construct edge embeddings, namely EdgeRWSE (via random walks on the edges) and Hodge1Lap (via spectral ... | Rebuttal 1:
Rebuttal: We thank the review for acknowledging our work and the useful advice. In the new revision of our paper, we will introduce RWMP in the main text. However, we believe that Section 6 is a necessary part. This section generalize our theories, showing that our method can be extended to any higher-order... | Summary: Positional and structural encodings have been widely studied, especially in the context of graph transformer architectures where such encodings often make up for the lack of inductive bias provided by the original graph, due to dense connectivity in the attention mechanism. RWSE (random walk structural encodin... | Rebuttal 1:
Rebuttal: Thank you for your constructive review. We address them all in the following.
Weaknesses:
(1) Typos: We will carefully proofread the paper, fix the typos and improve the writing.
(2) Comparisons to existing architectures based on simplicial complexes (e.g., [3], [4], [28]): First, [3, 4, 28] a... | Rebuttal 1:
Rebuttal: We sincerely thank all reviewers for their insightful and valuable review points. We add extensive experiments and report the results in the **pdf** attached to this common response. We answer some common questions in this response. Due to limited space, some experimental results and answers to th... | NeurIPS_2023_submissions_huggingface | 2,023 | Summary: Authors propose leveraging the random walk on different orders of the simplicial complex domain to enhance graph neural networks both in their performance and in the understanding of their theoretical expressivity.
$\textbf{Node level}$
Authors frame existing positional encoding (PE) and structure encoding ... | Rebuttal 1:
Rebuttal: Thank you for your valuable review and constructive points. We address them all as follows.
Weakness
1. Implementing higher-order and inter-order random walk: Thank you for your advice. We have newly implemented full EdgeRWSE and inter-order random walk in experiments, which we believe is huge c... | null | null | null | null | null | null |
On the Overlooked Structure of Stochastic Gradients | Accept (poster) | Summary: This paper studies two overlooked structures of stochastic gradients in deep learning. Specifically, the first part of the study conducts formal statistical tests on the distribution of stochastic gradients across parameters and iterations, and empirically showed that dimension-wise gradients exhibit power-law... | Rebuttal 1:
Rebuttal: We sincerely thank Reviewer AgKN for kind support and comments.
We also duly address your questions as follows.
Q1: It seems that the main observations of this paper also hold for random initialized deep models. While this shows that these structures may be a fundamental and common structure of ... | Summary: This paper focuses on investigating the stochastic gradient noise (SGN), defined as the difference between the full-batch gradient and the stochastic gradient. The authors conduct formal statistical tests to examine the distribution of SGN across both parameters and iterations. Their analysis reveals that dime... | Rebuttal 1:
Rebuttal: We sincerely appreciate Reviewer tTSf for kind support and helpful comments.
We duly address your concerns as follows.
Q1: The experiments conducted in this study primarily focus on LeNet (a CNN with approximately 60,000 parameters) and FCN models, especially when analyzing the covariance spectr... | Summary: This paper looks at the structure of stochastic gradient noise (SGN). They show that the noise for a single parameter, across minibatches/iterations is light-tailed, while looking across parameters, the noise is heavy-tailed. Further, they show that the spectrum across parameters is heavy-tailed.
Strengths:... | Rebuttal 1:
Rebuttal: We sincerely appreciate Reviewer 8wue for hard work and recognizing main contributions of our work.
We notice that Reviewer 8wue agrees on the contribution of our finding and the associated empirical and statistical evidences, while the main concern about lacking of theory is posted.
We frankly... | Summary: This paper studies some statistical properties of stochastic gradients, with a particular focus on their covariance.
The main finding is that the stochastic gradients have an approximately Gaussian distribution whose covariance spectrum follows a power law, which is robust to several changes in the setting.
E... | Rebuttal 1:
Rebuttal: We sincerely appreciate Reviewer jrig’s kind support and constructive comments.
We duly address your concerns as follows.
Q1: One main weakness is that I find it hard to understand what exactly is being plotted in Figure 1. The distinction between rows and columns of the gradient history matrix ... | null | NeurIPS_2023_submissions_huggingface | 2,023 | Summary: This paper makes two contributions. First, it conducts certain formal statistical tests on the distribution of stochastic gradients and gradient noise across both the parameters and iterations. Second, it further discover that the covariance spectra of stochastic gradients have the power-law structures overlo... | Rebuttal 1:
Rebuttal: We sincerely appreciate Reviewer HTRt’s kind support and constructive comments.
We duly respond to your comments as follows.
Q1: The experimental results are not so sufficient. The models utilized in the experiments are LeNet (LeCun et al., 1998), Fully Connected Networks (FCN), and ResNet18 (He... | null | null | null | null | null | null |
Grassmann Manifold Flows for Stable Shape Generation | Accept (poster) | Summary: This paper proposes a normalizing flow approach applicable to generate data on Grassmann manifold. The proposed method utilize the fact that Grassmann manifold is a quotient space of Stiefel manifold on which numerical calculation is feasible. Specifically, the method first constructs the flow from vector fiel... | Rebuttal 1:
Rebuttal: Thank you for the detailed comments and positive feedback indicating that the presented framework is non-trivial and novel!
We have prepared detailed responses to your questions below.
We sincerely hope that these responses will address and alleviate any concerns.
****
> **Question:1**
When we ... | Summary: This paper presents the application of continuous normalizing flow (CNF) to the Grassmann manifold for shape generation. The authors establish the theoretical foundations for learning distributions on the Grassmann manifold using CNF, enabling the generation of stable shapes. The proposed approach offers a pro... | Rebuttal 1:
Rebuttal: Thank you for your detailed comments and valuable feedback! We have prepared detailed responses to your questions below. We sincerely hope that these responses will address and alleviate any concerns.
****
> **Weakness:1. I have concerns regarding the significant contribution of the paper to the ... | Summary: This paper extends the previous Riemannian CNF to subspaces, i.e., Grassmann Manifold Gr(k,D), and tackle stable shape generation with nice transformation-invariant property. They derive the first and full mathematical formulation for learning invariant densities and flow on a Grassmann manifold. They successf... | Rebuttal 1:
Rebuttal: Thank you for the detailed comments and positive feedback indicating that the presented framework is non-trivial and novel!
We have prepared detailed responses to your questions below.
We sincerely hope that these responses will address and alleviate any concerns.
****
> **Weakness:1. From Figur... | Summary: The paper presents a novel concept: Continuous Normalizing Flows on a Grassmann Manifold (GrCNF), a model that presents a more efficient and accurate way to capture and represent the geometric features of data. The unique proposition of GrCNF is its application of the Grassmann Manifold to grasp data equivaria... | Rebuttal 1:
Rebuttal: Thank you for the detailed comments and positive feedback indicating that the presented framework is non-trivial and novel!
We have prepared detailed responses to your questions below.
We sincerely hope that these responses will address and alleviate any concerns.
****
> **Question:1. In the art... | Rebuttal 1:
Rebuttal: We have included a PDF with the figures as part of our response. Kindly review the PDF to access the pertinent information and visual illustrations.
Please note that the figures in the uploaded PDF are presented in a smaller size due to limitations in available space.
Rest assured, in the Camera-r... | NeurIPS_2023_submissions_huggingface | 2,023 | Summary: This paper is on developing techniques for sampling from probability distributions on Grassmann manifolds and utilizing it to generating shapes. The shapes are associated with the probability distributions on Grassmann and sampling from the distribution generates shapes. The sampling itself it based on flows u... | Rebuttal 1:
Rebuttal: Thank you for your detailed comments and valuable feedback!
We have prepared detailed responses to your questions below.
We sincerely hope that these responses will address and alleviate any concerns.
****
> **Weakness:1. I find this paper somewhat confusing to read. The presentation relies on ... | null | null | null | null | null | null |
Tree-Based Diffusion Schrödinger Bridge with Applications to Wasserstein Barycenters | Accept (spotlight) | Summary: The authors develop a novel method for approximating Wasserstein barycenters, which is based on diffusion models. This method combines ideas from the recently introduced Diffusion Schrodinger bridge and a framework for multi-marginal optimal transport with tree-structured costs.
Thus, although the paper focuse... | Rebuttal 1:
Rebuttal: We thank the reviewer for their thorough review of our work.
> My main question concerns the different update schemes mentioned in Section 3. The (mIPF) algorithm requires sampling from the whole tree [...]
We clarify our training procedure. In Proposition 2 and Proposition 3, we draw a bridge b... | Summary: Authors develop a new method for computing multi-marginal optimal transport in the case where the cost function is tree structured. Their algorithm cleverly exploits the dynamic formulation of the problem. The correctness of the algorithm is thoroughly demonstrated, and benefits are shown over a number of toy ... | Rebuttal 1:
Rebuttal: We thank the reviewer for their feedback and their positive evaluation.
> the paper is slightly hard to read [...]
While we acknowledge that the current paper might be heavy with mathematical notation, we thought it was necessary to include them to precisely describe our TreeDSB methodology (Se... | Summary: This paper generalizes the diffusion Schrödinger bridge framework to multidimensional problems with tree-structured quadratic cost function. The method builds on representing each egde in the graph as two SDEs (one for each direction) parameterized by neural networks. These are used to propagate the informatio... | Rebuttal 1:
Rebuttal: Thank you for your comments, we appreciate your acknowledgment of the paper’s merits.
> I think it would be nice with a little bit more information on the neural nets, about their structure and how they are used and updated.
All information on the neural nets used in our experiments are given i... | Summary: The authors consider the multi-marginal optimal transport (mOT) with entropic regularization and a tree-structured quadratic cost. The authors propose to address it by developing the Tree-based Diffusion Schrodinger Bridge (TreeDSB), i.e., dynamic and continuous state-space of the multimarginal Sinkhorn algori... | Rebuttal 1:
Rebuttal: Thank you for taking the time to review our submission and for your constructive feedback.
> The motivation [...] seem quite weak
We have provided a detailed answer to this point in the general response. First, TreeDSB introduces a new class of algorithms to solve EmOT using a dynamic framework.... | Rebuttal 1:
Rebuttal: We thank the reviewers for their insightful comments and are encouraged by their positive feedback regarding the proposition and soudness of our work. We provide detailed responses to each reviewer but summarize here their main feedback.
### 1. Motivation for the application to the Wasserstein Ba... | NeurIPS_2023_submissions_huggingface | 2,023 | Summary: Schröninger bridges (SB) can be understood as dynamical counterpart of the Entropic Optimal Transport problem.
This paper introduces the Tree Diffusion Schröninger Bridge (TDSB) model, which can be seen as a multi-marginal generalization of SB when the cost $c$ has a tree structure: given $\ell$ marginals $\... | Rebuttal 1:
Rebuttal: Thank you for your comments and thoughtful questions. Below we provide a detailed answer to the reviewer’s concerns.
> Why should one use this approach rather than existing ones?
We have provided a detailed answer to this point in the general response which we summarize here. First, contrary to ... | null | null | null | null | null | null |
Accelerated Training via Incrementally Growing Neural Networks using Variance Transfer and Learning Rate Adaptation | Accept (poster) | Summary: This paper introduces multiple synergistic tricks that allow for structurally growing networks to be competitive with and even sometimes outperform static networks. These methods include a batched schedule of layer widening, initializing opposing neurons, adapting the scale of the existing weights, handling op... | Rebuttal 1:
Rebuttal: # To Reviewer 8dnv
Thank you for the review and comments. We address your points individually below.
**Q: Formatting, margins, fonts.**
A: In the final version, we will correct margins and spacing in Tables 5 and 6, and increase font size in figures.
**Q: the caption of Figure 2 suggest you a... | Summary: This paper studies the growth of neural networks in order to reduce training time. Authors study various optimization challenges when training dynamically grown neural networks and propose some novel techniques to address these challenges. The paper is well written in general, though it requires few clarificat... | Rebuttal 1:
Rebuttal: # To Reviewer NBgL
Thank you for the review and comments. We address your points individually below.
**Q: Comparison with other methods in ImageNet is not ideal since baseline numbers might be different in different codebases; There is no limitation section at the moment. I think inability to m... | Summary: Starting from an optimization and training dynamics perspective, this paper proposes a new method to efficiently growing neural networks, based on a parameterization that dynamically stabilizes weight, activation, and gradient scaling as the architecture grows over time and different subnetworks take the lion'... | Rebuttal 1:
Rebuttal: # To Reviewer GgUS
Thank you for the review and comments. We address your points individually below.
**Q: The study is limited to simple, low-scale image classification and machine translation tasks, which aren't representative of the application domains often facing researchers and practitione... | Summary: The authors study the network growth problem and propose a novel problem to stabilize the optimization difficulty. The authors conduct experiments on various model architectures (including VGG, ResNet, MobileNet, and Transformer) and observe consistent performance improvements.
Strengths: The studied problem... | Rebuttal 1:
Rebuttal: # To Reviewer j2wb
Thank you for the review and comments. We address your points individually below.
**Q: The most important application for the proposed algorithm is large-scale training. However, the experiments discussed in this study are based on relatively small models, which left doubts o... | Rebuttal 1:
Rebuttal: # General Reply
We thank the reviewers for the detailed comments and answer specific questions in individual responses to each review below. We also conducted additional experiments on learning rate adaption (prompted by Reviewer 8dnv) and slightly longer training (prompted by Reviewer NBgL). P... | NeurIPS_2023_submissions_huggingface | 2,023 | Summary: This paper developed an approach to efficiently grow neural network architecture by parameterization scheme with regard to weight, activation, and gradient scaling. It employed a learning rate adaptation mechanism rebalancing the gradient contribution of the subnetworks. The experimental results show efficienc... | Rebuttal 1:
Rebuttal: # To Reviewer iBKk
Thank you for the review. However, it appears there is a serious misunderstanding as many of your questions make incorrect assumptions or are directly answered in the paper itself. Please see our clarifications in response to your individiual points.
**Q: Wondering if the au... | Summary: This paper presents a method for incrementally growing network width across stages. The reader is shown the details of the growing process using a 3-layer feedforward network as an example, with the appendix providing details for other types of networks. Additional algorithmic details, including learning rate ... | Rebuttal 1:
Rebuttal: # To Reviewer Tyrx
Thank you for the review and comments. We address your points individually below.
**Q: A few parts of the paper are a bit terse. In particular, Section 4 of the Appendix could benefit from additional exploration of, e.g., more alternatives or motivation/derivation for using th... | null | null | null | null |
CBD: A Certified Backdoor Detector Based on Local Dominant Probability | Accept (poster) | Summary: The paper presents a certified backdoor detector for backdoor poisoning attacks based on local dominant probability. Certified approaches for backdoor poisoning attacks have been studied recently. These approaches usually have low certified radii on large datasets. This paper goes in another direction by certi... | Rebuttal 1:
Rebuttal: # Response to Reviewer XxVd
We sincerely thank you for your time and effort in reviewing our paper and your positive comments. As you mentioned, we investigate certified backdoor defense in a different direction from existing ones by certifying whether the victim model has learned the backdoor. W... | Summary: The paper proposes a certified backdoor detector (CBD) based on a novel, adjustable conformal prediction scheme using a proposed statistic named local dominant probability. The proposed approach claims to be effective against multiple types of backdoor attacks on multiple datasets.
Strengths: The paper is wel... | Rebuttal 1:
Rebuttal: # Response to Reviewer xfbM
We sincerely thank you for your time and effort in reviewing our paper. We also thank you for recognizing the importance of our work, the soundness of our theoretical analysis, and the quality of our presentation. In the following, we address your concerns in detail.
_... | Summary: The authors propose a new method for whether machine learning detecting models have been trained with a backdoor trigger, with the following interesting property: if the trigger has sufficiently small L_2 norm, *AND* the trigger is sufficiently robust to Gaussian noise (i.e., the model will likely return the t... | Rebuttal 1:
Rebuttal: # Response to Reviewer fRFm
Thank you for your time and effort in reviewing our paper, and your positive comments on the novelty and the presentation of our work. We are especially encouraged by your appreciation of our certification approach as reflected in your accurate summary of our work. In ... | Summary: This paper proposes CBD, a certified detection method for backdoor attacks. It leverages a newly proposed statistic: local dominant probability to capture the target label probability increment in the neighborhood of infected instances. The theoretical analysis provides both detection inference and the conditi... | Rebuttal 1:
Rebuttal: # Response to Reviewer jLvZ
Thank you for your time and effort in reviewing our paper and your positive comments. In the following, we address each of your concerns in detail.
___
**Q1 :** Time efficiency of CBD
**A1 :** Thank you for your comment. We discuss the time efficiency of CBD from three... | Rebuttal 1:
Rebuttal: # Summary
We sincerely thank the reviewers for their constructive feedback and suggestions. We are glad that the reviewers found our paper novel and effective. Our responses are summarized as follows:
* Experiments added as suggested by the reviewers:
1. We showed the performance of our CBD wit... | NeurIPS_2023_submissions_huggingface | 2,023 | Summary: This paper proposes a certification method to detect backdoored models. It is based on the assumption that backdoor triggers are robust, meaning that adding random noises onto trigger-injected samples will not change their predictions (to the target label). It also assumes that the neighborhood of a given inpu... | Rebuttal 1:
Rebuttal: # Response to Reviewer MHNa
Thank you for your time and effort in reviewing our paper. In the following, we address each of your concerns in detail.
___
**Q1:** Why is the proposed method certified?
**A1:** Thanks for your question. Our certification consists of the following two parts.
For the... | null | null | null | null | null | null |
Training Fully Connected Neural Networks is $\exists\mathbb{R}$-Complete | Accept (poster) | Summary: The paper investigates the algorithmic complexity of training fully connected ReLU neural networks. It is shown that perfectly fitting two-dimensional data by even a one-hidden layer network with two output neurons is $\exists\mathbb{R}$-complete, namely it is as hard as finding the roots of a multivariate pol... | Rebuttal 1:
Rebuttal: We thank you for carefully reading and assessing our paper and for the valuable feedback.
We agree that our reduction only works for perfect fitting of the data and two-dimensional outputs, as we also discuss in the paper. While the first property is inherent to the nature of the existential theo... | Summary: The paper proves ER-completeness of training (empirical risk minimization) a fully connected two-layer ReLU network with two inputs and two outputs. The result contributes to the understanding of theoretical complexity of neural network training at an important, fundamental level. The authors also did a good j... | Rebuttal 1:
Rebuttal: We thank you for carefully reading and assessing our paper and for the valuable feedback.
As other reviewers had similar suggestions, we plan to enhance the proof ideas section, especially concerning the interplay of the gadgets as well as the equivalence of the training problem and the ETR-INV i... | Summary: This paper shows that training a neural network with two input neurons and two output neurons is ER complete where ER stands for the existential theory of the reals.
Strengths: A new idea (reduction) to prove hardness result. The paper is generally well written and provides a good cover of the (dense) related... | Rebuttal 1:
Rebuttal: We thank you for carefully reading and assessing our paper and for the valuable feedback.
We also thank you very much for the detailed list of suggestions to improve the presentation of our paper, which we happily take into account. See below for some detailed responses:
1+5: We agree and plan to... | Summary: This paper strengthens a result by Abrahamsen, Kleist and Miltzow [Neurips 21] by showing that empirical minimization problem of two-layer neural network is \exist R-hard. They further show that arbitrary algebraic number is required as optimal weights even with rational data points. The example include a ReLU... | Rebuttal 1:
Rebuttal: We thank you for carefully reading and assessing our paper and for the valuable feedback.
Our definition of neural networks as directed acyclic graphs is in line with the definition by Abrahamsen, Kleist and Miltzow (NeurIPS 21) whose result we strengthen. While this generality makes sense in the... | null | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Functional Renyi Differential Privacy for Generative Modeling | Accept (poster) | Summary: Standard differential privacy (v-DP) provides mathematical guarantees on the closeness of distributions of real-valued, randomized functions when they act on adjacent inputs. v-RDP is a variant of the standard DP definition which can be connected back to v-DP guarantees, and can provide tighter v-DP bounds thr... | Rebuttal 1:
Rebuttal: Due to the length limit, we will simplify the reviewer's question when we quote.
> Q1. Originality: The paper combines RDP and f-DP, and does not explore any new properties of proposed f-RDP.
**Response:** We believe that we have developed the necessary tools for the practical use of f-RDP, and w... | Summary: The work generalizes the DP to infinite dimensions. Based on the modeling of Hall et al., the authors propose some characteristics of functional DP aligning with the traditional DP. The most interesting part is adopting their proposed framework to RHKS for an infinite functional space. Some numerical studies i... | Rebuttal 1:
Rebuttal: > Q1. The technical contribution should be clearly stated. Some remarks about the different characteristics of the functional DP and traditional Dp in Subsample and others should be stated.
**Response:** Thanks for your suggestion. Our primary technical contribution is the functional RDP with it... | Summary: The authors extend the concept of RDP from vectors to functions. They introduce the notion of functional RDP and present a few properties of the functional RDP under composition / amplification by sampling. Finally, they demonstrate the functional RDP using the application in RKHS.
Strengths: 1. They gener... | Rebuttal 1:
Rebuttal: > Q1. My main concern is the novelty of the paper: Their results on RDP for functions (Definition 5) and subsampled RDP for functions directly generalize from the existing RDP literature, except that the output space might now have an infinite-dimension. However, assuming a bounded RDP divergence ... | Summary: Conceptually, this paper combines the idea behind RDP and f-DP and proposes f-RDP.
Theoretically, this paper proved several important properties of f-RDP (e.g., Composition, post-processing, and the privacy level for the widely-used Gaussian mechanism).
Experimentally, this paper showed f-RDP can be applied... | Rebuttal 1:
Rebuttal: > Q1. Some high-level explanations are missing and it's hard for the readers out of DP community to understand Section 2.
**Response:** Thanks for your suggestion. We will add more explanations and summarize a paragraph in the beginning of Section 2.
> Q2. May I know what is the high-level idea... | Rebuttal 1:
Rebuttal: We thank all reviewers for their insightful reviews. There is a doubt on the value of the proposed functional-RDP, especially in the DPGM application, e.g., why we need functional-RDP and why it is better than v-DP in this application. Besides, reviewer 1DXa raised a practical and interesting ques... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Distributional Learning of Variational AutoEncoder: Application to Synthetic Data Generation | Accept (poster) | Summary: In this paper, the authors present a novel nonparametric distributional learning method for VAE. They provide a comprehensive theoretical analysis of the proposed solution (DistVAE), as well as prove its superiority over CTGAN, TVAE, and CTAB-GAN when applied to synthetic data generation on real tabular datase... | Rebuttal 1:
Rebuttal: > Q1 (Question). The authors claim that their approach "extends model capacity [...] without sacrificing the computational advantages of the VAE framework". Perhaps the paper would benefit from an experimental comparison (in terms of capturing data distribution and computational efficiency) of the... | Summary: This paper introduces a novel distributional learning method of VAE, which aims to effectively capture the underlying distribution of the observed dataset using a nonparametric approach. The proposed decoder is composed of an infinite mixture of asymmetric Laplacian distribution, which possesses general distr... | Rebuttal 1:
Rebuttal: > Q1 (Weakness). As [15] indicated, “the model training becomes unstable if the estimated variance of the decoder shrinks to zero in Gaussian VAE” is not theoretically correct.
Thank you for bringing up the issue regarding the citation of related papers in our work. We sincerely appreciate your v... | Summary: In this paper the authors propose to define the decoder of a VAE using an infinite mixture of asymmetric Laplacian distribution, which results in distributional learning. Specifically, they design a non-parametric generative model in the VAEs. This class of VAEs estimates the conditional CDF while trained with... | Rebuttal 1:
Rebuttal: > Q1 (Question). My main confusion is about the 2 dimensional latent space in the evaluations. If I understand correctly, all the models are defined with 2D latent space. I wonder if this puts limits on the decoder capacity by itself, and I wonder how the model performs when we do higher latent di... | null | null | null | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Unified Segment-to-Segment Framework for Simultaneous Sequence Generation | Accept (poster) | Summary: This paper proposes a unified Seg2Seg framework for simultaneous sequence generation tasks. The framework introduces a latent segment as the pivot between the source and target, exploring all potential mappings via the proposed expectation training. With this approach, the authors achieve high-quality generati... | Rebuttal 1:
Rebuttal: Thanks for your careful and valuable comments, and we will refine the paper following your suggestions. Following, we will respond to your questions in detail.
Q1: It is unclear how the prediction of target tokens is learned?
A1: Seg2Seg leverages the Transformer as the backbone and addi... | Summary: This paper presents an approach (Seg2Seg) for simultaneous sequence generation, which attempts to balance high generation quality with low latency. Rather than processing the entire source sequence at once (incurring high latency), the model splits it into segments. Latent segments are used to aggregate source... | Rebuttal 1:
Rebuttal: Thanks for your careful and valuable comments, and we will refine the paper following your suggestions. Following, we will respond to your questions in detail.
Q1: About the notion of adaptive wait policies?
A1: Indeed, I agree with you that the concept of adaptive policies existed befor... | Summary: This paper proposes a method for simultaneous sequence to sequence modeling.
The proposed approach introduced a latent segment, which represents the number of consecutive reads followed by consecutive writes.
The method is applied to three simultaneous sequence to sequence modeling task, streaming ASR, simulta... | Rebuttal 1:
Rebuttal: Thanks for your careful and valuable comments, and we will refine the paper following your suggestions. Following, we will respond to your questions in detail.
Q1: The design choice of Emb(xj) instead of using encoder output states?
A1: Thank you for raising a question regarding the calc... | Summary: The paper addresses the task of simultaneous sequence generation in real-time scenarios, where the target sequence is generated while receiving the source sequence. The authors propose a unified segment-to-segment framework called Seg2Seg to tackle this challenge. Seg2Seg aims to learn the optimal moments for ... | Rebuttal 1:
Rebuttal: Thanks for your careful and valuable comments, and we will refine the paper following your suggestions. Following, we will respond to your questions in detail.
Q1: The difference and superiority of Seg2Seg compared to MMA?
A1: Compared with MMA, there are several fundamental distinctions... | Rebuttal 1:
Rebuttal: Thanks to all the reviewers for their careful and valuable comments.
Following the reviewers’ suggestion, we additionally report the results on a commonly used SimulMT benchmark IWSLT15 English→Vietnamese (En→Vi), to more comprehensively evaluate the performance of Seg2Seg on SimulMT. As ... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Robustifying Generalizable Implicit Shape Networks with a Tunable Non-Parametric Model | Accept (poster) | Summary: The paper presents a novel shape reconstruction method (point cloud-> reconstructed implicit shape) that combines a deep shape prior and a non-parametric interpolant. Existing deep shape reconstruction methods are typically trained on synthetic datasets (to obtain full supervision) and suffer from a domain shi... | Rebuttal 1:
Rebuttal: ### “The paper could be written more clearly”:
We will improve the writing of the paper for the next version.
### “The method seems to rely on a pretrained occupancy network. Does this affect the performance or expressiveness of the technique”:
The very objective (line #5 #39) of this work is ... | Summary: To address the poor performance of feed-forward models on noisy real data, the authors propose using test-time Kernel Ridge Regression (KRR) to fine-tune the results for implicit shape networks, specifically in the context of recovering 3D reconstruction surfaces from noisy point clouds. The method involves in... | Rebuttal 1:
Rebuttal: ### “1. The paper does not mention the overhead of the current method. Since the KRR model needs to be optimized for different point cloud inputs, it is crucial to report the optimization time in this paper”:
As reported line #213, for a 10k sized input point cloud, our method takes around 10 sec... | Summary: The paper proposes a method for improving shape reconstruction from a set of sparse and noisy unoriented point clouds. As data-driven trained networks can have failures on out-of-distribution examples, this method proposes to take the input point cloud, the predicted surface points, and the features generated ... | Rebuttal 1:
Rebuttal: ### “Q1/W1. Why is robustifying pre-trained shape reconstruction networks the desired approach to improving shape reconstruction, instead of directly designing a reconstruction method with kernels ? What benefit does this alternative direction offers - why post-correct rather than inducing these p... | Summary: This paper aims at the task of implicit surface reconstruction from noisy point clouds. It tackles the model generalizability and robustness by introducing a kernel ridge regression as a implicit function that maps a shape feature to occupancy.
Strengths:
The proposed method achieve higher accuracy compare... | Rebuttal 1:
Rebuttal: ### "a.What is the motivation behind KRR? How does it help the generalizability of the models? Why it is fitted for surface reconstruction from noisy point clouds? The authors should demonstrate the intuition behind the introduction of KRR, compared to other occupancy decoders."
While feedforward... | Rebuttal 1:
Rebuttal: We thank the reviewers for their valuable feedback. We will improve writing in the next version and correct typos pointed out by reviews. We will add the additional discussions and results provided in the rebuttal. We address the questions separately in the respective rebuttals and we summarize al... | NeurIPS_2023_submissions_huggingface | 2,023 | Summary: This paper proposes an efficient mechanism to address the generalization issues often encountered by generalizable feedforward models trained for implicit shape reconstruction from unoriented point clouds. The solution proposed in this work is to use kernel ridge regression in the network's feature space, tune... | Rebuttal 1:
Rebuttal: ### “What does the k_theta in L184 represent?”
As reported line #217, k_theta is the Gaussian kernel with a single length-scale for each data-dimension in our experiments. We will clarify this. We note that by using a Gaussian kernel we benefit from the universality properties of the associated ... | Summary: This paper aims to reconstruct smooth and high-quality surface from noisy point cloud. While previous works like Poisson surface reconstruction can extract surface from point cloud, but it's sensitive to input noise. The existing learning-based methods are able to handle noisy point cloud, but often suffer fro... | Rebuttal 1:
Rebuttal: ### “What's the fundamental difference between the proposed method and previous kernel-based neural surface reconstruction?”:
CVPR 23 Proceedings were not available yet at the Neurips 23 deadline. Existing learning based kernel reconstruction methods ([78] Neural splines CVPR 2021, [76] Neural ke... | null | null | null | null |
Joint Data-Task Generation for Auxiliary Learning | Accept (poster) | Summary: This paper tackles the problem of auxiliary learning, specifically for cases where the provided auxiliary data actually harms the primary task.
A joint data-task generation framework is proposed, which generates new auxiliary tasks during the optimization process.
The label generator generates data in the feat... | Rebuttal 1:
Rebuttal: Thank the reviewer for the detailed comments and constructive suggestions. We address your concerns point by point as follows.
1. **About the optimization strategy**: The bi-level optimization has been adopted by previous works[1,2,3], but the instance regularization we adopt is quite different f... | Summary: This paper focuses on Auxiliary Learning, a technique aiming to improve the model generalization ability on the primary task with the help of related auxiliary tasks, and proposes a novel joint data-task generation framework for auxiliary learning (DTG-AuxL). DTG-AuxL is designed to mitigate certain issues pre... | Rebuttal 1:
Rebuttal: Thank the reviewer for the detailed comments and constructive suggestions. We address your concerns point by point as follows.
1. **About Table 3 and Table 4**: We will make them better formatted in our final version. Thanks for your suggestion.
2. **More experiments**: We follow the reviewer's... | Summary: This paper studies the problem of learning with auxiliary tasks. Motivated by the observation that in most scenario the auxiliary data is hard to collect, the authors propose to train a model that can automatically generate auxiliary data (and labels) for any task at hand. Specifically, the authors propose a d... | Rebuttal 1:
Rebuttal: Thank the reviewer for the detailed comments and constructive suggestions. We address your concerns point by point as follows.
1. **Reply to Weakness 1**:
**The impact of generated data and task on the backbone**: The generated data not only influences the task-specific head but also the back... | Summary: Motivated by the manually collected auxiliary data and tasks in auxiliary learning may have negative impacts on the primary task, this paper proposes an auxiliary learning framework that jointly generates new auxiliary data and task. To generate new data for various auxiliary scenarios, the authors design an a... | Rebuttal 1:
Rebuttal: Thank you for your careful comments and constructive suggestions. We address your concerns point by point as follows.
1. **Reply to Weakness1**:
**Discussion about generation with stable diffusion**: Stable Diffusion is a pretrained text-to-image generation model, and it relies on large data... | null | NeurIPS_2023_submissions_huggingface | 2,023 | Summary: This paper proposes an auxiliary learning framework that jointly generates new auxiliary data and new labels, to tackle the negative transfer problem in auxiliary learning. The framework is designed to be generic to various auxiliary scenarios, both multiple input and single input, regression and classificatio... | Rebuttal 1:
Rebuttal: Thank the reviewer for the constructive comments and suggestions. We address your concerns point by point as follows
1. **Rely to Weakness 1**:
**Comparison with MixUp**:
We agree with the reviewer that comparison with MixUp is important in the shared-input scenario, so in Appendix 5.4... | null | null | null | null | null | null |
Robust covariance estimation with missing values and cell-wise contamination | Accept (poster) | Summary: The paper “Robust covariance estimation with missing values and cell-wise contamination” presents an unbiased estimator for covariance in the presence of missing values that do not require any imputation step, and the authors claim they achieve competitive results compared to other state-of-the-art methods. T... | Rebuttal 1:
Rebuttal: We thank the referee for the detailed feedback.
We would like to address first the weaknesses pointed out by the referee.
- The approximately low-rank assumption is standard in the literature on the theoretical analysis of covariance estimation. For many datasets including numerous UCI datasets... | Summary: The objective of this paper is to mitigate the influence of outliers and missing values in covariance estimation without resorting to imputations. A particular type of missing data that satisfies assumption (1) is considered and an unbiased estimator of the true covariance that achieves the minimax bound wit... | Rebuttal 1:
Rebuttal: We thank the referee for the careful reading of our paper.
We discuss first the weaknesses identified by the referee.
- W1. Regarding the first identified weakness, we decided to present first the classical MCAR setting for the sake of clarity. But we also covered the heterogeneous missingness ... | Summary: This work proposes an unbiased estimator for the covariance matrix under missing entries. Unlike most existing approaches, the proposed estimator does not require an imputation step to discard missing values before computing the covariance matrix. Interestingly, the proposed method employs an approach that avo... | Rebuttal 1:
Rebuttal: We thank the referee for the careful reading of our paper. As recommended by the referee, we will precise the definitions of the metrics in the revision of this paper.
- Q1. STD stands for standard deviation. We will precise it in the table caption.
- Q2. The formula for the relative spectral di... | Summary: This paper studies the problem of estimating a covariance matrix in the presence of missing observations and/or outliers.
The authors first consider the missing values framework and theoretically analyse the unbiased covariance estimator introduced in [Karim Lounici. High-dimensional covariance matrix estimati... | Rebuttal 1:
Rebuttal: Q1. We thank the referee for pointing out the issue in the proof. We actually found it after submitting the paper and found an alternative simpler argument leading to an improved bound.
Indeed Ineq. (29) in our paper should be:
$$
\Vert\widehat{\Sigma} - \Sigma\Vert \leq \delta^{-2} \Vert diag ... | Rebuttal 1:
Rebuttal: We wish to thank the reviewers for their insightful evaluation of our paper. We appreciate all the comments and remarks, which we will implement in our revision. We discuss here some points that were raised by several referees.
The goal of our paper is to study the robust covariance estimation p... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
REFINE: A Fine-Grained Medication Recommendation System Using Deep Learning and Personalized Drug Interaction Modeling | Accept (poster) | Summary: In this paper, the authors present a deep learning framework that utilizes patients' electronic health records and drug information to recommend safe and accurate medication combinations based on their symptoms and examination results. The proposed method is evaluated on two real-world datasets, MIMIC-IV and P... | Rebuttal 1:
Rebuttal: We would like to thank the reviewer for highlighting the strong points and shortcomings of our work. We have addressed the queries as follows (numbers are in order of query raised):
1) Evaluation in terms of Jaccard as well as three other metrics have already been provided in Table 1 under Append... | Summary: This paper proposes a new medication recommendation model that learns patient representations from diagnoses, lab tests, and medications. The new aspect of the paper is that it considers different severities of drug-drug interactions and can prevent overly conservative drug prescriptions.
Strengths: 1. The bi... | Rebuttal 1:
Rebuttal: We would like to thank the reviewer for the thoughtful reviews. We have addressed some of the queries and have summarized as follows:
1) The contribution of the paper is below the NeurIPS acceptance bar.
We would like to emphasize that our paper introduces an innovative fine-grained medication r... | Summary: This paper presents a medication recommendation system with 2 components: a GAT based medication interaction and a ViT-based patient embedding. The results on MIMIC and PRIVATE shows its superior performance.
Strengths: 1: I think this paper is well-presented. The structure is very clear and the writing is po... | Rebuttal 1:
Rebuttal: We would like to thank the reviewer for highlighting the strong points and shortcomings of our work. We have addressed some of the queries and have summarized them as follows:
1) Given that the proposed work is a concatenation of ViT and GAT, the level of novelty is questionable.
We would like t... | Summary: This manuscript proposes a new system for personalized medication recommendation. The method relies on the history of patient visits: diagnoses, lab responses and medication dosages across time. For obtaining safe recommendations, it uses severity level information of drug interactions. An important aspect of... | Rebuttal 1:
Rebuttal:
We would like to thank the reviewer for the thorough analysis and comments on our work. We have addressed the queries as follows:
1) Existing works have focused on coarse-grained medication recommendations at the class level since making drug-level predictions introduces more complexity. In the ... | Rebuttal 1:
Rebuttal: Response to ethical concerns.
We take data privacy and security very seriously, and we would like to reassure the reviewer that all necessary precautions were taken to ensure that no ethics regulations were violated in the use of the private dataset for this research. The data has been appropriat... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
STXD: Structural and Temporal Cross-Modal Distillation for Multi-View 3D Object Detection | Accept (poster) | Summary: This paper presents a novel cross-modal knowledge distillation framework for multi-view 3D object detection, namely STXD, with three effective strategies.
(1) correlation-based cross-modal feature distillation, which gets rid of the domain gap on the feature level during distillation.
(2) feature similarit... | Rebuttal 1:
Rebuttal: We appreciate the recognition from reviewer "U3eQ" for acknowledging our "new perspectives on knowledge distillation for 3D detectors," particularly our "temporal-based feature distillation" approach. "U3eQ" also noted the effectiveness of STXD, validated through extensive experiments, and its "pr... | Summary: This work presents a cross-modal knowledge distillation framework for the multi-view, called STXD. It tries to distill structural and temporal knowledge across different modalities (from LiDAR to Camera) from three aspects, namely Correlation Regularizing Distillation (CD), Temporal Consistency Distillation (T... | Rebuttal 1:
Rebuttal: We thank the reviewer "X2eU" for acknowledging our cross-modal distillation framework is "promising" and has "potential usage for application" as well as recognizing our "sufficient experiments" showing "significant improvements" over the baselines. We hope you find our responses provided below he... | Summary: This paper proposes a novel structural and temporal cross-modal knowledge distillation (STXD) framework for multi-view 3D Object Detection. STXD leverages the Correlation Regularizing Distillation loss to address the distributional divergences between non-homogeneous modalities. Besides, it effectively distill... | Rebuttal 1:
Rebuttal: We appreciate the acknowledgment from reviewer "XsGg" for recognizing "effectiveness" and "valuable insights" of our non-homogeneous cross-modal knowledge distillation framework, particularly the use of Correlation Regularizing Distillation (CD) as well as our comprehensive ablation studies. Below... | Summary: This paper introduces a cross-modal distillation framework named STXD, which addresses existing distillation problems from both structural and temporal perspectives. Structure-based distillation imposes constraints on the similarity matrix of the student and teacher, ensuring that corresponding positions have ... | Rebuttal 1:
Rebuttal: We thank the reviewer "1M3u" for taking time and effort to review our work. We also appreciate the reviewer's recognition of the competence of our knowledge distillation approach, clear presentation, and comprehensive component-wise analyses. Below, we provide responses to your concerns, and we ho... | Rebuttal 1:
Rebuttal: Dear AC and Reviewers,
First of all, we sincerely appreciate your time and effort to review our work. At the end of this comment, we have attached our **Rebuttal PDF**, which includes additional experimental results and diagrams illustrating the comparisons with prior methods. Tables and figures ... | NeurIPS_2023_submissions_huggingface | 2,023 | Summary: The authors propose a new method for LiDAR-to-Camera 3D object detection model distillation. They propose distillation along three avenues. First, referencing Barlow Twins, they consider redundancy reduction through cross-correlation between LiDAR and Camera features. Second, they enforce consistency between t... | Rebuttal 1:
Rebuttal: We express our gratitude to the reviewer "4Udp" for recognizing our fundamental contributions to feature-level distillation approaches, particularly highlighting the significance of Temporal Consistency Distillation, and acknowledging our thorough validation results on a component-wise basis. Belo... | null | null | null | null | null | null |
Demographic Parity Constrained Minimax Optimal Regression under Linear Model | Accept (poster) | Summary: This paper studies the linear regression problem under a new definition of $(\alpha, \delta)$-fairness consistency as a fairness constraint.
In particular, the authors derived that under the constraint of $(\alpha, \delta)$-fairness, the minimax optimal error is $dM/n$ when $\alpha$ is less than 1/2. Moreover,... | Rebuttal 1:
Rebuttal: > It would be great if the authors could discuss and provide more understanding on how restrictive the assumption that $\alpha\le 1/2$ is. Is it a necessary condition for achieving the minimax rate of dM/n? Also, how should one think about solving for the case that $\alpha > 1/2$?
Thank you for ... | Summary: This paper investigates the minimax error in regression under the constraint of demographic parity (DP). In contrast to earlier works, this paper studies both direct and indirect discrimination and their effects on the estimator. In addition to proposing an algorithm to construct a regressor achieving a desira... | Rebuttal 1:
Rebuttal: > In your model, you mentioned that indirect discrimination mainly impacts $\mu_S$ while direct discrimination is reflected via $\beta^*_S$. My understanding is that, direct discrimination is for the "model" itself, i.e. $\beta$, while indirect discrimination has influence on the data set. Is my d... | Summary: The authors study the problem of minimax optimality for the linear regression model under demographic parity constraints from the fairness literature. They provided matching upper and lower bounds for the model that is considered in the paper.
Strengths: Overall:
The paper is well-motivated by presenting bo... | Rebuttal 1:
Rebuttal: > (1) How does this compare to the results of Chzen and Schreuder is unclear. There is no unified setting which captures both models.
Thank you for drawing attention to the comparison between our work and that of Chzen and Schreuder. Your observation is astute, and we appreciate the opportunity ... | Summary: This paper studies the statistical minimax error of building a linear regression model under the constraint of demographic parity.
It generalizes the result in Chzhen and Schreuder [5] by allowing the slope coefficients to change with sensitive attribute rather than just
the intercepts. The authors provide bo... | Rebuttal 1:
Rebuttal: > The statement made in line 136-137 should be explained more clearly. In particular, why does enforcing strict demographic parity result in a constant function?
Thank you for your query regarding the explanation provided in lines 136-137. The occurrence of this particular phenomenon can be bette... | Rebuttal 1:
Rebuttal: We would like to express our gratitude to all the reviewers for dedicating their time to providing such careful reviews. We also greatly appreciate the high evaluations that our papers received from the reviewers. Please refer to the individual responses for our answers to the questions the review... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Score-based Generative Modeling through Stochastic Evolution Equations in Hilbert Spaces | Accept (spotlight) | Summary: In this paper, the authors introduce Hilbert Diffusion Model (HDM). This generative model is a generalization of diffusion models to infinite dimensional state space (in this case Hilbert spaces) through the use of generalized SDEs. The authors extend the notion of score and time-reversal to this setting. Thei... | Rebuttal 1:
Rebuttal: First of all, [Bond-Taylor & Willcocks, 2023] is a paper we had not cited at the time of submission, and we are very grateful to the reviewer for pointing it out. For the first main remark, Please see the **Author Rebuttal by Authors** for more details.
For the second remark, real applications c... | Summary: The authors propose a framework called the Hilbert Diffusion Model (HDM), which generalizes and unifies the time-reversal diffusion generative models with stochastic differential equations (SDEs) in Hilbert spaces. This model enables SDEs to be used in infinite-dimensional settings and to include stochastic pa... | Rebuttal 1:
Rebuttal: We sincerely appreciate the reviewer’s valuable feedback.
> ***This work is not the first to extend time-reversal diffusion models to infinite-dimensional space, and the concept of time-reversal diffusion on Hilbert space is already addressed in [1][2][3]. As these are mentioned in your citation... | Summary: This work studies continuous-time diffusion processes in Hilbert spaces, and their applications to generative modeling. In particular, a time-reversal formula is derived for Hilbert-space SDEs, allowing the continuous-time diffusion model framework to be generalized to infinite-dimensional Hilbert spaces. The ... | Rebuttal 1:
Rebuttal: We greatly appreciate the reviewer’s constructive feedback.
> ***The paper could be better contextualized within the existing literature on this topic, i.e. there is no related work section.***
In fact, we compare our approach to the existing literature in Appendix C.1, but we do not have a corr... | Summary: The paper proposes a time-dependent stochastic generative model that can be implemented in Hilbert spaces. The presented method unifies diffusion processes (commonly defined in diffusion models) with a more general stochastic evolution equation. A stochastic evolution equation is defined with satisfied conditi... | Rebuttal 1:
Rebuttal: Thank you for your valuable feedback. In the camera-ready version, we will supplement additional explanations regarding the dataset, Figures 1 and 3, and the theoretical background such as smooth cylinder functions and Fomin differentiable measures, which were briefly discussed due to the space li... | Rebuttal 1:
Rebuttal: We appreciate the reviewers' valuable comments. Since some reviewers question the theoretical difference between the proposed method and the existing methodologies, e.g., [Kerrigan et al., 2023, Lim et al., 2023, Pidstrigach et al., 2023], etc., we would like to address those issues. We promise to... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
$\varepsilon$-fractional core stability in Hedonic Games. | Accept (poster) | Summary: This paper introduces the concept of an $\epsilon$-fractional core stability, relaxing the standard notion of core stability by allowing at most an $\epsilon$ fraction of all possible outcomes to core-block. They show that for sufficiently small values of $\epsilon$, $\epsilon$-fractional core stability always... | Rebuttal 1:
Rebuttal: We thank the reviewer for carefully reading our paper and for the constructive remarks. In the following we will address the comments in the 'Weaknesses' and 'Questions' sections.
**Limitations of the definition:** A possible scenario where the notion of $\varepsilon$-FC stability shows its limit... | Summary: The manuscript at hand proposes a novel relaxation of core stability in coalition formation games, called eps-fractional core stability, which can be viewed as a requirement that a coalition selected uniformly at random from the set of all coalitions (conceptually choices according to some other distribution a... | Rebuttal 1:
Rebuttal: We thank the reviewer for the insightful comments. We reply here to the points raised in the 'Weaknesses' and explain how we will address them in our paper. Before providing our detailed answers, we want to highlight that while for simple FHGs our results hold only for the uniform distribution, fo... | Summary: This paper introduces and studies the concept of epsilon-fractional core stability in Hedonic Games (HGs), which is a natural relaxation of core stability where a small ratio (or a small probability mass) of coalitions is allowed to core-block. The paper shows that this notion is very flexible and resilient to... | Rebuttal 1:
Rebuttal: We thank the reviewer for the positive comments and constructive remarks. In the following, we will address the posed questions.
As we discussed in the paper, $\lambda$-bounded distributions are a very general class. Lemma 3.4 outlines what we can essentially say about the probability of a certai... | Summary: This paper studies the problem of finding a core-stable partition in the hedonic games. As core-stable partitions seldom exist and they are usually hard to compute, the authors propose a relaxed notion of epsilon-fractional core-stable partitions, in which at most epsilon-fraction of all possible coalitions ca... | Rebuttal 1:
Rebuttal: We thank the reviewer for the helpful comments. In the following, we will address the comment in 'Weaknesses' and the posed questions:
In the case of simple FHGs (i.e. Section 5), we assume that the distribution is uniform, thus Theorem 5.1 works under this assumption. This shows that the definit... | Rebuttal 1:
Rebuttal: We thank the reviewers for their valuable comments. We will take into account their feedback and revisit the paper accordingly. In what follows, we provide a general answer to the reviewers addressing some shared concerns raised in their comments.
The specific questions will be addressed in the de... | NeurIPS_2023_submissions_huggingface | 2,023 | Summary: This paper is a theory paper analyzing the stability characteristics of the Hedonic Game (HGs). HGs are a class of problems that partition agents considering their preferences. The core stability is the most important characteristic, which measures how well the agents' preferences are fulfilled. Specifically, ... | Rebuttal 1:
Rebuttal: We thank the reviewer for the positive comments and the useful feedback.
To answer effectively to the reviewer's questions, we would like to focus on a couple of points that appear to be unclear. We did not understand exactly if the reviewer here wanted to quote paper [18] instead of [27]. In any... | null | null | null | null | null | null |
Defending Pre-trained Language Models as Few-shot Learners against Backdoor Attacks | Accept (poster) | Summary: This paper introduces a novel approach to address the vulnerability of pre-trained language models (PLMs) utilized as few-shot learners in the face of backdoor attacks. The proposed defense mechanism, called MDP (masking-differential prompting), exploits the sensitivity of sentence masking to effectively diffe... | Rebuttal 1:
Rebuttal: We thank the reviewer for the valuable feedback on improving this paper! Please find below our response to the reviewer’s questions.
> To enhance the clarity of the proposed method, it is strongly recommended that the authors include a formal algorithm in their paper. Additionally, it would be be... | Summary: This work conducts a pilot study showing that PLMs as few-shot learners are highly vulnerable to backdoor attacks while existing defenses are inadequate due to the unique challenges of few-shot scenarios. To address such challenges, the authors advocate MDP, a novel lightweight, pluggable, and effective defens... | Rebuttal 1:
Rebuttal: We thank the reviewer for the valuable feedback on improving this paper! Please find below our response to the reviewer’s questions.
> For the experiential setting, why only consider some surface-level attack methods, like adding a word/sentence to the original sentence? Some more advanced and st... | Summary: The paper presents an investigation into the security vulnerabilities of pre-trained language models (PLMs) when used as few-shot learners, demonstrating that they are highly susceptible to backdoor attacks. The authors further note that existing defenses are inadequate due to the unique challenges presented b... | Rebuttal 1:
Rebuttal: We thank the reviewer for the valuable feedback on improving this paper! Please find below our response to the reviewer’s questions.
> Generalizability: While the paper demonstrates that the MDP defense mechanism works well for defending RoBERTa-large against backdoor attacks in a few-shot settin... | Summary: This paper proposes MDP(Masking-differential prompting), to defense PLMs as few-shot learners against backdoor attacks. The paper empirically validates that MDP outperforms baseline defenses.
Strengths: 1.The paper is well-written and easy to follow.
2.Investigating backdoor attack/defense under few-shot sc... | Rebuttal 1:
Rebuttal: We thank the reviewer for the valuable feedback on improving this paper! Please find below our response to the reviewer’s questions.
> Since the MDP is based on the observation: if the trigger is (partially) masked, the language modeling probability of a poisoned sample tends to vary greatly. A s... | null | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Exploiting Contextual Objects and Relations for 3D Visual Grounding | Accept (poster) | Summary: The paper proposes to supervise the detection of contextual objects and contextual relations in a given language sentence to help improve language grounding performance. Specifically, the authors devise three modules: a) Language guided detection module is responsible for detecting all mentioned objects. Since... | Rebuttal 1:
Rebuttal: We greatly appreciate your comprehensive feedback on our paper. Your questions and suggestions have been carefully addressed to improve the clarity and quality of the paper. Please find our responses below.
**About Text-guided detection module:**
**A1:**
We supervise the model with the ground... | Summary: This paper presents a transformer-based method for learning underlying spatial relationships in scenes for 3D visual grounding.
Specifically, three neural modules are proposed to tackle this challenge: 1) a language-guided object detector; 2) a relation matching network; 3) an attention-based target inference... | Rebuttal 1:
Rebuttal: We thank the reviewer for the detailed review and the positive feedback on our paper. We appreciate your insights and concerns, and below are our responses to the points you raised.
**Q1: The impact of designed spatial relation features (inputs to the relation matching module).**
**A1:**
The s... | Summary: This paper introduces the CORE-3DVG method, which consists of three consecutive modules: Text-guided object detection network for detecting all candidates, Relation matching network for matching the relationships among candidate objects, and Target inference network for selecting the referred object. The propo... | Rebuttal 1:
Rebuttal: We appreciate the constructive suggestions and comments. The following are our responses to the comments.
**Q1: The lack of essential analysis regarding pseudo labels.**
**(a) The accuracy of localizing objects other than the target.**
**A1 (a):** Thanks for your suggestions. For a language de... | Summary: This work presents a novel approach for learning 3D visual grounding: the task of localizing and identifying objects given natural language inputs. The language prompts in this problem often involve using other contextual objects and their relations to the target object. The proposed approach breaks down this ... | Rebuttal 1:
Rebuttal: We thank the reviewer for the positive feedback and insightful comments. Below, we address each of the concerns raised.
**Clarity**
**Response:**
Thanks for the kind reminder. We will revise the paper and move some details to the supplementary material to make it more concise and clear. The tex... | Rebuttal 1:
Rebuttal: We sincerely thank all reviewers for their constructive feedback and valuable insights. We have carefully addressed each concern raised and will incorporate the suggested changes to improve the clarity and quality of our paper.
In response to the concerns raised, we have made more comprehensive c... | NeurIPS_2023_submissions_huggingface | 2,023 | Summary: This paper proposes to explicitly model object relations mentioned in the language description for 3D visual grounding task. The proposed method designs a relation matching network and weakly-supervised contextual relation learning strategy to perform contextual learning. Finally, they design a target inferenc... | Rebuttal 1:
Rebuttal: We thank the reviewer for the valuable suggestions and comments. Below are our responses to the raised questions.
**Q1: Comparison with methods using ground truth object proposals.**
**A1:** In our approach, we focus on reasoning about the target objects from point clouds without relying on pre... | null | null | null | null | null | null |
Multitask Learning with No Regret: from Improved Confidence Bounds to Active Learning | Accept (poster) | Summary: This paper proposes refined confidence intervals for multitask kernel regression and applies this to the multitask online learning setting and the multitask active learning setting. The thrust of the idea is to further exploit the structure of the task kernel (Eq (2)) where the refined confidence bound is alwa... | Rebuttal 1:
Rebuttal: We thank the reviewer for raising these questions, which we address below.
> An implicit assumption is that each $f_t$ is a sample of the known $GP(k_X)$ that I feel the paper does not make it explicit enough.
In this work, we take a frequentist interpretation and assume the tasks$ f_i$ are fi... | Summary: This paper focuses on multi-task regression problems where they provide tighter confidence bounds that can be used to establish regret bounds in multi-task online learning and multi-task active learning. They also propose an algorithm that reduces the regret compared to previous bounds by adapting a task simil... | Rebuttal 1:
Rebuttal: We thank the reviewer for going through our work. We respond below to the raised weaknesses and questions.
**W1.**
We will use the extra page upon acceptance to introduce our notation in a separate paragraph.
**W2.**
We will use the extra page upon acceptance to move Algorithm 3 into the main ... | Summary: This work studies uncertainty quantification of estimated tasks in agnostic multitask regression. Through exploiting the structure of the task Gram matrix, This work proposes an improved bound on the multitask confidence interval that is tighter than a previous confidence interval [1] from by at most $\sqrt{N}... | Rebuttal 1:
Rebuttal: We thank the reviewer for the comments and suggestions. We address below the specific questions raised.
> It would be better to provide more intuition of deriving the improved confidence interval. For example, provide a proof sketch of Theorem 1.
We agree with the reviewer’s suggestion, we exp... | Summary: This paper focuses on multitask learning in the agnostic setting. It provides multitask confidence intervals and novel online learning algorithm that achieves such improved regret. The paper also introduces a novel multitask active learning setup where several tasks must be simultaneously optimized but only on... | Rebuttal 1:
Rebuttal: We thank the reviewer for appreciating our theoretical analysis and experimental evaluations. Below we address the raised points.
> Can you please also add other real-world applications where multi-task active learning can be extended?
We agree with the reviewer’s suggestion and we will add fu... | Rebuttal 1:
Rebuttal: We thank all the reviewers for going through our work and providing insightful suggestions and comments.
We have addressed each of the reviews individually, hoping that most of the reviewer’s concerns are resolved.
We are happy to provide additional insights and details based on the reviewers’ ... | NeurIPS_2023_submissions_huggingface | 2,023 | Summary: The primary objective of this paper is to analyze the confidence bound of the multi-task kernel regression model and explore its application in online learning and active learning scenarios. Initially, the authors introduce a novel confidence bound, revealing distinct regimes based on task similarity. Notably,... | Rebuttal 1:
Rebuttal: We thank the reviewer for the raised points, to which we would like to individually respond below.
\
\
**W1 (empirical results on 2 datasets only).**
We would like to point out that our work is primarily of theoretical nature and that experiments are essentially designed to support our theoretical... | null | null | null | null | null | null |
Neural Latent Geometry Search: Product Manifold Inference via Gromov-Hausdorff-Informed Bayesian Optimization | Accept (poster) | Summary: This paper addresses the problem of automatically identifying the optimal latent geometry for machine learning models. The neural latent geometry search (NLGS) searches for a latent geometry composed of a product of constant curvature model spaces with minimal query evaluations.The proposed method utilizes the... | Rebuttal 1:
Rebuttal: We thank the reviewer for the positive comments with regard to our work. We hope the additional explanations for some of the weaknesses identified will help to raise the score.
**[Weakness 1]** We thank the reviewer for this comment, but in our opinion, the proposed minimization problem would not... | Summary: The goal of this paper is to learn the geometry of underlying given data points after mapping them into latent space. Rather than learning a general (nonparametric) Riemannian manifold from the embedding data, the paper models this manifold as a product of certain prototypical manifolds with constant curvature... | Rebuttal 1:
Rebuttal: We thank the reviewer for the positive comments on our work. We hope that the following will address some of the issues raised.
**[Weakness 1]** We thank the reviewer for raising this concern. The reason why the manifold is fitted in the latent space is a consequence of the manifold hypothesis, ... | Summary:
The paper introduces an approach to estimate the best latent geometry structure for a machine learning model, trained on a data distribution and for a specific task. The approach assumes a product manifold structure of the latent space of subspace of constant curvature, being able to model product spaces of e... | Rebuttal 1:
Rebuttal: We thank the reviewer for the positive review. We hope that any outstanding concerns will be addressed with the answers provided below.
**[W1/Q1]** We would like to thank the reviewer for raising this concern. As mentioned in (L1) for Reviewer v4xj, there are no scalability issues associated with... | Summary: A variety of recent advances have provided evidence that learning in the
manifold setting (i.e., assuming that high dimensional data lie along
low-dimensional latent manifolds, aka. the manifold hypothesis) is advantageous. However, determining the optimal geometry of the latent space often requires expertise ... | Rebuttal 1:
Rebuttal: We thank the reviewer for appreciating the importance of the topic that we target in our work. We hope that the following responses to the weaknesses raised by the reviewer will merit a raise in the score.
**[W1]** We thank the reviewer for raising this issue. In the same spirit as our previous r... | Rebuttal 1:
Rebuttal: We thank all reviewers for their feedback. We are glad that they acknowledged the novelty of the problem being addressed and the proposed framework. In the overall response, we would like to address some of the concerns raised in multiple reviews, which we have added in different appendices.
**[... | NeurIPS_2023_submissions_huggingface | 2,023 | Summary: In the past few years, there's been an interesting trend in the ML literature where people have proposed non-Euclidean latent space geometries to improve network performance. However, to my knowledge geometries have always been picked on an ad hoc basis or tuned naively as a hyperparameter - the most common be... | Rebuttal 1:
Rebuttal: We thank the reviewer for the very positive comments regarding our work. We also believe that the concept of NLGS could have far-reaching ramifications in the context of model tuning. We hope that the following explanations and changes will be sufficiently convincing to raise the preliminary score... | null | null | null | null | null | null |
Multi-Player Zero-Sum Markov Games with Networked Separable Interactions | Accept (poster) | Summary: The paper considers a subclass of Markov Games (MG) that have local interactions and a zero-sum structure for their Q value functions. This subclass generalizes zero-sum polymatrix games. It's shown that the marginal distribution of a Markov CCE for these games is a Markov NE. It's then shown that computing a... | Rebuttal 1:
Rebuttal: Thank you for finding our analysis rigorous and non-trivial, and the technical parts well-discussed. We acknowledge your main comments regarding *motivations*, and sincerely hope our responses below could help you appreciate the significance of the class of games and equilibrium we consider.
**M... | Summary: The paper studies the problem of learning equilibria in multi-player zero-sum Markov games with networked local interactions (abbreviated as MZNMGs for short). In such games, the set of Markov CCEs collapses into that of Markov NEs. The paper shows that finding approximate Markov stationary CCEs is PPAD-hard i... | Rebuttal 1:
Rebuttal: Thank you for finding our paper sound, well-written and significant. We respond to your reviews point-by-point as follows.
**Originality.** Thank you for your comment. We agree that some of the ideas and techniques are based on existing works. However, non-trivial ideas and techniques are propose... | Summary: The paper focuses on the problem of computing both in a centralized and in a decentralized way equilibria for a special class of stochastic games. These are games that may be found in multiple states. At each state players have a set of available actions, that once played will define both the reward that each ... | Rebuttal 1:
Rebuttal: Thank you for finding our results well-motivated, contributions clearly stated, and the paper easy to follow. We respond to your reviews point-by-point as follows.
**The results are essentially negative and positive cases have limited applications.** Regarding Prop. 3, first, the result applies t... | Summary: The authors define a notion of Markov Games with local interactions with the extra assumption that are zero-sum (inspired from the polymatrix/graphical games that appeared in papers in Algorithmic Game Theory). They manage to show that Markovian Coarse Correlated Equilibria correspond to Markovian Nash Equilib... | Rebuttal 1:
Rebuttal: Thank you for finding our equilibrium collapsing result interesting, and the algorithmic idea natural and elegant. We respond to your reviews point-by-point as follows.
**Hardness result seems not surprising and straightforward from Daskalakis et al.** We agree that the result *relies on* Daskala... | null | NeurIPS_2023_submissions_huggingface | 2,023 | Summary: The paper studies multi-player zero-sum Markov games with networked interactions, generalizing the well-studied class of polymatrix normal-form games. They first show that Markov coarse correlated equilibria (CCE) collapse to Markov Nash equilibria (NE), allowing to focus on computing Markov CCE for which they... | Rebuttal 1:
Rebuttal: **Comparison with prior work**. Thank you very much for bringing up the work “Fictitious play in networks” to our attention. We were not aware of the paper, and will add the reference in our updated version. We will also revise the statement regarding the first FP in zero-sum polymatrix games in t... | null | null | null | null | null | null |
Bi-Level Offline Policy Optimization with Limited Exploration | Accept (poster) | Summary: This manuscript proposes a bi-level structure for solving offline RL problem. The authors design the lower level constructing a set of value estimates, while the upper level maximizing the conservative value estimate. Specifically, the author provides a penalized adversarial estimation algorithm. The superiori... | Rebuttal 1:
Rebuttal:
We thank the reviewer for the suggestive feedback, which greatly improves our paper! We address the reviewer's comments point by point below.
> **Weakness 1: Algorithm 1**
We feel sorry for the misleading. Due to the page limit, we actually provide a comprehensive discussion of Algorithm 1 in A... | Summary: Offline Reinforcement Learning suffers from distribution shift, arising as a consequence of limited exploration and function approximation. The paper aims to address these two challenges by proposing a bi-level hierarchical framework for policy optimization. The upper-level optimizes a policy improvement objec... | Rebuttal 1:
Rebuttal: We thank the reviewer for the suggestive feedback, which greatly improves our paper! We address the reviewer's comments point by point below.
> **Weakness and Question 1: Empirical evaluation**
Thanks for your great suggestions!
**For regret validation**:
We first would like to clarify that... | Summary: This paper presents a bi-level structured policy optimization algorithm for offline reinforcement learning (RL). The algorithm aims to address the distributional shift and function approximation challenges in offline RL by modeling a hierarchical interaction between the policy and the value function. The lower... | Rebuttal 1:
Rebuttal: We thank the reviewer for the suggestive feedback, which greatly improves our paper! We address the reviewer's comments point by point below.
**Weakness 1: Connection and central claim**
We feel sorry for the confusion and thanks for helping us improve the readability of the paper. The offline p... | Summary: This paper proposed a pessimistic offline RL method. Different from the previous method in the function approximation setting, the method proposed in this paper does not require coverage and completeness assumption. Moreover, in terms of algorithm design, this paper derives a lower confidence bound for RL obje... | Rebuttal 1:
Rebuttal: We thank the reviewer for the suggestive feedback, which greatly improves our paper! We address the reviewer's comments point by point below.
> **Weakness 1: Experiments**
Thanks for the suggestion! We evaluate our algorithm on walker2d, hopper, halfcheetah, and maze2D D4RL benchmarks. The resul... | Rebuttal 1:
Rebuttal: We are grateful to all reviewers for taking the time to review our paper and sharing insightful and positive feedback. We are encouraged that they found the submitted work to be novel (7QtD,UsZq,C4A2) and interesting (75WN, q6tp), addressing a challenging and meaningful research problem (7QtD,UsZq... | NeurIPS_2023_submissions_huggingface | 2,023 | Summary: This paper addresses the problem of offline reinforcement learning (RL). Offline RL involves computing a policy based on a static/pre-collected dataset, with no further interaction with the environment permitted. A significant hurdle in offline RL is the distributional shift resulting from insufficient explora... | Rebuttal 1:
Rebuttal:
We thank the reviewer for the suggestive feedback, which greatly improves our paper! We address the reviewer's comments point by point below.
> **Weakness 1: Complexity of the algorithm**
Thanks for raising this point. We would like to make a discussion on the three aspects.
**Complexity**: w... | null | null | null | null | null | null |
Combining Behaviors with the Successor Features Keyboard | Accept (poster) | Summary: The author(s) present a novel and efficient approach to knowledge transfer in long-horizon tasks, aiming to minimize interactions with the environment (less frames required in training). Their proposed method leverages Categorical Successor Feature Approximator (CSFA) within an enhanced deep OK network, result... | Rebuttal 1:
Rebuttal: Thank you for this review. Below we respond to specific questions.
**The authors did not clearly position their contribution in relation to other approaches commonly used in long-horizon tasks, such as hierarchical approaches, hindsight approaches and meta-learning approach [1-3]. The author's co... | Summary: This paper focuses on the Option Keyboard (OK) task which was recently proposed as a method for transferring behavioral knowledge across tasks. They propose a new method of Deep Option Keyboard (Deep OK), which enables transfer with discovered state features and task encodings. To enable discovery, they furthe... | Rebuttal 1:
Rebuttal: Thank you for your review. We are glad that you found the writing clear and approach well-reasoned. Below we respond to specific questions.
**Maybe a more general method than sampling queries from the task encoding space that is more fairly concentrated could be designed and at least provide some... | Summary: This work focuses on attempting something which no previous work has done before; learning task encodings, cumulants and successor features (SFs). The author(s) proposed a solution based on learning these three functions jointly. They considered a model that learns the SFs as probability mass function using t... | Rebuttal 1:
Rebuttal: Thank you for this positive review. We are happy you appreciated this submission. Below we respond in detail to your questions.
**Changing name to not give false impression of using options.**
Thank you for pointing this out. We have changed the name to "Successor Feature keyboard".
**Adding a ... | Summary: First, I must emphasize that I was new to the RL community before this NeurIPS. This is the first RL paper I have reviewed. I have been struggling to understand this paper. I tried my best. I hope the AC, reviewers, and authors may think me helpful.
This paper studies how to transfer a pre-learned RL system t... | Rebuttal 1:
Rebuttal: Thank you for taking the time to read this paper and for asking these clarifying questions. We have used your comments to improve the text. Due to space constraints, we respond to the most pressing comments and questions below.
# Environment is a bit narrow. Why only 3D case? Using only one datas... | Rebuttal 1:
Rebuttal: We thank the reviewers for their very constructive feedback which we have used to improve the clarity and accessibility of our paper. We are encouraged that reviewers found that we tackle a novel challenge [R2, R3] with a well-reasoned approach [R3], have significant [R4] and convincing results [R... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
PreDiff: Precipitation Nowcasting with Latent Diffusion Models | Accept (poster) | Summary: This paper presents the PreDiff for precipitation nowcasting. PreDiff adopts the latent diffusion for prediction. Besides, a special knowledge control design is used in a classifier guidance way to incorporate auxiliary prior knowledge, which can guide the model to satisfy conservation constrain. In both synth... | Rebuttal 1:
Rebuttal: Thanks for providing detailed feedback. We hope our response below can address your main concerns.
**W1: more explanation on novelty and prior knowledge for precipitation nowcasting**
Our literature review covers related works from three perspectives: deep learning for Earth system forecasting (... | Summary: This paper applies latent diffusion models to the problem of precipitation nowcasting, developing a modular and flexible method for enforcing external constraints during the generating process. Constraints are general in that they can be specified in terms of any function F from the data to the reals, F:(x,y)-... | Rebuttal 1:
Rebuttal: Thank you for providing detailed feedback. We hope that our response below can address your main concerns.
**Q1: multiple samples from the probabilistic models and the performance comparison with the deterministic models**
Thank you for the thoughtful feedback. To clarify, instead of sampling on... | Summary: This paper introduces a new method for doing precipitation nowcasting. The method is based on building a conditional diffusion model in a global latent space, which is found separately in an autoencoding fashion. Additionally, a “knowledge control” mechanism is introduced to help ensure that the generated samp... | Rebuttal 1:
Rebuttal: Thanks for providing detailed feedback. We hope our response below can address your main concerns.
**W1: the model itself is not the most novel**
Our literature review covers related works from three perspectives: deep learning for Earth system forecasting (line 30-42), diffusion models (line 43... | Summary: This work proposes a latent diffusion model called PreDiff for spatiotemporal forecasting. The main contributions of this paper are two-fold: 1) they use a conditional latent diffusion model to handle the uncertainty of spatiotemporal forecasting. 2) They introduce an explicit knowledge control mechanism to al... | Rebuttal 1:
Rebuttal: Thank you for providing detailed feedback. We hope that our response below can address your main concerns.
**W1.1: ablation studies**
we have conducted **several experiments serving the role of ablatiion studies**. We first ablate the two aspects that differentiate PreDiff from LDM [1]:
+ We rep... | Rebuttal 1:
Rebuttal: # General Response
Dear Reviewers,
We appreciate reviewers’ effort in evaluating this work. We are glad that they found our paper "well-written" (Ntzp,sDg9,yf9m,zn7Z), our PerDiff with novel knowledge control mechanism "interesting"/"practical"/"effective" (Ntzp/yf9m/zn7Z), our empirical studies... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
CoDrug: Conformal Drug Property Prediction with Density Estimation under Covariate Shift | Accept (poster) | Summary: The paper concerns the problem of conformal drug property prediction, where set valued predictions that should cover with high probability the true class are made. The authors argue that due to sampling design covariate shift tends to occur where test data has different input distribution than the training dat... | Rebuttal 1:
Rebuttal: We thank the author for their overall positive assessment of the paper. In our response, we would like to address the posed questions and elaborate on the stated weaknesses.
## Clarification regarding the exchangeability argument
A sequence of random variables is exchangeable if the joint probabi... | Summary: This paper studies conformal drug property prediction under covariate shift, i.e. the calibration data and the test data are not exchangable.
The proposed method, CoDrug, uses an energy-based model trained on both labeled and unlabeled data to model the conditional $p(y|x)$, and KDE to estimate the likelihood ... | Rebuttal 1:
Rebuttal: We thank the reviewer for their overall positive assessment of the paper. We would like to address the questions/concerns raised by the reviewer.
## Importance of ZINC unlabelled data
We report the importance of using ZINC-based unlabelled data in Table 3 and Section 5.2 of our paper. The "CoDru... | Summary: The paper is concerned with the problem of conformal drug property prediction by formulating and solving the problem as a conformal prediction problem under covariate shift. To this end, the paper proposed training an energy-based model using both labeled and unlabeled data. To derive conformal prediction und... | Rebuttal 1:
Rebuttal: We thank the reviewer for their feedback and would like to address the concerns regarding the novelty of our paper. Our work presents several contributions, which we believe are novel for various reasons:
- We focus on conformal prediction (CP), a powerful and increasingly popular tool for obtain... | Summary: The authors apply conformal prediction under covariate shift to molecular property prediction. They propose using KDE to estimate the densities for the calibration and test sets, and demonstrate that doing so is better than plain vanilla CP.
Strengths: ### Originality
This paper puts together existing ideas ... | Rebuttal 1:
Rebuttal: We thank the reviewer for the overall positive assessment of the paper. In our response, we would like to address the question about the applicability of KDE and elaborate on the limitations of the proposed framework.
- While we demonstrated that KDE can provide asymptotic coverage guarantees, t... | null | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Consistent Diffusion Models: Mitigating Sampling Drift by Learning to be Consistent | Accept (poster) | Summary: This paper introduced Martingale Property (MP) in diffusion model and the training strategy to apply this property to modify the score network outputs. In short, the martingale property states the condition that the SDE trajectory conditioned in a sample at time t is equal to the denoising function. After show... | Rebuttal 1:
Rebuttal: We thank a lot the Reviewer for their feedback! We are glad that the Reviewer appreciated certain aspects of our work. In what follows, we attempt to address some remaining concerns.
> In the date of submission, there is a concurrent work [1] that also modifies the score network output by distil... | Summary: Using approximate score functions and discretization can lead to compounding distribution shift as samples drift toward less likely regions of the training distribution. The authors propose to address this problem by enforcing that the learned denoiser satisfies an invariant they call the Martingale Property: ... | Rebuttal 1:
Rebuttal: We thank the Reviewer for their feedback. We are happy that the Reviewer appreciated the theoretical and experimental aspects of our work.
> The main theorem (Thm. 3.2) does not seem to be very actionable. The results assume that MP holds exactly, which is an unreasonable assumption when conside... | Summary: Diffusion models are trained to denoise images that have been generated using a particular schedule of additive Gaussian noise; however, when sampling from a trained diffusion model, one applies a chain of denoising steps, each one operating on the output of the previous step. This may lead to sampling drift, ... | Rebuttal 1:
Rebuttal: We thank the reviewer for the extremely detailed and constructive feedback. We will incorporate all the thoughtful comments regarding the presentation of our work.
We agree that there are some similarities with the concurrent CM work and we will gladly discuss it. The motivation is different: CM ... | Summary: This paper proposed a new method to address the issue of imperfect score-matching in diffusion models by enforcing a Martingale Property (MP) during training. This MP property ensures that the model's predictions on its self-generated data are consistent with the generated outputs. The key contributions includ... | Rebuttal 1:
Rebuttal: We thank a lot the Reviewer for appreciating the novelty, the theoretical and the empirical results and the presentation of our work!
We will compress the background section and combine Figures 1a, 1b into a single Figure as recommended by the Reviewer.
> The objective function presented in Sect... | Rebuttal 1:
Rebuttal: We thank the Reviewers for their constructive feedback. We reply separately to each of the Reviewers. Following the recommendation of Reviewer G7yx, we also include a formal statement of our training algorithm, given in the one-page PDF accompanying this rebuttal.
We thank again the Reviewers and... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Human-Aligned Calibration for AI-Assisted Decision Making | Accept (poster) | Summary: Communicating model uncertainty to human decision makers is critical for trustworthy deployment of assistive AI systems. Prior work, e.g., Vodrahalli et al 2022, have found that people may respond better to miscalibrated AI confidence scores. What underlies this odd behavior? In this submission, the authors ex... | Rebuttal 1:
Rebuttal: We thank the reviewer for their careful and insightful comments, which will help improve our paper. Please, find a point-by-point response below.
**[Discretization of human confidence]** As long as the human confidence scores are represented using a discrete ordered set, our theoretical results a... | Summary: This paper proposes a causal model of AI-assisted decision-making that considers both the classifier’s and decision maker’s confidence scores. The authors demonstrate theoretically why it may be challenging for the human to identify the optimal decision-making policy even if the classifier is perfectly calibra... | Rebuttal 1:
Rebuttal: We thank the reviewer for their careful and insightful comments, which will help improve our paper. Please, find a point-by-point response below.
**[Empirical results]** We acknowledge that our experiments do not empirically demonstrate that, **for a fixed task**, there is an improvement in $E_{\... | Summary: This paper theoretically analyzes human decision making in the presence of AI-provided predictions and confidence estimates. They do this by defining a causal model to characterize AI-assisted decision making along with a sensible assumptions about a rational decision-maker's policy: namely that their probabil... | Rebuttal 1:
Rebuttal: We thank the reviewer for their careful and insightful comments, which will help improve our paper. Please, find a point-by-point response below.
**[Empirical validation that human-alignment leads to higher human performance]** We acknowledge that our experiments do not empirically demonstrate th... | Summary: The paper studies a binary decision-making process augmented by AI assistance. The decision maker optimizes a utility function satisfying certain natural conditions. The decision rule depends on the decision maker's (DM) confidence and a classifier's confidence. It is shown that there is a lack of alignment be... | Rebuttal 1:
Rebuttal: We thank the reviewer for their careful and insightful comments, which will help improve our paper. Please, find a point-by-point response below.
**[Very simple setup]** Our theoretical results in Sections 3-4 and the multicalibration algorithms in Section 5 do not make any assumption about the c... | Rebuttal 1:
Rebuttal: We thank the reviewers for their careful and insightful comments, which will help improve our paper.
In addition to the point-by-point responses to each individual reviewer below, please, find attached a pdf containing updated figures for the “Art” and “Sarcasm” tasks as well as an updated Table... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
May the Force be with You: Unified Force-Centric Pre-Training for 3D Molecular Conformations | Accept (poster) | Summary: 1. proposing a pretraining framework for conformation space including equilibrium and off-equilibrium.
2. this method also shows highly accurate in MD or relax, including molecules and polymers.
3. providing the community with a diverse set of DFT dataset of polymers.
Strengths: 1. This paper is well-written ... | Rebuttal 1:
Rebuttal: Thank you for your valuable insight! In our separate work, we have indeed found that including partial charge in pre-training where available helps property prediction tasks. We are actively investigating this and more strategies to make the pre-training more efficient.
In the rebuttal period, w... | Summary: This paper incorporates off-equilibrium molecular conformations into consideration, in which the energy is not minimized. This paper defines pre-training objectives for force field learning on sampled off-equilibrium molecular conformations. Experimental results validate the effectiveness of pre-training on ad... | Rebuttal 1:
Comment: We would like to kindly remind the reviewer to read our response that we have made available for all reviewers in response to questions on the presentation of the paper. Thank you!
---
Rebuttal 2:
Comment: We would like to remind the reviewer to please take a look at our previous reply and commen... | Summary: The paper propose a set of additional objectives for training machine learning interatomic potentials, in particular for equilibrum conformations as well as a denoising term. They show that this improves stability, when pretrained on a large dataset. They also introduce a new dataset of polymers.
Strengths: T... | Rebuttal 1:
Comment: Dear Reviewer,
We appreciate the time you've taken to review our work. We kindly direct your attention to the points we made in our reply.
- Our contribution is to the pre-training framework, not the model architecture itself. ET-OREO has significantly improved the performance of ET, and as we ha... | null | null | Rebuttal 1:
Rebuttal:
First, we thank the reviewers for their time and effort to provide valuable feedback. Below, we take the opportunity to discuss your points of concern.
**Our contribution is focused on the pre-training paradigm, instead of the model architecture.**
NequIP has a great performance with limited tr... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Cal-QL: Calibrated Offline RL Pre-Training for Efficient Online Fine-Tuning | Accept (poster) | Summary: The paper introduces a method called Calibrated Q-learning (Cal-QL) for efficient online fine-tuning. This method learns an effective initialization from offline data, which enables fast online fine-tuning. It does so by underestimating the value of the learned policy from offline data while ensuring that the ... | Rebuttal 1:
Rebuttal: Thank you for your detailed feedback and for a positive assessment of our work. To address your concerns, we have now run experiments evaluating Cal-QL on Gym locomotion tasks and found that Cal-QL outperforms prior methods on these tasks, tied with CQL. We also added a comparison to SPOT (Wu et a... | Summary: The paper proposed a new approach to Offine-RL Online Fine-Tuning called calibrated Q-learning (Cal-QL). This approach aims to learn an effective initialization from offline data that enables fast online fine-tuning capabilities. Cal-QL accomplishes this by learning a conservative value function initialization... | Rebuttal 1:
Rebuttal: Thank you for your review and for a positive assessment of our work. To address your concern regarding comparisons, we have added an experiment to compare Cal-QL to SPOT (Wu et al 2022), as you suggested. We found that Cal-QL outperforms this prior method (more details below). We also clarify that... | Summary: The paper proposes a new offline to online algorithm, Cal-QL, that aims to mitigate the unlearning during the beginning of the online fine-tune stage. The paper first conducts empirical experiments to observe that the reason for the unlearning is due to overly pessimistic function values, and thus proposes the... | Rebuttal 1:
Rebuttal: Thank you for your feedback and a positive assessment of our work. We address your concerns below and will update the paper to clarify each of the questions.
**Please let us know if your questions are resolved, and if so, we would be grateful if you are willing to upgrade your score. We are happ... | Summary: Initially, this paper conducts an empirical investigation to understand why the offline policy derived from prior methods is inadequate as an initialization for online fine-tuning. Building upon these empirical findings and the derived hypothesis, the authors design Cal-QL, a method differing from CQL by only ... | Rebuttal 1:
Rebuttal: Thank you for your review and positive feedback on the paper. To address your concerns, we have now added comparisons on D4RL locomotion tasks, where we find that Cal-QL outperforms prior methods IQL, AWAC, Hybrid RL, performing similarly to the best method, CQL. We have also clarified other conce... | Rebuttal 1:
Rebuttal: Dear Reviewers and the AC,
We thank all the reviewers for their detailed and constructive feedback. We are glad that Reviewer vsVq found our explanations to be meticulous, Reviewer zucz found our algorithm to be simple, and all reviewers found our paper to be well-written and clear.
To address t... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Conditional Matrix Flows for Gaussian Graphical Models | Accept (poster) | Summary: This work proposed a framework for performing inference on Gaussian Graphical Models by approximating the posterior with a normalizing flow over PSD matrices. In this way, the authors can investigate $l_p$-norm regularized GGMs for any value of $p$ in an efficient way.
Strengths: The idea of using normalizing... | Rebuttal 1:
Rebuttal: We thank the reviewer for acknowledging the novelty of the proposed framework and for suggesting relevant experiments, which we have included in the global response. We hope to provide satisfying answers to the concerns raised.
- "The main weakness that I identified is the lack of comparison betw... | Summary: This paper proposes a method that can be used to infer conditional independencies in a Gaussian model. These conditional independencies are related to zeros in the precision matrix. Typically, sparse enforcing norms are used to estimate the precision matrix while enforcing zeros in the elements outside of the ... | Rebuttal 1:
Rebuttal: We thank the reviewer for the useful feedback on the experiment section and for acknowledging the clarity in the presentation of the proposed approach. We hope to provide satisfying answers to the concerns raised.
- "The proposed method is a combination of already known techniques."
We believe t... | Summary: This paper targets the structure learning problem in Gaussian Graphical Models via (Normalising) Flow-based Variational approximation of the elements of weight metrics that correspond to the Gaussian Bayesian network.
They use sub-l1 pseudo norms to penalize dense precision metrics (which correspond to graphs... | Rebuttal 1:
Rebuttal: We thank the reviewer for the very encouraging comments and for acknowledging the potentially large impact of the proposed model. We hope to provide satisfying answers to the concerns raised.
- "Most notably, the algorithm is compared to no alternative work (neither in the main paper nor in the s... | Summary: This paper concerns the estimation of precision matrix under $l_p$ norm sparcity penal. The solution is a variational inference through normalizing flow, which is a function of shrinkage parameter $\lambda$ and non-negative norm parameter $p$. It allows for straightforward computation of solution paths for the... | Rebuttal 1:
Rebuttal: We thank the reviewer for acknowledging the novelty and soundness of our work. We hope to provide satisfying answers to the concerns raised.
- "Using simulated annealing algorithm to recover a path of solutions for varying $\lambda$ and $p$ is useful, in particular for the case of $p$, as in case... | Rebuttal 1:
Rebuttal: ## Global response
We would like to thank all reviewers for the very encouraging comments and for acknowledging the novelty and the potentially large impact of the proposed approach. As all reviewers expressed one main concern, we try to address it in the global response. The main point of critici... | 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.