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 |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
DomainGallery: Few-shot Domain-driven Image Generation by Attribute-centric Finetuning | Accept (poster) | Summary: This paper tackles the problem of few-shot image generation, in a scenario where it is domain-driven and defined by several attributes shared among different target data.
Strengths: 1. The research topic is of general interest in the generative models, and this is a good bridge between subject-driven generati... | Rebuttal 1:
Rebuttal: ### Weakness 1
We are not quite sure whether the phrase *more latest and advanced T2I models* mentioned by the reviewer refers to the newer subject-driven methods after DreamBooth, or newer T2I models such as Stable Diffusion after version 1.4 we used in our work.
- **If it refers to newer subjec... | Summary: This paper introduces a few-shot domain-driven image generation method that finetunes pretrained Stable Diffusion in an attribute-centric manner. It features prior attribute erasure, attribute disentanglement, attribute regularization and attribute enhancement.
Strengths: 1. The results seem good without over... | Rebuttal 1:
Rebuttal: ### Weakness 1
With regard to the part of attribute regularization, we have utlized a variant of the similarity consistency loss in CDC [1]. However, our core contribution of this part is not the similarity consistency loss itself, but is the strategy to construct **paired** source/target codes i... | Summary: Text-to-image generation models pre-trained on large-scale datasets made progress but the models are still limited when we expect to generate images that fall into specific domain or style that are hard to describe or unseen to models. This paper proposed the DomainGallery, a few-shot domain-driven image gener... | Rebuttal 1:
Rebuttal: ### Weakness 1
Thanks for the advice. Here we recap the four technical challenges we have mainly focused on in our work:
- **Challenge 1:** Prior attributes of the identifier [V] may show up in the generated samples even if we have bound new domain attributes to it.
- **Challenge 2:** The identi... | Summary: This paper addresses the few-shot domain transfer problem in text-to-image diffusion generation, where the goal is to keep the style and attributes of the source examples while being able to generate images of potentially different subject matters. It binds the attributes to rarely-used tokens, but observes th... | Rebuttal 1:
Rebuttal: ### Weakness 1 & 2
We would like to thank the reviewer for raising concerns about the *unintended attributes*, which is indeed one of the core issues in few-shot image generation. Due to the lack of sufficient data samples, there is usually a gap, either large or little, between the distribution ... | Rebuttal 1:
Rebuttal: ### Recap
In this work, we propose DomainGallery, a method for few-shot domain-driven image generation, which analyzes the key issues that previous works failed to settle, and accordingly designs a series of attribute-centric finetuning techniques. With these novel and effective techniques, our D... | NeurIPS_2024_submissions_huggingface | 2,024 | null | null | null | null | null | null | null | null |
SDformer: Similarity-driven Discrete Transformer For Time Series Generation | Accept (poster) | Summary: This paper investigates the use of a vector-quantization transformer for time series generation. The authors propose a similarity-driven vector quantization method for time series modeling. After training the encoder and decoder structures for generating embeddings, the authors propose two ways to generate tim... | Rebuttal 1:
Rebuttal: Thanks for your constructive reviews and suggestions. In the following, we will answer your questions one by one.
**Q1.** I wonder about the motivation behind this work.
**A1.** Our motivations include the following three points:
* **Inference efficiency.** The inference time of previous SOTA mo... | Summary: In this paper, the authors propose a time series generation model named SDformer. The model is based on the discrete token modeling techniques and demonstrate empirically its feasibility for the time-series generation task. The method presented in the paper surpasses the current state-of-the-art models on mult... | Rebuttal 1:
Rebuttal: Thanks for your constructive reviews and suggestions. In the following, we will answer your questions one by one.
**Q1.** How did the authors choose the size of the codebook K.
**A1.** K is a hyperparameter that requires experimentation to determine the optimal value. Specifically, we choose fr... | Summary: This paper introduces SDformer, a method for time series generation, which addresses challenges in inference time and quality improvement. It utilizes a similarity-driven vector quantization technique to learn high-quality discrete token representations of time series. It employs a discrete Transformer for dat... | Rebuttal 1:
Rebuttal: Thanks for your constructive reviews and suggestions. In the following, we will answer your questions one by one.
**Q1.** The description of the proposed method is not very detailed.
**A1.** Thank you for your feedback. SDformer is a two-stage method.
* **First Stage:** In the first stage, we d... | Summary: This paper proposed a MAE type attention model for time series generation. One major advantage is its efficiency, which is much faster than the diffusion based method. Various experiments are conducted and the proposed model outperform all baselines.
Strengths: 1. Efficiency: The paper introduces SDformer, a ... | Rebuttal 1:
Rebuttal: Thanks for your constructive reviews and suggestions. In the following, we will answer your question.
**Q1.** One concern of the current submission is the lack of sample code provided to reviewers for verifying the reported experimental results.
**A1.** Thank you for your suggestion. We have pro... | null | NeurIPS_2024_submissions_huggingface | 2,024 | null | null | null | null | null | null | null | null |
MemVLT: Vision-Language Tracking with Adaptive Memory-based Prompts | Accept (poster) | Summary: This paper introduces a memory-based vision-language tracker, called MemVLT. Specifically, the memory storage and memory interaction modules are designed to storage and flexible interaction between short-term and long-term memories, achieving vision-language tracking. A large number of experiments demonstrate ... | Rebuttal 1:
Rebuttal: **Dear Reviewer jufM,**
Thanks for thoroughly reviewing our work. We appreciate your acknowledgment of our model's strong performance and the quality of our writing. Below, we address your concerns and provide detailed responses:
---
### **Weakness 1: novelty of our work**
(1) **Novelty in lon... | Summary: This paper presents a new visual-language tracking method, called MemVLT, which incorporates memory modeling to adjust static prompts.The design of MemVLT is inspired by the theory of Complementary Learning Systems (CLS), which adapts to changes in the target by mimicking the storage and interaction mechanisms... | Rebuttal 1:
Rebuttal: **Dear Reviewer unPL,**
Thanks for your careful review of our work. Your acknowledgment of our memory mechanism modeling and its promising performance is invaluable to us. We notice that you are particularly concerned about the model's ablation analysis. Therefore, we have conducted the correspon... | Summary: This paper introduces MemVLT (Memory-based Vision-Language Tracker), an approach to vision-language tracking (VLT) that incorporates memory mechanisms inspired by human cognition. The key innovation is the use of adaptive memory-based prompts to guide tracking, as opposed to relying solely on initial static pr... | Rebuttal 1:
Rebuttal: **Dear Reviewer GRLs,**
We sincerely appreciate your time and effort in reviewing our work. We are grateful for your recognition of our cognitive science-inspired modeling approach, strong performance, and broader impact discussion. We notice that you are particularly concerned about the differen... | Summary: This paper proposes a Memory-based Visual-Language Tracker based on the complementary learning system theory. Memory mechanism modeling facilitates the generation of adaptive prompts to effectively guide the tracking process. Extensive experiments demonstrate that MemVLT achieves new state-of-the-art performan... | Rebuttal 1:
Rebuttal: **Dear Reviewer BESZ,**
Thanks for taking the time to review our work. We appreciate your recognition of our interesting insights, good writing, and strong performance. We hope the following responses will address your concerns.
---
### **Weakness 1 & 2: comparison with recent SOTA trackers**
... | Rebuttal 1:
Rebuttal: **Dear Reviewers,**
We would like to thank all the reviewers for their feedback and help in improving our work. We are glad that they acknowledge the **interesting insights** from introducing CLS theory into the VLT task (Reviewers BESZ, GRLs, and unPL), **adequate ablation analysis** (Reviewer j... | NeurIPS_2024_submissions_huggingface | 2,024 | null | null | null | null | null | null | null | null |
DeTrack: In-model Latent Denoising Learning for Visual Object Tracking | Accept (poster) | Summary: The paper introduces bounding box denoising to visual object tracking. To make it more suitable for tracking, the paper devises an in-model denoising pipeline that reformulate the repetitive sampling process as multiple ViT blocks in one forward pass. It also proposes a compound memory strategy to assist long-... | Rebuttal 1:
Rebuttal: We are grateful for the reviewer's acknowledgment of our work. We have taken the time to thoroughly address your questions and have included detailed responses to clarify any uncertainties.
>***Q1**.Have the authors ever applied the pipeline of DiffusionDet directly to tracking? If practical, how... | Summary: The paper addresses the task of visual object tracking. The main idea is to pose tracking as a latent denoising problem where the bounding box from the previous frame is denoised to obtain the current frame box. The authors use a ViT to obtain embeddings using the search regions as well as a set of template im... | Rebuttal 1:
Rebuttal: We thank the reviewer for the acknowledgement of the main contributions of our work, the useful comments and the relevant feedback provided on the technical side. We have provided detailed responses to your questions.
>***Q1**.While the authors present the method as a latent denoising approach m... | Summary: In this paper, the authors reformulate the visual tracking problem using denoising learning process. Given noisy bounding box coordinates, the conditional denoising tracker model's task is to remove the bounding box noise for accurate target state estimation. For the denoising formulation, they introduce an in... | Rebuttal 1:
Rebuttal: We extend our heartfelt gratitude for your perceptive insights and your recognition of the unique and meaningful contributions made by our research. Your support is highly valued, and we would be honored if you could serve as an advocate for our work.
>***Q1**.The claim that the model only requir... | null | null | Rebuttal 1:
Rebuttal: We sincerely appreciate the thorough review provided by all the reviewers. The valuable feedback from the reviewers has significantly contributed to enhancing the quality of our manuscript. We extend our gratitude to Reviewer **JcZZ**, Reviewer **wkvN** and Reviewer **b9Bg** for acknowledging the ... | NeurIPS_2024_submissions_huggingface | 2,024 | null | null | null | null | null | null | null | null |
Learning to Learn with Contrastive Meta-Objective | Reject | Summary: This study presents a contrastive regularizer to improve meta-learning. Specifically, the authors propose to incorporate a contrastive meta-objective that improves the alignment and the discrimination abilities of meta-learners, leading to better task adaptation and generalization. The authors demonstrate empi... | Rebuttal 1:
Rebuttal: Thank you for dedicating your time to give valuable comments on our paper. We especially find [1,2] are reasonable and interesting related works and we would like to discuss and credit to them in later paper.
# Contrastive strategy, objective and hyper-parameters.
**Reply.**
Please refer to the *... | Summary: The paper aims to enhance the meta-learning process by implementing more robust supervision within the model space. Specifically, the authors seek to augment learning capabilities through model alignment and discrimination, aiming to approximate human-like rapid learning abilities. They propose that models tra... | Rebuttal 1:
Rebuttal: Thank you for dedicating your time to give valuable comments on our paper.
# The paper lacks validation on MetaDataset, which is a common large-scale dataset for few-shot learning tasks.
**Reply.**
As mentioned in *Common Reply 2*, currently only very few effort has been made to tune hyperparamet... | Summary: This paper deals learning to learn (meta-learning) problem, from the perspective of exploring inner-task and intra-task relationship. Specifically, this paper proposed a Contrastive meta-objective by exploring intra- and inter-task distances and severed as an additional term for training objective (in addtion ... | Rebuttal 1:
Rebuttal: Thank you for dedicating your time to give valuable comments and advises on our paper.
# Distinguishing ConML with conventional contrastive learning.
Conventional contrastive learning methods have already provided good insights to the representation learning and deep learning community. It has b... | Summary: The paper proposes a contrastive meta-objective that can be applied to various meta-learning methods. Also, interpreting in-context learning as a meta-learning formulation, extended the proposed method to in-context learning. Specifically, the objective is to contrast task identity obtained after episode optim... | Rebuttal 1:
Rebuttal: Thank you for dedicating your time to give valuable comments on our paper, especially for valuable advises about improving the presentation.
# In the case of optimization-based methods, should the method need to proceed the episode optimization twice? Then why does it only take 1.1~1.5x time only... | Rebuttal 1:
Rebuttal: Dear all reviewers,
Thank you for dedicating your time to give valuable comments on our paper!
Here we want to response globally for the primary contribution, and experiments about some detail settings about ConML.
# Common Reply 1:
## Primary contribution.
We regard the primary contribution of ... | NeurIPS_2024_submissions_huggingface | 2,024 | null | null | null | null | null | null | null | null |
OSLO: One-Shot Label-Only Membership Inference Attacks | Accept (poster) | Summary: In this work, the authors propose a novel label-only membership inference attack. This method involves adding an adversarial perturbation to compel a model, trained without the target data point, to misclassify the image. This approach requires only a single query to the model and achieves SOTA performance for... | Rebuttal 1:
Rebuttal: ---
**To Reviewer CtKQ:**
**Response to Weaknesses 1:**
Thank you for your feedback. We apologize for not including a comparison with YOQO [1] in our initial submission. Please note that OSLO and YOQO differ in several key aspects:
- **Effectiveness on Low FPR Settings:** The accepted norm in th... | Summary: The paper proposes a new label-only membership inference attack that only needs a single query to the target model to perform the attack. The general idea of the proposed attack is based on the observation that member samples need more adversarial perturbation to force a misclassification than non-member sampl... | Rebuttal 1:
Rebuttal: ---
**To Reviewer Vvp4:**
**Response to Weaknesses 1:**
Thank you for your insightful feedback. We apologize for not including a comparison with YOQO in our paper. Please note that the accepted norm in the MIA community is that achieving low FPR is key for MIAs to be practical [1][2]. Although YO... | Summary: The paper investigates the vulnerability of deep learning models to membership inference attacks (MIAs), focusing on label-only settings where only the predicted hard label is available. It introduces the One-Shot Label-Only (OSLO) MIA, which infers a sample's membership in the training set with a single query... | Rebuttal 1:
Rebuttal: ---
**To Reviewer B8Vg:**
**Response to Weakness 1:**
Thank you for your suggestion. We agree that deeper research into defenses against such attacks is a promising direction. However, due to space limitations, we focused our contribution on improving label-only MIAs in terms of query cost and at... | Summary: This paper presents a new label-only membership inference attack, which requires only a single query to the target model. Specifically, it leverages transfer-based black-box attacks to generate an adversarial perturbation. This perturbation is then added to the sample and input into the target model, and its p... | Rebuttal 1:
Rebuttal: ---
**To Reviewer uq6v:**
**Response to Weaknesses 1:**
Thank you for your feedback. While we acknowledge that training these surrogate models introduces additional computational overhead, it is a **one-time process** for each target model. Once completed, these models can be used to infer any ta... | Rebuttal 1:
Rebuttal: We thank the reviewers for their insightful feedback and constructive suggestions, which have helped us enhance the clarity and robustness of our paper. We are committed to addressing all concerns and providing a comprehensive and well-documented study in the revised version. Our responses to each... | NeurIPS_2024_submissions_huggingface | 2,024 | null | null | null | null | null | null | null | null |
DiffHammer: Rethinking the Robustness of Diffusion-Based Adversarial Purification | Accept (poster) | Summary: This work proposes a new attack evaluation for diffusion-based purification methods: the 1 + N evaluation, which incorporates expectation maximization-based attacks and N-time evaluation. This method is helpful to evaluate the worst-case robustness of stochasticity-based defense methods.
Strengths: 1. This pa... | Rebuttal 1:
Rebuttal: Thank you for the valuable feedback, here's our response to the concerned problems.
## Advantages and costs of N-evaluation
**Summary:** We were the first to propose using $N$-evaluation to enhance evaluation accuracy and improve attack effectiveness and efficiency. In terms of evaluation, we dem... | Summary: The paper proposes a new adversarial attack framework against diffusion-based purification defenses. The paper first explains the advantage of using N-time evaluation as the metric for randomized defenses. Then, it proposes an E-M based adversarial attack, which empirically shows SOTA ASR.
Strengths: 1. Attac... | Rebuttal 1:
Rebuttal: Thank you for the valuable feedback, here's our response to the concerned problems. We will describe our EM framework more clearly and rigorously in response to concerned problems, and we will also update this part of the exposition in our paper.
Given a sample $x\in\mathbb{R}^d$ and its label $... | Summary: Diffusion-based purification methods have gained recognition for their robustness against adversarial attacks. However, concerns arise regarding the adequacy of current evaluation methods, particularly in addressing the gradient dilemma inherent in these techniques. This paper introduces DiffHammer, an advance... | Rebuttal 1:
Rebuttal: Thank you for the valuable feedback, here's our response to the concerned problems.
## Validation of the gradient dilemma
**Summary:** By conducting cluster analysis on the gradients of the purification process, we observed that the gradients have multiple clustering centers with low inter-cluster... | Summary: This paper reveals two limitations of Eot-based attacks in diffusion-based purification: gradient dilemma and underestimation of resubmit attacks' risk. The authors introduce N-time evaluation to evaluate the risk of resubmitted attacks sufficiently. Then, they propose an EM-based attack to solve the gradient ... | Rebuttal 1:
Rebuttal: Thank you for the valuable feedback, here's our response to the concerned problems.
## Comparison with newer methods
**Summary:** We adhere to the literature's naming conventions for these methods, but our evaluation incorporates the enhancements proposed in \[1,2\] (2023), upgrading PGD and AA to... | Rebuttal 1:
Rebuttal: Thank you for the valuable feedback, here's our response to the concerned problems. We have provided some additional images attached to the pdf.
Pdf: /pdf/68ef61c5ec5126ac6acb299926dfcf8427e80dcf.pdf | NeurIPS_2024_submissions_huggingface | 2,024 | null | null | null | null | null | null | null | null |
Advancing Tool-Augmented Large Language Models: Integrating Insights from Errors in Inference Trees | Accept (poster) | Summary: This paper proposes a novel framework for improving tool-augmented large language models by incorporating insights from errors in inference trees. It constructs preference data from expert reasoning trajectories and finetune LLMs accordingly. Experiments demonstrate that the finetuned model significantly outpe... | Rebuttal 1:
Rebuttal: Many thanks for the constructive reviews! We will revise our paper accordingly.
------
**Q1**: "The paper can benefit from some case studies. Though numbers suggest that TP-LLaMA can master tool-usage instructions better and exhibits stronger generalization capabilities, case studies comparing t... | Summary: Empowering Large Language Models (LLMs) with external tools has been considered as an important research direction to extend LLMs' task scope. However, during the tool utilization of LLMs, it usually relied on successful tool execution path. Therefore, how to instruct LLMs to employ successful paths is an impo... | Rebuttal 1:
Rebuttal: Many thanks for the constructive reviews! We will revise our paper accordingly.
------
**Q1**: "In the proposed ToolPreference dataset, it just uses successful path and failure path as preferred and dispreferred data. However, I think such a selection is too easy as model will never choose any d... | Summary: The authors aimed to enhance the tool-augmented language model (LLM) by efficiently learning from the failed explorations that previous works had overlooked. Step-wise preference data were extracted by pairing the branch nodes with the corresponding child nodes from the successful path in the tool-using decisi... | Rebuttal 1:
Rebuttal: Many thanks for the constructive reviews! We will revise our paper accordingly.
------
**Q1**: "Computation Time Clarification: Could you specify the computation time on the mentioned infrastructure?"
**A1**: Below, we specify the computation time:
+ SFT training: 4.6h for 2 epochs.
+ DPO trai... | Summary: This paper proposes a method to collect preference data from decision trees, making use of the failed trajectories which are previously ignored. Based on ToolBench (https://arxiv.org/abs/2307.16789) and the decision trees provided in the original dataset, a preference dataset, ToolPreference, is built with th... | Rebuttal 1:
Rebuttal: Many thanks for the constructive reviews! We will revise our paper accordingly.
------
**Q1**. "The paper is related to the works on process reward, e.g., arxiv: 2310.10080, arxiv: 2312.08935, yet does not discuss about it in the related works section."
**A1**. Thank you for pointing out these ... | null | NeurIPS_2024_submissions_huggingface | 2,024 | null | null | null | null | null | null | null | null |
A Pairwise Pseudo-likelihood Approach for Matrix Completion with Informative Missingness | Accept (spotlight) | Summary: This paper describes a method for estimating missing values of a given non-completed matrix in a situation called MNAR, where the probability distribution for the observation depends on both the matrix index and the matrix value.
They propose a nonparametric method to prevent bias due to model selection. Whi... | Rebuttal 1:
Rebuttal: Thank you for the detailed review of our paper.
## Weakness
**W1: Please see our global rebuttal.**
**W2: It is worth highlighting that their contribution is to remove ill-posedness through the appropriate assumption. However, it is necessary to define the term "ill-posed" properly and to prove... | Summary: The paper proposes a novel matrix completion method where the entries are missing not at random. The author(s) tackle the problem using a penalized pairwise pseudo-likelihood approach. The key contributions of the paper are (a) a flexible separable observation probability assumption that depend on the measurem... | Rebuttal 1:
Rebuttal: We would like to thank you for the constructive feedback and review.
**Weakness: A bit more discussion on how the constants $\kappa$ and $\rho$ influence the right hand bound on equation (7) would be insightful.**
**A:**
Thanks for the suggestion. When $\kappa$ is small, it means there exists o... | Summary: Current matrix completion approaches mostly deal with non-uniform sampling scenarios. However, few approaches can allow the missingness to depend on the mostly unobserved matrix measurements. In this paper, the authors propose a regularized pairwise pseudo-likelihood approach for matrix completion in this sce... | Rebuttal 1:
Rebuttal: We greatly appreciate the effort you have spent reviewing our paper.
**Q1: Could the author give a more detailed definition of informative missingness and a concrete example?**
**A:** We used “informative missingness” to refer to what statisticians often called Missing Not At Random (MNAR) assum... | null | null | Rebuttal 1:
Rebuttal: We’d like to thank the efforts of all reviewers. The reviews are insightful and constructive to improve our work. We address some common questions in this global rebuttal and leave other questions in the detailed response to each reviewer.
We first discuss **the related works and evaluation of st... | NeurIPS_2024_submissions_huggingface | 2,024 | null | null | null | null | null | null | null | null |
SlimSAM: 0.1% Data Makes Segment Anything Slim | Accept (poster) | Summary: This paper introduces SlimSAM, a novel data-efficient method for compressing SAM. The authors propose an alternate slimming framework and disturbed Taylor pruning to enhance knowledge retention and compression performance with minimal training data. The method progressively prunes and distills distinct sub-str... | Rebuttal 1:
Rebuttal: ## **Q1: It would be better to analyze the impact of different Gaussian distributions on the pruning performance.**
Thanks for the valuable suggestions. The mean $\mu$ of Gaussian noise must be set to 0 so that the average of the sampled noise across the entire batch approximates zero. The Gaussia... | Summary: The authors present a data-efficient method for SAM 39 compression, SlimSAM, comprising of alternate slimming framework and the disturbed Taylor pruning. Comprehensive experiments are conducted, demonstrating the significant superior performance and data efficiency.
Strengths: 1. The paper is clearly written ... | Rebuttal 1:
Rebuttal: ## **Q1: There's a typo with $w_i$ on line 112.**
Thanks for the careful review. We will correct this error in the next version and check the rest of the submission.
## **Q2: The symbol *N* is assigned different definitions on lines 111 and 187.**
Thanks for the valuable feedback. We apologize fo... | Summary: This paper works on compressing the Segment Anything Model (SAM). The compression method has the advantage of using much fewer training data. It has an alternate slimming framework and the label-free importance estimation criterion, i.e., disturbed Taylor pruning. The alternate slimming framework minimizes the... | Rebuttal 1:
Rebuttal: ## **Q1: In Eq.(4), the paper introduces disturbed Taylor importance, with the Gaussian noise with zero mean and standard deviation. How will the deviation values influence the performance and why 0.01 is a reasonable setting?**
The noise deviation should be carefully chosen: a too-large deviatio... | Summary: In this paper, the authors propose a novel and data-efficient SAM compression method called SlimSAM, which progressively compresses the SAM model by alternately pruning and distillation. They introduce disturbed Taylor pruning to address the misalignment between the pruning objective and the training target, u... | Rebuttal 1:
Rebuttal: ## **Q1:the authors' experiments seem to compare only parameter count and computational cost, without testing the throughput of the compressed model in real-world scenarios. The reviewer hopes to see relevant discussions in the experiments to better demonstrate the method's practical value in real... | Rebuttal 1:
Rebuttal: Dear Reviewers, Chairs,
We sincerely appreciate the time and effort you have spent evaluating our submission, and we look forward to the discussion stage. We will include the review stage results in the appendix of our next version. | NeurIPS_2024_submissions_huggingface | 2,024 | Summary: previous works on SAM compression replace the heavy image encoder with lightweight counterparts which requires dealing with the trade-off between training cost and model performances. Thus the end performances are usually compromised. This paper proposes a sliming framework to perform pruning and knowledge dis... | Rebuttal 1:
Rebuttal: ## **Q1: how is the training set (0.1%) selected? are there any impacts if a different training set is sampled and used to train SlimSAM?**
Thanks for the comment. The training set is a **random subset** of the SA-1B dataset, containing 11k images, with 10k used for training and 1k for validation.... | null | null | null | null | null | null |
Recurrent Complex-Weighted Autoencoders for Unsupervised Object Discovery | Accept (poster) | Summary: The authors propose a new way to obtain binding through complex-valued activations in an autoencoder for object discovery. They introduce recurrence in the generated phase map and perform the operations using complex weights.
Their model reaches competitive performance on a SOTA benchmark for unsupervised obj... | Rebuttal 1:
Rebuttal: We thank the reviewer for their feedback on our work and appreciate the positive feedback on the quality of our presentation.
``` But this behavior can potentially also be obtained with real-valued weights. In fact, the weights are applied to the real and imaginary parts of a complex activity. T... | Summary: The paper aims to solve the problem of unsupervised object segmentation, which can be viewed as a feature binding problem when feature detectors process images at different locations in parallel. It proposes an improvement to synchrony-based approach which uses complex numbers as activation value for neurons: ... | Rebuttal 1:
Rebuttal: We thank the reviewer for their support of the motivation behind the use of complex weights and its rationale. We thank them for their support of the simplicity of our model which eliminates the need for contrastive training, $\chi$-binding and other heuristics.
``` I appreciate that the authors ... | Summary: The paper introduces SynCx, a fully convolutional auto-encoder with complex values that is applied iteratively. This recurrent model takes advantage of the ability of complex-valued weights to not only compute features but also bind similar features that can be encoded in the phase, creating a grouping effect.... | Rebuttal 1:
Rebuttal: We thank the reviewer for their positive comments regarding the simplicity of our model, the ease of understanding the rationale behind our design choices and for noting that the evaluations support the main claims in our paper.
``` It has been shown that TSNE have some limitations, specificall... | Summary: Summary:
The authors introduce SynCx, a fully convolutional recurrent complex-valued autoencoder (AE) designed for unsupervised object discovery. SynCx operates by iteratively refining a randomly initialized phase array, which is the same shape as the input image, to reconstruct the input image. At each stage... | Rebuttal 1:
Rebuttal: We thank the reviewer for their positive comments and support for the simplicity and elegance of our model/training, quality of our experimental validation/analysis and the overall presentation. We’re happy that the reviewer enjoyed reading our work.
``` However, it is unclear how well SynCx wo... | Rebuttal 1:
Rebuttal: Please find attached a PDF containing samples of the UMAP v/s t-SNE comparison for dimensionality reduction in our phase map visualization process.
Pdf: /pdf/a4cf995a032e36b9de934ce5a05364f19a7f8a85.pdf | NeurIPS_2024_submissions_huggingface | 2,024 | Summary: The authors propose SynCx: a model of synchrony-based perceptual grouping. SynCx is a recurrent autoencoder with complex-valued parameters. By design, phase-information in the model is carried through across model iterations allowing it to be "stateful". SynCx demonstrates competitive/comparable performance on... | Rebuttal 1:
Rebuttal: We thank the reviewer for acknowledging the simplicity of our approach which does not require sophisticated heuristics or additional training objectives or forms of supervision. We address your specific concerns below.
``` L178 "We use latent-level complex-valued features ... than simply color cu... | null | null | null | null | null | null |
Curvature Clues: Decoding Deep Learning Privacy with Input Loss Curvature | Accept (spotlight) | Summary: This paper conducts a theoretical and empirical investigation of the use of input loss curvatures in MIAs. They study the train-test input loss curvature scores and use this to develop a zero order input loss curvature black-box MIA.
Strengths: 1. This paper provides an interesting theoretical explanation of ... | Rebuttal 1:
Rebuttal: We thank the reviewer for their feedback, we address the weaknesses and questions below:
1. Missing some references, e.g. Li et al. 2023 or the original DetectGPT paper (Mitchell et al. 2023).
A: We thank the reviewer for the feedback, we will include these references in the revised version of t... | Summary: This paper builds the connection between privacy and input loss curvature. Core observation is that test samples often lie in higher curvature areas than those prototypical samples, so the input loss curvature can be a good metric to distinguish between train and test sets. Utilizing this observation, authors... | Rebuttal 1:
Rebuttal: We thank the reviewer for their feedback, we address the weaknesses and questions below:
1. What is $s_f$, $L_f$ and $c_f$ in Line 300? The notation seems to be missing from the context. Line 117 missing a .
A: $s_f$, $L_f$ and $c_f$ are the fitting coefficients for equation 8, we will clarify t... | Summary: This paper develop a theoretical framework to derive an upper bound on train-test distinguishability based on input loss curvature. The authors propose a new black-box membership inference attack utilizing input loss curvature, which surpasses existing methods in membership inference effectiveness. The paper a... | Rebuttal 1:
Rebuttal: We thank the reviewer for their feedback, we address the weaknesses and questions below:
1. In the theoretical analysis, the authors propose the train-test distinguishability distribution upper bound for both data and curvature. However, they compare the upper bounds directly and suggest that cur... | Summary: The paper explores the use of loss curvature, under the hypothesis that test samples lie in high curvature regions, to improve the performance of membership inference attacks. It estimated the curvature by measuring the trace of the Hessian of the loss values. It demonstrates the viability of the approach in b... | Rebuttal 1:
Rebuttal: We thank the reviewer for their feedback, we address the weaknesses and questions below:
1. While the loss value based MIA evaluation generalizes to pretty much all class of models, it’s unclear if loss curvature based MIA attack would enable a similar generalization, i.e., going beyond the small... | Rebuttal 1:
Rebuttal: We thank all the reviewers for their input, there were a few common questions and weaknesses that we discuss and address below
1. Performance of zero-order curvature estimation vs curvature.
A: We provide additional results on the effect of zero-order estimation. We run the proposed MIA attack o... | NeurIPS_2024_submissions_huggingface | 2,024 | null | null | null | null | null | null | null | null |
Fine-grained Control of Generative Data Augmentation in IoT Sensing | Accept (poster) | Summary: This work introduces fine-grained control to generative models for IoT data augmentation by defining a metric space with statistical metrics.
These metrics are essential for capturing the key features of the spectrograms and serve as strong constraints for the generative model, allowing for customization of ... | Rebuttal 1:
Rebuttal: ### **Author Reply to Reviewer eQAt (Part 1/3)**
Thank you for your approvals, critiques and insightful questions! We would like to address your concerns and provide extra evaluation results as follows.
> W1: How to balance the original generative model loss and the designed metric loss is not w... | Summary: The paper addresses the challenge of overfitting in Internet of Things (IoT) sensing models caused by data distribution shifts between training datasets and real-world scenarios. To enhance model robustness, the authors introduce a novel data augmentation paradigm that adds fine-grained control to generative m... | Rebuttal 1:
Rebuttal: Thank you for all the comments! We would like to add the following clarifications to address your concerns.
> W1: Table 2 (Section 4.5) Statistical Significance: Including statistical significance tests (e.g., p-values) for the performance differences would strengthen the claims and provide more ... | Summary: This paper aims to synthesize IoT data to augment sensing models with generative models. The authors designed metric spaces to model the conditions which control the generative models. Experiments involving different IoT tasks, generative models, and sensing models have been conducted, proving the effectivenes... | Rebuttal 1:
Rebuttal: ### **Author Reply to Reviewer 8SZC (Part 1/4)**
Thank you for your constructive suggestions and insightful questions! We would like to add further clarification and evaluation to address your concerns.
> W1: Although I understand the novelty of using the proposed metrics as conditions for gener... | Summary: This paper is a study on data augmentation techniques for IoT sensing applications. The authors first present work on the most effective data augmentation techniques in the field of IoT sensing data. Based on this, they propose that domain knowledge can be combined with these data augmentation techniques to gu... | Rebuttal 1:
Rebuttal: We appreciate your positive comments on the paper idea and the evaluations. We would like to make the following responses and hope they can address your concerns.
> W1: The motivation of this paper to propose a fine-grained control approach, incorporating domain knowledge into the data enhancemen... | Rebuttal 1:
Rebuttal: ## General Responses
We greatly appreciate the detailed feedback provided by the reviewers. In response to your comments and suggestions, we have made the following major modifications:
### Reviewer 7DMe
- **Clarification of Motivation**:
- Clarified our motivation by elaborating on the unique... | NeurIPS_2024_submissions_huggingface | 2,024 | null | null | null | null | null | null | null | null |
Interpretable Concept Bottlenecks to Align Reinforcement Learning Agents | Accept (poster) | Summary: This paper provides evidence for the benefits of concept-based models in RL tasks, especially for identifying issues like goal misalignment. The proposed SCoBot integrates relational concepts into decision processes with minimal feedback, which not only brings the advantage of enhanced human understanding (Obj... | Rebuttal 1:
Rebuttal: We first want to thank the reviewer for their work and their appreciation of our manuscript, notably for highlighting the novelty of our method and the clarity of our visualizations. We are also adding to our manuscript new ablation studies. The reviewer can find them in the general answer and in ... | Summary: The authors introduce SCoBots, an explainable RL approach that first extracts symbols from images, extracts relations between the objects, and then uses them to train a distilled decision tree policy. They demonstrate better performance than naive CNN-based deep RL agents in Atari games.
For full disclosure, ... | Rebuttal 1:
Rebuttal: We thank the reviewer, again, for their valuable time and feedback. We are glad that they found our paper _very well written_, our method _quite interesting_ and that our experiments _do a good job of demonstrating advantages of [our] approach_.
Let us now address the reviewer's concerns.
**Add m... | Summary: The paper introduces Successive Concept Bottleneck Agents (ScoBots) that utilizes concept bottleneck models to enhance interpretability and decision-making. SCoBots incorporate successive concept bottleneck layers to integrate relational and object-based concepts for action selection. Unlike traditional deep R... | Rebuttal 1:
Rebuttal: We thank the reviewer for their dedicated effort to help us improve our manuscript. Let us address the raised concerns.
**1. Limited evaluation (Atari)**: ALE environments are light, incorporate a diverse set of challenges, and are still heavily used by RL practitioners. E.g., searching "reinforc... | Summary: This paper proposes Successive Concept Bottleneck Agents that takes an object oriented view on the environment and construct interpretable policies based on objects and associated properties.
Strengths: - The proposed method is well motivated and has strong empirical performance.
- The object oriented idea ha... | Rebuttal 1:
Rebuttal: We thank the reviewer for their high appreciation of our paper and valuable feedback. We are glad that they found our method *well motivated* and it's *strong empirical performances*. We address the reviewers' raised points hereafter.
**Add an experiment with inaccurate object detectors:**
We fi... | Rebuttal 1:
Rebuttal: We first thank all the reviewers for their time and valuable feedback.
The reviews are highly appreciated as they raise very interesting points, which ultimately enable us to further improve our manuscript.
Reviewers **KTbS**, **Zj4A** and **udPa** point out the beneficial and interesting role ... | NeurIPS_2024_submissions_huggingface | 2,024 | null | null | null | null | null | null | null | null |
Online Control in Population Dynamics | Accept (poster) | Summary: This paper addresses the challenge of controlling partially observable linear dynamical systems (LDS) with adversarial disturbances. Traditional methods for controlling LDS assume full observability, but this assumption does not hold in many real-world applications. The authors propose the GPC-PO-Simplex algor... | Rebuttal 1:
Rebuttal: Thank you for your time and comments! Unfortunately we believe that the reviewer has missed the point of the paper, which is to extend non-stochastic control methods to *population dynamics*. See below for responses to specific criticisms:
**Idealized Assumptions?** The review claims that we assu... | Summary: The paper studies the problem of actively controlling a class of dynamical systems that are used to model population dynamics, like the SIR model or replicator dynamics. Building on prior work on non-stochastic / adversarial control of linear dynamical systems, the authors propose a new class of dynamics, simp... | Rebuttal 1:
Rebuttal: Thank you for your time and comments! We are glad that you appreciated our paper. We address your comments and questions below:
**Intuition about simplex LDS conditions.** The assumption that the control norm is independent of the state is indeed needed for technical reasons: without it, since $A... | Summary: This paper studies online control problems for the simplex Linear Dynamical System model. They propose a gradient-based control algorithm GPC-Simplex and show its regret bound. Experiments are conducted to validate the performance of GPC-Simplex.
Strengths: The paper is in general well-written and smooth to f... | Rebuttal 1:
Rebuttal: Thank you for your time and comments! We are glad that you found the problem of great interest, and found our paper easy to follow. We address your questions and concerns below:
**Novelty of the proposed algorithm.** First, we reemphasize that the original GPC algorithm does not work in the simpl... | Summary: The paper explores the problem of population control in more practical settings such as the presence of adversarial noise and time-varying cost functions by proposing a robust methodology derived from online non-stochastic control theory.
Strengths: The paper provides regret guarantees against policies that d... | Rebuttal 1:
Rebuttal: Thank you for your time and comments! Please see below for our responses to your concerns:
**Motivation for simplex LDS?** The basic motivation for the simplex LDS model, as discussed in the paper, is as a specialization/adaptation of the general LDS model to population dynamics, i.e. settings wh... | Rebuttal 1:
Rebuttal: We thank all the reviewers for their time and comments. For full disclosure, we would like to inform the reviewers that we discovered a technical error in one of the secondary results in the appendices (Theorem 29, an analysis of a proposed extension to GPC-Simplex for the partially observable set... | NeurIPS_2024_submissions_huggingface | 2,024 | null | null | null | null | null | null | null | null |
Unity by Diversity: Improved Representation Learning for Multimodal VAEs | Accept (poster) | Summary: This paper proposes a multimodal variational mixture-of-experts prior (MMVM) VAE to address challenges in learning representations of multimodal data. Instead of modeling the latent representation as a joint posterior, this work introduces a multimodal data-dependent prior to regulate the latent representation... | Rebuttal 1:
Rebuttal: Dear reviewer 6TPP,
Thank you again for your insightful review and feedback. In the following, we will address and answer your comments and questions.
---
> There is no direct evaluation of the learned representation and the data-dependent prior to support the "soft-sharing of information betw... | Summary: The paper introduces the multimodal variational mixture-of-experts prior (MMVM) VAE, a new approach to enhance representation learning for multimodal Variational Autoencoders (VAEs). The MMVM VAE overcomes the limitations of the alternatives that parameterize joint posteriors with restrictive assumptions by us... | Rebuttal 1:
Rebuttal: Dear reviewer eP7z,
Thank you again for your insightful review and feedback. In the following, we will address and answer your comments and questions.
---
> The experimental analysis presented in the main text is not fully convincing. While the evaluation is properly conducted on several datas... | Summary: In this work, the authors proposed a new multimodal VAE architecture which, quickly summarized, both encoders and decoders factorized across modality latent variables, and the prior for each modality latent, $z_m$, is a mixture-of-experts composed of all the encoders. With this probabilistic model choice, the ... | Rebuttal 1:
Rebuttal: Dear reviewer 56cj,
Thank you again for your insightful review and feedback. In the following, we will address and answer your comments and questions.
---
> The experiments should compare with more modern approaches like MMVAE+.
We show the results for the MMVAE+ in the rebuttal PDF. We excl... | null | null | Rebuttal 1:
Rebuttal: Dear Reviewers,
We would like to thank all reviewers for providing comprehensive and valuable feedback. We particularly value the reviewers' recognition of our effective, interesting, and inspiring idea [**56cj**, **eP7z**, **6TPP**], the solid theoretical framework [**56cj**, **eP7z**], and the ... | NeurIPS_2024_submissions_huggingface | 2,024 | null | null | null | null | null | null | null | null |
An efficient search-and-score algorithm for ancestral graphs using multivariate information scores | Reject | Summary: Authors propose a greedy structure learning algorithm in the context of ancestral graphs. The score function is based on the normalized likelihood. The score is decomposable using the concept of ac-components. Limited experimental results are presented as experimental evidence.
Strengths: - The multivariate c... | Rebuttal 1:
Rebuttal: __Q1.__ _Are the limitations related to the cap of two-collider paths only? Are there any issues related to computational complexity? This was partially referenced in the attached form._
→ We thank this reviewer for emphasizing the main strengths of our work. The cap of two-collider paths does no... | Summary: In this paper, the authors present a novel likelihood decomposition of ancestral graphs, which is based on a novel concept ac-connected subset. With this decomposition, the authors present an efficient hybrid causal discovery method. But it is worthy to note that in the implementation of the causal discovery m... | Rebuttal 1:
Rebuttal: __Q1.__ _Could authors elaborate more on where Theorem 1 can be applied?_
→ We would like to thank this reviewer for underlining the importance of the topic and the clarity of our paper. Concerning this first question, we are not sure to understand the meaning of “where Theorem 1 can be applied”?... | Summary: The paper defines a score for ancestral graphs through an inclusion-exclusion expansion and implements it into a hybrid search approach.
Strengths: The likelihood decomposition (Theorem 1) is nice and there seems to be a slight improvement in the empirical testing.
Weaknesses: There are close overlaps to pub... | Rebuttal 1:
Rebuttal: __Q1.__ _Can the authors more clearly delineate what is novel and different here as compared to the other recent works of search-and-score for ancestral graphs?_
→ Concerning the overlaps of our work with recent papers and preprints on search-and-score methods for ancestral graphs, we actually di... | Summary: The paper proposes a greedy search-and-score algorithm for causal structure discovery in ancestral graphs which allow for both directed and bidirected edges.The paper provides an explicit decomposition of the likelihood function of ancestral graphs
in terms of multivariate cross-information over relevant ‘ac-c... | Rebuttal 1:
Rebuttal: __Q1.__ _What is the main Novelty of Theorem 1 compared to [14]? Can the theoretical motivation for the proposed search-and-score algorithm be deduced from the results in [14]?_
→ Hu and Evans 2020 (ref 14) proposed polynomial algorithms to find the Markov equivalence class of MAGs and ADMGs. Thi... | Rebuttal 1:
Rebuttal: We would like to thank all reviewers for their time, expertise, and insightful comments and questions which we address in separate replies below.
In particular, we address the novelty of our work with respect to recent published and submitted papers in our replies to Reviewers NFNu and mHLp.
S... | NeurIPS_2024_submissions_huggingface | 2,024 | null | null | null | null | null | null | null | null |
Neural Flow Diffusion Models: Learnable Forward Process for Improved Diffusion Modelling | Accept (poster) | Summary: The paper introduces Neural Flow Diffusion Models (NFDM), which incorporate a learnable forward process to enhance diffusion modeling. Traditional diffusion models rely on a fixed forward process, often complicating the reverse process and resulting in costly inference. NFDM aims to address these challenges by... | Rebuttal 1:
Rebuttal: We are delighted to see that the reviewer finds our approach novel, the theoretical explanations solid, and the experiments robust, and even recognizes the potential to influence future research. Below, we address the questions and comments raised in the review.
**Weaknesses:**
1. We thank the r... | Summary: This work proposes a generic framework to make diffusion model's encoder learnable.
Strengths: The authors propose a neat framework to make the encoding part of the diffusion model trainable. They also explain the differences and relationships between their proposal and related works.
Weaknesses: 1. The work... | Rebuttal 1:
Rebuttal: We are glad the reviewer found NFDM framework as neat. Below, we address the questions and comments raised in the review.
**Weaknesses:**
1. We define the distribution $q_\phi(z_t|x)$ implicitly through the function $F_\phi(\varepsilon, t, x)$. Under certain assumptions, we can derive an ordinar... | Summary: The authors propose a variant of a diffusion model in which the forward process is parameterized by a neural network that maps from data, noise, and a timestep to the noised data. They show how, if this neural network has certain properties, the SDE and ODEs of the forward and reverse processes can be simulate... | Rebuttal 1:
Rebuttal: We are pleased that the reviewer appreciates the elegance of our framework and recognizes the potential impact of our research direction. In response to the comments and questions raised, we provide the following clarifications and commitments:
**Weaknesses:**
1. We acknowledge the lack of visua... | Summary: The paper introduces Neural Flow Diffusion Models (NFDM), a novel framework that enhances diffusion models by supporting a broader range of forward processes beyond the standard Gaussian. The proposed parameterization technique facilitates the learning of the forward process and minimizes a variational upper b... | Rebuttal 1:
Rebuttal: We appreciate the reviewer's positive feedback regarding the clarity and presentation of our paper. Below, we address the questions and comments raised in the review.
**Weaknesses:**
1. Please refer to the answer to Question 1 below.
2. Unfortunately, due to limited computational resources, we w... | Rebuttal 1:
Rebuttal: We thank the reviewers for their valuable feedback, which helps us to improve the paper. Below, we have responded individually to the specific comments and questions from each reviewer.
Here we would like to provide some visualizations of samples generated with the NFDM and NFBM models trained on... | NeurIPS_2024_submissions_huggingface | 2,024 | null | null | null | null | null | null | null | null |
pcaGAN: Improving Posterior-Sampling cGANs via Principal Component Regularization | Accept (poster) | Summary: To solve image inverse problems (denoising, super resolution, etc.), many works attempt to sample from the posterior distribution of ground truth images given a degraded measurement. One line of works propose to train a stochastic neural network as a CGAN (conditioned on the degraded image and an additional ra... | Rebuttal 1:
Rebuttal: We sincerely thank the reviewer for their valuable feedback and time spent reviewing.
- **Reviewer**: I disagree with most of the claims made regarding the shortcomings of point estimators. The perception-distortion (PD) tradeoff theorem in [6] does not prove anything specific about point estimat... | Summary: The authors address inverse problems and aim to improve GAN-based posterior samplers by adopting regularization to correct the principal components of the covariance matrix of the approximated posterior distribution.
Strengths: 1. The paper is well written and easy to follow.
2. The idea presented in the pape... | Rebuttal 1:
Rebuttal: We sincerely thank the reviewer for their valuable feedback and time spent reviewing.
- **Reviewer**: For each dataset in the empirical study, a different set of baseline methods was chosen.
**Response**: Yes, this was done because there exists no single baseline method that is state-of-the-... | Summary: The paper focuses on learning posterior samplers for inverse problems, specifically aiming to accurately estimate the principal components of the posterior covariance matrix. To achieve this, the authors consider training a conditional generative adversarial network (cGAN), building upon previous work [21]. Th... | Rebuttal 1:
Rebuttal: We sincerely thank the reviewer for their valuable feedback and time spent reviewing.
- **Reviewer**: No ablation study of the rcGAN parameter $P_{rc}$, especially since $P_{rc}=2$ seems insufficient to estimate statistics properly.
**Response**: We use $P_{rc}=2$ when training an rcGAN mode... | Summary: This work considers using posterior sampling to solve inverse problems via conditional GANs. Previous work has shown that cGANs are competitive in terms of reconstruction quality as compared to other generative models to sample from the posterior (e.g., diffusion models), while also having the advantage that, ... | Rebuttal 1:
Rebuttal: We sincerely thank the reviewer for their valuable feedback and time spent reviewing.
- **Reviewer**: Slight, potentially somewhat incremental, improvement to the prior rcGAN approach of [21].
**Response**: Please see the global rebuttal.
- **Reviewer**: Additional regularization terms are ... | Rebuttal 1:
Rebuttal: We sincerely thank all the reviewers for their valuable feedback and time spent reviewing.
In this global rebuttal, we address remarks made by multiple reviewers.
- **Reviewers n2zA and cM4t**: Incremental contribution over the prior work rcGAN [21]?
**Response**: No.
rcGAN enforces accu... | NeurIPS_2024_submissions_huggingface | 2,024 | null | null | null | null | null | null | null | null |
Conformal Alignment: Knowing When to Trust Foundation Models with Guarantees | Accept (poster) | Summary: The paper proposes to adapt conformal guarantees to large language models. They build upon existing work to provide bounds on FDR error for each model output by conformalizing the distribution of an additional alignment predictor trained on labeled aligned data. Experiments are conducted on two radiological re... | Rebuttal 1:
Rebuttal: Thank you for appreciating the importance of this topic, the writing quality, and the flexibility and versatility of our framework! We respond to your insightful questions in the following.
- **Q1: Technical novelty**. Thank you for the question! While this work applies the conformal selection fr... | Summary: This paper introduces Conformal Alignment, a framework designed to ensure that outputs from foundation models align with human values before deployment in high-stakes tasks, such as radiology report generation. By training an alignment predictor with reference data, the framework guarantees that a user-specifi... | Rebuttal 1:
Rebuttal: Thank you for recognizing the importance and relevance of the research question, the applicability and rigor of our framework!
- **Q1: Comparison with related works**. We would like to clarify that the emphasis of the current work is the novel instantiation of conformal selection for reliable use ... | Summary: This paper proposes a procedure similar to conformal prediction to select deployable units and control the FDR. Asymptotic result is provided to justify the correctness of the proposed algorithm. Real data experiments in QA and medical tasks are also provided to demonstrate the effectiveness of the proposed me... | Rebuttal 1:
Rebuttal: Thank you for appreciating the theoretical rigor and the writing of our paper! Please see below our detailed responses.
- **Q1: Use of the term “alignment”**. Thank you for pointing out the potential confusion! In fact, our proposed method can also be used for “alignment” in the strict sense: if w... | Summary: This work applies recent work on conformal outlier detection [1] in the context of LLM alignment. The authors assume that one, given input $X$ and output $Y$ of an LLM, has access to an alignment scorer $A$ that can then produce an alignment score of $A(X, Y)$. They treat the null hypothesis of being unaligned... | Rebuttal 1:
Rebuttal: Thank you for appreciating the rigor, power, and versatility of our framework! We address your comments in the following.
- **Q1: Novelty**. Thank you for raising this question! We would like to argue that the motivation of this paper is exactly to address the unique challenges of LLMs—the relia... | Rebuttal 1:
Rebuttal: We thank the reviewers for their valuable feedback and constructive suggestions for improvement. Overall, the reviewers find our work addresses an important problem [R1, R3, R4, R5], is well-written [R1, R5], versatile to broad contexts and criteria [R1, R2, R5], offers rigorous and practically us... | NeurIPS_2024_submissions_huggingface | 2,024 | Summary: The paper proposes to use conformal prediction to determine when to trust model outputs. To this end the authors apply conformal risk control on across unit instead of focusing on a single unit. This is especially relevant in their X-ray example where the goal is to give a set of trusted documents to the medic... | Rebuttal 1:
Rebuttal: Thank you for the insightful comments, and for appreciating the novelty and practicality of our framework, as well as our experimental results!
- **Q1, Technical challenges**: We would like to clarify that the emphasis of the current work is the novel instantiation of conformal selection (we als... | null | null | null | null | null | null |
Aggregate-and-Adapt Natural Language Prompts for Downstream Generalization of CLIP | Accept (poster) | Summary: This paper proposes a new method to enhance the downstream generalization of CLIP by distilling knowledge from LLM- or human-generated text prompts. The proposed method involves training a prompt generator to predicts prompt embeddings (AAPE) based on images.
Strengths: 1. The method aggregates and distills t... | Rebuttal 1:
Rebuttal: Thanks for your helpful suggestions to improve our work. Below is our point-by-point response.
**Whether the prompt aggregator can suppress noisy text prompts.**
The attached **Rebuttal.pdf** (Fig. 1) visualizes the attention score for some prompt samples. We do observe low scores for those nois... | Summary: The proposed method leverages language priors for better downstream adaptation and generalization of CLIP [37], which is similar to CuPL [36] that utilizes prompts generated by a LLM (e.g., GPT-3) for zero-shot image classification using CLIP. Unlike CuPL, the proposed method aggregates multiple LLM-generated ... | Rebuttal 1:
Rebuttal: Thanks for your recognition of our work and the detailed feedback! We respond to specific comments below.
**Does AAPE $h(x)$ mitigate the image-text modality gap?**
Great question! Short answer is yes. Essentially, AAPE belongs to those text prompt tuning methods that only operate at the text br... | Summary: The paper proposes a new prompt embedding named Aggregate-and-Adapted Prompt Embedding (AAPE), which improves prompt learning by distilling knowledge about more detailed descriptions of classes into prompt embeddings. Concretely, the “summary” prompt is obtained by aggregating diverse reference prompts. Then, ... | Rebuttal 1:
Rebuttal: Thanks for the positive feedback and constructive suggestions. For the efficiency concern, please refer to our **Response to Common Concern**. Answers to your other questions below.
**More details on the CLIP reward**
As mentioned in L160, we use the same CLIP reward as in [17], formulated as $\... | Summary: This framework first aggregates textual knowledge from human or large language model (LLM) generated prompts into a summary aligned with each input image. This is achieved using a prompt aggregator. A prompt generator is then jointly trained to create prompt embeddings that are close to this aggregated summary... | Rebuttal 1:
Rebuttal: **Compare with ArGue (CVPR 2024)**
ArGue and our AAPE both learn text prompts that distill language priors from LLM. However, the prompt learning mechanisms are different: ArGue learns individual prompt token vectors. They are combined with the embeddings of the class name and class-wise visual a... | Rebuttal 1:
Rebuttal: **Response to Common Concern**
Thanks to all reviewers for the thoughtful comments. Before responding to the questions raised by each reviewer, we first address the common concerns around efficiency.
The attached **Rebuttal.pdf** (Table 1) compares the inference compute cost between our AAPE and... | NeurIPS_2024_submissions_huggingface | 2,024 | Summary: The paper proposes a prompt learning method that distills the knowledge from pre-trained LLM while conditioning them on the image embedding. A Prompt Aggregator module combines the LLM generated prompts per image and the Prompt Generator module, generates a prompt from the image embedding. The modules are trai... | Rebuttal 1:
Rebuttal: Thank you for the constructive feedback on our work. Regarding the efficiency concern, please refer to our **Response to Common Concern**. For other comments, our point-by-point response is as follows.
**Compare and discuss how this work differs from ProText (arXiv:2401.02418)**
Thanks for bring... | null | null | null | null | null | null |
Cross-Device Collaborative Test-Time Adaptation | Accept (poster) | Summary: This paper proposes a novel Collaborative Lifelong Adaptation (CoLA) method for test-time adaptation (TTA) across multiple devices. The key idea is to accumulate and share knowledge learned during adaptation across devices, enabling more efficient and effective adaptation to distribution shifts at test time. T... | Rebuttal 1:
Rebuttal: > Q1. Comparisons with test-time federated learning.
A. Please refer to our General Response.
> Q2. More scalability analysis regarding CoLA.
A. CoLA scales well with an increasing number of devices, i.e., with more shared domain vectors. From Table A, ETA/SAR+CoLA consistently benefits from a... | Summary: This paper explores test-time adaptation (TTA) in scenarios involving varying computational resources across multiple devices. To address this, the paper proposes storing and sharing domain knowledge among devices to improve adaptation. For resource-abundant devices (referred to as principal agents), domain kn... | Rebuttal 1:
Rebuttal: > Q1. Comparison with Multi-Device Test-Time Adaptation.
A. Please refer to our General Response.
> Q2. Differences between CoLA and previous continual TTA work TTA-CDKM [A].
A. The main differences between our CoLA and TTA-CDKM [A] are in the following aspects:
1. **Our CoLA is more feasible... | Summary: This paper considers test-time adaptation in multi-device scenarios and proposes Collaborative Lifelong Adaptation (CoLA). Domain knowledge vectors, as interactive objects on devices, are maintained and stored by measuring KL-divergence between the current test batch and the online estimated distribution. This... | Rebuttal 1:
Rebuttal: > Q1. Ablation studies on the threshold z.
A. CoLA remains effective among a wide range of threshold z, as shown in Table A. From the results, while a stricter threshold saves more domain vectors, CoLA achieves a stable performance of around 64.7%. When threshold z increases, CoLA saves significa... | Summary: Prior test-time adaptation (TTA) methods focused on single-device setups only, despite models being deployed to multiple devices in the real world. This paper presents Collaborative Lifelong Adaptation (CoLA) to leverage this situation and improve performance. Specifically, this paper introduces Principal Agen... | Rebuttal 1:
Rebuttal: > Q1. Analyses on when prior knowledge helps and when it harms.
A. **Analyses on when prior knowledge helps**: In our paper, we seek to improve test-time model adaptation performance and efficiency by exploiting useful knowledge from multiple devices in the application environment. In practice, d... | Rebuttal 1:
Rebuttal: We deeply appreciate all reviewers for their valuable feedbacks and constructive comments on improving the quality of our paper. We would like to address your questions below.
> G1. Differences and advantages of our methods from/over federated TTA [A,B]:
- **Differences on problem settings and ... | NeurIPS_2024_submissions_huggingface | 2,024 | null | null | null | null | null | null | null | null |
The tree autoencoder model, with application to hierarchical data visualization | Accept (poster) | Summary: This paper proposes a dimensionality reduction approach that comprises a tree of local PCA projections. The authors make the following contributions:
1. A mathematical description of their approach.
2. An optimisation algorithm and pseudocode to fit their method to a training set, together with the computation... | Rebuttal 1:
Rebuttal: Thanks for your very thorough review.
- *Hyperparameter setting and sensitivity.*
In principle, since an autoencoder defines a supervised problem, the hyperparameters might be chosen using a hold-out validation set. However, for visualization purposes we set them manually as described next.
... | Summary: This study proposes PCA trees as a hierarchical data visualization tool. PCA trees hierarchically partition the data space using hyperplanes, with the partition pattern represented by a binary tree. PCA is then applied to each leaf node of this binary tree. The parameters of a PCA tree consist of coefficients ... | Rebuttal 1:
Rebuttal: Thanks for your review.
- *Does our PCA tree truly achieve dimensionality reduction in the strict sense?*
It does indeed, but we need to define the latent space and its coordinates accordingly. This latent "space" consists of 1) the discrete index of a leaf (local latent space) and 2) a contin... | Summary: The paper proposes a hierarchical data visualisation method that is a combined use of oblique decision tree and PCA over leaf nodes. In experiments, the authors demonstrate the visualisation results of the proposed method for a few datasets and compare the reconstruction errors with standard PCA.
Strengths: T... | Rebuttal 1:
Rebuttal: We thank you for your questions, but it seems there are several misunderstandings in your review. Hopefully our answers below will clarify them.
- *Very limited originality [...] The method description is quite hard to follow, which appears like a piece of overly complicated writing for simple me... | Summary: The authors nicely describe an extension of [5] to unsupervised dimensionality reduction.
The paradigm is quite different from regular DR (unique scatterplot) as it yields a tree and thus multiple leafs.
There are some similarities with hierarchical clustering (to some extent).
The authors also emphasize a con... | Rebuttal 1:
Rebuttal: Thanks for your review and in particular for the connections with other methods.
- *Novelty with respect to [5].*
Since trees making hard decisions are not differentiable, we cannot use gradient-based methods and instead we use alternating optimization over the nodes. While [5] also used alter... | null | NeurIPS_2024_submissions_huggingface | 2,024 | null | null | null | null | null | null | null | null |
Ferrari: Federated Feature Unlearning via Optimizing Feature Sensitivity | Accept (poster) | Summary: The paper introduces "Ferrari," a framework for federated feature unlearning that optimizes feature sensitivity to address the 'right to be forgotten' in machine learning. It allows clients to remove sensitive or biased features from a model without full retraining or other clients' data. Ferrari is efficient,... | Rebuttal 1:
Rebuttal: `` 1. Defend Adversarial Attack ``
Our framework effectively defends against Model Inversion Attacks (MIA) [1] by preventing the reconstruction of the target unlearn features. As shown in Tab 3 in main paper, the attack success rate with our framework is similar to that of the naive retrain metho... | Summary: This paper studies the feature unlearning problem in the context of federated learning. The key idea is to interpret feature unlearning into model sensitivity at specific features. The proposed method which is called Ferrari is evaluated on different tasks which involves sensitive, backdoor, and biased feature... | Rebuttal 1:
Rebuttal: `` 1. Challenge of paper ``
Thank you for your suggestions. We clarify the second challenge from the original manuscript as follows:
Firstly, previous feature unlearning methods [1, 2] in centralized settings are impractical for Federated Learning (FL), as they require access to all datasets and... | Summary: This paper introduces a novel framework named Ferrari, addressing two challenges for federated feature unlearning. Ferrari leverages Lipschitz continuity to minimize the feature sensitivity directly during the local unlearning process, demonstrating several key advantages, such as eliminating the requirement o... | Rebuttal 1:
Rebuttal: `` 1. Non-Lipschitz Analysis ``
We evaluate the Lipschitz loss function for optimizing feature sensitivity (Equation 6) and compare it with a variant lacking the denominator, termed the Non-Lipschitz loss (see Fig 1 of the rebuttal PDF). Results show that models using the Non-Lipschitz loss exhib... | Summary: The rise of Federated Learning (FL) has led to the importance of the 'right to be forgotten', inspiring a need for Federated Unlearning (FU). Feature unlearning, particularly for removing sensitive, backdoor, and biased features, has attracted significant interest. However, current methods relying on the influ... | Rebuttal 1:
Rebuttal: `` 1. Motivation ``
Thank you for raising this important question about the motivation for feature unlearning in Federated Learning (FL). In the case of IID data, where clients may share similar features, it is indeed valuable to consider the need for unlearning specific features. For example, a ... | Rebuttal 1:
Rebuttal: Dear all reviewers,
Thank you for your valuable comments and thoughtful inquiries regarding our work. We appreciate the opportunity to address your concerns and are committed to providing further clarity.
To better illustrate our responses, we have uploaded a PDF document containing new figures ... | NeurIPS_2024_submissions_huggingface | 2,024 | Summary: The paper identifies two main challenges, which is 1) evaluating the unlearning effectiveness without rebuilding the dataset without the unlearned feature, and 2) achieving feature unlearning in Federated Learning without requiring data or computational resources from other clients. The authors define a concep... | Rebuttal 1:
Rebuttal: `` 1. Large-Scale Dataset ``
Thank you. As requested, we conducted additional experiments using the ImageNet dataset to assess the generalization ability of our proposed framework. Below are the tables summarizing the results. These results demonstrate that our framework performs well even in la... | null | null | null | null | null | null |
DISP-LLM: Dimension-Independent Structural Pruning for Large Language Models | Accept (poster) | Summary: This paper concerns about the dimension dependence of hidden states across different layers. It have been argued that previous studies with dimension dependence would be inefficient due to additional parameters or architectural constraints. This work breaks the dimension dependence by introducing indexing oper... | Rebuttal 1:
Rebuttal: We want to thank the Reviewer BGu5 for your insightful and constructive comments. We will address each of your questions below. We use W1 to refer to weakness 1.
**W1. Although the proposed method allows asymmetric dimensions in module weights across layers, it also requires the dimensions of hid... | Summary: DISP-LLM introduces dimension-independent structural pruning for LLMs, breaking structural dependencies to allow flexible pruning across layers and dimensions. It uses a hypernetwork for pruning decisions and claims superior performance over existing methods on various LLMs without weight updates or additional... | Rebuttal 1:
Rebuttal: We want to thank the Reviewer AWnP for your insightful and constructive comments. We will address each of your questions below. We use W1 to refer to weakness 1 and Q1 to refer to question 1.
**W1. Limited novelty beyond combining existing concepts**
Our method is novel because it introduces a n... | Summary: This paper proposes a novel structural pruning method for LLMs, primarily based on SliceGPT. It characterised in (i) removing structural dependence by facilitating each block to possess varying widths along its input and output dimensions (ii) no need for introducing addition paramaters like sliceGPT. The emp... | Rebuttal 1:
Rebuttal: We want to thank the Reviewer 6CTn for your insightful and constructive comments. We will address each of your questions below. We use W1 to refer to weakness 1 and Q1 to refer to question 1.
**W1. More comprehensive literature review about model compressing, like sparse mechanism.**
Thanks for ... | Summary: Structural pruning is a method to prune the weights of large language models (LLMs) while keeping their original performance as much as possible it can. However, structural pruning has limitations caused by depending on the structure, like the residual connection of LLMs. In this work, the authors propose a ne... | Rebuttal 1:
Rebuttal: We want to thank the Reviewer FhgQ for your insightful and constructive comments. We will address each of your questions below. We use W1 to refer to weakness 1 and Q1 to refer to question 1.
**W1. Even though DISP-LLM does not update the original weights of LLMs, training is required.**
In fact... | Rebuttal 1:
Rebuttal: General Response to Reviewers:
We would like to thank all reviewers for their insightful comments. We are greatly encouraged to see that all of you hold a positive evaluation of our work. We will address some common comments in the general response. We also put part of the individual response her... | NeurIPS_2024_submissions_huggingface | 2,024 | null | null | null | null | null | null | null | null |
United We Stand, Divided We Fall: Fingerprinting Deep Neural Networks via Adversarial Trajectories | Accept (poster) | Summary: This work discusses a robust model fingerprinting method for protecting the intellectual property (IP) of deep neural networks (DNNs) using adversarial trajectories, referred to as ADV-TRA. Traditional single-point fingerprinting methods suffer from sensitivity to decision boundary changes, leading to high fal... | Rebuttal 1:
Rebuttal: Thanks for your careful reading and useful feedback. Below, we will address your specific questions.
- Q1: This is an interesting question. ADV-TRA's mechanism offers valuable insights for other domains or neural network architectures. For example, in NLP, generating a batch of continuous text is... | Summary: This paper introduces ADV-TRA, a fingerprinting scheme that uses adversarial trajectories to fingerprint DNN models. Unlike conventional methods that rely on single-point fingerprint samples, ADV-TRA generates a series of fixed-length trajectory samples, each showcasing varying levels of adversarial perturbati... | Rebuttal 1:
Rebuttal: Thanks for your careful reading and useful feedback. Below, we will address your specific questions.
- Q1: We really appreciate your thoughtful suggestions regarding the related works. We will include these three outstanding works to supplement the Related Work section in the next version of the ... | Summary: The paper proposes a novel fingerprinting scheme for DNN classifiers.
Strengths: S1: The number of queries is relatively small (around 100)
Weaknesses: W1: The authors see fingerprinting only as a technical means for IP protection. Fingerprinting may have different applications. And, as IP protection is conc... | Rebuttal 1:
Rebuttal: Thanks for your careful reading and useful feedback. We address your concerned points below.
- Q1&Q2&W3: We thank the reviewer for pointing this issue out. Regarding the main performance results in Table 1, we follow the evaluation methodology from previous works [1,2]. Specifically, We ran the ex... | Summary: The paper proposes a method to fingerprint deep neural networks in order to protect the model intellectual property. The authors focus on a technique post-training (without requiring changes to the training procedure), based on trajectories of adversarial examples called ADV-TRA. In contrast to other works tha... | Rebuttal 1:
Rebuttal: Thanks for your careful reading and useful feedback, and we address your specific questions below.
- Q1 & W3: Thank you for pointing out this issue. We have discussed the issue of time overheads in Appendix D.5, which reflects the actual computational costs during the execution process. In fact,... | Rebuttal 1:
Rebuttal: Dear PC,
We have some reservations about the claims made by Reviewer u7bH with a confidence level of 5. There are two apparent misunderstandings or errors in his/her review:
- First, Reviewer u7bH has confused the meaning of fingerprinting in the biological domain and the deep learning domain. I... | NeurIPS_2024_submissions_huggingface | 2,024 | null | null | null | null | null | null | null | null |
Randomized Strategic Facility Location with Predictions | Accept (poster) | Summary: The paper considers the strategic (single) facility location problem where the goal is to decide on the location of a single facility given the preferred locations {p_i} of n agents. The preferred locations p_i are private information and the goal is to design strategyproof mechanisms so that the agents reveal... | Rebuttal 1:
Rebuttal: Thank you for your helpful comments and suggestions.
* **Reviewer:** "The paper mainly gives lower bounds in the different settings. The only positive result is obtained via running the centroid mechanism on the (predicted) extreme agents."
**Response:** Although many of our results take the for... | Summary: In this paper, the authors study the metric space facility location problem with machine-learned predictions. Specifically, they focus on the setting that the mechanism has predictions of every agent’s preferred location and with randomized mechanisms. The goal is to find truthful mechanisms that minimize the ... | Rebuttal 1:
Rebuttal: Thank you for your helpful comments and suggestions.
* **Reviewer:** Minor comments.
**Response:** We agree with all the suggested changes and have applied all of them.
Regarding the first comment, we have edited the statement as follows:
> We then turn to mechanisms augmented with the stronge... | Summary: The paper revisits the strategic facility location problem, focusing on the role of randomization in designing truthful mechanisms enhanced by machine-learned predictions. The authors build on a recent framework that refines worst-case results by incorporating potentially incorrect predictions about agents' tr... | Rebuttal 1:
Rebuttal: Thank you for your helpful comments and suggestions.
* **Reviewer:** Lemma 1 should quantify $\alpha$ and $\beta$. Theorem 3 should quantify $\epsilon$.
**Response:** Thank you for pointing this out. We have now quantified all these parameters and we provide the reworded statements below. Furt... | Summary: The authors study the problem of designing incentive-compatible mechanisms for the facility location game, in particular a learning-augmented variant of ABGOT22 where an oracle returns (potentially inaccurate) predictions about optimal solutions. The paper focuses on considering when predictions are available ... | Rebuttal 1:
Rebuttal: Thank you for your helpful comments and suggestions. We actually think that our 1-dimensional results are quite significant on their own and much more than a warm-up for the two-dimensional ones. We provide some justification for this fact in response to the reviewer's question below.
* **Revie... | Rebuttal 1:
Rebuttal: We thank all the reviewers for the time and effort they dedicated to evaluating our paper. We have provided individual responses for their questions and would be more than happy to provide additional details if any of these responses do not fully address the reviewers' concerns. Attached to this r... | NeurIPS_2024_submissions_huggingface | 2,024 | null | null | null | null | null | null | null | null |
PV-Tuning: Beyond Straight-Through Estimation for Extreme LLM Compression | Accept (oral) | Summary: The paper provides a new QAT algorithm for extreme compression of LLMs under various discrete weight representations (including vector quantization, uniform quantization). Traditional optimization techniques rely on straight-through estimation (STE) for optimization in the presence of discrete parameters. Thi... | Rebuttal 1:
Rebuttal: We thank the reviewer for their feedback and address their concerns below.
> (Clarity) The presentation (especially the algorithmic, technical details) in the paper needs to be significantly improved.
We agree with your comments on the presentation, and will try to make the algorithm definition e... | Summary: This paper proposes a novel PV-tuning algorithm for extreme compression of LLMs. In each iteration, two steps (P step and V step) are performed, each aimed at reducing the loss function. Different than existing methods that use straight-through estimator (STE), PV-tuning employs a subspace search strategy at t... | Rebuttal 1:
Rebuttal: We thank the reviewer for their feedback and we are glad that they appreciate our technical contribution and experiments. Below we do our best to address their concerns and answer questions.
> (W1) PV-tuning is computationally more expensive than existing PTQ methods and requires more GPU resour... | Summary: This paper proposes a new PTQ fine-tuning algorithm called PV tuning. Recent SOTA LLM PTQ works include fine tuning steps to recover the original model on top of the actual quantization step. Fine tuning has been shown to be an effective and relatively cheap to run (vs full QAT) way to improve quantization qua... | Rebuttal 1:
Rebuttal: We thank the reviewer for their feedback and address their concerns below.
> (W1) This paper is written in a dense and arcane way and is hard to follow. While there are some theorems, they do not appear to be very useful (e.g. 3.1 just claims that PV tuning converges, but without any statements o... | Summary: The paper focuses on fine-tuning techniques over compressed weights and proposes PV-Tuning, a general framework that improves existing fine-tuning strategies and provides convergence guarantees. Experiments show that PV-Tuning outperforms prior techniques and achieves the first Pareto-optimal quantization for ... | Rebuttal 1:
Rebuttal: We are glad that the reviewer appreciates our technical contribution and extensive experiments. We do our best to address their concerns and answer questions below.
> (W1) The paper introduces the principle and specific implementation of PV-Tuning in detail, but does not compare the differences i... | Rebuttal 1:
Rebuttal: We thank the reviewers for providing valuable comments and suggestions. We are glad that the reviewers appreciate our strong empirical results (3n7c, 5n4w, p3Lv, NP8P) including the latest models (p3Lv). Reviewers also highlight our technical contribution (5n4w, 3n7c, NP8P) and theoretical guarant... | NeurIPS_2024_submissions_huggingface | 2,024 | Summary: The paper proposes the use of fine-tuning over highly-compressed models to achieve better model compression. By introducing the PV-tuning algorithm, the authors have achieved superior discrete parameter optimization compared to traditional STE methods. The effectiveness of PV-tuning has been validated by the a... | Rebuttal 1:
Rebuttal: We are glad that the reviewer appreciates our technical contribution and accuracy improvements. Below, we do our best to address concerns and answer questions raised in the review.
> (W1) What does “one shot” imply? Does it refer to layer-wise quantization?
Indeed, in L37-38, one-shot quantizati... | null | null | null | null | null | null |
Sample-Efficient Geometry Reconstruction from Euclidean Distances using Non-Convex Optimization | Accept (poster) | Summary: This paper proposed a method called MatrixIRLS to solve the EDG problem, an important problem in various fields, especially when pairwise distance is partially observed. The proposed method improves the number of random sampled distances to guarantee local convergence to the ground truth.
Strengths: The relat... | Rebuttal 1:
Rebuttal: Thank you for your careful feedback.
We appreciate your pointers regarding typos and your presentation suggestions. We will refer to the definition of the coherence $\nu$ at the first occurrence in the paper.
> Figure 2 looks great, but it will be even better if we can observe the claimed rate $... | Summary: The paper studies the under-constrained problem to reconstruct n ordered points in Euclidean space R^r, which are given by fewer than all n(n-1)/2 pairwise distances.
Strengths: The authors should be highly praised for clearly stating main Problem 1 on page 1 and for formulating main Theorems 4.2 and 4.5 almo... | Rebuttal 1:
Rebuttal: Thank you for your thoughts about our submission. Please consider our responses regarding your assessment:
> (Meaningfulness of rank min. for Problem 1)
Assuming the points matrix $P$ has rank $r$, it follows that the Gram matrix also has rank $r$. When $n \gg r$, this implies that the rank of ... | Summary: This paper tackles the problem of reconstructing geometric configurations from minimal Euclidean distance samples. The authors propose an innovative algorithm based on iteratively reweighted least squares (IRLS) within a non-convex optimization framework. They provide a local convergence guarantee for the meth... | Rebuttal 1:
Rebuttal: We appreciate your constructive and very detailed feedback to our submission. A point-by-point response and clarification regarding your comments follows below.
**Weaknesses**
> 1. The paper compares the proposed method with a few state-of-the-art algorithms. Including more diverse comparisons, ... | Summary: The paper examines the problem of recovering point locations based on incomplete pairwise Euclidean distance information. The problem is cast as a low-rank matrix recovery problem, and an algorithm based on iteratively-reweighted least squares is proposed. The authors derive a lower bound for the minimum numbe... | Rebuttal 1:
Rebuttal: We appreciate your constructive reviews our submission. We address your questions below.
> While number of steps/time to convergence is shown for ill-conditioned matrix experiments, they are not shown for real data experiments. How much faster does the proposed MatrixIRLS derivative converge comp... | Rebuttal 1:
Rebuttal: We appreciate the reviewers' constructive and very detailed feedback to our submission. A point-by-point response and clarification regarding their comments have been addressed separately to each reviewer, however, we highlight some key points below, many of which are supported by additional exper... | NeurIPS_2024_submissions_huggingface | 2,024 | null | null | null | null | null | null | null | null |
Schedule Your Edit: A Simple yet Effective Diffusion Noise Schedule for Image Editing | Accept (poster) | Summary: This paper focuses on improving the stability of DDIM inversion to produce a better noise space for image editing. They found that the key issue is the singularity schedule, and try to overcome this issue by adopting the logistic schedule rather than the linear schedule.
Strengths: 1 This paper found that the... | Rebuttal 1:
Rebuttal: We appreciate your thoughtful insights, and respond to the suggestions and concerns as follows:
**Response to Q1**
> ***Difference compared with sigmoid schedule [1,2] ....; can the sigmoid schedule also solve the proposed singularity problem?***
Thanks for providing this valuable work [1,2].
... | Summary: This paper introduces a novel noise schedule called Logistic Schedule for text-guided image editing. The authors identify that traditional DDIM inversion methods suffer from accumulated prediction errors, leading to suboptimal content preservation and edit fidelity. To address this, this paper proposes the Log... | Rebuttal 1:
Rebuttal: We appreciate your valuable feedback. Before addressing your inquiries, we wish to clarify certain weaknesses highlighted in the review that we believe require further elucidation.
**Response to W1**
> ***Limited improvements: Although there are improvements in the quantitative comparisons, the d... | Summary: The paper proposes a noise schedule "Logistic Schedule" to solve the singularity problem in the traditional noise schedules for text-guided diffusion models for image editing. It aims to improve the stability of the DDIM inversion process, reducing noise prediction errors and enhancing content preservation and... | Rebuttal 1:
Rebuttal: Thank you for your positive comments and helpful suggestions.
**Response to W1**
> ***Additional comparisons with other noise schedules (except linear and cosine) or inversion techniques ....***
We acknowledge the value of comparing our Logistic Schedule with other noise schedules and inversion... | Summary: The authors introduce the Logistic Schedule, a novel noise schedule designed to resolve the singularity problem in traditional noise schedules, thereby enhancing inversion stability and reducing noise prediction errors. This new schedule enables more faithful editing that preserves the original content of the ... | Rebuttal 1:
Rebuttal: We appreciate your positive feedback and attention.
**Response to Q1**
> ***Will the extended datasets be open-sourced?***
Yes, certainly. We will open source the extended datasets and our experimental code, including the validations, to facilitate further analysis and research.
---
Rebuttal C... | Rebuttal 1:
Rebuttal: ## **General Response to All Reviewers**
We thank all reviewers for their insightful feedback. We appreciate that the reviewers found our paper theoretically and methodologically novel, with strengths such as
- **thorough analysis** of DDIM inversion issues and **addressing singularity issues ef... | NeurIPS_2024_submissions_huggingface | 2,024 | null | null | null | null | null | null | null | null |
MoMu-Diffusion: On Learning Long-Term Motion-Music Synchronization and Correspondence | Accept (poster) | Summary: The authors proposed a novel framework to address the motion-to-music and music-to-motion tasks. They leveraged aligned latent spaces between motion and music, a multi-modal diffusion transformer, and a cross-guidance sampling strategy. Experiments were conducted to demonstrate that their approach outperforms ... | Rebuttal 1:
Rebuttal: We thank the reviewer for his/her insightful comments and for appreciating the technical contributions of our paper. Here, we respond to your questions as follows:
***1. More specific implementation details***
a) The motion vectors are flattened to $n \in \mathbb{R}^{T_m\times(J\times2)}$ first.... | Summary: MoMu-Diffusion is a motion-music co-generation model with an aim of improved temporal synchronization of the generated motion and music sequence, based on two major components:
1. a bidirectional contrastive rhythmic VAE (BiCoR-VAE) that provides aligned latent space of the motion and music through joint train... | Rebuttal 1:
Rebuttal: We thank the reviewer for his/her insightful comments and for appreciating the technical contributions of our paper. Here, we respond to your questions as follows:
***1. Comparison with other methods on the demo page.***
Thanks for your suggestions. Since the illustrated video samples are pretty... | Summary: The paper propose MoMu-Diffusion, that generates both music-to-motion and motion-to-music videos. MoMu-Diffusion achieves SOTA results and is able to generate joint distribution of music and motion instead of one way.
Strengths: 1. Generate joint distribution of music and motion
2. Propose rhythmic contrastiv... | Rebuttal 1:
Rebuttal: We thank the reviewer for his/her insightful comments and for appreciating the technical contributions of our paper. Here, we respond to your questions as follows:
***1. Ablation study of the hidden size and layers of Transformer***
We appreciate your suggestion on the ablation study of model si... | Summary: This paper proposes a framework that enables the generation of music from motion, motion from music, or both simultaneously while maintaining synchrony between the two. To achieve this, kinematic amplitude is extracted from the motion, and motion audio clip segments corresponding to different kinetic amplitude... | Rebuttal 1:
Rebuttal: We thank the reviewer for his/her insightful comments and for appreciating the technical contributions of our paper. Here, we respond to your questions as follows:
***1. How to handle the cases where the beat in motion or music is not clear and achieve/evaluate natural generation beyond beat alig... | null | NeurIPS_2024_submissions_huggingface | 2,024 | null | null | null | null | null | null | null | null |
Why Do We Need Weight Decay in Modern Deep Learning? | Accept (poster) | Summary: In this work, the role of weight decay (WD) in contemporary deep learning is investigated. The authors note that interpreting weight decay as a classical $L_2$ regularizer seems less relevant for modern deep learning. Instead, it is suggested to separately regard WD from positions of over- and under-training. ... | Rebuttal 1:
Rebuttal: We thank the reviewer for the detailed comments.
1\. The main focus of our work is classification because, at the core of our examination, are the unique properties of exponentially tailed loss functions, such as cross-entropy (CE), as explained in Section 2.1. Therefore, our reasoning does not e... | Summary: This paper investigates the role of weight decay (WD) in modern deep learning, differentiating its effects in over-training and under-training regimes. Through experiments with ResNets on vision tasks (over-training) and Transformers on text data (under-training), the authors demonstrate that WD’s primary bene... | Rebuttal 1:
Rebuttal: We thank the reviewer for their helpful comments. Below, we address the concerns raised by the reviewer in detail
>Although the paper provides a compelling narrative connecting WD's influence to optimization dynamics, it doesn't offer concrete, actionable guidelines for practitioners on how to tu... | Summary: In the paper the authors investigate WD in current deep learning practices. The authors argue that for deep networks used in vision tasks, WD enhances implicit regularisation by modifying the optimisation through loss stabilisation. They also argue that for LLMs, WD helps balance the bias-variance trade-off, l... | Rebuttal 1:
Rebuttal: We thank the reviewer for their comments. We acknowledge the lack of a detailed discussion of the limitations of our work. We did mention the main limitations in the conclusions (no truly large-scale experiments, no proofs of new theoretical results), we will improve upon this by adding a paragrap... | Summary: This paper studies the effect of weight decay (WD) in the over-training and under-training regimes. It hypothesizes and empirically verifies different roles that WD plays in these scenarios. Specifically, in the over-training regime, WD primarily induces the implicit regularization of SGD via controlling the n... | Rebuttal 1:
Rebuttal: We thank the reviewer for their helpful comments, which will allow us to improve the manuscript.
**Improved figures:** We thank the reviewer for suggesting improvements to our figures. We will make sure to reference specific lines in the text to improve the readability of the results.
**Additi... | Rebuttal 1:
Rebuttal: We thank the reviewers for their suggestions and feedback on our manuscript, which will help improve the quality and clarity of our work. We will incorporate your comments into the revised version of the manuscript.
We acknowledge that we did not adequately justify our approximations by citing th... | NeurIPS_2024_submissions_huggingface | 2,024 | Summary: The paper explores the role of weight decay (WD) in training modern deep neural networks, focusing on its impact on optimization dynamics in two distinct training regimes: over-training (multiple passes through the dataset) and under-training (limited passes due to large dataset sizes and computational constra... | Rebuttal 1:
Rebuttal: We appreciate the reviewer's suggestion to enhance the theoretical explanations in our paper. We understand the importance of providing thorough theoretical insights.
- **Conjecture:** We acknowledge that our results regarding implicit regularization are presented as conjectures. This decision wa... | null | null | null | null | null | null |
Stability and Sharper Risk Bounds with Convergence Rate $O(1/n^2)$ | Reject | Summary: This paper studies the generalization measured by gradients via a uniform gradient stability. For $\beta$-uniformly stable algorithms, the paper gives generalization bounds of order $O(1/n+\beta+\sqrt{E_Z[\|\nabla f(A(S);Z)\|_2^2]/n})$, which yields fast rates if $E_Z[\|\nabla f(A(S);Z)\|_2^2$ is small. The pa... | Rebuttal 1:
Rebuttal: Thank you for your time and thoughts. Below, we reply to all the issues.
Question 1: Issue in Weaknesses ``The high-probability analysis based on uniform stability ...''
Answer: Thank you for your valuable feedback and insights. We appreciate the opportunity to address the points raised and pr... | Summary: The work shows high probability excess risk bounds of $O(1/n^2)$ for several algorithms under strong convexity, smoothness, Lipschitz continuity and low noise assumptions using algorithmic stability.
Strengths: 1. The results of the paper are interesting, showing a risk bound of $O(\frac{1}{n^2})$ using algor... | Rebuttal 1:
Rebuttal: Thank you for your time and thoughts. Below, we reply to all the issues.
Question 1: Issue in Strengths ``The problem setup of the paper is not clearly ...''
Answer: Thank you for your valuable feedback. We appreciate your suggestion to collect and detail the assumptions used in our manuscript.... | Summary: This paper achieves the high probability excess risk bounds $\mathcal{O}(1/n^2)$ for empirical risk minimization, projected gradient descent and stochastic gradient descent under strong convexity, smoothness and Lipschitz continuity assumptions.
Strengths: Please see Summary.
Weaknesses: 1.The paragraph from... | Rebuttal 1:
Rebuttal: **PART I**
Thank you for your time and thoughts. Below, we reply to all the issues.
Q1: Issue in Weaknesses ``The paragraph from line 52 to line 65 ...''
A: As Reviewer Xifq pointed out in Strengths, our roadmap in this paper is to provide a finer-grained analysis of the generalization error o... | Summary: The paper considers the standard statistical learning setting and derives $\mathcal{O}(\frac{1}{n ^ {2}})$ ($n$ denotes the number of samples) high-probability bounds for the excess risk $F(A(S)) - \inf_{w} F(w)$ ($A$ denotes the algorithm and $S$ denotes the training set) of ERM, PGD, and SGD. The best-known ... | Rebuttal 1:
Rebuttal: Thanks for your time and thoughts. We will answer all your questions.
Question 1: Issue in Weaknesses ``I found several typos ...''
Answer:
Thank you for your correction. Here are some comments.
(a) ``the stability equation in Lemma 4 should be written ...''.
Comment: Sorry, this is a t... | Rebuttal 1:
Rebuttal: Thank you for the valuable feedback from the reviewers. We will summarize the main difficulties and innovations of the paper here.
The challenges are primarily rooted in three key aspects:
1. **Sharper Generalization via Stability in Gradients**:
The distinction between a vector’s $2$-norm be... | NeurIPS_2024_submissions_huggingface | 2,024 | null | null | null | null | null | null | null | null |
The Poisson Midpoint Method for Langevin Dynamics: Provably Efficient Discretization for Diffusion Models | Accept (poster) | Summary: The paper proposes a novel, multi-step discretization method for SDEs, with particular applications to the Langevin diffusions (overdamped and underdamped) and the denoising diffusion model SDE. The rates obtained are better than any other known guarantees in KL divergence.
Strengths: The discretization propo... | Rebuttal 1:
Rebuttal: We thank the reviewer for the great feedback and multiple interesting possibilities for extension. We will improve the exposition and readability of the paper in the revision based on your comments. We address the questions raised by the reviewer below.
**The main weakness** pointed out by the ... | Summary: In this work, the authors draw inspiration from the randomized midpoint method for Langevin dynamics and design a new method called the poisson midpoint method. They establish guarantees for this method relative to the EM discretisation with thinning, and discuss theoretical and practical applications to sampl... | Rebuttal 1:
Rebuttal: We thank the reviewer for the detailed comments about the paper. We will improve the typos and the inconsistent notations in the revised manuscript. Since the revised version allows for an additional page, we hope to have enough space for a detailed discussion of every aspect pointed out by the re... | Summary: Based on the Randomized Midpoint Method (RMM) proposed for simulating the Langevin Dynamics numerically, the authors proposed a variant called the Poisson Midpoint Method (PMM). Theoretical analysis is performed to bound the deviation between the continuous dynamics and the discretized dynamics. Moreover, when... | Rebuttal 1:
Rebuttal: We thank the reviewer for the great review, which has helped us improve our work. Our responses to the questions and concerns raised are given below.
### The Main Concern:
This is regarding the 14th moment assumption (i.e, Assumption 2) for the initialization. Below we discuss the validity of ou... | null | null | Rebuttal 1:
Rebuttal: We thank the reviewers for valuable feedback and have responded to the key technical concerns in our individual rebuttals. We are committed to enhancing the clarity and comprehensiveness of the paper by incorporating additional details, as suggested.
The reviewers unanimously recognize the novelt... | NeurIPS_2024_submissions_huggingface | 2,024 | null | null | null | null | null | null | null | null |
Minimum Entropy Coupling with Bottleneck | Accept (spotlight) | Summary: The submission formulates and provides an approximate solution for the entropy-bounded information maximization problem. The submission extends Markov coding games to a rate limited setting and illustrates how its entropy-bounded information maximization approach can be to address such settings.
Strengths: Th... | Rebuttal 1:
Rebuttal: >paper … [6] shows that [21]'s algorithm achieves within 𝑙𝑜𝑔2(𝑒)≈1.44 bits of optimal entropy for n variable coupling and 𝑙𝑜𝑔2(𝑒)/𝑒≈0.53 bits of optimal entropy for 2 variable coupling, which are stronger guarantees than that stated above.
We thank the reviewer for pointing this out. ... | Summary: The paper develops a lossy compression framework that uses logarithmic loss (using conditional entropy or mutual information). This framework is designed to handle scenarios where the reconstruction distribution diverges from the source distribution, making it particularly relevant for applications requiring j... | Rebuttal 1:
Rebuttal: >the full MEC-B problem is left open, while special cases of it have been solved
Our approach involves decomposing the full MEC-B problem into two subproblems: EBIM and MEC. By optimizing the encoder and decoder separately, we can reconstruct a solution to the full MEC-B problem. First, we optimi... | Summary: The authors introduce a novel lossy compression framework which introduces a bottleneck to the canonical minimum entropy coupling framework. The authors show that the encoder task can be solved approximately by a novel greedy algorithm with guaranteed performance, while the decoder problem reduced to a canonic... | Rebuttal 1:
Rebuttal: >in equation (1), R is not yet defined
Thank you for pointing this out. We have fixed this and introduced R before presenting equation (1).
>In terms of real-world impact, could the authors further elaborate how their EBIM algorithm would contribute to robust watermarking approaches?
We conside... | null | null | Rebuttal 1:
Rebuttal: We would like to thank the reviewers for their positive evaluation of our paper. We appreciate their thoughtful and constructive feedback. We have addressed their comments in detail in the responses provided below. | NeurIPS_2024_submissions_huggingface | 2,024 | null | null | null | null | null | null | null | null |
DreamScene4D: Dynamic Multi-Object Scene Generation from Monocular Videos | Accept (poster) | Summary: This work introduces DreamScene4D, a pioneering approach for generating realistic 4D scene representations from complex multi-object videos with significant object motions. It proposes a "decompose-recompose" strategy, where a video is first decomposed into individual objects and the background scene. Each com... | Rebuttal 1:
Rebuttal: Thank you for acknowledging that our paper is novel, well-written and effective with comprehensive comparisons. Here we respond to your insightful comments and questions.
> **Q1. This method is composed of multiple steps and requires much human involvement.**
The statement “our method requires m... | Summary: DreamScene4D is a novel approach to generating 3D dynamic scenes of multiple objects from monocular videos using 360° novel view synthesis. The method employs a "decompose-recompose" strategy, segmenting the video scene into background and object tracks and further decomposing object motion into object-centric... | Rebuttal 1:
Rebuttal: Thank you for acknowledging DreamScene4D being a novel approach with comprehensive experimental validation and superiority over existing methods, and for the writing being well-structured. Here we respond to your insightful comments and questions.
> **Q1. How does video scene decomposition separa... | Summary: The paper introduces a novel framework for reconstructing multi-object scenes from monocular videos by factorizing motion into three components: object-centric deformation, object-to-world-frame transformation, and camera motion. This method enhances stability in motion optimization and effectively captures l... | Rebuttal 1:
Rebuttal: Thank you for acknowledging that our paper is novel, effective with solid evidence and well-written. Here we respond to your insightful comments and questions.
> **Q1. In Table 1, the ablation study on $L_{flow}$ does not justify the need for this component.**
We measure the 4D generation perfor... | Summary: 1. The paper claims that their methodology represents the first approach capable of generating verisimilar four-dimensional scene representations derived from complex multi-object video sequences exhibiting substantial motion. This approach purportedly enables 360-degree novel view synthesis and precise motion... | Rebuttal 1:
Rebuttal: Thank you for recognizing DreamScene4D as an innovative approach that addresses the difficulties of multi-object scenarios and achieves significant improvements over current state-of-the-art baselines in view synthesis and motion accuracy. Here we respond to your insightful comments and questions.... | Rebuttal 1:
Rebuttal: We appreciate the positive and insightful comments from all four reviewers. We are glad that they found our paper "novel / innovative", "effective with comprehensive experiments" and "well-written". We will release our code to facilitate future research in fine-grained 4D understanding from videos... | NeurIPS_2024_submissions_huggingface | 2,024 | null | null | null | null | null | null | null | null |
Discrete-state Continuous-time Diffusion for Graph Generation | Accept (poster) | Summary: This paper is an extension of current graph diffusion generation to discrete-state and continuous-time version, and proposes DisCo model. The authors also prove permutation-equivariance/invariance of the proposed diffusion framework, sampling and training loss.
Strengths: 1. Continuous-time diffusion is a gen... | Rebuttal 1:
Rebuttal: Dear reviewer 9BTR, we appreciate your tremendous time and effort in reviewing our paper and are glad that you recognize the proposed method's reasonability and writing quality. Our point-to-point responses to your questions are as follows.
> **Q1.** The selected baselines are not enough. Only Di... | Summary: The paper presents DISCO (Discrete-State Continuous-Time Diffusion), a novel framework for graph generation. DISCO formulates the graph diffusion process in a discrete-state continuous-time setting, preserving the discrete nature of graph data while allowing for flexible sampling trade-offs between quality and... | Rebuttal 1:
Rebuttal: Dear reviewer XoYX, we appreciate your tremendous time and effort in reviewing our paper and are glad that you recognize its presentation, technical soundness, and importance to this research problem. Our point-to-point responses to your questions are as follows.
> **Q1.** The proposed method see... | Summary: This paper categorizes existing graph diffusion models into 4 types according to whether the space of states and time steps are discrete or continuous. The authors design DISCO, the first discrete-state continuous-time graph diffusion generative model. By incorporating continuous-time Markov chains while pres... | Rebuttal 1:
Rebuttal: Dear reviewer ohXc, we thank you for your tremendous time and effort in reviewing our paper, and we are glad that you recognize this paper's clear motivation and theoretical contribution. For your questions, our point-to-point responses are as follows.
> **Q1.** Some symbols in the formulas of th... | Summary: The paper introduces a novel graph generation method based on the continuous-time Monte Carlo Markov Chain by adapting the previous work [7]. The proposed architecture comes in two variants: one utilizing a graph Transformer and another using a regular message-passing neural network. The performance of the mod... | Rebuttal 1:
Rebuttal: Dear reviewer Tpg4, we appreciate your tremendous time and effort in reviewing our paper and are glad that you recognize our proposed model, theoretical contribution, and experiments. For your questions, our point-to-point responses are as follows.
---
> **Q1.** Baseline methods, e.g., GraphINVE... | Rebuttal 1:
Rebuttal: Dear reviewers, we thank you for your time and effort in providing valuable reviews for our paper. We appreciate the reviewers' recognition that our theoretic results support the proposed method (Tpg4 and ohXc), that our experiments are comprehensive (Tpg4), that our paper is clear and easy to und... | NeurIPS_2024_submissions_huggingface | 2,024 | null | null | null | null | null | null | null | null |
Preference Learning Algorithms Do Not Learn Preference Rankings | Accept (poster) | Summary: This paper investigates the effectiveness of preference learning algorithms, include the RLHF (PPO-based) and DPO in training LLMs. In particular, the authors study the non-regularized ranking accuracy, which is largely ignored in the previous RLHF literature. The authors approach this investigation through bo... | Rebuttal 1:
Rebuttal: We thank the reviewer for their detailed review and greatly appreciate their support of the acceptance of our paper. We address the feedback below:
**"First, while it is fair to say that DPO is not good at optimizing the ranking policy in terms of the non-regularized target, it is not immediately... | Summary: The paper investigates the effectiveness of preference learning algorithms, RLHF and DPO, in training LLMs to rank human-preferred outputs higher. It challenges the assumption that these algorithms can successfully teach models to rank outputs according to human preferences. Empirical findings show that most s... | Rebuttal 1:
Rebuttal: We thank the reviewer for recognizing the “clear implications” of our theoretical and empirical analysis. We answer additional questions below.
**"Specifically, the authors could benefit from a deeper analysis of the implications of using ranking accuracy versus win rate as measures of model perf... | Summary: This paper studies the preference learning algorithms like DPO in Large Language Models (LLMs). It focuses on the ranking accuracy and shows that existing algorithms fail to achieve high ranking accuracy. Then, the paper further proposes an idealized ranking accuracy and finds that there is an alignment gap.
... | Rebuttal 1:
Rebuttal: Thank you to the reviewer for your thoughtful comments and feedback. We provide responses to your concerns below:
**"The motivation for studying ranking accuracy is weak.”**
We study ranking accuracy because DPO directly optimizes for ranking accuracy. For example, the authors of the DPO paper s... | Summary: The paper empirically highlights a few potential flaws in RLHF and DPO that prevent preference-tuned models from achieving high ranking accuracy. It presents a collection of empirical and theoretical findings:
+ Existing preference-tuned models achieve low ranking accuracies.
+ The idealized policy and the DPO... | Rebuttal 1:
Rebuttal: We thank the reviewer for their valuable feedback and suggestions. We appreciate your recognition of the thoroughness of our work. We have included our responses to your questions below and new results in the PDF attached to the global rebuttal.
**"One concern is that the ranking accuracy conside... | Rebuttal 1:
Rebuttal: We are grateful to the reviewers for their detailed feedback – Reviewer yhyh notes the paper's "robust evidence supporting its claims," Reviewer 5qJ8 comments that our "findings have clear practical implications for the development of alignment techniques," and Reviewer uW8U notes "the authors ide... | NeurIPS_2024_submissions_huggingface | 2,024 | null | null | null | null | null | null | null | null |
TreeVI: Reparameterizable Tree-structured Variational Inference for Instance-level Correlation Capturing | Accept (poster) | Summary: The paper introduces a new type of variational inference that allows correlations across training samples in the form of a tree, making it suitable for applications with graph-structured data or explicit constraints. The method extends to multiple trees for capturing more complex correlations and includes a me... | Rebuttal 1:
Rebuttal: We thank the reviewer for the comments and suggestions. Below we respond to each of the raised questions.
**Q1: What do the numbers in the nodes represent in Figure 1?**
A1: Sorry for the confusing notations used in the Figure 1. In Figure 1 (b) and \(c), the numbers 1, 2, ... 5 in the trees sh... | Summary: The authors propose to use a variational approximation that captures tree-structured correlations as a middle ground between scalable fully-factorized approximations and more expressive approximations which aim to capture the full correlation structure. They show how these tree structured correlations give ris... | Rebuttal 1:
Rebuttal: We thank the reviewer for the comments and suggestions. Below we respond to each point raised by the reviewer:
**Q1: It is unclear how important the initialization is and if there always exists a straightforward initialization procedure.**
A1: Below we show the constrained clustering accuracies ... | Summary: This paper proposes Tree-structured Variational Inference (TreeVI), a novel method for capturing instance-level correlations in the posterior distribution of latent variables. TreeVI represents correlations between latent variables using a tree structure. This enables reparameterization of latent variables usi... | Rebuttal 1:
Rebuttal: We thank the reviewer for the comments and suggestions. Below we respond to each point raised by the reviewer:
**Q1: The study could be enhanced by providing some insights derived from the tree structures obtained through training.**
A1: Thanks for the suggestions. Our method is mainly used to c... | Summary: The paper proposes to perform amortized variational inference, where the variational approximation has a tree-dependence structure across the instance-level latent variables. For this, the scale is decomposed into variance and correlation, where both are amortized through a neural network. A non-convex constra... | Rebuttal 1:
Rebuttal: We appreciate your detailed comments, but we believe that you misunderstood our paper deeply. We hope our clarifications could help you recognize our contributions correctly.
You mistakenly think that our main contribution is simply to Cholesky decompose the correlation matrix $\mathbf{R}$ as $\m... | Rebuttal 1:
Rebuttal: We thank the reviewers for their careful reading and detailed feedback.
Please refer to the attached PDF for new results. We note that we have provided additional insights about our learnt tree structure obtained through training. In Figure 1, we plot the tree structure learnt over constrained cl... | NeurIPS_2024_submissions_huggingface | 2,024 | null | null | null | null | null | null | null | null |
AWT: Transferring Vision-Language Models via Augmentation, Weighting, and Transportation | Accept (poster) | Summary: This paper introduces an adaptation framework AWT to enhance vision-language models. The authors propose to augment inputs from visual and textual perspectives. The augmented inputs are dynamically reweighted based on the prediction entropy. Finally, the authors propose to use optimal transport to find the bes... | Rebuttal 1:
Rebuttal: We sincerely appreciate your constructive feedback. Please allow us to address your concerns below.
#### **W1: I'd like to see performance-efficiency trade-off under few-shot settings.**
Thank you for your suggestion. We presented this experiment in Figure 2 of our attached PDF. In comparison to... | Summary: The paper considers a task of low-shot (zero-shot and few-shot) adaptation of vision-language models (VLMs). The proposed approach is based on diversifying the inputs of VLMs using augmentations for images and LLMs for class names. Each input is then weighted based on its prediction entropy, while optimal tran... | Rebuttal 1:
Rebuttal: We sincerely appreciate your insightful feedback. Please allow us to address your concerns below.
#### **W1: 1) The dataset description is a slight limitation. 2) Results for prompting without dataset description. 3) The LLM version for AWT, VisDesc [1], and CuPL [2] might differ.**
Thank you fo... | Summary: This paper proposes a novel adaptation framework for vision-language models, which replaces point-to-point alignment of text and images with set-to-set alignment. Specifically, the authors use image transformation and LLM generation to create augmentation sets. After weighting the augmentations based on entrop... | Rebuttal 1:
Rebuttal: We appreciate the time and effort you have dedicated to reviewing our paper. Please allow us to address your concerns below.
#### **W1: Relationship and differences between prior OT methods and AWT.**
Thanks for your comment. For a detailed response, please refer to our responses to common revie... | Summary: This paper proposes AWT, an adaptation framework that can boost pre-trained vision-language models (VLMs) for understanding new concepts on new classes. To summarize, AWT first do augmentations on both visual images and textual class names. Then an entropy metric is employed for weighting multiple augmented da... | Rebuttal 1:
Rebuttal: Thank you for your detailed review and insightful comments! Please allow us to address your concerns below.
#### **W1 and Q1: 1) Random resized cropping and flipping is naïve. 2) Cannot see any improvement by visual augmentations. 3) What about applying more complex visual augmentations?**
Thank... | Rebuttal 1:
Rebuttal: Dear Reviewers,
We sincerely appreciate the time and effort you have invested in reviewing our submission. This paper received 4 review comments, and 3 reviewers gave positive scores. All reviewers acknowledged the strong writing, the excellent benchmark performance, and the extensive experiments... | NeurIPS_2024_submissions_huggingface | 2,024 | null | null | null | null | null | null | null | null |
Linearly Decomposing and Recomposing Vision Transformers for Diverse-Scale Models | Accept (poster) | Summary: This paper presents a novel approach to adapting Vision Transformer (ViT) models for diverse computational environments. The key idea is to linearly decompose a large ViT model into a set of basic components during training, which can then be flexibly recomposed to create smaller, pre-initialized models suitab... | Rebuttal 1:
Rebuttal: To W1 \& Q1: Details of decomposition
Thank you for your question. Due to the parameter gap, the 12-layer model with only one component $A_1$ in the first stage does not perform as well as a typical 12-layer model. However, as more components are trained, the performance becomes comparable to th... | Summary: This paper proposes a novel approach to generate diverse-scale Vision Transformer (ViT) models with varying computational complexity, aiming to address the deployment challenges posed by the fixed architectures of standard ViT models. The key idea involves linearly decomposing a pre-trained ViT model into basi... | Rebuttal 1:
Rebuttal: To 1: Components sharing approach
Thank you for your insightful question. The component sharing in our model is applied within each individual layer, unlike the cross-layer sharing seen in MiniViT. Additionally, the forward process follows the typical ViT architecture, such as DeiT, during traini... | Summary: This paper proposes a novel method to efficiently adapt Vision Transformer (ViT) models to devices with diverse computational resources by linearly decomposing a ViT into basic components that can be flexibly recomposed into ViTs of various scales. The key ideas are:
- Inspired by polynomial decomposition, a ... | Rebuttal 1:
Rebuttal: To 1: Details of recomposition
Thank you for your feedback. In the revision, we will provide a more detailed explanation of the recomposition process and update Fig.3 in the paper to illustrate the initialization of different model architectures and the various training methods.
In Fig.D of the P... | Summary: Inspired by polynomial decomposition in calculus, the authors propose linearly decomposing the ViT model into a set of components during element-wise training. These components can be recomposed into differently scaled, pre-initialized models to meet different computational resource constraints. This decomposi... | Rebuttal 1:
Rebuttal: To 1: Swin Transformers and MobileViT
Thank you for your question, which inspired us to explore the applicability of our method to Swin Transformers and MobileViT.
Swin Transformers consist of four stages, where different stages have transformer layers with different resolutions, resulting in di... | Rebuttal 1:
Rebuttal: We greatly appreciate the insightful and constructive feedback provided by the reviewers. We have responded in detail to the concerns and questions raised, and have attached a PDF with figures and tables. Below is a brief overview of the PDF:
1. Fig. A: Shows the parameters of each layer of the p... | NeurIPS_2024_submissions_huggingface | 2,024 | null | null | null | null | null | null | null | null |
Graph-based Uncertainty Metrics for Long-form Language Model Generations | Accept (spotlight) | Summary: In this work, the authors design graph-based uncertainty metrics for long-form generation. Specifically, they sample multiple responses to a query and decompose them into claims. They then create edges between these responses and claims based on their entailment relationships. Using the constructed graph, they... | Rebuttal 1:
Rebuttal: Thank you for your valuable feedback and suggestions. We appreciate that you acknowledged our method “novel” and our evaluation “thorough”. We would like to address your remaining questions and concerns in the following response.
### Computational cost
> *”Both the uncertainty estimation method... | Summary: To achieve accurate and fine-grained uncertainty estimation for long-form generation of large language models (LLMs). the authors propose a new framework to measure the uncertainty by building a bipartite graph between the randomly sampled responses from LLMs and the set of the claims within these responses fi... | Rebuttal 1:
Rebuttal: Thank you for your valuable feedback and suggestions. We appreciate that you acknowledged that our method approaches an important problem and offers satisfactory results. We would like to address your remaining questions and concerns in the following response. **Due to the space limitation of the ... | Summary: This paper introduces Graph Uncertainty, a novel framework for estimating granular, claim-level uncertainty in long-form LLM outputs using semantic entailment graphs and graph centrality metrics. The authors claim the proposed method can go beyond existing works that can only be applied to multiple-choice ques... | Rebuttal 1:
Rebuttal: Thank you for your valuable feedback and suggestions. We appreciate that you acknowledge our work is “novel and addresses a critical gap in LLM uncertainty estimation”. We would like to address your remaining questions and concerns in the following response.
### Computational cost
> *”The propose... | null | null | Rebuttal 1:
Rebuttal: We thank all reviewers for their helpful feedback and suggestions.
We are glad that the reviewers found our work our method approaches an important problem [29kR], proposes a novel and interesting method [29kR, 2gkm], offers thorough empirical results [29kR, 2gkm], and “addresses a critical gap ... | NeurIPS_2024_submissions_huggingface | 2,024 | null | null | null | null | null | null | null | null |
Physics-Regularized Multi-Modal Image Assimilation for Brain Tumor Localization | Accept (poster) | Summary: The authors proposed a discrete physic-based loss function which can be minimized to infer the spatial distribution of tumor cell distribution, which can not be assessed by imaging data only. The loss maps imaging data to a tumor tissue biomechanical model to solve inverse problems and to better understand ind... | Rebuttal 1:
Rebuttal: **Thank you for your thoughtful feedback and diligent comment regarding the numbers in [16]. Please note our global response comment with additional results and explanations.**
**Q1**: *Could the authors highlight the differences between [16] and their paper? Does addition of the tissue deformat... | Summary: The authors propose a method using soft physics-based regularization to predict the distribution of brain tumor. The main contribution is a novel discretization scheme of physics equations to model brain tumor.
Strengths: 1. This work is an improvement for physics-based approaches for medical imaging
2. The... | Rebuttal 1:
Rebuttal: **Thank you for your thoughtful feedback regarding the ground truth labels and limited amount of available data. Please note our global response comment with additional results and explanations. Below we address your specific questions.**
**Weakness**: *Lack of ground truth data to train and eval... | Summary: This paper presents a new approach to integrate physics-based tumor growth constraint with multi-modal imaging data to predict tumor cell distribution and thus enhance tumor treatment planning for glioblastomas. The core of the methods includes a discrete physics residual and initial assumptions encoding initi... | Rebuttal 1:
Rebuttal: **Thank you for your thoughtful feedback regarding the dataset size and lack of statistical tests, despite relatively high uncertainties. We appreciate acknowledging the challenges of the task. Please note our global response comment with additional results and explanations.**
**Q1:** *Clarifica... | Summary: This paper presents a method for brain tumor modeling through a joint data-driven and physics-based approach. The proposed approach leverages a multi-task loss that incorporates physical assumptions and prior healthy anatomy to more faithfully model the tumor cell distribution. This method outperforms existing... | Rebuttal 1:
Rebuttal: **Thank you for your thoughtful feedback regarding the recurrence coverage. Please note our global response comment with additional results and explanations. Below, we address your specific questions.**
**Q1**: *According to line 207, recurrence coverage is the “percentage of the tumor segmentat... | Rebuttal 1:
Rebuttal: Dear Reviewers,
**Please see the one-page PDF with additional ablation studies, synthetic data results (with ground truth), and statistical tests for the recurrence coverage results.**
We thank all reviewers for their constructive feedback that helped us improve the paper. We are encouraged by ... | NeurIPS_2024_submissions_huggingface | 2,024 | null | null | null | null | null | null | null | null |
Collaboration! Towards Robust Neural Methods for Routing Problems | Accept (poster) | Summary: This paper focuses on the robustness of neural combinatorial optimization and proposes a new framework to robustify neural models for routing problems. The core idea is to train a set of models and to let them collectively make the best predictions. The authors have conducted experiments to benchmark the propo... | Rebuttal 1:
Rebuttal: We genuinely appreciate the effort the reviewer dedicated to evaluating our paper and providing insightful feedback! Here are our detailed responses to your comments, where `W` denotes Weakness (W1-W4) and `Q` denotes Question (Q1-Q2).
**W1: An experiment on independent AT.**
Actually, the basel... | Summary: This paper presents an ensemble-based Collaborative Neural Framework for for vehicle routing problems to improve robustness. Multiple models are adversarially trained for better robustness against attacks and improve generalization, and a neural router routes instances to models for better load balancing and e... | Rebuttal 1:
Rebuttal: We are grateful to the reviewer for dedicating your precious time to review our paper and offer feedback. Here are our detailed responses to your comments, where `W` denotes Weakness and `Q` denotes Question (Q1-Q6).
**W: The paper is not organized or written well; This paper is not technically s... | Summary: This paper develops a ensemble-based method CNF for adversarial training to enhance the robustness of NCO solvers. CNF demonstrates outstanding performance compared to all current adversarial training frameworks.
Strengths: 1. The effectiveness of CNF is outstanding.
2. The motivation, statement, and objectiv... | Rebuttal 1:
Rebuttal: We greatly appreciate the reviewer's time and effort in reviewing our paper and providing constructive comments. Here are our detailed responses to your review, where `W` denotes Weakness and `Q` denotes Question (Q1-Q5).
**W: The trade-off demonstrated in Figure is not clear; Figures 2 and Algor... | Summary: The paper proposes a Collaborative Neural Framework (CNF) to improve the robustness of neural methods for vehicle routing problems (VRPs) against adversarial attacks. CNF introduces two key innovations: first, a global attack strategy that generates diverse adversarial instances by attacking the best-performin... | Rebuttal 1:
Rebuttal: We sincerely appreciate the reviewer for spending valuable time reviewing our paper and providing positive comments. Here are our detailed responses to your comments, where `W` denotes Weakness.
**W: Theoretical analysis or guarantees.**
Thanks for your comment. The majority of the work within t... | Rebuttal 1:
Rebuttal: ## **General Response**
We extend our heartfelt gratitude to all reviewers for their valuable comments. We are pleased to see that the reviewers have recognized our research topic is **crucial** (`DziC`), our method is **novel** (`d2Cb`) and **effective** (`AuRq, d2Cb`), our paper is **well-writt... | NeurIPS_2024_submissions_huggingface | 2,024 | null | null | null | null | null | null | null | null |
CRONOS: Enhancing Deep Learning with Scalable GPU Accelerated Convex Neural Networks | Accept (poster) | Summary: - This paper builds on work on reframing the optimization of 2-layer ReLU networks as convex optimization
- The first contribution of this work is the use of the ADMM algorithm for solving the constrained optimization task from equation 4
- Noting that need of repeatedly solve the inner problem in the ADMM met... | Rebuttal 1:
Rebuttal: **Weaknesses**
1. This is an excellent point that was unclear in the original submission. We have included figures for CIFAR-10 and ImageNet in the rebuttal document which include D-Adapted Adam (DAdam) in the comparison. The difference in performance between CRONOS-AM and DAdam are visible on t... | Summary: The authors introduce a reformulation of optimizing a two-layer neural network (one neuron in the second layer) to a convex optimization problem and suggest an algorithm to solve it.
This is done via reformulating a convex program based on sampled P activation patterns from the set of all possible ReLU activa... | Rebuttal 1:
Rebuttal: **Weaknesses**
1. Thank you for raising the point! We understand that the paper covers a wide range and utilizes ideas from diverse subjects such as deep learning, convex optimization, randomized numerical linear algebra, and high-performance computing. Unfortunately given the page constraints, ... | Summary: This work investigated the optimization of convex reformulated neural networks. The authors proposed to solve the sub-sampled convex optimization problem (for a two-layer ReLU network) with operator splitting (using ADMM), and used conjugate gradient method with Nystrom preconditioning to solve one subproblem ... | Rebuttal 1:
Rebuttal: **Weaknesses**
1.) This is an excellent point! It is important to have guidance on how to select the hyperparameter $P$ in practice.
Fortunately, large $P$ is not necessary to ensure the optimal values are comparable.
The recent work [1] (see Theorem 2.1) on convex neural networks has shown th... | Summary: In the paper, two neural network optimization methods, CRONOS and CRONOS-AM, are proposed. The authors provide theoretical complexity analysis and convergence proof of the algorithms, and perform experiments on large datasets of image and language tasks to verify the effectiveness of the algorithm. As can be s... | Rebuttal 1:
Rebuttal: ** Weaknesses **
1. Thank you for pointing this out! We have fixed this in the revision.
**Questions**
We apologize for the typo and have corrected this for IMDB-DA. Thank you for raising the question! The Domain Adaptation experiments were aimed at minimizing the distribution shift between sour... | Rebuttal 1:
Rebuttal: We would like to sincerely thank all the reviewers for their thoughtful feedback and suggestions, which will greatly improve the submission.
In addition to our point-by-point response to each reviewer, we wish to underscore our contributions, describe the content of the rebuttal pdf, and address... | NeurIPS_2024_submissions_huggingface | 2,024 | null | null | null | null | null | null | null | null |
realSEUDO for real-time calcium imaging analysis | Accept (poster) | Summary: The authors extend an available algorithm called SEUDO for extracting time traces from two-photon imaging data robustly to the real-time close-loop setting. The algorithm can simultaneously extract ROIs and time traces. The algorithm performs on par with state-of-the-art offline algorithms and better than stat... | Rebuttal 1:
Rebuttal: **Weaknesses:**
1. **The main improvement come from engineering the implementation of SEUDO such as to run in close to real time (is that a weakness or strength?).**
**Our response:** Thank you, and we believe that our improvements represent a strength of our approach as we are able to successful... | Summary: In this study, the authors created a real-time version of the “Sparse Emulation of Unused Dictionary Objects” (SEUDO) algorithm termed realSEUDO, that can process multi-photon calcium imaging (CI) data in real time. The goal of this approach is to overcome the limitations of batch processing methods commonly e... | Rebuttal 1:
Rebuttal: **Weaknesses: While the performance with a small number of neurons is impressive there are two drawbacks. First, the testing datasets only include a very limited number of neurons and the scalability with much larger datasets are not presented. For example, current methods such as with a mesoscope... | Summary: The authors present an online method for neuron profile and trace estimation in multi-photon calcium imaging. Their approach enhances the SEUDO algorithm by running it on a frame-by-frame basis and periodically updating profiles using a patching mechanism across space and time. The implementation leverages dif... | Rebuttal 1:
Rebuttal: **Weaknesses:**
1. **The work largely relies on previously described algorithms like FISTA and SEUDO. Since a significant portion of the method describes these algorithms in detail, it is unclear what the key contributions of this work are.**
**Our response:** While we leveraged prior development... | Summary: The paper presents a novel online method for cell detection and fluorescence estimation from streaming calcium imaging (CI) data, called realSEUDO. Building upon the SEUDO algorithm, realSEUDO introduces significant advancements in real-time cell identification and fluorescence estimation. The authors effectiv... | Rebuttal 1:
Rebuttal: **Weaknesses: While the paper presents a novel and efficient approach with realSEUDO, several limitations are apparent. The reliance on specific hardware configurations, such as Linux CPU performance modes, could limit accessibility and generalizability. Additionally, the implementation primarily ... | Rebuttal 1:
Rebuttal: We thank the reviewers for their time and effort in reviewing our manuscript. We are encouraged that the reviewers recognized the relevance, capabilities, and potential impact of our work. We take the time here to respond to the primary concerns the reviewers raised, with more detailed responses t... | NeurIPS_2024_submissions_huggingface | 2,024 | null | null | null | null | null | null | null | null |
GenRec: Unifying Video Generation and Recognition with Diffusion Models | Accept (poster) | Summary: This paper explores the use of video diffusion models for joint generation and recognition. The authors introduce a framework called GenRec, based on Stable Video Diffusion, that aims to unify video generation and recognition. GenRec employs a random-frame conditioning process to learn generalized spatial-temp... | Rebuttal 1:
Rebuttal: Thank you for your valuable comments. We will address your concerns point by point.
**Q1: Provide more details on the derivation of Eq.14-16.**
The derivation of Eqs. 14-16 is similar to Eq. 12 in **[A]**. Here are the details.
Firstly, Eq. 13 in our paper shows the formulation of the unconditi... | Summary: The paper proposes to unify video recognition and generation with a masked finetuning of SVD (diffusion video model). By training with both classification and generation objectives, the model learns to do both. Rather than training with first frame only for conditioning, the masking is employed so that the cla... | Rebuttal 1:
Rebuttal: Thanks for the positive feedback! We address your questions below.
**Q1: SOTA selection and completion.**
We greatly appreciate your suggestion. We agree that including more comparable methods is very meaningful. **We have updated the Table 1 in the uploaded PDF** to include more comparable meth... | Summary: This paper investigates to unify the video generation and recognition tasks into the same diffusion model. The authors propose a conditional feature mask mechanism to unify the two learning tasks. Experiments are conducted on several tasks to validate the effectiveness of the proposed method.
Strengths: The i... | Rebuttal 1:
Rebuttal: Thank you for your valuable comments. We will address your concerns point by point.
**Q1: Why the loss weight of the recognition loss is not related with the noise level?**
Thanks for your valuable feedback. Adding loss weighting related to noise is standard for diffusion training, but is unexpl... | Summary: Recently, diffusion models have been released that can generate high quality video clips, indicating that they strongly capture natural appearance and motion in video. This paper seeks to adapt one recent model, Stable Video Diffusion (SVD), to explore if what it has learned can be leveraged to improve video r... | Rebuttal 1:
Rebuttal: Thanks for the positive feedback! We address your questions below.
**Q1:SoTA results completion.**
Thank you for your valuable feedback. We have updated the Table 1 in the **uploaded PDF**, which includes more comparisons with methods like InternVid, InternVid2, OmniVec and OmniVec2. In addition... | Rebuttal 1:
Rebuttal: We thank the reviewers for their detailed, thoughtful and valuable feedback. We are grateful that the reviewers identified our work unifies the video generation and video recognition (Reviewer caZh), provides "exciting result" (Reviewer PRtT), are interesting and necessary for the computer vision ... | NeurIPS_2024_submissions_huggingface | 2,024 | Summary: This work proposes a video diffusion model to unifies the video generation and video recognition tasks. The proposed GenRec, during training, jointly optimize the generation objective and the classification objective. During inference, GenRec can deal with both the video generation conditioned on frames or cla... | Rebuttal 1:
Rebuttal: Thank you for your valuable comments. We will address your concerns point by point.
**Q1: Concern about the novelty compared with [8,37,46].**
Thanks. Previous works [8, 37, 46] applied diffusion models to understanding tasks, but there are some limitation that mainly differs from our paper:
- ... | null | null | null | null | null | null |
Can Learned Optimization Make Reinforcement Learning Less Difficult? | Accept (spotlight) | Summary: This paper presents OPEN, a meta-learning approach with a learned optimiser. The approach is based on addressing three challenges for RL: 1) the non-stationary nature of input and output distributions; 2) the loss of plasticity; and 3) exploration. The approach performs better than a traditional Adam optimise... | Rebuttal 1:
Rebuttal: Dear reviewer qjkS,
Thank you for your in-depth review and recommendation of acceptance. We greatly appreciate that **you find our approach performant, scalable and novel**. We address your concerns below.
# Lack of Empirical Detail
>The techniques used for the empirical comparisons are not des... | Summary: This paper presents OPEN, a meta-learned optimizer for reinforcement learning gradient updates. The update rule for OPEN is split into 3 stages, one for each difficulty in RL they define: non-statiarity, plasticity loss and exploration. They train their optimizer with an evolutionary strategy. To analyze their... | Rebuttal 1:
Rebuttal: Dear reviewer KVk8,
Thank you so much for your very positive review and recommendation of strong acceptance to the conference. While you have **highlighted many positives of our work, such as the quality of our writing, experiments and ablations**, you do still discuss some weaknesses of our pape... | Summary: This paper explores whether learned optimization can address specific challenges in reinforcement learning (RL), such as non-stationarity, plasticity loss, and the need for exploration. The authors propose an optimizer, OPEN, which meta-learns update rules related to these issues. OPEN demonstrates strong perf... | Rebuttal 1:
Rebuttal: Dear Fr37,
Thank you very much for your review. We are glad that you appreciate our **clear writing and strong evaluation** throughout the paper, and were happy to see this reflected in your recommendation of acceptance. Below, we detail answers to your queries.
# Insufficiency of Dormancy
> Th... | Summary: This paper presents OPEN, which meta-learns an update rule (a learned optimizer) for reinforcement learning. The authors test in single-task and multi-task settings, and show good generalization to tasks outside of the training set.
Strengths: This is a good paper on the application of learned optimization to... | Rebuttal 1:
Rebuttal: Dear sKXo,
We would like to thank you for your positive review of our paper. In particular, we are grateful that you appreciate our **contribution based on conditioning directly on dormancy and other theoretically-grounded inputs** and recognize that this opens up huge potential opportunities in ... | Rebuttal 1:
Rebuttal: Dear reviewers,
# Thank you
We are very grateful to have received four high quality reviews, and appreciate the time and effort you all spent with our paper. We are glad that there was consistency among reviewers that our paper was easily understood and presented strong results, and that **every... | NeurIPS_2024_submissions_huggingface | 2,024 | null | null | null | null | null | null | null | null |
Robust Sparse Regression with Non-Isotropic Designs | Accept (poster) | Summary: The authors provide a computationally efficient estimator for robust and sparse linear regression under non-isotropic covariance matrices. Their first result achieves $O(\sqrt{\epsilon})$ error with state-of-the-art sample complexity under a weaker noise assumption than prior work. Their second result is the f... | Rebuttal 1:
Rebuttal: Dear Reviewer 3A8o,
Thank you very much for your review! We appreciate your evaluation of our work.
Regarding your question: We believe that moment assumptions without sum-of-squares are not enough. For robust mean estimation, similar lower bound was shown in [HL19]. However, we believe that it... | Summary: This paper studies sparse linear regression $y^* = X^* \beta^* + \eta$ (where $\beta^*$ is $k$-sparse) in the presence of two types of adversaries.
- First, the noise vector $\eta$ is sampled before $X^*$ in an (obliviously) adversarial way. Then $X^*$ is sampled independently of $\eta$ with i.i.d. rows and ... | Rebuttal 1:
Rebuttal: Dear Reviewer e1Aa
Thank you very much for your review! We will address all the typos and enhance the writing and notation as per your suggestions.
1) *“The sample complexity bounds are written in a wacky way. Why don't we fix the goal to be achieving $\varepsilon$ error and write the sample com... | Summary: It developed efficient estimators for sparse linear regression in the presence of oblivious and adaptive adversaries. It presents a robust algorithm that outperforms sota under desired conditions on the moment of the distribution.
Strengths: - The paper introduces several robust algorithms that outperform the... | Rebuttal 1:
Rebuttal: Dear Reviewer mVbA,
Thank you very much for your review! Our focus was on proving the strongest possible theoretical guarantees in polynomial time. There is extensive literature in learning theory, particularly in robust estimation and linear regression, that lacks empirical validation, and we be... | Summary: The paper focuses on robust estimation of sparse linear regression in the presence of both oblivious and adaptive adversaries. It claims to offer polynomial-time algorithms capable of recovering a sparse coefficient vector with high probability. The paper includes theoretical analyses using the standard techni... | Rebuttal 1:
Rebuttal: Dear Reviewer sWxz,
Thank you very much for your review! We appreciate your feedback and would like to address the points raised, as well as answer the questions and discuss limitations.
**Regarding the weaknesses:**
*Existing literature on the adaptive adversary:* While [PJL20] inspired the id... | null | NeurIPS_2024_submissions_huggingface | 2,024 | Summary: This paper studies sparse linear regression with oblivious and adaptive adversaries. The design matrix is random, the noise is chosen by an adversary, and the goal is to find the optimal k-sparse weight vector. The results are as follows: to achieve $O(\sqrt{\varepsilon})$ error, a sample complexity of $\widet... | Rebuttal 1:
Rebuttal: Dear Reviewer JPFs,
Thank you very much for your review! We appreciate your evaluation of our work and will improve the techniques section in the final version according to your suggestions.
Regarding your question: Indeed, it is a typo, there should be $\beta^*$ instead of $\hat{\beta}$ in line... | null | null | null | null | null | null |
VidMan: Exploiting Implicit Dynamics from Video Diffusion Model for Effective Robot Manipulation | Accept (poster) | Summary: The submission explores video pre-training for manipulation. Specifically, it trains a language-conditioned video prediction model to imagine future frames for the given task and an inverse dynamics diffusion model to predict actions, conditioned on the imagined future. The paper’s idea is that the first model... | Rebuttal 1:
Title: Response to Reviewer Yu5P [1/3]
Comment: Dear Reviewer,
Thanks for your constructive comments. Below, we would like to address all the weaknesses and questions in details.
**Q1:** `... by changing the number of cameras used ... the number of seeds used ...
[ACT3D, RVT] achieved big leaps in 2023 an... | Summary: This paper introduces a novel framework that utilizes a pre-trained video diffusion model to learn dynamic information and then adapt it to output actions for manipulation tasks by incorporating an action-aware adapter. In the second stage, this framework can transfer the learned dynamic knowledge for action l... | Rebuttal 1:
Title: Response to Reviewer 5fDE [1/2]
Comment: Dear Reviewer,
Thanks for your constructive comments. Below, we would like to address all the weaknesses and questions in details.
**Q1:** `Ablate the choices of adaptation, such as an output projector.`
**A1:** In our approach, we've adopted layer-wise ada... | Summary: This paper proposes a framework for video diffusion in robot manipulation (VidMan), which contains two stages. In the first stage, VidMan adopts large-scale Open-X datasets for pertaining an Open-Sora-like architecture for video prediction. In the second stage, an adapter is introduced to learn the inverse dyn... | Rebuttal 1:
Title: Response to Reviewer NqNP [1/2]
Comment: Dear Reviewer,
Thanks for your constructive comments. Below, we would like to address all the weaknesses and questions in details.
**Q1:** `Similar with diffusion-diffuser.`
**A1:** For trajectory diffusion for planning and inverse dynamics for action predi... | Summary: This paper presents a method called Video Diffusion for Robot Manipulation (VidMan) that uses a two-stage training mechanism. In the first stage, the VidMan model is trained to perform future state prediction (i.e. dynamics learning) on a large diverse dataset (Open X-Embodiment). In the second stage, the mode... | Rebuttal 1:
Title: Response to Reviewer 11a6 [1/2]
Comment: Dear Reviewer,
Thanks for your constructive comments. Below, we would like to address all the weaknesses and questions in details.
**Q1:** `In-depth analyses for the effectiveness of the adopted two-stage strategy.`
**A1:** We greatly appreciate the thought... | Rebuttal 1:
Rebuttal: We extend our heartfelt thanks to the reviewers for their time, thoughtful suggestions, and invaluable feedback. We are honored by the positive recognition from the reviewers regarding the technical contribution (all reviewers), method insight (Reviewer yiak, 11a6, NqNP), thorough ablation (Revewe... | NeurIPS_2024_submissions_huggingface | 2,024 | Summary: This paper introduces VidMan, a two-stage framework for robot manipulation using a pre-trained video diffusion model. The first stage involves training on an actionless video dataset to predict future visual trajectories using denoising diffusion. The second stage adapts this model into an inverse dynamics mod... | Rebuttal 1:
Title: Response to Reviewer yiak [1/3]
Comment: Dear Reviewer,
Thanks for your constructive comments. Below, we would like to address all the weaknesses and questions in details.
**Q1:** ` Compare with more related SOTA methods.`
**A1:** Thank you for your insightful suggestions. In response to the revie... | null | null | null | null | null | null |
A Full-duplex Speech Dialogue Scheme Based On Large Language Model | Accept (poster) | Summary: This paper provides a new approach for using language model assistants, specifically enabling real-time interactions with the language model in a full-duplex setting.
In detail, the authors achieved this by additionally training the Llama-3-Instruct model and integrating it with ASR and TTS modules, creating ... | Rebuttal 1:
Rebuttal: We thank the reviewer for their review and positive feedback. We are glad the reviewer like our demo.
> **Performance Degradation.**
In this experiment, we used only the full-duplex conversation data for SFT, which is why other capabilities were compromised. However, if full-duplex conversation... | Summary: This work presents a generative dialogue system capable of operating in a full-duplex manner. The model is based on Large Language Model with the same next token prediction loss, with adapted input sequence to account for possible speaker turn or interruption. This model is designed to work with external ASR a... | Rebuttal 1:
Rebuttal: Thank you for the thoughtful review of our work! Please allow us to address your concerns and answer the questions.
> The proposed method is purely based on text tokens...
Please refer to the overall comments.
> The evaluation in this work is limited...
In the system prompt we used during our... | Summary: The work introduces an engineering effort which makes the generative system capable of speak and listen without waiting for the end of the entire sequence.
The main contribution is combining one LLM with instruction tuning and a careful prompt with two ASR and TTS models.
The system demonstrates significant i... | Rebuttal 1:
Rebuttal: We appreciate that the reviewer understands and recognizes the contributions of this work. We address the main concerns as follows.
> W1. It's not clear to me...
Please refer to the overall comments.
> W2. The evaluation metrics...
In practical scenarios, indeed, more issues should be conside... | null | null | Rebuttal 1:
Rebuttal: We thank the reviewers for their careful reading, valuable comments, kind words and recognition of the contributions. We first answer a question raised by multiple reviewers.
> One big unified model instead of integrating with some external ASR/TTS models
Training an end-to-end unified multimoda... | NeurIPS_2024_submissions_huggingface | 2,024 | null | null | null | null | null | null | null | null |
Mitigating Partial Observability in Sequential Decision Processes via the Lambda Discrepancy | Accept (poster) | Summary: The authors propose the lambda discrepancy, a measure of partial observability based on TD-lambda. Importantly, this metric can be estimated from samples of POMDP observations (and a bootstrapped value function estimator) without relying on a transition model, access to POMDP states, or even knowledge of what ... | Rebuttal 1:
Rebuttal: Thank you for the kind words and feedback. We're glad you had such a positive impression of the idea and see potential for very high impact. You cited clarity as your biggest concern, so we'll address that here and put our other comments in the global response.
**Markovian Representation**
We ap... | Summary: This paper studies policy planning in a partially observable Markov decision process (POMDP). It has been acknowledged that planning in a POMDP is computationally challenging in general. This paper attempts to address this challenge by proposing novel planning heuristics First, the authors formalize the $\lamb... | Rebuttal 1:
Rebuttal: Thank you for the feedback and questions.
We will group two of your comments together, and respond to the remainder in order.
> *As for Theorems 1 and 2, do the policies π have to be Markov policies only taking the current state S as input? If so, the underlying environment is expected to be a b... | Summary: The authors consider a problem of reinforcement learning in partially observable systems. Instead of constructing an information state using the entire trajectory history with some peripheral objective, the authors propose using discrepancy in the values of TD$(\lambda)$ estimates for different values of $\lam... | Rebuttal 1:
Rebuttal: Thank you for your review and feedback. It seems that your main concern is the misconception that LD “requires operating over the entire state, observation and action space for its computation.” We apologize if this was confusing. While the *derivation* of the metric uses tensors that operate over... | Summary: The Authors address the problem of missingness or partial observability in Markov Decision Problems (MDPs), where the Markov assumption does always hold. They introduce a measure, the Lambda Discrepancy (LD), which serves as an indicator for how closely a Markovian state representation is given.
The LD learns... | Rebuttal 1:
Rebuttal: Thank you for your response and feedback! We are glad to hear you thought the writing was “clear and to the point.”
We will respond to your comments in order.
---
> *The experiments also use a recurrent PPO and compare it to recurrent PPO with lambda discrepancy but it is not explained how they... | Rebuttal 1:
Rebuttal: Thank you all for your feedback and suggestions. Your reviews helped us improve the quality of our submission.
We've made the following requested changes (**see PDF**):
1. Experiment varying “Markovness” of environment (new Fig 1)
2. Visual comparison of learned memories for PacMan (new Fig 2, ri... | NeurIPS_2024_submissions_huggingface | 2,024 | null | null | null | null | null | null | null | null |
Can Custom Models Learn In-Context? An Exploration of Hybrid Architecture Performance on In-Context Learning Tasks | Reject | Summary: The authors explore a number of different attention-based architectures along with Mamba on a series of in-context regression tasks. The architectures vary in their choice of normalization, positional encodings, activations, and hybridization with Mamba. They discover some varying capacities for the different ... | Rebuttal 1:
Title: Author Response
Comment: We thank you for your thoughtful feedback and helpful comments. We briefly respond to some of the above questions and apologize for the otherwise low quality submission.
The reviewer above highlights the limited scientific findings of the paper and proposes some avenues for ... | Summary: The authors build on the ICL work of [1] and [2], wherein networks are "trained to in-context learn" several tasks of varying complexity (e.g., Linear Regression, Sparse Linear Regression, Vector MQAR, etc.). Carrying on from [2], the main contribution of the presented work is the combination of various permu... | Rebuttal 1:
Title: Author Response
Comment: We appreciate your diligent feedback and practical critique. We attempt to clarify some details missed by the submission below.
We trained all checkpoints used in our analysis from scratch on synthetic data generated by sampling parametrized functions and x values from the d... | Summary: This work presents a codebase for benchmarking the in-context learning ability of language models, especially for hybrid models. In addition, several empirical results are presented to show that some model architectures fail entirely or have suboptimal performance on specific in-context learning tasks.
Streng... | Rebuttal 1:
Title: Author Response
Comment: We thank you for the earnest review and encouraging feedback. We otherwise apologize for the unpolished writing.
The reviewer above identifies the primary contribution of this paper to be the associated code but finds that the empirical analysis is lacking systematism, motiv... | Summary: This work presents an analysis of in-context learning (ICL) for a variety of hybrid architectures (composed of different blocks from preexisting large language model architectures) on different regression tasks. The experiments are built on top of a couple of prior works [1, 2] that also explored ICL in simila... | Rebuttal 1:
Title: Author Response (part 1/2)
Comment: We are grateful for your detailed feedback and thorough critique of the submission. We provide a clarification to questions about motivating the submission in this response and reserve responses to specific questions for another response. We apologize for the poor ... | null | NeurIPS_2024_submissions_huggingface | 2,024 | null | null | null | null | null | null | null | null |
Why Not Transform Chat Large Language Models to Non-English? | Reject | Summary: This paper introduces the TransLLM framework to transform English-centric chat LLMs to non-English languages, addressing the challenges of transferring advanced abilities without supervised data and preventing catastrophic forgetting of original knowledge. Key contributions include using the Translation Chain-... | Rebuttal 1:
Rebuttal: Thank you for the positive comments and constructive suggestions.
> W1: The experiments are primarily conducted on transforming LLaMA-2-chat-7B to Thai, which may limit the generalizability of the findings to other non-English languages and other models...
Please refer to the responses for all r... | Summary: This paper proposes a transfer pipeline, TransLLM, which uses a translation chain-of-thought (TCOT) to adapt English-centric large language models (LLMs) to low-resource languages. This pipeline consists of pre-training and supervised fine-tuning (SFT) phases. During the pre-training stage, the authors select ... | Rebuttal 1:
Rebuttal: Thank you for the positive comments and constructive suggestions.
> W1: Flexibility of Methodology
1. Transfer requires resources: Knowledge transfer across languages relies on semantic alignment, which inevitably needs some translation resources. As we discuss in the related works, most existin... | Summary: In this work, the authors present a method for transforming a chat-based English LLM to Non-English (Thai is the only language experimented with) based on a series of steps that teach the LLM to take in a non-English query and respond in non-English for that query. The methods presented is referred to as Trans... | Rebuttal 1:
Rebuttal: Thank you for your time and comments.
> W1: The experiments are done on only one LLM and on only one language (Thai). This severely constrains the extend to which the results could be verified.
Please refer to the responses for all reviewers.
> W2: The novelty of the proposed method is very thi... | Summary: This paper focuses on a scenario as transforming a English-centric **chat** large language model to a non-English chat large language model (or not just en-centric). The authors want to address the catastrophic forgetting problem where further tuning on the original En-chat LLM without reusing their original S... | Rebuttal 1:
Rebuttal: Thank you for your time and comments.
> W1: To be honest, I am not sure whether this paper has good starting point. They focus on a scenario where you want to further fine-tune the chat model in English to a chat model in other languages. Could you provide more concrete application scenarios in y... | Rebuttal 1:
Rebuttal: We thank all reviewers for your time and contributions.
> Generalizability
Most reviewers express concerns about the generalizability of TransLLM, given that our experiments have only involved transforming LLaMA-2-chat-7B to Thai. We would like to address these concerns in the following aspects:... | NeurIPS_2024_submissions_huggingface | 2,024 | null | null | null | null | null | null | null | null |
Learning Neural Contracting Dynamics: Extended Linearization and Global Guarantees | Accept (poster) | Summary: The paper proposes a novel neural network architecture, Extended Linearized Contracting Dynamics (ELCD), for learning dynamical systems with global contractivity guarantees. The authors leverage the concept of extended linearization to factorize the vector field and enforce negative definiteness of its symmetr... | Rebuttal 1:
Rebuttal: **It is unclear why the joint training of the encoder and the model is possible for ELCD but not in NCDS.**
It is possible to jointly train the encoder and the model for NCDS as well. The authors of NCDS treat the encoder as a Variational Autoencoder (VAE) and train it by minimizing the evidence ... | Summary: The paper presents Extended Linearized Contracting Dynamics (ELCD), a dynamical system with neural network components that has global contraction guarantees. They demonstrate improved efficiency and performance on the trajectory fitting LASA dataset (upto 8 dim), pendulum dataset (upto 16 dim), Rosenbrock data... | Rebuttal 1:
Rebuttal: First, we would like to thank you for finding our work to have a strong theoretical basis, well-written,
and easy-to-read with reasonable experiments. Regarding larger-scale experiments, we have been unable to
perform them at this time. We agree that they would add value and we envision executing ... | Summary: The paper presents a method for learning contracting representations from a dynamical system. The novelty of the method lies in learning a linear map and a coordinate transform map (Diffeomorphism), which extends the types of contracting systems that can be learned. The experiments show that the proposed metho... | Rebuttal 1:
Rebuttal: **The novelty of the method lies in learning a linear map...**
Thank you for your time in reviewing our paper. Respectfully, we would like to clarify that we do not learn a linear map, but rather a nonlinear matrix-valued map
$x \mapsto A(x,x^*)$ via the parametrization (11) to allow for nonlinea... | Summary: This paper proposes a novel parameterization of the extended linearisation form of a dynamical system that guarantees global contractivity. Whilst the most basic form of this parameterization only ensures contractivity in some (implicitly defined) metric, a latent space version is also proposed that enables mo... | Rebuttal 1:
Rebuttal: **Introduction**
Thank you for clarifying your thought process. We will make the following changes to the introduction to make things clearer:
We will replace line 17 onward in the first paragraph with:
"Beyond approximating the vector field f, it is desirable to ensure that the learned vector ... | Rebuttal 1:
Rebuttal: **Explanation of Updated Results**
After submission, we improved the performance of our method by implementing a change in the training loss. We had previously trained the model $f$ and the encoder $\phi$ to output $x_{t+1}$ from $x_{t}$ using a single Euler integration step: $x_{t+1} = \phi^{-1}... | NeurIPS_2024_submissions_huggingface | 2,024 | null | null | null | null | null | null | null | null |
Warm-starting Push-Relabel | Accept (poster) | Summary: This paper studies the push-relabel algorithm with warm-starts for the max-flow problem. Despite its empirical efficiency, the standard analysis of the algorithm is known to result in a somewhat pessimistic bound of $O(n^2m)$. The current paper considers a situation where the algorithm takes a predicted pseudo... | Rebuttal 1:
Rebuttal: Please see our responses to all reviewers for your questions on comparing warm start PR to warm start FF, as well as computing $\eta$.
Comment: *I would like to know whether we can develop a better method for learning $\hat{f}$ based on the error measure in Definition 1.*
Response: It is an inte... | Summary: Authors provide the first theoretical guarantees for the Push-Relabel (PR) algorithm coupled with the gap relabeling heuristic, to address max flow problems, while using an arbitrary (pseudo-flow) initialization such as a predicted flow. The main result relates to proving a worse case complexity of $O(\eta * n... | Rebuttal 1:
Rebuttal: Please see our responses to all reviewers for your questions on comparing warm start PR to warm start FF.
Comment: *visualizations corresponding to values reported to Table 2 / 5 could have been interesting to really observe the quadratic behavior.*
Response: Please see the pdf attached in the r... | Summary: This paper provides the running-time complexity analysis for warm-start Push-Relabel algorithm for the fundamental max-flow/min-cut problem. In particular, they study learning-augmented version of the Push-Relabel algorithm, where the algorithm can start from a pseudo-flow with error bounded by $\eta$. The mai... | Rebuttal 1:
Rebuttal: Please see our responses to all reviewers for your questions on comparing warm start PR to warm start FF, as well as computing eta.
Comment: *The algorithm (Algorithm 2) requires a bound on the error $\eta$ of the prediction, which can be hard to estimate in practice.*
Response: We refer the re... | Summary: This paper propose and analyzes a warm starting scheme for the classical
push-relabel algorithm for max-flow problems. The basic approach is to convert
the warm-starting flow into a cut-saturating flow, at which point a clever
push-relabel scheme is used to ensure that the source-side of the cut has only
exces... | Rebuttal 1:
Rebuttal: Please see our responses to all reviewers for your questions on comparing warm start PR to warm start FF, as well as computing eta.
Q: *Can $\eta$ ever exceed O(m)? That is, are the circumstances where the warm-started version of push-relabel has worse asymptotics than cold-start push relabel?*
... | Rebuttal 1:
Rebuttal: We thank the reviewers for their careful reading of our paper! We are glad they agree that warm-starting the push relabel algorithm to find a max flow/min cut is an interesting theoretical question with potentially useful practical applications.
Please see comments on questions brought up by seve... | NeurIPS_2024_submissions_huggingface | 2,024 | null | null | null | null | null | null | null | null |
Fairness without Harm: An Influence-Guided Active Sampling Approach | Accept (poster) | Summary: In this paper, the authors propose Fair Influential Sampling(FIS) to achieve a better Pareto Frontier of the fairness-accuracy tradeoff through sampling training data. To maintain fairness without diminishing the accuracy, training data examples are validated through a validation dataset. In this case, privacy... | Rebuttal 1:
Rebuttal: We want to thank the reviewer for their positive feedback and comments. We will address individual comments below.
**Response to the refinement of validation set**: Thank you for raising this concern. For this issue, we did some ablation study in Section 6.3 to explore the impact of the validatio... | Summary: This paper addresses the challenge of achieving fairness in machine learning (ML) without compromising model accuracy. The authors propose a novel active data sampling algorithm that mitigates group fairness disparity by acquiring more data without requiring sensitive attribute annotations for training, thus p... | Rebuttal 1:
Rebuttal: We want to thank the reviewer for their positive feedback and comments. We will address individual comments below.
**Re to W1**: Thank you for raising this concern. We want to clarify that the definitions of well-known fairness metrics, such as DP and EOd, are provided in Section 3 (Preliminaries... | Summary: Group/statistical notions of fairness are known to be at odds with the accuracy metric. The authors propose an active sampling strategy to improve the tradeoff between fairness and predictive performance. This sampling approach only requires the sensitive information to be known in a small validation set. They... | Rebuttal 1:
Rebuttal: We want to thank the reviewer for their detailed feedback and comments. We will address individual comments below.
**Re to W1**: We want to clarify that we do not assert that "Those approaches do not strive to achieve uniform accuracy over all groups." Rather, our claim is that these approaches ... | Summary: The authors propose a method called Fair Influential Sampling (FIS), which selects candidate data points for training based on their estimated improvement on both fairness and accuracy. This estimation considers the influence of a new data point as the average gain in accuracy and fairness achieved by performi... | Rebuttal 1:
Rebuttal: We want to thank the reviewer for their positive feedback and comments. We will address individual comments below.
**Response to weakness**: Thank you for raising this good point. We appreciate the feedback and apologize for not emphasizing this aspect in the main text of our paper. We have condu... | null | NeurIPS_2024_submissions_huggingface | 2,024 | null | null | null | null | null | null | null | null |
Optimal Top-Two Method for Best Arm Identification and Fluid Analysis | Accept (poster) | Summary: This paper considers the top-2 algorithm for identifying the best arm in multi-armed bandits. The authors introduce a new approach for determining the optimal $\beta$ for sampling the empirical best and best challenger arms. This novel method relies on a function of allocations anchored at a threshold, which i... | Rebuttal 1:
Rebuttal: We thank the reviewer for the suggestions and provide our responses below.
We first address the weaknesses.
**[Use of forced exploration in analysis and experiments]** We address this weakness in **(1)** of the global rebuttal and in our rebuttal to Reviewer 4f2z at "Use of foced exploration in... | Summary: The paper investigates the problem of fixed-confidence BAI in stochastic multi-armed bandits. Simply stated, the problem entails finding the best arm (the arm with the largest mean reward) as quickly as possible, subject to an upper bound on the error probability. There exist a plethora fo works that investiga... | Rebuttal 1:
Rebuttal: We thank the reviewer for the suggestions and provide our responses below.
The typos have been corrected in the revision. Thanks.
**Response to the questions:**
**[Intuitive reason of using the anchor function]** The proposed algorithms are motivated by the first order conditions which uniquel... | Summary: This paper considers the problem of identifying an optimal "top-2" algorithm for the best arm identification problem in the bandits framework. The question of how to allocate pulls between incumbents and challengers is considered. An algorithm for achieving the best balance is described.
Strengths: "Top-2" ap... | Rebuttal 1:
Rebuttal: We thank the reviewer for the suggestions and provide our responses below.
We first address the weaknesses pointed out.
In the revision we have attempted to improve the exposition, including the mathematical presentation. We would appreciate further specific suggestions that you may have to i... | Summary: This paper focuses on best arm identification (BAI) under the fixed confidence setting. It assumes the underlying distributions are from the single parameter exponential function. It discusses the problem of how to find the optimal $\beta$ for Top-Two type of algorithm for BAI, where $\beta$ denotes the probab... | Rebuttal 1:
Rebuttal: We thank the reviewer for the suggestions and provide our responses below.
We first address the weaknesses pointed out.
**[Dynamics of the ODEs]** We illustrate the evolution of anchor function and indexes in the fluid dynamics in Figure 2 of the pdf attached with the global response. In the fig... | Rebuttal 1:
Rebuttal: Here we address questions/weaknesses raised by more than 1 reviewer. The manuscript will be updated as per our reponses.
**1) Conjectures on sufficient exploration and numerical experiments supporting them:** As reviewer 4f2z rightly pointed out, AT2 needs forced exploration. Otherwise, for insta... | NeurIPS_2024_submissions_huggingface | 2,024 | Summary: The paper is about fixed confidence bandit best arm identification. The authors study the top-two family of algorithms. Those algorithms identify at each time step a leader and a challenger arm, and sample one of those two arms: typical top-two algorithms in the literature sample the leader with a fixed probab... | Rebuttal 1:
Rebuttal: We thank the reviewer for the suggestions and provide our responses below.
We first address the weaknesses.
**[Asymptotic results of the paper]** We thank the reviewer for pointing us to the recent reference [Jourdan and Degenne, '23]. We respond to it in **(3)** of the global rebuttal.
**... | null | null | null | null | null | null |
Transductive Learning is Compact | Accept (poster) | Summary: The paper looks at the question of whether or not the transductive sample complexity is compact in the sense that the transductive sample complexity globally can be deduced from understanding the tranductive sample complexity of finite instances of transductive learning. More formally transdictive learning is ... | Rebuttal 1:
Rebuttal: Thank you for your time and attention in reviewing the paper. We will address your questions in order.
Line 48: $H|_S$ can indeed be infinite when $Y$ is infinite. For this reason, we refer only to *finite subsets* of $H|_S$ as finite projections of $H$. That is, a finite projection of $H$ refers... | Summary: The paper studies transductive learning with general real-valued loss functions. The paper shows that:
(1) for proper metric loss functions and continuous loss functions defined on compact spaces, the sample complexity of (realizable and agnostic) transductive learning a class H is exactly the same as the samp... | Rebuttal 1:
Rebuttal: We thank the reviewer for their time and attention in reviewing the paper.
We note that $H|_S$ is the collection of all functions in $H$ restricted to the datapoints $S$. Indeed $H|_S$ can easily be infinite, in which case it is not a finite projection of $H$. We instead refer to the finite subs... | Summary: This work studies the transductive learninng model. In this model, given a domain $X,Y$ and $H\subset Y^X$ the adversary choses data $(x_i,y_i)$ (in the realizable setting the adversary can chose the labels after the reveal to the learner). The adversary then uniformly at random hides one data point. The goa... | Rebuttal 1:
Rebuttal: We thank the reviewer for their time and attention in reviewing the paper.
> The authors do not convinced me about the importance of this model. The authors should explain why this model is important, what it explains that other models does not (i.e., the classical PAC learning).
We argue that ... | Summary: This document explores the concept of compactness in the context of transductive learning, a model closely related to the PAC model in supervised learning. The authors demonstrate that for a broad class of loss functions, a hypothesis class can be learned with a specific transductive sample complexity if and o... | null | null | NeurIPS_2024_submissions_huggingface | 2,024 | null | null | null | null | null | null | null | null |
Knowledge Circuits in Pretrained Transformers | Accept (poster) | Summary: This paper tries to study how language models (LMs) recall factual knowledge by finding circuits within the model internals in a mechanistic way, on a dataset of (subject, relation object)-like sentences. The discovered circuits help understand the mechanisms of LM's knowledge recall, and also enable some expl... | Rebuttal 1:
Rebuttal: Dear reviewer 8n8x,
Many thanks for your detailed and constructive comments. We hope that this reply answers all your questions, and we look forward to further discussion.
**Addressing Weaknesses and Questions:**
> **WK1&WK2: The findings in this paper are not particularly surprising, and the i... | Summary: The paper introduce a critical subgraph named Knowledge Circuit in the language model to understand how LLMs store and express the knowledge. Through extensive experiments and cases analysis, the authors show the significance of this proposed Knowledge Circuit, which can unveil implicit neural knowledge repres... | Rebuttal 1:
Rebuttal: Dear reviewer 8w5y,
Many thanks for your detailed and constructive comments. We appreciate that you are optimistic about the impact of the work.
**Response to weakness and questions:**
> **WK: It would be better if the authors analyzed more knowledge editing methods.**
Thanks for your advice,... | Summary: This paper focus on the problem of interpreting the knowledge storage mechanism of the large language models. The authors proposed a new perspective, which uses the knowledge circuit to understand how the language model stores and expresses the knowledge. Knowledge circuit is a subgraph of the computation grap... | Rebuttal 1:
Rebuttal: Dear reviewer 7LFR,
Many thanks for your detailed and constructive comments. We recognize and agree with the limitations of our work and address the specific comments and improvements here.
**Response to weakness and questions:**
> **WK1: Relation with reference [36]**
Thank you for bringing ... | Summary: The authors propose utilizing techniques from mechanistic interpretability to explore the connections between model components involved in factual recall. They systematically ablate connections between model components in a reverse topological order, maintaining a list of the connections that most significantl... | Rebuttal 1:
Rebuttal: Dear reviewer jmbU,
We are very happy that you are optimistic about the impact of the work. In this comment, we respond to the question and the suggested improvements.
**Response to weakness & questions:**
> **WK1&Q1: About the editing layer we select in our experiments.**
The layer in curre... | Rebuttal 1:
Rebuttal: Dear Reviewers,
We thank you all for the detailed and constructive comments.
*All* reviewers together found our contribution “a novel perspective on interpreting, points us in the right direction.” (Njw6) and “first to provide a mechanistic interpretation of the impact of knowledge editing tech... | NeurIPS_2024_submissions_huggingface | 2,024 | Summary: This paper presents a circuit-based analysis of factual recall and model editing in language models. The authors make two key claims about knowledge storage and retrieval: 1) Factual recall is compatible with circuit-based frameworks. 2) Traditional model editing methods influence behavior change within the mo... | Rebuttal 1:
Rebuttal: Dear reviewer Njw6,
Thank you very much for your well-considered review! We are very happy that you appreciate the work we put into exploring a novel perspective on interpreting and understanding the subject. Thank you also for the insightful questions, and please let us know if you have any more... | null | null | null | null | null | null |
SpatialPIN: Enhancing Spatial Reasoning Capabilities of Vision-Language Models through Prompting and Interacting 3D Priors | Accept (poster) | Summary: This paper introduces SpatialPIN to enhance the spatial reasoning capabilities of Visual Language Models (VLMs) by explicitly incorporating 3D priors from 3D foundation models. Extensive experiments demonstrate that SpatialPIN is effective for spatial Visual Question Answering (VQA) and robotic pick-and-place ... | Rebuttal 1:
Rebuttal: **Q1**: Sections 3.1 and 3.2 mix method descriptions with prompts and corner cases, making it challenging for readers to understand.
**A1**:
Thank you for pointing this out! Since our methodology involves VLMs interacting with 3D priors, the output of VLMs sometimes serves as input for 3D founda... | Summary: This paper presents SpatialPIN, a framework designed to enhance the spatial reasoning capabilities of Vision-Language Models (VLMs) through prompting and interacting with 3D priors from multiple foundation models in a zero-shot, training-free manner. The authors argue that current state-of-the-art spatial reas... | Rebuttal 1:
Rebuttal: **Q1**: The combination of techniques used in SpatialPIN may not be entirely new, as it builds on existing 3D and VLM methodologies.
**A1**: While using existing 3D and VLM methods, composing the framework of SpatialPIN isn't trivial and requires many thoughtful designs to first understand 2D obj... | Summary: This paper presents a pipeline designed to equip 2D Vision Language Models with the capability to understand 3D spatial relationships. The key benefits of this framework are its zero-shot, training-free nature. The effectiveness of this approach has been validated through experiments on spatial Visual Questio... | Rebuttal 1:
Rebuttal: **Q1**: Ambiguity in motivation. The authors assert that "high-level 3D-aware tasks are underexplored," yet there is notable prior work such as SpatialVLM and a series of studies involving GPT-4V for robotics that address high-level 3D-aware tasks.
**A1**: We thank the reviewer for highlighting ... | null | null | Rebuttal 1:
Rebuttal: We thank all reviewers for their constructive comments. We appreciate the reviewers for recognizing the novelty and empirical evaluation of our work i.e., "the approach is novel, addresses an important gap in current VLMs' spatial reasoning capabilities" (K1EW), "experiments are comprehensive" (C3... | NeurIPS_2024_submissions_huggingface | 2,024 | null | null | null | null | null | null | null | null |
Reawakening knowledge: Anticipatory recovery from catastrophic interference via structured training | Accept (poster) | Summary: This paper explores the training dynamics of neural networks, particularly language models in a structured training setting where examples are presented cyclically. The paper discovered a phenomenon called "anticipatory recovery," where models in each training epoch recover from forgetting of an example right ... | Rebuttal 1:
Rebuttal: Thank you for the detailed feedback.
**Re: Unfair comparison for catastrophic forgetting and anticipatory recovery.**
> We agree with the reviewer that there is no catastrophic forgetting if all tasks are trained in every epoch, or if training passes over all tasks many times. However, in our wor... | Summary: The authors identify a phenomenon termed anticipatory recovery, where if the model is trained on a fixed shuffled data sequence for multiple epochs, the loss of a training sample will first increase and then decrease again right before it is met again. Extensive experiments have been conducted to show the phen... | Rebuttal 1:
Rebuttal: Thank you for the detailed feedback.
**Re: “In Figure 1(b), how exactly is the loss computed?”**
> The loss in figure 1(b) is computed by replicating figure 1(a) on each document in the training sequence, and re-aligning these curves so that 0 on the x-axis always represents the moment before the... | Summary: The paper studies a specific kind of structured training dubbed cyclic training, where documents (or tasks) are presented cyclically in a fixed, repeated sequence. The paper reveals that LLMs exhibit anticipatory recovery behavior, i.e. they begin to recover from forgetting a document before encountering it ag... | Rebuttal 1:
Rebuttal: Thank you for the detailed feedback.
**Re: “its practical value remains unclear.” and “cyclic training appears less natural compared to continual learning”.**
> We would like to argue the opposite, that traditional continual learning is less natural than cyclic training. Most standard benchmarks ... | Summary: This paper takes a deeper look at the training dynamics of neural networks such as LLMs in a particular kind of non-IID learning setting when tasks or data is iterated through cyclically in a fixed repeated sequence. The authors find evidence in this setting of an emergent behavior that they call "anticipatory... | Rebuttal 1:
Rebuttal: Thank you for the detailed feedback.
**Re: Writing improvement and summary of main findings.**
> We have a summary of main findings in the final paragraph of section 1 but we agree some elaboration will help readers anticipate what is coming. We thank the reviewer for the writing feedback, and wi... | Rebuttal 1:
Rebuttal: **The practicality of cyclic training and relevance of prequential evaluation:**
We appreciate the reviewers' recognition of the technical soundness and overall quality of our manuscript. While reviewers dpPf and gu9E have appreciated the contribution of our findings, we understand that reviewe... | NeurIPS_2024_submissions_huggingface | 2,024 | null | null | null | null | null | null | null | null |
Deterministic Policies for Constrained Reinforcement Learning in Polynomial Time | Accept (poster) | Summary: The authors propose an algorithm that efficiently computes near-optimal deterministic policies for constrained reinforcement learning (CRL) problems. The proposed algorithm comprises three ideas: 1) value-demand augmentation, 2) action-space approximate dynamic programming, and 3) time-space rounding. Under so... | Rebuttal 1:
Rebuttal: Thank you for your comments! We address the concerns about the contributions of our paper. Our paper is highly focused, addressing only the question proposed on line 29 of our intro: “Can near-optimal deterministic policies for constrained reinforcement learning problems be computed in polynomial ... | Summary: The paper proposes a novel algorithm for constraint MDPs whose constraints have TR properties. The algorithm converts the original problem to an unconstraint MDP. To address the challenge of high complexity of the proposed approach, the method discovers that for constraints with TSR properties, it can accelera... | Rebuttal 1:
Rebuttal: Thank you for your comments! Although we introduce TSR constraints, they generalize the main constraints studied by the constrained reinforcement learning (CRL) literature, including expectation, almost sure, and anytime constraints. CRL is an entire subarea of RL that was born as early as 1981 [K... | Summary: The paper presents an algorithm that computes deterministic policy for constraint RL. The work is majorly theoretical. The proposed algorithm for the worst-case analysis achieves the best approximation guarantees. Their approach incorporates ideas from state augmentation with value function, approximate dynami... | Rebuttal 1:
Rebuttal: Thank you for your comments!
We first address the weakness comments:
1. We do not believe our motivation is weak as it comes from two longstanding lines of work: (1) deterministic policies, which are critical for multi-agent coordination and for autonomous vehicles [Hong, Geißer], and (2) Anytim... | Summary: The paper provide a novel polynomial-time algorithm for finding near-optimal deterministic policies in constrained MDPs with a large family constraints beyond expectation. The algorithm first considers in the dual covering problem minimizing cost for constrained value. The new problem is then viewed as solving... | Rebuttal 1:
Rebuttal: Thank you for your comments! The main techniques used in this paper are packing-covering equivalence (essentially, strong duality), dynamic programming, and rounding. These are general techniques found in standard approximation algorithms textbooks such as the book “The Design of Approximation Alg... | Rebuttal 1:
Rebuttal: Since the reviewers agree with the mathematical correctness of our results, we emphasize the significance of our contributions.
To summarize, our general framework provides answers to not just one but three open complexity questions spanning two longstanding lines of work: we prove polynomial-tim... | NeurIPS_2024_submissions_huggingface | 2,024 | null | null | null | null | null | null | null | null |
Cross-model Control: Improving Multiple Large Language Models in One-time Training | Accept (poster) | Summary: The development of the Cross-model Control method, which enables the use of fine-tuning outcomes from one model to improve other models through a portable tiny language model, effectively reducing training costs and computational resources.
The introduction of a novel token mapping strategy called prefix match... | Rebuttal 1:
Rebuttal: We deeply value the time and effort you've invested in offering us constructive feedback. In response, we are endeavoring to address your points as follows:
**Q1: Scalability and Efficiency: The computational requirements and potential bottlenecks of implementing the PM-MinED strategy, especially... | Summary: Large language models face a series of common optimization requirements under specific applications or ethical standards. Existing methods only optimize one target model at a time, which requires changing model parameters or adding new parameters. Authors found that the logit changes of different models before... | Rebuttal 1:
Rebuttal: We sincerely appreciate your valuable comments and feedback. We will describe more details of the experiments if more pages are provided.
**Q1: "more hyper-parameter details" & "impacts of training epochs"**
A1: **We selected the hyperparameters that achieve the best performance for the differe... | Summary: This paper proposes a novel training method named CMC, designed to improve performance for multiple LLMs in one-time training, thereby reducing training costs by reusing fine-tuning outcomes. The core approach of the paper consists of three steps:
1. Introducing a portable tiny language model (delta model) wit... | Rebuttal 1:
Rebuttal: We deeply value the time and effort you've invested in offering us encouraging feedback. In response, we are endeavoring to address your points as follows:
**Q1：During the inference stage, if the user LLM generates a token that does not exist in the delta model's vocabulary, what is the input to ... | Summary: This paper investigates an interesting problem in fine-tuning LLMs: how to reuse the fine-tuning outcomes of one model for other models to reduce training costs. The authors provide an important empirical finding: different models exhibit highly similar logit shifts before and after the same fine-tuning proces... | Rebuttal 1:
Rebuttal: We sincerely appreciate your recognition of our work and the valuable suggestions. We would like to address each of the questions as follows:
**Q1: Because the findings in Figure 2 are crucial for proposing the method, I would like to see not only the qualitative analysis provided in Figure 2 but... | null | NeurIPS_2024_submissions_huggingface | 2,024 | Summary: The paper introduces Cross-model Control (CMC), a novel method to improve multiple large language models (LLMs) in a single training session using a portable tiny language model. The core idea hinges on the observation that the logit shifts before and after fine-tuning are similar across different models. Leve... | Rebuttal 1:
Rebuttal: We deeply appreciate your encouraging feedback regarding the originality and significance of our work. We are pleased to address your inquiries in detail as follows:
**Q1: While PM-MinED is introduced to handle different vocabularies, its effectiveness might be limited when dealing with highly di... | null | null | null | null | null | null |
Synergistic Dual Spatial-aware Generation of Image-to-text and Text-to-image | Accept (poster) | Summary: Given that previous works for standalone SI2T (Scene-to-Image-to-Text) or ST2I (Scene-to-Text-to-Image) perform imperfectly in spatial understanding due to the difficulty of 3D spatial feature modeling, this paper proposes tp model SI2T and ST2I together under a dual learning framework. Within this dual framew... | Rebuttal 1:
Rebuttal: We are grateful that you acknowledge the strengths of our work. We've conducted additional experiments and tried every effort to address your concerns. We hope you'll reconsider the evaluation if you find our responses are effective. Following are our outputs.
***
**Q1:** Although the proposed me... | Summary: This paper presented a novel model for the SI2T and ST2I tasks. The proposed model combines the two dual tasks and let them mutually learn via intermediate feature sharing. Through this framework, both SI2T and ST2I are enhanced. The author also provide analysis that how this method works.
Strengths: 1. The p... | Rebuttal 1:
Rebuttal: We sincerely thank you for you for your time and provide us with rich and constructive feedback on our paper. And we believe it will surely improve our work. Following we extract your concerns into points and try to address them one by one.
***
**Q1:** I wonder if 3D modeling is necessary for the... | Summary: This paper presented a novel dual learning framework for spatial image-to-text and spatial text-to-image generation. The main model is a combination of three discrete diffusion models, where an intermediate 3D presentation is first generated and then the image and text outputs are generated based on the 3D fea... | Rebuttal 1:
Rebuttal: Thank you for going through our paper so deep and carefully. We appreciate that you acknowledge the novelty of our proposed task and the comprehensive experiments. All your possible concerns are addressed as follows.
***
**Q1:** How many node categories does the 3DSG has and how to define the hig... | Summary: The paper introduces a dual learning framework and a 3D scene graph representation for enhancing spatial image-to-text and text-to-image tasks in visual-spatial understanding. The proposed Spatial Dual Discrete Diffusion (SD3) system outperforms existing methods on the VSD dataset, demonstrating the effectiven... | Rebuttal 1:
Rebuttal: We sincerely thank you for your time and the careful review for our paper. Your suggestions will definitely help improve our paper. We try to address your questions point by point as follows.
***
**Q1:** Could the authors consider prioritizing the most relevant or impactful references and possibly... | Rebuttal 1:
Rebuttal: # General Response to All Reviewers
***
Dear Reviewers,
Thanks for all of your insightful and cheerful comments on our manuscript. Your feedback will greatly assist us in enhancing the quality of our paper, and we are committed to incorporating your suggestion in our revision process.
Meanwhil... | NeurIPS_2024_submissions_huggingface | 2,024 | null | null | null | null | null | null | null | null |
Initialization is Critical to Whether Transformers Fit Composite Functions by Reasoning or Memorizing | Accept (poster) | Summary: The work mainly investigates whether transformers indeed learn primitive functions or simply learn surficial correlations (i.e., "memorizing") when trained on a mannually-designed toy compositional task, which contains a set of composite numerical functions (e.g., $g_1(x)=x+5, g_2(x)=x+1,...$).
Authors find ... | Rebuttal 1:
Rebuttal: $\textbf{Weakness 1.1 and Q 1.1}$
The experiment setting of this work might be over-simplified...
$\textbf{Reply}$
We appreciate you suggesting of the SCAN and COGS datasets. We have conducted experiments on these two tasks, and the results are consistent with our conclusions. Additionally, we ... | Summary: This paper studies the role of transformer initialization scale on its ability to compositionally generalize. Using a simple arithmetic task, the authors find that when initialization scale is low, the model learns inferential (or compositional) solutions, whereas when the initialization scale is high, models ... | Rebuttal 1:
Rebuttal: $\textbf{Weakness 1}$
Other studies demonstrate that this type of arithmetic generalization can be achieved...
$\textbf{Reply}$
Arithmetic generalization is a widely studied problem, and several works have improved performance primarily on arithmetic tasks by manipulating input token forms or p... | Summary: Transformers show remarkable capabilities but struggle with compositional tasks, raising questions about their true understanding versus input-output mapping. The authors investigates transformers' generalization to unseen compositional tasks using anchor functions, revealing that initialization scale impacts ... | Rebuttal 1:
Rebuttal: $\textbf{Point 1}$
While the problem and findings are new, the approach lacks novelty.
$\textbf{Reply}$
We appreciate your recognition of the novelty in our problem and findings. While our approach may appear straightforward, we hope reviewers appreciate the challenges we overcame in experiment... | null | null | Rebuttal 1:
Rebuttal: We thank the reviewers for your thoughtful and insightful comments. We have addressed every comment, and believe that, taken together, the reviewers' comments have improved the manuscript significantly. To address reviewers' common concerns, we have supplemented relevant experimental results, adde... | NeurIPS_2024_submissions_huggingface | 2,024 | null | null | null | null | null | null | null | null |
On the Limitations of Fractal Dimension as a Measure of Generalization | Accept (poster) | Summary: This study revisits recent claims about persistent homology dimension (in the limit a fractal dimension) and the generalization gap of supervised learning systems. The essential theoretical background is summarized, and then three main experiments are conducted, concerning 1) (partial) correlation of the conce... | Rebuttal 1:
Rebuttal: We thank this referee for their careful, detailed feedback. We are pleased to see that they found our experiments “conducted with high quality”, the results “insightful”, “useful” and “relevant” to the community, and that they appreciated our discussion on limitations as a “role-model” on how thes... | Summary: This work empirically studies the connection between the generalization gap and the fractal dimension of a neural network's optimization trajectory. A recent line of research has established such connections in a theoretical manner through upper generalization bounds. Specifically, three types of experiments a... | Rebuttal 1:
Rebuttal: We thank this reviewer for the careful review provided. We are very happy to see that they also appreciated the importance of evaluating existing work to “build upon solid foundations”. We would like to begin by addressing the weaknesses they identify.
1. **Discrepancy between our correlation coe... | Summary: This paper performs experiments to evaluate the correlation between the generalization gap and the persistent homology (PH) dimension, a measure of fractal dimension deriving from topological data analysis. The authors identify that the $\ell^2$ norm of the final parameter iterate correlates more strongly with... | Rebuttal 1:
Rebuttal: We thank this reviewer for their feedback and are happy to see that they found our conclusions “well-supported” and the text “well-written and organized.”
We would like to first address the concern regarding the clarity of the PH dimension definition. We acknowledge that the concept of fractal di... | Summary: The paper investigates the effectiveness of using fractal dimensions, particularly the Hausdorff dimension and persistent homology dimension, as measures for predicting the generalization gap in neural networks. It demonstrates that fractal dimensions fail to predict generalization for models trained from poor... | Rebuttal 1:
Rebuttal: We thank the reviewer for their time and feedback on our work. We are pleased to see that they found our evaluation “critical;” our statistical analysis “rigorous”, and our demonstration of the shortcomings of the existing theory “convincing”. We also appreciate the recognition of the opportunitie... | Rebuttal 1:
Rebuttal: We thank all the referees for the careful and detailed reviews, and for their encouraging words, describing our contribution as “critical”, “rigorous” and “convincing” (rTx2); “well-supported” and “well-written and organized” (whkD); “necessary in order to establish a solid foundation future work ... | NeurIPS_2024_submissions_huggingface | 2,024 | null | null | null | null | null | null | null | null |
Single Image Unlearning: Efficient Machine Unlearning in Multimodal Large Language Models | Accept (poster) | Summary: The authors propose Single Image Unlearning (SIU), to unlearn the visual recognition of a concept in Multimodal Large Language Models (MLLMs). They also introduce MMUBench, a new benchmark for machine unlearning in MLLMs.
Strengths: The paper introduces a new approach to machine unlearning in MLLMs. The conce... | Rebuttal 1:
Rebuttal: We are greatly encouraged by your positive comments. Thanks a lot for all the appreciation.
>**Q1**: The main limitation of the paper is (as acknowledged by the authors) that they only use a single LLM type (LLAVA).
**A1**: Thank you for your rigorous review. Please refer to general response Q... | Summary: In this paper, the authors proposed a single image unlearning (SIU) approach, which aims at unlearning the concepts recognized from a single training image. A Dual Masked KL-divergence (DMK) loss was introduced to be jointly trained with the cross-entropy loss to mitigate the degradation of MLLMs. Besides, to ... | Rebuttal 1:
Rebuttal: We are grateful for your attentive comments and providing thoughtful feedback on our work. We will provide our insights point by point below.
>**Q1**: It is not clear whether the incorrect answer is caused by unlearning or hallucinations.
**A1**: Thank you for your insightful comment. The incorr... | Summary: This paper proposes an algorithm for the unlearning of the visual recognition of concepts from single images in Multimodal Large Language Models (MLLMs) and introduces a benchmark dataset for evaluation. To achieve single image unlearning in MLLMs, the following process is conducted. First, based on the given ... | Rebuttal 1:
Rebuttal: Thank you for your critical feedback and suggestions. We address your thoughts point by point below.
>**Q1**: The structure of the paper needs improvement.
**A1**: Thank you for your suggestion. We agree that the structure of the paper can be improved. In the final version, we will improve the p... | null | null | Rebuttal 1:
Rebuttal: We first thank all the reviewers for their insightful feedback and suggestions. Our work is a pioneering study in the field of machine unlearning in Multimodal Large Language Models (MLLMs). In this paper, we constructed a comprehensive benchmark to evaluate machine unlearning methods in MLLMs. Mo... | NeurIPS_2024_submissions_huggingface | 2,024 | null | null | null | null | null | null | null | null |
A Cat Is A Cat (Not A Dog!): Unraveling Information Mix-ups in Text-to-Image Encoders through Causal Analysis and Embedding Optimization | Accept (poster) | Summary: This papers address the bia problem of text-to-image generation. Unlike the previous work which focus on processing of the attention map, this work explores the text embedding, which is a key factor. Authors conduct expensive experiments to analysis the text embedding, then propose a interesting method to add... | Rebuttal 1:
Rebuttal: Dear reviewer,
We appreciate your valuable suggestions and taking the time to review our work! We have carefully considered your feedback, and we would like to address your concerns. In the following section, we respond to each weakness, while weakness 3 is included in the overall rebuttal at the... | Summary: The paper conducts an in-depth analysis of the impact of causal mechanisms in text encoders of text-to-image (T2I) diffusion models, which can lead to information bias and loss. The authors propose the Text Embedding Balance Optimization (TEBOpt) method, a training-free method to optimize text embeddings, resu... | Rebuttal 1:
Rebuttal: Dear reviewer,
We appreciate your valuable feedback and taking the time to review our work! We have carefully considered your suggestions and we would like to address your concerns. In the following section, we respond to each weakness and question, while weaknesses a and c are included in the ov... | Summary: This paper proposes Text Embedding Balance Optimization (TEBOpt) to enhance the distinctiveness between text embeddings of equally important objects when using texts as conditions for diffusion models. It begins with an intriguing and informative experiment to investigate why a prompt like "An <object1> and an... | Rebuttal 1:
Rebuttal: Dear reviewer,
We sincerely appreciate your valuable suggestions and supportive feedback, as well as the time you took to review our work! We love your summary of our work, demonstrating a thorough understanding. In the following section, we respond to each weakness and question, while weakness 3... | Summary: Background: Text-to-image diffusion models, such as Stable Diffusion, frequently encounter difficulties in accurately generating images from textual descriptions, especially when there are multiple objects -> object mixing and missing.
This paper focuses text embeddings to solve the problems.
Hypothesis: sel... | Rebuttal 1:
Rebuttal: Dear Reviewer,
Thank you for taking the time to review our work and providing us with your valuable suggestions and insightful comments. We have carefully considered your feedback, and we would like to address your concerns. In the following section, we respond to each weakness (W) and question (... | Rebuttal 1:
Rebuttal: Dear all reviewers,
We greatly appreciate the insightful suggestions and valuable comments from each reviewer. These have been immensely helpful and enlightening for refining this paper. In this section, we respond to common weaknesses and provide the PDF to include more qualitative and quantitat... | NeurIPS_2024_submissions_huggingface | 2,024 | null | null | null | null | null | null | null | null |
UNIT: Unifying Image and Text Recognition in One Vision Encoder | Accept (poster) | Summary: In this paper, the authors propose a novel training framework aiming at unify image and text recognition within a single model. The framework contains both visual and language decoders. In addition, the authors adopt a two-stage training, where the pre-training stage uses intra-scale data while the finetuning ... | Rebuttal 1:
Rebuttal: Thank you for recognizing the contribution of our model to the community. We will make it publicly available in the coming days. We understand that your main concerns and rating stem from some technical details and test settings. Below, we will address these issues one by one. We sincerely hope yo... | Summary: This paper presents a new visual backbone training framework named UNIT, which integrate image and text recognition simultaneously. The UNIT framework leverages text recognition, visual feature construction and image captioning as the pre-training tasks. The training pipeline includes two stage: intra-scale pr... | Rebuttal 1:
Rebuttal: Thank you for the constructive comments. We understand your main concerns are related to resolution settings and scaling behaviors. We will address these issues one by one and hope you can recognize the significance of our work.
## **Q1. Training with different resolutions**
Our model can be tr... | Summary: The paper proposes to enable vision encoders to support general image recognition and text recognition at the same time, in which a lightweight language decoder is proposed for predicting text outputs and a lightweight vision decoder is proposed to prevent catastrophic forgetting of the original image encoding... | Rebuttal 1:
Rebuttal: Thank you for your comprehensive and encouraging review. We are pleased to hear that you appreciate the technical contributions and the significantly improved performance of our proposed visual encoder, UNIT, along with the intra/inter-scale training paradigm. In the following, we will respond to ... | null | null | Rebuttal 1:
Rebuttal: We thank all the reviewers for their time, insightful suggestions, and valuable comments. We respond to each reviewer's comments in details. These issues will be addressed in the revised manuscript, and we believe this makes our paper much stronger.
We promise to release all the pre-trained mode... | NeurIPS_2024_submissions_huggingface | 2,024 | null | null | null | null | null | null | null | null |
DiffPano: Scalable and Consistent Text to Panorama Generation with Spherical Epipolar-Aware Diffusion | Accept (poster) | Summary: This paper introduces the DiffPano framework, designed to generate consistent panoramic images from multiple viewpoints based on a text description of a scene. The authors first create a panoramic video-text dataset from 3D scenes using Habitat Simulator, BLIP2, and an LLM. Using this dataset, they fine-tune t... | Rebuttal 1:
Rebuttal: We thank Reviewer Yygs for his constructive suggestions and appreciate comments like "interesting task", "useful for practical applications such as VR environments", "new dataset quite meaningful", and "effectively adapts to panoramic images".
We will gladly incorporate all the feedback in the rev... | Summary: The paper focusing on the simulation of panoramic images and the annotation of text descriptions. The primary contribution is the introduction of a LoRA-based fine-tuning technique, aimed at enhancing the performance of benchmark datasets. The authors propose a pipeline that integrates simulated panoramic data... | Rebuttal 1:
Rebuttal: We thank Reviewer K8Nd for his constructive suggestions and appreciate comments like "interesting idea", "new way to enhance 3D Vision", "well-structured", and "clear explanation". However, the reviewer seems to have misunderstood the core content of this paper,
the problem it solves, and the cor... | Summary: This work proposes a text-driven panorama generation framework to achieve scalable, consistent, and diverse panoramic scene generation. In particular, a spherical epipolar attention module with relative poses is designed to ensure multi-view consistency. Moreover, a comprehensive panoramic video-text dataset i... | Rebuttal 1:
Rebuttal: We thank reviewer 7HQE for his constructive suggestions and appreciate comments like “pioneer work in scalable multi-view panorama generation", "diverse and rich panoramic video-text dataset promote the community of panoramic generation", "compact and practical framework", "well-structured and eas... | Summary: The paper presents a novel framework called DiffPano for scalable and consistent text-to-panorama generation. The authors first build a panoramic video-text dataset, then propose a spherical epipolar-aware diffusion model to generate multi-view consistent panoramic images, addressing the limitations of existin... | Rebuttal 1:
Rebuttal: We thank the reviewer 2LuU for his constructive suggestions and
appreciate comments like "well-written", "easy to follow",
"first work for text to mv panorama", "valuable panoramic video-text dataset",
"novel and technically sound approach of spherical epipolar-aware diffusion",
"effectiveness... | Rebuttal 1:
Rebuttal: We thank all reviewers for their constructive suggestions and appreciate comments like
“well-written, well-structured and easy to follow, clear" (R2LuU, R7HQE, RK8Nd), "first work, pioneer work, novel, technically sound, interesting" (R2LuU, R7HQE, RK8Nd, RYygs), "valuable, promote, enhance, usef... | NeurIPS_2024_submissions_huggingface | 2,024 | null | null | null | null | null | null | null | null |
Transformers Represent Belief State Geometry in their Residual Stream | Accept (poster) | Summary: This work uses a new theoretical framework from Computational Mechanics to analyze representations in transformers. Using these tools, the authors can convert between a data-generating structure, such as one modeled with a Hidden Markov Model, and a geometric structure in a probability simplex that describes t... | Rebuttal 1:
Rebuttal: We are deeply grateful for your insightful review of our paper. Your recognition of our work as "well executed" with "clear writing and ideas" is greatly appreciated. We're particularly pleased that you found our examples and figures helpful in explaining complex concepts like mixed state represen... | Summary: The authors present evidence that belief states are linearly represented in the residual stream of transformer-based large language models. They also argue that belief state geometry represents information beyond next-token prediction and can be spread across the residual streams of multiple layers. This work ... | Rebuttal 1:
Rebuttal: ## Weaknesses
We have revised Sec. 2.2 for clarity (see responses to questions below). For issues dealing with simplicity of our experiments and applicability to LLMs, please see global response.
Regarding belief states affecting next token prediction making the results unsurprising: It is true ... | Summary: The paper studies the geometric structure of how belief states are represented in the residual stream of transformers using relevant theory from computational mechanics. The main object of investigation of the paper is a transformer model trained on data generated using an HMM. Following this, the theory of op... | Rebuttal 1:
Rebuttal: We sincerely thank you for your thoughtful review of our paper. We are particularly grateful for your recognition of our work as offering a "fresh perspective" and presenting a "well-thought-out toy model." Your positive comments on our visualizations are also much appreciated, as we care deeply a... | Summary: This paper proposes that the mixed-state presentation that is grounded in optimal prediction can explain the representations and computational steps in transformers over a sequence generated by some simple hidden Markov chains. This is done by first training the transformer until convergence, and then fitting ... | Rebuttal 1:
Rebuttal: Thank you for your insightful review of our paper. We appreciate your recognition of our work as "extremely interesting" and "well-verified," as well as your acknowledgment of the paper being "really well-written" with "very insightful" discussion. Your positive remarks about the importance of our... | Rebuttal 1:
Rebuttal: Dear Reviewers,
We sincerely thank you all for your reviews on our paper; we believe the paper will be much stronger because of them. We were quite happy by the positive recognition of our work, described as “very important for the interpretability of transformers,” a “fresh perspective,” and “a ... | NeurIPS_2024_submissions_huggingface | 2,024 | null | null | null | null | null | null | null | null |
Make Your LLM Fully Utilize the Context | Accept (poster) | Summary: This paper introduces INformation-INtensive (IN2) training, a data-driven approach to address the "lost-in-the-middle" problem in large language models (LLMs) with long context windows. The authors hypothesize that this issue stems from insufficient explicit supervision during long-context training. They creat... | Rebuttal 1:
Rebuttal: Thanks for your encouraging feedback on our work. We hope the following response could further address your concerns.
> W1: Since the data is synthesized and the chunk is randomly injected into the Doc, would the model memorize such synthetic data but lose the generalizability on other long-conte... | Summary: The paper argues that long-context models suffer from a “lost in the middle” phenomenon due to the infrequency of important data at any one position in the middle of the window during training. To correct for this, the authors propose IN2, a method of training on synthetic long-context QA data to artificially ... | Rebuttal 1:
Rebuttal: Thanks for your detailed review and constructive suggestions. We hope the following response and our additional experiments have addressed your concerns.
> W1 and Q4: Require experimental comparison with normal instruction tuning.
>
> R1: We present further experimental results to demonstrate tha... | Summary: In this paper, the authors propose two types of training to mitigate the lose-in-the-middle issue of Large Language Models (LLMs). The first one creates some QA pairs. They first pick 128-token segments with which they use LLM to generate QA pairs and mix these 128 tokens with other contexts to form long conte... | Rebuttal 1:
Rebuttal: Thanks for your insightful questions about our work. The following are our responses to these questions.
> Q1: What are the reasons for using synthetic data rather than existing natural long context datasets during fine-tuning?
>
> A1: There are mainly two reasons.
> - First, many existing natur... | Summary: This work tackles an essential problem in LLMs, called lost-in-the-middle, that causes the lack of information in the middle of lengthy input. As a data-driven solution, the authors propose a new training approach, INformation-INtensive (IN2). IN2 aims to enhance the use of context information in an unbiased p... | Rebuttal 1:
Rebuttal: Thanks for your valuable feedback on our work. We hope the following response could further address your concerns.
> W1: From the viewpoint of Scaling Laws, comparing Mistral-7B and its further trained model FILM-7B is unfair.
>
> R1: We present further experimental results to compare IN2 traini... | Rebuttal 1:
Rebuttal: We thank all reviewers for their constructive comments and questions on our paper.
We appreciate the recognition of the important and challenging research question we explored, the innovative idea and the effectiveness method we proposed, and the thorough evaluations we conducted.
We summarize ... | NeurIPS_2024_submissions_huggingface | 2,024 | null | null | null | null | null | null | null | null |
Towards Learning Group-Equivariant Features for Domain Adaptive 3D Detection | Accept (poster) | Summary: In this paper, the authors utilize a Gaussian Mixture Model (GMM) based grouping & exploration module on the object descriptors extracted from foreground points, then fed the group features to the Group-Correlation module and fuse with RPN to selectively detects objects similar to the individual group, which i... | Rebuttal 1:
Rebuttal: We thank the reviewer for the constructive assessment of our work. In the following, we address the concerns point by point. Please feel free to use the discussion period if you have any additional questions.
1) Lack of variety?
We address the multi-class adaptation in the main author rebuttal. I... | Summary: The performance of 3D object detection in large outdoor point clouds suffers in unseen environments due to inter-domain gaps. Existing domain adaptation methods, focusing on single factors like object size or shape, still leave substantial gaps. This work proposes a grouping-exploration strategy framework to a... | Rebuttal 1:
Rebuttal: We thank the reviewer for the constructive assessment of our work.
We address the concern you raise in the main author rebuttal. In particular, Table A2 shows the experimental results for the multi-class adaptation setting.
Please feel free to use the discussion period if you have any additiona... | Summary: This paper deals with the domain adaptive 3D detection based on point clouds, where labels are available for source domain and not for target domain. The authors noticed that previous works tried to model and minimize the domain gap in terms of one specific factor, like shape, density, etc., and proposed to cl... | Rebuttal 1:
Rebuttal: We thank the reviewer for the constructive assessment of our work. In the following, we address the concerns point by point. Please feel free to use the discussion period if you have any additional questions.
(1) How does the proposed method correct the bias in self-training?
We would like to let... | Summary: The authors propose a new framework that leverages a grouping-exploration strategy to address the inter-domain gap observed in 3D object detection. The core of their approach is to divide the available labels into multiple clusters, ensuring equal learning attention across these groups using group-equivariant ... | Rebuttal 1:
Rebuttal: We thank the reviewer for the constructive comment on our work.
We would like to let you know that the concern that you raised is addressed in the author rebuttal, which is visible to all reviewers.
In particular, Tables A1 and A2 in the uploaded rebuttal pdf show the experimental results for o... | Rebuttal 1:
Rebuttal: Dear Reviewers,
We greatly appreciate your valuable and constructive reviews for our paper. Most of the reviews are very helpful. We thoroughly noted your feedback and will make sure to include all the discussions that we have during the rebuttal in the final version of the paper. We summarized t... | NeurIPS_2024_submissions_huggingface | 2,024 | null | null | null | null | null | null | null | null |
Regularized Conditional Diffusion Model for Multi-Task Preference Alignment | Accept (poster) | Summary: This paper presents a novel approach using a regularized conditional diffusion model to align with preferences in multi-task reinforcement learning (RL) scenarios. The proposed method generates preference representations that effectively distinguish between various task trajectories and returns. By maximizing ... | Rebuttal 1:
Rebuttal: Thanks for taking the time to provide feedback on our paper. We appreciate your valuable comments and would like to address each of your concerns.
**1. Could you provide an additional ablation study on the influence of the number of tasks? Specifically, does the representation for task A differ w... | Summary: The paper proposes a novel method called the Regularized Conditional Diffusion Model for multi-task preference alignment, addressing the challenge of generating trajectories that align with human preferences in multi-task settings. This method introduces preference-based representations and a mutual informatio... | Rebuttal 1:
Rebuttal: Thanks for taking the time to provide feedback on our paper. We appreciate your valuable comments and would like to address each of your concerns.
**1. It is recommended that the authors emphasize these distinctions and update the contribution list.**
Thanks for your suggestion. We will emphasi... | Summary: This paper proposes a regularized conditional diffusion model which aligns with preferences for multi-task scenarios. To achieve this, learnable representations for preferences are aligned with preference labels, which can be adopted as condition inputs to guide the generation process of diffusion models. Mean... | Rebuttal 1:
Rebuttal: Thanks for taking the time to provide feedback on our paper. We appreciate your valuable comments and would like to address each of your concerns.
**1. Table 1 presents a large variation in performance for different tasks with the same implementation method. What factors contribute to this phenom... | null | null | Rebuttal 1:
Rebuttal: # Global Rebuttal
We appreciate all the reviewers for their dedication and recognition of our work, including acknowledging our motivation and novelty [L2wb, Y24x], the effectiveness and improvements of our method [L2wb, 3Xmu], and our presentation [L2wb, Y24x].
We have attempted to address eac... | NeurIPS_2024_submissions_huggingface | 2,024 | null | null | null | null | null | null | null | null |
Acoustic Volume Rendering for Neural Impulse Response Fields | Accept (spotlight) | Summary: This paper addresses the challenge of novel sound synthesis from arbitrary positions. To achieve this, acoustic fields are modeled using an implicit neural representation, integrating acoustic wave propagation rules. Subsequently, a specifically designed volume rendering technique is employed to ensure consist... | Rebuttal 1:
Rebuttal: Thank you for valuable feedback and acknowledgement of our work. We answer questions point by point below.
**W1: Network is not very big (6 layers). Why it takes 24 hours to train the model, while the Instant-NGP is fast? Could the training time be reduced?**
Most of the runtime for our method i... | Summary: This method proposes a physics-based approach for learning neural impulse response field through implicit wave propagation modeling. This is the first work that really incorporates acoustic wave propagation principles for constructing an impulse response field, leveraging volume rendering techinques in analagy... | Rebuttal 1:
Rebuttal: Thank you for a thoughtful review, valuable feedback and acknowledgement of our work. We provide point by point clarifications and answer questions below.
**W1: Summarize the key characteristics of the simulator in the main text.**
Thank you for this suggestion, we will expand the description of... | Summary: This work presents a method to render a neural impulse response field to generate room impulse responses at novel positions after being trained with a number of limited RIR samples. The method works similar to a Nerf, rays are cast from a sphere and sampled at discrete positions along the rays. Here the signal... | Rebuttal 1:
Rebuttal: Thank you sincerely for the thoughtful review and valuable feedback. Below, we address the questions raised and outline the revisions that we will include in the camera-ready version of the manuscript.
**W1: Binaural sound rendering.**
Please refer to **Zero-shot binaural audio rendering** in th... | Summary: The authors reformulate neural fields to model the spatial interactions of sound, by learning to predict impulse response in a fixed transmitter / arbitrary receiver pose geometry. The method substantially outperforms contemporary work on simulated and real world datasets, and lays the groundwork for new immer... | Rebuttal 1:
Rebuttal: Thank you for a thoughtful review and valuable feedback. We address the questions related to our work below.
**Q1: Computing efficiency and runtime of framework.**
Please refer to **Computing efficiency** in the global rebuttal.
**Q2: Include user study of qualitative evaluations about the zero... | Rebuttal 1:
Rebuttal: We would like to express our gratitude to all the reviewers for their insightful comments and feedback. Below, we address two common questions raised by multiple reviewers: computing efficiency and zero-shot binaural audio rendering. The rest of the questions and comments are addressed individuall... | NeurIPS_2024_submissions_huggingface | 2,024 | null | null | null | null | null | null | null | null |
Finding NeMo: Localizing Neurons Responsible For Memorization in Diffusion Models | Accept (poster) | Summary: This paper introduces NeMo, a straightforward and easy-to-understand method. NeMo identifies neurons responsible for memorization and deactivates them, offering potential applications in the protection of sensitive or copyrighted data.
Strengths: This topic is highly interesting as it enhances interpretabilit... | Rebuttal 1:
Rebuttal: **Providing more compelling reasons for the unique position of this method in mitigating memorization.**
> The novelty and positioning of our research is multi-faceted: NeMo is the first method to localize memorization in diffusion models down to individual neurons. Previous research has generally... | Summary: This paper introduces NEMO, a method that identifies and deactivates specific neurons responsible for memorizing training data in DMs. This approach prevents the reproduction of sensitive images, enhances output diversity, and mitigates data leakage, enabling more responsible use of DMs.
Strengths: 1. It expl... | Rebuttal 1:
Rebuttal: **The author should specify their novelty compared with [1]. The novelty seems limited considering the conclusion in [1].**
> We believe there is a reference missing here in the review. Could you please specify which paper we should compare our method to? Reference [1] in our paper does not s... | Summary: The paper localize memorized samples to neurons in the cross attention layers in diffusion models. By deactivating neurons responsible for memorization, the proposed method enables models to generate diverse images different from training images.
Strengths: 1. The authors propose the first method to localize ... | Rebuttal 1:
Rebuttal: **My expectation from "localization of memorization" was that there is a specific subset of neurons that, when blocked, avoid replication of training data. Such a localization would not require the authors to develop a pipeline to detect memorization, as those neurons can be pruned.**
> It would ... | Summary: This paper introduces NEMO, a method for localizing memorization in diffusion models (DMs) down to the level of individual neurons. The authors empirically evaluate NEMO on the Stable Diffusion model and demonstrate that deactivating the identified memorization neurons effectively mitigates memorization, incre... | Rebuttal 1:
Rebuttal: **Comparison to other mitigation methods?**
> We have already quantitatively compared our method with the state-of-the-art mitigation approach by Wen et al. (ICLR 2024) in Section 4.1 and Table 1. Additionally, in Appendix C.9, we compared our method to the random token mitigation strategy pr... | Rebuttal 1:
Rebuttal: We thank all reviewers for their time spent reviewing our paper and their valuable feedback. We very much appreciate all the reviewers tend to accept the paper and think that the paper proposes an innovative and novel approach (Rr34, zaL8), presents interesting findings (zaL8, RyXV), provides a be... | NeurIPS_2024_submissions_huggingface | 2,024 | null | null | null | null | null | null | null | null |
EduQate: Generating Adaptive Curricula through RMABs in Education Settings | Reject | Summary: This paper proposes a variant of Multi-Armed Bandits (MAB) named EdNetRMABs that model the interdependencies between learning contents to meet the real-world education scenarios. Subsequently, the authors introduce EduQate, an interdependency-aware Q-learning algorithm to optimize content recommendation given ... | Rebuttal 1:
Rebuttal: We regret that Reviewer MTkZ did not find our work to be a meaningful contribution. We hope that our responses have addressed the reviewer's concerns. We welcome the opportunity for further discussion to clarify any doubts and provide additional insights into our work. Additionally, we would like ... | Summary: The paper introduces **EduQate**, a system that generates adaptive educational curricula using restless multi-armed bandits (RMABs). This method aims to efficiently achieve mastery across multiple interdependent educational contents. Unlike traditional methods that assume learning contents are independent, Edu... | Rebuttal 1:
Rebuttal: We thank the reviewer for their feedback and aim to address them. We thank the reviewer for pointing out formatting errors and will clarify them in the future iteration.
## Questions
1. On simplified modelling:
There seems to be a misunderstanding here. In our formulation of EdNetRMABs, arms on... | Summary: The paper proposes EduQate, an innovative framework that leverages EdNetRMABs to achieve interdependence among knowledge points. By using Q-learning, EduQate implements optimal strategies for personalized learning, offering optimality guarantees without needing explicit knowledge of transition functions govern... | Rebuttal 1:
Rebuttal: We thank the reviewer for the valuable feedback and the formatting concerns will be fixed.
## Questions
1. On implementing more RL baselines:
Please note that a key component of problems of interest are the presence of the arms (topics). Modeling such problems in RL is not trivial, as the sta... | Summary: This paper proposes a solution to generate personalized learning curricula in educational settings, focusing on the challenge of accounting for interdependencies between learning topics. It argues that existing approaches, often based on the Restless Multi-Armed Bandit (RMAB) framework, fall short by assuming ... | Rebuttal 1:
Rebuttal: We thank the reviewer for the positive feedback and comments. We address some of their questions and concerns below:
## Questions
1. Can EduQate's recommendations be easily interpreted?
Reviewers point on interpretability is well noted. Given that RMABs are typically simple models, where arms h... | Rebuttal 1:
Rebuttal: We thank the reviewers for their time and positive feedback! We are pleased to know that the reviewers found our work significant and an important contribution to the field of education technology (Reviewers 4CnK, Yn9L, fqUb, KG3r).
We wish to point out that the IB metric is not introduced by us... | NeurIPS_2024_submissions_huggingface | 2,024 | Summary: This paper presents a way to extend RMAB with Q-learning to accommodate the fact that some items/arms belong to a group. RMAB can be considered as a weakened version of contextual bandit CB but also a strong version of CB since it considers state transitions (explicitly defined on arms).
Strengths: This is a ... | Rebuttal 1:
Rebuttal: We thank the reviewer for his positive feedback and comments, and address their questions below:
##Questions
1. On releasing code/simulators
We will share our code for community use.
2. On Equation (1): Does summing over the pseudo-actions change the optimality of Whittle Index? Is Theorem... | null | null | null | null | null | null |
MetaLA: Unified Optimal Linear Approximation to Softmax Attention Map | Accept (oral) | Summary: Proposes a unifying view of recent linear attention/SSM/linear RNN methods and compares these models to Softmax Attention. Proposes MetaLA to address the shortcomings of the prior methods in approximating Softmax attention. Performs experiments on MQAR, language modeling, image classification and LRA.
Strengt... | Rebuttal 1:
Rebuttal: Thanks for your insightful feedback. We will outline your suggestions and questions, followed by our detailed responses. We hope our answers address your concerns.
>*Weakness1:* The general presentation of the "optimal linear approximation" and theoretical analysis is dense and confusing and obsc... | Summary: The paper presents a theoretical analysis of existing linear attention methods such as LinFormer, SSM, and LinRNN. Building on this analysis, the authors propose a unified framework that combines the strengths of these methods. Utilizing this framework, authors develop a novel linear attention model called Met... | Rebuttal 1:
Rebuttal: Thanks for your insightful feedback and your time in reading our paper.
>*Weakness1:* More long sequence validation besides the LRA benchmark. This is not sufficient to justify the scalability of the method for real LLMs.
**A:** Thank you for the suggestion. In addition to the LRA benchmark, w... | Summary: They proposed MetaLA, which solved problems of previous attention alternatives (LinRNN, SSM, LinFormer). They start to build their MetaLA by deriving general form of linear alternative of softmax attention.
1. They remove the K matrice redundant parameters and achieve better training efficiency
2. Add self-au... | Rebuttal 1:
Rebuttal: Thank you for your insightful feedback. Your questions are crucial and something we have been thinking about.
>*Weakness1:* The experiment is not done with various scales. It is hard to know the effect of the scaling model and training dataset with MetaLA.
**A:** Based on your suggestions, we co... | Summary: This paper discusses the development and evaluation of linear complexity models for Transformers, which aim to replace the conventional softmax attention mechanism. The authors first unify existing models including LinFormer, SSM, and LinRNN into the framework of Linear Attention. Then they establish three (ac... | Rebuttal 1:
Rebuttal: Thanks for your insightful feedback and your time in reading our paper.
>*Weakness 1:* Concerns about the optimality for linear attention.
**A:** Your concern is valid. There is no definitive evidence to prove that MetaLA is the optimal approximation of self-attention. Please allow me to explain... | Rebuttal 1:
Rebuttal: Dear ACs and Reviewers,
We would like to extend our sincere gratitude to all the reviewers for taking the time to read our paper and offering insightful suggestions. Linear models have emerged as a promising alternative to transformers, garnering significant interest within the foundational model... | NeurIPS_2024_submissions_huggingface | 2,024 | null | null | null | null | null | null | null | null |
AgentPoison: Red-teaming LLM Agents via Poisoning Memory or Knowledge Bases | Accept (poster) | Summary: The paper proposes a new setting for red teaming against RAG empowered agents. Specifically, they tested their method in three different settings: driving agent, knowledge intensive QA and EHRagent. The method they proposed is to use constrained optimization to jointly optimize for two objectives: "a) the retr... | Rebuttal 1:
Rebuttal: Dear Reviewer dD1S,
Thank you for the valuable questions and comments! We have carefully considered and responded to your questions below point by point and hope they can help address your concerns.
**W1**:
Actually, $\eta_{tar}$ and $\eta_{coh}$ do have an intuitive value and can be adaptab... | Summary: This paper proposed an attack against the RAG-based LLM agents. Specifically, it proposes a constrained trigger optimization to search the trigger so that any queries included can be mapped to a certain compact cluster while keeping the coherence and attack success rate. The experiments show that the proposed ... | Rebuttal 1:
Rebuttal: Dear Reviewer XDTd,
We are really grateful for the valuable questions and comments you raised to help us improve the work! Below we have detailed our response point by point and hope they can help address your concerns.
**W1:**
Firstly, the goal of AgentPoison is to red-team LLM agents to asse... | Summary: The paper titled "AGENTPOISON: Red-teaming LLM Agents via Memory or Knowledge Base Backdoor Poisoning" introduces a novel backdoor attack method targeting large language model (LLM) agents. These agents leverage a memory module or a retrieval-augmented generation (RAG) mechanism to retrieve knowledge and past ... | Rebuttal 1:
Rebuttal: Dear Reviewer UuDZ,
Thank you very much for your appreciation of the novelty and effectiveness of our work and its contribution to uncovering the safety and trustworthiness of LLM agents! To answer your question, we have provided below a theoretical analysis of the sample complexity of approximat... | Summary: The paper presents AGENTPOISON, a novel red-teaming approach aimed at exposing vulnerabilities in LLM agents by poisoning their long-term memory or RAG knowledge base. Unlike conventional backdoor attacks, AGENTPOISON does not require additional model training or fine-tuning and ensures high attack success rat... | Rebuttal 1:
Rebuttal: Dear Reviewer 2UC7,
Thank you for your appreciation of our work! Below we have provided a point-by-point response to your questions and hope they could help address your concerns.
**W1**:
We totally agree with the reviewer that a more comprehensive discussion on possible mitigation strategies a... | Rebuttal 1:
Rebuttal: Dear reviewers, area chairs, and program chairs,
Thank you for your interest and appreciation of our work and valuable comments to help improve our paper! We have addressed the concerns and questions of each reviewer point-by-point in each of our individual rebuttal responses. To provide more con... | NeurIPS_2024_submissions_huggingface | 2,024 | Summary: The paper presents a novel red-teaming approach that targets the vulnerabilities of large language model (LLM) agents by poisoning their long-term memory or retrieval-augmented generation (RAG) knowledge base. The primary contribution is the development of AGENTPOISON, a method to inject backdoor triggers into... | Rebuttal 1:
Rebuttal: Dear Reviewer jtC2,
Thank you very much for your appreciation of the novelty and effectiveness of our work and its contribution to uncovering the safety and trustworthiness of LLM agents! In fact, you have accurately summarized our paper and followed the key ideas that we proposed.
In addition, ... | null | null | null | null | null | null |
Interactive Deep Clustering via Value Mining | Accept (poster) | Summary: This paper proposes to incorporate user interactions to deal with the hard samples in deep clustering. By asking the user for the cluster affiliations of boundary samples, the proposed framework improves the performance of existing clustering methods.
Strengths: 1. The hard boundary samples are indeed the per... | Rebuttal 1:
Rebuttal: We appreciate your positive feedback and suggestions. Below we provide the point-by-point response to your concerns.
**Weakness 1: Inconsistent $\lambda_1, \lambda_2$ for different clustering models**
We agree that it is daunting to tune the hyper-parameters for unsupervised tasks. In the previ... | Summary: This paper proposes a way to integrate user-feedback to fine-tune clustering models and improve their performance on hard samples, i.e. samples near cluster boundaries. In particular, the authors propose a score to value and select samples for user-interaction. This score is a combination of *hardness*, *repre... | Rebuttal 1:
Rebuttal: Thanks for your recognition of the problem this work aims to address, as well as our simple yet effective idea. We sincerely appreciate your insightful and constructive comments. Below are the point-by-point responses to your concerns raised in weaknesses, questions, and limitations.
**Weakness 1... | Summary: This work proposes an interactive deep clustering framework IDC, which could be integrated with existing deep clustering methods. The key idea is to adjust the decision boundary by querying the cluster affiliations of high-value samples. The authors applied IDC to two pre-trained clustering models on five data... | Rebuttal 1:
Rebuttal: We appreciate your positive feedback and kind suggestions for our work. Below are the point-by-point responses to your concerns mentioned in the weaknesses and limitations sections.
**Weakness 1 & Limitation 1: Imperfect user feedback**
According to your advice, we asked three people to answer ... | Summary: In this work, the authors propose a plug-and-play module to boost the clustering performance of existing methods through user interaction. A sample value evaluation criterion is designed to propose valuable user queries of a high performance-to-cost ratio. Experiments show that the proposed module could signif... | Rebuttal 1:
Rebuttal: Thanks for your positive feedback and detailed suggestions for our work. Below are the point-by-point responses to your concerns.
**Weakness 1: The relative performance improvement**
According to your comments, we added a baseline by manually correcting the cluster assignments of the 500 query s... | Rebuttal 1:
Rebuttal: We appreciate the reviewers for their insightful and constructive comments. We supply a PDF containing the supplementary figures and provide the point-by-point responses below.
---
*Due to the space limitation, we put part of the responses to Reviewer 2n5n in this section:*
**Weakness 3: Dataset... | NeurIPS_2024_submissions_huggingface | 2,024 | null | null | null | null | null | null | null | null |
Reprogramming Pretrained Target-Specific Diffusion Models for Dual-Target Drug Design | Accept (poster) | Summary: This is an interesting and promising work that combines state-of-the-art techniques in structure-based drug design, such as diffusion models, to address the challenging but important problem of dual-target drug design. The main contributions of the paper are:
1.Carefully curated a dataset based on synergistic... | Rebuttal 1:
Rebuttal: Thank you for your feedback. Please see below for our responses to the comments.
**Q1: Formal definition of the dual-target drug design problem, especially the pocket alignment, and consideration of different ligands and the quality of the alignment.**
A1: We have provided the definition of the ... | Summary: This paper addresses the challenge of designing dual-target drugs, a promising strategy in overcoming drug resistance in cancer therapy. The authors propose leveraging the success of deep generative models in structure-based drug design, formulating dual-target drug design as a generative task. They introduce ... | Rebuttal 1:
Rebuttal: Thank you for your detailed feedback. Please see below for our responses to the comments.
**Q1: "The proposed method involves complex geometric deep learning techniques, which may pose implementation challenges for researchers less familiar with these methods."**
A1:
To understand our method, r... | Summary: The paper presents a comprehensive approach for dual-target drug design by reprogramming pretrained single-target diffusion models. It introduces a curated dataset derived from synergistic drug combinations, and two methods, COMPDIFF and DUALDIFF. The experimental results demonstrate the outperformance of the ... | Rebuttal 1:
Rebuttal: Thanks for your positive feedback. Please see below for our responses to the comments.
**Q1: Discussion about the generalization of this framework.**
A1:
Our proposed framework is **general** and **any types of pretrained generative models** for structure-based drug design can be applied to dua... | Summary: This work studies a dual-target drug design using diffusion models trained on single-target protein-ligand complex pairs.
There two proposed methods, COMPDIFF and DUALDIFF, both are aim to align and generate dual-target ligands using SE(3)-equivariant composed message passing. The paper also introduces a cur... | Rebuttal 1:
Rebuttal: Thank you for your positive feedback. Please see below for our responses to the comments.
**Q1: Further studies of the alignment motivated by measurement theories or so.**
A1:
The alignment of the dual pockets is dependent on the molecule selected to compute the alignment transformation (i.e., ... | null | NeurIPS_2024_submissions_huggingface | 2,024 | null | null | null | null | null | null | null | null |
SampDetox: Black-box Backdoor Defense via Perturbation-based Sample Detoxification | Accept (poster) | Summary: The paper addresses the issue of backdoor threats in third-party Machine Learning as a Service (MLaaS) applications. The authors propose a novel two-stage method called SampDetox to defend against both visible and invisible backdoor attacks without requiring access to the model's parameters or training data. T... | Rebuttal 1:
Rebuttal: **AW1:**
Thanks for your suggestion. We conducted new experiments to evaluate the performance of SampDetox considering the case where pre-training datasets for diffusion models are out-of-distribution from their target classification tasks. The detailed experimental setups are as follows:
* We co... | Summary: The paper addresses the significant issue of backdoor defense in a black-box setting during the inference phase of MLaaS applications. It proposes a novel technique called SampDetox, where the method uses diffusion models for detoxification by injecting noise to disrupt backdoors and retrieve clean instances.
... | Rebuttal 1:
Rebuttal: **AW1:**
Thanks for your suggestion.
First, it should be clarified that the defender completely controls the design, training, and utilization of diffusion models. Therefore, it is almost impossible for attackers to inject backdoors into diffusion models in practice.
In addition, even if the trai... | Summary: This paper proposes a two-stage black-box method to defend against the backdoor in DNN. The first stage mitigates the backdoor triggers with low visibility and the second stage mitigates the trigger with higher robustness.
Strengths: 1. The two-stage design is novel and effective.
2. The evaluation is compre... | Rebuttal 1:
Rebuttal: **AW1:**
Indeed. In Appendix D.3 (Limitations of SampDetox), we have discussed a type of backdoor attack that SampDetox cannot defend against. Specifically, when dealing with triggers derived from semantic features that are part of the original classified features and whose distribution is the sam... | Summary: This paper introduces a novel black-box backdoor purification technique consisting of two stages. In the first stage, it eliminates global perturbation by adding lightweight noise to the inputs and then leverages DDPM for denoising. Subsequently, to address visible and robust triggers, it identifies the local ... | Rebuttal 1:
Rebuttal: **AQ1:**
Thanks for your suggestion.
Since the semantic features adopted by Reflect [1] and DSFT [2] are neither part of the original classified features nor whose distribution is the same as the distribution of the training dataset of the pre-trained diffusion model, the triggers based on such se... | null | NeurIPS_2024_submissions_huggingface | 2,024 | Summary: This paper proposes SampDetox, a novel black-box backdoor defense method using perturbation-based sample detoxification. The key contributions are:
1. A preliminary study revealing the correlation between trigger visibility and poisoned sample robustness.
2. A two-stage defense approach combining lightweight ... | Rebuttal 1:
Rebuttal: **AW1:**
Indeed, the effectiveness of SampDetox depends on our observed correlation. However, this does not mean that our approach cannot prevent more sophisticated attacks. This is because our observation has good generalization ability, since our observation is based on ten SOTA backdoor attacks... | null | null | null | null | null | null |
Harnessing small projectors and multiple views for efficient vision pretraining | Accept (poster) | Summary: This paper investigates a theoretical formulation of the contrastive loss used in self-supervised pretrainging of vision models. (One of the assertions of the paper is that all SSL (self-supervised learning) losses are variants of the same loss). The authors propose a more compute-efficient version of the loss... | Rebuttal 1:
Rebuttal: Thank you for your thorough review and valuable suggestions for improving our paper. We appreciate the time and effort you've invested in providing this feedback. We acknowledge the shortcomings in our initial presentation, particularly the absence of a comprehensive overview of existing self-supe... | Summary: The paper conducts a theoretical study the training dynamics of self-supervised learning (SSL) methods by using an idealized loss function that resembles the Barlow Twins/VICReg loss. The idealized loss function is built on the idea of forward data augmentation (\emph{DAF}) graph kernel. The paper shows how th... | Rebuttal 1:
Rebuttal: We thank the reviewer for their thorough and insightful comments. We are pleased that the reviewer found our analysis of the SSL loss formulation and the training dynamics exciting and relevant to the community. Below, we address the questions raised by the reviewer with enhanced clarity and addit... | Summary: This paper identifies the implicit bias of non-contrastive SSL loss and optimization, and proposes two ingredients to improve SSL learning: 1) Low-dimensional projectors can yield good representations; 2) Multiple augmentations improve kernel approximation. Further, the authors propose that In a low-data regim... | Rebuttal 1:
Rebuttal: We would like to thank the reviewer for their thorough and insightful comments. We are also glad the reviewer found our proposal theoretically motivated and the empirical validation satisfactory. Below, we address the questions raised by the reviewer:
1. **Definitions of weak/strong orthogonality... | Summary: This work introduces a new loss for self-supervised retraining that is functionally equivalent to existing methods but is computationally more efficient. The method identifies a key equivalence between existing SSL objectives (VICReg and BarlowTwins) features when employing augmentation kernel. Furthermore, th... | Rebuttal 1:
Rebuttal: Thank you for your thorough and insightful review of our work. We greatly appreciate your time and effort in assessing our paper and providing constructive feedback.
We are delighted you found the paper well-written, clearly presented, and understandable to a broad audience. It's delightful to hea... | Rebuttal 1:
Rebuttal: We would like to thank all the reviewers for their thorough and insightful comments and suggestions. Below, we summarize the major points addressed in individual responses to the reviewers' comments.
1. **Inclusion of Results for ImageNet-100:**
We have added results for ResNet-18 pretrained usi... | NeurIPS_2024_submissions_huggingface | 2,024 | null | null | null | null | null | null | null | null |
Fast Rates in Stochastic Online Convex Optimization by Exploiting the Curvature of Feasible Sets | Accept (poster) | Summary: - This paper studies online convex optimisation on feasible sets with curvature conditions. They introduce a new way to exploit the curvature of the feasible set through sphere enclosure. For feasible sets satisfying this condition, they leverage results from universal online learning to show that there exists... | Rebuttal 1:
Rebuttal: We appreciate your valuable time and thorough review. We will correct the minor comments and typos in the revised version. Here are our responses to the review.
**Rebuttal to Weaknesses**
> The improvement over FTL is mostly in the stochastic environment. In Theorem 13, if we consider the fully ... | Summary: In this work, the authors study stochastic online convex optimization (SOCO), and establish a tighter regret bound of $O(\log T)$ for sphere-enclosed sets. Specifically, such a regret bound is derived by exploiting an immediate result of several existing universal online learning algorithms. Moreover, the auth... | Rebuttal 1:
Rebuttal: Thank you very much for your valuable time to give the insightful review. Below are our responses to the review.
**Rebuttal to Weaknesses**
> First, compared with stochastic (offline) optimization, the main challenge of online convex optimization (OCO) is how to deal with a fully adaptive advers... | Summary: The paper relates the regrets of (stochastic) online convex optimization to the geometry of the feasible set. Precisely, the paper proposes the sphere-enclosed-feasible-set property which measures the local curvature condition of the feasible set at the offline optimal solution, and shows that many existing on... | Rebuttal 1:
Rebuttal: We are grateful for your valuable time and insightful review.
Below are our responses to the review.
> In addition, the lower and upper bounds provided appear to hold under slightly different conditions (see the question below).
and
> I’m wondering if there is any connection between the \lambd... | Summary: This paper considers the influence of feasible set geometry on the OCO regret. For a class of algorithms satisfying certain regret upper bounds (including MetaGrad and Maler), the authors provide new fast-rate results in the stochastic convex loss setting and the corrupted convex loss setting. Their results al... | Rebuttal 1:
Rebuttal: Thank you for your valuable time and careful review. We will address the minor comments and typos in the revised version. Below are our responses to the review.
**Rebuttal to Weaknesses**
> One possible limitation of the results is that, unlike the OLO results, the action set curvature-dependent... | null | NeurIPS_2024_submissions_huggingface | 2,024 | null | null | null | null | null | null | null | null |
A Simple yet Universal Framework for Depth Completion | Accept (poster) | Summary: This paper proposes a novel depth completion framework that leverages a large model for monocular depth estimation (i.e., MiDAS, called also Visual Foundation Model -- VFM) to achieve depth completion across a wide range of scenes and sensors (here defined as "UniDC problem").
The UniDC baseline architecture a... | Rebuttal 1:
Rebuttal: ### [bXoG] W.1 Comparison with VFM-less Methods with
We acknowledge the reviewer's concern regarding the fairness of comparing our method with VFM-less approaches. We recognize that Visual Foundation Model (VFM) features provide a significant advantage due to their extensive pre-training on large ... | Summary: This paper proposes Universal Depth Completion as a new task and provides a solution that tackles the problem. The proposed approach utilizes a monocular depth foundation model for its general understanding of 3D scenes and then completes depth information from different sensors with a learned affinity map. Th... | Rebuttal 1:
Rebuttal: ### [AS7d] W.1 Experiments on other RGBD datasets
We appreciate your feedback regarding the limited use of datasets in our evaluation. In response to concerns about unseen domain generalization, we have expanded our experimental evaluation to include an additional dataset, SUN-RGBD, which contains... | Summary: The paper introduces a universal depth completion framework that aims to resolve two challenges: "generalizable knowledge" of unseen scenes and "adaptation ability" to arbitrary depth sensors.
Unlike the previous method relying on extensive pixel-wise labeled data, the proposed method 1) utilizes a foundation... | Rebuttal 1:
Rebuttal: ### [g2HF] W/L. Comparison with other methods and experiments about unseen domain generalization.
As requested by the reviewer, we conduct additional experiments on diverse datasets beyond KITTI DC and NYU v2. We include tests on the SUN RGB-D, which show different environmental conditions and var... | Summary: This paper proposes a universal depth completion method (UniDC) to address generalization issues in unknown scenes and arbitrary depth sensors. The method utilizes depth information extracted from a pre-trained monocular depth estimation model to generate pixel-wise affinity maps, which adjust sparse depth. Ad... | Rebuttal 1:
Rebuttal: ### [tAUP] W.1 Online KITTI depth completion benchmark.
We thank the reviewer for the comments about the outstanding performance and generalization ability of our method. Following your comment, we report the performance of our work in the KITTI benchmark, which is reported in Figure.A of the uplo... | Rebuttal 1:
Rebuttal: We thank all reviewers for their helpful comments which make our paper polish up. If reviewers want to see additional results and analysis, please let us know it and we are always welcome to discuss.
For additional experiments requested by multiple reviewers, we report them here.
- [Table.A] Com... | NeurIPS_2024_submissions_huggingface | 2,024 | null | null | null | null | null | null | null | null |
The Importance of Being Scalable: Improving the Speed and Accuracy of Neural Network Interatomic Potentials Across Chemical Domains | Accept (poster) | Summary: This paper introduces the Efficient Graph Attention Potential (EGAP), a new architecture for Neural Network Interatomic Potentials (NNIP) designed for scalability and efficiency. The authors investigate scaling strategies for NNIP models and propose a model that leverages optimized self-attention mechanisms. T... | Rebuttal 1:
Title: Rebuttal by Authors
Comment: Thank you for your constructive feedback and valuable suggestions! We address your concerns as follows:
**Question: dataset split and inconsistent dataset usage of MD22**
Thank you for the suggestions. We have revised our MD22 experiments to have the same train/val spli... | Summary: Following an investigation into the scaling of an attention-based NNIP model (EquiformerV2), this paper finds that it is better to increase the number of parameters in the attention mechanism than to increase the number of parameters using higher-order tensors. Based on these results, the paper proposes a mode... | Rebuttal 1:
Rebuttal: Thank you for your constructive feedback and valuable suggestions! We address your concerns as follows:
**Question: scalability of the model, and performance**
Thank you for the suggestions. We agree with your definition of scalability, and we will clarify it further in the introduction of the f... | Summary: Authors explore scaling strategies for ML interatomic potentials, which are alternatives to the increase of spherical order L, which is costly. By leveraging the insights they extract from ablations, and avoiding the inefficiency of equivariance, they build a new invariant model based on features containing an... | Rebuttal 1:
Rebuttal: Thank you for your positive feedback and valuable suggestions! We address your concerns as follows:
**Question: more experiments and baselines**
We have updated the results of EGAP on OC20 2M and All+MD split, with comparisons of relevant baselines. We also added another experiment on the OC22 d... | Summary: The paper presents a novel approach to predict properties of some 3d molecules based on efficient transformer-based graph neural networks. The paper investigates the computational bottlenecks of the previous approaches and addresses them by using simpler features and designing a novel graph attention architect... | Rebuttal 1:
Rebuttal: Thank you for your constructive feedback and valuable suggestions! We address your concerns as follows:
**Question: inefficient attention over edges**
We clarify the details about the architecture of our model: in the realm of NNIPs, most of the molecular graph is constructed by a radius graph w... | Rebuttal 1:
Rebuttal: # Overall Response to All Reviewers
We thank all the reviewers for their constructive feedback. In our general response, we address some common issues discussed by the reviewers and provide updated experimental results for these points:
**State-of-the-art results on OC20-2M and OC20-All+MD.** To... | NeurIPS_2024_submissions_huggingface | 2,024 | null | null | null | null | null | null | null | null |
Federated Learning from Vision-Language Foundation Models: Theoretical Analysis and Method | Accept (poster) | Summary: In this papaer, the authors proposed a method to finetune vision-language model (CLIP) via prompt learning. Specifically, the author optimize a global textual prompt which will be optimized via FedAvg, while each client additionally optimize a local textual prompt. The proposed method is a combination of the e... | Rebuttal 1:
Rebuttal: Thank you for your feedback and for mentioning that our theoretical analysis is "thorough and clear."
> Weakness 1: The proposed method is basically a combination of both CoOp and PromptFL, or to be more specific, optimizing both client-specific and client-agnostic textual prompt. The novelty of ... | Summary: This paper discusses the integration of pretrained vision-language foundation models, such as CLIP, into federated learning. The idea is to use prompt-based federated learning to minimize the communication and computational costs. The authors go into the theoretical analysis to understand the performance of th... | Rebuttal 1:
Rebuttal: Thank you for your appreciation of our approach and consider that our paper is "useful for the community". Below, we address each of your points in detail.
> Q1(1): While the overall approach is quite interesting, the authors restrict their investigation in a much simpler classification task.
F... | Summary: This paper studies theoretical properties of prompt based federated learning methods for visual language models. Following the proposed theoretical results, a new algorithm named Global-local Prompt Portfolio for Federated Learning (PromptFolio) was proposed. The proposed algorithm was examined in image classi... | Rebuttal 1:
Rebuttal: We greatly appreciate your feedback and inquiries. We're thankful that you consider our examination of the connection between federated prompt learning and portfolio optimization to be "a nice idea".
> Question 1: In the algorithm, each client sends prompts to a server. How do assure privacy prot... | Summary: The proposed methodology offers a novel take on analyzing federated learning using prompt learning (for foundation models) via feature learning theory. At its core, the idea is to identify and monitor task-relevant and task-irrelevant features, and leveraging inspiration from portfolio optimization which says ... | Rebuttal 1:
Rebuttal: Thank you for your comment. We will provide a detailed, item-by-item response to these concerns.
> Abstract/ Conceptual Question 1: Could the authors comment about the few-shot learning aspect of foundation models (+prompt learning) and how the optimal mixing coefficient would be affected in that ... | Rebuttal 1:
Rebuttal: We sincerely thank all the reviewers for their time and effort in evaluating our work and providing constructive comments. We appreciate that the reviewers consider our theoretical framework for federated prompt learning to be "pretty solid" (BdA2), "thorough and clear" (YEHg), and "novel" (bKME).... | NeurIPS_2024_submissions_huggingface | 2,024 | Summary: This paper tackles the interesting problem of analyzing federated learning (FL) from vision-language foundation models like CLIP. The authors develop a theoretical framework based on feature learning theory to understand how prompt-based FL works. They introduce a new algorithm called PromptFolio that mixes gl... | Rebuttal 1:
Rebuttal: We sincerely thank you for your comments and questions. We appreciate that you found our feature learning approach "solid," the connection between prompt mixing and portfolio optimization "novel," and our ablation studies "informative."
> Weakness 1: This reviewer is confused by the feature learn... | null | null | null | null | null | null |
Preventing Dimensional Collapse in Self-Supervised Learning via Orthogonality Regularization | Accept (poster) | Summary: Driven by the dimensional collapse phenomenon observed in self-supervised learning, this research adopts a methodology designed to prevent the reduction in dimensionality of weight matrices. The empirical findings indicate a reduction in dimensional collapse in certain instances.
Strengths: The proposed metho... | Rebuttal 1:
Rebuttal: We appreciate your comments and feedback. In addition to the general response, we address your itemized concerns here.
>The proposed method lacks novelty, essentially applying the loss formulation from Barlow Twins to the weight matrix.
Methods like Barlow Twins try to avoid dimensional collapse... | Summary: In order to combat the problem of dimensionality collapse in self-supervised learning, the authors of this work introduce orthogonal regularization on encoders during pretraining. Taking inspiration from previous work on supervised orthogonality regularization, the authors study the impact of OR on eigenvalues... | Rebuttal 1:
Rebuttal: We appreciate your comments and feedback. In addition to the general response, we address your itemized concerns here.
>I would like to point out the improvements of OR on SSL methods is not as major on the simple datasets such as Cifar100 set according to Table 1\. However, the changes in larger... | Summary: This work proposes a method to reduce the dimensional collapse issue in self-supervised pretrained networks. Dimensional collapse occurs when a few large eigenvalues dominate the eigenspace. While previous research focused solely on the output representations of the pretrained networks, this work shows that di... | Rebuttal 1:
Rebuttal: We appreciate your comments and feedback. In addition to the general response, we address your itemized concerns here.
>(W1)
Thank you for your very constructive suggestions\! Following your suggestions, we added more experiments of BYOL with vic loss (feature whitening their predictor's output)... | Summary: This paper aims to improve self-supervised learning (SSL) by introducing orthogonality regularization (OR) during pretraining to prevent dimensional collapse. OR improses orthogonality constraint on weight matrices, not only preventing dimensional collapse of representation but also regularizing the weights fr... | Rebuttal 1:
Rebuttal: We appreciate your comments and feedback. In addition to the general response, we address your itemized concerns here.
>More discussion on the difference between the proposed orthogonal regularization methods would have been helpful. On the surface, the two methods are almost equally good.
Thank... | Rebuttal 1:
Rebuttal: We thank all reviewers for their questions and constructive feedback. Here, we respond to the five core issues of common interest:
**Solo-learning baseline settings**
The baseline results of our report are the results of the newest official config file, which are different from the checkpoint ... | NeurIPS_2024_submissions_huggingface | 2,024 | Summary: Dimensional collapse, characterized by the dominance of a few large eigenvalues, is a phenomenon in self-supervised learning (SSL) that can lead to redundant features and weight matrices. To address this, the authors propose orthogonal regularization (OR) during pretraining to promote orthogonality in weight m... | Rebuttal 1:
Rebuttal: We appreciate your comments and feedback. In addition to the general response, we address your itemized concerns here.
> The soundness of some experiments
Thanks for pointing out this. Solo-learn only provides the official config file (yaml) for resnet18 on cifar100, on which we replaced the back... | null | null | null | null | null | null |
Conformal Inverse Optimization | Accept (poster) | Summary: The paper proposes a robust optimisation method for decision problems that provides not only a point estimate but also an uncertainty set. The methodology is tested on decision making tasks such as shortest path and the knapsack problem and shows favourable results when compared to classical inverse optimisati... | Rebuttal 1:
Rebuttal: - **Data availability.** We thank the reviewer for recognizing the importance of the topic and the clarity and novelty of our paper. However, we would like to point out that data availability should not be a bottleneck in practice. For example, consider the ridesharing example we describe in Secti... | Summary: This paper proposes conformal inverse optimization (CIO) which uses conformal prediction to learn uncertainty sets for parameter estimates in IO. These uncertainty sets are then used to formulate a robust version of the forward optimization problem. The authors provide intuition for their method and also provi... | Rebuttal 1:
Rebuttal: - **Running example.** We thank the reviewer for this suggestion. Throughout the paper, we can think of the forward problem as a shortest path problem where the decision $x$ corresponds to a path from an origin to a destination specified by $u$ and the parameters $\theta$ can be treated as the tra... | Summary: This paper considers a robust inverse optimization problem. The author provides an uncertainty set with probabilistic guarantees for the estimated parameters, illustrates the advantages of robust inverse optimization over deterministic inverse optimization through an example, and demonstrates its superiority i... | Rebuttal 1:
Rebuttal: We thank the reviewer for recognizing the novelty and clarity of our paper.
- **Uncertainty Set justification.** We explain the two constraints in Equation (9) below.
- **l2 norm = 1**. We point the reviewer to the discussion under “objective function” in Section 3.1. We focus on objective f... | Summary: In some supervised decision making, one usually leverage solution of optimization problem sequentially for future decisions. The authors argue that a single point estimation might not be enough and rather suggest estimating a whole uncertainty set on the model parameter and instead suggest a minimax robust opt... | Rebuttal 1:
Rebuttal: 1. Presentation.
- **The goal of Example 1 is to get through the intuition**. Since both reviewers UvTc and ggLM recognize its usefulness, we view the simplicity of example 1 as a strength. We suggest keeping the Lemmas and Proposition as they help to organize our results. Putting them under ... | null | NeurIPS_2024_submissions_huggingface | 2,024 | null | null | null | null | null | null | null | null |
SpeedLoader: An I/O efficient scheme for heterogeneous and distributed LLM operation | Accept (poster) | Summary: This paper propose a new scheduling mechanism via effective sub-batch centric computation, to facilitate the tensor computation and exchange, which is proven to particularly minimize the excessive communication overhead.
Strengths: 1. The idea of redesign the data flow and sharded model training under restric... | Rebuttal 1:
Rebuttal: ### Response to Question 1
Thank you for your question. We have prepared a detailed explanation of the one-shot hyperparameter tuning process below. Please review it and let us know if you have any questions, so we can integrate it into the camera-ready version if the paper is accepted.
The prima... | Summary: This paper proposes SpeedLoader, a system for offloading parameters and activations under restricted resources for distributed training and inference. It utilizes a tensor exchange manager to minimize the communication overheads. Emperically, a larger proportion of time is spent on computation rather than co... | Rebuttal 1:
Rebuttal: Response to Question
Thank you for your valuable feedback. While the diagrams may appear similar, SpeedLoader is fundamentally different from FlexGen. FlexGen is a significant work that substantially improves the inference throughput of LLMs by efficiently managing activation and cache tensors. ... | Summary: This paper proposes a new paradigm called SpeedLoader for large language models (LLMs) inference and training to offload and reload the layer weights and activations to and from CPU memory. Compared to prior works, SpeedLoader can process multiple batches with the current active layer, which thus able to reduc... | Rebuttal 1:
Rebuttal: ### Response to Question 1
Thank you very much for the kind and insightful review. Currently, 7B models at bfloat16 precision is very challenging to be trained on a 40GB A100. The model and optimizer states alone can take up to 56GB memory. One of the methods to deal with it is to shard the model... | Summary: This paper introduces an innovative compute strategy that minimizes I/O between peers and devices in heterogeneous LLM training, and a high-efficiency tensor manager that optimizes transfers between device and host, reducing fragmentation and redundancy. These enhancements lead to superior inference efficiency... | Rebuttal 1:
Rebuttal: ### Response to Question 1
Thank you very much for the thoughtful insights. The key concept behind SpeedLoader is to achieve **equivalent computation of a large batch** with a small memory consumption. An important feature of SpeedLoader is its management of activations. For instance, processing ... | Rebuttal 1:
Rebuttal: Dear Reviewers,
Thank you for your detailed and insightful feedback. We have conducted additional experiments to enhance the evaluation of SpeedLoader, making it more credible and comprehensive. Below, we address your concerns and clarify some points:
1. **Impact of Different Sequence Lengths Un... | NeurIPS_2024_submissions_huggingface | 2,024 | null | null | null | null | null | null | null | null |
Long-Horizon Planning for Multi-Agent Robots in Partially Observable Environments | Accept (poster) | Summary: The paper introduces the Long-Horizon Planner for Multi-Agent Robotics (LLaMAR), a cognitive architecture leveraging Language Models (LMs) for planning tasks in partially observable environments. Unlike traditional methods, LLaMAR uses a plan-act-correct-verify framework, enabling real-time self-correction and... | Rebuttal 1:
Rebuttal: Dear Reviewer,
Thank you for your comprehensive review and insightful feedback on our paper. We appreciate your thoughtful comments and suggestions, and we have provided our responses and clarifications below.
### Computational complexity
Yes, we agree that having multiple LM-based modules woul... | Summary: This paper presents a multi-agent planning and execution engine called LLaMaR that is based on several multi-modal foundation models such as GPT4V, i.e., large models (LMs) that can reason about text and visual inputs. LLaMaR has 4 LMs: planner, actor, corrector, and verifier. The planner decomposes the task ... | Rebuttal 1:
Rebuttal: Dear Reviewer,
Thank you for your comprehensive review and insightful feedback on our paper. We appreciate your thoughtful comments and suggestions, and we have provided our responses and clarifications below.
### LLaMAR explanation
The specific order of the modules is due to natural causal rel... | Summary: This paper focuses on the long-horizon multi-agent planning tasks. They propose a new agent structure composed of four language models (Planner, Actor, Corrector, Verifier), named Language-Model-based Long-Horizon Planner for Multi-Agent Robotics (LLaMAR). LLaMAR operates without prior knowledge or privileged... | Rebuttal 1:
Rebuttal: Dear Reviewer,
Thank you for your comprehensive review and insightful feedback on our paper. We appreciate your thoughtful comments and suggestions, and we have provided our responses and clarifications below.
### Single-agent comparison in MAP-THOR:
We perform experiments with just a single agen... | Summary: This paper explores the use of Language Models (LMs) to generate task plans for autonomous robots, emphasizing adaptability across diverse tasks without the need for extensive customization. The authors introduce LLaMAR, an LM-based agent framework that employs a “plan-act-correct-verify” strategy to manage lo... | Rebuttal 1:
Rebuttal: Dear Reviewer,
Thank you for your comprehensive review and insightful feedback on our paper. We appreciate your thoughtful comments and suggestions, and we have provided our responses and clarifications below.
### SentenceBERT
Yes, we agree with the concern about potential performance issues an... | Rebuttal 1:
Rebuttal: Dear reviewers,
We would like to thank you all for your comprehensive reviews and insightful feedback on our paper. We appreciate your thoughtful comments and suggestions. A few of the reviewers pointed out that some of the tasks do not require explicit cooperation within the multi-agent team and... | NeurIPS_2024_submissions_huggingface | 2,024 | Summary: The paper addresses the problem of long-horizon household tasks in a multi-agent setup using LLMs and for this, the authors propose LLaMAR. This framework integrates planning, acting, correcting, and verifying parts. LLaMAR plans a high-level action sequence, assigns them to multiple agents, and checks if acti... | Rebuttal 1:
Rebuttal: Dear Reviewer,
Thank you for your comprehensive review and insightful feedback on our paper. We appreciate your thoughtful comments and suggestions, and we have provided our responses and clarifications below.
### Justification for the Proposed Framework
The specific order of the modules is due... | null | null | null | null | null | null |
xMIL: Insightful Explanations for Multiple Instance Learning in Histopathology | Accept (poster) | Summary: This work proposed an LRP-based extension for MIL setup in histopathology. It focuses on explanation quality, not derivation of novel architectures.
Strengths: It is quite well written in terms of language.
Images are nicely depicting the key ideas.
Weaknesses: This paper does not follow the MIL paradigm. ... | Rebuttal 1:
Rebuttal: > Limited Novelty
Regarding your concerns about the novelty of applying existing XAI methods to MIL, we would like to clarify that it was not the aim of our work to develop a new explanation method for MIL. Instead, the novelty of our work extends across other dimensions:
1. **Methodical novelty... | Summary: The authors propose xMIL-LRP which is an explainable MIL method to overcome the shortcomings of previous instance score-based methods (such as attention-based). This can provide positive/negative evidence towards the label and can be slotted into existing MIL methods, since this is a post-hoc interpretability ... | Rebuttal 1:
Rebuttal: > Details on LRP
We acknowledge that some descriptions are too condensed. Therefore, we will add a section in the appendix with text and figures to properly describe LRP.
> Faithfulness experiments
We will clarify the faithfulness experiments and their motivation in the manuscript, with algori... | Summary: The paper introduces xMIL-LRP, an efficient solution for explainable Multiple Instance Learning (MIL) by incorporating layer-wise relevance propagation (LRP) into MIL. The authors further employ the AH-rule to define the "propagation rules" for relevance. Experiments are conducted on three toy datasets and fou... | Rebuttal 1:
Rebuttal: > Limited Novelty
Regarding your concerns about the novelty of applying existing XAI methods to MIL, we would like to clarify that it was not the aim of our work to develop a new explanation method for MIL. Instead, the novelty of our work extends across other dimensions:
1. **Methodical novelty... | null | null | Rebuttal 1:
Rebuttal: **Overview**
We thank the reviewers for their comments and valuable feedback. We have addressed their questions and concerns and made the following additions and changes to our manuscript:
* More clarity regarding the contributions of the work (reviewers utz5 & DaUT): We discussed our contributi... | NeurIPS_2024_submissions_huggingface | 2,024 | null | null | null | null | null | null | null | null |
Tracing Hyperparameter Dependencies for Model Parsing via Learnable Graph Pooling Network | Accept (poster) | Summary: This paper studies the model parsing (MP) task [1], which predicts the hyperparameters of the generative model with a generated image input. The authors propose a Learnable Graph Pooling Network (LGPN) method to explore hyperparameter dependencies in the generative model. The LGPN consists of two core designs:... | Rebuttal 1:
Rebuttal: We thank the reviewer for constructive feedbacks. Reviews are positive about our empirical performance and the learnable pooling algorithm, and we address all concerns raised as follows:
**Brief Recap:**
We adhere to the problem statement of model parsing [R3], in which the algorithm takes one i... | Summary: This paper examines the challenge of predicting the hyperparameters of a generative model from its output, specifically a generated image. The proposed approach introduces a learnable graph pooling technique, framing the problem as a node classification task. Experiments across several tasks demonstrate the ef... | Rebuttal 1:
Rebuttal: Thank the reviewer for recognizing our interesting research topic and easy-to-implement algorithm. In this response, we present clarification of the coordinated attack detection (**Q1**) and additional empirical results (**Q2**) that will be added in the revised version.
**Q1: Coordinated attack... | Summary: This paper focuses on the problem of model parsing, which aims to analyze the hyperparameters of the model that generates a given image. Different from existing model parsing methods, the proposed Learnable Graph Pooling Network (LGPN) exploits the dependencies among hyperparameters to better perform model par... | Rebuttal 1:
Rebuttal: We thank the reviewer for acknowledging our practical and novel research topic and the effectiveness of the proposed LGPN. In this response, we offer additional experimental results regarding GTC's effectiveness (**Q1**), a discussion of mistaken labeling (**Q2**), and clarification on the hyperpa... | Summary: This paper solves the model parsing task that requires models to predicting hyperparameters of the generative models (GMs) through one input image. The proposed model LGPN predict hyperparameters and their dependencies via directed graphs. LGPN incorporates a learnable pooling-unpooling mechanism to convert th... | Rebuttal 1:
Rebuttal: We appreciate the reviewer's praise for our method's effectiveness and reasonable algorithm design. Here, we present answers to all concerns raised.
**Q1: Difference between ours and two previous works:**
We highlight three key differences between our approach and those in references [R1, R2], ... | Rebuttal 1:
Rebuttal: We want to thank all four reviewers for their valuable comments. We are delighted to see (a) **all reviewers** recognize our proposed LGPN's effectiveness and high generalization ability, (b) our research topic of model parsing is interesting and novel (**Reviewer born** and **Q98w**), (c) praises... | NeurIPS_2024_submissions_huggingface | 2,024 | null | null | null | null | null | null | null | null |
Soft ascent-descent as a stable and flexible alternative to flooding | Accept (poster) | Summary: Flooding method is a previous method which aims at improving the generalization performance by making the average loss equal to a given threshold. This paper makes two changes of the 'flooding' method: 1. Rather than making the average loss equal to a given threshold, they make the point-wise loss equal to a g... | Rebuttal 1:
Rebuttal: Thanks for taking the time to read our paper, write a review, and for the positive outlook on our work. Below, we respond to the question you raised.
> Although the generalization gap of the proposed method is smaller (shown in Table 1), the accuracy is not larger than SAM (shown in Figure 4). Wh... | Summary: The flooding method sets a threshold for the average surrogate loss within a mini-batch during training: if above it do gradient descent as usual but if below it switch to gradient ascent. This paper updates the flooding method in 2 ways: 1) it invert the order of the sign function and the aggregation making i... | Rebuttal 1:
Rebuttal: We thank the reviewer for their time and constructive feedback provided in the review. Below, we respond to the points raised.
__Regarding iFlood (ICLR 2022):__
Thank you very much for this reference. We were completely unaware of this work, but it is a very nice paper, and as the reviewer point... | Summary: This paper proposes a method called SoftAD (soft ascent-descent) which aims to improve the "flooding" method. Specifically, it downweights points on the borderline, limits the effects of outliers, and retains the ascent-descent effect of flooding, with no additional computational overhead. Stationarity guarant... | Rebuttal 1:
Rebuttal: Thanks to the reviewer for their time in reading the paper and writing the review. Below, we respond to the points raised.
> Please give the situations where we are interested in other metrics such as model complexity or average surrogate loss at test time.
This is an important point, thanks. Wh... | Summary: The paper proposes a new optimization technique, Soft Ascent-Descent (SoftAD), aimed at enhancing the stability and performance of machine learning models. The method balances loss minimization and model complexity, positioning itself as an improvement over existing techniques like Flooding and SAM. The paper ... | Rebuttal 1:
Rebuttal: We thank the reviewer for their time in carrying out this review. Below, we respond to the questions raised.
---
> 1. How do you suggest optimizing the threshold parameter $\\theta$ in practice? Are there any heuristics or automated methods that can be employed?
The ideal setting in practice wi... | null | NeurIPS_2024_submissions_huggingface | 2,024 | null | null | null | null | null | null | null | null |
Scaling Laws and Compute-Optimal Training Beyond Fixed Training Durations | Accept (spotlight) | Summary: The authors present an empirical study on using constant learning rates (plus a short cooldown) instead of cosine schedules for training LLMs. The authors show that:
- Constant LR + cooldown roughly matches cosine schedules. (Fig 3,4)
- SWA of a long schedule almost matches the performance of shorter schedule... | Rebuttal 1:
Rebuttal: Thank you very much for taking the time to read our work and the detailed questions. You can find the replies below:
> **On the experimental setup and perplexity (weaknesses and Questions 1-2):**
Thank you for the feedback -- we have made the setup and pointers to the exact details more clear in... | Summary: Scaling Laws and Compute-Optimal Training without Fixed Training Duration
Summary
A major weakness of the cosine annealing learning rate schedule, one of the most prevalent learning rate schedules in LLM training, is that for optimal performance the cycle length must be adjusted based on the training duration... | Rebuttal 1:
Rebuttal: Thank you very much for the positive and detailed feedback to our work!
We aim to clarify your individual questions below:
> **Spikes in Figure 1**
For all the figures in our paper, we plot the validation perplexity of a fixed set of sequences in the dataset; this therefore entails very smooth ... | Summary: The paper focuses on learning rate schedules in scaling large language models (LLMs). Traditionally, LLMs are trained with a cosine learning rate schedule, which requires a pre-specified training duration for learning rate decay. This makes it difficult to dynamically adjust the training duration, as early sto... | Rebuttal 1:
Rebuttal: We thank you very much for your valuable comments and kind words!
> **Expansion to larger models**
Please see the general response above for the results on both a 8B and a 1B model, which we were able to train without issues out of the box. This makes us confident that the match of performance o... | null | null | Rebuttal 1:
Rebuttal: We thank everyone for taking the time to read our work and the positive & constructive feedback. We are very encouraged by the positive comments of the reviewers to our findings and the writing.
We also thank all reviewers for the useful comments with detailed questions, to which we reply in indi... | NeurIPS_2024_submissions_huggingface | 2,024 | null | null | null | null | null | null | null | null |
Goal Reduction with Loop-Removal Accelerates RL and Models Human Brain Activity in Goal-Directed Learning | Accept (spotlight) | Summary: The paper provides a novel way to find subgoals just from past experiences (without any prior knowledge), named goal-reducer. Goal-reducer works by finding “loops” in trajectories and remove them. The authors extend the concept of loops in the continuous cases by defining a filtration radius (distances of two ... | Rebuttal 1:
Rebuttal: We appreciate the feedback and valuable insights provided by you. For your questions, here are some of our thoughts:
1. Applicability issue: We agree that when a detour is necessary, things could be confusing. However, to our knowledge this usually depends on the problem formulation. For example,... | Summary: The paper proposes a goal reduction mechanism utilizing a novel loop-removal technique to derive subgoals from arbitrary and distant original goals. This method, called goal-reducer, distills high-quality subgoals from a replay buffer without prior global environmental knowledge. It accelerates performance in ... | Rebuttal 1:
Rebuttal: Thank you for the insightful feedback as well as pointing out pieces we missed in the original writing. Regarding your questions,
1. State distance estimate issue: Yes we used Euclidean distance to measure the similarity between different states representations, however, this does not necessarily... | Summary: This paper introduces a new automated method to identify subgoals in reinforcement learning. The method is based on a replay buffer of agent trajectories and proceeds by identifying and eliminating “loops” (points where the agents revisit almost the same state), in effect attempting to identify efficient short... | Rebuttal 1:
Rebuttal: Thank you for your detailed and constructive feedback. Regarding equation 4-5, they represent the Bellman part in RL, equation 6 and 7, together, shows how a goal reducer can accelerate RL by using the generated subgoal to regularize (KL divergence is used) the learning of value of the original go... | null | null | Rebuttal 1:
Rebuttal: We appreciate the reviewers' thoughtful feedback and the opportunity to clarify several key aspects of our work. Individual rebuttals are made below to avoid confusion. In this global rebuttal, we aim to address the concerns raised by reviewer [ww9y](https://openreview.net/forum?id=Y0EfJJeb4V¬e... | NeurIPS_2024_submissions_huggingface | 2,024 | null | null | null | null | null | null | null | null |
TAIA: Large Language Models are Out-of-Distribution Data Learners | Accept (poster) | Summary: The paper introduces TAIA, a novel method for enhancing the performance of large language models (LLMs) in data-scarce domains with domain-mismatched data. The authors identify that during fine-tuning, only attention parameters significantly contribute to downstream task performance when training and test sets... | Rebuttal 1:
Rebuttal: Thank you for your review and important questions! We are glad to hear that you find our newly proposed TAIA useful novel and that our analysis on self-attention and feed-forward networks is thorough. Below we address the questions and concerns raised in the review.
> For example, figure 2 could ... | Summary: This paper proposes the hypotheses that under some circumstances LLMs are better trained by ft all parameters and then dropping the learnt FFNN layers. The proposed method is named TAIA from training all parameters but inferring with only Attention.
A comparison of TAIA's the performance together with anothe... | Rebuttal 1:
Rebuttal: Thank you for your insightful reviews and comments! We are glad to hear that you find the proposed methodology useful and that it significantly improves over the previous OOD approaches. It is also encouraging to see that you find our experiments extensive, the analysis insightful and the explanat... | Summary: This paper proposes a simple but effective strategy to improve fine-tuning of LLMs: fine-tune as normal but at inference time, only use the fine-tuned attention parameters but use the pre-trained MLP parameters. The intuition for this is that knowledge is generally stored in the MLP layers, and fine-tuning int... | Rebuttal 1:
Rebuttal: Thank you very much for your detailed reviews, comments, and suggestions. We are glad to hear that you find the TAIA a valuable method and that the experiments are sufficient to validate the method. Below we address the concerns and questions raised in the review.
> Confusion about Section 4.2.
... | Summary: The paper proposes a novel inference-time intervention method that trains all model parameters but retains only the self-attention updates for inference. This approach, named "Training All parameters but Inferring with only Attention" (TAIA), optimizes performance across a variety of downstream and closed-book... | Rebuttal 1:
Rebuttal: Thank you for your review and important questions! We are glad to hear that the reviewer found our proposed TAIA innovative and robust. Below we address the concerns and questions raised in the review.
> 1.Fine-tuning of FFN leads to distribution shift
Response: To empirically show the distribut... | Rebuttal 1:
Rebuttal: We offer a possible explanation of TAIA.
Suppose an LLM $p_{\theta}$ containing pretrained weight $\theta_0$, the vanilla LoRA-tuned model yields $\Delta \theta_0$ weight which is to be merged back to pretrained weight and has a relatively small norm. Suppose a simplified neural network layer wi... | NeurIPS_2024_submissions_huggingface | 2,024 | null | null | null | null | null | null | null | null |
MultiPull: Detailing Signed Distance Functions by Pulling Multi-Level Queries at Multi-Step | Accept (poster) | Summary: The authors propose a method to reconstruct neural multi-scale SDFs from point clouds using an optimization-based approach. The multi-scale SDF is represented with a network architecture based on previous work [11] and optimized using a new iterative pulling approach, where the SDF for each scale is optimized ... | Rebuttal 1:
Rebuttal: Thank you for your acknowledgements in our method and results. In the following, we respond to the comments with respect to the weaknesses and questions in turn. All experiments presented in the rebuttal are based on the FAMOUS dataset.
##### **(1) Further comparison with related works.**
> The ... | Summary: This paper presents a model for reconstructing SDF from point clouds. The proposed approach introduces components, including Frequency Feature Transformation and Multi-Step Pulling, to iteratively refine the reconstructed SDF. The experiments demonstrate that this method performs well across multiple 3D object... | Rebuttal 1:
Rebuttal: Thank you for your acknowledgements in our method and results. In the following, we respond to the comments with respect to the weaknesses and questions in turn. All experiments presented in the rebuttal are based on the FAMOUS dataset.
##### **(1) Motivation of FFT.**
> The rationale for using ... | Summary: This paper proposes MultiPull, a method for reconstructing a surface model by SDF from a 3D point cloud containing only the coordinates of each point. The paper proposes a method for estimating the SDF using the Fourier transform of the surface model predicted from the point cloud only at multi-scale. It also ... | Rebuttal 1:
Rebuttal: Thank you for your acknowledgements in our method and writing. In the following, we respond to the comments with respect to the weaknesses and questions in turn. All experiments presented in the rebuttal are based on the FAMOUS dataset.
##### **(1) Concern about frequency feature consistency.**
... | Summary: This paper proposes to learn multi-scale implicit fields from 3D point clouds for accurate optimization of SDFs in a coarse-to-fine manner. The spatial query points are first mapped to frequency features through the FFT module. Then the MSP module is designed to exploit the multi-level frequency features to pr... | Rebuttal 1:
Rebuttal: Thank you for your acknowledgements in our method. In the following, we respond to the comments with respect to the weaknesses and questions. All experiments presented in the rebuttal are based on the FAMOUS dataset.
##### **(1) Novelty.**
> In general, the proposed techniques are just an increm... | Rebuttal 1:
Rebuttal: We thank all the reviewers for their thoughtful comments and valuable feedback. We sincerely thank reviewers uTM5 and NaoM for their supplementary requirements on the paper details and our contribution experiments. Through these experiments, we further demonstrate the scalability and unique contri... | NeurIPS_2024_submissions_huggingface | 2,024 | null | null | null | null | null | null | null | null |
Hierarchical Hybrid Sliced Wasserstein: A Scalable Metric for Heterogeneous Joint Distributions | Accept (poster) | Summary: This work focuses on introducing a Sliced-Wasserstein distance between heterogeneous joint distributions, i.e. distributions whose ambient space lies in a product space, with possibly different geometries for each space. This new distance relies on a new hierarchical hybrid Radon transform. The authors derive ... | Rebuttal 1:
Rebuttal: We would like to thank the reviewer for the time and constructive feedback.
**Q17**. First, the applications on 3D meshes are nice and large scale (10000 points) but only in few dimensions ... It would have been nice to compare it with e.g. a spherical Radon transform.
**A17**. For the low-dimen... | Summary: The paper proposes a new metric for comparing heterogeneous joint distributions, called Hierarchical Hybrid Sliced Wasserstein (H2SW). The proposed metric is tested for shape deformation, for training shape autoencoder and comparing datasets.
The proposed metric is based on a new slicing operator (Hierarchical... | Rebuttal 1:
Rebuttal: We would like to thank the reviewer for insightful comments and constructive feedback. We would like to answer questions from the reviewers as follow:
**Q11**. The third contribution listed (lines 81-82) is confusing (and in my opinion should be rephrased/tone down): the paper ignores other attem... | Summary: Wasserstein distance measures the distance between two distributions and is particularly important in machine learning, especially in generative models. However, computing the Wasserstein distance is computationally intensive. For homogeneous distributions, Sliced Wasserstein (SW) and Generalized Sliced Wasser... | Rebuttal 1:
Rebuttal: We would like to express our gratitude for the time and feedback from the reviewer. We would like to answer questions from the reviewers as follows:
**Q6**. The paper needs to clearly explain heterogeneous joint distributions, such as by detailing why 3D shape deformation is considered a heteroge... | Summary: To address the issue that Sliced Wasserstein (SW) and Generalized Sliced Wasserstein (GSW) GSW are only defined between distributions supported on a homogeneous domain and using SW and GSW directly on the joint domains cannot make a meaningful comparison, this paper first extends the partial Radon Transform to... | Rebuttal 1:
Rebuttal: We would like to thank the reviewer for the time and the feedback.
**Q1**. The authors did not clarify the main challenges of designing a sliced Wasserstein variant for heterogeneous joint distributions in the Introduction section. Additionally, they failed to explain how the proposed H2SW addres... | Rebuttal 1:
Rebuttal: Dear chairs and reviewers,
First, we would like to thank the reviewers for spending time reviewing our paper and providing constructive feedback. We would like to summarize some additional results and common questions from the reviewers:
* As suggested by Reviewer **XX3o**, we added an experimen... | NeurIPS_2024_submissions_huggingface | 2,024 | 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.