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 |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
An Improved Relaxation for Oracle-Efficient Adversarial Contextual Bandits | Accept (poster) | Summary: This paper investigates the contextual bandit problem where the contexts are sequentially sampled from a known i.i.d. source, and losses are generated adversarially. The primary focus is on developing oracle-efficient algorithms that minimize the expected regret (the expectation over both the context's randomn... | Rebuttal 1:
Rebuttal: Thank you for your comments.
We are glad that
you think our ideas and techniques are of independent interests.
Please find below our response to the questions and comments mentioned in your review.
**Questions**
Q1: We initially obtain a "Rademacher style" bound (Equation 8 in our paper) in our ... | Summary: In this paper, the authors consider a classic contextual bandit problem with adversarial loss and stochastic context. Specifically, they proposed a relaxation-based algorithm which achieves $O(K^{1/3}T^{2/3}\log^{1/3}|Pi|)$ expected regret bound, improving upon the best known $O(K^{2/3}T^{2/3}\log^{1/3}|Pi|)$ ... | Rebuttal 1:
Rebuttal: Thank you for your comments. We are glad that you found the writing of our paper to be clear.
Please find below our response to the questions and comments mentioned in your review.
**Questions**
Q1: We note that
the regret lower bound of $\sqrt{TK\log(|\Pi|)}$ holds for all (not necessarily e... | Summary: The paper studies the problem of minimizing regret for online adversarial contextual bandits using an ERM oracle. If one doesn't care about oracle efficiency, then the well known EXP4 algorithm achieves the optimal rate. However, the best previous oracle-efficient result is due to [Syrgkanis et al. '16]. This ... | Rebuttal 1:
Rebuttal: Thank you for your comments.
We are glad that you found our paper to be "exceptionally well written".
Please find below our response to the questions and comments mentioned in your review.
**Questions**
Q1: The oracle complexity is equivalent to that of Syrgkanis et al. 2016 and
our main result... | Summary: This paper proposed a new algorithm for Oracle-Efficient Adversarial Contextual Bandits that achieves the best known regret bound and improve upon the previous best bound in its dependency on the action set size $K$.
Strengths: The algorithm is elegant. The paper is well-written and pleasant to read. The theo... | Rebuttal 1:
Rebuttal: Thank you for your comments. We are glad that you found the paper well-written and that you believe our algorithm is elegant.
Please find below our response to the questions and comments mentioned in your review.
> any work that make effort towards some type of computational efficiency ... experi... | Rebuttal 1:
Rebuttal: We thank the reviewers for their comments.
We here respond to the concerns that were shared by
more than one reviewer.
**Relaxing the assumption of the setting (Reviewers yx68, gZc2, and 2pq4)**
Our focus on this paper was improving the existing regret bounds in
the setting based on prior work (... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
ToolkenGPT: Augmenting Frozen Language Models with Massive Tools via Tool Embeddings | Accept (oral) | Summary: The authors propose "ToolkenGPT." This method allows language models to take advantage of external tools without full model fine-tuning, and with the ability to show more tools and with greater detail than in-context learning typically allows. They show that use of "toolkens" (tool tokens) improves performance... | Rebuttal 1:
Rebuttal: We thank reviwer QLCo for acknowledging that our approach is new, innovative and interesting, brining up new and useful ideas to the field. Below we address all the reviewer's concerns and questions.
(a) Comparing with full-model fine-tuning. We appreciate that the reviewer acknowledge that our p... | Summary: This paper proposes a new method to help LLM utilize tools. Essentially, the method trains an additional output embedding token (toolken) for each tool. During inference, if LLM outputs a toolken, a special routine of using the corresponding tool is triggered and the returned result of the tool is then replace... | Rebuttal 1:
Rebuttal: We thank Reviewer 5YmU for acknowledging that our proposed method is much more efficient, easy to implement, and can be applied to various domains to foster more research in the field of tool-augmented LLMs. We also thank the reviewer for spotting several typos and we will fix them all in the revi... | Summary: This paper presents ToolkenGPT, a novel approach to generalizing LLMs to massive external tools. The key is to represent each tool as a "token", such that when LLMs generate the tool token(toolken), the associated tool will be invoked and executed. The learning of toolken is easy and efficient since there is n... | Rebuttal 1:
Rebuttal: We thank reviewer Mpoh for acknowledging that our proposed method is novel and elegant to learn massive tools efficiently, and makes a good contribution to the field of tool learning. We address the reviewer's comments on the evaluation as follows:
(a) Comparing with full-model fine-tuning. Thank... | Summary: The paper presents ToolkenGPT, a framework for extending LMs with tool use. For each new tool, a new token is added to the output vocabulary of the LM and the embedding for that token is trained with annotated or synthetic examples. When the LM generates that token, the model is switched to a different mode an... | Rebuttal 1:
Rebuttal: We thank the Reviewer xBHJ for acknowledging that our method is novel and useful, and will push forward the field of tool-augmented LLMs. We appreciate the reviewer's great suggestions and questions. We will further proofread the paper and fix the typos in the final version. Herein, we address the... | Rebuttal 1:
Rebuttal: We are grateful to all reviewers for their detailed and constructive feedback! We are encouraged to see that reviewers find:
- Our proposed ToolkenGPT is an "innovative, novel and useful", "a novel yet elegant way" (R-C7TV, R-xBHJ, R-QLCo, R-Mpoh), which "brings up new ideas", "fosters more resear... | NeurIPS_2023_submissions_huggingface | 2,023 | Summary: This study introduces an approach to teach Large Language Models (LLMs) to utilize numerous tools through learning embeddings for these tools. The embeddings are incorporated into the LLMs' output layer, invoking the use of tools at relevant steps. The inference process alternates between reasoning and tool-us... | Rebuttal 1:
Rebuttal: We thank the reviewer C7TV for acknowledging that our method is innovative, effective and scales well to massive tools with limited examples. Below we address the reviewer's comments and questions point by point.
(a) Computation costs and performance. Thanks for the great comment! We understand t... | null | null | null | null | null | null |
Regularity as Intrinsic Reward for Free Play | Accept (poster) | Summary: This work proposes an exploration approach that prioritizes "regularity" in the exploration behavior, in contrast to typical novelty-seeking exploration. The objective is defined as the minimization of entropy over user-defined, object-centric features. The method is primarily evaluated on block rearrangement ... | Rebuttal 1:
Rebuttal: We thank the reviewer for their detailed feedback on our work. We address their concerns in the following.
# Generality of RaIR
We have provided new experiments in new domains in the rebuttal pdf (Fig. D2-D4). As the reviewer suggested we have included RaIR generated arrangements in locomotion... | Summary: This paper presents a new approach that utilizes a regularity as an intrinsic reward, which aims to regularize, or constrain, the search space of exploration towards state regions where the entities are more likely to have structured patterns. To this end, the paper first formulates how to define regularity as... | Rebuttal 1:
Rebuttal: We thank the reviewer for their assessment and feedback.
# Generalizability of RaIR
We have now added new experiments where we showcase RaIR in different scenarios. In order to address the concern of focusing on one specific type of structuredness, we have tested a variant of ShapeGridWorld, whe... | Summary: The paper draws inspiration from a child’s development cycle and proposes the use of regularity as an intrinsic reward to help guide exploration in RL. It shows that existing metrics of state novelty can be combined with the proposed regularity metric to enable construction of regular structure during free pla... | Rebuttal 1:
Rebuttal: We thank the reviewer for their feedback and address their questions and comments in the following.
# Clarification on the Graph Architecture
We use a fully-connected Graph Neural Network (GNN).
*How did the authors decide on this specific state representation? Does it have any advantages tha... | Summary: Inspired by the fact that humans prefer symmetric patters, the author proposed to use regularity as intrinsic reward to encourage an agent to discover interesting environmental patterns. One example of regularity is to put boxes in a symmetric pattern. The results showed that this kind of intrinsic reward allo... | Rebuttal 1:
Rebuttal: We thank the reviewer for their review and feedback. Our response follows:
# Baselines:
In the newly uploaded pdf, we have included experiments with RND. Here we run RND with the same model-based planning backbone (GNN ensemble together with iCEM as in CEE-US) and only alter the intrinsic rewar... | Rebuttal 1:
Rebuttal: We thank all reviewers for their constructive feedback and appreciate that they found our method to be "novel and intuitive”, the experiments “comprehensive” with “impressive results for stacking towers in various configurations solely from intrinsic rewards” and the paper “generally well written”... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Large Language Models Are Latent Variable Models: Explaining and Finding Good Demonstrations for In-Context Learning | Accept (poster) | Summary: The paper presents a new perspective to understand the in-context learning behavior of large language models (LLMs) from the angle of latent concept learning resembling topic models. Based on the generative process defined by input data, latent concepts, and labels, the authors propose a two-stage algorithm to... | Rebuttal 1:
Rebuttal: Thank you for your review. Below is our response to your comments:
1. **About how realistic is the problem setting**: We want first to reiterate that our problem setting is chosen to empirically prove the correctness of our proposed theory of in-context learning. We want to reiterate that improvi... | Summary: The paper introduces a novel demonstration selection method aimed to enhance performance of (few-shot) in-context learning. The approach is characterized by a two-stage process that includes latent concept learning and demonstration selection, each holding a unique significance. In the latent concept learning ... | Rebuttal 1:
Rebuttal: Thank you for your positive review. Below is our response to your comments:
1. **About optimal latent**: Yes, in practice, the learned latent is not (even never) optimal, but approximate. The assumed optimality is for deriving the upper bound theory that the in-context learning classifier can be ... | Summary: This paper describes a framework to select demonstrations for in-context learning by using bayseian formulation for the data generation process. Similar to topic models, the formulation uses a "concept" variable and words in this sequence are conditioned over this concept variable and are conditionally indepen... | Rebuttal 1:
Rebuttal: Thank you for your review. Below is our response to your comments:
1. **About topic model assumption**: We want to clarify that the topic model here is in a more general sense, which is not equivalent to LDA [1], but similar to the modern neural topic models proposed in [2,3]. In this more genera... | Summary: This paper tries to study in-context learning through a Bayesian lens, namely treating LLMs as implicit topic models. It proposes an algorithm to select optimal demonstrations from a set of annotated data with a small LLM and then use the selected demonstrations with larger LLMs. 12.5% improvement compared to ... | Rebuttal 1:
Rebuttal: Thank you for your review. Below is our response to your comments:
1. **About our claims**: We would like to clarify that we are not claiming that LLMs are topic models. Our claim is that LLMs are IMPLICITLY topic models, that learn and infer a latent task/topic variable for solving each query. W... | Rebuttal 1:
Rebuttal: We want to first thank all the reviewers for taking the time to review our paper. We provide some clarification and additional results below:
1. **About topic modeling assumptions**: We want to clarify that the topic model here is in a more general sense, which is not equivalent to LDA [1], but s... | NeurIPS_2023_submissions_huggingface | 2,023 | Summary: This work aims at proposing a demonstration example selection algorithm in in-context learning, using a formulation of topic models for language models. Specifically, the proposed algorithm applies prompt tuning on some prefixing learnable tokens to obtain latent concepts, and select demo examples that maximiz... | Rebuttal 1:
Rebuttal: Thank you for your detailed review. Below is our response to your comments:
1. **Takeaway of our paper**: We would like to clarify that our study's focus is on elucidating the underlying mechanisms of in-context learning and presenting a novel approach to demonstration selection that empirically ... | null | null | null | null | null | null |
Predicting mutational effects on protein-protein binding via a side-chain diffusion probabilistic model | Accept (poster) | Summary: The paper proposes SidechainDiff a diffusion model over the torsion angles of the protein sidechains and uses the learned representations of this model to predict the mutational effects of protein-protein binding.
Strengths: The paper shows a wide variety of promising experimental results ranging from standa... | Rebuttal 1:
Rebuttal: We sincerely appreciate the reviewer's constructive feedback. We have carefully addressed each comment and made the necessary revisions accordingly.
__Q1:__
1. Thanks for your advice, We have incorporated references to existing published methods [1, 2] in the related work section.
2. Comprehensi... | Summary: The paper introduces a new representation learning-based approach called SidechainDiff that predicts how amino acid mutations influence protein-protein binding using Riemannian diffusion model. It's very important problem in the field of structural biology, for example, in evaluation of antibody variants impac... | Rebuttal 1:
Rebuttal: We sincerely appreciate the reviewer's constructive comments on our work. In response, we have carefully addressed their concerns and provided clarification on various issues as follows:
__Q1:__
To obey the rules of the double-blind reviewing policy, we initially don't provide a GitHub URL. But S... | Summary: The paper introduces DiffAffinity, a novel method for predicting the effects of mutations on protein-protein interactions. The authors leverage a diffusion model to learn representations that aid in predicting changes in binding affinity (∆∆G). They compare DiffAffinity's performance with existing methods such... | Rebuttal 1:
Rebuttal: We appreciate the reviewer's approval and constructive comments on our work. In response, we have carefully addressed their concerns and provided clarification on various issues as follows:
__Q1:__
The affinity of the complex is derived from the dynamic conformation of the protein rather than its... | Summary: The paper studies the problem of mutational effect prediction on protein-protein binding. To address the data scarcity issue, the authors propose to first pre-train their model on PDB using diffusion models to fit the protein side-chain distribution. Then, the learned representations are used for predicting th... | Rebuttal 1:
Rebuttal: We thank the reviewer for constructive comments on our work. Below are our responses to the reviewer's concerns.
__Q1:__
We realize that too many notations similar to the paper of the Riemannian Diffusion Model may be misleading. We agree with the reviewer in this aspect. Following the reviewer’s... | Rebuttal 1:
Rebuttal: In our “global” response, to address reviewers’ main concerns, we elucidate our motivation, delineate the distinctions between our approach and previous methods (including the original Riemannian diffusion model and torsion diffusion model), and present new results (including comparison with ESM2... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Score-based Source Separation with Applications to Digital Communication Signals | Accept (poster) | Summary: 1. In this paper authors proposed a method for separating superimposed sources using diffusion based generative models. The proposed model derived a new objective function based on maximum a posterior and alpha posterior.
2. The application of the proposed work is clearly mentioned by considering the existing ... | Rebuttal 1:
Rebuttal: We thank the reviewer for the feedback and comments. We will address the raised questions and concerns here.
**Regarding the reviewer’s comments under “weaknesses”:**
Due to space considerations, we were not able to include all details in the main manuscript, but we have included them in the sup... | Summary: This paper proposed a new Bayesian method to separate a mixture of two signals $\mathbf{s}$ and $\mathbf{b}$ from the mixture signal $\mathbf{y}=\mathbf{s}+\kappa \mathbf{b}$ where $\kappa \in \mathbb{R}_+$ is known. The new method leverages the score from pre-trained diffusion models to extend MAP estimation ... | Rebuttal 1:
Rebuttal: We thank the reviewer for their comments regarding our technical contributions. Below we address the raised questions and concerns.
**Regarding the reviewer’s comments under “weaknesses”:**
1. **Theory:** Due to space considerations, we included our theoretical results in the supplementary materi... | Summary: This paper focuses on the problem of single-channel source separation (SCSS) for RF signals with discrete nature. The authors propose to solve the SCSS problem using MAP and use pre-trained diffusion models to approximate the scores for both the source signal and inference signal. To avoid being stuck in a loc... | Rebuttal 1:
Rebuttal: We thank the reviewer for the feedback and helpful questions. We shall now address the raised questions and concerns.
**Regarding the reviewer’s comments under “weaknesses”:**
1. **Knowledge about the scaling factor**: *This is an excellent point and we hope to address it in future work by introd... | Summary: A new method for separating superimposed sources using diffusion-based generative models is proposed (alpha-RGS).
The method relies on separately trained statistical priors of independent sources and is guided by maximum a posteriori estimation with an α-posterior.
Experimental results with RF mixtures demonst... | Rebuttal 1:
Rebuttal: We thank the reviewer for the feedback and helpful questions. We shall now address the raised questions and concerns.
**Regarding the questions:**
1. **Statistics of noise:** Thank you for the question. We would first like to clarify that in a typical communication setup with a transmitter, chann... | Rebuttal 1:
Rebuttal: We thank the reviewers for taking the time to review our work proposing a novel algorithm that leverages diffusion-based priors and randomized levels of Gaussian smoothing with applications source-separation in the RF domain, a relatively new domain area to the ML community. We appreciate all the... | NeurIPS_2023_submissions_huggingface | 2,023 | Summary: This paper investigates a novel Bayesian approach for separating the superposition of two sources based on diffusion generative models. The problem is motivated by application to the spectrum of radio-frequency (RF) communication systems. Several experiments using real datasets on RF mixtures demonstrate that... | Rebuttal 1:
Rebuttal: We thank the reviewer for taking the time to understand the technical contributions in our work. Below we shall address the raised questions and concerns.
**Regarding the reviewer’s comments under “weaknesses”:**
1. **Multi-source Separation:** To make the exposition clearer, we focus on the two-... | null | null | null | null | null | null |
HyP-NeRF: Learning Improved NeRF Priors using a HyperNetwork | Accept (poster) | Summary: In the paper authors propose a novel method for NeRF based model which able to generalize. Authors use hypernetwork paradigm and Multi-Resolution Hash Encodings.
The paper is interesting but has two main problems:
- two-stage training, where the second stage is applied to generate NeRF representation.
- a... | Rebuttal 1:
Rebuttal: We would like to thank the reviewer for their thoughtful comments.
1. **Related works...**: We would like to thank the reviewer for the categorization. However, there could be several criteria for categorizing different techniques, like based on the downstream tasks the technique enables. Furthe... | Summary: In this paper, the authors present HyP-NeRF, a framework based on meta-learning principles, tailored for the learning of category-level NeRF priors. This is achieved through conditioning on latent codes with the assistance of hypernetworks. The NeRF model struggles with generalization across categories because... | Rebuttal 1:
Rebuttal: We sincerely appreciate the reviewer's thorough analysis of our work and for raising their questions and concerns. In the global and local rebuttal, we have addressed the concerns and questions and provided additional comparisons and visualizations (rebuttal PDF).
1. To demonstrate **shape and co... | Summary: The paper proposes HyP-NeRF, a method using hypernetworks to learn generalizable category-level priors, addressing the limitations of existing work on generalization. Specifically, the proposed hypernetwork-based method predicts the parameters of NeRF and multi-resolution hash encodings, and further incorpora... | Rebuttal 1:
Rebuttal: We would like to appreciate the reviewer's efforts in studying our work and for their positive assessment of our contributions. Below, we elaborate on the questions raised, and provide additional comparisons to address all the concerns.
1. **There are not enough details on why fine-tuning proced... | Summary: This paper proposes a way to encode a categorical prior for NeRFs. The method train an auto-decoder, which optimize together both the latent codes (one for every instance in the dataset) and the decoder parameters. The decoder will take the latent code and predict parameters for an instance-NGP backbone suppor... | Rebuttal 1:
Rebuttal: We are delighted to receive such a positive assessment of our work and are grateful for it. Below, we provide the answers to the questions that were raised.
1. **Require separate pipeline for generation ... this second stage of the pipeline will predict something that's aware of the conditional ... | Rebuttal 1:
Rebuttal: We thank the reviewers for their feedback and positive assessment of our contributions. We address major factual errors and broad reviewer concerns in this global response and specific reviewer concerns/questions in the local response. We are confident that we will have addressed all their concern... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
TensorNet: Cartesian Tensor Representations for Efficient Learning of Molecular Potentials | Accept (poster) | Summary: The paper presents TensorNet, a novel O(3)-equivariant message-passing neural network for efficient representation of molecular systems in scientific research. This model utilizes Cartesian tensor atomic embeddings to simplify feature mixing via matrix product operations. By decomposing tensors into rotation g... | Rebuttal 1:
Rebuttal: We thank the reviewer for their positive feedback on the model. We now address the questions raised.
**The proposed architecture (illustrated in Fig. A1) appears quite intricate. Could you provide more insight or intuitive reasoning behind the specific design choices in the model? What fundamenta... | Summary: This paper proposes TensorNet, an O(3)-equivariant neural network architecture for molecules. Using the decomposition of a 3x3 matrix into a scalar, vector, and matrix shown in Eq. (2), TensorNet efficiently computes the interaction of O(3)-equivariant features up to l=2, where l is the degree (frequency) of t... | Rebuttal 1:
Rebuttal: We thank the reviewer for their positive feedback on the novelty of the approach, completeness of the validation, and results. We here answer the further points raised.
**The paper has room to improve in terms of presentation. I'm unfamiliar with the chemistry (molecule) domain, and some parts s... | Summary: This paper proposes a cartesian tensor representation for efficient learning of molecular potentials. It enables the feature mixing process to be a simple matrix product operation. In addition, the matrix product operation is simplified by cost-effective decomposition techniques. Experimental results demonstra... | Rebuttal 1:
Rebuttal: We thank the reviewer for the nice feedback on the strengths of TensorNet. We clarify here the other points raised.
**The presentation is not that easy to follow and the notation is a little bit complicated, which brings additional hardness for readers to understand the core idea. In addition, al... | Summary: The paper introduces Tensor3 Net, a message-passing neural network architecture designed for molecular systems representation. Tensor3 Net leverages rank-2 Cartesian tensor representations and O(3)-equivariance. The tensors are decomposed into rotation group irreducible representations, enabling separate proce... | Rebuttal 1:
Rebuttal: We thank the reviewer for the feedback on the model in terms of its validity compared to e3nn. Here we address the main point raised.
**The one main weakness is that the author only model rank-2 tensors, which is relatively arbitrary. What it should be done is to introduce a network with general... | Rebuttal 1:
Rebuttal: The questions raised by the reviewers have been addressed in their individual rebuttals. However, since some of them expressed concerns about the clarity of exposition, we use this general rebuttal to attach a figure that will be added to the manuscript, which we hope provides clarification on bot... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Multi-Object Representation Learning via Feature Connectivity and Object-Centric Regularization | Accept (spotlight) | Summary: This paper presents a novel approach towards an unsupervised object-centric representation learning framework. Diverging from previous methodologies that prioritize an input reconstruction scheme, this method emphasizes feature connectivity to cluster neighboring pixels, employing two object-centric regulariza... | Rebuttal 1:
Rebuttal: Thank you for your detailed review and for the encouraging comments. We provide some responses to the questions as follows:
**Q1: The scalability of this method to real-world scenes of higher complexity remains a point of ambiguity. Further testing and discussion are necessary to ascertain the ef... | Summary: The paper presents a novel approach to object-centric (multi-object) representation learning. In contrast to conventional sequential or iterative attention-based methods, the proposed method employs a pixel-wise shortest distance-based clustering technique. Moreover, the paper emphasizes the use of regularizat... | Rebuttal 1:
Rebuttal: Thank you for your detailed review and for the insightful comments especially on scalability. We provide some responses to the questions as follows:
**Q1: The reviewer questions the difference between using the original RGB input image (with proper normalization) and the feature map generated by ... | Summary: The paper presents a method for learning pixel representations for multi-object segmentation. The method works by projecting pixels using $1 \times 1$ convolution into a feature space. The graph is defined on top of pixels using 8-connectivity. A modified Dijkstra's algorithm assigns pixels to objects based on... | Rebuttal 1:
Rebuttal: Thank you for your detailed review and for the extensive pointers to related works. We provide some responses as follows:
**Q1: Algorithm 1 suggests that the object sets are not disjoint and may contain the same pixels as other objects. How are the masks obtained in that case, e.g. for evaluation... | Summary: The paper discusses a novel method for learning object-centric representations from images, the new approach uses feature connectivity to group together pixels likely to be part of the same object and employs two object-centric regularization terms for refining the representations.
The method has been tested... | Rebuttal 1:
Rebuttal: Thank you for your detailed review and for the encouraging comments. We provide some responses to the questions as follows:
**Q1: Lack of results on more challenging datasets with complex scenes, object types, object shapes, and object occlusions.**
We have conducted additional experiments on th... | Rebuttal 1:
Rebuttal: We would like to thank the reviewers for the helpful feedback and insightful comments.
We are encouraged that our contribution of an unsupervised object discovery network which leverages on feature connectivity and designed object-centric regularization terms is described as "*novel*" (TJh8, T4Hn... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Selective Amnesia: A Continual Learning Approach to Forgetting in Deep Generative Models | Accept (spotlight) | Summary: This paper draws inspirations from continual learning and proposes a Selective Amnesia method to selectively forget concepts in pre-trained generative models. The proposed method is implemented in a widely-used framework of Bayesian continual learning, together with generative replay to optimize the objective.... | Rebuttal 1:
Rebuttal: We thank the reviewer for acknowledging the quality, strong experimental results and interesting approach of our work. Below we provide our responses to the concerns raised by the reviewer.
> Therefore, a comparison of the proposed Selective Amnesia and AFEC is necessary.
Thank you for bringing ... | Summary: The paper aims to forget selected concepts from deep generative models trained with variational inference with a focus on the trustworthiness such as malicious prompts. The paper introduces a continue learning approach that combines EWC and GR to maximize the likelihood (or ELBO) on the remembered set. Experim... | Rebuttal 1:
Rebuttal: We thank the reviewer for acknowledging the motivations and usefulness of our work. Below we provide our responses to the reviewer's concerns.
> ...detailed explanation on the selection of q
The choice of $q$ can be understood theoretically from Corollary 1-- a greater difference between $q$ and... | Summary: In this work, authors introduce Selective Amnesia - a new method for selective forgetting of particular concepts in generative modeling without access to the training data. Authors propose to use Continual Learning methods - namely, Elastic Weights Consolidation and Generative Replay to retrain the base model ... | Rebuttal 1:
Rebuttal: We would like to thank the reviewer for acknowledging the strong motivation and contributions of our work, as well as our extensive experimental results. Below we provide our responses to the concerns raised by the reviewer.
> What happens if the concept we want to forget is highly correlated wit... | Summary: This paper presents a technique called Selective Amnesia, which enables controllable forgetting of concepts in pretrained deep generative models. The authors show that this technique can be applied to a variety of models, including text-to-image and VAEs, and can be framed from the perspective of continual lea... | Rebuttal 1:
Rebuttal: We would like to thank the reviewer for acknowledging the strong motivations, novelty and quality of our work. Below we provide our responses to the concerns raised by the reviewer.
> ...extending the quantitative evaluation to more challenging datasets would provide a more comprehensive perspect... | Rebuttal 1:
Rebuttal: Thank you to the reviewers for their comments and feedback. We appreciate the positive comments and that the reviewers find our continual learning approach to concept forgetting to be meaningful and technically interesting. Based on the comments, we have revised the paper to include (i) additional... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Subspace Identification for Multi-Source Domain Adaptation | Accept (spotlight) | Summary: This paper considered multiple-source domain adaptation problems in the context of causal representation learning. Specifically, this paper considered quite general scenarios in the domain adaptation by figuring four different representations.
- Domain specific and label irrelevant
- Domain specific and labe... | Rebuttal 1:
Rebuttal: We are very grateful for your valuable comments, helpful suggestions, and encouragement. Below please see our point-to-point responses to your comments and suggestions.
>**Q1**: Theorem 1 about theoretical assumptions. In general, I could follow the proof, where intuitions within these assumption... | Summary: This paper studies the problem of multi source domain adaptation where we have access to multiple labeled source domains and unlabeled target domain. Authors consider a novel data generation process by modeling them through 4 new latent variables (i.e. combinations of domain specific or domain invariant and la... | Rebuttal 1:
Rebuttal: We are very grateful for your valuable comments, helpful suggestions, and encouragement. Below please see our point-to-point responses to your comments and suggestions.
>**Q1**: This framework is somewhat similar to earlier work on domain generalization(https://arxiv.org/pdf/2102.11436.pdf). I ag... | Summary: In this paper, the authors investigate the problem of multi-source domain adaptation. In detail, the authors first devise a causal generation process, which consider the domain-specific and label-irrelevant, domain-specific and label-relevant, domain-invariant and label-invariant, domain-invariant and label-ir... | Rebuttal 1:
Rebuttal: We would like to express our deep gratitude for the valuable feedback and helpful suggestions provided on our paper, as well as for the time you devoted to reviewing it. Below, we have addressed each of your comments and suggestions.
>**Q1**: According to the theoretical results, the authors give... | Summary: This paper proposes a novel framework to tackle the problem of multi-source unsupervised domain adaptation. Existing methods often have stringent requirements, such as a large number of domains and invariant label distributions. To address these limitations, this paper proposes a subspace identification theory... | Rebuttal 1:
Rebuttal: We highly appreciate the valuable commands and helpful suggestions on our paper and the time dedicated to reviewing it. Below please see our point-to-point responses to your comments and suggestions.
>**Q1**: I am unsure what is the intuitive reasoning behind the "domain-invariant label-irrelevan... | Rebuttal 1:
Rebuttal: Dear Reviewers Yx4K, nEWG, pec6, and bTZb:
Thanks for the thoughtful and constructive review. It is encouraging that the reviewers think SIG is novel (Reviewer Yx4K and pec6), interesting (Reviewer Yx4K and nEWG), and solid ( Reviewer bTZb). We here provide a general response to summarize the mod... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Sampling from Structured Log-Concave Distributions via a Soft-Threshold Dikin Walk | Accept (poster) | Summary: This paper studies the problem of sampling from a log-concave distribution, parametrized by $\exp(f)$ where $f$ is $L$-Lipschitz, over a $d$-dimensional polytope specified by $m$ halfspaces. Assuming the polytope is contained in a ball of radius $R$, the authors develop an algorithm that generates a sample fro... | Rebuttal 1:
Rebuttal: Thank you for your valuable comments and suggestions. We are glad that you appreciate our state-of-the-art results, and its many applications to training differentially private and Bayesian ML models, and we thank you for supporting our paper. We answer your specific questions below.
**“…genera... | Summary: The paper proposes a sampling method for log-concave distribution with bounded polytope. The methods can be used in privacy preservation. The theoretical results show the guarantee of the sampling methods with some accepted error magnitude. The comparisons for other methods are also detailedly proposed, showin... | Rebuttal 1:
Rebuttal: Thank you for your valuable comments and suggestions. We are glad you appreciate our theoretical results and the applications to differential privacy, and thank you for supporting our paper. We are sorry for any difficulty understanding our presentation. We answer your specific questions below.
*... | Summary: This paper studies the problem of sampling from distribution of the form $\pi(\theta) \propto e^{-f(\theta)}$, restricted to a polytope $K$. Here, $f$ is either Lipschitz or smooth convex. To this end, the authors propose to use _Dikin walk_ Markov chain (Kannan and Narayanan, 2012), which was originally propo... | Rebuttal 1:
Rebuttal: Thank you for your valuable comments and suggestions. We are glad that you appreciate our proof overview, and thank you for supporting our paper. We answer your specific questions below.
**"regularizer that depends on $\theta$..."**
Thank you for the suggestion. While there are settings where a... | Summary: The paper studies the question of efficient sampling from a log-concave distribution from a constrained polytope. The main idea is to use the well known Dikin's random walk with a soft threshold. The threshold is important to ensure a far more efficient convergence to the desired distribution. This soft thresh... | Rebuttal 1:
Rebuttal: Thank you for your valuable comments, and for taking the time to review our paper. We are glad that you appreciate the strength of our results and we thank you for supporting our paper. | null | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
CamoPatch: An Evolutionary Strategy for Generating Camoflauged Adversarial Patches | Accept (poster) | Summary: Existing methods often produce clearly visible distortions since they do not consider the visibility of the patch. To address this, the authors propose a novel method for constructing adversarial patches that approximates the appearance of the area it covers. They achieve this by using a set of semi-transparen... | Rebuttal 1:
Rebuttal: **W#1:** We argue the innovation of our work lies in its consideration of visibility to adversarial patches as well as its methodology for generating adversarial patches. Specifically, previous works do not consider the visibility of a patch resulting in clear distortions to the image as shown in ... | Summary: This paper explores black-box score-based patch attacks on image classification. The previous patch attacks are relatively visible to the naked eye, so the authors hope to have a camouflaged adversarial patch. Therefore, the authors propose a novel method for constructing adversarial patches that approximates ... | Rebuttal 1:
Rebuttal: **W#1:** We address the reviewer's concerns from the following two aspects.
- In fact, We evaluate the proposed method by conducting adversarial attacks on four, including two adversarial and two non-adversarial, classifiers.
- In this work we compare the proposed method with the TPA algorithm of... | Summary: Adversarial examples for a DNN that are not visible to humans are generated with Evolutionary Strategies. The examples are semi-transparent RGB-valued circles. The shape and position are optimized.
Strengths: - ES and SA are gradient free methods
- Patch detectability is taken into account
Weaknesses: - How... | Rebuttal 1:
Rebuttal: **W#1:** The norm of a perturbation, including $l_2$, has proven to be a reliable metric of evaluate the visibility of a perturbation, with previous works either constraining (lines 58-60; reference [2]) or minimizing (section 2.2; references [45, 33, 29, 35, 8]) its value to ensure or promote the... | Summary: This paper proposes a evolutionary strategy based framework for generating imperceptible adversarial patch. Compared with existing methods, the proposed method can generate invisible adversarial patch by iteratively optimizing both the patch and its position on the image.
Strengths: 1. The proposed method is... | Rebuttal 1:
Rebuttal: **W#1, Q#1, L#1:** As shown in Table 2 of our manuscript, we compare the runtime of different hyper-parameter configurations. This partially covers the evaluation of the efficiency of our proposed method. On the other hand, for the sake of fair comparisons, we fixed the computational budget to be ... | Rebuttal 1:
Rebuttal: Attached is the manuscript containing the tables and figures of reviewer requested experiments and plots.
Pdf: /pdf/ff405f80ec68960de0dee1e0d2ca7c3c048753a1.pdf | NeurIPS_2023_submissions_huggingface | 2,023 | Summary: This paper presents a novel method for generating adversarial patches that approximates the appearance of the area it covers, achieving better or comparable performance to state-of-the-art methods while minimizing the visibility of the patch. The method uses semi-transparent, RGB-valued circles that blend into... | Rebuttal 1:
Rebuttal: **W#1:** As per requested by the reviewer, we report the original accuracy, after-attack accuracy, non-normalized residual (nq) and $l_2$ norm metrics of conducted targeted and non-targeted attacks in **Table 2 and Table 1 of the global 'author rebuttal PDF'**.
**W#2:** The reason for the small n... | null | null | null | null | null | null |
Optimization and Bayes: A Trade-off for Overparameterized Neural Networks | Accept (poster) | Summary: The paper presents a novel approach that bridges the gap between optimization and Bayesian learning problems. It introduces an interpolation technique that connects optimization and Bayesian learning. Additionally, the author proposes a generalization bound for infinite wide neural networks and establishes a c... | Rebuttal 1:
Rebuttal: Thank you for your detailed feedback. We appreciate the time and effort you've spent in reviewing our work. Below, we've addressed your points in the order presented.
> It would be beneficial to elucidate the motivation behind the sampling efficiency, denoted as $\text{eff}_{\lambda}$ in Equation... | Summary: The paper provides a perspective interpolating optimization and Bayesian inference. The gradient based optimization procedure gives an output distribution $q(x)$ and the Bayesian inference procedure gives a posterior $p_\lambda(x)$, the two distributions are connect by the weight $w_\lambda$. Then the paper pr... | Rebuttal 1:
Rebuttal: We sincerely appreciate your constructive feedback and recognition of our work. Your comments have helped us to improve our manuscript.
> Weaknesses
Thank you for the suggestion. We will improve the presentation of the mentioned equations for clarity.
> Some of the main theoretical result still... | Summary: This paper proposes a framework to bridge the gap between ERM and Bayesian learning problems.
The authors derived the algorithm-dependent PAC-Bayesian generalization bound for infinitely wide networks where the KL divergence between the posterior distribution obtained by infinitesimal step size gradient descen... | Rebuttal 1:
Rebuttal: Thank you for taking the time to review our work and for your insightful comments and questions. We appreciate your recognition of the novel insights and contributions of our work in bridging the gap between ERM and Bayesian learning.
> Would you write explicitly what the final resulting of the n... | Summary: The paper introduces a novel learning algorithm, Transformative Bayesian Learning (TransBL), which aims to navigate the balance between empirical risk minimization (ERM) and Bayesian learning in the context of overparameterized neural networks. The authors establish a PAC-Bayesian bound for these networks and ... | Rebuttal 1:
Rebuttal: Thank you for the detailed and constructive feedback.
> Strengths
We are grateful for your recognition of the strengths of our paper, especially our theoretical contributions in areas including the interpolation mechanism, usage of PAC-Bayesian bounds, and analysis of energy and entropy change.
... | null | NeurIPS_2023_submissions_huggingface | 2,023 | Summary: The paper derives an algorithm-dependent PAC-Bayesian generalization bound for infinitely wide neural networks that is based on the KL divergence between posterior obtained by gradient flow and a Gaussian prior. The work shows how to transform optimization into importance sampling through interpolation, coined... | Rebuttal 1:
Rebuttal: We deeply appreciate your recognition of the clarity, motivation, and contributions of our work.
> It was sometimes hard to follow the setup of the experimental section (maybe dedicate a separate section in the appendices for details). Having a better picture of the details (especially how infere... | null | null | null | null | null | null |
An Alternative to Variance: Gini Deviation for Risk-averse Policy Gradient | Accept (poster) | Summary: The paper points to limitations of the mean-variance criterion for risk-averse RL, and of the methods that optimize it. Instead, the paper proposes an alternative risk measure relying on Gini Deviation. The paper shows how to derive the PG for this metric, and demonstrates that it learns risk-averse policies m... | Rebuttal 1:
Rebuttal: We thank the reviewer for the review. We address main concerns below.
**For Weaknesses**
1. **Regarding the notion of coherence.**
Thanks for pointing out the coherent risk measure. We are aware of this category of risk as mentioned in Section 1. However, as indicated by our title and abstract,... | Summary: This paper considers the problem of risk-sensitive reinforcement learning in the policy optimization setting, under the mean-variance risk criteria. They provide background on the problem and demonstrate the issues with learning an optimal policy due to the estimation of the variance requiring double samples. ... | Rebuttal 1:
Rebuttal: We thank the reviewer for the positive review. We address main concerns below.
+ **What is the space $M^\infty$ introduced in Definition 1?**
We apologize for this typo. It should be $L^\infty$ to represent the set of bounded random variables, e.g., see original definition in Equation 1 of the p... | Summary: The paper proposes to restrict the Gini deviation of the return for risk-averse reinforcement learning (RARL). This is motivated by a discussion of the shortcomings of variance-based RARL methods, such as scale dependence and instability due to large squared returns. A derivation of an expression for the gradi... | Rebuttal 1:
Rebuttal: We thank the reviewer for the positive review. We address main concerns below.
**For Weaknesses**
1.**The current version does not really discuss the proposed algorithm but jumps right into the experiments.**
The full algorithms are in Section 9 in appendix due to page limit.
The procedures of... | Summary: This paper presents a new approach to measure the risk for risk-averse reinforcement learning problems.
To be more precise, traditional methods usually use the variance of total return or per-step reward as the measurement of risk, since it is known that the variance is easy to calculate. Though easy to calcu... | Rebuttal 1:
Rebuttal: We thank the reviewer for the positive review. We find the questions in "Weaknesses" and "Questions" are closely related. We combine the related ones and address main concerns below.
1. **Regarding the draw back, bias and variance of the sample based estimator**
As discussed in the limitation se... | Rebuttal 1:
Rebuttal: We include a PDF file which contains additional results requested by reviewer cQhS, regarding the sensitivity to $\lambda$
Pdf: /pdf/c71d4827f335a541ccb794d3c3c105e769180bd8.pdf | NeurIPS_2023_submissions_huggingface | 2,023 | Summary: This paper discusses the challenges of mean-variance RARL methods, and proposes the Gini deviation as a substitute to variance penalization.
Strengths: The paper is clearly written and well-motivated (by the disadvantages of return and instantaneous reward variance). The proposed method is evaluated in severa... | Rebuttal 1:
Rebuttal: We thank the reviewer for the positive review. We address main concerns below.
+ **It seems the proposed algorithm has to sort its samples in every iteration (L245-248) which seems computationally expensive, and which may be sensitive to the parameters chosen for $n$?**
The sample here correspo... | null | null | null | null | null | null |
POMDP Planning for Object Search in Partially Unknown Environment | Accept (poster) | Summary: This paper addresses the problem of finding efficient policies for a robot to search for a known target object in a physical environment. The robot can find the target object by either moving around to change its field of view or by moving objects that occlude the target object to reveal the target object. Wh... | Rebuttal 1:
Rebuttal: $\textbf{Response about weakness}$:
W1. This paper...
$\textbf{Answer}$: We appreciate the reviewer's insightful feedback regarding the Prior Work section. Apologies for overlooking the concept of 'Mechanical Search' in our initial version. We have revised the paper and connected our framework w... | Summary: This paper presents a POMDP based approach for finding objects in partially unknown environments. They present a planning algorithm that reuses the belief tree and uses a fake object to guide the search. The system is tested in four simulated scenes in Gazebo and outperforms a POMCP baseline.
Strengths: - ... | Rebuttal 1:
Rebuttal: $\textbf{Response about weakness}$:
W1. The main experiment of the paper is a set of comparisons to POMCP [13], which was published in 2010. The related works section of the proposed work mentions several other more modern approaches. It is not clear why these works were not also compared to.
$\... | Summary: This paper proposes a method for mobile robots to efficiently search for objects in complex environments. They use a Partially Observable Markov Decision Process (POMDP) formulation and a planning algorithm called Growing Partially Observable Monte-Carlo Planning (GPOMCP) to improve the robot's success rate an... | Rebuttal 1:
Rebuttal: $\textbf{Response about weakness}$:
W1. The whole...
$\textbf{Answer}$: Thank you for providing valuable comments. While some parameters require manual configuration, it's worth noting that the majority of these parameters do not significantly impact the final performance of the framework. Most ... | Summary: The paper presents a full object search pipeline based on a MCTS solver for POMDPs. The object are represented by a octo-grid containing the odds of the object being is the target or an obstacle. A 3D point cloud map of the environment *without any object* is given to the algorithm, allowing for easier object ... | Rebuttal 1:
Rebuttal: $\textbf{Response about weakness}$:
W1. The are not many other baselines......
$\textbf{Answer}$: Thank you for your valuable comments. We appreciate the suggestion to consider more baselines for comparison. Please refer to the “Baseline methods” section of the global comments file for more deta... | Rebuttal 1:
Rebuttal: $\textbf{Baseline methods}$
$\textbf{Answer}$: We appreciate the suggestion to consider more baselines for comparison. However, as a system-level application, it is generally fair because many object search approaches are based on the pure POMCP method with different problem settings, eg.[1] [2].... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
A Batch-to-Online Transformation under Random-Order Model | Accept (poster) | Summary: This is a paper about online learning in the random arrival/order model, i.e., the adversary chooses samples but they are randomly permuted before being presented to an online algorithm. The authors propose a general tool that establishes a reduction to the offline/batch setting. The bounds that one can obtain... | Rebuttal 1:
Rebuttal: We thank the reviewer for the insightful comments. In the following, we address the raised concerns point by point. Minor comments and typos have already been addressed and corrected in our manuscript.
# 1. Technical Novelty
We respectfully **disagree** with the reviewer's comment suggesting tha... | Summary: This paper provides a general framework for how to use good offline methods to construct online methods under the random-order setting. It points out the importance of the average sensitivity in the transformation from good offline methods to good online methods.
Strengths: 1. The framework is very general f... | Rebuttal 1:
Rebuttal: We thank the reviewer for the insightful and positive comments! In the following, we address the raised concerns point by point.
# 1. Recover $\epsilon = 0$ (exact regret)
We can obtain exact regret for our algorithms. Using the definition of approximate regret, we can rearrange the terms and ob... | Summary: The paper presents a general method for converting an offline approximation to a learning problem into an online algorithm for the random-order problem which enjoys low (epsilon-approximate) regret. Specifically, the regret of this method depends on the sensitivity of the offline approximation; intuitively, th... | Rebuttal 1:
Rebuttal: We thank the reviewer for the positive comments! We hope to address the concerns raised in the following response.
> The offline-to-online conversion itself is quite simple and does not introduce any algorithmic techniques (arguably, this is the first conversion that comes to mind).
We believe ... | Summary: The authors present a framework for reducing online to offline learning. More presicely, they consider random order online learning where at each round a random point of some domain $X$ is presented to the learner (as opposed to some adversary choosing the example). They first show that an offline algorithm wi... | Rebuttal 1:
Rebuttal: We thank the reviewer for the insightful comments!
# Applicability in adversarial setting and comparison to the previous bounds
We thank the reviewer for the constructive advice, we added a comparison table for the results we obtained in our manuscript.
We believe that the online-to-offline red... | Rebuttal 1:
Rebuttal: We thank the reviewers for their insightful comments and constructive advice!
We would first like to highlight that our work is primarily theoretical in nature, as our focus lies in exploring and analyzing the theoretical foundations of the general framework proposed. Moreover, the related works ... | NeurIPS_2023_submissions_huggingface | 2,023 | Summary: The paper considers the random order model originally proposed by Garber et al. '20, and presents a generic framework to take any (approximate) offline algorithm and via usage of coresets turn it into an online algorithm with low average sensitivity, which can be then used to obtain an (approximate) regret bou... | Rebuttal 1:
Rebuttal: We thank the reviewer for the insightful comments! Due to the space limit, we address the major comments here.
# 1. Comparison to previous works.
It is possible to convert our results to standard regret bounds. Using the definition of approximate regret, we can rearrange the terms and obtain $\m... | null | null | null | null | null | null |
Universal Gradient Descent Ascent Method for Nonconvex-Nonconcave Minimax Optimization | Accept (poster) | Summary: The paper proposes the doubly smoothed gradient descent ascent method and studies its convergence under nonconvex-KL and nonconvex-concave minimax problems.
Strengths: This paper proposes a novel model to address the limiting cycle phenomena in nonconvex-nonconcave minimax optimization. The theoretal analysis... | Rebuttal 1:
Rebuttal: Thanks for your comments. We hope the clarification can clear your concerns and may reevaluate our contribution.
**Q1: Superiority of DSGDA compared with other methods.**
DSGDA can directly achieve both $\epsilon$-GS and $\epsilon$-OS points at the same rate. Furthermore, its convergence rate s... | Summary: This paper addresses the nontrivial general minimax problem by proposing a novel algorithm called Doubly Smoothed Gradient Descent Ascent (DSGDA). The authors provide both experimental and theoretical evidence to support their claims.
Strengths: 1. Overview I enjoyed reading this work. It's well written and o... | Rebuttal 1:
Rebuttal: Thanks for your comments. We are happy that the reviewer enjoys our paper and would like to thank you for your insightful comments. Here are some illustrations based on your questions/suggestions.
**Q1: Superiority of the algorithm under other settings.**
We kindly request the reviewer to confi... | Summary: This paper introduces a novel single-loop algorithm called the doubly smoothed gradient descent ascent method (DSGDA). DSGDA effectively balances the primal and dual updates, eliminating limit cycles in various challenging nonconvex-nonconcave scenarios in the literature. The paper establishes that under a one... | Rebuttal 1:
Rebuttal: We are happy that the reviewer enjoys our paper and would like to thank you for your insightful comments. Below we provide a point-by-point response to your comments and questions.
**Q1: Connection between theory and practice: It would be beneficial to establish a clearer connection between the e... | Summary: Constructing a method that finds an optimal point of nonconvex-nonconcave problems is of interest. Built upon the "one-sided" smoothed GDA (SGDA) that exploits the Moreau-Yosida smoothing technique, this paper studies the "doubly" smoothed GDA (DSGDA), applying the smoothing to both the primal and dual variabl... | Rebuttal 1:
Rebuttal: Thanks for your constructive comments. We hope the clarification that we emphasize may help for a better understanding of our contributions.
**Q1: Does SGDA converge under the one-sided KL setting? As there other methods that work under the one-sided KL setting?**
Recent work [1] has demonstrat... | Rebuttal 1:
Rebuttal: We appreciate and thank the reviewers for their instructive comments. Below we provide our response to some common comments and questions. We believe that our responses will effectively address your major concerns, leading to a re-evaluation of our paper's contribution and quality.
**Q1: Is there... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Communication-Efficient Federated Bilevel Optimization with Global and Local Lower Level Problems | Accept (poster) | Summary: The paper studied the federated bilevel optimization problems, in which the upper level losses are non convex and lover level losses are strongly convex. Two algorithms, named as FedBiOAcc and FedBiOAcc-Local, were proposed for whether the lower level problems requires to be consensus or not. The authors prov... | Rebuttal 1:
Rebuttal: Thanks for spending time reviewing our manuscript. Below are our responses to your concerns and questions:
Response to Weakness:
Our main contributions are three folds: Firstly, we view the estimation of hyper-gradient as solving a federated quadratic problem and therefore can get the unbiased e... | Summary: This paper proposes an algorithm named FedBiOAcc that solves the federated bilevel optimization problem. In particular, the hyper-gradient of the bilevel optimization problem is evaluated by solving a reformulated quadratic federated optimization problem. The upper-level problem, lower-level problem and the hy... | Rebuttal 1:
Rebuttal: Thanks for spending time reviewing our manuscript and providing insightful comments. Below are our responses to your questions:
Response to Weakness:
Many real-world applications in FL exhibit a nested structure, and therefore can be solved as a federated bilevel problem. Firstly, the data-clean... | Summary: This paper focuses on the problem of bilevel optimization in a federated learning or FL environment. Bilevel optimization has various FL applications, and a few recent works proposed versions of bilevel optimization schemes for FL. A challenging step in bilevel optimization is the computation of the "hypergrad... | Rebuttal 1:
Rebuttal: Thanks for your detailed and insightful comments.
Response to Weakness:
We agree with the reviewer over the increased hyper-parameter space. In fact, the optimal convergence rate of the momentum-based variance reduction technique is achieved by carefully balancing the learning rate, momentum co... | Summary: this paper proposed new algorithms for federated bilevel optimization. The authors apply the idea of [1] to the federated setting, which views the hypergradient computation as solving a quadratic subproblem. Combining with the STORM algorithm, the author proposed fedbioacc with a convergence guarantee based on... | Rebuttal 1:
Rebuttal: Thanks for spending time reviewing our paper and provide comments. Below are our responses to your concerns:
**W1**: We agree with the reviewer that bilevel optimization is harder than the minimax optimization problems. However, our three-level perspective of the bilevel optimization can eliminat... | null | NeurIPS_2023_submissions_huggingface | 2,023 | Summary: This paper proposed a communication efficient federated bilevel algorithm in both global and local lower-level problem setting that can achieve the best theoretical communication complexity. Extensive numerical experiments are provided to test the effectiveness of the proposed algorithm.
Strengths: They prov... | Rebuttal 1:
Rebuttal: Thanks for your insightful comments. Below are our responses to your questions and concerns:
Response to weakness:
**W1** : Although the idea of viewing hyper-gradient estimation as solving a quadratic problem is investigated in several papers, these paper only consider the non-distributed setti... | null | null | null | null | null | null |
Online Nonstochastic Model-Free Reinforcement Learning | Accept (poster) | Summary: This paper introduces an extension to model-free RL that has its roots in non-stochastic control. The main idea is to adjust a nominal control of a base policy (from e.g. a RL algorithm) by a control signal that can be computed from estimates of the process noise. As a main contribution, the authors present th... | Rebuttal 1:
Rebuttal: Thank you for your time and thoughtful review. We address the main points below:
Weaknesses:
Our contribution is mostly conceptual and theoretical, thus we focused on methods with provable guarantees. Using a learned model to compute disturbances could be interesting, but we don’t see how to de... | Summary: This work introduces the notion of disturbance-based policies for model-free reinforcement learning, as opposed to traditional state-based policies. The disturbances capture unmodeled deviations from observed dynamics. In the model-free setting, since these disturbances are not known to the learner, the paper ... | Rebuttal 1:
Rebuttal: Thank you for your time and thoughtful review. We address the main points below:
We note that PD2 is actually quite general, and does not require specific domain-engineered signals from the environment. For example, in all our experiments, we use the coordinates of the state for each auxiliary ... | Summary: This work considers model-free RL in the setting with additive disturbances in the environments' forward dynamics. In particular, it focuses on "disturbance-based policy" which adds a correction policy to the vanilla state-based policy. Because the disturbances are unknown in model-free RL, the correction poli... | Rebuttal 1:
Rebuttal: Thank you for your time and thoughtful review. We address the main points below:
Weaknesses:
We acknowledge that paper would be stronger with non-uniform disturbances. We will add detailed experiments for sinusoidal disturbances and adversarial sign gradient disturbances.
Benchmarking against... | null | null | Rebuttal 1:
Rebuttal: **Please see attached pdf file.**
Dear reviewers, we thank you for your time and effort in reviewing our paper!
The main weakness jointly pointed by the reviewers is lack of experimentation with more varied noise patterns, as opposed to stochastic noise.
We attach below in pdf format two exper... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Optimal Time Complexities of Parallel Stochastic Optimization Methods Under a Fixed Computation Model | Accept (poster) | Summary: Consider the classical statistical risk minimization framework with stochastic gradients, and propose new lower bounds on the optimal time complexity. Also propose a new method Rennala SGD that attains the lower bound. This work makes the case for asynchronous methods.
In particular, this work proposes a new ... | Rebuttal 1:
Rebuttal: Thank you for the interesting comments!
> Lacking empirical evaluations, the only such experiments are on quadratics, buried deep inside the appendix. While these plots are compelling, I would encourage the authors to consider small-scale numerical results on other machine learning workloads.
Ou... | Summary: This paper proposes a protocol that generalizes the classical oracle framework approach. Using this protocol, it establishes minimax complexities for parallel optimization methods that have access to an unbiased stochastic gradient oracle with bounded variance.
Strengths: - The motivation of this paper is cle... | Rebuttal 1:
Rebuttal: Thank you for the review!
> The challenge of theoretical analysis should be made more clear, which highly relates to the contribution of this paper.
Let us briefly explain the main challenges of the paper:
1. Even before we started proving the lower bounds, the main challenge was to understand h... | Summary: This paper studies time complexity of parallel stochastic optimization, establishes lower bounds, and proposes algorithms achieving matching complexity results.
Strengths: Based on my reading of the main paper (I haven't read the proofs), this is a solid theoretical work. The theoretical results improve widel... | Rebuttal 1:
Rebuttal: Thank you for the comments!
> The organization of the paper is a bit reader unfriendly. The assumptions are mentioned at multiple places, starting with page 1, but are formally introduced almost at the end (page 8). This makes the reading tedious.
We define assumptions about our main problem (1)... | Summary: This paper studies the minimax complexities of distributed asynchronous stochastic optimization methods. By extending the oracle framework previously used in the literature, it establishes new (lower) lower bounds for parallel optimization methods. Based on the insights from their proof, they propose a new alg... | Rebuttal 1:
Rebuttal: Thank you for your positive review!
> Your model assumes fixed delays per worker: could your analysis also hold with stochastic ones?
> You experimented with a fixed delay of ... Did you also experiment with real-life delays experienced naturally in the distributed asynchronous setting...
Great... | null | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Grammar Prompting for Domain-Specific Language Generation with Large Language Models | Accept (poster) | Summary: This paper studies generating strings from highly structured languages for large language models (LLMs). The authors propose grammar prompting to enhance LLMs to use external knowledge and domain-specific constraints, expressed through a grammar expressed in Backus–Naur Form. Serving as intermediate reasoning ... | Rebuttal 1:
Rebuttal: We thank the reviewer for the feedback and helpful comments.
We would like to emphasize the contribution and novelty of our work: (1) We focus on enabling LLMs to generate highly structured languages from just a few exemplars, a task where neither the standard nor the chain-of-thought prompting p... | Summary: The authors investigate the effectiveness of grammar prompting as a simple approach to help LLMs utilize external knowledge and domain-specific constraints. This approach is motivated by the goal of enabling LLMs to generate DSL outputs that differ significantly from those encountered during pretraining. The a... | Rebuttal 1:
Rebuttal: We thank the reviewer for the positive feedback and constructive suggestions.
> Minor: The contributions and novelty should be stated at the end of the introduction section.
We will make the contributions/novelty clearer in the introduction.
> Have you explored other grammar forms?
We explored... | Summary: This paper presents an approach for improving few-shot prompting of LLMs for tasks that produce structured outputs, where the structure can be described by a grammar in BNF form. The approach is similar to chain-of-thought prompting: for a given input x, the LLM is prompted to generate the minimal subset of gr... | Rebuttal 1:
Rebuttal: We thank the reviewer for the positive feedback and the recognition of the strengths of our work.
> “One minor weakness is that it would benefit the paper to also show results on other LLMs …”
In our recent experiments with PaLM2-Large using grammar prompting, we utilized the same 100 examples... | Summary: This work explores grammar prompting as a simple approach for enabling LLMs to use external knowledge and domain-specific constraints, expressed through a grammar expressed in Backus–Naur Form (BNF), during in-context learning. Grammar prompting augments each demonstration example with a specialized grammar th... | Rebuttal 1:
Rebuttal: We thank the reviewer for the positive feedback and the recognition of the strengths of our work.
> “In experiment part, how do you ensure the results are enough to verify the proposal? For example, the superiority in molecule generation is not well established compared with the other graph-based... | Rebuttal 1:
Rebuttal: We appreciate all reviewers' time and efforts in reviewing our paper. We are glad to find that reviewers generally recognized our key contributions and clear presentation of our paper:
**Method**: “The approach is elegant and seems simple to implement” [rfkj], “Novel method for prompting and cons... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Constrained Proximal Policy Optimization | Reject | Summary: This paper introduces the constrained version of PPO that is devised for solving constrained MDP problems in the discounted reward setting with discounted cost constraints.
Strengths: The paper presents the constrained PPO. It also gives some analytical results and presents a heuristic algorithm that is see... | Rebuttal 1:
Rebuttal: **We would like to express our sincere gratitude to you for valuable and constructive comments. Our responses are as follows.**
Some comments from reviewers are not displayed in full but replaced by ... to save space.
**W1**: Inadequate literature survey, ... , in the function approximation sett... | Summary: The paper introduces Constrained Proximal Policy Optimization (CPPO) for Constrained Reinforcement Learning (CRL). The CPPO method is designed to overcome the limitations of existing methods by offering a first-order feasible region method that doesn't require dual variables or second-order optimization, and i... | Rebuttal 1:
Rebuttal: **We would like to express our sincere gratitude to you for valuable and constructive comments. Our responses are as follows.**
**W1**: Clarity: the paper could benefit from improvements in clarity and readability. The presentation of their ideas is somewhat dense and could be difficult for reade... | Summary: This paper proposes a novel first-order feasible method, CPPO, for efficient constrained reinforcement learning. The proposed approach integrates the Expectation-Maximization (EM) framework to solve the policy optimization problem by treating the CRL as the probabilistic inference. In the E-step, CPPO calculat... | Rebuttal 1:
Rebuttal: **We would like to express our sincere gratitude to you for valuable and constructive comments. Our responses are as follows.**
**W1**: I suggest further polishing and improving the mathematical formulation and notations to make them more rigorous. For example, the objectives and constraints in (... | Summary: This paper proposes a novel CPPO method to solve the contained RL (CRL) problem. Specifically, CPPO leverages the probabilistic inference and converts the CRL problem formulation based on the probabilistic ratio, resulting in the first-order optimization solution. CPPO also develops the recovery update method ... | Rebuttal 1:
Rebuttal: **We would like to express our sincere gratitude to you for valuable and constructive comments. Our responses are as follows.**
**W1**: As stated in Section 3.2.1, CPPO builds on CVPO and has two main differences (using advantage instead of q and using the probability ratio instead of directly ca... | Rebuttal 1:
Rebuttal: **We would like to express our sincere gratitude to reviewers for valuable and constructive comments.**
Many reviewers have raised questions regarding the novelty of our work. It is important to emphasize that the proposed CPPO method is an extension of PPO in constrained RL field, rather than a ... | NeurIPS_2023_submissions_huggingface | 2,023 | Summary: This paper studies constrained reinforcement learning problems. It proposes a new EM-type algorithm and designs a heuristic version for practical use. The authors also conduct some numerical experiments to validate the performance of the algorithm.
Strengths: 1. The algorithm is first-order and thus computati... | Rebuttal 1:
Rebuttal: **We would like to express our sincere gratitude to you for valuable and constructive comments, thanks for your support. Our responses are as follows.**
**W1**: The algorithm looks very similar to CVPO and only has some small modifications.
**R1**: Thanks for your comment. We believe the propose... | Summary: This paper focuses on constrained reinforcement learning and proposes a method called Constrained Proximal Policy Optimization (CPPO). Experimental results demonstrate the improved performance in terms of episodic return and episodic cost.
Strengths: + The proposed CPPO achieves improved balance between retu... | Rebuttal 1:
Rebuttal: **We would like to express our sincere gratitude to you for valuable and constructive comments, thanks for your support. Our responses are as follows.**
**W1**: The novelty and contribution of this work is not clear. The difference between the proposed CPPO and CVPO does not seem to be very signi... | null | null | null | null |
Fast Approximation of Similarity Graphs with Kernel Density Estimation | Accept (spotlight) | Summary: Given a set of n points and a pairwise kernel k(x,y), the fully connected similarity graph K is the weighted graph that has an edge between every pair x,y with weight k(x,y). The paper presents an algorithm for computing a sparse graph G that approximates K in a certain sense that is useful for spectral cluste... | Rebuttal 1:
Rebuttal: Many thanks for your positive and detailed report, and valuable comments. We will take these into account when preparing the final version of our paper. Here are our responses to your questions and comments.
**Question 1.**
We agree that there is some similarity between our algorithm and the one ... | Summary: This work presents an algorithm framework for creating a (weighted) sparse similarity graph, which is an important object for a variety of ML problems. The algorithm samples pairs of points to include in the graph using a sampler based on kernel density estimation (KDE). The idea is to recursively split the da... | Rebuttal 1:
Rebuttal: Many thanks for your positive and detailed report, and valuable comments. We will take these into account when preparing the final version of our paper. Here are our responses to your questions and comments.
**Question 1: KDE preprocessing and query time.**
In our paper, $T_{\mathrm{KDE}}$ inclu... | Summary: The paper proposed an efficient approximation of the similarity graph using the Kernel Density Estimation (KDE) method as a black box. The proposed framework can preserve the clustering structure and can be run in nearly linear time. Lastly, experiments on synthetic and image segmentation datasets are provided... | Rebuttal 1:
Rebuttal: Many thanks for your positive and detailed report, and valuable comments. We will take these into account when preparing the final version of our paper. Here are our responses to your questions and comments.
**Question 1.**
Our method constructs a sparsifier which has the same cluster structure a... | Summary: Authors proposes a fast approximation method for constructing sparse similarity graphs from data for spectral clustering. They leverage Sun and Zanetti (2019)'s sampling-based approach to find "cluster-preserving sparsifier" for similarity graphs. This is combined with black box KDE and binary search algorithm... | Rebuttal 1:
Rebuttal: Many thanks for your very positive report. Here are our responses to your questions.
**Question 1.**
In Figure 1 of the attached PDF file, we provide an updated version of the experiment on the two moons dataset to include the results of the SZ algorithm.
The SZ algorithm iterates over every ed... | Rebuttal 1:
Rebuttal: We thank all reviewers for their positive and detailed reviews. We respond to their specific questions individually, with reference to some additional figures and table in the attached PDF.
Pdf: /pdf/17fc5a9b4d9bcc04ab397d481f3c85f7d9cf248a.pdf | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Recurrent Temporal Revision Graph Networks | Accept (poster) | Summary: Temporal graphs are more accurate for modeling real-world scenarios compared to static graphs, but the current approach of extending neighbor aggregation from static graphs to temporal graphs is computationally expensive when considering all historical neighbors. To address this, the authors propose a novel fr... | Rebuttal 1:
Rebuttal: Thank you for your constructive feedback which we will consider seriously in our revised paper.
**Regarding your comment on weakness 1.** Thank you for this valuable feedback. We will add a simple explanation of what heterogeneity means in the context of temporal revision. In simple terms, it m... | Summary: The paper studies the problem of temporal graph learning. The authors propose the recurrent temporal revision (RTR) layer, which involves learning a hidden state for each node and utilizing recursive neural networks to integrate information from all neighbors. Additionally, the authors introduce heterogeneity ... | Rebuttal 1:
Rebuttal: Thank you for your insightful feedback which we will consider seriously in our revised paper.
**Regarding your comment W1.** We would like to emphasize that long-term user behavior has been a major focus in recommender systems. Numerous studies, such as [R1, R2, R3, R4, R5], have demonstrated it... | Summary: In this paper, the authors propose a novel aggregation layer based on a recurrent neural network dubbed recurrent temporal revision (RTR) for temporal graph networks to solve the dynamic temporal graph embedding problem. Specifically, a new aggregation function is proposed, which encodes neighbor state informa... | Rebuttal 1:
Rebuttal: Thank you for your constructive feedback which we will consider seriously in our revised paper.
**Regarding your confusion about the correspondence between the ablation experiments and the two proposed measurements.** We would like to clarify that while the first two experiment sections (5.1 and... | Summary: The paper introduces a novel framework called Recurrent Temporal Revision (RTR) to address the challenge of neighbor subsampling in temporal graphs, an issue commonly encountered in real-world applications like social networking, recommender systems, traffic forecasting, and crime analysis. RTR serves as a sta... | Rebuttal 1:
Rebuttal: Thank you for your insightful feedback which we will consider seriously in our revised paper.
**Regarding your consideration of the rationale behind our new definition of expressiveness.** We develop the set of new expressiveness measures mainly for the following two reasons. First, as you have ... | Rebuttal 1:
Rebuttal: We sincerely thank all the reviewers for their constructive feedback. We provide the extended Table 3 in the attached PDF, while all other concerns reviewers raised are addressed in their respective rebuttals.
Pdf: /pdf/bdf7fd885b9edf50ce797a9df9523e84a47f580b.pdf | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Counterfactually Comparing Abstaining Classifiers | Accept (poster) | Summary: This paper outlines a method for estimating the counterfactual performance of an abstain classifier on cases where it abstained (what would have happened if it had not abstained?). Using tools from causal inference and the potential outcomes framework in particular, they cast this is a missing data problem, an... | Rebuttal 1:
Rebuttal: Thank you for your review. We sincerely hope that our point-by-point responses below address your concerns.
---
> Motivation
Please see our common response to all reviewers for this point.
---
> Novelty
First, highlighting and formalizing the connection between abstaining classifier evaluatio... | Summary: This paper proposes to compare black-box abstaining classifiers (neither the base classifier nor the abstention mechanism is known to the evaluator) by the counterfactual score, that is the expected score of the abstaining classifier had it not been given the option to abstain. They prove that this quantity is... | Rebuttal 1:
Rebuttal: Thank you for your feedback on our work and for acknowledging the practical importance of our work and the problem. See below for our responses to each of your concerns and questions.
----
> As stated in the paper, the counterfactual score is unidentifiable if abstentions are deterministic whi... | Summary: The authors focus on the problem of evaluating abstaining classifiers while also taking account their counterfactual predictions had they not abstained on certain data points. As an evaluation metric they consider the expected accuracy of the classifiers over both their actual and counterfactual predictions, w... | Rebuttal 1:
Rebuttal: We sincerely appreciate your positive and comprehensive feedback on our work.
In our revision, we incorporated all of your editorial notes, including moving Table 1 to the beginning of Section 2 and having all table captions appear before the tables themselves. | Summary: The paper proposes a new way to evaluate counterfactually, abstain classifiers. By considering the task of selective classification, the authors reformulate the problem into a Missing data problem which can thus be seen as a causal inference (Counterfactual) problem.
The authors also state the corresponding as... | Rebuttal 1:
Rebuttal: Thank you for your review of our work and for acknowledging the novelty and clarity of our exposition. See below for our responses to each of your concerns.
----
> My biggest concern of the paper is the motivation. As stated in their example 1.1 the example they have given is quite contrived and... | Rebuttal 1:
Rebuttal: We appreciate all of our reviewers for their feedback on our work. We are particularly thankful to them for acknowledging the problem’s significance as well as the novelty of the connections we elucidate between abstaining classifiers, black-box evaluation, and causal inference.
A main concern s... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Parameterizing Context: Unleashing the Power of Parameter-Efficient Fine-Tuning and In-Context Tuning for Continual Table Semantic Parsing | Accept (poster) | Summary: The author presents a novel approach that combines PEFT and ICT to train a continual table semantic parser. The proposed solution is based on a teacher-student framework. Two task streams are developed from WikiSQL and Spider semantic parsing benchmarks. Experiments using C3 and baselines are conducted using t... | Rebuttal 1:
Rebuttal: We are very grateful to you for providing us with valuable feedback and suggestions for our paper. We will provide explanations and clarifications for each weakness and question.
#### For Weakness 1:
Before us, the problem formulation in $\rm{Section} \ 2.2$ was proposed by [7], so we did not inc... | Summary: The paper introduces a method combining parameter-efficient fine-tuning (PEFT) and in-context tuning (ICT) to address the issues of overfitting and catastrophic forgetting in training a continual table semantic parser with limited training examples. Through a task-adaptive PEFT framework and a teacher-student ... | Rebuttal 1:
Rebuttal: We are very grateful to you for providing us with valuable feedback and suggestions for our paper. We will provide explanations and clarifications for each weakness and question.
#### For Weakness 1:
Our method is fundamentally different from that of [Gao et al.] for the following two reasons:
1.... | Summary: This paper introduces a novel method for training a continual table semantic parser, which aims to translate natural language into SQL based on task-specific tables with limited training examples. The proposed method integrates parameter-efficient fine-tuning (PEFT) and in-context tuning (ICT) to overcome the ... | Rebuttal 1:
Rebuttal: We are very grateful to you for providing us with valuable feedback and suggestions for our paper. We will provide explanations and clarifications for each weakness and question.
#### For Weakness 1:
We chose GRAPPA as the backbone PLM for baselines for two main reasons:
1. Unlike T5, GRAPPA is p... | Summary: The paper proposes a new continual learning method C3 for table semantic parsing which combines parameter-efficient fine-tuning (PEFT) and in-context tuning (ICT). The C3 framework contains a teacher network (ICT) that extracts contextual information from demonstrations and a student network (PEFT) that learns... | Rebuttal 1:
Rebuttal: We are very grateful to you for providing us with valuable feedback and suggestions for our paper. We will provide explanations and clarifications for each question.
#### For Question 1:
We hypothesize that when the teacher model and student model share the same architecture (T5), the greater the... | null | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Partial Counterfactual Identification of Continuous Outcomes with a Curvature Sensitivity Model | Accept (spotlight) | Summary: This paper studies the partial identification of counterfactual probabilities from observational data. More specifically, the authors consider the standard bandit model containing a treatment A and a reward Y; exogenous variables U exist affecting both A and Y. The treatment A is a binary variable; while the r... | Rebuttal 1:
Rebuttal: We are grateful for the review. It is great to hear that you found our paper well-written and well-organized, and that you think that the paper will have an impact on the causal inference community. We would like to address mentioned weaknesses and open questions.
### Response to weaknesses
We wa... | Summary: The authors study the problem of counterfactual identification of continuous outcome variable with a specific causal graph. They show that the expected counterfactual outcome of (un)treated has non-informative bound when the function class is arbitrary smooth. Then, they introduce a new model called CSM that g... | Rebuttal 1:
Rebuttal: Thank you for your review! We appreciate that you found our paper interesting, important, and well-written. We would like to respond to the mentioned weaknesses.
- Regarding the generalizability of our method to the more complex graphs: Our CSM generalizes in theory to other Markovian-SCMs (see d... | Summary: This paper studies the partial identifiability of counterfactual queries and proposes an interesting proof of general non-identifiability resulting in non-informative bounds. The authors build upon the intuition used for this proof to propose a new class of structural causal model using principal curvatures. U... | Rebuttal 1:
Rebuttal: Thank you for the positive evaluation of our work. We are happy that you consider our paper a novel, systematic, and rigorous study. We would like to discuss the mentioned weaknesses and questions in the following.
- We agree that the $\kappa$ in our CSM may be perceived — in parts — as an abstra... | Summary: This paper presents a method for obtaining bounds on continuous counterfactual outcomes. In general, level-3 counterfactual outcomes (what would have happened under action $a'$, given knowledge of what did happen under action $a$) are not identifiable from either interventional or observational data. Prior w... | Rebuttal 1:
Rebuttal: We are very grateful for such a rigorous, helpful, and in-depth review. We also appreciate that the reviewer found our paper clearly written, our method novel, and our contribution well-placed in the related work. Here is our answer to the questions:
- (1) We stress that counterfactual inference ... | Rebuttal 1:
Rebuttal: We are very grateful for the positive and in-depth feedback from the reviewers. We have carefully addressed all of the questions in the individual responses below. We will incorporate all changes (labeled with **Action**) into the camera-ready version of our paper.
We have additionally uploaded ... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Model-Free Reinforcement Learning with the Decision-Estimation Coefficient | Accept (poster) | Summary: The paper studies model-free RL algorithm using DEC as the complexity measure.
Strengths: Extending previous model-based algorithm to model-free algorithm is a nice contribution.
Weaknesses: I believe this paper provides a nice contribution for the DEC family of papers. But I would like the authors to disc... | Rebuttal 1:
Rebuttal: Thank you for the review and useful suggestions!
> I believe this paper provides a nice contribution for the DEC family of papers. But I would like the authors to discuss more on the implication side. Why we care about extending model-free algorithms to bilinear classes? There are too many comple... | Summary: This paper studies a variant of the recently proposed framework of Foster et al. for sequential decision making, based on a concept called "decision estimation coefficient" (DEC). The variant proposed here is based on enhancing the estimation step in the "estimation-to-decisions" (E2D) with an optimistic bias ... | Rebuttal 1:
Rebuttal: Thank you for the positive comments and helpful suggestions!
> While the authors discuss this limitation quite openly, I would have appreciated some more discussion as to where this relatively poor scaling comes from. One contributing factor is certainly the use of batched estimation steps. My un... | Summary: This paper contributes to a line of research on decision-estimation coefficients (and related algorithms). An optimistic variant of the original DEC is introduced and it is shown that this variant can lead to a new related meta-algorithm with optimism. This new structure and algorithm are shown to provide non-... | Rebuttal 1:
Rebuttal: Thank you for the positive comments and helpful suggestions!
> Does optimistic DEC subsume all of the original DEC results? Are there any downsides compared to the original E2D?
Yes, the optimistic DEC—when equipped with Hellinger distance—recovers all of the main guarantees based on the DEC in ... | Summary: This paper adapted the DEC framework and E2D algorithm to incorporate model-free reinforcement learning. Specifically, this paper proposed the optimistic DEC and proved that it appears in the regret bound for the optimistic E2D algorithm with an optimistic estimation oracle. The example of bilinear class is p... | Rebuttal 1:
Rebuttal: Thank you for your positive comments!
> Am I correct that oDEC is always smaller than DEC, and DEC serves as a lower bound, so it is not meaningful to discuss oDEC as a lower bound?
The relation between the oDEC and the DEC depends on the choice of divergence and the model class. As shown by Co... | null | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
ProteinNPT: Improving Protein Property Prediction and Design with Non-Parametric Transformers | Accept (poster) | Summary: This paper introduced ProteinNPT, which is a transformer based semi-supervised learning model. By combining MSA transformer and Non-parametric transformer(NPT), it outperformed their baselines on property and mutation effect predictions. Then they tested their model on the protein design task and showed its po... | Rebuttal 1:
Rebuttal: **C1. The authors mention their novelties in different aspects. In fact, the dedicate a whole subsection for it, which helps detail the novelty. However, beyond the additional data and testing that come with it, most of the novelties appear to be subtle changes or directly borrowed ideas from MSA ... | Summary: The paper researches for the purpose of iterative protein design. It first collects 13 additional fitness metrics to the ProteinGym benchmark. Then, it proposes ProteinNPT, an MSA-Transformer-based model for fitness prediction, and achieved state-of-the-art on ProteinGym. Finally, it applies the model to itera... | Rebuttal 1:
Rebuttal: **C1. Novelty: The way adapts Transformer to property prediction is straightforward. The novelty is limited.**
We would like to first clarify that we do not adapt the standard Transformer architecture to property prediction, but rather apply a non-trivial variant of axial transformer as described... | Summary: his paper introduces a non-parametric transformer variant called ProteinNPT, which is utilized for protein property prediction and design tasks. The study compares ProteinNPT with re-implemented top-performing baselines on the extended ProteinGym benchmark.
The objective of the paper is to address challenges... | Rebuttal 1:
Rebuttal: **C1. The difference with axial transformer [...] better suited within the main body of the paper rather than relegating it to the Appendix.**
We have integrated the points made in Appendix B6 within Sec. 2 as suggested.
**C2. It would be beneficial to include additional mathematical formulations... | Summary: The authors propose to use non parametric transformers to model the ‘fitness’ landscape and the protein primary sequences with a Bert-like model. Both tokens and continuous attributes can be masked in the modeling. Additionally, the authors introduce several novel datasets to test and benchmark their model. Fi... | Rebuttal 1:
Rebuttal: Thank you for the very thoughtful comments and suggestions. We address each of your questions below, including several additional analyses which we believe significantly strengthen our submission.
**C1. I am disappointed about the lack of description of the novel datasets in the main part of the ... | Rebuttal 1:
Rebuttal: Dear reviewers,
We sincerely thank you for the time spent engaging with our paper and really appreciate the thoughtful comments. Based on your feedback, we have conducted additional experiments to further explore the strengths of our proposed approach, and have also clarified all points you had r... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Extraction and Recovery of Spatio-Temporal Structure in Latent Dynamics Alignment with Diffusion Models | Accept (spotlight) | Summary: The authors address the problem of aligning behavior-related neural population dynamics, either within-subject but across different experimental sessions, or between subjects. This problem exists due to inter-subject variability in terms of which neurons are recorded or drift in the recording and is an importa... | Rebuttal 1:
Rebuttal: Thank you for your positive review and insightful questions. We provide clarifications and new results to address your questions below.
> Authors should state whether existing methods for discovering latent structures using diffusional models exist.
To the best of our knowledge, no existing meth... | Summary: Brain-computer interfaces require recalibration to accommodate drifts in the recorded neural populations over time. While there has been some success in aligning neural recordings based on their latent dynamics, deep learning-based alignment methods have also gained popularity since they ignore some of the imp... | Rebuttal 1:
Rebuttal: Thank you for your detailed and constructive comments. We would like to make the following clarifications. Hopefully these will resolve most of your concerns, and they can be taken into account when deciding the final review score.
> They overlooked other methods based on the alignment of latent ... | Summary: * One of the key challenges in analyzing neural recordings is the scalability of models that link behavior and neural population activity across recording sessions or in inter-subject settings.
• When analyzing single-trial neural population activity, past studies have pointed out that neural activity can be u... | Rebuttal 1:
Rebuttal: Thank you for your positive review and insightful questions. We provide clarifications and new results that we have generated to address your questions below. Hopefully these will resolve most of your concerns, and they can be taken into account when deciding the final review score.
> The authors... | Summary: This paper proposes a distribution alignment method (ERDiff), which combines extraction of spatio-temporal structures in latent dynamics from the source distribution and maximum likelihood alignment procedure on the target domain. The proposed method was evaluated on both synthetic and real data (neural recor... | Rebuttal 1:
Rebuttal: Thank you for your positive review and insightful questions. We provide clarifications and new results that we have generated to address your questions below.
> Although the methods have been shown to outperform some of the current techniques, advantage of the proposed approach is not well demons... | Rebuttal 1:
Rebuttal: We would like to express our sincere gratitude to all the reviewers for their insightful feedback and suggestions. We appreciate the positive comments which characterized our work as having a `"clear motivation"` (2uTS), `"novel"` methodological progression (9MfT, wx95), being `"technically solid"... | NeurIPS_2023_submissions_huggingface | 2,023 | Summary: The paper focuses on aligning highly variable neural population activities across days and subjects to stabilize the learning proposes and advance applications such as the brain computer interface. The idea is to train a VAE on one dataset that is hopefully self-consistent and then using a VAE trained on a dat... | Rebuttal 1:
Rebuttal: Thank you for your valuable comments. We would like to make the following clarifications. Hopefully these will resolve most of your concerns, and they can be taken into account when deciding the final review score.
> A cellular neuroscience focused paper with no additional effort to demonstrate a... | Summary: Inter-individual and inter-session variability significantly complicate direct comparison of neural recordings collected over time, degrading trained behavioral models. This can be cast as a more general distribution alignment problem, shared across unsupervised learning. To address neural distribution alignme... | Rebuttal 1:
Rebuttal: Thank you for your detailed and constructive comments. We would like to make the following clarifications. Hopefully these will resolve most of your concerns, and they can be taken into account when deciding the final review score.
> **(1)** It is not clear to what extent the presented findings m... | null | null | null | null |
Calibrate and Boost Logical Expressiveness of GNN Over Multi-Relational and Temporal Graphs | Accept (poster) | Summary: The paper deals with classifying multi-relational graphs, and specifically, the R-GNN architecture, which deals with more than one type of edge by producing a different neighbor set for each of them, aggregating the messages of each neighbor, and then merging them together in a final update layer. The first re... | Rebuttal 1:
Rebuttal: We sincerely thank Reviewer pnMq for the recognition of our theoretical results on both static and temporal settings as well as our proposed novel transformation. Below, we would like to give detailed responses to each of your comments.
___
> **Q1**: Please justify why a study on R-GNN is importa... | Summary: This paper justifies the logical expressiveness of R^2-GNN from the perspective of "universal" graph classes, including multigraphs and infinite graphs. Specifically, the R^2-GNN seems to be identical to the previous ACR-GNN in [1]. The engagement of graph classes makes theoretical statements in finer settings... | Rebuttal 1:
Rebuttal: We sincerely thank Reviewer KFPi for acknowledging the novelty and theoretical significance of our findings within multi-relational and temporal context. In particular, we are enthused by your thought-provoking comments, which motivate us to have deeper contemplation. Below, we would like to give ... | Summary: The paper introduces a new connection between the power of the relational GNNs (working on multi-relational graphs) with a class of first-order logic functions. This class of functions named $\mathcal{FOC}_2$ is a subset of first-order logic functions constrained to having only two variables, but includes coun... | Rebuttal 1:
Rebuttal: We sincerely thank Reviewer a55p for acknowledging the merits of our paper in both theoretical and experimental aspects. In particular, we thank you for your interesting and insightful reviews, which motivated us to formally prove and add the following results (see below). We will add these result... | Summary: The paper extends the logical characterization of GNN --- node classification, by investigating the relational case i.e. multi-relational graphs. The paper generalizes the work of Barcelo et al. [ICLR 2020], who provided the logical characterization of GNNs for node classification in terms of First Order Logic... | Rebuttal 1:
Rebuttal: We sincerely thank Reviewer Yror for the recognition of our original and theoretical results analyzing the expressivity of graph neural networks in the multi-relational and temporal setting. Below, we would like to give detailed responses to each of your comments.
> **Q1**: When using many predi... | null | NeurIPS_2023_submissions_huggingface | 2,023 | Summary: In this paper the authors provide a theoretical analysis of graph neural network (GNN) expressivity on multi-relational and temporal graphs. Specifically, they extend the multi-relational graph convolution network R-GCN to a class of architectures they call R-GNN, which allows for arbitrary aggregation (poolin... | Rebuttal 1:
Rebuttal: We sincerely thank Reviewer p5D8 for the recognition of our original and theoretical results analyzing the expressivity of graph neural networks in the multi-relational setting as well as their potential contribution to the graph neural network community. Below, we would like to give detailed resp... | Summary: This work calibrate the logic expressiveness of R2-GNNs as node classifiers on multi-relational graphs. Motivated by some negative results, authors boost R2-GNNs with a graph transformation, which enables R2-GNN to capture FOC2 formulas. They further extend expressiveness results and graph transformation to te... | Rebuttal 1:
Rebuttal: We sincerely thank Reviewer Hbfa for the recognition of our theoretical results on both static and temporal settings as well as our proposed novel transformation. Below, we would like to give detailed responses to each of your comments.
___
> **Q1**: There is no scalability comparison between the ... | Summary: This paper proposes the R^2-GNN that captures the logic features in the graph. The proposed methodology largely resembles the previous work of ACR-GNN, where a GNN variant is proposed to capture the logic feature for graphs with no relation types. The proposed R^2-GNN extends the previous method by considering... | Rebuttal 1:
Rebuttal: We sincerely thank Reviewer auoR for the positive feedback and insightful comments. Below, we would like to give detailed responses to each of your comments.
> **Q1**: Why the authors pick the R-GNN as the base model. The authors should evaluate their method on different base models.
**A1**: It... | null | null |
FeCAM: Exploiting the Heterogeneity of Class Distributions in Exemplar-Free Continual Learning | Accept (poster) | Summary: This paper addresses the problem of continual learning, which holds great importance in the machine learning field. The authors argue that using the Mahalanobis distance is more optimal than the Euclidean metric for handling new classes. The proposed method is evaluated on various continual learning settings a... | Rebuttal 1:
Rebuttal: Thanks for your feedback and suggestions.
__W1:__ We agree that usually one method sticks to one setting. Exemplar-free methods use 50% of data in the first task as equally splitting is a much more challenging setting which is usually tackled by storing exemplars or by expanding the network in ne... | Summary: The authors claim that feature space of a learnt network with old classes is more scattered in newly learned classes and this prevents Euclidean distance to represent the distribution properly. Therefore, they have adopted Mahalanobis distance instead to handle the heterogenety of the newly coming class distri... | Rebuttal 1:
Rebuttal: Thanks for your comments and suggestions.
__W2:__ This is not the norm. We now clarified this in the paper. We use |A| to indicate the cardinality of a set A. Here, we used | | to signify the number of classes seen till that task.
__Q1:__ Yes. Whenever the classification can benefit from having ... | Summary: This manuscript introduces a Bayes classifier method FeCAM to address the feature distribution shift within the realm of Continual Incremental Learning (CIL). This method uses Mahalanobis distance formulation and additionally uses some techniques like correlation normalization, covariance shrinkage, and Tukey’... | Rebuttal 1:
Rebuttal: Thanks for the inputs and the suggestions.
__W1:__ We clarify that in the main text to “50 initial classes and 10 IL steps of 5 classes”.
__W2:__ We train FeTrIL with more epochs, which results in improved accuracy in Table 1 compared to the original paper. We recompute FeTrIL as this is the sta... | Summary: The authors study classifiers in the incremental learning scenario with a fixed strong pre-trained feature extractor. They point out the limitations of the Euclidean distance-based nearest class mean classifier in continual learning and demonstrate the benefits of Mahalanobis distance-based classifier. Several... | Rebuttal 1:
Rebuttal: Thanks for the inputs and suggestions.
__W1, Q1:__ Similar to [2], we perform experiments using the DeiT-S/16 vision transformer pretrained on the ImageNet data with different pre-training data splits and then evaluate the performance of NCM (with euclidean distance) and the proposed FeCAM method... | Rebuttal 1:
Rebuttal: We thank the reviewers for their insightful comments and sincerely appreciate their efforts in providing valuable feedback. The reviewers agree that the paper is very well-written and is easy to follow (MMgd, nDen), has clear and logical theoretical reasoning (YVv6), along with a good motivation w... | NeurIPS_2023_submissions_huggingface | 2,023 | Summary: The paper presents a method for class-incremental learning when a strong base classifier is available and incrementally classes are added to the classifier. The base classifier is kept frozen and another classifier is used on top of the features from this classifier. Instead of using deep networks, authors pro... | Rebuttal 1:
Rebuttal: Thanks for all the feedback and suggestions.
__W1:__ FeCAM is independent of random initialization. Since we do not train any deep neural network after the first step, the method is deterministic and thus we do not see any variation in the results on multiple runs.
__W2:__ As a consequence of ch... | null | null | null | null | null | null |
Neural Lyapunov Control for Discrete-Time Systems | Accept (poster) | Summary: The paper proposes a Lyapunov control method for discrete-time systems, in contrast to previous approaches targeted at continuous-time systems. The paper proposes a mixed-integer linear programming approach for verifying stability conditions in discrete-time systems, a technique for computing sub-level sets wh... | Rebuttal 1:
Rebuttal: We thank the reviewer for the thoughtful and engaging comments. Below we provide our responses.
1. ***Comment:*** *The Heuristic Counterexample Generation cannot guarantee all counterexamples are found. It relies on sampling in the state space, but sampling, the method itself, cannot exhaust the ... | Summary: The stability of nonlinear systems has long been a challenge in the field of control systems, with current methodologies employing Lyapunov stability theory to derive control policies. However, finding suitable Lyapunov functions for these systems is notably complex. To address this, researchers have started u... | Rebuttal 1:
Rebuttal: We thank the reviewer for the comments. Our detailed responses are below.
1. ***Comment:*** *The loss function is then a weighted sum of several terms, which is hard to balance.*
***Response:*** It is common for loss function to be the weighted sum of several terms, as those weights are hyp... | Summary: This paper extends the neural Lyapunov control method to the discrete-time dynamical systems and propose two tricks to quickly find the counterexamples to accelerate the training process. The proposed method is tested in several standard tasks with comparison to the SOTA methods.
Strengths: [1] The proposed m... | Rebuttal 1:
Rebuttal: We thank the reviewer for the thoughtful comments. We respond to these below.
1. ***Comment:*** *The extension of neural Lyapunov control from continuous-time dynamical systems to discrete-time dynamical systems has little contribution to the existing neural control community because the Lyapunov... | Summary: This paper studies learning stabilizing controllers with neural Lyapunov functions in discrete-time nonlinear systems. Previous works on neural Lyapunov control focused on continuous-time systems, and this is the first work that studies discrete-time systems. The outline of the proposed approach is similar wit... | Rebuttal 1:
Rebuttal: We appreciate the reviewer's thoughtful and detailed comments and questions, to which we respond in detail below.
1. ***Comment:*** *It is not clear to me which challenges are specific to the discrete-time system.*
***Response:*** There is a consequential distinction in the definition of Lya... | null | NeurIPS_2023_submissions_huggingface | 2,023 | Summary: This paper presents an algorithmic framework for learning neural Lyapunov control with stablizing control policy for discrete-time dynamic systems. It proposes to verify the stability condition with mixed-integer linear programming and speed up Lyapunov function learning with gradient based approximation. On ... | Rebuttal 1:
Rebuttal: We appreciate the reviewer's thoughtful comments and suggestions. Our detailed responses are below.
1. ***Comment:*** *The paper provides 2 options for setting the bias term in the training loss function. However, it is not crystal clear the impact of this term over the learning. An ablation stu... | null | null | null | null | null | null |
Resetting the Optimizer in Deep RL: An Empirical Study | Accept (poster) | Summary: The authors argue that Adam's internal parameters should be reset with each iteration. The authors demonstrate the effectiveness of this approach in the Atari domain.
Strengths: - Results are convincing for Rainbow.
- Novelty is very low, but potential impact is high, if the result generalizes, there is litt... | Rebuttal 1:
Rebuttal: We appreciate the reviewer's reading our paper as well as the overall quite positive assessment. Thanks for pointing out that the paper's potential impact is high.
- As mentioned by the authors, novelty is low compared to Bengio et al.
Note that Bengio et al. proposed an approach to account for ... | Summary: The paper addresses the issue of using modern optimizers, such as Adam, which maintain internal parameters that are updated over time, potentially contaminating the optimization process. To mitigate this effect, the paper proposes a simple strategy of resetting the internal parameters of the optimizer at the s... | Rebuttal 1:
Rebuttal: We first would like to thank the reviewer for mentioning that our paper is high quality, facilitating a smooth and engaging reading experience. We also appreciate that the reviewer provided a detailed review.
- While I appreciate the proposed method's ease of use, I believe that the authors could... | Summary: This paper questions the standard use of Adam-type optimizers in deep RL. The paper argues that solution methods in deep RL are best thought of as solving a sequence of optimization problems. And that the standard use of optimizers leads to "contamination" of the optimizer's internal parameters. The paper then... | Rebuttal 1:
Rebuttal: We thank the reviewer appreciating our idea. In what follows, we address the particular weaknesses and questions raised.
- The paper does not describe what exactly this contamination means, and neither does it show the presence of any contamination.
Please take a look at the new cosine similarit... | Summary: The paper studies optimization in value-based deep reinforcement learning. The key insight is that when using target networks for action-value function training, changes in the target parameters yield a change in the optimization problem the online parameters are solving. Because of that, the authors argue tha... | Rebuttal 1:
Rebuttal: We thank the reviewer for the time spent carefully reviewing the paper and for appreciating our work. Please find below some clarification regarding your questions.
- "Granted, the median human-normalized performance improves from 1.75 to 2.25, however, per-game effects from resetting the optimiz... | Rebuttal 1:
Rebuttal: We appreciate our reviewers for their thoughtful feedback. Despite their constructive criticism, we believe that all reviewers see positive aspects in our work. In particular, reviewer Ks5C agrees that our approach is motivated and well-described and that we ran our main experiments on all 55 Atar... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Causal Context Connects Counterfactual Fairness to Robust Prediction and Group Fairness | Accept (poster) | Summary: This work aims to provide guidance for contextually-informed algorithmic fairness criteria selection. The work motivates counterfactual fairness from its legal and ethical foundations. The primary contribution of the work is to detail a set of cases (data generating processes and selection mechanisms) where co... | Rebuttal 1:
Rebuttal: Thank you for the close read. We appreciate your time, and we believe that we have addressed your two major concerns by adding an experimental evaluation and clarifying the limitations (and purpose) of working in the purely spurious setting.
> There are no experiments with either simulated or r... | Summary: The authors show that under various causal graphs involving protected attributes, features and labels, that a counterfactually fair classifier achieves a specific association-based fairness metric. From this, they suggest a pipeline to detect counterfactual fairness.
Strengths: The authors do a really good j... | Rebuttal 1:
Rebuttal: Thank you for the support and feedback on framing and content, which we believe has improved the paper.
> The organization of the paper can be improved. For example, a lot of the beginning of Section 4 would make more sense in the introduction. Figure 2 and its associated introductory text would ... | Summary: This work investigated two important coutnterfactual fairness problems: (1) When will the counterfactually fair predictor also be the risk minimizer on the target distribution? (2) How do the classic statistical fairness metrics correspond to counterfactual fairness? In response to these questions, this work c... | Rebuttal 1:
Rebuttal: Thank you for the support, corrections, and provocative questions.
> I did not fully check the proof details but only read the sketch. It looks like the statements heavily depend on the three causal DAGs in Figure 2. For the theoretical results, the assumptions for each statement can be made more... | Summary: The paper explores a connection between counterfactual fairness and correlation-based fairness metrics, exhibiting causal graph settings in which counterfactual fairness implies specific correlational metrics are satisfied. The paper further connects this mapping to the accuracy-fairness tradeoff.
Strengths: ... | Rebuttal 1:
Rebuttal: Thank you for your support and your useful comments.
> I think the paper itself foresees my largest concern: it's not clear what settings these results apply in. … This exists, but way down in Section 5. Help your reader by making these connections right away.
We agree that the mapping of causal... | Rebuttal 1:
Rebuttal: We thank the reviewers for their time and thorough commentary. We are glad to hear the core technical results were seen as valuable contributions (“interesting way to connect…”, “technically sound”, “useful”, “novel”, “the theoretical results are of high quality”, “expand upon the prior work meani... | NeurIPS_2023_submissions_huggingface | 2,023 | Summary: This paper provides a new way to analyze the fairness-accuracy tradeoff in many fairness definition through the lens of counterfactual fairness. The authors prove that, under certain conditions, the counterfactually fair predictor is optimal when the target is OOD generalization. The theoretical analyses build... | Rebuttal 1:
Rebuttal: Thank you for your time and useful suggestions, which significantly improve the manuscript.
> W1. Some statements/assumptions need more comprehensive illustration.
We have added the more comprehensive illustrations in response to your comments below.
> W3/Q3. The authors use impossibility theor... | null | null | null | null | null | null |
You Only Condense Once: Two Rules for Pruning Condensed Datasets | Accept (poster) | Summary: This paper proposes a combination method for dataset pruning and dataset condensation. Both methods could reduce the size of training datasets. Experiments show that the proposed method achieves state-of-the-art performance on several datasets regardless of networks. The proposed rules are simple and effective... | Rebuttal 1:
Rebuttal: Thanks a lot for your comment. We will answer the questions below.
>1. In Table 1, the ImageNet-10 on the DREAM method is not listed.
DREAM uses tiny-ImageNet (resolution: 64x64) instead of a subset of ImageNet (resolution: 224 x 224). This is because DREAM performs clustering during condensatio... | Summary: This paper target at the dataset condensation for on-device scenarios.
It proposes two novel rules to prune the condensed dataset. The first rule is to rank the condensed images. The rule believes easy examples are better for pruning condensed dataset. The second rule is to choose balanced samples for differen... | Rebuttal 1:
Rebuttal: Thank you for your high score, and we'd like to answer your question below.
> 1. The LBPE score and balanced construction. Which rule is more important?
Thanks for the question. Table 2 in the paper provides an ablation study on the importance of LBPE score and Balanced Construction. These resul... | Summary: This paper employs a dataset pruning approach on a condensed dataset. Specifically, the intent is to evaluate and rank the data within this condensed dataset based on their respective significance. This enables users to selectively choose any number of condensed datasets, offering both efficacy and flexibility... | Rebuttal 1:
Rebuttal: Thank you for your comments regarding the evaluation section of our paper. We will explain your questions point by point.
> 1. The main weakness is the evaluation part. This paper compares their method: (1). dataset condensation + dataset pruning with the previous method (2). dataset condensation... | Summary: --- Post-Rebuttal Edit ---
After author rebuttals, I have updated my rating from 5 to 6 and confidence from 2 to 3.
--- End Post-Rebuttal Edit ---
This paper considers a practical problem in dataset condensation/distillation, where edge devices have varying constraints that need to be flexibly met by a data... | Rebuttal 1:
Rebuttal: Thank you very much for your insightful comments and for raising concerns about the clarity of the motivation and the terminology used in our paper.
> 1. What is an “on-device scenario?”
We acknowledge that the term "on-device scenarios" may not be clear to all readers. What we mean by this ... | null | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Phase diagram of early training dynamics in deep neural networks: effect of the learning rate, depth, and width | Accept (poster) | Summary: The paper studies the effect of depth, width and learning rate on the early training dynamics of DNNs. The authors first describe four phases throughout, and then focus on the first two phases.
In the first phase, they identify three ranges of learning rates, according to how they affect the evolution of the ... | Rebuttal 1:
Rebuttal: We thank the reviewer for their encouraging comments. Below are our responses to the comments:
> The authors define a large number of regimes acoording to some rather subtle changes in behavior sometimes, and it is not always clear whether these regimes are also correlated with other quantities t... | Summary: The paper identifies new phenomena involving the dynamics of the sharpness (largest eigenvalue of the Hessian) in the early stages of neural net training, for ReLU nets of various architectures that are initialized using He initialization. Previously, it was known that for very wide networks, there are two '... | Rebuttal 1:
Rebuttal: We thank the reviewer for their encouraging comments. Here is our response to the question on loss connectivity:
> Have you looked at loss connectivity between the final parameters (at the end of training) and the initialization?
As the main focus of this work is the early training dynamics, we... | Summary: In this paper, the authors studied the effects of learning rate, width and depth of neural networks for early training dynamics. Specifically, the authors focused on the value of loss and the maximum eigenvalue of Hessian of loss and found that there are 4 different possible training dynamics in the early trai... | Rebuttal 1:
Rebuttal: We thank the reviewer for their comments and careful reading. Below are our responses to the questions and comments:
> The current paper seems to have lots of results and experiments in the main text. As a reader, it is not very easy for me to get the main conclusion for each section/set of exper... | Summary: This work mainly investigates the learning dynamics depending on the learning rate and categorizes its qualitative behaviors in some phases. In particular, the authors identify new phases; sharpness reduction, and loss-sharpness catapult. They investigate the dependence of the critical learning rates (mainly, ... | Rebuttal 1:
Rebuttal: We thank the reviewer for taking the time to read our manuscript and for providing detailed comments. Below are our responses to the comments provided:
> Connection to the infinite width limit:
In our study, we are looking at the behavior of the optimization dynamics at training timescales $t_\s... | Rebuttal 1:
Rebuttal: We would like to express our gratitude to all the reviewers for taking the time to review our paper and providing detailed comments. Attached below is the PDF containing the figures required for some responses.
Pdf: /pdf/6454cba2e59bb7937f38dde0a452c80e73c4e977.pdf | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Towards Anytime Classification in Early-Exit Architectures by Enforcing Conditional Monotonicity | Accept (poster) | Summary: This paper studies an interesting problem of early-exit classification networks when applied to anytime prediction tasks. Specifically, the authors empirically demonstrate that in an early-exit network, the prediction confidence scores of its different exits for the same sample are not monotonically increased.... | Rebuttal 1:
Rebuttal: We appreciate your time and efforts in reviewing our work.
> the authors empirically demonstrate that in an early-exit network, prediction confidence scores [...] for the same sample are not monotonically increased.
We would point out that our findings indicate the ground-truth class probability... | Summary: This submission focuses on multi-exit early networks, from the perspective of anytime inference that can accommodate varying computational budgets. Such models produce a progressive refinement of the final prediction, offering the opportunity to "exit-early", providing a meaningful prediction. The main motivat... | Rebuttal 1:
Rebuttal: We would like to thank you for your valuable review and many interesting questions.
> how the proposed methodology would behave in frozen backbone early-exit models
> Is the proposed methodology still relevant if applied on early-exit models trained via the frozen-backbone methodology
**Yes, ou... | Summary: The authors propose a new approach, called Product Anytime (PA), for monotonic confidence estimation in early-exit neural networks. Inspired by the Product-of-Experts (PoE) approach, PA takes the normalized confidence with respect to the product of ReLU thresholded prediction logits. The authors show through e... | Rebuttal 1:
Rebuttal: We thank you for your time in reviewing our work. We address your concerns below.
> The authors propose [PA] for monotonic confidence estimation
We would just point out that our model primarily targets monotonicity in the ground-truth class probability, denoted as $p_m(y^* | \boldsymbol{x})$, an... | Summary: The paper first reviews the definition & key properties of anytime models (Zilberstein, 1996) and argues to focus on equipping the EENNs with conditional monotonicity (in addition to their built-in interruptibility, Sec 2). The paper then verifies that existing EENNs lack this critical property (Fig 1), propos... | Rebuttal 1:
Rebuttal: We thank you for your feedback and help in improving our work.
> Given the criticality of the relaxation/approximation of the Heaviside function, it’s highly desirable to see more thorough analyses of and comparisons between different approximations, e.g. sigmoid …
We fully agree with your obse... | Rebuttal 1:
Rebuttal: We thank all reviewers for their valuable time and feedback; it’s much appreciated.
We are encouraged that the reviewers found “the paper well written and easy to follow” (G2Sm), thought that “the studied problem is interesting and important” (NWJy), “often overlooked by the relevant literature” ... | NeurIPS_2023_submissions_huggingface | 2,023 | Summary: This paper proposes a post-hoc modification of early-exit neural networks. Specifically, the authors focus on the conditional monotonicity out of four properties presented in a prior work. A training-based method is also proposed to improve the model confidence.
Strengths: The proposed method based on product... | Rebuttal 1:
Rebuttal: Thank you for your feedback and help in improving our paper.
> Some writing is somewhat verbose/redundant. For example, while properties of anytime models take a considerable amount of space, they are not that meaningful at last.
While we focus exclusively on the monotonicity property, we have ... | null | null | null | null | null | null |
Strategic Data Sharing between Competitors | Accept (poster) | Summary: This paper uses algorithmic game theory to study potential data sharing between competing firms that train machine learning models with similar goals. The authors propose a new framework that can be used to analyze trade-offs faced by data-using agents who make decisions about whether to collaborate (via data ... | Rebuttal 1:
Rebuttal: Thank you for your valuable and constructive feedback. In the following we address the weaknesses and questions raised in your review.
**Weaknesses**
**In my view, the main threat to the validity and impact of the work is the reliance on assumptions from conventional economics … The current draf... | Summary: The authors present a trainable model for optimising the benefits of data sharing based valuation of data, benefits of the shared data sets for the inference model, in a conventional market model. The authors point out that the model can be used to find an optimal data sharing strategy.
The model is nicely d... | Rebuttal 1:
Rebuttal: Thank you for your valuable and constructive feedback. In the following we address the weaknesses and questions raised in your review.
**Weaknesses**
**The assumption that the collaborative partners share the same data distribution and the data is modeled as iid samples of a common distribution.... | Summary: This paper explores the dilemma faced by firms when considering strategic data sharing with their competitors. It introduces a framework to investigate the incentives for data sharing and examines its impact on collaboration and profitability. The author discusses the barriers to data sharing, such as privacy ... | Rebuttal 1:
Rebuttal: Thank you for your valuable and constructive feedback. In the following we address the weaknesses and questions raised in your review.
**Weaknesses**
**It would be great to see a discussion of the various real-world examples in the context**
Thank you for your valuable suggestion. We will add m... | Summary: * This paper analyzes the economic consequences of data sharing between competitors.
* Interaction is modeled as a market:
* Market has $m$ firms ($F_1,\dots,F_m$), where firm $i$ produces $q_i$ units of good $G_i$ at price $p_i\ge 0$, and quality $v_i$.
* In addition, there are “outside goods” $\{1,\dots,... | Rebuttal 1:
Rebuttal: Thank you for your valuable and constructive feedback. In the following we address the weaknesses and questions raised in your review.
**Weaknesses**
**Parametric assumptions for main theorems are not validated against real-world data. Not clear which parameter regimes are likely in practice.**
... | Rebuttal 1:
Rebuttal: We thank all reviewers for their valuable and constructive feedback. Below we provide a general response to several recurring comments. We also offer individual answers for all remaining questions.
**Real-world evaluation of the market and data impact models.**
Multiple reviewers brought up the ... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
A Deep Instance Generative Framework for MILP Solvers Under Limited Data Availability | Accept (spotlight) | Summary: The paper discusses the recent surge in using machine learning techniques for solving combinatorial optimization problems, particularly mixed-integer linear programs (MILPs). However, the limited availability of real-world instances often hinders optimal decision-making and unbiased solver assessments. To addr... | Rebuttal 1:
Rebuttal: We thank the reviewer for the insightful and valuable comments. We respond to your comments as follows and sincerely hope that our rebuttal could properly address your concerns. If so, we would deeply appreciate it if you could raise your score and your confidence. If not, please let us know your ... | Summary: The authors propose a deep generative framework for MILP(mixed integer programs) instances. The work has applications in enhancing MILP solvers under limited data availability. MILP instances are represented as weighted bipartite graphs. The authors propose a masked variational autoencoder (VAE) method for ... | Rebuttal 1:
Rebuttal: We thank the reviewer for the insightful and valuable comments. We respond to your comments as follows and sincerely hope that our rebuttal could properly address your concerns. If so, we would deeply appreciate it if you could raise your score and your confidence. If not, please let us know your ... | Summary: The paper generates MIP instances with the help of VAEs. Experiment results show that the generated samples remain the characters of the training dataset from several aspects (e.g. the graph statistical value, the solving time of the generated samples, and the number of branching nodes of the generated samples... | Rebuttal 1:
Rebuttal: We thank the reviewer for the insightful and valuable comments. We respond to your comments as follows and sincerely hope that our rebuttal could properly address your concerns. If so, we would deeply appreciate it if you could raise your score and your confidence. If not, please let us know your ... | Summary: The paper addresses the challenge of limited availability of real-world mixed-integer linear programs (MILPs) instances for machine learning techniques employed in combinatorial optimization problems. The scarcity of these instances often leads to sub-optimal decisions and biased solver assessments. Existing s... | Rebuttal 1:
Rebuttal: We thank the reviewer for the positive and valuable comments. We respond to your comments as follows and sincerely hope that our rebuttal could properly address your concerns. If so, we would deeply appreciate it if you could raise your score. If not, please let us know your further concerns, and ... | Rebuttal 1:
Rebuttal: Dear reviewers,
We sincerely thank all reviewers' insightful and constructive comments, which helped to significantly improve our work. We respond to the comments given by each reviewer in detail, and in this global response, we answer several key issues raised by multiple reviewers.
**1. Contr... | NeurIPS_2023_submissions_huggingface | 2,023 | Summary: This paper is focused on using machine learning techniques to address combinatorial optimization problems, particularly mixed-integer linear programs (MILPs). Particularly, this paper proposed G2MILP, which is a generative model for generating realistic MILP instances. Without prior expert knowledge, G2MILP ca... | Rebuttal 1:
Rebuttal: We thank the reviewer for the insightful and constructive comments. We respond to your comments as follows and sincerely hope that our rebuttal could properly address your concerns. If so, we would deeply appreciate it if you could raise your score. If not, please let us know your further concerns... | Summary: The paper proposed a deep generative framework G2MILP based on a masked variational autoencoder (VAE) for mixed-integer linear programs (MILP) instances. The proposed framework G2MILP corrupts and replaces constraints to generate new MILP instances while keeping the variables unchanged. Experiments show that G... | Rebuttal 1:
Rebuttal: We thank the reviewer for the insightful and valuable comments. We respond to your comments as follows and sincerely hope that our rebuttal could properly address your concerns. If so, we would deeply appreciate it if you could raise your score. If not, please let us know your further concerns, an... | null | null | null | null |
Rethinking SO(3)-equivariance with Bilinear Tensor Networks | Reject | Summary: The paper proposes n permutation and SO(3) equivariant network for b-tagging in high energy physics. Permutation equivariance is required since the input data is a _set_ of track particle features, while SO(3) equivariance is required since these features are geometrically scalars or vectors whose rotational t... | Rebuttal 1:
Rebuttal: We thank the reviewer for a thoughtful response to our paper. We address the reviewer's specific questions below:
* "How can the SO(2)-equivariant layers be applied..."
They can be used in exactly the same way. Because each parallel result of the subnetwork $\Phi$ is a proper vector quantity, s... | Summary: The paper presents a method to handle complex geometric structured data, specifically SO(3) representations, through SO(3)-equivariance and judicious symmetry breaking. The technique improves computational efficiency and enhances the performance of a network operating on these representations. When applied to ... | Rebuttal 1:
Title: late rebuttal
Comment: Apologies for the late response. We thank the reviewer for a thoughtful response to our paper. We address points below:
* Some explanation in the paper...
The b-tagging problem is an example of a class of symmetry-related problems, which are of interest to a much wider audien... | Summary: The authors propose a SO(3)-equivariant network operating on scalars, vectors and 2-tensors. They identify the corresponding equivariant linear layers and come up with a mixing strategy to mix different representations. Furthermore, SO(2)-equivariant linear layers are proposed to allow the scenario that SO(3) ... | Rebuttal 1:
Rebuttal: We thank the reviewer for a thoughtful response to our paper. We address points below:
* "The idea of tensor-product-based neurons is not new..."
We thank the reviewer for brining this work (and others) to our attention, and would be add a brief discussion comparing and contrasting our method. ... | Summary: This work presents a lightweight architecture based on scalars, vectors, and tensors for learning in 3D, subject to $SO(2)$-equivariance in a known direction that varies by sample. They are motivated by the jet-tagging problem in high energy physics, in which a given batch is equivariant with respect to a cert... | Rebuttal 1:
Rebuttal: We thank the reviewer for a thoughtful response to our paper. We address specific points below:
* The baselines and experiments are not sufficiently developed...
This is a misunderstanding; the baseline model _does_ receive as input the standard projections z_0 and d_0, cf. L234, in addition to ... | null | NeurIPS_2023_submissions_huggingface | 2,023 | Summary: Before starting, I must mention that I am not a physicist, and therefore, I have focused on the machine learning aspects of the paper.
### Summary
This paper proposes the use of SO(3) equivariant neural networks for B-tagging. The method proposed builds upon Deep Sets and provides a quite constrained formula... | Rebuttal 1:
Rebuttal: We thank the reviewer for a thoughtful response to our paper. We address specific points below:
* "My main concern is that..."
We agree that this point is of central importance to the entire work.
The short answer is that the network is, in fact, globally equivariant w.r.t. SO(3).
The reason for... | null | null | null | null | null | null |
Optimal Excess Risk Bounds for Empirical Risk Minimization on $p$-Norm Linear Regression | Accept (poster) | Summary: The paper studies the excess risk of empirical risk minimization on the $p$-norm linear regression. The asymptotic bound of the excess risk of ERM for $p$ in ($1,\infty$) is known by prior work and more recent work had given high probability excess risk bounds for $p=2$ that match that asymptotic rate. This pa... | Rebuttal 1:
Rebuttal: We thank the reviewer for their positive evaluation and constructive feedback. We will make sure to add a definition of $\rho$ and $\sigma_p^2$ inside the statements of the theorems to make them more readable.
---
***”What do the three regimes represent in Theorem 3?”***
They are an artifact of... | Summary: This paper studies the performance of empirical risk minimization (ERM) for $\ell_p$-norm linear regression, in both the realizable and non-realizable (say, noisy measurement) settings. While the $p=2$ case has been studied extensively in the literature, not as much is known for the general $p \in (1,2) \cup (... | Rebuttal 1:
Rebuttal: We thank the reviewer for their valuable feedback. Please find our detailed response below.
---
***”My main concern/confusion is the message of the paper.”***
The goal and achievement of the paper is the study of the performance of ERM non-asymptotically and under weak moment assumptions for th... | Summary: This paper studies the empirical risk minimization of linear regression with the p-norm loss function, where p ∈ (1,∞).
Main contributions:
- The authors provide a high probability excess risk bound on the empirical risk minimizer, for p ∈ (1,∞), where for p ∈ [2,∞) only weak moment assumptions are needed, an... | Rebuttal 1:
Rebuttal: We thank the reviewer for their valuable feedback. We kindly ask the reviewer to re-evaluate the paper in light of our response below.
---
***”The proof part is poorly written.”***
We thank the reviewer for pointing us to the typos in Corollary 1 and Proposition 1; we fixed these in the supplem... | Summary: The paper considers linear regression with inputs in dimension d and with n data points. It consider the Lp loss |.|^p with a focus on p different from the classical value 2. The paper provides non-asymptotic risk excess bound for this Lp loss, the excess being between the optimal linear combination w^* and th... | Rebuttal 1:
Rebuttal: We thank the reviewer for their positive evaluation. We will be happy to answer any questions or comments that may arise during the discussion period. | null | NeurIPS_2023_submissions_huggingface | 2,023 | Summary: The paper studies excess risk bounds for empirical risk minimization for linear regression under $\ell_p$-norm. It extends existing work $p=2$ to general $p\in (1,\infty)$.
Sampling complexity bounds are given for recovering the true parameter in the realizable case.
Additive error excess risk bounds are giv... | Rebuttal 1:
Rebuttal: We thank the reviewer for their positive evaluation and detailed feedback. We address the reviewer’s concerns below.
---
***”Optimality is a strong claim. In what sense can your bounds be considered optimal?”***
Our bounds are optimal in the sense that the leading term in them matches the asymp... | null | null | null | null | null | null |
Multi-task Representation Learning for Pure Exploration in Bilinear Bandits | Accept (poster) | Summary: This paper studies multi-task representation learning for the problem of pure exploration in bilinear bandits. The authors design algorithm GOBLIN that uses an experimental design approach to optimize sample allocations for learning the global representation as well as minimize the number of samples needed to ... | Rebuttal 1:
Rebuttal: Thank you for reviewing our paper. Below, we clarify the key novelty in our approach and compare our work with [Du et al. 2023].
*Weakness*
1) Even though we use existing ideas in SVD, G-optimal design phased elimination in linear bandit setting, it is highly non-trivial to extend that analysis... | Summary: The authors study the problem of pure exploration in bilinear bandits, where the reward is a bilinear function of the feature vectors of two types of arms. They consider both single-task and multi-task settings, where the latter involves learning a shared low-dimensional representation across multiple tasks. T... | Rebuttal 1:
Rebuttal: We thank the reviewer for appreciating the soundness and contribution of our work. We answer the questions raised below.
*Weakness*
1) Thanks for pointing out these papers. We will discuss the papers pointed out in the revision. Here we provide a brief comparison of our work with [1][2]. [1] stu... | Summary: This paper studies pure exploration of bilinear bandits with multi-task representation learning, i.e., different tasks share a common low-dimensional linear representation.
As an intermediate step, the authors propose the first single-task bilinear bandit pure exploration algorithm, based on G-optimal design. ... | Rebuttal 1:
Rebuttal: We thank the reviewer for appreciating the soundness and contribution of our work. We answer the questions raised below.
*Weakness*:
1) Thank you for raising this point. There is no lower bound for the single task setting for pure exploration in bi-linear bandits with unknown underlying structur... | null | null | Rebuttal 1:
Rebuttal: References:
- KS Jun, R Willett, S Wright, R Nowak, Bilinear bandits with low-rank structure, ICML 2019
- T Fiez, L Jain, K G. Jamieson, L Ratliff, Sequential Experimental Design for Transductive Linear Bandits, NeurIPS 2019
- Y Du, L Huang, W Sun, Multi-task Representation Learning for Pure Expl... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
To Repeat or Not To Repeat: Insights from Scaling LLM under Token-Crisis | Accept (poster) | Summary: This paper conducts an empirical study on the scaling of transformer models under limited training data, which they call a *token crisis*. They show that under token crisis, training T5 models for multiple epochs results in the degradation of pre-training and downstream task performance. They also show that th... | Rebuttal 1:
Rebuttal: > Q1: This paper only uses a single task as the downstream task, so it is unclear if the result will hold for other datasets.
Thank you for your insightful suggestion. We conducted more downstream evaluation on BoolQ and RTE datasets, which are two widely used datasets in SuperGLUE benchmark. As ... | Summary: The paper presents an empirical study on the effect of training on multiple epochs in the data limited regime. They show that Chinchilla's scaling laws holds for T5 style models. The authors show that repeated tokens result in degradation of accuracy. The also study some of the factors contributing to this deg... | Rebuttal 1:
Rebuttal: > Q1: Some of the conclusions are trivial. For example, larger models are more susceptible to overfitting.
Thanks for the suggestion. We argue that, although some of our conclusions are intuitively trivial, these conclusions play an important role in studying the token-crisis problem. For instanc... | Summary: The authors propose a concept called token-crisis, which means the growth rate of high-quality text data available is much slower than the growth rate of data required by LLMs. This paper is the first empirical study of the repeating pre-training data for the token-crisis problem. Some major findings are: larg... | Rebuttal 1:
Rebuttal: > Q1: Why there are no experiments testing for a smaller number of data repetitions? The smallest number of repetitions in the paper is 2^8, which is very large.
Thank you for your question. In our plots, we actually have the experiments testing for a small number of data repetitions. For instanc... | Summary: This paper delves into the token-crisis issue in language models, a situation characterized by performance decline when the same pre-training data is used across multiple epochs. The authors scrutinize several methods of training language models with recurring tokens, encompassing regularization strategies and... | Rebuttal 1:
Rebuttal: > Q1: All the experiments were conducted on the T5-style masked language modeling, which is opposite to the recently surging GPT-style (i.e., causal) language modeling. And it's unclear whether all the sights/findings are applicable or transferable to the CLMs.
Thank you for your insights. We tot... | Rebuttal 1:
Rebuttal: Dear reviewers and AC:
We thank all the reviewers for their feedback. Two main concerns and concise arguments are summarized below. The full version can be found in the corresponding response of different reviewers.
* Use Decoder-only Model instead of Encoder-Decoder Model. => This project star... | NeurIPS_2023_submissions_huggingface | 2,023 | Summary: Data is one of the most important factors for training a well-performed Large Language Model(LLM). There are not enough studies that are deep into the effect of data. This paper addresses this key problem and analyzes the effect from the important aspects including the effect of pre-training, the effect of dow... | Rebuttal 1:
Rebuttal: > Q1: Use GPT instead of T5 as backbone.
Thank you so much for your valuable insights. As we discussed in Appendix A.1, the differences between the encoder-decoder and decoder-only models might not be as big as our community thought. There are two main distinctions to consider: (1) The encoder-d... | null | null | null | null | null | null |
No-Regret Online Prediction with Strategic Experts | Accept (poster) | Summary: This submission studies the problem of online decision-making using predictions given by experts with incentives to behave strategically. In particular, at each time-step, K experts hold a belief about a binary outcome. Each expert reports a prediction to a learner, who picks a set of m experts. The learner th... | Rebuttal 1:
Rebuttal: First, we would like to thank the reviewer for their constructive feedback and comments. We hope our responses below may help and perhaps convince the reviewer to raise their score. We are happy to answer any further questions during the author-reviewer discussion period.
$\bullet$ **Q:** The lac... | Summary: This paper generalizes the problem of online binary prediction with expert advice, where at each round the learner can pick m>1 experts and the overall utility is a modular or submodular utility of the chosen experts. The experts are strategic and wish to be selected by the algorithm as often as possible hence... | Rebuttal 1:
Rebuttal: First, we would like to thank the reviewer for their constructive feedback and comments. We hope our responses below may help and perhaps convince the reviewer to raise their score. We are happy to answer any further questions during the author-reviewer discussion period. Also, we have provided ad... | Summary: - The paper studies the problem of online prediction with expert advice. In their setting, a learner at each round selects $m$ experts, and the loss is determined by a submodular or modular function of the beliefs reported by the selected experts. Each expert, being strategic, may intentionally misrepresent th... | Rebuttal 1:
Rebuttal: First, we would like to thank the reviewer for their constructive feedback and comments. We hope our responses below may help and perhaps convince the reviewer to raise their score. We are happy to answer any further questions during the author-reviewer discussion period. Also, given the size limi... | Summary: The paper broadly focuses on the design and analysis of no-regret and incentive-compatible algorithms for the m-experts problem (pick a subset $S_t \subseteq [K]$ of $m$ experts in each round $t$ to obtain a utility value $f_t(S_t)$ which is a function of the losses $\ell_{i,t} \in [0,1]$ for all $i \in S_t$),... | Rebuttal 1:
Rebuttal: First, we would like to thank the reviewer for their constructive feedback and comments. We hope our responses below may help and perhaps convince the reviewer to raise their score. We are happy to answer any further questions during the author-reviewer discussion period.
$\bullet$ **Q:** The FTP... | Rebuttal 1:
Rebuttal: In the attached file, we have included two additional figures plotting the running average of regret $\frac{1}{t}\mathbb{E} \big[\max_{S\subseteq [K]:|S|=m}\sum_{\tau=1}^t f_{\tau}(S)-\sum_{\tau=1}^t f_{\tau}(S_{\tau})\big]$ for $1\leq t\leq T$ for our proposed algorithms and also the FiveThirtyEi... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
A Unified Solution for Privacy and Communication Efficiency in Vertical Federated Learning | Accept (poster) | Summary: This research proposes a novel cascaded hybrid optimization approach named VFL-CFOZO for VFL that combines zeroth-order (ZO) gradient and first-order (FO) gradient optimization techniques. The critical output layer of the clients uses the ZO gradient, while other parts utilize the FO gradient. This approach en... | Rebuttal 1:
Rebuttal: >Weakness 1 (Experiment with more clients needed):
We conduct experiments with 4 and 8 clients to further assess the performance of our framework on a larger scale. The results are presented in the Figure 1&2, Table 1&2 in the **attached PDF (in the topmost rebuttal box)**. It is essential to no... | Summary: The paper proposes a solution for privacy and communication efficient vertical FL training framework by utilizing zero-order-optimisation (ZOO). Since the convergence of ZOO-based VFL is significantly slower than standard gradient-based VFL, the paper proposes to only implement ZOO on the cut layer.
Strength... | Rebuttal 1:
Rebuttal: >W1&Q1 (The choice of compressor):
Yes, applying different compressors may lead to different communication costs. We will add an experiment on the Top-K compressor in the appendix to enrich the discussion. However, comparing the performance of different compressors is **not the focus** of our pa... | Summary: The paper is aiming at solving two critical problems of Vertical Federated Learning (VFL), the convergence rate of ZOO-based VFL and the privacy guarantee of ZOO-based VFL. This study provides a simple solution of using different optimizations i.e., the first-order optimization (FOO) and zeroth-order optimizat... | Rebuttal 1:
Rebuttal: >Weakness 1 & Question1 (The contribution of ZOO and compression in reducing communication cost):
Yes, both methods contribute to enhancing communication efficiency. However, typically ZOO makes a more significant contribution to reducing the communication cost.
Take the experiment presented i... | Summary: This paper presents a pioneering Zero-Order Optimization (ZOO)-based VFL algorithm that effectively ensures privacy preservation while significantly enhancing communication efficiency. Regarding privacy, the paper demonstrates theoretically that ZOO can inherently offer $(\epsilon,\delta)$-differential privacy... | Rebuttal 1:
Rebuttal: >Weakness 1& Question 1& Limitation 2 (Experiment on more clients needed):
We conduct experiments with 4 and 8 clients to further assess the performance of our framework on a larger scale. The results are presented in the Figure 1&2, Table 1&2 in the **attached PDF (in the topmost rebuttal box)*... | Rebuttal 1:
Rebuttal: **We would like to thank the reviewers for dedicating their time and effort to assess the manuscript.**
**Attached is the PDF for the figures and tables.**
Pdf: /pdf/478498f259b2a4437fce9c9e7d80e0426fe01cf1.pdf | NeurIPS_2023_submissions_huggingface | 2,023 | Summary: The paper introduces a hybrid Federated Learning (FL) framework, named VFL-CZOFO, which aims to provide intrinsic privacy protection while also significantly improving the convergence rate when compared to existing ZOO-based frameworks.
Strengths: 1. Faster Convergence: The paper demonstrates that VFL-CZOFO a... | Rebuttal 1:
Rebuttal: >Weakness 1 (Experiment on more clients needed):
We conduct experiments with 4 and 8 clients to further assess the performance of our framework on a larger scale. The results are presented in Figure 1&2, Table 1&2 in the **attached PDF (in the topmost rebuttal box)**. It is essential to note tha... | null | null | null | null | null | null |
Two Heads are Better Than One: A Simple Exploration Framework for Efficient Multi-Agent Reinforcement Learning | Accept (poster) | Summary: This paper propose a novel and compute-efficient exploration method: COIN, which can incorporate Curiosity-based and Influence-based exploration. For influence-based exploration，COIN quantify how each agent’s action affect the other agents and use the influence degree as the intrinsic reward to promote the exp... | Rebuttal 1:
Rebuttal: **Q1:** It lacks a certain degree of novelty, and just do some combination of previous work such as MI and prediction error.
**A1:** Off course, MI and prediction error are commonly used in RL.
**In this paper, our main motivation is to highlight that different kinds of exploration play a differe... | Summary: This paper investigates exploration methods in multi-agent reinforcement learning (MARL). Specifically, among the two main families of exploration methods, namely curiosity-based methods and influence-based methods, the paper reports that they could serve complementary roles to each other. Subsequently, the pa... | Rebuttal 1:
Rebuttal: **Q1:** The results of QMIX + ours in Aloha seem to be quite odd given the early burst and drop with the plateaued bad results. Any intuition or explanation for the phenomenon?
**A1:** In Aloha, messages collided when two adjacent agents send messages simultaneously.
Each successful transmission ... | Summary: In this paper the authors propose a new framework to improve exploration in MARL. The method COIN improves exploration by combining the concepts in the literature of curiosity and influence.
Strengths: This paper proposes a method that combines in a framework two popular concepts in the reinforcement learning... | Rebuttal 1:
Rebuttal: **Q1:** ''Our influence-based intrinsic reward promotes the agents to affect each other and hence brings diversity." needs further evidence.
**A1:** The parameter-sharing paradigm leads to the same behavior of different agents in Google Research Football(GRF). After watching the replay, we find t... | Summary: This paper looks at two different exploration strategies within Multiagent Reinforcement Learning. The two exploration strategies are curiosity-based and influenced based, where the latter is unique to MARL and takes into account the mutual information of one agents actions on the trajectory of another agent. ... | Rebuttal 1:
Rebuttal: **Q1:** The lack of clear exploration in the hyperparameter space. The continuum of mixings between the rewards would be most interesting. Quantitative set of measures of the hyperparameter sweeps.
**A1:** The importance of two kinds of exploration is various in different time steps. Thus, it is ... | Rebuttal 1:
Rebuttal: Thanks for your careful reviews and discerning comments. We will explain all the mentioned questions in the following, and some new experiments are added to the affiliated file.
Pdf: /pdf/97fd42a2492bd1fc4bc520b4ee8fd55320a39c43.pdf | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Every Parameter Matters: Ensuring the Convergence of Federated Learning with Dynamic Heterogeneous Models Reduction | Accept (poster) | Summary: This work establishes a general convergence analysis for a heterogeneous federated learning (FL) algorithm that trains a shared global model using a sequence of time-varying and client-dependent local models. In particular, this work establishes sufficient conditions for the convergence of such a heterogeneous... | Rebuttal 1:
Rebuttal: We thank the reviewer for their constructive and insightful comments and suggestions. Following are our responses to your questions and concerns.
> Q1. The last term on the right-hand side of the inequality in Theorem 1 (and similarly for Theorem 2) corresponds to the average of the norms of the ... | Summary: This paper studies federated learning with heterogeneous client models and non-iid client data. By assuming that the client models are pruned versions of a common global model and using the notion of minimum covering index, the paper provides the convergence of FedAvg under client model pruning. Theoretical re... | Rebuttal 1:
Rebuttal: Thank you for the constructive comments. We will follow these helpful comments in our revised version. Following are our responses to your questions and concerns.
> Q1. Term $E\Vert \theta_q\Vert^2$ in the bound. This term appears in all theorems and lemmas in the main paper. However, it is uncl... | Summary: This paper focuses on the cross-device federated learning setting. The authors introduce a general theoretical framework for analyzing FedAvg with masks on local pruned models. This analysis is particularly valuable in establishing the convergence of federated schemes when the global model is distributed acros... | Rebuttal 1:
Rebuttal: Thank you for your comments and for confirming the contributions of our paper. We provide clarification to your questions and concerns as below.
> Q1. It is unclear whether $ \Gamma_q^{(i)}$ is defined or if it refers to $N_{q}^{(i)}$ mentioned on line 177. Similarly, does $\Gamma^*$ denote th... | Summary: In this work, the authors provide a general theoretical framework to analyze the convergence of Federated training schemes conducted over local models of heterogeneous network structures. Such structures can usually be obtained through different model reduction methods, such as model pruning/sparsification or ... | Rebuttal 1:
Rebuttal: Thanks for your review!
> Q1. From my understanding, your framework seems to consider reduced model structures when these structures are obtained when the global model is reduced (pruned). Will your analysis hold even in the case of local model reduction ... How would that affect your analysis if... | Rebuttal 1:
Rebuttal: Dear reviewers, thank you all again for your valuable time and positive suggestions that will definitely make our paper stronger. We have provided each reviewer with our responses in a Q&A format. We are happy to answer further questions if any. | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Enhancing User Intent Capture in Session-Based Recommendation with Attribute Patterns | Accept (poster) | Summary: The paper presents a transformer-based method, namely Frequent Attribute Pattern Augmented Transformer (FAPAT), that considers attribute patterns as supplementary information. More specifically, FAPAT builds attribute transition graphs, mine frequent attribute patterns, and match attribute patterns to better c... | Rebuttal 1:
Rebuttal: **W1. Clarifications for pattern acquisition**
A1. End-to-end retrieval becomes a viable option in the absence of resource constraints. However, given the multitude of substructures—numbering in the millions or even billions—within industrial sessions, we must ponder efficiency. Data mining metho... | Summary: This paper proposes FAPAT, which augments session-based recommendation models using item attributes. Specifically, FAPAT constructs attribute graphs and models user intent using pattern mining and an improved graph attention mechanism.
Strengths: 1. The paper studies an important application task, i.e., sessi... | Rebuttal 1:
Rebuttal: **W1. Over-complicated method**
A1. Differing Perspective on GRU4Rec. I must respectfully disagree with the reviewer's perspective. The counter-example of GRU4Rec showcases remarkable performance on diginetica (mean length: 4.850). Nevertheless, its effectiveness significantly wanes on Tmall (mea... | Summary: This paper presents a novel framework for session-based recommendations. The code idea of the method is to extract highly frequent attribute patterns from graphs to augment session sequence encoding. Specifically, it first leverages frequent graph pattern mining for attribute pattern retrieval. Then it applies... | Rebuttal 1:
Rebuttal: **W1. Clarifications for pattern acquisition and baselines**
A1. Upon the motifs containing three or four nodes, we exclusively consider those featuring a circle or a triangle. These two structural types play a crucial role in diminishing randomness and enhancing robustness. In response to the re... | Summary: This paper studies session-based recommendation in E-commerce. The authors propose to enhance user intent identification by constructing attribute transition graphs and using frequent attribute patterns as memory to augment session representations. The study employs frequent graph pattern mining algorithms to ... | Rebuttal 1:
Rebuttal: **W1. Unclear motivation**
A1. In E-commerce services, like scenarios involving new users or those in private mode, session-based recommendation (SBR) is challenging as it doesn't incorporate user profiles. Solely relying on temporal data is a limitation in prior research. Simultaneously, constru... | null | NeurIPS_2023_submissions_huggingface | 2,023 | Summary: The paper proposes a framework that effectively utilizes attribute graph patterns to enhance anonymous sequence encoding for session-based recommendations. Given the lack of personal information in session-based recommendation scenarios, making accurate item suggestions within the session is crucial. The autho... | Rebuttal 1:
Rebuttal: **W1. Unclear novelty**
A1. The session-based recommendation (SBR) does not provide the user profile in E-commerce services, such as for new or private mode users. Solely relying on temporal data remains limited in prior research. Simultaneously, constructing graph topologies might introduce nois... | null | null | null | null | null | null |
A Reduction-based Framework for Sequential Decision Making with Delayed Feedback | Accept (poster) | Summary: The paper studies the problem of sequential decision-making problems with delayed feedback. The authors propose a reduction from standard sequential decision-making algorithms to the following problem providing regret guarantees in many settings: linear bandits, tabular RL, linear and general function approxim... | Rebuttal 1:
Rebuttal: Thanks for your careful reading and valuable comments!
Q1: Presentation of the framework
A1: In fact, the framework we propose in Algorithm 2 is already for the general setting, which covers both single-agent and multi-agent, both tabular and linear. This framework can transform any multi-batche... | Summary: This submission studies the problem of learning from delayed feedback in various online learning settings including multi-armed bandits, linear bandits, linear MDPs, RL with function approximation, and various markov game settings (among others). Delays are assumed to be i.i.d. random variables. Additionally, ... | Rebuttal 1:
Rebuttal: Thanks for your careful reading and valuable comments!
Q1: Writing and readability
A1: To comply with the page limit, we have moved the details about multi-agent Markov Games to the appendix. We will polish the contents to ensure clarity and coherence for the readers in the revision. We apprecia... | Summary: The paper presents a new framework for handling stochastic delayed feedback in general sequential decision-making problems, encompassing bandits, single-agent Markov decision processes (MDPs), and Markov games (MGs). The authors introduce a novel reduction-based framework that converts any multi-batched algori... | Rebuttal 1:
Rebuttal: Thanks for your careful reading and valuable comments!
Q1:Is the contribution rather incremental? Explain specific limitations of the current state-of-the-art methods.
A1:We respectfully disagree with the reviewer’s argument that our contribution is rather incremental. Please see our **common re... | Summary: This paper provides a general framework for analysing any ‘mini-batch’ or rarely switching algorithm in the presence of delayed feedback. This approach covers bandits, finite horizon MDPs and finite horizon Markov games. The results provided match or improve on the best-known results for delayed feedback in se... | Rebuttal 1:
Rebuttal: Thanks for your careful reading and valuable suggestions!
Q1. Writing and clarity
A1. To comply with the page limit, we have moved the details about Markov Games to the appendix. We will arrange the contents to ensure clarity and coherence for the readers. We appreciate your valuable feedback.
... | Rebuttal 1:
Rebuttal: ## Common rebuttal for the contributions of our paper:
We respectfully disagree with the reviewers’ argument that our contribution is rather incremental. Here we describe some drawbacks of previous works, and how our work addresses them.
The previous works suffer from several drawbacks: (i) they... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Dense and Aligned Captions (DAC) Promote Compositional Reasoning in VL Models | Accept (spotlight) | Summary: This paper studies the problem of getting models like CLIP to perform compositional reasoning. One observed problem with these models is that they devolve into "bags of objects" models, and this work seeks to address their variable-binding ability towards identifying more complex relationships. This paper take... | Rebuttal 1:
Rebuttal: We thank the reviewer for the insightful comments. In the following, we provide a response to the questions raised in the review:
1. **On the most cost-effective caption enrichment strategy:**
Great question! In the ablation section 5, in Table 4 of the paper, we evaluated the contributions of di... | Summary: The current vision-language (VL) models suffer from "object bias" issues, this paper proposes two components to improve the compositional reasoning capability of the existing VL models. It employs the pretrained LLMs to augment the captions of the images, in terms of quality and density, then finetune the CLIP... | Rebuttal 1:
Rebuttal: We thank the reviewer for the time and effort spent reviewing our paper. In the following, we provide a response to the questions raised in the review:
1. **On the paper contributions:**
Our paper establishes caption quality and caption density to be very important factors in finetuning VL models... | Summary: This paper highlights two problems with existing image-text datasets that make them unsuitable for use as pre-training datasets for evaluating performance on attributes and relations, and claims that the models simply act as bag of words when trained on these datasets. The first problem they identify is that o... | Rebuttal 1:
Rebuttal: We thank the reviewer for the insightful comments. We address all the reviewer comments below:
1. **On the advantage of the caption quality improvement:**
We performed a user study to complement the evidence provided in the paper (CLIP score analysis - Fig 3a; qualitative examples - Fig 3b; the i... | Summary: The authors propose a method for data augmentation on image-text web
corpora they call DAC. The main idea is to run both a segmentation
model and an image captioning model, and then use those outputs to
generate captions that are likely to describe the image using an
LLM. Finally, a MIL loss + LoRA over CLIP i... | Rebuttal 1:
Rebuttal: We thank the reviewer for the insightful comments. In the following, we provide a response to the questions raised in the review:
1. **On caption quality references in L43-44:**
Thanks for pointing this out! We **will revise** the L43-44 caption quality reference to be more like the way caption q... | Rebuttal 1:
Rebuttal: We thank all the reviewers for their efforts in reviewing our paper and for
providing helpful and insightful feedback.
We are happy to see that they found our work: **interesting and creative** `(93oA, bpWN, Fh7T)` and **addressing important questions** `(Fh7T)`.
Furthermore, we also thank them... | NeurIPS_2023_submissions_huggingface | 2,023 | Summary: The paper proposes methods to improve performance on compositional reasoning tasks — by improving the caption quality (alignment) and density (description bias). The paper observes that the poor compositional task performance is due to these limitations in pretraining / fine-tuning data. The approach itself, i... | Rebuttal 1:
Rebuttal: We thank the reviewer for the insightful comments. In the following, we provide a response to the questions raised in the review:
1. **On filtering over-segmentation:**
Thank you for this suggestion! It is indeed interesting to explore filtering beyond the proposed MIL losses. While we would pref... | null | null | null | null | null | null |
GPT is becoming a Turing machine: Here are some ways to program it | Reject | Summary: Authors propose a way to prompt GPT-3 to exhibit behavior simulating execution of iterative programs. Authors propose the following prompt constructs: providing structured examples of program execution; using fragments of execution; not using self-attention on some parts of the generated text. Authors compare ... | Rebuttal 1:
Rebuttal: For any program, we prompt the LLM with an execution path described well, but using arbitrary language (keywords and such). Therefore, an automatic way of creating IRSA prompts would be to simply generate an execution path programmatically, but following the rules of general repetitive structure, ... | Summary: Making LLMs follow procedural rules precisely, as done when executing a program, is been a challenging task. In this paper, the authors introduce iterations by regimenting self-attention (IRSA), a prompting technique to make large-language models (LLMs) *execute* a hand coded programs (on novel inputs) precise... | Rebuttal 1:
Rebuttal: The versions of GPT we had available at the time of submission had 4k token limit. The problem with LCS is that its state is proportional to the product of sequence lengths (the DP matrix). With newer models, this will indeed be less of an issue (see the response to the previous reviewer, too). Ne... | Summary: The paper proposed several novel prompting methods that could trigger GPT-3 to perform iterative behavior for executing algorithms with loops. The main technique the paper presented, IRSA, is to use highly structured prompts that contains information about unrolled execution trace, program states, and a detail... | Rebuttal 1:
Rebuttal: We thank reviewer 2ucu for acknowledging that our work showed great originality and insights, our prompt methods showed state-of-the-art results on multiple tasks, and the proposed Skip attention and Fragmented prompting are promising techniques as a solid base for future work. Below we address al... | Summary: The paper explores the use of regimented self-attention (IRSA) to prompt GPT-3 to perform iterative behaviors necessary for executing programs involving loops. The authors investigate three approaches to trigger the execution and description of iterations. The results suggest that IRSA leads to larger accuracy... | Rebuttal 1:
Rebuttal: We thank the reviewer for acknowledging that our problem and approach are interesting and the potential applications of IRSA in education. Below we address all the reviewer's concerns and questions.
(a) **Reliability**. We are using much stricter attention control than usual CoT prompts, howe... | Rebuttal 1:
Rebuttal: We thank the reviewers for the great comments and constructive suggestions. We are encouraged that reviewers agree that the paper provides "Interesting problem and approach,” that fragmented prompting and skip attention, "which pair well together", are original and can have applications beyond pro... | NeurIPS_2023_submissions_huggingface | 2,023 | Summary: This work introduces Iterations by Regimenting Self-Attention (IRSA) which is a set of LLM prompting techniques for producing repetitive, algorithm-like behavior that can be useful for a range of tasks, such as carrying out a sorting algorithm or solving a logic puzzle. There are 3 techniques discussed: 1) "Ba... | Rebuttal 1:
Rebuttal: Thanks a lot for acknowledging that our contributions of skip attention and fragments this would be "valuable work for the NeurIPS community", our approaches are novel and original, and our evaluation is reasonable that the skip attention method can work great and generalize to very long sequences... | null | null | null | null | null | null |
Differentiable Sampling of Categorical Distributions Using the CatLog-Derivative Trick | Accept (poster) | Summary: This paper proposes a variance-reduced statistic for $\mathbb{E}_{x \sim p(x)} [f(X)]$, which could be the gradient of a loss for example. The main idea (known as Rao-Blackwellization) is simple to understand and can be summarized thusly: the quantity can be rewritten to distinguish any dimension $d$:
$E_{(x_... | Rebuttal 1:
Rebuttal: We appreciate the concise explanation of our method and the positive assessment. Please, find below a remark about the formulas summarizing our method and the answer to your question.
**Remark**
Please, note that the CatLog-Derivative trick can be expressed in the following way (using your same ... | Summary: The paper discusses improved gradient estimation through multivariate categorial probability distributions. A special case of independent random variables is discussed more thoroughly and experiments are conducted only for this case. To summarize for independent variables i.e. $p(X) = \prod_{i \in [N]} p(X_i)$... | Rebuttal 1:
Rebuttal: We thank the reviewer for the time dedicated to review our paper. Please, find below a discussion about the mentioned weaknesses and the answers to the questions.
# Discussion
The improvements with respect to RLOO-F are statistically significant, as can be seen from the reported standard error b... | Summary: In this paper the authors provide an alternative to the log derivative trick, e.g. the REINFORCE estimator, for categorical distributions, which is unbiased and has provably lower variance than the REINFORCE estimator; the estimator is called the CatLog-Derivative trick. For the case D-dimensional multivariate... | Rebuttal 1:
Rebuttal: Thank you for the enthusiasm and the constructive suggestions. Please find below a comment about the weaknesses and the answers to your questions.
# Comment
We acknowledge that we have primarily devoted attention to IndeCateR, instead of the more general CateR. Our primary focus was on the most ... | Summary: The authors derive a gradient estimator for discrete random variables that sums out one dimension while keeping a sample for the other dimensions fixed. Their estimator works for generally fully factorised distributions, and they derive a variant for fully independent distributions, with which they perform exp... | Rebuttal 1:
Rebuttal: We thank the reviewer for the dedicated time and for bringing the related work on local expectation gradients (LEG, NeurIPS 2015) [1] to our attention. We also thank them for their appreciation about the clarity, significance and quality of our work. Please, find below a discussion on the relation... | Rebuttal 1:
Rebuttal: We thank all reviewers for their valuable feedback and want to specifically draw attention to the comparison between CatLog and Local Expectation Gradients (LEG) [1] as brought up by reviewer JDGY. Please find our theoretical analysis that formally proves the differences in the added pdf.
Indeed,... | NeurIPS_2023_submissions_huggingface | 2,023 | Summary: This paper develops a new efficient estimator for gradients of an expectation computed over a multivariate discrete random variable. The gradients are computed wrt the continuous parameters of the discrete distribution of interest.
Previous work consists of two major threads
- unbiased estimators, such as REI... | Rebuttal 1:
Rebuttal: Thanks for the detailed and thoughtful review and the extra effort to provide feedback aiming at improving the quality of the presentation and at sharpening some statements. Please, find below a discussion to your raised points and the answers to your questions.
# Discussion about weaknesses
**A... | null | null | null | null | null | null |
When Do Transformers Shine in RL? Decoupling Memory from Credit Assignment | Accept (oral) | Summary: The paper proposes a set of metrics to measure and isolate the memory dependency of partially observable environments.
The authors illustrate that some currently existing benchmarks do not sufficiently isolate the memory of an agent.
Based on their analysis they propose two versions of a T-Maze environment, on... | Rebuttal 1:
Rebuttal:
> Does computing memory lengths involve human demonstrations?
We here clarify that computing these metrics *does not* require any human demonstrations. For some tasks like T-Maze, we obtain the precise numbers with careful analysis of the task structures. For other complex tasks like Memory Maze... | Summary: This work explores the impact of memory and credit assignment in decision transformer architectures, presenting several significant claims. Regarding memory, the authors establish an upper bound on the memory length required for an optimal policy. In terms of credit assignment, they provide a lower bound on th... | Rebuttal 1:
Rebuttal: > This work explores the impact of memory and credit assignment in decision transformer architectures.
We would like to clarify that our experiments evaluate the *online model-free RL* setting, not the offline RL setting that is related to Decision Transformers (DT). Nevertheless, extending our e... | Summary: This work aims to answer how the parametrization of policy in RL affects the performance of sparse-reward tasks that require memory and credit assignment. In particular, the authors are interested in when transformers perform well. The authors define in the POMDP settings their own criteria of how to quantify ... | Rebuttal 1:
Rebuttal:
> For policies with memory (i.e. recurrent), the definition seems less relevant since the policy can remember parts of all the information starting from time 0.
Recurrent policies can *take* all the information starting from time 0 as inputs, but not necessarily *remember* it starting from that ... | Summary: In this paper the authors consider the effectiveness of transformers for solving two kinds of RL problems: credit assignment and memory. To achieve their goal the authors develop rigorous definitions for memory and credit assignment so that different RL tasks can be understood and compared in these terms. The ... | Rebuttal 1:
Rebuttal: > What potential solutions could be to the credit assignment problem when using Transformers?
We think the credit assignment problem might be easier to solve with Transformers by using them to explicitly redistribute reward signals [Liu et al., 2019, Ferret et al., 2020]. Specifically, Transforme... | Rebuttal 1:
Rebuttal: # General Response
First, we would like to thank all four reviewers for their positive and constructive feedback on our work!
During the rebuttal period, we conducted additional experiments to address some of the questions posed by the reviewers.
**mXpQ (see Figure 1 in the PDF)**
> Can we si... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Linguistic Binding in Diffusion Models: Enhancing Attribute Correspondence through Attention Map Alignment | Accept (oral) | Summary: This paper presents an approach called “SynGen” to improve attribute binding between nouns and their modifiers for text-to-image latent diffusion models, specifically Stable Diffusion. They propose a two-part loss on attention maps that (1) encourages attention maps of nouns and their modifiers to be similar, ... | Rebuttal 1:
Rebuttal: **W1: Writing doesn’t highlight the connection to A&E. Especially section 2:** Thank you for pointing out that the connection to A&E did not come across clearly. We will revise section 2 to give just credit and clarify the relation to A&E.
**W2: "Concept Separation" is too general. Consider colle... | Summary: A frequent issue in text-prompted image generation (and in many other grounded language scenarios) is that a model will treat text akin to a bag of words and ignore syntactic relationships, such as which adjective attaches to which noun -- this is referred to as a problem with lingusitic binding. The paper pro... | Rebuttal 1:
Rebuttal: Thank you for your supportive review of this work. The reviewer raises an important point in designing future systems using SynGen. Future work will need to take into account failures of the parser and develop ways to handle them.
---
Rebuttal Comment 1.1:
Comment: I have read the other reviews,... | Summary: The paper focuses on attribute binding/leak/neglect problems in text-to-image models. The authors propose, SynGen, a method that utilizes dependency trees combined with cross-attention map optimization to achieve better attribute binding results. Extensive experiments are conducted to show the effectiveness of... | Rebuttal 1:
Rebuttal: **W1: What is the speed SynGen compared to the original SD?** Syngen is about the same speed as A&E (~10% slower). This is about twice slower than vanilla SD. We did not invest in performance tuning, so we assume speed can be improved considerably once we do. We will add this information to the Li... | Summary: This paper approaches the binding problem in text to image diffusion models that is marked by the inability of the models to appropriately identify which modifiers are attached to which nouns in the input text. This often results either in images that mix up the attributes of the different nouns that are menti... | Rebuttal 1:
Rebuttal: **Provide qualitative examples having uncommon modifiers vs different number:** Following this suggestion. We will show qualitative examples of SynGen and our baselines on prompts with a varying number of modifiers. Specifically, with 2, 3, 4, 5, and 6 modifiers. Note that the DVMP dataset does no... | null | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
3D-Aware Visual Question Answering about Parts, Poses and Occlusions | Accept (poster) | Summary: This paper introduces the concept of 3D-aware Visual Question Answering (VQA) and presents a new dataset called "Super-CLEVR-3D". It also proposes a model called "PO3D-VQA" that combines probabilistic neural symbolic program execution with deep neural networks using 3D generative representations of objects. Th... | Rebuttal 1:
Rebuttal: We appreciate the time and effort you put into reviewing our paper. Here we address the questions you raised.
**Q1: Why not use point cloud or multi-view images, instead of single-view images, as input?**
A1: We would like to highlight the focus of our paper: 3D-aware VQA from 2D images, i.e. a... | Summary: 1, The paper extends 2D VQA to the task of 3D-aware VQA, which requires the understanding of the 3D structure of visual scenes and includes knowledge of 3D object pose, parts, and occlusions.
2. Introducing the Super-CLEVR-3D dataset, which contains questions about object parts, their 3D poses, and occlusions.... | Rebuttal 1:
Rebuttal: We appreciate your thoughtful and positive feedback on our paper. Please find our responses below:
**Q0 & A0: Please refer to the general response for generalization to real images.**
**Q1: Lack of some statistics about the constructed Super-CLEVR-3D dataset.**
A1: We include the dataset statis... | Summary: This work introduces a framework to Visual Question Answering (VQA) that incorporates understanding of the 3D structure of scenes, a significant leap from traditional 2D-based models. To evaluate the task, they proposed a new dataset, Super-CLEVR-3D, designed specifically for 3D-aware VQA, containing questions... | Rebuttal 1:
Rebuttal: Thank you for your feedback. Here we address the questions:
**Q0 & A0: Please refer to the general response for generalization to real images.**
**Q1: About 3D navigation and manipulation: Are there questions directly related to navigation and manipulation in 3D space?**
A1: There are no questi... | Summary: This paper is seeks to improve the VQA models' understanding of 3D structure of images, particularly parts, poses, and occlusions. There are two main contributions:
1. Super-CLEVR-3D: a dataset that contains questions about parts, poses, and occlusions.
2. PO3D-VQA: a 3D aware VQA model that combines nuerosy... | Rebuttal 1:
Rebuttal: We sincerely thank you for the detailed discussions and the appreciation of our work. We would like to address the concerns and questions you raised:
**Q0 & A0: Please refer to the general response for generalization to real images.**
**Q1: There aren't questions about “z” direction (“above/bel... | Rebuttal 1:
Rebuttal: # General Response
We thank all the reviewers for the feedback. We are glad that all four reviews are positive and acknowledge both of our contributions: (1) The Super-CLEVR-3D dataset, i.e. a new VQA dataset that introduces an important new task (HP3a, NU9U) that requires reasoning over object p... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Boosting Adversarial Transferability by Achieving Flat Local Maxima | Accept (poster) | Summary: The authors empirically observed that the adversarial data lies on the flat local maxima yields enhanced attack transferability. Inspired by this observation, this paper proposes a regularization to help the gradient-based attacks to find the adversarial data at the flat local maxima. This regularization penal... | Rebuttal 1:
Rebuttal: Dear Reviewer 2kUx,
Thanks for your valuable and insightful comments. We address your concerns as follows:
**Q.1.** Could the authors provide the error bar?
> **A.1.** Following your suggestions, we will add all error bars to the final version of the paper. Here we provide part of the error bars ... | Summary: The paper proposes a method called Penalizing Gradient Norm (PGN) to improve the transferability of adversarial perturbations. The method is motivated by the observation that encouraging the flatness of the local landscape for adversarial examples can lead to better transferability, and thus PGN regulates the ... | Rebuttal 1:
Rebuttal:
Dear Reviewer 8Bpx,
Thanks for your valuable and insightful comments. We address your concerns as follows:
**Q.1.** PGN v.s. RAP
>**A.1.** As you suggested, we discuss the similarities and differences between PGN and RAP methods as follows.
>1) **Empirical verification**. Both RAP and PGN aim to... | Summary: This paper aims to boost adversarial transferability by using the Penelizing Gradient Norm, which can restrict adversarial examples located in flat regions. The writing is good, and it is easy to read. The experiments demonstrate that the proposed method achieves good results.
Strengths: - The motivation is ... | Rebuttal 1:
Rebuttal: Dear Reviewer xHRb,
Thanks for your valuable and insightful comments. We address your concerns as follows:
**Q.1.** In Theorem 1, the authors briefly ... key design of PGN actually works well.
>**A.1.** As you suggested, we first theoretically analyze the acceleration effect of the finite differ... | Summary: In this work, the authors first assume and empirically validate that adversarial examples at flat local minima tend to have better adversarial transferability. Based on this finding, they introduce a regularizer on the gradients in the neighborhood of the input sample to achieve flat local minima. To make the ... | Rebuttal 1:
Rebuttal: Dear Reviewer PZXp,
Thanks for your valuable and insightful comments. We address your concerns as follows:
**Q.1.** Results on vision transformers, such as ViT, Swin, etc. would better be included.
> **A.1.** As you suggested, we further adopt four mainstream vision transformers to evaluate the ... | Rebuttal 1:
Rebuttal: Dear Reviewers and Area Chairs:
We sincerely appreciate all of your precious time and constructive comments. The results of the partial response data (Tables) are in the pdf file.
Pdf: /pdf/c43b63a8954f8e1968febe3320dc44341fd31911.pdf | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Fantastic Robustness Measures: The Secrets of Robust Generalization | Accept (poster) | Summary: This paper assesses the correlation between existing measures and the robust generalization gap in various experimental settings. Specifically, the author evaluates the relationship between measures based on weight-norm, margin, smoothness, flatness, gradient-norm, and robust generalization under different con... | Rebuttal 1:
Rebuttal: We appreciate the reviewer's insightful assessment of the strengths and weaknesses of our paper. Your feedback provides valuable insights that help us improve the quality and clarity of our work. Below, we provide a detailed response addressing each of the weaknesses [W#] raised with all referenc... | Summary: This paper studies the correlation between previously proposed robustness measures and the robustness generalization gap. It studies a large number of models with different architectures and training parameters and shows that some prior beliefs about the usefulness of popular robust generalization measures are... | Rebuttal 1:
Rebuttal: We appreciate the reviewer's insightful assessment of the strengths and weaknesses of our paper. Your feedback provides valuable insights that help us improve the quality and clarity of our work. Below, we provide a detailed response addressing each of the weaknesses [W#], questions [Q#], and lim... | Summary: The authors propose a large-scale empirical analysis of the impact of several potential indicators of robust generalisation. This paper can be seen as a metaanalysis of many of the indicators proposed in the literature.
Strengths: 1. I appreciate a paper aiming to test and/or reproduce at a larger scale of ex... | Rebuttal 1:
Rebuttal: We appreciate the reviewer's insightful assessment of the strengths and weaknesses of our paper. Your feedback provides valuable insights that help us improve the quality and clarity of our work. Below, we provide a detailed response addressing each of the weaknesses [W#], questions [Q#], and lim... | Summary: The paper conducts a large scale correlation analysis for adversarially trained models to identify which existing robustness-related measures are predictive of robust overfitting. The correlation analysis is performed over large set of relevant dimensions (including model architectures, training procedures, at... | Rebuttal 1:
Rebuttal: We appreciate the reviewer's insightful assessment of the strengths and weaknesses of our paper. Your feedback provides valuable insights that help us improve the quality and clarity of our work. Below, we provide a detailed response addressing each of the weaknesses [W#] and questions [Q#] raise... | Rebuttal 1:
Rebuttal: **Dear all,**
We would like to thank the editor and the reviewers for their careful comments and suggestions.
We summarize the reviews according to our own perspective.
****
**Strengths.**
We are glad that Reviewers PVgT, 2BXc, and vGwc found that our results **“uncover some interesting findin... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Active Bipartite Ranking | Accept (poster) | Summary: The paper proposes an elimination active learning criterion for the bipartite ranking problem. The criterion account for the local smoothness of a data sample’s posterior distribution. The author proves the sample complexity and the PAC learnability of the proposed algorithm under some assumptions.
Strength... | Rebuttal 1:
Comment: Please see the general rebuttal in regards to your first question.
Indeed, due to the interconnected nature of the problem the decision to remove a point may depend on points already removed from the active set, a problem not found in multi armed bandits, see the discussion at the end of section ... | Summary: The paper proposes an active learning for the bipartite ranking problem in which the goal is to provide a ranking for a set of points rather than given positive or negative labels. The proposed algorithm called active-rank tries to optimize the ROC curve, evaluated by the distance to the optimal one. The autho... | Rebuttal 1:
Comment: Please see the general rebuttal in regards to your first and second questions.
The definition of scenario 1 was wrong, the correct version is,
Scenario 1: $\mu_i = 0$ when $i\leq 11$, $\mu_{12}=\mu_{13} = 0.23$, $\mu_{14}=\mu_{15} = 0.33$, $\mu_{16}=0.35$ and $K = 16$.
This scenario aims at ... | Summary: This paper develops an active learning framework for the bipartite ranking problem. A selective sampling procedure or strategy is necessary for lebeling data points sequentially. The proposed method is called active-rank, which aims to minimize the distance between the ROC curve of the ranking function built a... | null | Summary: An active bipartite ranking algorithm is proposed in this article for one-dimensional data with a posterior probability $\eta(x) = {\rm P}(Y =+1\vert X =x )$ piecewise constant on a grid of size $K$. An upper bound is provided on the number of queried to achieve a controlled error of type PAC$(\epsilon,\delta)... | Rebuttal 1:
Comment: I thank the authors for the discussion on the comparison to passive learning, which is helpful and worth developing in the main text. My score is increased. | Rebuttal 1:
Rebuttal: We thank all reviewers for the time taken to read our paper and for their constructive feedback. Some topics come up for several reviewers.
Performance of passive approach in comparison to our active algorithm:
For a uniform/passive sampling strategy to be PAC$(\epsilon,\delta)$ one would have... | NeurIPS_2023_submissions_huggingface | 2,023 | Summary: - This paper studies sample complexity bounds for bipartite ranking under an active learning paradigm. The paper gives an algorithm called “active-rank” to solve the variant of the problem under the assumption that the posterior is a piecewise constant function over a grid of size K. The sample complexity boun... | Rebuttal 1:
Comment: Please see the general rebuttal in regards to your first question.
Line 159 is a typo, it should read "for all $j: |\mu_i - \mu_k| \geq \Delta_i$".
The x-axis in figure 4 is indeed stopping time, this will be clarified.
Aside from scenario one, for larger stopping times, i.e. greater than 500... | null | null | null | null | null | null |
DIFFER:Decomposing Individual Reward for Fair Experience Replay in Multi-Agent Reinforcement Learning | Accept (poster) | Summary: This paper proposes to train a mixing network and agent networks separately for multiagent reinforcement learning when the system is based on the centralized training and decentralized execution (CTDE) paradigm. The authors derive an individual reward that maintains invariance between the gradients of the mixi... | Rebuttal 1:
Rebuttal: Dear Reviewer:
Thanks for your valuable comments.
**Q1: Is $\widetilde{Q_i}$ necessary?**
A1: Your understanding is indeed correct.
1) In the computation of the TD error for the agent networks, there is no need to explicitly determine the value of $\widetilde{Q_i}$ because it cancels out withi... | Summary: This is a very interesting and strong paper on MARL that seeks the strategy of learning individual policies rather than team policies in order to break symmetry and learn identity and contextual influences on the individual policies. In order to do so, it solves the technical problem of decomposing a team re... | Rebuttal 1:
Rebuttal: Dear Reviewer:
Thanks for your valuable comments. We sincerely appreciate your positive evaluation and recognition of our research.
To enhance our work’s robustness and comprehensiveness, we plan to incorporate additional experiment results, intuitive visualization, and theoretical analysis in ... | Summary: This paper introduces a method for decomposing shared rewards into individual rewards within value-based multi-agent reinforcement learning (MARL) methods. By decomposing the rewards, the method enables the prioritization of crucial individual experiences. Experimental results demonstrate the effectiveness of ... | Rebuttal 1:
Rebuttal: Dear Reviewer:
Thank you for the constructive comments.
**Q1: Intuitive visualization of the decomposed individual rewards.**
A1: We have included visualizations displaying the individual rewards for each agent, along with corresponding screenshots, which can be found in the Figure 3 of Rebutta... | Summary: The paper presents a new algorithm DIFFER to help individual agents to learn individual TD-error such that it improves the overall (team) model performance. The paper's proposed method is different from an existing method MASER in that it is maintaining the overall team objective. Overall model performance com... | Rebuttal 1:
Rebuttal: Dear Reviewer:
Thanks for your valuable suggestion.
1) We conducted experiments to compare the performance of DIFFER and MASER in SMAC environments. According to your suggestion, we have conducted additional experiments to meticulously evaluate the performance of DIFFER in relation to MASER on t... | Rebuttal 1:
Rebuttal: We extend our sincere gratitude to all the esteemed reviewers for their invaluable and constructive comments.
The consistently positive reviews our submission received from all the Reviewers have filled us with delight.
We are greatly encouraged by their recognition of the originality (Review f7X... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Training Energy-Based Normalizing Flow with Score-Matching Objectives | Accept (poster) | Summary: The authors consider normalizing flows with linear transformations. Computing the determinant of their Jacobian is expensive, limiting their applicability when training via maximum likelihood. Therefore, the authors introduced an energy-based method to train normalizing flows with linear transformation, which ... | Rebuttal 1:
Rebuttal: We appreciate the reviewer’s time and effort spent on the review, and would like to respond to the reviewer’s questions as follows. Please note that parts of the responses are provided in the [global comment](https://openreview.net/forum?id=AALLvnv95q¬eId=9f9rk2L3tW) due to the word limit.
---... | Summary: Normalizing flows are generative models trained by maximizing the log-likelihood of the dataset given the model, which is expressed as the sum of the base probability of a transformed variable plus a series of log-Jacobian terms that keep track of the volume distortion factors. Evaluating the log-Jacobian is u... | Rebuttal 1:
Rebuttal: We would like to thank the reviewer for thoroughly summarizing our paper, and would like to respond to the reviewer’s questions as follows.
---
### **Comments**
---
**C1.**
Major points: **(a)** In general, I am not convinced that this approach solves a major problem in the current literature. In... | Summary: The authors use a connection between flows and EBMs to devise a new training method for flows via score-matching.
Their training procedure improves efficiency via avoiding computation of the Jacobian determinants that are independant to values of the samples (i.e. only contribute to the normalizing constant of... | Rebuttal 1:
Rebuttal: We sincerely thank the reviewer for the time and effort spent on the review, and would like to respond to the reviewer’s questions as follows.
---
### **Comments**
---
**C1.**
**(a)** For sample generation by inverting the flow, the inverse matrix for the linear flow layers will cost $D^3$ - thi... | Summary: This paper proposes a new flow based approach to generative modeling called EBFlow. EBFlow uses training objectives from EBMs to reduce training cost for flow based methods unlike previous approaches that mostly achieved this via different architectures or biased estimation methods. The main insight for EBFlow... | Rebuttal 1:
Rebuttal: We appreciate the reviewer’s time and effort spent on the review, and would like to respond to the reviewer’s questions as follows.
---
### **Comments**
---
**C1.** I am not sure how to interpret the empirical results. It is obvious from the plots in Fig 2 that the proposed method is much faster... | Rebuttal 1:
Rebuttal: This global comment includes extended discussions of the questions raised by reviewers htQa and fofM. The attached PDF file contains additional experimental results.
---
### **Reviewer htQa (Cont'd)**
---
**C8.** I think the author intends to model $g(\theta)$. Therefore, it is better to avoid us... | NeurIPS_2023_submissions_huggingface | 2,023 | Summary: This paper has proposed an energy-based normalizing flow model, where the computaiton of Jacobin determinants for linear transformations can be skiped with score-matching objectives. This could enable deeper layers of linear transformation and make the flow model more efficient.
Strengths: This paper has prop... | Rebuttal 1:
Rebuttal: We appreciate the reviewer’s time and effort spent on the review, and would like to respond to the reviewer’s questions as follows. Please note that parts of the responses are provided in the [global comment](https://openreview.net/forum?id=AALLvnv95q¬eId=9f9rk2L3tW) due to the word limit.
---... | null | null | null | null | null | null |
Universal Online Learning with Gradient Variations: A Multi-layer Online Ensemble Approach | Accept (spotlight) | Summary: In this work, the authors aim to develop an algorithm for online convex optimization that can adapt to the specific curvature of the loss function without knowing it in advance. The goal is to develop an algorithm that can adapt to strongly convex, exp-concave and convex loss functions dynamically, while al... | Rebuttal 1:
Rebuttal: Thanks for the constructive feedback and appreciation of our work! Below we answer your questions on the lower bounds and the possibility of further improvements.
**Q1**. It would be interesting to get comparisons with lower bounds, in particular for the small loss bounds that depend in F and V.
... | Summary: This paper put together and designed new techniques to reach a high level of adaptivity in online convex optimization. It is a timely contribution that is useful to put together various algorithms optimal for different problem setups and to better adapt to the characteristic of each setup.
Strengths: This is ... | Rebuttal 1:
Rebuttal: Thanks for the valuable feedback and appreciation of our work! We believe that our research occupies a nonnegligible position in universal online learning, especially given the significance of the gradient-variation quantity coupled with the novel techniques we've introduced.
The gradient variati... | Summary: This paper proposes algorithms that have data-dependent regrets and adaptive to the loss function classes. The algorithm improves on Zhang et al., 2022 in the sense that 1) it also achieves a gradient variation bound for convex Lipschitz losses and 2) it requires only one gradient evaluation at each round. The... | Rebuttal 1:
Rebuttal: Thanks for the valuable feedback and appreciation of our work! Given that all three questions (computational complexity, base learners setup, and number $N$) relate to the base learner configuration, we provide a comprehensive explanation for it first, followed by the answers to your questions.
-... | Summary: A new algorithm for online learning is provided, together with improved regret bounds compared to the state-of-the-art:
The new regret guarantees are summarised in table 1 on page 2: in the case of a strongly convex function, the regret guarantee is of $O(min(\log(V_T),\log (F_T)))$ where $V_T$ and $F_T$ are... | Rebuttal 1:
Rebuttal: Thanks for the valuable feedback and very careful check of our paper! We will carefully revise it according to the mentioned points. Due to the 6,000-character limit of this year's rebuttal, we first answer your questions about [Chen et al., 2023] and [Zhang et al., 2022] and then respond to prese... | null | NeurIPS_2023_submissions_huggingface | 2,023 | Summary: This paper proposes an online convex optimization algorithm that is adaptive in two ways to the environment.
Specifically, it can achieve logarithmic regrets for problems with good properties such as strong convexity and exp-concavity,
as well as $\sqrt{T}$ regret in the worst-case.
Further,
it achieves improv... | Rebuttal 1:
Rebuttal: Thanks for the valuable feedback. We believe that our work has solid contributions and holds significant interest to the community. The submitted version may not have effectively conveyed our findings clearly, leading to the reviewer's perception. We will carefully revise the paper to clarify the ... | Summary: Addressing to the drawback of Zhang et al. [2022], which does not enjoy gradient-variation bounds for convex functions, using an optimistic version of ADAPT-ML-PROD with a second-order bound of [Wei et al., 2016],three-layer online ensemble structure with two-layer meta learner running MSMWC [Chen et al., 2... | Rebuttal 1:
Rebuttal: Thanks for the insightful feedback. Below we clarify our unique contributions and differentiate our approach from existing ones, hoping to address your concerns.
---
**Q1**. The derived results substantially depend on [Zhang et al., 2022], [Chen et al., 2021], [Zhao et al., 2021], and [Wei et al... | null | null | null | null |
NAS-X: Neural Adaptive Smoothing via Twisting | Accept (poster) | Summary: The paper proposes NAS-X, a novel approach to estimate the posterior expectations in Reweighted Wake Sleep (RWS) architectures. Instead of a traditional estimation using self-normalized importance sampling (SNIS), and similar to Neural Adaptive Sequential Monte-Carlo (NASMC), the proposed method employs a Sequ... | Rebuttal 1:
Rebuttal: Thanks for your comments and feedback. Just to be clear on terminology --- In your review you refer to “RWS architectures”, but we weren’t sure if you meant model architecture or a general framework for developing new methods. We take the later interpretation and refer to it as the “RWS approach” ... | Summary: The paper tackles the problem of learning sequential latent variable models using a new method called NAS-X. It combines two previous research: 1. The smoothing with twisting that’s applied in variational inference (SIXO in particular). It approximates smoothing distribution as targets for proposal learning, i... | Rebuttal 1:
Rebuttal: We thank the reviewer for their comments and constructive feedback. Below we respond to the questions/concerns raised.
**The paper lacks discussions of future work and limitations of the proposed method.**
One limitation is the NAS-X does not have a unified objective (i.e., the proposal and mode... | Summary: An algorithm based on reweighted wake-sleep (RWS) is applied to sequential latent variable models $p_\theta(x_{1:T}, y_{1:T})$. Model parameters are fit by maximizing the evidence $p_\theta(y_{1:T})$ in $\theta$, and posterior inference is performed by minimizing the forward KL divergence $KL(p_\theta(x_{1:T} ... | Rebuttal 1:
Rebuttal: We thank the reviewer for their comments and feedback. In addition to the small changes you suggested, we also respond to your main questions and concerns:
**Lack of comparison to existing methods.**
In new experiments (see PDF), we compare NAS-X to several more baselines, including FIVO, SIXO, ... | Summary: The authors propose to use SIXO particle approximation to calculate the expectations in reweighted wake-sleep algorithm for state space models.
Strengths: The paper is written in a clear way and all required background is well-explained (but description of SIXO and twists is too short).
The idea of replacin... | Rebuttal 1:
Rebuttal: We thank the reviewer for their positive comments and constructive feedback. Below we respond to the questions/concerns raised.
**I would lengthen the description of SIXO and twists.**
Thank you for the suggestion. We have added a detailed discussion of the technical details behind SIXO and twi... | Rebuttal 1:
Rebuttal: We thank the reviewers for their detailed feedback. We respond to reviewers individually and provide a general response below. We have strengthened our submission with several new experiments (see figures in PDF) and theoretical analyses (see below). If the reviewers feel that the new experimental... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
AGD: an Auto-switchable Optimizer using Stepwise Gradient Difference for Preconditioning Matrix | Accept (poster) | Summary: The paper proposes a new gradient based optimization called AGD. The idea is to utilize gradient difference between the current and previous steps in the preconditioning matrix which is related to Hessian. Also, an autoswitching function is proposed to switch between SGD and the adaptive optimizer. As mentione... | Rebuttal 1:
Rebuttal: Thanks for your valuable review.
**Q1: Figure 3 suggests AGD is faster to converge, and Table 3 suggests AGD results in better performance. Could we make such a conclusion that it should be always used instead of SOTA optimization methods?**
AGD is at least the optimal method in the datasets and... | Summary: This paper introduces a novel optimizer, namely, Auto-switchable optimizer with Gradient Difference of adjacent steps. The authors propose a different way of approximating information from the Hessian for faster and better convergence in a way that a preconditioning matrix is computed using successive optimiza... | Rebuttal 1:
Rebuttal: Thanks for your constructive review.
**Q1: In my humble opinion, the issue with the current idea lies in terms of the number/width of the optimization steps and how far the solution lies compared to the initialization point in the parameter space ... That means that for real-world multi-dimension... | Summary: The paper proposes a new optimizer based on finite difference approximation to obtain the inner product between the Hessian row vector and the parameter vector difference from gradients of succeeding steps and an auto-switching function to switch between SGD and the adaptive optimizer. It uses an exponential m... | Rebuttal 1:
Rebuttal: Thanks for your thoughtful review.
**Q1: The explanation of related works is minimum. It needs to be more concretely clarified what problem of existing second-order methods it addresses.**
Second-order methods entail the computation or approximation of the Hessian, which can be computationally d... | Summary: The paper proposes a new optimizer called AGD that integrates the information of the Hessian into the preconditioning matrix and switches dynamically between SGD and the adaptive optimizer. The authors establish theoretically proven convergence guarantees in both non-convex and convex stochastic settings. The ... | Rebuttal 1:
Rebuttal: Thanks for your insightful review.
**Q1: The convergence rate of AGD depends on the hyperparameters δ and β1. If these hyperparameters are not chosen carefully, AGD may not converge to the optimal solution. Looking forward to seeing more comprehensive robustness evaluation using different network... | Rebuttal 1:
Rebuttal: #### Global Response
1. **AGD performance on large models**
We conduct optimizer experiments on the GPT2 (124M) and GPT2-Large (774M) models using the OpenWebText dataset. Our code is based on https://github.com/karpathy/nanoGPT. We compare AGD with the default AdamW optimizer. The results a... | NeurIPS_2023_submissions_huggingface | 2,023 | Summary: This paper proposed AGD, a novel optimiser that can dynamically switch between an adaptive optimiser and SGD. To achieve this, an auto-switching function is introduced to change the diagonal elements of the preconditioning matrix based on the gradients of adjacent steps. Both theoretical analysis on the conver... | Rebuttal 1:
Rebuttal: Thanks for your constructive review.
**Q1: Optimisers can perform differently on large-scale and small-scale experiments. Although good PTB results are given in the paper, a larger-scale language modelling experiment may still be useful.**
Please refer to [Global Response 1](https://openreview.n... | null | null | null | null | null | null |
On the Importance of Feature Separability in Predicting Out-Of-Distribution Error | Accept (poster) | Summary: This paper proposes an easy-to-use dataset-level method for predicting OOD scores. The authors analyze two desiderata in representation learning: high inter-class dispersion and high intra-class compactness. Through some experiments on CIFAR and TinyImageNet, the authors reveal that the inter-class dispersion ... | Rebuttal 1:
Rebuttal: 1. **About the OOD setting in the experiments**
Thank you for pointing out the potential for misunderstanding. We would like to clarify that there are different definitions of "OOD" in various related areas. In OOD detection, OOD examples are those test samples with different label space from the... | Summary: In this paper, the authors focus on the task of predicting model performance on unseen/shifted datasets, without support of annotations. To do so, they first show the connection between feature separability and test accuracy, with an intuitive example and theoretical explanation. Based on the analysis, they pr... | Rebuttal 1:
Rebuttal: 1. **Improving quality of some figures.**
Thanks for your constructive suggestions. We will improve the quality of those figures in the revised version, as you suggested.
2. **The reason why intra-class compactness dose not work.**
Thank you for the thought-provoking question. Previous works in... | Summary: This research focuses on predicting test accuracy on shifted datasets without access to ground-truth labels. The authors begin by analyzing the potential issue of the existing methods which were based on the shift distance and point out that the previous distributional distances were not always correlated high... | Rebuttal 1:
Rebuttal: 1. **Can improving the quality of pseudo labels further enhance performance of OOD error estimation?**
To verify whether improving quality of pseudo labels can further improve the estimation performance, we conduct experiments on Tiny-ImageNet with ResNet18 by filtering out low-confidence predict... | Summary: This work aims to predict classifier accuracy on unlabeled test samples. To achieve this goal, this paper proposes a feature separability-based dataset-level score to check whether features have high inter-class scatter. This feature separability score is calculated by measuring how far the centroids of featur... | Rebuttal 1:
Rebuttal: Great thanks for your constructive comments and suggestions! Please find our response below.
1. **Improving Figure 1 by adding the two matrics.**
Thank you for the suggestion. In the final version, we will add the corresponding metrics (shown in table 1) on the figures. We list the numerical det... | Rebuttal 1:
Rebuttal: ## General Response
We thank all the reviewers for their time, patient, and valuable comments. We are glad that all reviewers agree that our work is *well-motivated and well-organized, clear, and concise*. Reviewers HxyT, BweV, and ALm9 appreciate that *the experimental results are strong and con... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Image Captioners Are Scalable Vision Learners Too | Accept (oral) | Summary: This paper shows that training a ViT with the image-to-text generation (i.e., image captioning) objective is an effective way to learn good visual representations for downstream tasks. The paper systematically compares between the image captioning and the image-text contrastive learning (CLIP) objectives, wher... | Rebuttal 1:
Rebuttal: **Minor limitation: Frozen adaption may not justify the advantage of Cap-ViT**
We actually tried fine-tuning the image encoder along with the decoder for both CLIP* and Cap/CapPa models and did not obtain an improvement for any of the models. This is consistent with prior work which did not obser... | Summary: This paper shows that the captioning loss is a competitive alternative to pretrain image backbones compared with contrastive loss (CLIP) when using the same training budget and training data. The standard next token prediction objective on the full caption sequence (Cap) is complemented with a parallel predict... | Rebuttal 1:
Rebuttal: **The layout of the tables and figures is hard to follow**
Thank you for bringing this to our attention. We will improve the alignment of tables and figures with the text flow in the next revision.
\
**How do the compute requirements (e.g. gpu memory) of the CLIP\* with 8k/16k batch size compare... | Summary: This paper revisits image captioning as a pretraining task for learning general vision encoders from web image-text pairs. Surprisingly, the empirical study shows that captioning pre-trained vision encoders is competitive or better than contrastively pre-trained ones on image recognition and vision-language ta... | Rebuttal 1:
Rebuttal: **CLIP-style models are convenient for image-text retrieval/computing image-text similarities and offer a more efficient computation.**
This is indeed a downside of captioning-based approaches. On the upside, Cap and CapPa are much better than CLIP-style models in zero-shot classification tasks w... | Summary: In this paper, the authors more a controlled comparison between two pretraining approaches to learning visual representations from language supervision:
contrastive (CLIP-style) and generative (= image captioning, e.g. VirTex, SimVLM, BLIP, CoCa), etc.
The goal of this paper is not to advance the state-of-the-... | Rebuttal 1:
Rebuttal: Thank you for suggesting additional experiments, many of which we were able to address.
\
\
**Performance of the exact autoregressive language model used in CLIP paper**
We trained this exact model (with a ResNet-50 encoder as in the CLIP paper) for 900M examples seen and found that it performs s... | Rebuttal 1:
Rebuttal: We thank all the reviewers for carefully reading our paper. We appreciate their thoughtful comments and the overall very favorable assessment. In particular, we liked the many suggestions for additional experiments, some of which we were able to run. We believe these experiments will make the pape... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
On Convergence of Polynomial Approximations to the Gaussian Mixture Entropy | Accept (poster) | Summary: This paper presents new methods for approximating the entropy of a mixture of Gaussians/cross-entropy between mixtures of Gaussians. By using deterministic approximations based on power series/orthogonal polynomials expansions, the authors are able to obtain similar results as MCMC approximations for only a fr... | Rebuttal 1:
Rebuttal: **line 183 : what do you mean by convergence rate of $\beta \alpha^n + \eta$ ?**
From Taylor's theorem, a function, $f$, can be represented as $f=T_n+R_n$ where $T_n$ is the $n^{th}$ order Taylor series and $R_n=\mathcal{O}(\alpha^n)$ is the remainder. Rewriting this function in terms of the Tayl... | Summary: The paper provides an improved approximation to the entropy of Gaussian Mixture Models (GMMs). The proposed approximation is a lower bound to the entropy and is more accurate compared to related methods.
Strengths: 1. Accurate computation of the GMM entropy has implications on variational inference (VI) as we... | Rebuttal 1:
Rebuttal: **While the proposed approximation establishes a lower bound for the GMM entropy, it is not clear whether the bound is tight. Thus, applying the approximation to VI may not yield monotonically increasing ELBO**
The Taylor series bound on GMM entropy becomes tight as the order of the approximation... | Summary: This paper looks at the convergence of an existing approximation of the GMM entropy. They show conditions, which commonly occur, in which the approximation diverges. The authors then present several new approximations in which conditions can be chosen that guarantee convergence. The authors demonstrate this bo... | Rebuttal 1:
Rebuttal: **I have reviewed a previous version of this paper**
We sincerely thank the reviewer for their effort and helpful insights on an earlier submission. The early feedback has strengthened the present work.
**The computational complexity is a concern for this method**
For similar polynomial orders... | Summary: The authors show that a common entropy approximator, based on Taylor series approximators, for Gaussian mixture models is not necessarily convergent. They propose an alternative approach based on the Taylor expansion of the logarithm, whose efficiency hinges on the fact that moments of Gaussians can be comput... | Rebuttal 1:
Rebuttal: **"Many of the lemmas and theorems are known results..."**
Key results of this paper are divergence of the Huber et al. approximation (Thm. 3.1) along with our Taylor and Legendre approximations and their convergence (Thm. 4.2 & 4.5). To the best of our knowledge these results have not previousl... | Rebuttal 1:
Rebuttal: We thank all reviewers for their time and useful insights. We have provided individual responses to each reviewer below. For brevity we do not address minor typos, all of which will be corrected in a final version. | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
One-2-3-45: Any Single Image to 3D Mesh in 45 Seconds without Per-Shape Optimization | Accept (poster) | Summary: This paper introduces a novel approach for converting a single image into a 3D textured mesh without the necessity for per-shape optimization. The method employs Zero-123 for generating novel views of the input image, which is then utilized by a generalizable MVSNeRF to reconstruct the 3D shape. To tackle issu... | Rebuttal 1:
Rebuttal: Thank you for your insightful comments and valuable suggestions. We will revise our paper based on your feedback. Here are our responses to your comments:
**Weighted sum of projected colors**
Thanks for pointing this out. We agree that computing the point colors as a weighted sum of projected co... | Summary: The authors tackle the problem of 3D reconstruction from a single image, which is challenging due to the lack of 3D information. They propose a novel method that combines a 2D diffusion model, Zero123, with a cost-volume-based 3D reconstruction technique, SparseNeuS, to generate a 360-degree 3D textured mesh i... | Rebuttal 1:
Rebuttal: Thank you for your insightful comments and valuable suggestions. We will revise our paper based on your feedback. Here are our responses to your comments:
**Number of shapes for evaluation**
We first would like to clarify that we used 40 shapes (20 from GSO and 20 from Objaverse, excluding the L... | Summary: The proposed method overcomes the challenges of lengthy optimization time, 3D inconsistency, and poor geometry that are common in existing methods. This method uses a view-conditioned 2D diffusion model, Zero123, to generate multi-view images from a single input image, and then lifts these images to 3D space. ... | Rebuttal 1:
Rebuttal: Thank you for your insightful comments and valuable suggestions. We will revise our paper based on your feedback. Here are our responses to your comments:
**Natural extension or improvement of Zero123?**
We would like to emphasize our contribution in proposing a new paradigm for generalizable si... | Summary: This paper tackles the task of single image 3D reconstruction, where given a single input image of an object, the proposed method aims to generates a 360 degree 3D textured mesh.
Compared to prior works tackling a similar problem setting, this paper primarily aims to improve the quality of the reconstruction ... | Rebuttal 1:
Rebuttal: Thank you for your insightful comments and valuable suggestions. We will revise our paper based on your feedback. Here are our responses to your comments:
**How do errors in one module affect the final performance?**
Correct segmentation mask and pose estimation are important for the final gener... | Rebuttal 1:
Rebuttal: Dear Reviewers,
Thank you for dedicating your time to review our paper and offering insightful feedback. We sincerely appreciate your efforts to help enhance the quality of our research. We are also pleased to note that all reviewers were supportive of our work:
(a) Recognize our contribution in... | NeurIPS_2023_submissions_huggingface | 2,023 | Summary: This paper introduces an efficient method for single-image 3D reconstruction that significantly improves upon previous techniques by generating textured mesh in one feed-forward pass. The authors employ a view-conditioned 2D diffusion model, Zero123, to create multi-view images from a single input view and the... | Rebuttal 1:
Rebuttal: Thank you for your insightful comments and valuable suggestions. We will revise our paper based on your feedback. Here are our responses to your comments:
**Perform outside 1,156 categories?**
Yes, our method can generate meshes for images beyond the scope of the 1,156 training categories, as s... | null | null | null | null | null | null |
Generalized Semi-Supervised Learning via Self-Supervised Feature Adaptation | Accept (poster) | Summary: The authors claim to have addressed the new problem of semi-supervised learning with feature distribution mismatch (FDM-SSL) and to have proposed a new method called Self-Supervised Feature Adaptation (SSFA) for this problem.
The proposed SSFA consists of semi-supervised learning module and a feature adaptati... | Rebuttal 1:
Rebuttal: Thank you for your comments. We now answer them point by point.
**W A. Problem Novelty: A discussion of several closely related problems ([a], [b] and [c]) is lacking.**
Our proposed FDM-SSL problem setting is different from the previous works, we summarize the similarities and differences as fo... | Summary: The paper challenges the traditional assumption of semi-supervised learning: the feature distributions of labeled and unlabeled data are consistent. It claims that this assumption rarely holds in realistic scenarios, as: (a) unlabeled samples could contain various corruptions; (b) unlabeled samples could conta... | Rebuttal 1:
Rebuttal: Thanks for your useful comments! We see that your main concerns are novelty, experimental details and performance. We will address your concerns.
**W1. The proposed method has limited novelty. Incorporating self-supervised learning into semi-supervised learning, or leveraging self-supervised lea... | Summary: In this work, the authors propose a generalized Self-Supervised Feature Adaptation (SSFA) framework for FDM-SSL (Feature Distribution Mismatched Semi-Supervised Learning) that does not require prior knowledge of the distribution of unlabeled data. The SSFA framework aims to address distribution mismatch by dec... | Rebuttal 1:
Rebuttal: Thanks for your valuable comments! Below are the responses to your comments.
**W1. Lemma 1 is fine but convexity is a very strong assumption, which does not hold true in real-world deep loss functions. Some empirical evidence is provided, but it is somewhat counterintuitive to think that minimizi... | Summary: The authors propose an interesting and novel semi-supervised learning method with auxiliary self-supervised feature adaptation for addressing issues with heterogeneity in labelled and unlabelled data. In fact, the authors focus on heterogeneity across domains, which is not something traditionally considered in... | Rebuttal 1:
Rebuttal: Thank you for your constructive comments! We address all of your concerns point by point as below.
**W1. However, the caveat is that the baseline methods cannot do both SSL and UDA, which makes comparisons of performance metrics harder. However, SSFA does seem to be a novel method bridging the tw... | Rebuttal 1:
Rebuttal: We are grateful to all reviewers and ACs for the generous effort they have invested in reviewing our work! Some important or common questions are answered.
**Q1. The novelty of the method may be limited.**
Thanks for your comments.
According to Occam's Razor, *Entities should not be multiple un... | NeurIPS_2023_submissions_huggingface | 2,023 | Summary: This paper focuses on Feature Distribution Mismatch SSL (FDM-SSL), which considers a mismatch between the labeled and unlabeled data. In this task setting, the challenges include the scarcity of labeled data and the mixed distributions of the unlabeled data. The paper proposes a framework named Self-Supervised... | Rebuttal 1:
Rebuttal: Thank you for your valuable comments! We see that your main concerns are novelty, more explanations and some details. We answer your questions point by point.
**W1. I acknowledge that this work presents some novelties in proposing and addressing the challenges in domain-shifted SSL. However, the ... | null | null | null | null | null | null |
Keep Various Trajectories: Promoting Exploration of Ensemble Policies in Continuous Control | Accept (poster) | Summary: This paper addresses the problem of exploration in ensemble-based reinforcement learning. It introduces a novel approach called Trajectory Aware Ensemble Exploration (TEEN) that aims to enhance exploration by increasing diversity among the sub-policies. This diversity is measured by the KL divergence between t... | Rebuttal 1:
Rebuttal: We thank you for your detailed feedback and insightful suggestions, please see the following for our response.
> Q1: There appears to be an inconsistency between Line 18 in Algorithm 1 and the text. Specifically, Line 18 suggests that all sub-policies are updated sequentially, while the "recurren... | Summary: This paper presents a new ensemble RL algorithm. The main motivation of this work is to increase the variation between sub-policies, which is measured by the distance between the corresponding state-action visitation distribution. It formulates the loss function by connecting the visitation distribution distan... | Rebuttal 1:
Rebuttal: We would like to extend our sincere gratitude for your insightful feedback on our submission. Your valuable comments have greatly contributed to enhancing the quality of our work. In response to your suggestions, we have carefully revised the manuscript to address the points you raised.
> Q1: Is ... | Summary: This paper presents an algorithm called trajectories-aware ensemble exploration (TEEN) for sample efficient ensemble reinforcement learning. The authors point out that the existing ensemble methods do not effectively address the required diversity in exploration, which TEEN is designed to tackle better. It ach... | Rebuttal 1:
Rebuttal: We would like to express our heartfelt appreciation for the invaluable feedback you provided on our submission. Your insightful comments and suggestions have been instrumental in refining our research. We clarify your concerns as follows.
> W1: First, in RQ1 in evaluating the performance against ... | Summary: The paper presents an efficient ensemble learning strategy termed Trajectories-awarE Ensemble exploratioN (TEEN), which promotes exploration by diversifying the state-action visitation distribution of multiple sub-policies. The proposed method introduces a discrepancy measure that tries to maximize the differe... | Rebuttal 1:
Rebuttal: We thank you for the your detailed feedback and insightful suggestions, please see the following for our response.
> W1: There might be states not captured by the state-action visitation of one particular sub-policy(say $\pi_{a}$). In that case, the contribution of $\rho^{\pi_{a}}(s,a)$ to $\rho(... | Rebuttal 1:
Rebuttal: We greatly appreciate the reviewer's thorough reading and detailed comments about our submission. Your constructive reviews strengthen our draft. And some major concerns might be due to potential misinterpretations of the paper and we would like to clarify them in the responses.
We updated Appendi... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Effective Bayesian Heteroscedastic Regression with Deep Neural Networks | Accept (poster) | Summary: The authors consider the task of properly modeling heteroscedastic observation noise with Bayesian Neural Networks (BNNs).
They identify several shortcomings in prior work concerning variance modeling and propose to rely on parameterization
via natural gradients instead. Together with a Laplace approximation t... | Rebuttal 1:
Rebuttal: Thank you for the overall positive evaluation of our paper and your detailed and constructive comments and questions!
> Section 5.2 reads more like an afterthought that is added to the paper to have an additional experiment.
Thank you for your comment. This is mostly due to lack of space. We wil... | Summary: This paper extends the technique of the Laplace Approximation to incorporate heteroscedastic aleatoric uncertainty. The proposed technique achieves this by exploiting the natural parameters of the Gaussian likelihood. Main motivation is the coupling of the mean and the input dependent variance in the Gaussian ... | Rebuttal 1:
Rebuttal: Thank you for your detailed feedback and constructive questions which will help inprove our manuscript.
> The paper mainly focuses on the regression problems. [...] Would it be possible to include the term "regression" in the title?
Thank you for pointing out that heteroscedastic does not imply ... | Summary: This article focuses on refining the techniques used to manage two types of uncertainties in complex regression tasks utilizing deep neural networks. The uncertainties, known as aleatoric (arising from inherent randomness in the data) and epistemic (originating from the model's limitations), are critical to ad... | Rebuttal 1:
Rebuttal: Thank you for your positive remarks! Below we would like to clarify your concerns regarding computational cost of our method.
> How large is the computation cost and time of the proposed method in comparison to the rival methods?
1. Using the natural parameterization of the likelihood does not c... | Summary: This paper introduces a new method for fitting heteroscedastic regression models with neural networks. In contrast to previous work that highlights the deficiencies of MLE and indirectly hint at regularization, this paper suggests using natural gradients during training and combines Bayesian ideas for regulari... | Rebuttal 1:
Rebuttal: Thank you for your positive evaluation of our work and your constructive comments and questions.
**Weaknesses**
> It was odd to see Problem 2.1 presented so high up in the paper.
Currently, Section 2 serves as a theoretical comparison to the most related works. Within this context, we introduce... | Rebuttal 1:
Rebuttal: We thank the reviewers for their time and constructive feedback on our manuscript. It is encouraging to see that the reviewers rated our paper overall positively pointing out that our paper *"identifies prior weaknesses, and offers a principled solution"* (Reviewer 7fHN), that Reviewer HwYi *"appr... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Combinatorial Group Testing with Selfish Agents | Accept (poster) | Summary: The paper studies Combinatorial Group Testing (CGT) problem in a game-theoretic setting. The setup consists of a set $[n]$ of $n$ agents, of which a set of $k$ agents, denoted by $K$ are `active'. The goal is to reveal their identities in the minimum time possible via set of queries. Each query $Q$ is a subset... | Rebuttal 1:
Rebuttal: **Weakness 1**: *The paper … builds up on ideas of Chionas et al. [2023]. It also seems like extension of their setups, which is not entirely bad thing; however, it limits the extent of contributions. The main algorithmic contribution of the paper is basically the BS_Jumps algorithm.*
Reply:
... | Summary: This submission applies and extends a very recent game theoretic perspective, introducing in particular a notion of strategy equilibrium called adversarial equilibrium, on contention resolution games to the more general setting of combinatorial group testing. In combinatorial group testing, a (typically small)... | Rebuttal 1:
Rebuttal: We thank the reviewer for valuable feedback. Below we give our answers. If the reviewer believes that we have addressed adequately all of her/his concerns, we would very much appreciate if she/he reconsiders upgrading the score for our paper.
**Question 1:**
*To me the motivation of requiring adv... | Summary: This paper introduces a game theoretic framework in the context of combinatorial group testing. In detail, in a population of $n$ elements, there are $k<n$ agents that are considered active (e.g. they are positive to some disease) and the goal is to use queries on subsets of the $n$ elements to reveal who tho... | Rebuttal 1:
Rebuttal: We thank the reviewer for valuable feedback. Below we give our answers. If the reviewer believes that we have addressed adequately all of her/his concerns, we would very much appreciate if she/he reconsiders upgrading the score for our paper.
**Weakness:**
*The main weakness is the fact that the ... | Summary: In this paper, the authors study the problem of combinatorial group testing in a new framework, when the elements to be found are selfish players. The authors propose an algorithm such that no agent is willing to deviate from the algorithm strategy (being in this way in an adversarial equilibrium).
Strengths... | Rebuttal 1:
Rebuttal: We thank the reviewer for valuable feedback. Below we give our answers. If the reviewer believes that we have addressed adequately all of her/his concerns, we would very much appreciate if she/he reconsiders upgrading the score for our paper.
**Question 1:**
*It is not very clear to me which is t... | Rebuttal 1:
Rebuttal: We would like to thank the reviewers for their effort and valuable feedback. Below we address the main two requests for providing more applications and describing technical challenges comparing to Chionas et al. IJCAI 2023. Other questions we address as individual rebuttals to the reviewers.
**Mo... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Maximize to Explore: One Objective Function Fusing Estimation, Planning, and Exploration | Accept (spotlight) | Summary: This paper studied reinforcement learning with general function approximation setting. The authors proposed a new MEX framework, which unifies the exploration and exploitation within a unconstrained optimization objective. Under the structural assumptions about low GEC, they established regret upper bound for ... | Rebuttal 1:
Rebuttal: **Weakness 1: For the model-free setting, the objective (3.1) may not be as easy to optimize as it looks like. By (3.3), the definition of L_h^k, solving (3.1) requires to solving a mini-max optimization problem. On the other hand, those constrained optimization objectives in previous literature (... | Summary: This theory paper proposes an algorithmic framework where the "hypothesis" (model for MB methods, Q function for MF methods) for each iteration is chosen by maximizing one objective: the hypothesis likelihood (i.e., NLL for MB methods, TD error for MF methods) plus the expected returns (value function at the i... | Rebuttal 1:
Rebuttal: **Originality: Compare MEX with [1] on Gridworld environments.**
Based on [3], we compare MEX on Gridworld with [1] in Figure 2 of **General Response**.
**Clarity: Comparison to related works and organization of Section 5.**
Please see **General Response Q3** for a brief comparison. We will add ... | Summary: The paper presents a unified objective for optimizing regret in online RL algorithms with theoretical justification and experimental evaluation. The objective crucially dos not dependent on posterior estimation or bi-level optimization, making it easier to implement in practice than previous theory motivated a... | Rebuttal 1:
Rebuttal: **Weakness 1: The presentation is necessarily dense, making it hard to grasp a couple of intuitive concepts.**
Thanks for the constructive comments! We address them in the following.
- *Function approximations:* we will present the concrete realizations of the general function approximations e... | Summary: This paper proposes an RL framework, maximize to explore (MEX), that requires solving an unconstrained optimization which encompasses estimation, planning, and exploration. It proposes both model-free and model-based algorithms and further extends this framework to zero-sum Markov game setting. All these algor... | Rebuttal 1:
Rebuttal:
**Weakness 1: Lacking discussions of previous research that have similar ideas.**
Thanks very much for pointing these related works out! We will add these works and discuss them in detail in the revision. In the following, we compare our work with these works briefly.
[1]: While this work first... | Rebuttal 1:
Rebuttal: **General Response:**
Thank each reviewer for the review. We provide comments on some common questions here.
**Q1: Explanations on the Bellman completeness assumption (Assumption 3.1) and the choice of loss function (3.3), especially the second term, for the model-free case.**
**A1:** **About t... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Koopa: Learning Non-stationary Time Series Dynamics with Koopman Predictors | Accept (poster) | Summary: This paper tackles the task of non-stationary time series forecasting with Koopman operator. Specifically, the authors propose Koopa Blocks, which uses Fourier transform to disentangle the time series into time-variant and time-invariant components, and leverages Koopman
operator to model their dynamics respe... | Rebuttal 1:
Rebuttal:
Many thanks to Reviewer pXJ3 for providing a detailed review and insightful questions.
**Q1:** About the performance of the proposed method.
To our best knowledge, the benchmark of time series forecasting has been excavated for a long time. And PatchTST as the SOTA forecasting model has surpas... | Summary: This work propose a new architecture inspired by Koopman theory for time-series forecasting task. To address the non-stationarity problem, this work follow the idea of disentangle local and global time-series representation and forecasting model respectively. Particularly, this work propose to use Fourier filt... | Rebuttal 1:
Rebuttal: Many thanks to Reviewer EGYL for the detailed and insightful review.
**Q1:** Technique novelty of proposed method.
As clarified in $\underline{\text{Q5 of the reviewer M3Xz}}$, we sum up "how existing work of Koopman theory tackles non-stationary problems" and highlight our difference with previ... | Summary: The authors address the issue that most real-world time-series are non-stationary. To tackle this problem they introduce Koopa, a globally learned (time-invariant) and localized (time-variant) linear Koopman operators, to exploit respective dynamics underlying different components. In practice, Koopa is compos... | Rebuttal 1:
Rebuttal: Many thanks to Reviewer a2Yj for providing thorough detailed comments.
**Q1:** Hyperparameter sensitivity of the segment length $S$.
As per your suggestion, we have checked the robustness of the proposed method with respect to the hyperparameter $S$, which varies in $\{H/8,H/4,H/2,H\}$, where th... | Summary: This paper addresses non-stationary time series forecasting, in which case the temporal distribution of time series is changing over time. Koopman theory is introduced in this paper for its fundamental capacity on modeling dynamical systems and a Koopman forecaster called Koopa is proposed. Concretely, in this... | Rebuttal 1:
Rebuttal: Many thanks to Reviewer M3Xz for providing a detailed review.
**Q1:** Novelty of disentangling module.
We'd like to emphasize our proposed method differs from previous works as follows:
- It considers time-variant and invariant disentanglement, which is essential for Koopman forecasters, and fe... | Rebuttal 1:
Rebuttal: ## Global Response to All Reviewers
We sincerely thank all the reviewers for their insightful reviews and valuable comments, which are instructive for us to improve our paper further.
To cope with non-stationary time series forecasting, this paper proposes Koopman-based forecasters (Koopa) to p... | NeurIPS_2023_submissions_huggingface | 2,023 | Summary: The paper studied time series forecasting problem for weather forecasting, energy consumption, and financial assessment. To propose a model that generalizes on varying distribution, the authors proposed Koopa which is composed of modular Koopman predictors to describe and advance forward series dynamics. The m... | Rebuttal 1:
Rebuttal: Many thanks to Reviewer 36zV for providing thorough insightful comments.
**Q1:** The convergence of the proposed method and compared methods.
As shown in $\underline{\text{Figure 1 of the global response}}$, we provide the model training curves to check the convergence of our proposed model and... | null | null | null | null | null | null |
Oracle Complexity of Single-Loop Switching Subgradient Methods for Non-Smooth Weakly Convex Functional Constrained Optimization | Accept (poster) | Summary: This paper studies switching subgradient method for solving weakly-convex objective (weakly-)convex constrained problems. The oracle complexity for finding a nearly stationary point is derived for both convex constraint case and weakly-convex constrained (with good initialization). Numerical experiments valida... | Rebuttal 1:
Rebuttal: Thank you so much for reviewing our paper and providing valuable feedback that helps us further improve our work!
**Question 1: For the weakly-convex constrained case, if we cannot have feasible initial point, does the algorithm SSG converge? If so, what kind of point will it arrive? As we know,... | Summary: The paper considers a weakly-convex constrained optimization problem and establishes the oracle complexity of the switching subgradient method for finding a nearly stationary point.
Strengths: 1. Considering a single-loop algorithm, the paper can establish the complexity $O(1/\epsilon^4)$ for a case of weakly... | Rebuttal 1:
Rebuttal: Thank you so much for reviewing our paper and providing valuable feedback that helps us further improve our work!
**Comment: In the numerical experiments, the paper compares a single-loop method with double-loop methods but reports the evolution of objective functions over iterations. This is not... | Summary: This paper examines the oracle complexity of the switching subgradient method used for solving non-convex constrained optimization problems. The target functions are weakly convex, while the constraint functions are either convex or weakly convex. Remarkably, this method matches the complexity of double-loop m... | Rebuttal 1:
Rebuttal: Thank you so much for reviewing our paper and providing valuable feedback that helps us further improve our work!
**Comment: ... there seems to be an overemphasis on marginally relevant references. The broad coverage of optimization beyond weakly convex settings, although interesting, does not di... | null | null | Rebuttal 1:
Rebuttal: We would like to thank all reviewers for reviewing our submission and providing valuable feedbacks. We have provided answers to each reviewer's questions separately. Reviewer tsNX suggested us to report the evolutions of the curves over time instead of over iterations. We have included the new fig... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Tuning Multi-mode Token-level Prompt Alignment across Modalities | Accept (poster) | Summary: The paper aims at improving few-shot classification based on CLIP-ViT-B/16 using prompt tuning. The authors introduced a multi-modal prompt tuning framework with token-level alignment and distribution matching. The prompts are tuned on ImageNet and then the model is benchmarked on 15 image datasets.
Strengths... | Rebuttal 1:
Rebuttal: We thank reviewer dk2b for the comments and suggestions. Below, we address the concerns raised in your review point by point. Please let us know if you have any further concerns or whether this adequately addresses all the issues that you raised with the paper.
> The writing appears rushed and so... | Summary: This work aims to overcome the limitations of previous works in prompt tuning for vision-language models. Unlike prior approaches that focus on single modality or holistic prompt alignment, the paper proposes a multi-mode token-level tuning framework that leverages optimal transportation to align prompt tokens... | Rebuttal 1:
Rebuttal: We thank reviewer Uuq5 for providing positive feedback and helpful suggestions. Below are our responses.
> Lacks discussion of computing cost and Q1: How about the comparison of computing cost between your method and the baselines?
We have reported the detailed comparison in the global comment ... | Summary: This paper introduces a multi-mode token-level alignment framework for multi-modal prompt tuning, which improves the representation of visual and textual modalities and can be used to improve existing methods. The task is formulated as a distribution matching problem, addressed using prompt and token-level opt... | Rebuttal 1:
Rebuttal: We thank reviewer hEEC for providing positive feedback and helpful suggestions. Below are our responses.
> Analysis of the additional time and memory costs has not been provided.
We have reported the detailed comparison in the global comment section. Please check for more discussion.
> Addition... | Summary: The paper proposes a prompt-learning method for adapting CLIP to few-shot classification. The proposed prompt is multimodal, i.e., both vision and language encoder is adaptable, and multimode, i.e., each modality is assigned with several prompts for diverse representation. Experiments are conducted on few-shot... | Rebuttal 1:
Rebuttal: We thank reviewer S87N for the comments and suggestions. Below, we address the concerns raised in your review point by point. Please let us know if you have any further concerns or whether this adequately addresses all the issues that you raised with the paper.
> Incremental novelty of the paper
... | Rebuttal 1:
Rebuttal: We thank all the reviewers for the time and expertise they have invested in these reviews and for their valuable comments. We are encouraged that all reviewers noted the impressive results achieved by the paper, and reviewers TnsF, hEEC, and Uuq5 praised the novelty and effectiveness of our method... | NeurIPS_2023_submissions_huggingface | 2,023 | Summary: This paper argues that existing prompt turning fails to caption the sample diversity, learning to sub-optimal prompt discovery. To this end, they propose a multi-mode token-level alignment for multi-modal prompt tuning. Specifically, they formulate the prompt turning as the hierarchical optimal transportation ... | Rebuttal 1:
Rebuttal: We thank reviewer TnsF for providing positive feedback and helpful suggestions. Below please find a response.
> Time Complexity Analysis
We have provided a detailed experimental comparison at the global comment. Please refer to the above analysis for more details.
>More Cases for Visualization... | null | null | null | null | null | null |
Accelerating Molecular Graph Neural Networks via Knowledge Distillation | Accept (poster) | Summary: * This paper explores knowledge distillation (KD) on speeding up molecular GNNs, which has challenges of being regression instead of classification, and having both scalar and vector outputs
* They tried a few different methods on a few different teacher-student combinations, with the best results closing 65% ... | Rebuttal 1:
Rebuttal: We want to thank the reviewer for their comments and feedback! We are extremely happy to hear they appreciate the importance of our work, the unique challenges we address in the paper, and the quality of the presentation. We also appreciate your insightful comments. We carefully respond to your co... | Summary: This paper aims to improve the performance of resource-efficient GNNs for molecular simulation, an area where top models are becoming increasingly larger and more cumbersome.
Several knowledge distillation approaches are proposed with the aim of regressing the smaller student model's embeddings onto those of ... | Rebuttal 1:
Rebuttal: We are grateful for the detailed feedback! Happy to hear you appreciate the importance and novelty of our work, and the quality of our empirical results. We also appreciate your insightful comments. We carefully respond to your concerns below.
### Technical novelty of our work and contextualizat... | Summary: This paper investigates the use of knowledge distillation (KD) to accelerate molecular dynamics using graph neural networks (GNNs) and improve their predictive accuracy. They describe their experiments with various KD protocols such as node to node, edge to node, and vector to vector KD for four different mole... | Rebuttal 1:
Rebuttal: We thank the reviewer for their feedback and are glad to hear that they acknowledge the importance of the problem we are tackling! We also appreciate your insightful comments. We carefully respond to your concerns below.
### We have an additional benchmark dataset: COLL (W2+Q3)
We wish to highlig... | Summary: The main contribution of this work is the exploration of knowledge distillation as a means to enhance the performance and scalability of GNNs for molecular simulations. The authors introduce custom KD strategies, namely node-to-node, edge-to-node, and vector-to-vector
distillation, to overcome limitations of... | Rebuttal 1:
Rebuttal: We thank the reviewer for the concise, positive review! We are glad to hear that they appreciate the presentation of our work and acknowledge its importance!
### Applying KD to tensorial models
We also thank the reviewer for their suggestion to apply our methods to tensorial models like e3nn, but... | Rebuttal 1:
Rebuttal: We wish to take this opportunity to thank the reviewers for their time and effort in assessing our work. We are deeply grateful for the reviews we have received. We were extremely happy to hear that the reviewers recognized the importance of our work and the quality of our results and presentation... | NeurIPS_2023_submissions_huggingface | 2,023 | Summary: the paper proposes new knowledge distillation strategies to enhance the hidden representation in the molecular GNN, downstream regression task of energy and force predicition shows promising results
Strengths: 1. the paper is well written and easy to follow
2. the experiments show the knowledge distillation c... | Rebuttal 1:
Rebuttal: We thank the reviewer for their feedback. We are glad to hear that you like the presentation of our work and highlight the improvement in performance we achieve! We also appreciate your insightful comments. We carefully respond to your concerns below.
### We have an additional benchmark dataset: ... | null | null | null | null | null | null |
k-Median Clustering via Metric Embedding: Towards Better Initialization with Differential Privacy | Accept (poster) | Summary: This paper studies $k$-median problem in a discrete metric space. Here there is a metric space $(U, \rho)$ where $|U| = n$, and we are given a multiset $D$ of datapoints (which is a subset of $U$). The goal is to select $k$ centers $F$ from $U$ that minimizes the sum of the distances of points in $U$ to their ... | Rebuttal 1:
Rebuttal: Dear Reviewer MjrB
Thanks for your feedback.
Questions:
In your experiments, have you considered running the different DP algorithms using different number of iterations? In particular, is it possible that more iterations help for other (worse) iterations?
Answer: We note that, the major goal ... | Summary: The paper considers the problem of finding good initializations for k-median clustering algorithms with and without differential privacy. To this end, the authors propose NDP-HST and DP-HST, two methods based on Hierarchically Well-Separated Trees to find a good initialization of k points. The theoretical appr... | Rebuttal 1:
Rebuttal: Dear Reviewer JWq6
Thanks for your feedback.
Questions 1:
Why does Algorithm 4 guarantee differential privacy? From my understanding, $\hat{N_v}$ constitutes a count query that should be protected with the Laplace mechanism (or Geometric mechanism). I don't understand why $2^{L-h_v}$ is the sens... | Summary: The paper studies the problem of finding an initial solution for k-median with privacy, which
can be improved using local search methods. This is a standard approach, and privacy is well
motivated in this setting. The authors develop a new algorithm for initialization using
Hierarchically Well-Separated Tree (... | Rebuttal 1:
Rebuttal: Dear Reviewer vfbx,
Thanks for your feedback.
Question 1:
The differences between the problem definitions in (1) (2) are not completely clear. Shouldn't the k-median cost be defined with respect to all the data points? That should be same in the private and non-private versions. For the private ... | Summary: The submitted paper considers fast initializations for k-median. They proceed in a fairly natural manner:
(1) Do tree embedding.
(2) Run simple constant-factor approximation on the tree embedding.
A cool advantage with their simple algorithm on the tree embedding is that the authors show that they can make t... | Rebuttal 1:
Rebuttal: Dear Reviewer 5iV1,
Thanks for your feedback.
Q1. k-median++ (at least k-means++) initialization is known to be not always great empirically. That's why the authors already in the first paper proposed a greedy variant (which is the one that is the default in the scikit implementation). Why not c... | null | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Three Iterations of (d − 1)-WL Test Distinguish Non Isometric Clouds of d-dimensional Points | Accept (poster) | Summary: This paper studies the completeness of $l$-WL test for Euclidean point sets. It shows an algorithm that certifies two Euclidean point sets of dimension $d$ are isometric using $(d-1)$-WL test, where only three iterations suffice. The results extend to $d$-WL test which only requires one iteration.
The punchli... | Rebuttal 1:
Rebuttal: Reviewer's comment: "I think overall the paper is not hard to follow, but the presentation can be better
The contents on page 4 and 5 would be a lot easier to understand if there is a picture.
The algorithms deserve a box highlighting each step either in description or pseudocode, though I person... | Summary: The authors rigorously prove that applying the d-1 WL to the distance matrices of point clouds in d dimensions is complete: that is, two point clouds are related by an isometry (and relabeling of the points) if and only if they are not separated by the d-1 WL test
Strengths: The questions the authors address ... | Rebuttal 1:
Rebuttal: Reviewer's comment: "In lines 43-49 in the discussion of Hordan et al.: The authors of that paper discuss not only a 3-WL algorithm but also what the call in the abstract `a 2-WL-like algorithm', when d=3. To avoid confusion it may be helpful to address this claim in the discussion somehow.”
RESP... | Summary: The expressive power of GNNs has long been a central topic in the GNN community. WL test serves as a fundamental algorithm in guiding designing expressive GNNs. For each k-WL, it has been shown that there always exist non-isomorphic graphs that cannot be distinguished by k-WL, and thus k-WL is incomplete for a... | Rebuttal 1:
Rebuttal: Reviewer's comment: "This is mainly a theoretical paper, without any experimental evaluation. However, I understand that showing experimental result is not very necessary given this paper focuses on a fundamental theoretical problem."
RESPONSE: Indeed, as the referee points out, the aim of the pa... | Summary: The paper addresses the question of testing the existence of a one-to-one application between two point clouds such that the distances are preserved. This information can be used, for example, to build better structured graph neural network architectures to optimize their performance. The starting point is tha... | Rebuttal 1:
Rebuttal: Reviewer's comment: "the authors stressed the importance of their results in improving the design of neural net architecture. This is unfortunately not clarified and it should be nice if the authors could comment more on that."
RESPONSE: What we mean is simply that, when working with cloud poin... | Rebuttal 1:
Rebuttal: GENERAL RESPONSE: Many thanks for your comments and suggestions. Regarding presentation, we acknowledge that adding a figure to illustrate the main idea behind the proof of our main result, as well as giving a high-level description of the underlying algorithm in a framed box, will greatly improv... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
MultiFusion: Fusing Pre-Trained Models for Multi-Lingual, Multi-Modal Image Generation | Accept (poster) | Summary: This paper presents MultiFusion, a multilingual multimodal image generation model, which can be effectively trained by fusing existing pre-trained visual models, language models, and stable diffusion models. Using a multilingual autoregressive language model as a bridge, MultiFusion follows MAGMA to enable mul... | Rebuttal 1:
Rebuttal: We thank reviewer 2Kwy for their feedback and suggestions. Below we address each weakness and question separately. Tab. 5-7, as well as Fig. 11 and 12, can be found in the supplementary PDF of this review.
### W1) Architectural details
We thank the reviewer for the suggestions and incorporated thi... | Summary: In this paper, the authors present a novel approach to expressing complex concepts with arbitrarily interleaved multimodal and multilingual input. Their approach leverages pre-trained models and allows an efficient fusion of different component without training a model from scratch.
Strengths: 1. The paper is... | Rebuttal 1:
Rebuttal: We thank the reviewer for their feedback and subsequently explain the motivation of MultiFusion in more detail.
### W1) Motivation and comparison.
**Motivation**
There exist several motivations for using interleaved multimodal inputs. As demonstrated by the experiments in the paper, reference i... | Summary: In this paper, the authors introduce MultiFusion, a novel approach that enables the expression of complex and nuanced concepts in text-to-image diffusion models (DM) through arbitrarily interleaved inputs of multiple modalities and languages. The “fusion” concept is at the core of the whole work: to fuse modal... | Rebuttal 1:
Rebuttal: We thank the reviewer for their constructive feedback. Tab. 5-7 as well as Fig. 11 and 12 can be found in the supplementary PDF of the rebuttal.
We agree on the importance of design justification. We reiterate on design choice details for clarity in our global response under the section "Reiterati... | Summary: This work proposed a novel method to build multilingual multimodal generation models that supports prompts composed of interleaved text and image. It combines a strong pre-trained multilingual language model with the image generation model from Stable Diffusion (SD) and achieves alleviated capabilities such as... | Rebuttal 1:
Rebuttal: We thank the reviewer for their feedback and subsequently address each limitation and question separately. [Tab. 5-7, as well as Fig. 11 and 12, can be found in the supplementary PDF of this review]
### W1) Training details
We supply further information on the training data sizes, parameter coun... | Rebuttal 1:
Rebuttal: We thank all the reviewers for their detailed and helpful feedback.
We are encouraged that they found our solution for expressive image generation to be well-motivated (8eHc), novel (sjMA, g96C), and well-written (sjMA, 5v9N). Reviewers highlighted the efficient fusion of pre-existing models for s... | NeurIPS_2023_submissions_huggingface | 2,023 | Summary: This paper presents an approach for creating a model that can take interleaved sequences of images and multi-lingual text as input, and generate novel images as output, by fusing together pre-trained models: (1) a ResNet image encoder from CLIP (2) an encoder-decoder text Transformer LM and (3) a Stable Diffus... | Rebuttal 1:
Rebuttal: We thank the reviewer for their thorough review and insightful questions and will address them in the following. We strongly believe that the changes made during the review process have improved the quality of the paper. [Rebuttal Pdf contains Tab. 5-7 as well as Fig. 11 and 12]
### Q1) Further ... | null | null | null | null | null | null |
Probabilistic Inference in Reinforcement Learning Done Right | Accept (poster) | Summary: The authors propose a model-based RL algorithm for MDPs with unknown rewards and transition dynamics, which approximates the posterior probability of an action being optimal in a given state to ensure efficient exploration. The approach is mostly contrasted with model-free RL as inference, but there is also an... | Rebuttal 1:
Rebuttal: We thank the reviewer for their feedback. In this rebuttal we hope to convince you that our work is adequately positioned with respect to existing works on Bayesian model-based RL, and that its framing in the light of ‘RL as inference’ is relevant.
__Where does this paper stand with respect to e... | Summary: The paper views reinforcement learning as a Bayesian variational inference problem, and proposes VAPOR, an algorithm which computes an approximately optimal occupancy measure and its corresponding policy. Bayesian regret of VAPOR is analyzed to have a sub-linear bound. In numerical experiments, VAPOR shows goo... | Rebuttal 1:
Rebuttal: We thank the reviewer for their thorough review. When we encounter a reviewer that has clearly taken the time to read the paper in detail, as you have, and given the paper a low score, that is a signal to us that the paper is unclear in the highlighted areas — so thank you for bringing these issue... | Summary: The paper provides a Bayesian treatement for the reinforcement learning as probabilistic inference framework for the discrete state/action space. As extension to the standard formulation, posterior probabilities of state-action optimalities are regarded and formally defined. For tractability, an approximation ... | Rebuttal 1:
Rebuttal: We thank the reviewer for their valuable comments and suggestions (such as expanding on the future work and limitations), which we will incorporate in the revised version.
__No code.__ We will include more details on the implementation of both VAPOR and VAPOR-lite, such as detailed pseudo-code (i... | Summary: This paper undertakes a rigorous Bayesian treatment of the posterior probability of state-action optimality. It proposes a variational approach to approximate the state-action optimality. The proposed method involves a tractable convex optimization problem and is provably efficient. This paper also conducts ex... | Rebuttal 1:
Rebuttal: We thank the reviewer for their valuable comments and for highlighting connections to previous work, on which we will expand in the revised version. We answer their queries below:
__Discussion of O’Donoghue et al.__ Thanks for highlighting that O’Donoghue et al., 2020 argue that TS solves at each... | null | NeurIPS_2023_submissions_huggingface | 2,023 | Summary: The paper identifies "the posterior probability of state-action optimality" denoted $\mathbb{P}_{\Gamma^*}$ as a key object for inference and control and provides a variational optimization approach to estimate it. Clear presentation, insightful analysis, and experiments on GridWorld, DeepSea, and Atari are pr... | Rebuttal 1:
Rebuttal: Thank you for the complimentary review! We are delighted that you think our work is valuable. We welcome any further comments. | null | null | null | null | null | null |
Diversify Your Vision Datasets with Automatic Diffusion-based Augmentation | Accept (poster) | Summary: The paper proposes a generative data augmentation method, ALIA, which utilizes large image captioning and language models to summarize the domain description and employs language-guided image editing methods to create augmented training data. Experiment results show that ALIA outperforms recent data augmentati... | Rebuttal 1:
Rebuttal: We appreciate your feedback and agree that our method is dependent on the ability of the pretrained models we use. We address your other questions and concerns below.
**Datasets.** Our datasets consist of similar objects because we focused on the difficult setting of fine-grained classification.... | Summary: The paper introduces ALIA to generate domain descriptions of the dataset from captions of each image. These captions are then used to generate more data using Stable Diffusion. There is also a filtering process to remove corrupted images or those with minimal edits. ALIA improves performance over the baselines... | Rebuttal 1:
Rebuttal: Thank you for your insightful review. We hope that we have adequately addressed your concerns below:
**Data and Computational Efficiency.** We generate 2 edits per image per prompt in the training set for ALIA. Depending on the output of the LLM, the number of prompts vary from 4-10. Using the H... | Summary: The paper proposes a method to augment an existing vision dataset with samples that are likely reflective of task-relevant variations that are potentially missing in the same. To do this, the authors propose a method, ALIA (Automated Language-guided Image Augmentation), which utilizes off-the-shelf image-gener... | Rebuttal 1:
Rebuttal: Thank you for your thoughtful feedback and suggestions, we hope to have addressed each of your concerns with the following:
**More clarity around domain shift.** We agree that more clarity around what each domain shift is, in addition to quantitative proof that ALIA improves accuracy under this d... | Summary: This paper explores improving robustness to variations in domains with the use of modern large vision and language models. They propose ALIA (Automated Language-guided Image Augmentation) which generates automated descriptors of data domains and augments training data with language-guided image editing. They u... | Rebuttal 1:
Rebuttal: We appreciate your appreciation of our work! We agree that ALIA is not suitable for cases were there are unseen domains which are significantly different from the domains seen in training or cases where the class relevant features need to be significantly augmented (e.g. pose change). As we will d... | Rebuttal 1:
Rebuttal: We thank all reviewers for their detailed reviews. We are delighted that they found the paper easy to follow and saw value in our discussions surrounding failed edits and different data augmentation techniques.
As requested, we have run a few more experiments to ensure a comprehensive evaluation... | NeurIPS_2023_submissions_huggingface | 2,023 | Summary: This paper presents a data augmentation methods. It first leverages the advanced image captioning and large language models to generate prompts. Next, it uses diffusion-based models to generate more images. In addition, it trained a model to filter failure or similar cases during image generation. Extensive e... | Rebuttal 1:
Rebuttal: We appreciate your review and you bring up important questions on how our work differs from the several works that explore using diffusion generated data to improve "general" classification tasks. We aimed to address this in the introduction of our paper and we hope to provide more clarity below. ... | null | null | null | null | null | null |
Unbiased Compression Saves Communication in Distributed Optimization: When and How Much? | Accept (poster) | Summary: The authors consider the distributed convex optimization problem in the centralized setting. In this setting, the authors provide the new lower bounds on the total communication cost under the assumption that nodes send unbiased independent compressed vectors to a master. They also provide improved analysis of... | Rebuttal 1:
Rebuttal: We thank the reviewer for the positive comments and the valuable question.
As the reviewer accurately points out, the convergence analysis presented in our paper is contingent upon the condition on local smoothness constants. Under the difference of local and global smoothness, it remains uncert... | Summary: The authors study the extent to which unbiased compression can reduce the total communication cost of optimizing smooth convex functions. Additionally, they refine the analysis of ADIANA ([25]) to show that the bounds are near-optimal.
Strengths: Quantifying how much unbiased compression helps is an interesti... | Rebuttal 1:
Rebuttal: We thank the reviewer for the detailed comments. We have clarified all questions as clearly as possible, and will be glad to address any further comments or questions.
**1. Not surprising that unbiasedness cannot help without independence**
We respectfully disagree with this opinion. Lots of lit... | Summary: This paper investigates the communication saving of unbiased compression on distributed optimization. The main contributions are:
i) a communication lower bound of distributed optimization algorithms with (not necessarily independent) unbiased compression;
ii) a communication lower bound of distributed opti... | Rebuttal 1:
Rebuttal: We thank the reviewer for the detailed comments. All questions have been clarified as best as we can. We are glad to address any further comments or questions.
**1. Section 1 claims that quantization and sparsification are modeled as unbiased compressors in some papers. But they can be biased too... | Summary: The paper analyzes the communication cost of unbiased compression algorithms for distributed optimization and proves that unbiased compression alone cannot reduce communication cost since whatever is saved via compression is lost due to an increase in the number of communication rounds needed for convergence. ... | Rebuttal 1:
Rebuttal: Thank you for your positive feedback and valuable comments. Below are our answers to the concerns you proposed.
**1. I would recommend moving the content of Appendix A which provides intuition on why independence helps to the main pape. Algorithm 1 can be moved to the Appendix to make space, sinc... | Rebuttal 1:
Rebuttal: We thank all the reviewers for their careful review and valuable feedback. We have addressed each question raised by the reviewers in the separate rebuttals below and are glad to address any further concerns. As several matters have been brought up by multiple reviewers, we present a global respon... | NeurIPS_2023_submissions_huggingface | 2,023 | Summary: Several communication compression strategies have been proposed in the last few years to improve distributed optimization. The authors consider the tradeoffs between lowering the communication costs per round vs the number of communication rounds and understand how they affect total communication costs. They t... | Rebuttal 1:
Rebuttal: We thank the reviewer for the positive comments. All questions have been clarified below.
**1. The lower bounds are proved only for certain classes of algorithms.**
We agree that our lower bounds only apply to the class of certain algorithms described in Section 2.3. However, these settings are... | Summary: The paper explores the conditions of unbiased compression to reduce the communication cost in distributed optimization. Specifically, the paper first presents a theoretical formalization of the total communication cost (TCC) in distributed optimization. With this formulation, the paper proves that unbiased com... | Rebuttal 1:
Rebuttal: We thank the reviewer for the positive comments and valuable suggestions. The datasets used in our current manuscript follow the experimental settings in prior works (e.g. [25, 26]) in the literature. According to the reviewer's suggestion, we conducted additional experiments with a larger dataset... | Summary: The paper investigated the overall communication costs for federated learning algorithms. A lower bound on the per-round communication cost is presented, then an analysis of an existing algorithm, ADIANA, is provided to obtain an improved upper bound.
Strengths: The general ideas for the settings and the resu... | Rebuttal 1:
Rebuttal: We thank the reviewer for the detailed comments. All questions have been clarified as best as we can. We are glad to address any further comments or questions.
**1. There is an absence of several crucial details, or perhaps typos, which makes it difficult to concretely understand the results.**
... | null | null |
Dynamic Regret of Adversarial Linear Mixture MDPs | Accept (poster) | Summary: This paper studiess adversarial RL under episodic linear mixture MDPs with adversarial full-information reward and stationary unknown transition kernels. This paper first proposes a new algorithm to deal with the adversarial reward and then provides an upper bound of dynamic regret for the proposed algorithm. ... | Rebuttal 1:
Rebuttal: Thanks for your valuable comments. We will address your concerns and questions below.
---
In particular, we believe there are some misunderstandings about the distinction between the non-stationary *stochastic* MDPs (studied in other works) and the non-stationary *adversarial* MDPs (this paper)... | Summary: This paper explores the problem of reinforcement learning in non-homogeneous MDPs with adversarial full-information reward feedback and unknown transition kernels. The authors propose a new algorithm that has advantages in dynamic regret and does not require prior knowledge as input. The algorithm achieves opt... | Rebuttal 1:
Rebuttal: Thanks for your helpful review. We will answer your questions individually. If your concerns are properly addressed, please consider updating the score. Thanks!
---
First, the reviewer may have an important misunderstanding on the two different settings: (i) non-stationary *stochastic* MDPs (oth... | Summary: The paper "Dynamic Regret of Adversarial Linear Mixture MDPs" studies reinforcement learning in episodic non homogeneous MDPs with adversarial rewards in full information feed back setting. The motivation for this work is that In many applications, reward functions may be picked adversarially, and changing ove... | Rebuttal 1:
Rebuttal: We greatly appreciate the insightful and constructive comments from the reviewer. In the following, we will respond to each of the questions raised.
---
Q1: “This work considers full information setting, bandit setting is not considered, and much more complex.”
A1: It would indeed be very inter... | Summary: This study analyzes the problem of online learning episodic inhomogeneous MDPs with _full-information_ reward functions (changed over episodes $k$ and step $h$) and the unknown transition kernel (unchanged over episodes). The MDPs are linear mixture MDPs and the performance measure is _dynamic_ regret (against... | Rebuttal 1:
Rebuttal: Thanks for your insightful comments! Below we will address your questions.
---
Q1: “Can you provide some intuition of what the switching number of best base-learner, $S_T$, measures? Are there some regimes for this parameter that are interesting (and provide intuitive connections to existing st... | null | NeurIPS_2023_submissions_huggingface | 2,023 | Summary: The paper studies RL in the adversarial episodic inhomogeneous MDP setting with transition kernels. It should be noted that the paper considers full-information feedback & linear mixture MDPs (both are very strong assumptions). The authors propose an algorithm drawing inspiration from policy optimization and p... | Rebuttal 1:
Rebuttal: Thanks for your valuable comments and suggestions. We provide our response to each question below.
----
Q1: "Could the authors highlight what are the key innovations (either in the algorithms or in the analysis) that lead to this improvement?"
A1: The key innovations over the prior works (espec... | null | null | null | null | null | null |
Gradient-Free Kernel Stein Discrepancy | Accept (poster) | Summary: The paper provides a method to estimate the kernel Stein discrepancy without gradients.
Strengths: The paper provides a method to estimate the kernel Stein discrepancy without gradients.
Weaknesses: The paper is not clearly written for someone who is not familiar with the field. After reading, I'm still co... | Rebuttal 1:
Rebuttal: Thank you for taking a look at our manuscript.
Please see our global response to all Reviewers on the topic of how our manuscript has been presented.
> After reading, I'm still confused about whether someone else has published a similar method before.
Thank you for the opportunity to clarify th... | Summary: This paper explores the use of Stein discrepancies in scenarios where the score function of the target distribution is unavailable or computationally impractical to evaluate. The authors introduce a novel approach called gradient-free kernelized Stein discrepancy (GF-KSD), which leverages a Stein operator deve... | Rebuttal 1:
Rebuttal: Thank you for your kind comments and for your eloquent assessment of our manuscript. As the reviewer most familiar with the literature on KSD, we sincerely hope that you will champion our work for acceptance, in light of the rather mixed scores we have received.
> To further validate the applica... | Summary: This paper proposed a posterior approximation using a new Stein discrepancy, which does not require derivatives of the statistical model. For that purpose, the authors derived the new discrepancy, called gradient-free KSD, and studied its statistical and convergence behaviors theoretically.
Then the authors de... | Rebuttal 1:
Rebuttal: Thank you for taking a look at our manuscript.
Please see our global response to all Reviewers on the topic of how our manuscript has been presented.
> In Definition 2, sup is undefined
The symbol "$\sup$" is the supremum; our understanding is that we can assume familiarity with "$\sup$" for Ne... | Summary: The authors study the kernel Stein discrepancy based on a Stein operator, introduced previously in [Liu 2018], which does not require access to the gradient of the target density. As their main theoretical result, they prove that, under certain assumptions on the target distribution (and the auxiliary distribu... | Rebuttal 1:
Rebuttal: Thank you for your thoughtful feedback, especially at short notice, which is greatly appreciated.
> This all means that $p$ must not be heavy-tailed if the assumptions of Theorem 2 are to be satisfied.
This is completely correct - but we emphasise that the standard kernel Stein discrepancy (KSD)... | Rebuttal 1:
Rebuttal: __On the Presentation of our Manuscript__
Our manuscript is theoretical in nature and is written for readers who have some familiarity with the concept of kernel Stein discrepancy (KSD). This is a small but important subsection of the NeurIPS community, who we believe will be the readers most in... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Efficient Adversarial Attacks on Online Multi-agent Reinforcement Learning | Accept (poster) | Summary: This paper proposes a series of novel attacks against MARL under different assumptions. From the theoretical perspective, it then analyzes the efficacy of the proposed attacks.
Strengths: + The paper proposes a series of novel attacks against online MARL and provides a decent theoretical analysis for the effi... | Rebuttal 1:
Rebuttal: > Why the proposed assumptions that the attackers can freely alter the target agents' actions and rewards are practical and realistic?
In some applications of RL models, action decisions and reward signals may need to be transmitted over communication links. When data packets containing the rewar... | Summary: This paper presents a novel approach to adversarial attacks in Multi-agent Reinforcement Learning (MARL), offering significant insights into various attack settings. The authors explore different attack strategies, including reward poisoning, action poisoning, and a combination of both, across various setting... | Rebuttal 1:
Rebuttal: > The related work section of the paper does not seem to cover the literature comprehensively. It would be beneficial for the authors to provide a more thorough review of the existing literature, especially focusing on the works that have addressed similar problems in the past, how they did it (fo... | Summary: This paper investigates the impact of adversarial attacks on Multi-Agent Reinforcement Learning (MARL) models. The authors propose an attacker who can manipulate rewards and actions to guide the agents into a target policy or maximize cumulative rewards under a specific reward function. The paper presents an a... | Rebuttal 1:
Rebuttal: > "Limited comparison with existing work"
Thank you very much for your comment. We will add the following discussion to the revised paper.
Adversarial attacks on single agent RL have been studied in various settings [Behzadan and Munir, 2017, Huang and Zhu, 2019, Ma et al., 2019, Zhang et al., 2... | Summary: This paper studies adversarial attacks in online multi-agent RL, focusing on reward and action poisoning attacks. It provides a set of characterization results for three different attack modalities: white-box, gray-box and black-box attacks. The authors first discuss the limitation of action-only or reward-on... | Rebuttal 1:
Rebuttal: > "Specifically for MARL, the most important references seem to be covered. However, there are a couple of recent references that could be included in the list (see below). I encourage the authors to do so."
Thank you very much for bringing these very interesting papers to our attention. We will ... | Rebuttal 1:
Rebuttal: We thank the reviewers for their valuable feedback, suggestions, and time invested. Here, we answer the reviewers' common concerns.
**Related works.** Due to the page limit of the main paper, we do not provide a comprehensive comparison with prior research on adversarial attacks. We will add the ... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Generalized Logit Adjustment: Calibrating Fine-tuned Models by Removing Label Bias in Foundation Models | Accept (poster) | Summary: This paper addresses biases in foundation models, such as CLIP, which due to the imbalanced training datasets, are skewed towards frequent semantics. The authors propose a Generalized Logit Adjustment (GLA) method, an optimization-based approach for debiasing these models. Despite inherent challenges, the GLA ... | Rebuttal 1:
Rebuttal: **Q1** The validity and effectiveness of the GLA framework with NLP models.
**A1** We appreciate the reviewer for the comment. Applying this algorithm to other tasks and fields is also our future research direction. For now, we primarily focus is on enhancing the fine-tuning performance for disc... | Summary: This paper studies the label bias in Foundation Models, like CLIP. Specifically, authors model the estimation of the label bias as a constrained optimization problem, and propose a Generalized Logit Adjustment (GLA) method to debias foundation models. Experimental results demonstrate the effectiveness of propo... | Rebuttal 1:
Rebuttal: **Q1** Discussion with existing label estimation methods[1][2].
**A1** [1-2] and our GLA work under different scenarios and require different forms of input.
[1-2] face a problem where the test prior is unknown but training data are given. In the context of foundation model adaptation, [1-2] req... | Summary: This paper proposes a distribution estimation method that estimates the distribution of pre-training data for large VLMs like CLIP. Combined with downstream data distributions, a logit adjustment is applied to the outputs of both the zero-shot model and the fine-tuned model. By using a simple ensemble of the z... | Rebuttal 1:
Rebuttal: **Q1**The biggest weakness of this paper is the improvements are very limited.
\end{point}
**A1**We hold a different view. For ImageNet-1K full training with 1,000 classes, improving $1.5 \%$ accuracy over end-to-end fine-tuning is a noteworthy improvement. For ImageNet with 16 training shots, we... | Summary: Logit adjustment method has been widely used in long-tailed recognition, in which the source dataset is available, thus the class margin can be set in advance. However, the large-scaled foundational data cannot be accessed. To reduce the recognition bias of foundational models on downstream tasks, this paper p... | Rebuttal 1:
Rebuttal: **Q1** The method is simple, and the proofs look a little reluctant.
**A1** We hold a different view. **A simple method with statistical grounding is our strength rather than weakness**. As Reviewer *wcSK* claimed ``The proposed method is both simple to understand and implement. This means that ... | Rebuttal 1:
Rebuttal: # Response to All Reviewers
Dear Program Chair, Senior Area Chair, Area Chair, and Reviewers,
First of all, we gratefully thank all the reviewers for their thoughtful comments and feedback.
In this paper, we identify label bias in foundation models like CLIP and underscore its adverse effects o... | NeurIPS_2023_submissions_huggingface | 2,023 | Summary: This paper proposes a method for combining the predictions of zero-shot and few-shot classifiers. The method removes the need for a weighting hyper-parameter to interpolate between the predictions. Furthermore, it also removes a source of bias (due to the frequency of words in the pre-training dataset).
Str... | Rebuttal 1:
Rebuttal: **Q1** Additional experimental and theoretical analysis showing that $\log \mathbf{q}=\pi_p$
**A1**
To further address the concerns about our bias estimation, we further provide explanation of our Proposition 2 and some empirical evidences.
In Proposition 2, we proved that when given the outputs ... | null | null | null | null | null | null |
NeRF-IBVS: Visual Servo Based on NeRF for Visual Localization and Navigation | Accept (poster) | Summary: This paper addresses visual localisation (i.e. the estimation of the
6-degrees-of-freedom pose of a camera) from scene coordinate regression.
The problem tackled is the need for a large amount of pose and depth labels to
train scene coordinate regression networks.
The paper proposes a solution to reduce the ... | Rebuttal 1:
Rebuttal: Thank you for the constructive comments and suggestions.
**W1: Paper writing could be improved.**
Thanks for the valuable suggestion. We have polished the whole paper to make it more smooth and concise for readers. Please refer to **Q2** in the **"General Response"** for specific modification de... | Summary: The paper presents a novel visual localization method combined with the NeRF technique to address the issue of training a large number of pose images in existing visual localization models. The proposed method has two advantages: (1) the paper trains a coordinate regression network using a few posed images wit... | Rebuttal 1:
Rebuttal: Thank you for the constructive comments and suggestions.
**W1: The authors claim the proposed method is able to localize the camera with fewer posed images, can the authors give a rigorous description of it? For instance, how many posed images can be used, or can the authors show the reduced rati... | Summary: This paper proposes a visual localization pipeline that uses NeRF and image-based visual servoing. The main contribution is that the proposed method uses much fewer images with pose annotations and no 3D ground-truth labels (only uses pseudo ground-truth provided by NeRF to train a coordinate regression networ... | Rebuttal 1:
Rebuttal: Thank you for the constructive comments and suggestions.
**W1: The proposed method mentions that only renders a single image $C_r$, but how do you ensure that $C_r$ has visual overlap with the query? How often does this occur? Why not sample around the coarse pose T and generate multiple images?*... | Summary: This paper tackles the task of visual localization, i.e. estimating a camera pose from a query image. Approaches in the literature require a large number of posed images and even dense 3D supervision for some of them. Authors propose to leverage Neural Radiance Fields (NeRF) to solve the problem of visual loca... | Rebuttal 1:
Rebuttal: Thank you for the constructive comments and suggestions.
**W1 and Q1: Paper writing could be improved.**
Thanks for your valuable suggestion. We have polished the whole paper to make it more smooth and concise for readers. Please refer to **Q2** in the **"General Response"** for specific modific... | Rebuttal 1:
Rebuttal: # General Response:
We thank all reviewers for their insightful and constructive suggestions, which help a lot in further improving our paper.
**Q1: The number of training data.**
We notice that several reviewers are concerned about the number of data required to train the proposed method comp... | NeurIPS_2023_submissions_huggingface | 2,023 | Summary: The paper introduces image based visual servo to refine the coarse poses estimated using a coordinate regression network trained on neural rendering based 3D labels. The paper also uses correspondences in depth between neural rendered image and query image to iteratively refine the pose of the image. The paper... | Rebuttal 1:
Rebuttal: Thank you for the constructive comments and suggestions.
**W1 and W2: The paper trains the coordinate regression network using the approximate 3D labels provided by NeRF which is not ideal. It is not clear how the authors decide which pixels in the NERF-rendered images are used to train the coord... | null | null | null | null | null | null |
Geometric Transformer with Interatomic Positional Encoding | Accept (poster) | Summary: The paper proposes a Transformer based architecture incorporating interatomic information upon learned Atomic Cluster Expansion, integrated into the self-attention and refined residually over the network.
The method obtains state-of-the-art performance on a majority of prediction tasks on two popular molecular... | Rebuttal 1:
Rebuttal: ## Response
Thanks for your valuable comments. We provide the point-to-point responses in the following.
### Weakness 1
1. Sorry for the lack of literature review torward the Transformer-based architecture. The works you mentioned are pioneers in the molecular modeling. We introduce them in the... | Summary: The authors propose an Interatomic Positional Encoding (IPE), and introduce the Geoformer. IPE is motivated by Atomic Cluster Expansion (ACE), which describes the many-body contributions in transformers for molecular modeling. The authors first derive integrated cluster potentials for atom pair i,j in a molecu... | Rebuttal 1:
Rebuttal: ## Response
Thanks for your valuable comments. We provide the point-to-point responses at length in the following.
### Weakness 1
1. The distance filter (Radial Basis Fuctions, RBF) in TorchMD-Net is a sub-class of Atomic Cluster Expansion (ACE), which could be considered as the 2-body expansio... | Summary: The paper introduces Geoformer, a new geometric transformer for molecular property prediction. The authors argue that while transformers have been dominant in various data modalities, their application to molecular modeling has been limited. To address this, the authors propose Interatomic Positional Encoding ... | Rebuttal 1:
Rebuttal: ## Reponse
Thanks for raising some valuable concerns. We provide the point-to-point response in the following.
### Weakness 1
After conducting a literature review on point cloud analysis, we discovered related work which incorporates geometric information in the point cloud modeling [1-5]. Thou... | Summary: The paper introduced a Transformer-based model with interatomic positional encoding (IPE) for molecular modeling, which the authors termed "Geoformer". Geoformer incorporated novel positional encoding derived from empirical physical knowledge (atomic cluster expansion) and was able to achieve comparable or eve... | Rebuttal 1:
Rebuttal: ## Response
Thanks for your comments and acknowledgment of our work. We provide the point-to-point responses in the following.
### Question 1
We appreciate your suggestion and have provided more explanation of the construction of the B-basis in Theorem 2 and the Appendix for better clarity. Ind... | null | NeurIPS_2023_submissions_huggingface | 2,023 | Summary: The paper proposes a geometric Transformer called Geoformer for geometric molecular modeling. It designs Interatomic Positional Encoding (IPE), taking 3D geometric information into account to parametrize atomic environment for positional encoding in Transformer architecture. Geoformer adopts a learned Interato... | Rebuttal 1:
Rebuttal: ## Response
Thanks for your valuable comments. We provide point-to-point responses in the following:
### Weakness 1
The term **cluster** is a concept in the original Atomic Cluster Expansion (ACE) theory, which represents the local chemical environment of centered atoms. A cluster $\alpha$ cont... | null | null | null | null | null | null |
Beyond Invariance: Test-Time Label-Shift Adaptation for Addressing "Spurious" Correlations | Accept (poster) | Summary: This paper proposes to perform test-time adaptation to exploit the existence of spurious correlations in a setting where $p(x | y, z)$ remains constant, but when the prior $p(y, z)$ changes based on the target task.
The authors evaluate this new method on various tasks, including standard image/text benchmar... | Rebuttal 1:
Rebuttal: Thank you for your review. Response to your points is below.
“**There are no comparisons to other test-time adaptation methods.**”
All the test-time adaptation methods we are aware of are focused on covariate shift, or label shift, but no methods (AFAIK) can handle shift in the correlation betw... | Summary: This paper introduces a new method named Test-Time Label Shift Adaptation(TTLSA) to manage distribution shifts in machine learning, specifically focusing on spurious correlations.
Rather than eliminating these spurious relations as mentioned in other papers, TTLSA utilizes the spurious features to adjust the ... | Rebuttal 1:
Rebuttal: Thank you for your thoughtful review. Response to your points is below.
“**The TTLSA method assumes that the generative distribution p(x|y, z) is preserved across domains. Here, the authors only address z as a one-dimensional feature. This assumption can be restrictive and may not hold true in ma... | Summary: **A brief summary:**
The paper proposed "Test-Time Label-Shift Adaptation" (TTLSA), that utilizes rather than eliminates spurious correlations. TTLSA adapts to changes in marginal distribution using the Expectation Maximization algorithm. Experiments on several datasets show that TTLSA outperformed traditional... | Rebuttal 1:
Rebuttal: Thank you for your review. Response to your points is below.
“**The training on the discriminative classifier requires a large labeled dataset from the source distribution…. still relies on a trained discriminative classifier, which is used to predict p(y, z|x)**”
We agree this is a weakness.... | Summary: - This paper tackles the test-time adaptation problem to mitigate spurious correlation. By assuming the pervasiveness of $p(x|y, z)$, the authors derived an alternative method for predicting the target distribution and proposed Test-Time Label-Shift Adaptation (TTLSA) to optimize it. The TTLSA involves two ste... | Rebuttal 1:
Rebuttal: Thank you for your review. Response to your points is below.
“**One of my concerns is the assumption of having an annotated training dataset.**"
We agree this is a weakness. However, as we show in Appendix C.5 (in supplemental), it is possible to apply our method with a very small amount of labe... | Rebuttal 1:
Rebuttal: We respond to each reviewer separately below. However, a common concern was our dependence on a training set which has y and z labels. We agree this is a weakness. However, as we show in Appendix C.5 (in the supplemental material), it is possible to apply our method with a very small amount of lab... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Multi-body SE(3) Equivariance for Unsupervised Rigid Segmentation and Motion Estimation | Accept (poster) | Summary: The paper describes a method to simultaneously segment rigid parts and estimate part-level transformations for multi-body point cloud inputs. Given the input point cloud, equivariant features are first extracted and aggregated to invariant features for part segmentation. To estimate the transformations between... | Rebuttal 1:
Rebuttal: Thank you for your valuable feedback. We appreciate your recognition of our use of equivariant backbones & their invariant counterparts, our solution to the cold-start problem, and our extensive and compelling experiments. We would like to address each of your concerns in detail below.
**Q1:** Th... | Summary: This paper studies the problem of rigid part segmentation and motion segmentation. The proposed network first uses a global SE(3) equivariant backbone to extract point features, then a part segmentation head will predict invariant zero-order segmentation and a motion head will predict one-order transformations... | Rebuttal 1:
Rebuttal: We appreciate your thorough feedback and are grateful for your recognition of our exploitation of applying equivariance to segmentation and our convincing experiments. We have carefully considered each of your concerns and have provided detailed responses to each one below.
**Q1:** The “equivaria... | Summary: This paper proposes a unified framework to simultaneously handle rigid segmentation and motion estimation. Specifically, an probabilistic SE(3)-equivariant network is presented to obtain transformation-invariant features, which is followed by two heads for rigid segmentation and motion estimation. Due to the i... | Rebuttal 1:
Rebuttal: Thank you for your constructive feedback. We are grateful for your recognition of the significance and originality of our training strategy based on part-level SE(3)-equivariance, clarity, and experiments. We have carefully reviewed each of your concerns and all the additional experiments would be... | Summary: This paper proposes a part-level SE(3)-equivariant framework for multi-body rigid motion modeling. The two heads that take in the SE(3) features, segmentation and motion estimation, are carefully crafted based on their inherit invariant and equivariant characteristics. The relationship of scene flow, rigid seg... | Rebuttal 1:
Rebuttal: We appreciate your valuable feedback and are grateful for your recognition of our writing and contributions presented in our paper, such as the utilization of SE(3) equivariance, experiments, and model size. We have carefully considered your main concern and have provided detailed responses below.... | Rebuttal 1:
Rebuttal: We thank all reviewers for their valuable comments and suggestions. We would like to highlight some strengths of our paper pointed out by reviewers, including a **well-motivated** (sjc7) and **original** (Ztt7) idea, **extensive and convincing experiments with impressive results surpassing state-o... | NeurIPS_2023_submissions_huggingface | 2,023 | Summary: Modeling multi-body rigid movements involves - rigid segmentation which is category-agnostic and motion estimation which is open-set. This work proposes a part-level SE(3) equivariant network that takes in point-clouds corresponding to video frames and can effectively estimate category-agnostic open-set motion... | Rebuttal 1:
Rebuttal: Thank you for your valuable feedback. We appreciate your recognition of our improvements on feature equivariance, the high performance and low FLOPs, and our use of interdependence for unsupervised training. Please find our responses to each concern below.
**Q1:** The paper writing could be bette... | Summary: This paper presents a method that estimates the part segmentation, and the rigid motions of the parts, for an articulated object (with multiple rigid parts), given as input a sequence of point clouds of that object. The method works by first computing SE3-equivariant features for the points, and then processin... | Rebuttal 1:
Rebuttal: Thank you for your valuable feedback! We are grateful for your acknowledgment of the motivation, novelty, and contributions of our model in terms of its parameters, complexity, and accuracy. While we understand your concern regarding the clarity of certain sections, we would like to address each o... | null | null | null | null |
RangePerception: Taming LiDAR Range View for Efficient and Accurate 3D Object Detection | Accept (poster) | Summary: This paper studies range-view-based LiDAR 3D object detection. It proposes Range Aware Kernel and Vision Restoration Module to handle the problems of range view representation, such as inconsistent range distribution and boundary splitting. The main contributions are as follows:
- The proposed module boosts t... | Rebuttal 1:
Rebuttal: Thank you for your positive recognition of our work. We hope the discussion below could address your concerns.
## Response 1.1 Extensive Ablation Study for Range Aware Kernel (Question 1 & Weakness 2)
We have conducted thorough comparison experiments during the development process of the Range A... | Summary: The paper presents an efficient and accurate RV-based 3D object detection framework called *RangePerception*. The framework addresses two critical challenges impeding the performance of existing RV-based methods: 1) a natural domain gap between the 3D world coordinate used in output and 2D range image coordina... | Rebuttal 1:
Rebuttal: We appreciate your recognition of our work. We hope the following discussion adequately addresses your concerns.
## Response 2.1 Occlusion and Scale Variation Phenomena (Question 1)
As is illustrated in our main paper L202-204 \& Fig.2, RangePerception learns the pixel-wise dense predictions, n... | Summary: This paper proposes RangePerception, an RV-based 3D detection framework that aims to address the issues of Spatial Misalignment and Vision Corruption in range-view representation. The Range Aware Kernel (RAK) disentangles the range image space into multiple sub-spaces, and overcomes the Spatial Misalignment is... | Rebuttal 1:
Rebuttal: Thank you for recognizing our work. We aim to address your concerns adequately in the following discussion.
## Response 3.1 Occlusion and Scale Variation Phenomena (Question 1)
This topic is carefully discussed in Response 2.1 to reviewer PXtx.
## Response 3.2 Performance on nuScenes Dataset (W... | Summary: This paper proposed a new range view based 3D object detection algorithm, which achieved on-par performance compared to sota BEV/voxel based algorithm.
Strengths: 1. The paper starts from first principle, analyzes the possible information losses during the lidar data collection/analysis/transformation process... | Rebuttal 1:
Rebuttal: We appreciate your positive acknowledgment of our work. We hope that the subsequent discussion effectively addresses the concerns you have raised.
## Response 4.1 Incorporation of Multi-return Range Images (Question 1)
The incorporation of multi-return LiDAR data is extensively investigated and ... | Rebuttal 1:
Rebuttal: Thank you for acknowledging our work positively. We are pleased to provide the supplemental responses for part of the individual responses.
## Response 5.1 Supplemental Response for Reviewer GJQ1
### Table A: Ablation study of Range Aware Kernel, measured by 3D AP/APH on WOD validation set.
| ... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Hypervolume Maximization: A Geometric View of Pareto Set Learning | Accept (poster) | Summary: This manuscript introduces a method for multi-objective optimization and mainly contributes to modeling the Pareto set. Compared to previous methods which usually focus on limited and discrete Pareto solutions, the method in this manuscript is capable of modeling the full Pareto set in a continuous way. They a... | Rebuttal 1:
Rebuttal: We sincerely appreciate your valuable feedback. We hope that our following response can address your concerns.
**W1. The text on the figures is too small to read. Maybe considering to add a description in the caption.**
Thank you for your guidance. In the revised edition, we have implemented a l... | Summary: This paper presents a novel perspective of hypervolume maximization in Pareto learning. The proposed method is largely interpretable based on this perspective, and could outperform many peer competitors considering efficiency and solutions.
Strengths: 1. The hypervolume analysis from the polar coordinate per... | Rebuttal 1:
Rebuttal: We sincerely thank you for the valuable comments on our work. We hope to address your concerns with the following answers.
**W1. The innovations are not emphasized. For example, if the benefit of the proposed method based on geometric analysis includes improved efficiency compared with other Pare... | Summary: This paper proposes a geometric perspective for Pareto set learning by establishing an equivalence between learning the complete Pareto set and hypervolume maximization. A theoretical analysis is provided to examine the gap between the estimated hypervolume and the true hypervolume of the Pareto set and empiri... | Rebuttal 1:
Rebuttal: We sincerely appreciate your valuable feedback. We hope that we can address your concerns with the following answers more or less.
**W1. More Problems: The empirical studies need to involve more problems, e.g., those from the popular benchmarks DTLZ and WFG, which have been shown to represent so... | Summary: The author presents a novel approach to multiobjective algorithms aimed at modeling the Pareto set using neural networks. The proposed approach in the manuscript allows for the direct modeling of the entire Pareto set, and it also establishes an equivalence between learning the complete Pareto set and maximizi... | Rebuttal 1:
Rebuttal: We are grateful for your constructive comments on our work. We hope that our following response can address your concerns more or less.
**Q1. Is the performance of the network influenced by its structure? Are there any anticipated performance differences when using alternative architectures?**
T... | Rebuttal 1:
Rebuttal: **General response**
We sincerely appreciate the diligent efforts of all the reviewers. In response to their concerns, we have conducted three additional experiments.
- - -
> **1. Many-Objective Problem**
In our first experiment, we evaluate the performance of our proposed method on many-objecti... | NeurIPS_2023_submissions_huggingface | 2,023 | Summary: The paper presents an algorithm to learn a continous approximation of the Pareto frontier in multi-objective reinforcement learning (MORL). The idea of the algorithm is to train a neural network to produce solutions such that their hypervolume is maximized.
The idea iteself is not new, but the proposed method... | Rebuttal 1:
Rebuttal: Thank you for your helpful feedback. We greatly appreciate it. In response, we have carefully considered your concerns and incorporated them into our revisions.
**W1-2. Missing Highly Relevant Work: The idea of learning a continuous Pareto frontier was first proposed --even though without a neura... | Summary: This paper presents a novel approach to multiobjective algorithms that allows for the direct modeling of the entire Pareto set. The authors present a novel approach to Pareto set learning (PSL) from a geometric perspective, distinguishing it from existing methods that treat all preferences equally, resulting i... | Rebuttal 1:
Rebuttal: We sincerely appreciate your valuable feedback. We have taken your concerns, and we hope our following answers can address your concerns more or less.
**W1. There is a limited discussion regarding the design of the neural network structure, and the machine learning perspective could benefit from ... | null | null | null | null |
Rewrite Caption Semantics: Bridging Semantic Gaps for Language-Supervised Semantic Segmentation | Accept (poster) | Summary: This paper first points out the semantic gap problem of the existing language-supervised semantic segmentation method. It shows that not all visual elements are included in the corresponding language annotations. Then the paper proposes Concept Curation (CoCu), which includes Vision-driven Expansion, Text-to-V... | Rebuttal 1:
Rebuttal: We thank for your valuable suggestions and response as follows for your concern.
**Q1: Writing in Sec. 3.3?**
Thanks for pointing it out. As suggested, we will clarify the motivation of our method at the beginning of Section 3.3 as well as its relation with Section 3.1.
**Q2: Online or Offline?... | Summary: This paper proposed a novel data curation/argumentation process, named Concept Curation (CoCu), for language-supervised semantic segmentation. In the setting of language-supervised semantic segmentation (e.g. GroupViT), the network is trained with image-text contrastive loss on large-scale image-text pairs. Au... | Rebuttal 1:
Rebuttal: We thank you for your value of our efforts and helpful suggestions. Please check our clarification below regarding your concerns.
**Q1: More qualitative comparison.**
Please refer to the Appendix Figure 2 for additional qualitative comparisons (heatmap). The figure shows that CoCu learns visual... | Summary: This paper targets learning unsupervised semantic segmentation from image-text pairs. The authors specifically address the issue of training data quality in previous method Group-ViT and propose an approach to enhance the captions with additional visual concepts through an automated pipeline. The paper demonst... | Rebuttal 1:
Rebuttal: We thank you for your valuable suggestion. Below please find our clarification regarding your concern.
**Q1: Does it work by distilling CLIP Model? Replace CLIP with GroupViT?**
No, CoCu does not rely on vast knowledge encoded in CLIP [32] but more on our novel pipeline design. To verify this, w... | Summary: Current VLMs suffer from a noticeable semantic gap between visual and textual modalities since many visual concepts present in images are easily missed in their paired captions. This work proposes Concept Curation (CoCu), a pipeline that leverages CLIP to compensate for the missing semantics. For each image-te... | Rebuttal 1:
Rebuttal: Thank you for your comments. We appreciate the value you see in our innovative motivation and competitive experimental results. Below, please find clarifications regarding your concerns below.
**Q1: In Table 1, are both CoCu and GroupViT trained for 30 epochs?**
Yes, in manuscript Table 1, pre-t... | Rebuttal 1:
Rebuttal: Please check global_response.pdf for our new Figures.
Pdf: /pdf/ab1c4584b82b029b287583a1ef997cbff32e3d2b.pdf | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
(Almost) Provable Error Bounds Under Distribution Shift via Disagreement Discrepancy | Accept (poster) | Summary: This paper try to propose a new metric on evaluation of the distribution shift of dataset. The authors first prove the proposed error estimation bound with some assumptions. Then they demonstrate the effectiveness of their method through training a surrogate model maximizing the disagreement discrepancy. The a... | Rebuttal 1:
Rebuttal: We thank the reviewer for their feedback.
In your review, you evaluated the soundness as “poor”. We are not sure we understand the reason for this evaluation. Our understanding is that soundness is meant to evaluate the correctness of claims in a given work. Do you believe that our bound or our ... | Summary: The paper proposes a method that under certain assumptions provides upper bounds on the accuracy under distribution shift when provided with unlabelled test data from the shift's target distribution.
Strengths: The assumptions necessary to obtain the bounds are clearly stated and discussed.
The problem of es... | Rebuttal 1:
Rebuttal: Thanks for your feedback! We address your comments below.
> **The bounds…rely on assumptions…confirmed empirically after the true labels…are known**
You’re correct that we can’t confirm our assumption until the true labels are known. **It’s important to remember that this is true of *every* met... | Summary: The paper aims to propose a way to characterize the error bounds under distribution shift with provable error guarantees. Based on several assumptions, the authors show that the target error bound can be reasonably estimated by maximizing the agreement within the desired classifier on source distribution and m... | Rebuttal 1:
Rebuttal: Thanks for your thorough review! We appreciate that you have recommended acceptance, and that you consider the presentation and contribution of this work to be of high quality.
> **“There could be cases when the assumptions fail to hold”**
We agree that there are cases where the assumption might... | Summary: This paper provides new error bound of distribution shift based on a notion called disagreement discrepancy. By assuming that the model class has enough expressiveness, it is theoretically proven that the error can be bounded by the worst-case disagreement discrepancy. Empirical investigation shows that such a... | Rebuttal 1:
Rebuttal: Thanks for your feedback! To address your points:
> **“briefly introduce $\mathcal{H}$- and $\mathcal{H}\Delta\mathcal{H}$-divergence”**
We do roughly describe them in Section 3.1, but we will add a more detailed description with the additional space.
You’ve stated the writing could use improve... | Rebuttal 1:
Rebuttal: Thanks to all reviewers for their helpful comments and suggestions!
We worked very hard to present this work as clearly as possible, so we are glad to hear that **reviewers xsfT, rQbE, TEVz, and uhaU all found the writing clear and easy to follow (even “enjoyable to read”!)**
A few reviewers poi... | NeurIPS_2023_submissions_huggingface | 2,023 | Summary: This paper describes a novel method to estimate the error of a classifier under a shift of distribution. It relies on finding a worst case "critic" classifier which maximizes disagreement on some unlabeled target domain and hence bounds the disagreement with the true labeling function. The authors apply severa... | Rebuttal 1:
Rebuttal: Thanks so much for your thoughtful feedback! **We are very happy to hear that you appreciated the presentation; we took great pains to present this work as clearly as possible.**
We will add the suggested reference [1] to the related work section. We remark that it would not make sense to include... | Summary: The authors propose a novel approach to estimate error bounds for deep neural networks under distributions shift. The proposed bounds remedy the limitations of exiting bounds that either provide vacuous bounds or underestimate the error for a big fraction of shifts while also often requiring access to test lab... | Rebuttal 1:
Rebuttal: Thanks for your feedback! You’ve stated that you are mostly concerned with missing experimental results, which we are happy to add. **We put preliminary results in our pdf response, which we'll expand on in the paper.**
We want to briefly explain why these results were not originally included. Yo... | null | null | null | null |
Django: Detecting Trojans in Object Detection Models via Gaussian Focus Calibration | Accept (poster) | Summary: The paper proposes the first object detection backdoor detection framework Django (Detecting Trojans in Object Detection Models via Gaussian Focus Calibration). It leverages a dynamic Gaussian weighting scheme that prioritizes more vulnerable victim boxes and assigns appropriate coefficients to calibrate the o... | Rebuttal 1:
Rebuttal: **[W1] Clarification of the Django**
We are sorry for the confusion. [51, 49, 15, 31] are initially designed for detecting backdoor on image classification models. According to our evaluation(main text, line 268-line 291, Table 1), existing trigger inversion techniques all have limited performan... | Summary: This paper investigates the problem of detecting trojans in the context of Objection Detections. The authors first observed that the poison effect can vary significantly across bounding boxes in object detection models due to its dense prediction nature, which leads to a misalignment issue for existing trigger... | Rebuttal 1:
Rebuttal: **[W1] Comparison with non reverse-engineering based backdoor detection methods**
Please refer to **global response A1**.
------
**[W2] Theoretical comprehension of the choice of gaussian**
Please refer to **global response A2**.
------
**[Q1] Correlation between weight schemes and comprom... | Summary: This paper proposes a novel trigger detection method that works on object detection tasks. The proposed method leverages a novel Gaussian Focus Loss to calibrate the misaligned loss during trigger inversion. The evaluations presented in the paper demonstrate that the proposed method is effective in different s... | Rebuttal 1:
Rebuttal: **[W1] No limitation discussed**
As discussed in the main text (lines 107-113), our primary focus in this paper is on attacks that use static polygon triggers, which are more feasible in real-world scenarios. How to effectively inject more complex attacks such as WaNet[38], DFST[7] and dynamic at... | Summary: The key idea of this paper is to propose Django, an adaptation of the trigger reverse-engineering technique used for detecting backdoored models in classification to object detection models. The paper shows that existing trigger reverse-engineering techniques are ineffective in object detection and finds that ... | Rebuttal 1:
Rebuttal: **[W1] More comprehensive related work in object detection**
We thank the reviewer for the detailed suggestion regarding related work. We will add a new section to introduce more detail regarding backdoor attack and defense in the context of object detection.
-----
**[W2] Comparison with non ... | Rebuttal 1:
Rebuttal: **Global Response**
We extend our sincere gratitude to all the esteemed reviewers for their invaluable and astute feedback. In the Global Response section, we shall diligently address the common questions raised and supplement the document with additional figures and tables to enhance clarity. Be... | NeurIPS_2023_submissions_huggingface | 2,023 | Summary: In this paper, the author proposes a new method for detecting backdoors in object detection models. The author finds that directly applying backdoor detection methods for classification models to object detection models results in a loss misalignment problem. Therefore, the author suggests assigning different ... | Rebuttal 1:
Rebuttal: **[W1] Hyperparameter settings and sensitivity**
In Section 4.2 (lines 326-332), we discuss and evaluate the sensitivity of hyperparameters in the pre-processing procedure ($h$ and $\omega$), and the main results are shown in the main text Figure 5b. Recall the values of $h$ and $\omega$ strik... | null | null | null | null | null | null |
DPM-Solver-v3: Improved Diffusion ODE Solver with Empirical Model Statistics | Accept (poster) | Summary: The paper proposes a new sampling algorithm, DPM-Solver-v3, to solve the diffusion ODE. Applying the Rosenbrock-type exponential integrators, DPM-Solver-v3 pre-computes several coefficients (empirical model statistics) to minimize the norm of the gradient of non-linear term, which leads to reduced discretizati... | Rebuttal 1:
Rebuttal: Thank you for the interest and acknowledgment of our theoretical and empirical contributions. Below are our explanations to the questions, which we hope may clarify some misunderstandings. We kindly request that you consider raising the score accordingly if you are satisfied.
*W1: The method need... | Summary: This work proposes a method to speed up the DPM solver by using empirical model statistics (pre-trained model) as well as employing multistep methods with a predictor-corrector framework for improving sample quality. Promising results were reported.
Strengths: - This work addresses an important problem of fas... | Rebuttal 1:
Rebuttal: Thank you for recognizing our work's importance and theoretical neatness. However, we feel that there are some fundamental misunderstandings. We will address your questions and concerns below. We kindly request that you consider raising the score accordingly if you are satisfied.
*W1: The improve... | Summary: This work proposes DPM-Solver-v3, an ODE solver for diffusion model inference. The method builds upon existing formulations involving exponential integrators, where the linear term of the ODE is canceled and only the noise predictor needs to be approximated. A new model parameterization is derived which involv... | Rebuttal 1:
Rebuttal:
Thank you for your detailed review, especially for reading the proofs and ablations in the Appendix. Below we provide detailed responses for the questions and clarify misunderstandings. We kindly request that you consider raising the score accordingly if you are satisfied.
*W1 & W2: Some tuning ... | Summary: This paper proposed an improved/generalized version of DPM-solver. It include three more coefficients based on the semi-linear ODE solution proposed by DPM-solver, accounting for minimizing the 'linearity' for the nonlinear component of the ODE and minimizing the first-order discretization error respectively. ... | Rebuttal 1:
Rebuttal: Thank you for your positive comments. Below we provide detailed responses for the questions and clarify misunderstandings. We kindly request that you consider raising the score accordingly if you are satisfied.
*W1: Equation 10 is wrong, the first-order discretization error may be amplified.*
**... | Rebuttal 1:
Rebuttal: We sincerely thank all reviewers' effort for the detailed and insightful suggestions. First we'd like to present all the paper revisions.
### Summary of paper revisions
**Theories**
- Theoretical justification of why minimizing first-order discretization error can help higher-order solvers (Rev... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Inner Product-based Neural Network Similarity | Accept (poster) | Summary: This paper studies neural network similarity which is an important problem in many areas such as federated learning and continue learning. In this paper, the authors develop a new method to reduce NN representational similarity to filter subspace distance. Moreover, they present the effectiveness and efficienc... | Rebuttal 1:
Rebuttal: Thank you for your supportive comments.
**Q1:** Is it possible to extend this algorithm to other models, such as transformer-based model?
**A:** We can extend our method to other models, including transformer-based models. The proposed method relies on the decomposition of weight matrices, and s... | Summary: In their paper, the authors introduce a novel approach to significantly decrease the computational cost of representational similarity analysis in CNNs by transitioning it to filter subspace distance evaluation. Their proposed filter subspace-based similarity is both theoretically and empirically demonstrated ... | Rebuttal 1:
Rebuttal: Thank you for your supportive comments.
**W1:** Could you elaborate on how Assumption 2.6 is applicable to complex real-world datasets? Consider a simple matrix $\begin{pmatrix} 0 & 0 \\\\ 1 & 0.9 \end{pmatrix}$, the assumption does not hold true.
**A:** Our assumption states that different chan... | Summary: This paper presents a new approach for measuring the similarity between neural network models based on a new efficient metric defined layer-wise on filter atoms.
In particular, the metric is defined as the cosine similarity between this filter atoms that can be implemented as the normalised inner product of t... | Rebuttal 1:
Rebuttal: Thank you for your supportive comments.
**W1:** Comparison only with CCA and CKA-based similarity metrics
**A:** We want to claim that CCA and CKA are two representative feature-based network similarity measures that are adopted in many works [1,2,3]. Showing a strong linear correlation with CCA... | Summary: This paper proposes an approach for evaluating neural network similarity by decomposing convolution layer into linear combination of atom filters. The basic idea derives from [33] and the paper extend the approach for continual learning (CL), to federated learning (FL). The paper provides both empirical and th... | Rebuttal 1:
Rebuttal: Thanks for your constructive comments. We address all your concerns in the following.
**W1.1** The proposed method is restricted to only convolutional layers, and it needs different NNs using the same atom coefficients.
**A:** Although our method focuses on convolutional filter subspace, it can... | Rebuttal 1:
Rebuttal: ### General Responses
*We provide figures for the additional required experimental results in the attached PDF.*
**GR1: Extend our algorithm to transformer-based models.**
Although our method focuses on convolutional filter subspace due to the highly compact size of resulting filter subspace ele... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Reinforcement-Enhanced Autoregressive Feature Transformation: Gradient-steered Search in Continuous Space for Postfix Expressions | Accept (spotlight) | Summary: Feature transformation is an effective way to improve downstream task. This paper aims to improve the search efficiency of the optimal feature space while ensuring the stability and robustness of the transformation. They formulate the discrete Automated Feature Transformation (AFT) problem as a continuous opti... | Rebuttal 1:
Rebuttal: Dear Reviewer GfrD
Thank you for your detailed feedback and observations on our work. We highly appreciate the recognition of the strengths of our paper, and we acknowledge the points raised in the weaknesses and questions sections. We will address each of them comprehensively:
**1. Regarding th... | Summary: This paper introduces a succinct yet effective framework for automatic feature transformation. The authors mapped decision sequences collected from reinforcement learning into a high-dimensional vector space. Optimization is carried out through the gradient direction provided by an evaluator, and a sequence re... | Rebuttal 1:
Rebuttal: Dear Reviewer yQs6,
We sincerely appreciate the reviewer's time and effort in assessing our paper. We are pleased to find that the design and potential of the framework were recognized as strengths. We want to address the highlighted weaknesses and answer the posed questions to improve our paper... | Summary: The authors propose a feature transformation method that reformulates the problem as a continuous space optimization task and utilizes a reinforcement-enhanced autoregressive framework for gradient-steered search. The method involves four steps: (1) reinforcement-enhanced data preparation, (2) feature transfor... | Rebuttal 1:
Rebuttal: Dear Reviewer miDK,
We want to thank you for acknowledging this paper's novelty, effectiveness, extensive experiments, and reproducibility.
**1. Regarding the inferior performance of DIFER on the openML dataset:**
Response: OpenML616 and OpenML637 are two artificially constructed datasets, som... | Summary: Distinct from existing work, this study collects feature transformation operation sequences, which have a well-researched background based on reinforcement learning. It then obtains hidden representations of these limited numbers of sequences in a self-supervised manner and finally optimizes this continuous ve... | Rebuttal 1:
Rebuttal: Dear Reviewer MJDJ,
Thank you for acknowledging that our paper has a solid research background, well-written, easy-to-understand, and extensive experiments to support the claimed research insights and technical contributions. For your mentioned issues:
**1. Regarding the typos, name inconsistenc... | null | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.