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 |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
STREAMER: Streaming Representation Learning and Event Segmentation in a Hierarchical Manner | Accept (poster) | Summary: This paper introduces STREAMER, a self-supervised architecture for hierarchical representation learning and segmentation of streaming perceptual inputs. The model adapts to the complexity of the input domain through layer-by-layer training, aiming to accurately predict future events while providing necessary i... | Rebuttal 1:
Rebuttal: ### Weaknesses:
1. We will clarify the terminologies in the camera ready submission. For clarification:
- *Event model* and *representation*: As defined in cognitive psychology literature, "An event model is a representation of what is happening now, which is robust to transient variability ... | Summary: This paper tackles the task of event segmentation in videos, with a particular emphasis on trying to perform this task in a self-supervised setting. The authors draw inspiration from event segmentation theories from human perception to design a model, STREAMER, that semantically groups inputs at varying length... | Rebuttal 1:
Rebuttal: Thank you for your remarks on the qualitative results and highlighting the strengths of our idea.
### Weaknesses:
1. Building higher-level hierarchical events in a self-supervised manner requires long videos. Existing unsupervised methods, like ABD, are clustering-based methods which do not requ... | Summary: This paper introduces a self-supervised method to learn the hierarchical representation of streaming video events through predictive learning. This method segment videos into multiple chunks based on different semantic scale, from low-level fine-grained segmentation to high-level conceptual understanding. Mean... | Rebuttal 1:
Rebuttal:
### Weaknesses:
1. Thank you for the suggestion. Although NewsNet and MovieNet are large datasets with long enough videos, they have not yet been released. NewsNet was published very recently at CVPR23; it is not clear when it will be released to the public. Furthermore, in addition to the unrel... | Summary: This paper presents an algorithm for unsupervised temporal segmentation of videos in an online setting, by modeling actions hierarchically. The paper is based on the idea that prediction errors arise when a new action comes up, which makes the existing context insufficient (within a certain abstraction level).... | Rebuttal 1:
Rebuttal: Thank you for your general remarks on the idea and pointing out its strengths, such as avoiding memory constraints in long-term events modeling.
### Strengths:
3. Thanks for this remark. We just wanted to clarify that the qualitative videos were not cherrypicked, but selected at random.
### Wea... | Rebuttal 1:
Rebuttal: We thank the reviewers for their comments and useful suggestions. We think that these comments have helped us better refine the paper.
A common concern expressed in the reviews is evaluation only on egocentric dataset. We address this concern here in addition to individual rebuttals.
Our method,... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Speculative Decoding with Big Little Decoder | Accept (poster) | Summary: This paper presents a new adaptive computation method to be used for text generation tasks. The method works by applying a small model, and if needed, revising its predictions using a larger, more expensive model. The resulting model obtains a far better speed-accuracy curve compared to the baseline model.
To... | Rebuttal 1:
Rebuttal: ### W1. Elaboration on non-autoregressive use of the model
We can use the same decoder model, but rather in a non-autoregressive (parallel) fashion by providing the model with multiple tokens and processing them simultaneously.
This mechanism is identical to how decoder models can simultaneously ... | Summary: The paper proposes BiLD, a framework to coordinate a large model and a small model such that the large model is only executed infrequently and efficiently in a non-autoregressive manner to refine the small model's inaccurate predictions. The proposed two policies allows the small model to hand control over to ... | Rebuttal 1:
Rebuttal: ### W1. Evaluation on QA Benchmarks
We acknowledge the comment that our paper lacks evaluations on QA tasks.
However, we highlight that **QnA tasks are not the ideal benchmarks for assessing our methodology's effectiveness**. This is because the main focus of our framework lies in enhancing *dec... | Summary: The paper presents an approach to improve decoding using a smaller model to create drafts, and fallback to a larger model whenever the smaller model predicts any token with probability smaller than a fixed threshold. Additionally, during this fallback, if the larger model produces a token that is not in sync w... | Rebuttal 1:
Rebuttal:
### W1. Insights on Better Latency and Performance
Quantitative analysis reveals a consistent trend where our method, when compared to speculative decoding [1, 2], effectively enhances both text generation quality and latency. We provide insights and explanations into why our approach surpasses ... | Summary: This paper presents an efficient generation approach that utilizes a smaller model during decoding to reduce computational requirements. In essence, the large model is only employed when the small model's generation exhibits low confidence (fallback policy) or when there is a significant distribution differenc... | Rebuttal 1:
Rebuttal: ### W1. Batch sizes
While we only evaluate the use case of batch size 1, it is **not a limitation and it is straightforward in modern LLM serving systems to extend the BiLD framework to multi-batch applications**. This is because the modern LLM serving systems (especially for multi-batch applicat... | Rebuttal 1:
Rebuttal: We appreciate all the reviewers for taking the time to review our work and providing us with their valuable feedback. We provided responses to the questions that each of the reviewers has commented on.
Additionally, we present a comprehensive comparison with previous works [1, 2] as a global rebut... | NeurIPS_2023_submissions_huggingface | 2,023 | Summary: The paper proposes Big Little Decoder (BiLD), a framework to accelerate autoregressive text generation by using a small decoder model to generate most tokens efficiently and invoking a large decoder model occasionally to refine inaccurate predictions. BiLD manually designs two policies to invoke refinement, 1)... | Rebuttal 1:
Rebuttal: ### W1. BiLD with Random Sampling
Our approach **isn't restricted to greedy decoding, but it seamlessly extends to sampling methods.** Even though the fallback policy is based on the maximum prediction probability of a small model, we can still perform decoding based on random sampling. The only ... | null | null | null | null | null | null |
SKI to go Faster: Accelerating Toeplitz Neural Networks via Asymmetric Kernels | Reject | Summary: The present paper offers a Toeplitz matrix architecture which can handle sequence modeling. The architecture comes in two flavors. The first flavor is a fast version that is most useful for bi-directional tasks. It speeds up previous Toeplitz networks by using an interpolation and low rank approximation scheme... | Rebuttal 1:
Rebuttal: Thank you for your helpful critiques and suggestions. We will use them to improve clarity.
# Weaknesses
## Broader Comments
* Speedups arise from not multiplying the RPE by a decay bias.
* Table 1: FD-TNN doesn't add parameters (48.58 million vs TNN's 48.59). We are faster than TNNs with similar ... | Summary: The authors of the paper propose two modifications of a recently published alternative to attention mechanism, Toeplitz Neural Operator (TNO), which constitutes the most important part of Toeplitz Neural Networks. The application of TNO is the multiplication of the input sequence by a Toeplitz matrix. Paramete... | Rebuttal 1:
Rebuttal: Thank you for your comments and overall positive assessment, as well as your suggestions for improving the clarity.
# Weaknesses
1. We admit that we do not have results for pathfinder-X, the most challenging dataset and task. We had difficulty achieving good performance on it for the model sizes ... | Summary: The paper presents several techniques to speed up Toeplitz neural networks (TNNs). In particular, TNNs use convolution of length n (sequence length of the input) and so scales as O(n log n), and TNNs have many calls to the MLP that generate relative positional encoding (RPE) and decay bias. To reduce the time ... | Rebuttal 1:
Rebuttal: Thank you for your comments, which have helped us create plots that we believe will strengthen the paper. We answer your questions below, and hope they are adequate to consider altering your assessment.
# Strengths
1. Thank you. Note that the relationship is actually reversed: it is the decompos... | Summary: The paper proposes to reduce the computational complexity of Toeplitz neural networks. TNNs are a new form of network for sequence modeling that reduces space complexity of the attention matrix to allow for longer sequences.TNN model consists of a stack of Gated Toeplitz Units (GTU) that includes TNO (Toeplitz... | Rebuttal 1:
Rebuttal: Thank you for your comments. We address them based on the relevant section.
# Weaknesses
* While SKI-TNN has slightly worse predictive performance than TNN, in most cases (2/3 of the LRA cases and the bidirectional pre-training), the degradation is minimal, while the speed improvements are drama... | Rebuttal 1:
Rebuttal: We thank the reviewers for their helpful critiques and comments, including “The article creatively combines together a large body of previous work” and “The current paper presents a nice idea.” There were two areas where multiple reviewers had critiques, and we address them here. The first is that... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Isometric Quotient Variational Auto-Encoders for Structure-Preserving Representation Learning | Accept (poster) | Summary: This paper considers representation learning and generative models that can learn symmetry and isometric information about the data manifold. Some types of data, especially image data, can have semantic meanings that are invariant under some group of symmetry transformations. Past work has explored models that... | Rebuttal 1:
Rebuttal: We sincerely appreciate your time and efforts in reviewing our paper, and the constructive comments. Below are our responses to your questions, along with the global response.
***
**Q1. QVAE loss.**
QVAEs, a probabilistic auto-encoder that can learn group invariance, might be seemed to be a sim... | Summary: This paper introduces a new VAE with the modified objective; they call it improved Quotient Variational Auto-Encoder or IQVAE. This new structure-preserving VAE incorporates an additional regularization term, the Riemannian isometry loss, to optimize the decoder. This new term enables the decoder to preserve R... | Rebuttal 1:
Rebuttal: We sincerely appreciate your time and efforts in reviewing our paper, and the constructive comments. Below are our responses to your questions, along with the global response.
***
**Q1. SO(3) group beyond SO(2).**
We would like to emphasize that we have conducted experiments with two datasets (... | Summary: The authors proposed isometric quotient VAEs (IQVAEs), to extract the quotient space and learn the Riemannian isometry. Empirical proof-of-concept experiments are conducted to demonstrate the performance of the method.
Strengths: A new Auto-Encoders approach is proposed and its effectiveness is demonstrated t... | Rebuttal 1:
Rebuttal: We sincerely appreciate your time and efforts in reviewing our paper, and the constructive comments. Below are our responses to your questions.
***
**Q1. Details of the downstream classification tasks.**
For the downstream classification model, we utilized the default hyper-parameters provided ... | Summary: This paper proposes an autoencoder framework with additional features of (simultaneously) (1.) An invariance/equivariance of the latent space to a chosen group symmetry (2.) a near local isometry of the latent space. The authors build on the Quotient Autoencoder (QAE) framework that accounts for the group symm... | Rebuttal 1:
Rebuttal: We sincerely appreciate your time and efforts in reviewing our paper, and the constructive comments. Below are our responses to your questions, along with the global response.
***
**Q1. Combination of already known components.**
We would like to emphasize that our approach of combining both sym... | Rebuttal 1:
Rebuttal: Dear Reviewers,
We would like to express our heartfelt appreciation for taking the time to review our paper. Your insightful comments and suggestions have been tremendously valuable to us. Below are our responses to some of the frequently asked questions you've shared.
***
**Q1. Mix-up.**
We u... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
RRHF: Rank Responses to Align Language Models with Human Feedback | Accept (poster) | Summary: The paper introduces a new learning paradigm called RRHF for aligning large language models with human preferences. The authors propose a ranking-based approach that leverages sampled responses from various sources to align the probabilities of generated responses with human preferences. They demonstrate that ... | Rebuttal 1:
Rebuttal: W1: Thanks for your suggestion. We will add related works of instruction-tuning including SELF-INSTRUCT.
W2: The results of Online Sampling RRHF (OP RRHF) have good reward scores and converge quicks while it tends to reward-hacking. Online sampling RRHF without KL term improves (based on reward) ... | Summary: This paper proposes RRHF that ranks responses to align language model with human feedback. contributions include:
- algorithms to leverage various responses from multiple LMs to align with human preferences. drawing connection of viewing LM for both generation and a reward model for scoring.
- the engineering ... | Rebuttal 1:
Rebuttal: W1: We evaluate different sampling settings to show how sampling procedures affect the RRHF performance. We have shown in Table 6 that sampled responses with large rewards lead to better alignment performance. Including iterative sampling (IP) shows that RRHF can boost during iteration. The purpos... | Summary: The paper proposes RRHF, which adapts the BRIO (ranking loss) framework to the LLM alignment problem. RRHF can train a model to match with human rewards via a ranking loss defined by the margin between the normalized likelihood. Experiments show RRHF matches PPO's performance on the Anthropic's Helpful and Har... | Rebuttal 1:
Rebuttal: W1: SFT/RRHF/PPO only generates once during inference, while best-of-n needs to sample n times during inference. Best-of-n slows n times naturally. More discussions in W2.
W2: In this paper, we mainly compare to PPO instead of best-of-n sampling since best-of-n slows n times naturally in inferenc... | Summary: This paper proposes a learning algorithm called RRHF. The idea is to produce an artifact similar to what Reinforcement Learning from Human Feedback (RLHF) produces, except with a simpler learning algorithm (that might not require so many copies of the base model, e.g. for simultaneously the policy, reward, and... | Rebuttal 1:
Rebuttal: W1: We add an experiment based on Alpaca and random sample 4 response for each query on HH dataset. Using RRHF, we obtain PPL of 14.41 and reward score of -0.96. The results outperform beam search+RRHF and random sampling+PPO regarding reward scores of -1.02 and -1.03 respectively. This proves tha... | null | NeurIPS_2023_submissions_huggingface | 2,023 | Summary: This work introduces Reinforcement Learning from Human Feedback (RLHF) to improve alignment between large language models and human preferences, enhancing interactions between humans and models. The proposed method, RRHF (Ranking-based Reinforcement Learning from Human Feedback), efficiently aligns language mo... | Rebuttal 1:
Rebuttal: W1: The requirement of a reward model is not mandatory in RRHF, as it has the capability to learn directly from human preference data. This is highlighted in Figure 1 and Section 3, where we discuss the option of utilizing either a trained neural network reward model or a dataset of human preferen... | Summary: The paper proposes another way to learn from feedback / preference data. They use the log probabilities of the candidates and try to train the model in a way that tries to increase the log probability of the chosen / better candidate and decrease the log probability of the non-chosen candidate.
They show that... | Rebuttal 1:
Rebuttal: W1 & Q1: The contrastive loss does not constitute the primary contribution of our paper. It should be noted that this loss function was not initially proposed by SLiC, but rather follows prior works. Both SLiC and our related work section acknowledge the utilization of similar functions by BRIO [1... | null | null | null | null |
Better Correlation and Robustness: A Distribution-Balanced Self-Supervised Learning Framework for Automatic Dialogue Evaluation | Accept (poster) | Summary: This paper proposes a new framework to enhance prior TDEMs by utilizing a self-supervised learning approach. The authors introduce a simple and intuitive framework called BCR. The BCR framework employs two approaches: TSR, which reconstructs the training set using a heuristic to balance the coherence distribut... | Rebuttal 1:
Rebuttal: We thank the reviewer for the positive comments as well as constructive suggestions! Below, we discuss each of the reviewer's concerns in detail.
> **The main assumption of this paper is that the test set consists of a uniformly distributed coherence, but I am not clear about the justification for... | Summary: This paper introduces a turn-level dialogue evaluation model (TDEM) and focuses on the self-supervised learning (SSL) framework, which has recently demonstrated state-of-the-art (SOTA) performance in open-domain dialogue evaluation.
The authors identify two potential problems with existing SSL methods for tr... | Rebuttal 1:
Rebuttal: We thank the reviewer for the positive comments as well as constructive suggestions! Below, we discuss each of the reviewer's concerns in detail.
> **This paper explores exclusively the narrow topic of turn-level dialogue evaluation. I understand that dialogue evaluation is a crucial research top... | Summary: This paper proposes a distribution-balanced self-supervised learning framework called Better Correlation and Robustness (BCR) to tackle the challenges of weak performance in medium coherence samples and nonuniform score distribution. Specifically, BCR contains an effective training set reconstructing method to... | Rebuttal 1:
Rebuttal: We thank the reviewer for the valuable comments as well as constructive suggestions! Below, we discuss each of the reviewer's concerns in detail.
> **The research challenges tackled by this paper are limited to a specific framework (i.e., SSL) for a specific task (i.e., dialogue evaluation). Thus... | Summary: This manuscript aims at two problems in turn-level dialogue evaluation models (TDEMs), including lower correlations between human judgments and the self-supervised learning (SSL) framework on medium coherence samples, and the nonuniform score distribution. The authors also provide a theoretical analysis to dem... | Rebuttal 1:
Rebuttal: We thank the reviewer for the positive comments as well as constructive suggestions! Below, we discuss each of the reviewer's concerns in detail.
> **The novelty is not very significant. The techniques of TSR are commonly used in dialogue systems.**
We highly value your feedback and would also li... | Rebuttal 1:
Rebuttal: We sincerely thank all the reviewers for their valuable feedback. We have provided detailed responses to these suggestions, and if there are still any questions that have left you confused, we eagerly look forward to providing further clarification for you.
As demonstrated by our presentation, ou... | NeurIPS_2023_submissions_huggingface | 2,023 | Summary: This paper proposes a novel turn-level dialogue evaluation framework, called Better Correlation and Robustness (BCR), which introduces pseudo-negative samples with real-valued pseudo labels to train a model that evaluates responses given a context and query. The authors also propose a new loss function that mo... | Rebuttal 1:
Rebuttal: We thank the reviewer for the positive comments as well as constructive suggestions! Below, we discuss each of the reviewer's concerns in detail.
> **Can we explore alternative strategies to create augmented samples and soft pseudo labels? For instance, [1] considers not only the speaker but also ... | null | null | null | null | null | null |
Mobilizing Personalized Federated Learning in Infrastructure-Less and Heterogeneous Environments via Random Walk Stochastic ADMM | Accept (poster) | Summary: Authors consider in this paper a Federated Learning framework with a mobile server. They propose an algorithm named Random Walk Stochastic Alternating Direction Method of Multipliers (RWSADMM) for networks with instability and limited capacity. To overcome the non-IID nature of clients' data, they ensure local... | Rebuttal 1:
Rebuttal: We appreciate your diligent inquiry and thoughtful comments. In the subsequent sections, we will provide a detailed response to each of your questions.
Weakness 1: model is over simplified:
Answer 1: In our model, we enhance the conventional random walk by introducing a time-inhomogeneous Markov... | Summary: This paper proposes random walk stochastic method based on traditional ADMM, which enables personalized federated learning considering the mobility of clients and using a mobilized server. The authors solve the optimization of the stored client’s model parameters through Lagrangian and dual problem. The theore... | Rebuttal 1:
Rebuttal: We sincerely value your positive review of our work and the words of encouragement you've provided. We are open and receptive to any further insights you might have to offer. We will respond to each question individually.
Weakness 1: Wandering in the dynamic geometric network topology:
Answer 1:... | Summary: This paper proposes a new FL setting that involves a dynamically moving server among clients residing in different regions and receiving updates from clients of the selected region through wireless links. The proposed algorithm RWSADMM addresses personalization and data heterogeneity by minimizing the local pr... | Rebuttal 1:
Rebuttal: Thank you for your follow-up questions and comments. Below we address each question and minor questions in a point by point fashion.
Weakness 1: Realism and Transmission burden:
Answer 1: a) Realism of the Proposed Setting:
Since we know that traditional federated learning may not be practical i... | null | null | Rebuttal 1:
Rebuttal: We extend our gratitude to the reviewers for their insightful feedback, which has provided us an opportunity to clarify and refine our work. In response to the raised concerns, we present a comprehensive overview of our approach's enhancements and differentiators, as well as clarifications on crit... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Optimal Treatment Allocation for Efficient Policy Evaluation in Sequential Decision Making | Accept (poster) | Summary: The paper designs methods for taking actions in multiple sequential settings (Markovian, time-varying Markovian, non-Markovian) to generate high-quality data for the estimation of average treatment effect.
Strengths: 1. Focuses on improving ATE estimation by improving the data collection process, as opposed t... | Rebuttal 1:
Rebuttal: * **Connections with bandits.** This is an excellent comment! In the following, we discuss the similarities and differences between the "explore than commit" strategy in the bandits literature and our proposal.
- **Similarities**: In the "explore then commit" strategy, a system initially navi... | Summary: This paper studies optimal designs for allocating treatments so that treatment effects can be accurately estimated in online experiments. The focus of the paper is designing behavior policies for data generation to minimize the MSE of the ATE estimator when we know that dynamics may evolve through an NMDP, TMD... | Rebuttal 1:
Rebuttal: * **Choice of class $\Pi^b$.**
- Many thanks for this constructive comment! First, we would like to clarify that it is an exceptionally challenging task to directly identify the optimal behavior policy $\{ \pi^b\_t \}\_t$ in TMDPs/MDPs that minimizes the following efficiency bound for estimat... | Summary: The paper studies optimal treatment allocation aiming to maximize the obtained information from online experiments to estimate treatment effects accurately. The authors propose optimal allocation strategies in a dynamic setting. These strategies are designed to minimize the variance of the treatment effect est... | Rebuttal 1:
Rebuttal: * **Related Work on Experimental Design.** Thank you for highlighting these pertinent references! Regrettably, they were overlooked in our initial manuscript submission. Should our paper be accepted, we will certainly incorporate and discuss these references. Specifically:
- Bojinov et al. (ht... | Summary: This paper considers A/B tests which are often employed for evaluating new products/treatments/policies against existing baselines. The goal is to study the optimal design of such A/B tests to maximize the information obtained and estimate treatment effect more accurately. The paper considers three possible se... | Rebuttal 1:
Rebuttal: * **Imbalance in Allocation.** This is an excellent comment. Indeed, a significant variance disparity between the two groups would inevitably lead to a marked imbalance in the treatment allocation. This is because in the design we proposed, the probabilities of treatment assignment are dictated by... | Rebuttal 1:
Rebuttal: * We greatly appreciate all the reviewers' critical assessments of our work, many of which will lead to a more readable and self-contained version of our paper. We have worked very hard to address all comments and concerns, all discussions will be incorporated if our paper gets accepted.
* Rebutt... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Learning Time-Invariant Representations for Individual Neurons from Population Dynamics | Accept (poster) | Summary: This paper introduces NeuPRINT, an unsupervised approach that simultaneously learn (i )a dynamical model of recorded neural activity and (ii) a time invariant embedding of individual neurons using transformers.
Using a downstream transcriptomic classification task, the authors show that the learnt embedding ... | Rebuttal 1:
Rebuttal: We thank the reviewer for the constructive comments on improving the writing and presentation of the paper. Meanwhile, we encourage the reviewer to also check the results in the overall response and attached pdf.
**“The statistics used to build C and S are never defined.”**
We provide the prec... | Summary: In this manuscript the authors describe an approach to predicting the transcriptomic identity of a neuron from its activity dynamics that they call NeuPRINT. Specifically, the authors take a dataset with calcium imaging and transcriptomic identification of the imaged neurons. First, they train up a model – eit... | Rebuttal 1:
Rebuttal: **“Is the boost in performance with NeuPRINT solely coming from the population activity and behavioral covariates?”**
First, to address the reviewer’s concern on the role of population activity in the improved accuracy, we performed multiple new experiments to also add population activity to LOLC... | Summary: Understanding neural dynamics is crucial for studying brain function, but most existing methods assume that all neurons are functionally identical, which is not the case at the biological level. The biological variability among neurons can influence their dynamics, and since this variability is genetically enc... | Rebuttal 1:
Rebuttal: We thank the reviewer for their constructive review.
**“the dataset used only contains spontaneous neural activity”**
We believe this is not a limitation of the proposed method. Our method could be directly applied to neural recordings under different conditions, stimuli, and tasks. To address t... | Summary: The authors evaluate several approaches for identifying neural types and subtypes (e.g., GABAergic cells are commonly divided into several subclasses Lamp5, Pvalb, Vip, Sncg, Sst). They used a dataset from a recently (2022) published neuroscience paper and compared their approaches to a recent study that had t... | Rebuttal 1:
Rebuttal: We thank the reviewer for their constructive review.
**"The evaluation was limited to a single dataset."**
We believe our general response (“Uniqueness of the dataset and study” and “Robustness of the method”) addresses the key points identified as weaknesses by this reviewer. Please refer to th... | Rebuttal 1:
Rebuttal: We thank the reviewers for their detailed comments. We would like clarify two key points that we thought are at the heart of much confusion:
**Uniqueness of the dataset and study**: Population recordings with calcium imaging are typically performed on mouse lines that are selective for a subset o... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Disentangled Wasserstein Autoencoder for T-Cell Receptor Engineering | Accept (poster) | Summary: The paper proposes to disentangle the structural from functional representations when performing machine learning (ML) assisted protein engineering. The authors propose a Wasserstein autoencoder with auxiliary classifier to learn disentangled embeddings for the structural and functional components. Once disent... | Rebuttal 1:
Rebuttal: Thanks for your comments. Here we respond to your points and questions in the sequential order:
**1. Explicit and implicit**
By explicit and implicit, here we are talking about the nature of the latent space. For some generative models, such as diffusion models, there is no “explicit” latent s... | Summary: In many protein engineering problems, we want to maintain the overall fold of the protein while changing its functional properties. This is analogous to style transfer in computer vision. The paper uses a disentangled Wasserstein autoencoder to learn separate functional and structural embeddings for T-cell rec... | Rebuttal 1:
Rebuttal: Thanks for your comments. Here we respond to your points and questions in the sequential order:
**1. Limited scope**
We focus on TCR here because it is easier to define a quantifiable “function” (binding vs non-binding). For many other scenarios, such a definition might be ambiguous. Also, there... | Summary: In this work, the authors develop an approach to T-cell receptor (TCR) generation using disentangled representation learning (DRL). They introduce a disentangled Wasserstein autoencoder with an auxiliary classifier, and show through a series of experiments that they can separate the “structural” and “functiona... | Rebuttal 1:
Rebuttal: Thanks for your comments. Here we respond to your points and questions in the sequential order:
**1. Validity metrics**
We follow the same evaluation metrics as in [1]. Specifically, with a trained autoencoder, we use (1) reconstruction accuracy (2) the likelihood in the distribution of the embe... | Summary: This paper proposes to use disentangled representations for protein engineering which enables preserving the generic structure while making functional modifications to the protein of interest. To disentangle functional patterns from the rest, they propose to use disentangled Wasserstein auto encoders with auxi... | Rebuttal 1:
Rebuttal: Thanks for your comments. Here we respond to your points and questions in the sequential order:
**1. Additional results for Table 1**
The other six peptides were discarded because of poor classification (with AUC < 0.8). As we rely on the classifier to control the information content of the fun... | Rebuttal 1:
Rebuttal: We really appreciate the reviewers’ valuable comments. Here we would like to address some common concerns that several reviewers have brought up.
We have also attached a pdf with the figures for the rebuttal, and refer to them as “Attached Fig XX”.
**1. The usage of “structure embedding”**
We a... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Context-lumpable stochastic bandits | Accept (poster) | Summary: The paper considers contextual bandit problem with S contexts and K actions, under
the assumption that the contexts can be lumped into $r \leq \min\{S,K\}$ groups such that the mean reward for the various actions is the same for any two contexts that are in the same group. The paper considers the task of lear... | Rebuttal 1:
Rebuttal: We thank the reviewer for their valuable comments. Below we address the concerns raised by the reviewer.
- **Comparison with [1], and (3) and (4) seem trivial:** The paper [1] the reviewer mentioned falls into the category of “oracle-based” contextual bandits. In oracle-based contextual bandits,... | Summary: This work proposes a PAC learning algorithm and a regret minimization algorithm in the case when the contexts can be lumped into smaller partitions of contexts. The PAC algorithm achieves minimax sample complexity, while the regret minimization algorithm achieves a regret upper bound that could be less than th... | Rebuttal 1:
Rebuttal: We thank the reviewer for providing their valuable feedback and raising important questions. Below we answer the reviewers’ questions.
- **Knowledge of $r$:** We agree that the requirement of knowing r is not very satisfactory. We have some ideas on removing this requirement (see the paragraph u... | Summary: This paper considers the problem of context-lumpable bandits, which consists of contextual bandits with $K$ actions and $S$ contexts that can be partitioned into $r \\le \\min\\{S, K\\}$ blocks. The authors consider the case when mean rewards of the actions are fixed over different contexts that belong to the ... | Rebuttal 1:
Rebuttal: We thank the reviewer for their detailed review and insightful comments. Getting a gap-dependent regret bound is undoubtedly an interesting question. As the reviewer points out, our algorithm detects large gaps and eliminates suboptimal arms accordingly. Therefore, assuming the rank is constant an... | Summary: This paper studies the context-lumpable stochastic bandits, which can be understood as stochastic contextual bandits with low rank structure. A specific case of this problem where the contexts are linear was studied in previous literature, but there wasn't much study on this general problem. For this problem, ... | Rebuttal 1:
Rebuttal: We thank the reviewer for their constructive feedback. Below we reply to the reviewer’s comments and questions.
- **$\widetilde{\Omega}$ and $\lg$:** We thank the reviewer’s suggestion on the notations. We will change them accordingly.
- **Stochastic contexts and suboptimal regret:** We agree th... | null | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Reward Scale Robustness for Proximal Policy Optimization via DreamerV3 Tricks | Accept (poster) | Summary: This paper explores implementation details from Dreamer-v3 in the context of PPO to examine the generality of these implementation tricks to DeepRL. They identify 5 unique implementation details from Dreamer-v3 and do a thorough examination of its effects on both continuous and discrete control environments (i... | Rebuttal 1:
Rebuttal: Thank you for taking the time to thoroughly consider our work. Based on your feedback, we will make the following clarifications and additions to the paper:
Hyperparameters: Symlog does not have any hyperparameters and we don’t experiment with changing the unimix ratio from 0.01. The Critic EMA d... | Summary: This paper aims to assess the generality of DreamerV3's stabilization techniques by adapting them to the widely-used model-free algorithm PPO. The authors then proceed to conduct a thorough ablation study across Atari games and the DeepMind Control Suite to evaluate their effectiveness.
Strengths: - The paper... | Rebuttal 1:
Rebuttal: Thank you for your detailed review of our work. We have no objections and will fix the minor writing details, thank you for pointing those out. As a point of discussion, we would like to add some context as to the impact of our work. The DreamerV3 paper was received with significant and sustained ... | Summary: This paper empirically tests the effectiveness of the DreamerV3 tricks on PPO. The authors conduct a comprehensive study on each trick, especially, extensive ablation studies, including add-one and drop-one ablations. Based on these experiments, the authors give some insights into the effect of the DreamerV3 t... | Rebuttal 1:
Rebuttal: Thank you for your time and effort in reviewing our work. We certainly agree that there are areas where our focus on PPO limits generalization to other RL algorithms, but we believe it is important to emphasize the significance of PPO as an algorithm. There is a large body of work aimed at studyin... | Summary: The authors investigate the architectural and algorithmic improvements proposed in the DreamerV3 paper, but applied to PPO. The tricks are each compared on the DMC and Atari suites.
Strengths: Thorough benchmark papers are a valuable asset to the RL community due to the difficulty in accurately comparing them... | Rebuttal 1:
Rebuttal: Thank you for your thorough review and suggestions. First, we would like to clarify a bit of nuance surrounding the paper. The five algorithmic tricks were presented by DreamerV3 as techniques to improve performance and decrease/eliminate the need for hyperparameter tuning. This sparked a tremendo... | Rebuttal 1:
Rebuttal: Thank you to all of the reviewers for their comments and suggestions about our work. We’d like to address our choice of PPO because several reviewers suggested other potential starting algorithms for this work (DreamerV3, A2C, SAC).
To start, this work is not a reproduction of DreamerV3 but inste... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Generative Neural Fields by Mixtures of Neural Implicit Functions | Accept (poster) | Summary: This paper proposes a new representation of neural fields based on a mixture of neural fields. Neural fields is another name for implicit neural representation (INR) which is a representation of data points (like images or 3d scenes), that maps an input of 2d or 3d position into the output of the RGB color and... | Rebuttal 1:
Rebuttal: Thank you for complimenting our modulation scheme, mixtures of neural implicit functions for neural fields. Following part, we will describe the properties of our model to solve the raised concerns by the reviewer.
## W1 Comparison to previous work
In our study, we focus on algorithms for gene... | Summary: This paper deals with the problem of learning the latent space of data through the concept of meta-learning INRs (or neural fields). In contrast to prior works that consider either the adaptation of whole neural field parameters or some linear modulation parameters, the proposed method considers some basis neu... | Rebuttal 1:
Rebuttal: Thank you for complimenting the proposed model, experiments and writing quality. Our model can achieve both flexibility on model capacity and inference efficiency based on linear combination of model weights of neural fields, which is related to question 2. We answered your questions as below. We ... | Summary: This paper proposes an approach to increase the learning capacity of implicit neural networks (INRs). The idea is to define a set of basis implicit functions that can be mixed to produce an instance-specific INR for each instance in a dataset. Notably, the approach does not affect the inference performance as ... | Rebuttal 1:
Rebuttal: Thank you for emphasizing the strengths of the proposed method and complimenting the current version of the work.
## Q1. additional interpolation plots
We visualize additional inter-class interpolation results in Figure B of attached one-page PDF. We conduct bilinear interpolation with samples f... | Summary: The paper proposes a mixture of neural implicit functions for generating neural fields via model averaging of implicit representations trained through context adaptation (meta-learning/auto-decoding paradigm) and subsequently task-specific generalization (additional sampling strategy with respect to the latent... | Rebuttal 1:
Rebuttal: We are glad to hear your positive comment. We will reflect the minor corrections you mentioned in the final version. The other concerns will be answered as below.
> Experiments lack quantitative evaluation/baseline. The authors are requested to clarify/highlight baseline models in proposed evalua... | Rebuttal 1:
Rebuttal: We sincerely appreciate your valuable comments to improve our work. We presume that there are no strong negative comments and we earnestly responded to your concerns; please see the respective comments. Before answering your questions and concerns, we would like to highlight our contributions by q... | NeurIPS_2023_submissions_huggingface | 2,023 | Summary: This paper introduces an approach for learning generative neural fields using linear combinations of implicit basis networks, called Mixture of Neural Implicit Functions (mNIF). The basis networks and their coefficients in mNIF can be learned through meta-learning or auto-decoding techniques. By this means, th... | Rebuttal 1:
Rebuttal: Thank you for your constructive comments on improving the writing and analysis of the proposed method. We answered your two concerns on missing baselines and analysis on inference cost as below.
## W1. missing related works and baselines
### INR with mixture of experts
Thank you for suggesting t... | null | null | null | null | null | null |
Better Private Linear Regression Through Better Private Feature Selection | Accept (poster) | Summary: the submission proposed a private feature selection algorithm that can be directly applied as a preprocessing step before any private linear regression algorithm. The algorithm itself is based on private top-k selection with the Kendall rank correlation, which has a constant sensitivity. Further experiments on... | Rebuttal 1:
Rebuttal: > It would be great if the authors could provide recommendations on when to use the proposed feature selection and how to set the number of features to maintain
We agree that more nuanced guidance for applying DPKendall would be useful. As a first cut, a reasonable heuristic may be to use it when... | Summary: This work studies a differentially private feature selection method based on Kendall rank correlation as plug-and-play to make differentially private linear regression more suitable to high-dimensional problems without requiring users to set data bounds or algorithmic hyperparameters. This work provides e a ut... | Rebuttal 1:
Rebuttal: > The current version of the paper provides the computation cost of DP-Kendall, DP-Lasso, and standard OLS. I wonder if it is possible to compare the utility guarantee with previous works.
Ideally we would have a more direct comparison between DPKendall and SubLasso, but this is difficult, as the... | Summary: The paper under consideration delves into the challenge of performing linear regression under differential privacy, especially in scenarios where the data is unbounded. This issue arises primarily due to the current private linear regression mechanisms that require tuning a set of hyperparameters for unbounded... | Rebuttal 1:
Rebuttal: > Could you provide a precise definition of what constitutes a "plug-and-play" algorithm?
See the response to Comment 1 in the general response.
> The proposed method necessitates the selection of parameter k, and also the budget split. How does this aspect impact the method's practical applicab... | Summary: This paper studies the problem of feature selection for linear regression subject to differential privacy constraints. One of the benefits of this work is that the feature selection does not require setting bounds on features a priori, thus allowing for fewer hyper-parameters which are difficult to set in pra... | Rebuttal 1:
Rebuttal: > The application of the exponential mechanism which balances the two desiderata (features that correlate with label and features that do not correlate with already selected features) is nice. Can other works for feature selection (e.g. PCA) be summarized in a similar way?
In some sense, yes. For... | Rebuttal 1:
Rebuttal: We thank the reviewers for their comments. A few comments appeared in multiple reviews, so we address them in this "global" response; we would be happy to add similar clarifying language to the next version of the paper. Other specific comments are addressed in the responses to individual reviews.... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
SlotDiffusion: Object-Centric Generative Modeling with Diffusion Models | Accept (spotlight) | Summary: This paper presents SlotDiffusion, an object-centric Latent Diffusion Model (LDM) designed for image and video data. The authors focus on improving slot-to-image decoding, a critical aspect of visual generation. SlotDiffusion outperforms previous slot models in unsupervised object segmentation and visual gener... | Rebuttal 1:
Rebuttal: Thank you for your review and insightful comments.
**1. Are performance gains mainly due to the use of diffusion models?**
A: Yes, the performance gains indeed mainly come from the use of an LDM-based slot decoder, as diffusion models have both strong modeling capacity and generation quality. W... | Summary: The paper presents, SlotDiffusion, an object-centric model that incorporates a latent diffusion model as an image decoder. Extracted slots are used as a conditioning signal for the cross-attention layers of the diffusion UNet. Other than that, the model follows the formulation of SA/STEVE/SAVi models. SlotDiff... | Rebuttal 1:
Rebuttal: We thank the reviewer for the constructive comments. We are glad to see the positive assessment of our paper and appreciate the detailed feedback.
**1. Parameter & size comparison of SlotDiffusion and baselines.**
A: We list the parameter size of SlotDiffusion and baselines on the MOVi-E dataset... | Summary: A latent diffusion-based slot decoder is introduced to improve the segmentation and reconstruction quality for slot-based autoencoders. This decoder addresses issues with previous advanced Transformer-based decoders such as SAVi and STEVE which i) do not preserve the spatial dimension of images and ii) lack an... | Rebuttal 1:
Rebuttal: We thank the reviewer for the positive feedback and valuable comments.
**1. Scaling to real-world data requires complex efforts and designs.**
A: We would like to note that the weights of DINO ViT and the architecture of VQ-VAE are directly adopted from prior works with minimal modifications. A... | Summary: The paper proposes a diffusion-based decoder for slot-centric models. Previously slot-centric models are trained with spatial broadcast or transformer autoregressive decoders, however these decoders result in a comparitively poor image generation or segmentation. The paper shows that using diffusion based deco... | Rebuttal 1:
Rebuttal: We thank the reviewer for their detailed feedback and encouraging comments.
**1. Performance of spatial broadcast decoder based models in Table 1 (video prediction on Physion).**
A: We train SlotFormer over slots extracted by SAVi, and report its video prediction results below. SlotDiffusion + ... | Rebuttal 1:
Rebuttal: We would like to thank the reviewers for their helpful feedback and insightful comments.
We are glad that the reviewers find our paper “*well written*” (QakT), “*easy to follow*” (GFM9), and “*clear*” (jjQP), our diffusion-based slot decoder as “*sound*” (GFM9) and “*novel*” (jjQP). Also, our exp... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
AttrSeg: Open-Vocabulary Semantic Segmentation via Attribute Decomposition-Aggregation | Accept (poster) | Summary: The paper proposes a decomposition-aggregation framework to address the problem of open-vocabulary semantic segmentation. The framework consists of a decomposition stage, where pre-trained language models (LLMs) such as ChatGPT are used to extract textual attributes for class names, and an aggregation stage, w... | Rebuttal 1:
Rebuttal:
## Reviewer PB8W
Thank you for your valuable comments and suggestions, as well as acknowledging the novelty and results of the idea.
We address individual questions and concerns below.
> **1. It would be beneficial to explore ways to regulate the risk introduced during the decomposition stage... | Summary: [Task] This paper presents a novel decomposition-aggregation framework for open-vocabulary semantic segmentation, which addresses the challenges of segmenting novel object categories during inference. The existing vision-language pre-training approaches make unrealistic assumptions about low-quality textual ca... | Rebuttal 1:
Rebuttal: ## Reviewer mZMs
Thank you for your appreciations on the novelty, contribution, and results of the idea, as well as the detailed and constructive review.
We address individual questions and concerns below.
> **1. compare with some latest works**
Thanks for the suggestions, we have added the... | Summary: Previous work does not address the issue of ambiguity or potentially incomplete textual categories and directly incorporates them into vision-language models without any post-processing. This approach hinders the text encoder's ability to fully comprehend the category. To overcome this limitation and provide c... | Rebuttal 1:
Rebuttal:
## Reviewer fEW5
Thank you for your appreciations on the novelty, motivation, and results of the idea, as well as the detailed and constructive review.
We address individual questions and concerns below.
> **1. Missing citations**
Thanks for pointing out.
We will add these literatures in rela... | Summary: This paper proposes a framework for decomposing class names into attributes for open vocabulary semantic segmentation. The authors obtain attributes from LLMs or hand-curate for unheard-of classes, then aggregates these attributes for obtaining a general descriptor for each classes. Moreover, the authors colle... | Rebuttal 1:
Rebuttal: Thank you for acknowledging our novelty and motivation, we address your kind concerns below.
> **1.1. COCO/PASCAL have well-known classes without neologism etc**
There may be some misunderstandings. We admit COCO/PASCAL have well-known classes. However, the aim of using them are **_not_** intend... | Rebuttal 1:
Rebuttal: ## General Response and Clarifications
We kindly thank the reviewers for their detailed reading, and considerate feedback. We are also grateful of the appreciations by the reviewers in these aspects:
(1) **Novel and well-motivated idea**: "motivation seems solid" (R-cgtT); "sound motivation and ... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
PrimDiffusion: Volumetric Primitives Diffusion for 3D Human Generation | Accept (poster) | Summary: This paper introduces a method to generate various 3D human geometry and appearance. To this end, it introduces a new design of diffusion model that learns to generate the parameters of 3D primitives of humans: a generalizable encoder constructs motion and appearance features and their attentions are fed to pr... | Rebuttal 1:
Rebuttal: Thank you for your careful review and constructive suggestions. We summarize and answer your questions as follows:
**Q1: Fundamental motivation of the proposed method.**
Please kindly refer to the above post named "Response to the common issue of the novelty and motivation".
**Q2-1: Clarificati... | Summary: The paper proposes a generative model for 3D humans. The work is based on volumetric rendering of geometric primitives and diffusion model. Given multi-view images of humans as training data, the method first learns a volumetric primitives representation, and then applies diffusion model framework to model the... | Rebuttal 1:
Rebuttal: Thank you for your careful review and constructive suggestions. We summarize and answer your questions as follows:
**Q1: The bibliography needs to be better formatted.**
Thank you for the great suggestion. We will revise all references in a unified format.
**Q2: Fidelity evaluations of ground t... | Summary: The paper presents a diffusion-based approach, PrimDiffusion, for 3D human generation. The method leverages the representation of MVP (mixture of volumetric primitives) and a design of volumetric primitives diffusion to generate high-quality 3D human models using multi-view images as training data. The authors... | Rebuttal 1:
Rebuttal: Thank you for your careful review and constructive suggestions. We summarize and answer your questions as follows:
**Q1: The proposed method is a 2-stage approach that requires multi-view images.**
Thanks for your feedback. Indeed, this is a limitation of our method which we will discuss in our ... | Summary: The paper presents, PrimDiffusion, an approach for generative modeling of 3D humans. PrimDiffusion is a diffusion-based model that operates on a set of volumetric primitives representing the human body. Each volumetric primitive consists of rotation, translation, scale, radiance, and color information. PrimDif... | Rebuttal 1:
Rebuttal: Thank you for your careful review and constructive suggestions. We summarize and answer your questions as follows:
**Q1: Novelty of the proposed method.**
Thanks for your comments. We highlight our key contributions in the above post named "Response to the common issue of the novelty and motivat... | Rebuttal 1:
Rebuttal: **Common response**
We sincerely thank all reviewers for the constructive comments and recognition of our work. We are encouraged that reviewers find our proposed method "innovative" (Reviewer oDDi), "novel" (Reviewer PWaT), and "effective" (Reviewer GMsy); our results "have superior performance"... | NeurIPS_2023_submissions_huggingface | 2,023 | Summary: The paper introduces a novel framework called PrimDiffusion, which utilizes diffusion models to generate 3D human models. The framework leverages a Mixture of Volumetric Primitives (MVP) to represent 3D humans, enabling efficient rendering and incorporating human priors for retargeting. This representation off... | Rebuttal 1:
Rebuttal: Thank you for your careful review and constructive suggestions. We summarize and answer your questions as follows:
**Q1: The reasons behind the omission of delta translation and delta rotation.**
Thanks for your comments. There are two reasons for predicting delta scale factor only: 1) better ge... | null | null | null | null | null | null |
H3T: Efficient Integration of Memory Optimization and Parallelism for Large-scale Transformer Training | Accept (poster) | Summary: Large language models have made significant advancements in terms of accuracy and representativeness across various applications. However, their large memory consumption and intensive computation size pose critical challenges for existing hardware platforms.
This paper proposes a memory-latency co-optimizatio... | Rebuttal 1:
Rebuttal: Thanks for your thorough and contributive review. Here are the replies to your questions:
1. There are limited details about the proposed solvers, especially the DP solver.
We are sorry for the lack of details in the main paper. But actually, we show many details about our solvers in the appendi... | Summary: For efficient training of large language models, how parallelism and memory optimization are jointly configured is essential. Since the search space is too broad to explore, each has been studied separately. However, the separate search for memory optimization and parallelism could lead to a suboptimal solutio... | Rebuttal 1:
Rebuttal: Thanks for your thorough and contributive review. Here are the replies to your questions:
Q1: Although the authors say that H3T covers model parallelism including pipeline parallelism, the paper does not explain how pipeline parallelism is incorporated.
A1:
We are sorry for your misunderstandi... | Summary: This paper presents a framework to automatically find an optimized integration of memory and parallelism for training large language models (LLMs). To this end, this paper first formulates the memory optimization strategies and parallelism techniques, and then solves the equation to find the best configuration... | Rebuttal 1:
Rebuttal: Thank you for your appreciation of our work. We notice that your primary concern is about the configuration of DeepSpeed used in our experiment. We briefly describe it in the appendix (section C.3, line 108): "For fairness, we turn on optimizer offloading and rematerialization for Megatron-DeepSpe... | Summary: This paper considers large model training and proposes a search algorithm for various memory optimizations.
Strengths: Please see the "Questions" section.
Weaknesses: Please see the "Questions" section.
Technical Quality: 2 fair
Clarity: 2 fair
Questions for Authors: - Lines 89-91: In what sense are you u... | Rebuttal 1:
Rebuttal: Thanks for your thorough and contributive review. Here are the replies to your questions:
1. The misuse of "dynamic" in lines 89-91.
Thanks for pointing out this mistake. Our point is that these approaches are based on static rules which cannot dynamically adapt to the changes in model structure... | null | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Byzantine-Tolerant Methods for Distributed Variational Inequalities | Accept (poster) | Summary: **Summary:** The authors propose robust optimization methods for solving distributed variational inequalities (VI). The authors provide theoretical guarantees for numerous popular algorithms for solving min-max (more generally VI) problems. The authors corroborate their theoretical findings with experiments on... | Rebuttal 1:
Rebuttal: We thank the reviewer for the detailed feedback and a positive evaluation of our work. Below we address the reviewer’s questions and concerns.
>**Large batch sizes requirement.**
Part of our results (Theorems 1-3) indeed require large batches: for SGDA-RA and SEG-RA this limitation cannot be byp... | Summary: This paper studies convergence guarantees of Byzantine tolerant methods in the context of distributed variational inequalities.
Strengths: The paper seems to contain a lot of work.
Weaknesses: In my opinion, the main weakness of the paper is its writing. I acknowledge that I have not read the papers most rel... | Rebuttal 1:
Rebuttal: We thank the reviewer for the detailed feedback. Below we address the reviewer’s questions and concerns.
>**The main weakness of the paper is its writing.**
Thank you for your multiple suggestions and comments. We will apply all necessary changes to the final version of the paper to make it easi... | Summary: This paper studies the problem of Byzantine-robust distributed learning with variation inequalities, e.g., max-min form. Three different robust methods for this problem are developed from the concepts of robust aggregation, client momentum, and checks of computations, respectively. For each of the proposed met... | Rebuttal 1:
Rebuttal: We thank the reviewer for the feedback and a positive evaluation of our work. Below we address the reviewer’s questions and concerns.
>**Motivation of tackling variational inequalities.**
In our work, we are mostly motivated by min-max problem formulations that form a quite broad class of proble... | Summary: In this paper, the authors provide robust variants of some popular optimization algorithms in the presence of Byzantine workers. They also use client momentum and random checks of computations to enhance the performance further. They provide theoretical guarantees of converges and empirically test their method... | Rebuttal 1:
Rebuttal: We thank the reviewer for the feedback. Below we address the reviewer’s questions and concerns.
>**Concerns about the novelty.**
Our paper indeed relies on the existing methods like SGDA and SEG for variational inequalities and existing techniques for achieving Byzantine robustness studied for m... | Rebuttal 1:
Rebuttal: We thank the reviewers for their feedback, time, and positive evaluation of our work. We appreciate that the reviewers acknowledged several strengths of our work:
- The paper is well-written (Reviewers zd6L and RWvG)
- The ideas are presented clearly (Reviewer zd6L)
- The problem is interesting... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
AUDIT: Audio Editing by Following Instructions with Latent Diffusion Models | Accept (poster) | Summary: This paper proposes a zero-shot generative editing model for audio signals. The text-based user edits include adding, dropping, replacing, inpainting, and perceptually upsampling audio signals. The audio is not limited to any particular audio domain (e.g., speech or music). Comparisons to recent text-to-audio ... | Rebuttal 1:
Rebuttal: Thanks for your careful and valuable comments! We will explain your concerns point by point.
> **1. This is not “the first audio editing model based on human instructions” (line 43-44) as any editing model is technically following the user’s instructions. Consider “the first model to perform sema... | Summary: This work proposes a new editing technique to audio data.
Strengths: The authors introduce a new task and a base method to tackling it.
Weaknesses: **Conditioning:** The authors concatenate the input audio to the sampled noise, why did you pick this option on top of sampling a Gaussian around the input samp... | Rebuttal 1:
Rebuttal: Thanks for your careful and valuable comments! We will explain your concerns point by point.
> **1. conditioning.**
Thank you found the alternative conditioning approaches helpful. Using Gaussian sampling around the input sample as the starting point for diffusion denoising (replacing the conven... | Summary: This paper proposes AUDIT, an instruction-guided audio editing model based on latent diffusion models. AUDIT is trained on triplet data consisting of an instruction, an input audio, and an output audio. The instruction is used to guide the diffusion model to only modify the audio segments that need to be edite... | Rebuttal 1:
Rebuttal: Thanks for your careful and valuable comments! We will explain your concerns point by point.
> **1. Please add ablation analysis for each component.**
We have supplemented the impact of the text encoder on the performance of the text-to-audio generation model.
| Text Encoder | FD($\downarrow$) ... | Summary: This paper proposes AUDIT, an instruction-guided audio editing model based on latent diffusion models. The model is trained with triplet training data (instruction, input audio, output audio) for
different audio editing tasks and is based on a diffusion model using instruction and input
(to be edited) audio as... | Rebuttal 1:
Rebuttal: Thanks for your careful and valuable comments! We will explain your concerns point by point.
>**1. It is unclear how it compares to task-specific models such as NU-Wave2 [1] for audio super-resolution.**
Thanks for your valuable suggestion! Since our approach is based on a diffusion model, our p... | Rebuttal 1:
Rebuttal: Thank you very much for the valuable suggestions and insightful questions you've provided! We have addressed each reviewer's concerns point by point.
Furthermore, we would like to reiterate some common points of discussion (of course, we have also responded individually to each reviewer's specifi... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Neural Graph Generation from Graph Statistics | Accept (poster) | Summary: This paper proposes a deep graph generative model from aggregate graph statistics, to protect privacy. The authors designed an inference model to learn latent variables from graph statistics. And the probabilistic model keeps permutation invariance.
Strengths: 1. To learn the intractable latent variables, the... | Rebuttal 1:
Rebuttal: We appreciate the reviewer's attention to the paper. Our responses to Weaknesses (W) and Questions (Q) are provided below; all references and citations refer to the paper.
**W1.** We agree that calculating graph statistics has computational overhead. However, the training time of GenStat, *which ... | Summary: This paper considers a new setting that learns a deep graph generative model from aggregate graph statistics rather than graph adjacency matrix. The motivation for generating graphs from statistics is privacy preservation. For this task, the authors propose a simple VAE model, where the encoder takes graph sta... | Rebuttal 1:
Rebuttal: We appreciate the reviewer's attention to the paper. Our responses to Weaknesses (W) are provided below; all references and citations refer to the paper.
**W1.a and W1.b. Task Setting and Application of Generated Graphs.** As we explain in related work (lines 55-61 and Figure 5), we start with *r... | Summary: The paper proposes a new approach to learning deep graph generative models that can generate more realistic graphs while protecting local privacy. The proposed GenStat architecture uses graph statistics instead of the entire adjacency matrix to train deep generative models.
Strengths: - The paper presents a n... | Rebuttal 1:
Rebuttal: We appreciate the reviewer's attention. Our responses to Weaknesses (W) and Questions (Q) are provided below; all references and citations refer to the paper.
**W1.Privacy Issues in Standard GGMs:** We provide a clear motivation for the problem in the introduction (lines 23-38). Graph Generative ... | Summary: The authors introduce a deep graph generator model based on graph statistics. Instead of encoding the graph itself, deterministic high-level graph statistics are first computed from the graph, and these are passed to a neural encoder and graph decoder. The stated main motivation for this approach is for increa... | Rebuttal 1:
Rebuttal: We appreciate the reviewer's positive feedback, especially regarding their observation that "*The contribution itself is interesting and likely to impact graph learning beyond the specific application (graph generation) presented in the paper.*"
**Questions.** Can the authors comment on my observ... | null | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
A Long $N$-step Surrogate Stage Reward for Deep Reinforcement Learning | Accept (poster) | Summary: The article considers a new way to formulate the n-step rewards in deep reinforcement learning. Specifically, while the normal n-step reward discount the Q-value by \gamma^n, the authors propose to only discount the Q-value by \gamma while using the normal n-step reward (times a constant) for r. For a simplifi... | Rebuttal 1:
Rebuttal: We greatly appreciate the reviewer's comments and suggestions. We'll properly address the questions in our revision if the paper is accepted.
>Q1. The author selected a few tasks for a few algorithms
Thank you for this important comment. We believe that the number of tasks and the baseline algo... | Summary: This paper introduces a new stage reward estimator called the long N-step surrogate stage (LNSS) reward for deep reinforcement learning (RL). The LNSS aims to mitigate the high variance problem in deep RL, which has been shown to impede successful convergence of learning, hinder task performance, and limit the... | Rebuttal 1:
Rebuttal: We greatly appreciate the reviewer's critical input, which has helped us to think about LNSS from a general perspective in order to make it more clear and highlight its novelty.
>Q1 compare with common $n$-step method, the difference is only the larger $n$? lack originality
Thank you for the que... | Summary: The paper propose a new reward estimator for off-policy RL to reduce the variance of the Q-value target estimate in bellman backup. Specifically, the method works by computing the discounted return of a trajectory segment and then augment such reward along with the original transition (state, action, next stat... | Rebuttal 1:
Rebuttal: We greatly thank the reviewer for reading our paper and for the favorable comments. Your questions have helped us further improve the paper.
> Q1 About estimation bias and learning performance and the LNSS reward estimate assumes the trajectory is filled with a constant reward
Please refer to o... | Summary: Authors proposed a long N-step surrogate stage (LNSS) reward for DRL which mitigates the high variance problem in continuous control problems. It is shown that the proposed LNSS can provide consistent performance improvement measured by average reward, convergence speed, learning success rate, and variance red... | Rebuttal 1:
Rebuttal: We greatly thank the reviewer for reading our paper and for the favorable comments. Your questions have helped us further improve the paper.
>Q1 How to justify smaller N not working as well as larger N? Also how to show this in baseline comparisons (using n=5 and N=5) and ablation studies (N=5, ... | Rebuttal 1:
Rebuttal: # General Response: Insights on LNSS, its essence, and novelty
We first thank all reviewers full-heartedly for their insightful comments. We will integrate some key points below into the revision if our paper is accepted.
1) LNSS is about a new, surrogate reward $r_k’$, which relies on multiple (l... | NeurIPS_2023_submissions_huggingface | 2,023 | Summary: This paper proposes a simple but effective surrogate reward, which uses the future n-step cumulative rewards (with some coefficient).
The theoretical results show that the proposed method has a smaller variance than the classical one-step method.
The experimental results show the effectiveness of the proposed... | Rebuttal 1:
Rebuttal: We thank the reviewer for carefully reviewing our paper and for the penetrating questions. We also thank the reviewer for recognizing that our paper addresses an important problem, and our LNSS is simple and potentially very effective.
>Q1 Learns exactly the same solution as the one using the ori... | null | null | null | null | null | null |
LayoutPrompter: Awaken the Design Ability of Large Language Models | Accept (poster) | Summary: This paper presents a set of few-shot techniques to enable large language models to generate geometric layouts, particularly in the domain of User interfaces, posters and documents. These techniques include dynamic prompting and layout ranker modules, which allows the LLMs to learn from relevant examples in th... | Rebuttal 1:
Rebuttal: **Weakness1. Reliant on existing dataset's coverage of the target task to achieve good performance.**
To understand the effect of dataset's coverage on model performance, we regard the size of the layout pool that prompt UIs can be sourced from as a proxy for coverage, and we study how the model ... | Summary: The authors propose LayoutPrompter, which is a new approach for generating high-quality, cross-domain graphic layouts without any model training or fine-tuning. It can handle a variety of conditional layout generation tasks and outperforms existing methods on some metrics. The approach is competitive with stat... | Rebuttal 1:
Rebuttal: **Weakness1. The details of WebUI dataset are missing.**
Thanks for your advice. We'll include a section in the supplementary material to describe the dataset details.
Here, we first provide you with the information.
1) Source.
We mainly collect the company homepages of the Fortune 500 companies... | Summary: In this paper, the authors introduce LayoutPrompter, a novel approach for conditional graphic layout generation using
large language models with minimal demonstrations.
The authors frame the graphic layout generation task as a sequence generation problem. They represent layouts as HTML
files, leveraging the f... | Rebuttal 1:
Rebuttal: **Limited novelty in techniques.**
Leveraging large language models for layout generation is by no means an easy task.
It requires a deep understanding of the layout domain and the according innovations, which can be grounded in our three design choices.
First, layout representation.
Due to the i... | Summary: This paper leverages large language models to solve the layout generation tasks. It shows a great performance when generalizing to other domains or other tasks. Since in-context learning is used, there is no need to train the model on a vast amount of data.
Strengths: 1. The prompting method shows a good perf... | Rebuttal 1:
Rebuttal: **Weakness1. How other LLMs perform with the proposed method.**
Initially, we tried Alpaca 7B.
However, the model is prone to generate heavily overlapping layouts, and sometimes the output is not even in the right format.
We then switched to text-davinci-003 and were successful.
We agree that it ... | Rebuttal 1:
Rebuttal: First, we thank the reviewers for recognizing our simple yet effective work that enables LLMs to generate graphic layouts in a training-free manner. We also appreciate all valuable and constructive comments. We attach our added quantitative and qualitative results here.
Pdf: /pdf/71c2683cd285fdd73... | NeurIPS_2023_submissions_huggingface | 2,023 | Summary: The authors propose a new conditional layout generation method using LLMs with few-shot learning. The method uses a pre-trained GPT-3 with two techniques: dynamic prompting and layout ranker. Dynamic prompting is to provide few-shot examples with similar constraints to the input query using ConsSim, and layout... | Rebuttal 1:
Rebuttal: **Weakness1. It will be better to compare with other content-aware layout generation works, such as CanvasVAE and CGL-GAN.**
We clarify that content-aware layout generation refers to generating layouts based on the image content, i.e., avoid putting graphic elements on top of main objects (e.g., ... | null | null | null | null | null | null |
Harnessing the power of choices in decision tree learning | Accept (poster) | Summary: The paper is about a new method for machine learning using decision trees. Namely, the authors propose an algorithm that can span the spectrum of the methods that are currently used:
(i) traditional decision tree learning algorithms that choose features and splitting points based on a criterion that is optimi... | Rebuttal 1:
Rebuttal: Reviewer 8rCb Comment 1: The paper is not based on any striking new idea. While the novelty is limited, it is nevertheless interesting.
Response to Reviewer 8rCb Comment 1: We believe the simplicity of our ideas, combined with theoretical performance guarantees and the fact that these ideas work ... | Summary: In the paper authors propose a new algorithm that allows balancing between two approaches to decision tree learning: fully greedy optimization, and exactly optimal decision tree construction. Authors discuss the long standing line of research that focuses on either ends of this balance, and propose a novel al... | Rebuttal 1:
Rebuttal: We thank Reviewer B1BN for their insightful comments and helpful feedback.
Reviewer B1BN Comment 1: The paper focuses on a binary classification setting. And features are also binary in the theorems and definitions. This was confirmed in the experimental section, where categorical features have b... | Summary: The paper presents a novel generalization of classic greedy decision tree algorithms called Top-k, which considers recursively the top k attributes for possible splits rather than the single best attribute. It enhances traditional decision tree learning algorithms such as ID3, C4.5, and CART, and is expected t... | Rebuttal 1:
Rebuttal: We thank Reviewer UM1K for their insightful comments and helpful feedback.
Reviewer UM1K Comment 1: While Top-k demonstrates stronger performance than classic greedy algorithms for the same tree depths, an additional, more fair setup would be comparing the test accuracy achieved in comparable tra... | Summary: The paper studies an extension of greedy decision tree algorithms. The main idea is, instead of simply greedily selecting the best feature, to consider several features. This can also be seen as a restriction of optimal decision tree algorithms, so in some sense the studied approach lies somewhere in between g... | Rebuttal 1:
Rebuttal: We thank Reviewer mdbV for their helpful feedback. We provide a point-by-point response to Reviewer mdbV’s comments below.
Conceptual
W1: We have synthesized a body of additional related work and would be happy to provide it in an additional “Official Comment” if requested. We have currently omi... | Rebuttal 1:
Rebuttal: Attached please find new experimental results which we refer to in our rebuttals to reviewers.
Pdf: /pdf/03bb1a4cd71a148c58685ec90fba0f6991c623a5.pdf | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Auditing Fairness by Betting | Accept (spotlight) | Summary: The paper proposes a safe anytime-valid inference (SAVI) testing procedure to audit the potential unfairness of a classifier or regression model. The paper posits group-conditioned fairness as a test of equality of means under some finite set of conditions, this definition includes several common fairness noti... | Rebuttal 1:
Rebuttal: We thank the reviewer for their interest in the paper!
We agree that exact equality of means may be too precise a condition in practice (though we note that most other group fairness papers work with these definitions). It is in fact possible to handle a null of the form $H_0: |\mu_1-\mu_0|<\eps... | Summary: This work introduces practical and efficient methods for auditing the fairness of deployed classification and regression models. The methods are sequential, allowing for continuous monitoring of incoming data, and can handle distribution shifts resulting from changes in the model or the underlying population. ... | Rebuttal 1:
Rebuttal: We thank the reviewer for posing several thoughtful questions.
Regarding the weaknesses, Section 4 now includes several new experiments that employ various data collection policies on a new dataset (an insurance dataset) and with a new model (logit). We test three different data collection polic... | Summary: This work proposes an approach based on recent results of super-martingales to detect when a certain predictor is demographically unfair. Studying different formulations of group fairness (such as equal opportunity, demographic parity, etc), the authors frame the problem as that of sequential hypothesis test, ... | Rebuttal 1:
Rebuttal: We thank the reviewer for their enthusiasm regarding the work! To address the main weakness, we have added experiments for various data collection policies. We discuss the details in point 3 below, and respond to the other questions in order:
1. We can handle equalized odds by playing two games... | Summary: This paper clarifies the issue of fairness auditing, wherein an auditor verifies the fairness of a system through sequential interaction with it. The goal is to develop a methodology that facilitates anytime-valid tests of system fairness. Essentially, this involves devising a sequence of test statistics for t... | Rebuttal 1:
Rebuttal: We thank the reviewer for their kind words on the paper.
Regarding time-varying data collection, we agree that the density, $\rho_t$, will not always be known. However, we had in mind situations in which knowing the density is feasible. Consider, for instance, algorithms deployed across a databa... | Rebuttal 1:
Rebuttal: Thanks to all reviewers for their thoughtful questions and comments!
Pdf: /pdf/a0cc141eb16187c0f1c71a1fbc507fb2876e73f4.pdf | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
PICProp: Physics-Informed Confidence Propagation for Uncertainty Quantification | Accept (poster) | Summary: The paper presents methods to estimate confidence sequences for the solution of a differential equation given noisy boundary conditions. Various methods are proposed of different complexity and the confidence sequences are illustrated for the solution of three different problems.
Strengths: The overall probl... | Rebuttal 1:
Rebuttal: Thanks for your thoughtful feedback. We will address your comments point by point below, and we hope that our answers can ease your concern.
**To Weakness 1:** In terms of not considering epistemic uncertainty, the reviewer is right and we have already clarified in the paper that we consider only... | Summary: The paper focuses on a novel algorithm for confidence interval estimation for deterministic partial differential equations. It provides leverages bi-level optimization to compute a valid CI, that is supported by both theoretical and numerical investigations, focusing on physics-informed learning.
Strengths: T... | Rebuttal 1:
Rebuttal: Thank you for your thoughtful and overall positive feedback. Here, we provide answers to your concerns point by point.
**To Weakness 1 and Q1**: We thank the reviewer for pointing out the absence of some results. We itemize our response to this question as follows:
(a) We will show SimPICProp r... | Summary: Authors present a method to propagate confidence intervals from some locations to the entire domain. In specific, they consider the problem of propagating uncertainty / confidence internals at the boundary, in partial differential equations based systems, to the rest of the domain. The proposed method PICProp ... | Rebuttal 1:
Rebuttal: Thank you for your thoughtful and overall positive feedback. Here, we provide answers to your concerns point by point.
**To Weakness 1:** The dimensionality of PDE/ODE problems is typically N+1, where N is the number of spatial dimensions and 1 accounts for the temporal dimension. In this regard,... | Summary: The paper discusses the limitations of current approaches for uncertainty quantification in deep learning and physics-informed learning, such as the need for strong assumptions regarding data likelihood and the computational cost associated with approximating the posterior. The paper introduces a novel problem... | Rebuttal 1:
Rebuttal: Thank you for your thoughtful feedback. Here, we provide answers to your concerns point by point.
**To Q1:** We assume that the reviewers are referring to the exhaustive search (ES) method as the baseline, of which the cost is acceptable.
We would like to clarify that the cost of ES is acceptab... | null | NeurIPS_2023_submissions_huggingface | 2,023 | Summary: The paper introduces a PICProp, a method to propagate uncertainty through spatiotemporal nonlinear PDE systems. PICProp (1) is more valid and makes less assumptions compared to NN based UQ methods such as dropout and deep ensembles, and (2) can be modified to be more computationally efficient by learning a met... | Rebuttal 1:
Rebuttal: Thank you for your thoughtful feedback. Here, we provide answers to your concerns point by point.
**To Weakness 1:** We would like to clarify that none of the statements in our paper are intended to give readers the impression that "wider CI = safety." In fact, in practical applications, construc... | null | null | null | null | null | null |
Bandit Task Assignment with Unknown Processing Time | Accept (poster) | Summary: The paper studies the *bandit task assignment* problem.
In this problem, a player chooses tasks sequentially,
each task taking some random time to complete and giving some random reward.
Both random quantities are sampled from a priori unknown distributions.
In each round, the player may start a new task, as l... | Rebuttal 1:
Rebuttal: > The techniques in the paper seem to be mostly based on prior work, though I think this is a minor point.
Technical challenges in this paper are summarized in Lines 109--124. To cope with the challenge due to conflicting start and end times of tasks, we employed a phased-update approach, instead... | Summary: This paper introduces a novel bandit learning scenario, in which the learner can select in any time step tasks, each of which leads to stochastic rewards and processing time that blocks the task for several rounds. At any round, the tasks in progress (i.e., both newly started or still running ones) are constra... | Rebuttal 1:
Rebuttal: > In Thm. 5.1, the gap-free bound is of the form $O(X + \frac{C_u^2}{C_l^2}NM \ln T)$. However, in the later discussion (e.g. in line 326), this second summand seems to be ignored. Why? Note that its dependence on $M$, $N$ and $C_u$ is worse than that of term $X$. And in line 355 you explicitly me... | Summary: This paper addresses thel problem of bandit task assignment, where a set of tasks need to be sequentially selected with uncertain processing times. The paper presents a UCB-type algorithm designed for bandit task assignment and provides theoretical analyses on upper bounds of the proposed algorithm and a regre... | Rebuttal 1:
Rebuttal: > There exist related works that have not been addressed in the work, including: Learning to control renewal processes with bandit feedback, S Cayci, A Eryilmaz, R Srikant, 2019. Budget-constrained bandits over general cost and reward distributions, S Cayci, A Eryilmaz, R Srikant, 2020.
Thank you... | Summary: The paper introduces a new bandit problem where at each round the agent choses a subset of tasks to run, where the length of tasks and their rewards are unknow and revealed only after the task is done. A UCB style algorithm is proposed and shown to achieve almost-optmial regret bounds.
Strengths: 1. The paper... | Rebuttal 1:
Rebuttal: We sincerely appreciate the time and effort you have invested in this review.
> The presentation could improve by moving the related works and side discussions to the later sections and instead focusing on the main setting introduction. It is hard to understand the setting since it is long laid ... | Rebuttal 1:
Rebuttal: We deeply appreciate the reviewers' thoughtful and comprehensive comments and feedback.
The reviewers' insights have greatly contributed to the improvement of our work, and we sincerely appreciate the time and effort you have invested in this review.
We hope our response below addresses your conce... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Visual Explanations of Image-Text Representations via Multi-Modal Information Bottleneck Attribution | Accept (poster) | Summary: Visual-language models (VLMs) are increasingly used for multiple vision-language problems like object recognition, image classification, image captioning, etc. Formally, these models are trained in a self-supervised fashion using large datasets comprising millions of image-text pairs, such that they learn gene... | Rebuttal 1:
Rebuttal: Thank you for your thoughtful and constructive questions and suggestions!
We agree that **"one of the main highlights of the work is that it does not require the ground-truth label"** and were pleased you found our method **"clear"** and **"simple"**.
Please let us know if you have any remaining... | Summary: This paper aims to enhance the interpretability of vision-language pretrained models like CLIP. To achieve this, the authors propose a multi-modal information bottleneck (M2IB) objective that reduces irrelevant and noisy information while retaining important visual and textual features. M2IB is used for attrib... | Rebuttal 1:
Rebuttal: Thank you for your thoughtful and constructive questions and suggestions!
We were pleased to read that you agree that our paper **"is tackling a significantly important research question"** and that you found our paper to be **"well-structured, with clear organization and sufficient methodologica... | Summary: This paper introduces a feature attribution method for vision language models. The method is based on previous works on information bottleneck attributions. It incorporates Gaussian moment-matching estimator techniques to make the objective function tractable. The experimental results show that the method out... | Rebuttal 1:
Rebuttal: Thank you for your thoughtful and constructive questions and suggestions!
We were pleased that you appreciated that our contribution is **"novel"** and that our evaluation uses a set of **"diverse evaluation metrics"**.
We conducted several additional experiments to address your comments.
Pleas... | Summary: This paper proposes an adjusted information-bottleneck-based approach for interpreting multi-modal models. Specifically, the approach was appropriately adapted from the information bottleneck method that has been widely used for interpreting image models. Through various approximations, the method was successf... | Rebuttal 1:
Rebuttal: Thank you for your positive review and for your thoughtful and constructive questions and suggestions!
We were pleased to read that you found that **"qualitative results demonstrate the method's ability to capture the mutually related information between text and images"**, that **"the quantitati... | Rebuttal 1:
Rebuttal: # General Comment to All Reviewers and Area Chair
We thank all reviewers for their thoughtful and constructive feedback!
## Summary
In this general response, we will
1. clarify that **the proposed method, M2IB, does not require (re-)training a pretrained model**: The parameters of the pretrain... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Prototype-based Aleatoric Uncertainty Quantification for Cross-modal Retrieval | Accept (poster) | Summary: The paper focuses on the unreliable prediction of cross-modal retrieval methods caused by low-quality data. The authors introduce evidential theoretical framework to quantify the uncertainty arising from inherent data ambiguity. To provide trustworthy uncertainty predictions, they propose a novel Prototype-ba... | Rebuttal 1:
Rebuttal: We thank the reviewer hqQV for the positive, valuable and detailed review, as well as the constructive suggestions for improvement. Our responses to the reviewer’s questions are below:
***Q1 : Some details about the number of training epochs, how to choose parameters, and code are scarce.***
**A... | Summary: This paper addresses the aleatoric uncertainty quantification in cross-modal retrieval. The authors point out that the model predictions are often unreliable due to the aleatoric uncertainty, which is induced by low-quality data. They first define the inherent aleatoric uncertainty in multi-modal data. Then K ... | Rebuttal 1:
Rebuttal: We would like to thank the reviewer b4gq for providing professional and valuable feedback, as well as encouraging comments. In what follows, we wish to address some of the questions you raised as below.
***Q1 : How learnable parameters $\beta_1$ and $\beta_2$ are learned in PAU?***
**A1:** Thank... | Summary: This paper focuses on improving video-text retrieval by considering the uncertainty (or confidence) when computing image-text similarity. The high-level idea is to estimate the data uncertainty and then perform re-ranking to improve the performance of retrieval.
The authors designed a PAU framework, Prototype... | Rebuttal 1:
Rebuttal: We thank the reviewer tGbU for providing constructive feedback. In what follows, we wish to address the questions you raised as below.
***Q1 : The authors seem to argue that a single-scene video should match only a single word (i.e., talking), which is counter-intuitive.***
**A1:** We are sorry ... | Summary: This paper proposes prototype-based aleatoric uncertainty quantification (PAU) to represent aleatoric uncertainty (i.e., data uncertainty) for cross-modal retrieval tasks. The proposed method uses the CLIP backbone, where the initial similarity matrix M is computed by the output of the CLIP encoders. Then, "ev... | Rebuttal 1:
Rebuttal: We thank the reviewer VBVJ for the positive, patient, professional review and the valuable suggestions for improvement. Our responses to the raised questions are below:
***Q1 : Compared with probabilistic-based uncertainty estimation.***
**A1:** Thanks for your professional and valuable comments... | Rebuttal 1:
Rebuttal: # Common Response #
We appreciate all reviewers for their kind reviews and recognizing our work is helpful to multi-modal learning.
According to reviews, **highlights of this paper include**:
- The problem this paper addressed is interesting, novel, and important. (**VBVJ, tGbU, b4gq, hqQV**)
- ... | NeurIPS_2023_submissions_huggingface | 2,023 | Summary: This paper presents the Prototype-based Aleatoric Uncertainty Quantifications (PAU) framework, which enhances the reliability of a pre-trained vision-language model (CLIP) by leveraging evidential knowledge. The paper introduces learnable prototypes as semantic categories and maps images and texts to these cat... | Rebuttal 1:
Rebuttal: We thank the reviewer qbVB for the valuable and constructive feedback. We will address the raised questions point-by-point as below.
***Q1 : The loss function defined in Equation 5 raises confusion. Why uncertainty computed by K category domain and mean similarity computed by N category domain ar... | null | null | null | null | null | null |
Zero-Regret Performative Prediction Under Inequality Constraints | Accept (poster) | Summary: The paper provides a new algorithm for performative prediction under general inequality constraints. The proposed algorithm has sublinear regret with respect to the performative optimum if the distribution map follows the location family model. The theory is supported by simulations.
Strengths: The motivation... | Rebuttal 1:
Rebuttal: Thanks a lot for your review and for recognizing our contribution to the research field of performative prediction. Our point-to-point responses to concerns on weaknesses and questions are given below:
**Reply to Weaknesses:**
**1.** Thank you for the comment. The main contributions of this pape... | Summary: The paper attempts to study the problem of learning where the problem instances arise from a distribution. What is special is that this distribution is dependent on the decision-maker. The specific problem is when the distribution is given by a linear shift over a fixed distribution where the degree of shift i... | Rebuttal 1:
Rebuttal: Thank you for your review and for recognizing our contribution to the research field of performative prediction. Our point-to-point responses to concerns on weaknesses and questions are given below:
**Reply to Weaknesses:**
1. > The paper promises a study in the direction of general optimization... | Summary: The paper studies the problem of performative prediction under inequality constraints and gives an no-regret learning algorithm that obtains $O(\sqrt{T})$ regret and uses $\sqrt{T} + 2T$ samples.
In the problem of performative prediction, the data distribution $Z\sim D(\theta)$ depends on the choice $\theta$ ... | Rebuttal 1:
Rebuttal: Thank you for your review and for recognizing our contribution to the research field of performative prediction. Our point-to-point responses to concerns on weaknesses and questions are given below:
**Reply to Weaknesses:** Thank you for the comment. The primal-dual method is widely used in the l... | Summary: This paper studies performative prediction under inequality constraints. In particular, the paper develops a robust primal-dual framework that requires only approximate gradients up to a certain accuracy but achieves the same order of performance as the stationary stochastic primal-dual algorithm even without... | Rebuttal 1:
Rebuttal: Thank you for your review and for recognizing our contribution to the research field of performative prediction. Our point-to-point responses to concerns on weaknesses and questions are given below:
**Reply to Weaknesses:**
1) >The major weakness of the paper is the assumptions are very strong... | null | NeurIPS_2023_submissions_huggingface | 2,023 | Summary: This paper studies the problem of performative prediction under inequality constraints. The authors propose an adaptive, robust primal-dual algorithm that achieves sublinear regret and constraint violations, where the regret is with respect to the performative optima (instead of just a performative stable poin... | Rebuttal 1:
Rebuttal: Thank you for your review and for recognizing our contribution to the research field of performative prediction. Our point-to-point responses to concerns on weaknesses and questions are given below:
**Reply to Weaknesses:**
**1. Reply to Weakness 1:** Thank you for the comment. Our algorithm on... | null | null | null | null | null | null |
(Out-of-context) Meta-learning in Language Models | Reject | Summary: This work introduces the phenomenon of out-of-context meta-learning in large language models (LLMs). It demonstrates, through carefully designed experiments, that LLMs have the ability to internalize the semantic content of the text that appears to be from a reliable source.
Specifically, they focus on explor... | Rebuttal 1:
Rebuttal: $\let\n\newcommand \n{q}{\mathtt{QA}}\n{\d}[3]{\mathtt{#1 D^{#2cons}_{#3}QA}} \n{\do}{\mathtt{\dot D}}\n{\a}{\alpha}$Thanks for the detailed feedback and for highlighting parts of our presentation you found unclear! We understand that our experimental setup is intricate, and want to make our work ... | Summary: The paper shows the existence of a phenomenon that the authors refer to as out-of-contect meta learning in large language models. The authors design experiments that show that this phenomenon causes the internalization of text that is broadly useful, meaning that the LLM is more likely to treat this content as... | Rebuttal 1:
Rebuttal: Thank you for you review. We are glad you found the phenomenon and our explanations for it interesting and sound, and our paper well-written and easy to follow.
We also believe we're in full agreement on the limitations of this work, given that they overlap with our limitations section. We think ... | Summary: The paper argues for the existence of ‘out-of-context meta-learning’ as a characteristic of LLMs. The authors support this claim with cleverly designed experiments on QA using a 2.3B parameter pretrained Pythia model. They argue that the presented experiments demonstrate out-of-context meta-learning. They perf... | Rebuttal 1:
Rebuttal: Thank you for your very thoughtful, clear, and in-depth response. We want to ensure that our work is clear to you and to other readers, and are eager to address any issues with the current framing and experiments. We also want to ensure we understand your position correctly.
We believe your simp... | Summary: The authors assert that "out of context metalearning makes LLMs better at internalizing useful information for understanding." They intuitively frame understanding as "treating content as true in question answering." They analyze its application to tasks such as mapping novel phrases or words to attributes and... | Rebuttal 1:
Rebuttal: Thank you for your review! We are glad you liked our experimental setup, found our analysis detailed, and our results interesting.
Regarding the functional decision theory citation, we agree that a reference to a peer-reviewed venue would be more appropriate. There is indeed a more appropriate ci... | Rebuttal 1:
Rebuttal: We would like to thank all reviewers for their constructive reviews, and are looking forward to further discussion. Summarizing the contents of the attached response PDF:
- A table explaining notation and data subset structure, as suggested by reviewers 1 and 3, and aimed at improving clarity as ... | NeurIPS_2023_submissions_huggingface | 2,023 | Summary: The authors introduce the phenomenon of internalization in LLMs, specifically "weak internalization" and "strong internalization" (out-of-context meta-leanring). Weak internalization refers to LLMs' improved performance on questions with consistent definitions rather than with inconsistent ones. Strong interna... | Rebuttal 1:
Rebuttal: Thank you for a thoughtful review and suggestions. We'll aim to address and resolve them below:
> It is unclear if the size of a model affects the internalization phenomenon. Ablations of a few models [of] varying size would help [...].
We agree that an ablation of the model size against the ext... | null | null | null | null | null | null |
Learning from Visual Observation via Offline Pretrained State-to-Go Transformer | Accept (poster) | Summary: This paper proposes a novel approach to deal with the "learning from expert visual observation" setting, lacking action labels. The proposed method first pretrains State-to-Go Transformer (STG), which takes the sequence of encoded visual observations as inputs and predicts next-step embeddings under causal mas... | Rebuttal 1:
Rebuttal: We thank the reviewer for careful review and constructive suggestions, and for praise our originality "incorporating sequence modeling and temporally-aligned regressor into LfVO" "novel and insightful". We address the questions and concerns below.
`Q1.`"Maybe ELE + temporally-aligned regressor wo... | Summary: This paper studies the problem of learning from visual observation. While most adversarial imitation learning methods learn intrinsic reward online, in order to improve the sample-efficiency of online learning, this work proposes to offline pre-train a state-to-go prediction transformer and a discriminator, bo... | Rebuttal 1:
Rebuttal: We thank the reviewer for careful review and valuable feedback, and for the positive comments that “offline-pretraining framework for adversarial imitation learning is novel and intuitive”. We address the questions and concerns below.
`Q1.` "More empirical evaluations to support the superiority o... | Summary: This paper presents a new method, called state-to-go (STG) transformer, for learning from visual observation (LfVO), which aims to learn strong policies given the access to (i) offline datasets without actions and rewards and (ii) following online interaction. The main idea of STG transformer is to learn repre... | Rebuttal 1:
Rebuttal: We thank the reviewer for detailed comments and valuable feedback, and for pointing out the novelty of our pre-training scheme. We provide a point-by-point response below.
`Q1.`"Include more investigations, e.g., experiments with pre-trained encoders for all the methods." and "Investigate the imp... | Summary: Paper presents a novel method for learning from visual observations. Specifically, a state encoder is first trained in an offline fashion with two tasks: 1) predicting the normalized relative distance between two states; 2) predicting the relative difference between two consecutive states. Then the whole agent... | Rebuttal 1:
Rebuttal: We thank the reviewer for detailed comments and insightful review. We are encouraged that the reviewer finds "the idea technically sounds well-motivated in distinguishing states" and "our rationales are clear". We address the questions and concerns below.
`Q1.`"Include some baselines based off be... | Rebuttal 1:
Rebuttal: We sincerely thank the four reviewers for their thoughtful comments! We have completed some supplements according to reviewers' suggestions, and we summarize the major changes as follows:
1. We have added three baselines including IDM-based method **BCO**, "advanced-GAIfO" **AMP** and IRLfO metho... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
On the Asymptotic Learning Curves of Kernel Ridge Regression under Power-law Decay | Accept (poster) | Summary: This paper provides a rigorous analysis of the asymptotic rates of kernel ridge regression. Building on works which analyzed the Gaussian design case, this work recovers the same rates using less severe restrictions on the distribution of features. The novel contribution here beyond similar works of Li et al 2... | Rebuttal 1:
Rebuttal: Thanks for your positive feedback on our work.
### Weaknesses:
> **Obtained learning curves pre-exist in the literature**
>
> The rates obtained match prior non-rigorous analyses and so do not exhibit *new* behaviors over what has been studied in the prior works. That said, the rigorous proof un... | Summary: This paper analyzed generalization error rates for the Kernel Ridge Regression (KRR) problem. The authors make use of several generic assumptions, including (i) eigenvalue decay, (ii) Embedding index, (iii) source condition, and (iv) Holder continuity of the kernel. It is worth noting that assumptions (i), (ii... | Rebuttal 1:
Rebuttal: Thanks for your positive feedback on our work.
### Weaknesses:
> *"1. The contribution section appears to be highly ..."*
**Our response:**
We agree that our contribution is indeed relatively technical and thanks for your helpful suggestions.
Following your advice, in the new revision we w... | Summary: The authors focus on the learning curve of kernel ridge regression, an important topic in theoretical machine learning, which is also used to approximate the generalization ability of 'lazy trained/NKT regime' neural networks.
The authors provide a nearly full characterization of the learning curve in the se... | Rebuttal 1:
Rebuttal: Thanks for your positive feedback on our work.
Questions:
> *1."I would be interested to know more about the limitation of the predictions. Removing the Gaussian assumption is quite a feat, but are they any restrictions on the data structure except the power law distribution/embedding dimension?... | Summary: The paper studies the asymptotics of kernel ridge regression learning curves when the kernel eigenvalues and the ridge parameter both obey a power law.
Strengths: 1. The paper is readable and well written.
2. It derives asymptotic results for KRR for a specific choice of kernel eigenvalues and regularizatio... | Rebuttal 1:
Rebuttal: Thanks for your carefully review.
The KRR learning curve, recently proposed in Bordelon et al. [2020], Cui et al. [2021], aims to depict the interaction between the source condition ($s$), capacity ($\beta$), the choice of the regularization parameter ($\lambda$), and the noise level ($\sigma^... | null | NeurIPS_2023_submissions_huggingface | 2,023 | Summary: This paper studies the generalization error/excess risk of the kernel ridge regression (KRR). Using the classical bias-variance decomposition, the authors gives bounds on the bias term, which is claimed to be tighter than previous literatures. Together with the variance term results from another paper "Yichen... | Rebuttal 1:
Rebuttal: > My biggest concern would be that this paper is merely an incremental paper from "Yicheng Li, Haobo Zhang, and Qian Lin. Kernel interpolation generalizes poorly, March 2023a."
>
> According to the remarks 3.3 and 3.5, the main theorems 3.2 and 3.4 basically recover the results from previous works... | null | null | null | null | null | null |
Strategic Distribution Shift of Interacting Agents via Coupled Gradient Flows | Accept (poster) | Summary: The authors consider a the problem where the change in a distribution for an objective can be modeled as a set of coupled nonlinear parabolic PDEs. These methods have nonlocal interactions and describe a model of the influence of the model on the population and vice-versa. Additionally, these correspond to the... | Rebuttal 1:
Rebuttal: Thank you for your detailed, specific, and very helpful feedback on our work.
**Question 1:** Thank you for catching this; we had a typo. The equation should be $$\underset{z\in\mathbb{R}^d}{\operatorname{argmin}}J(z,x)$$ because this is modeling an individual agent, whose data is given by $z$, o... | Summary: This paper studies the long-term behavior of a dynamical system where there are distribution shifts in response to algorithmic decision-making. The authors derive PDEs to capture distributional changes and describe the result of the algorithm for both the coorperative setting and the competitive setting. Under... | Rebuttal 1:
Rebuttal: Thank you for your helpful feedback and questions.
**Weaknesses:** We agree that more of a discussion about insights would improve our paper and we can add this in the next version. One key insight is the identification of a large classes of input functions that lead to exponential convergence wi... | Summary: This paper presents a framework for analyzing dynamics of distribution shift that captures feedback loop between learning algorithms and the distributions that they are deployed on using a coupled differential equations model. The paper considers both co-operative and competitive settings and suggests that a l... | Rebuttal 1:
Rebuttal: Thank you for your review and feedback.
**Questions:** In comparison with prior work that addresses learning in the face of distribution shift, we can analyze with the methods used in [MPZ21], wherein the learning algorithm fits a linear mean-shift model to the algorithm parameters. Our framework... | Summary: The paper proposes a new framework for analyzing the interaction and feedback loop between learning algorithms and the distributions on which they are deployed. The framework in the paper is more general than prior models, allowing to capture more complex interactions and shifts in the algorithm and distributi... | Rebuttal 1:
Rebuttal: Thank you for your detailed response and helpful comments.
**Model Optimization:** The setting where the dynamics are not exactly minimizing a cost functional is an interesting idea for future work; our work can be seen as a necessary first step for building more general frameworks in this direc... | Rebuttal 1:
Rebuttal: See figures in attachment.
Pdf: /pdf/f866e917a390229e65bdbfd2464591e4d0fe8b05.pdf | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Reward Imputation with Sketching for Contextual Batched Bandits | Accept (poster) | Summary: This paper focuses on contextual batched bandit (CBB) problems, where only partial information (the rewards of executed actions) is available. The authors proposed an efficient reward imputation approach in CBB, which completes the unobserved reward, and the new objective can be solved by the imputation regula... | Rebuttal 1:
Rebuttal: # Response to Reviewer et1K
Thank you for your reviews and suggestions.
**1. Response to Weakness**
We appreciate the reviewer's insightful suggestion regarding the high-level idea. At first glance, the concept of reward imputation might appear counter-intuitive, as imputed rewards are derived ... | Summary: This paper uses imputation to improve the learning of the reward parameters, even for nonexecuted steps. This results in better variance and empirical benefits over many existing algorithms.
Strengths: 1. The empirical benefit seems quite strong over existing algorithms on real datasets.
2. The proposed idea... | Rebuttal 1:
Rebuttal: # Response to Reviewer mYi5
We are grateful for the time you took to review our paper and grasp the gist of its content.
**1. Response to Weakness 1: Notations**
We have checked and refined the order of symbol introductions and will include a new notation introduction in the final version. In ... | Summary: This paper studies the contextual batched bandit (CBB) problem, where the learner interacts with the environment in an online manner through $N$ episodes. In episode $n$, the learner selects the policy $p_n$, receives $B$ contexts $s_{n,b}\in \mathbb{R}^d$, decides the action $A_{I_{n,b}}$ according to $p_n(... | Rebuttal 1:
Rebuttal: # Response to Reviewer wPev
Thanks for your comments and suggestions.
**1. Response to Weakness 1: Regret Comparison**
In the discussion of the regret bound theory, we compare our proposed "policy with reward imputation" to a "policy without the reward imputation" that corresponds to the Batch ... | Summary: This paper studies the contextual batched bandits (CBB) setting, where compared to classical online CB, a batch of actions is played and the reward being observed in an episode. Motivated from the benefit of full-information feedback, it tries to perform the reward imputation for unlabelled data, and optimizes... | Rebuttal 1:
Rebuttal: # Response to Reviewer PXsH
Thank you for your comments and suggestions.
**1. Response to Weakness 1: Setting of Imputation Rate $\gamma$**
Indeed, we have identified the consideration you mentioned regarding the setting of the imputation rate $\gamma$. We have introduced rate-scheduled version... | Rebuttal 1:
Rebuttal: Thank you for your feedback and valuable suggestions. We have carefully considered and incorporated all the comments and recommendations provided. Additionally, we have uploaded a new PDF file containing supplementary experimental results. We kindly request the reviewers to bring to our attention ... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
VeriX: Towards Verified Explainability of Deep Neural Networks | Accept (poster) | Summary: This paper presents an elegant approach for calculating the optimal robust explanation for a DNN prediction. Intuitively, an explanation is robust if small perturbations of the the features not included in the explanation do not modify the model output. A robust explanation is optimal if small perturbations to... | Rebuttal 1:
Rebuttal: We would like to thank the reviewer for their recognition of our work.
Answers to Questions:
## 1. Scalability of the approach & Evaluation using other verification tools
Due to space limitation, we address this question with all the details in the **“global” response** to all reviewers at the ... | Summary: The paper contributes with proposing a robustness-based method to compute formal explanations that irreducible (minimal subset of input data, often referred as sufficient explantions) for feedforward neural networks (NNs). Formal robustness checker is used in the experiments but the method allows to use any f... | Rebuttal 1:
Rebuttal: Responses to Weaknesses:
## 1&2. Algorithm 1, Theorems 3.3 & 3.4, and Proposition 3.5
We didn’t claim to be the first to discover the so-called deletion-based algorithm. Rather, we explore its use in Algorithm 1 to compute *optimal robust explanations* in the context of formal explainable AI. W... | Summary: This paper proposes a feature sensitivity evaluation approach to enhance neural network explainability. The proposed approach aims to provide an optimal robust explanation, which is translated as a minimal combination of features that preserves invariance in the network's prediction. Neural network verificatio... | Rebuttal 1:
Rebuttal: We would like to thank the reviewer for their recognition of our work.
Answers to Questions:
## 1. Runtime performance & Comparison to other NN verification tools
Due to space limitation, we address this question with all the details in the **“global” response** to all reviewers at the top of t... | Summary: The paper presents a system called VERIX for producing robust explanations and generating counterfactuals for deep neural network decisions. VERIX addresses the problem of explainability in AI by applying constraint solving techniques and a sensitivity-based heuristic to iteratively create explanations and cou... | Rebuttal 1:
Rebuttal: We would like to thank the reviewer for their recognition of our work.
Answers to Questions:
## 1. Scalability of VeriX for more complex models
Due to space limitation, we address this question with all the details in the **“global” response** to all reviewers at the top of this page. Specifica... | Rebuttal 1:
Rebuttal: We would like to thank all reviewers for their comments and suggestions.
Due to space limitations, below in **1** we address the common questions from more than one reviewer about the *runtime performance and scalability* of our approach, in **2** we respond to a crucial misunderstanding from Re... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Federated Conditional Stochastic Optimization | Accept (poster) | Summary: This work proposes the first provably efficient federated conditional stochastic optimization solutions with a biased gradient estimator, including 3 variants: FCSG, FCSG-M (with momentum), and Acc-FCSG-M (with a variance reduction technique). The sample complexity and communication complexity are analyzed for... | Rebuttal 1:
Rebuttal: I really appreciate your time and comments. I would follow your suggestion and add more details in final version.
The technical contribution in the proof that are absent in prior works is listed here:
First, we are the first work to do federated conditional stochastic optimization and we not onl... | Summary: This paper explores different algorithms for optimization problems in the FL setting with 2 layers of stochasticity, where the inner stochasticity is dependent on the outer layer. An interesting application for these algorithms is the MAML problem, where we aim to find a global initialization over clients that... | Rebuttal 1:
Rebuttal: Thanks a lot for your time and valuable review.
> Q1. "The proposed method is only evaluated on problems where it is assumes multiple tasks can be sampled at clients instead of one task being assumed for each client. "
Our paper focus on conditional stochastic optimization and FL MAML is an exa... | Summary: In this paper, the authors delve into the complexities of nonconvex federated conditional stochastic optimization problems. A suite of algorithms, ranging from the simplest form to accelerated versions that meet the ideal lower bounds, is presented. The study also investigates the sample and communication comp... | Rebuttal 1:
Rebuttal: Thanks a lot for your time and comments.
1. "A bunch of assumptions are imposed here for the theoretical analysis, like Lipschitz continuity ... "
As you mention in Q2, we follow the assumptions in the previous work and do not add any extra assumptions. For Assumption Lipschitz continuous objec... | Summary: This paper proposed algorithms for federated conditional stochastic optimization problem. Noteably, the sample complexity is of $O(\epsilon^{-5})$ and the communication complexity is of $O(\epsilon^{-2})$ to obtain $\epsilon$-stationary point. This complexity matches by the single-machine algorithm BSpiderBoos... | Rebuttal 1:
Rebuttal: Thanks for your time and valuable comments.
Q1. "... the reference [13] actually introduced MLMC method that has sample complexity ..."
[13] utilized the Monte Carlo method to achieve better results compared with the vanilla stochastic gradient method. In our paper, we focus on the stochastic g... | null | NeurIPS_2023_submissions_huggingface | 2,023 | Summary: This research delves into the realm of conditional stochastic optimization within the framework of federated learning. It begins by introducing the Federated Conditional Stochastic Gradient (FCSG) algorithm, which utilizes a conditional stochastic gradient estimator. Subsequently, the algorithm is refined by i... | Rebuttal 1:
Rebuttal: We are deeply thankful for your appreciation and valuable comments.
> Q1. "in equation (6), it is unclear which variable or function each gradient corresponds to ..."
Thanks. I will fix these typos in the final version. Since our objective function is a CSO and its gradient follow the chain rul... | Summary: Authors propose methods for problem of stochastic optimisation of composition of functions' expected values, which can be addressed with conditional gradient concept. Federated algorithm, which uses biased gradient estimation, and its modifications with momentum and acceleration are proposed. Theoretical guara... | Rebuttal 1:
Rebuttal: We are deeply thankful for your comments. I will add deviation of trajectory for different runs with different realisations of randomness in our final version.
---
Rebuttal Comment 1.1:
Comment: Dear authors, thank you for your work on the final version of your paper! The rebuttal has clarified ... | null | null | null | null |
Modelling Cellular Perturbations with the Sparse Additive Mechanism Shift Variational Autoencoder | Accept (poster) | Summary: This paper presents a generative model for single cell RNA sequence data that models perturbations as sparse linear offsets in latent space. They derive a variational inference algorithm for their model and evaluate it on a couple of perturb-seq datasets where the perturbations are CRISPR knockouts of genes. T... | Rebuttal 1:
Rebuttal: We thank the reviewer for their deep feedback and thoughtful response about the core of our proposed model. We believe we can address your concerns resoundingly and hope to convince you to raise your score. Please read our general response since it addresses your main concern as well.
**Results v... | Summary: The paper presents an extension of the VAE model by incorporating sparse perturbations of the latent variables, based on the sparse mechanism shift assumption. The proposed method models the latent state of perturbed samples as a combination of a local latent variable capturing sample-specific variation and sp... | Rebuttal 1:
Rebuttal: We thank the reviewer for their feedback and hope to convince them to improve their score:
**Benefits compared to baselines:**
We apologize for the confusion, we must have not explained the results well enough in our experimental section. CPA-VAE is a model we propose to turn CPA into a generativ... | Summary: This paper proposes a new VAE-based model for learning disentangled representation, dubbed Sparse Additive Mechanism Shift Variational Autoencoder (SAMS-VAE). The model is specifically designed for datasets under perturbation; the dataset has the form of $(x_i,d_i)$ where $d_i \in \{0,1\}^T$ indicates the pert... | Rebuttal 1:
Rebuttal: We thank the reviewer for their positive review and their detailed suggestions. We will address their feedback in the following:
**Disentanglement/Self-Contained:**
While we believe that we cite a rich set of papers explaining the setting we are based on, in particular the recent SVAE+ by Lopez ... | Summary: This paper proposes a new method, SAMS-VAE for learning disentangled representations leveraging the sparsity mechanism assumption. SAMS-VAE is bridging the gap between two existing models - CPA and SVAE+, hence resembles both. In simulated and real (single-cell) data experiments, the authors claim improvement ... | Rebuttal 1:
Rebuttal: We thank the reviewer for their positive review and time spent on out paper.
**Derivation of Eq.4,5:**
Eqs 4 and 5 indicate choices of variational families we design to perform inference in the model. As such, they are proposed by us and form a key part of our contribution in modeling the posteri... | Rebuttal 1:
Rebuttal: We thank our 4 reviewers for their feedback and questions.
The reviewers agreed that the paper is sound, interesting, and mostly clear and well-written.
While we addressed many of the individual concerns, we also want to summarize some of the shared discussion.
(I) we received questions about ho... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Convergence Analysis of Sequential Federated Learning on Heterogeneous Data | Accept (poster) | Summary: This paper presents an intriguing exploration of the counterintuitive comparison results between Sequential SL (SSL) and Federated Averaging (FedAvg) in the context of heterogeneous data. The authors attribute the surprising performance of SSL over FedAvg in worst-case scenarios to the assumption of data heter... | Rebuttal 1:
Rebuttal: This part contains 4 responses, denoted by R22-R25.
##### **R22: (Q1) Extend the experimental evaluation to include larger datasets and models.**
As suggested, we have added ResNets (He et al., 2016) [Ref 7] for larger models and CINIC-10 (Darlow et al., 2020) [Ref 8] for larger datasets. The te... | Summary: This paper studies the Sequential SL algorithm in which the at each round, sequentially, each machine performs some local SGD iterations, and then the final iterate is passed on to the next machine. Such an algorithm may be useful in cross-device, resource constrained settings, where local iterations are expe... | Rebuttal 1:
Rebuttal: This part contains 7 responses, denoted by R15-R21.
##### **R15: (W1) The two last paragraphs on page 1 are somewhat confusing.**
As suggested, we will streamline these paragraphs in the revision. We would like to clarify that SFLV1 shares the same update rule with FedAvg despite SFLV1 $\neq$ Fe... | Summary: The authors provides the convergence analysis of a FL-like approach where clients updates the model sequentially based on local datasets. Results are provided for the strongly convex, convex, and non-convex case. Interestingly, the analysis shows that in some setting the sequential approach can outperform the... | Rebuttal 1:
Rebuttal: This part contains 8 responses, denoted by R7-R14.
##### **R7: (W1) Does not deal with split computing.**
Due to space limitation, please refer to R1 (the global response) for details.
---
##### **R8: (W2) No intuitive explanation about why this may be the case.**
*The Intuition about why SSL... | Summary: This paper provides the first theoretical convergence rates for sequential split learning (SSL) in the presence of heterogeneous data. In each round of SSL, a random permutation of the machines is computed. According to this permutation, each machine performs $K$ local gradient steps on the model sequentially.... | Rebuttal 1:
Rebuttal: This part contains 5 responses, denoted by R2-R6.
##### **R2: (W1) Analysis of Split learning without any explicit splits.**
As suggested, we would like to clarify that the operation details (e.g., the form of model split) would not affect the convergence analyses of SSL. To this end, we stated ... | Rebuttal 1:
Rebuttal: ##### **R1: There is no explicit explanation of "Split" Learning (or split computing).**
***Clarification for no explicit explanation of Split Learning***. We would like to clarify that, due to the page limitation, we only state the basic model update rule of Sequential Split Learning (SSL) in Al... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Distribution Learnability and Robustness | Accept (poster) | Summary: This paper studies the interaction between learnability and several notions of "robust" learnability, wherein candidate distributions can be corrupted in various ways. The authors show an equivalence between learnability and a robust version of learnability in which a contaminated distribution can be added to... | Rebuttal 1:
Rebuttal: Thank you for the detailed and positive review! We are glad to hear the reviewer appreciates the breadth and technical quality of the results we present. We acknowledge the fair criticisms made regarding writing, and appreciate the many concrete suggestions offered by the reviewer to remedy these ... | Summary: This paper studies distribution learnability, with a focus on the relationship between the realizable and the agnostic/robust settings. This relationship is also studied in the context of sample compression schemes and differential privacy.
In particular, the authors show separation results: learnability in t... | Rebuttal 1:
Rebuttal: Thank you for the kind review! We are glad the reviewer believes that our results would be of interest to the learning theory community – highlighting the breadth of settings considered and the significant conceptual contribution.
In the following, we address specific points raised by the reviewe... | Summary: This paper study the connection between learnability and robust (agnostic) learnability for general classes of distributions in the framework of distribution learning. Differ from the PAC learnablity, they show that in the distribution learning setting the learnability does not always implies agnostic learnabl... | Rebuttal 1:
Rebuttal: Thank you for the kind review! Regarding your questions:
**On the definitions of adaptivity:**
In Definition 1.4 (1), the sample $S'$ of size $\eta m$ is arbitrary. Similarly, the sample $S’$ in (3) of the same definition is obtained by arbitrarily replacing a fixed fraction of instances in the ... | null | null | null | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Retinotopy Inspired Brain Encoding Model and the All-for-One Training Recipe | Reject | Summary: The authors make two claimed contributions:
1. They propose an architecture for an encoding model. This architecture consists of three key components.
* A frozen DINO trained ViT-B backbone + learnable convolutional layers on top of ViT feature maps
* A differentiable spatial sampling layer (implemented via py... | Rebuttal 1:
Rebuttal: - RE: On clarity
“details on all-for-one training”: see G2.
“do you share weights”: see G4. We share the backbone for voxels (in each veROI model), while the last linear regression is unique for each voxel. In addition, each voxel’s input feature $\mathbf{h}_i$ is affected by RetinaMapper and L... | Summary: This paper develops an all-for-one training model to address the challenge of one big-model problem by converting it into multiple small models, in which the small models aggregate the knowledge while preserving the distinction between the different functional regions. With the proposed method, biological know... | Rebuttal 1:
Rebuttal: Q1: ``what's the novelty of this paper in terms of machine learning”,
The brain prediction task could be considered a few-shot dense 3D prediction task. This task requires generalization with high SNR data. In building a per-voxel prediction, there are far more parameters than training samples... | Summary: The manuscript addresses the challenge of generating brain encoding models (specifically for visual stimuli) - which seeks to predict brain responses at the voxel level to visual stimuli. A challenge facing brain encoding is heterogeneity in data modality, individual variability, and functional differences acr... | Rebuttal 1:
Rebuttal: W1: “compare SOTA with computational speed”: see G1. We build on top of the state-of-the-art retinotopy module and remove the need for an additional mechanism to infer individual voxel's receptive size. Our LayerSelector significantly reduced the computation cost.
W2: “improvement in correlation... | Summary: The authors tackle the brain encoding task, which predicts brain voxel-level responses to image stimuli. The authors aim to train a comprehensive brain encoding model using the vast amount of public data from diverse imaging modalities and numerous participants. The proposed method, the All-for-One training re... | Rebuttal 1:
Rebuttal: S1: Thank you for the nice compliment on our figures!
W1: “how all-for-one works”: see G2 and attached PDF.
W2: “how RetinaMapper works”: see G1. We will update Figure 4 caption “RetinaGrid color corresponds to that color in the cortex, RetinaMap color does not match the color in RetinaGrid.”.
... | Rebuttal 1:
Rebuttal: # General response
## G1.
RetinaMapper is a crucial innovation of our method. It is an instance of a retinotopy module in the following discussion.
### A. state-of-the-art retinotopy module (RetinaMapper):
1) Attention mask per voxel, GNet in [1], requires way more learning parameters than ou... | NeurIPS_2023_submissions_huggingface | 2,023 | Summary: This paper focuses on the task of brain encoding model, which aims to predict brain voxel-wise responses to stimulus images. First, the paper proposes the All-for-One (AFO) training recipe, which enhances interactions among multiple ROI models to handle the large diversity within the data. Second, the paper in... | Rebuttal 1:
Rebuttal:
Q1A: “why add non-trainable variance to 'u’”: see G4. Non-trainable variance is a reparameterization trick.
Q1B: “why add global features to 'm^l’”: see G4. High-level visual voxels are mapped to the center point on the image grid, but their receptive field should cover the whole image; adding '... | null | null | null | null | null | null |
Curvature Filtrations for Graph Generative Model Evaluation | Accept (poster) | Summary: In this manuscript, the authors proposed a special curvature-related graph descriptors for the evaluation of graph generative model. The key idea is to evaluate the discrete curvatures on graphs and construct a curvature-based filtration process. The corresponding persistent Landscape based metric is used for ... | Rebuttal 1:
Rebuttal: Thanks for the positive assessment of our work! We will include the suggested citations and are very happy to provide additional clarifications!
> The advantage to use curvature as a filtration parameter is not clear. The advantage of the special filtration process and final vectorization approac... | Summary: The paper proposes a new set of graph features: first, 3 variants of graph curvature are calculated. Second, persistence landscapes are calculated over edge-based filtrations via curvature. Third, sets of graphs are compared via comparing their landscapes.
Theoretical properties of landscapes (stability, expre... | Rebuttal 1:
Rebuttal: Thank you for your positive assessment of our work! We are looking forward to clarifying any additional questions you might have!
> While the paper provides versatile approaches to comparison of graph datasets, the single end-to-end algorithm to evaluate graph generative models and corresponding ... | Summary: This paper proposed a TDA-based graph generation model evaluation method using the discrete Ricci curvature of the graph as a filtration function. The proposed comparison between the two graphs is theoretically shown to have good properties for graph analysis, and the comparison is verified in experiments.
St... | Rebuttal 1:
Rebuttal: We thank the reviewer for taking time in their review and providing valuable feedback. We have tried to address the weaknesses that you have highlighted. We believe most of the comments arise from the confusion between traditional distances between *graphs* and a distance between *distributions of... | null | null | null | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Ecosystem-level Analysis of Deployed Machine Learning Reveals Homogeneous Outcomes | Accept (poster) | Summary: This paper studies how particular examples in different benchmark datasets perform across multiple models using the HAPI and DDI datasets. It finds first that many examples are “systemic failures,” meaning that they are misclassified by all of the commercial models evaluated. Then, it also finds that model imp... | Rebuttal 1:
Rebuttal: Thank you to reviewer KctL for a very thorough and insightful review. First, we will address the issue of the expected failure rates and discuss what we hope this rate captures and what it does not; then we answer the specific questions raised by the reviewer.
> Though the authors do a good job i... | Summary: The paper empirically investigates ecosystems consisting of many machine learning models. Rather than analyze each model in isolation, the paper instead focuses on the extent to which all models in an ecosystem incorrectly classify an individual (the systemic failure rate), building upon recent work on *algori... | Rebuttal 1:
Rebuttal: Thank you to reviewer HXE5 for a thoughtful review and set of questions. At the heart of the matter, we see two core concerns raised in the review:
1. **Finding 4.** We agree that systemic failures are definitely a phenomenon of direct interest, but even beyond systemic failures, the racial dispar... | Summary: The paper introduces ecosystem-level analysis where instead of a single model, a collection of models will be evaluated together and their common failures/successes will be identified. Across three modalities, the paper shows a pattern of polarization where an instance is either more likely to be misclassified... | Rebuttal 1:
Rebuttal: We appreciate reviewer N5w8’s thorough review, highlighting the core framework and associated findings as key strengths. The questions they pose are definitely of interest but, overall,we unfortunately cannot answer several of them due to the lack of available data. Nonetheless, we provide more sp... | Summary: This paper introduces the concept of "ecosystem-level analysis," which evaluates the performance and impact of a collection of machine learning models in a given context rather than analyzing individual models in isolation. This approach could significantly contribute to understanding the broader impacts of de... | Rebuttal 1:
Rebuttal: Thanks to reviewer kuru07 for their helpful review. Overall, we firmly disagree with many of the vague claims made in the review, and ask the reviewer to provide specific evidence for us to be able to better engage with their claims. For example, at many points there are claims of lacking empirica... | null | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Faster Query Times for Fully Dynamic $k$-Center Clustering with Outliers | Accept (poster) | Summary: This paper studies dynamic algorithms for k-center with outliers in a general metric space with bounded doubling dimension, denoted as dim. In the k-center with z outliers problem (which they call (k, z)-center problem), the goal is to find a set of k points C, such that the maximum distance from the data poin... | Rebuttal 1:
Rebuttal: **Comment:** I don't find a statement of the main theorem. A rigorous statement may be helpful for people who would like to cite/use your paper without going too much into detail. Also, this makes it somewhat difficult to read your proofs, since it's easy to get lost in the series of lemmas and I ... | Summary: The main contribution of this paper is the proposal of a new approximate algorithm for "fully dynamic k center with outliers" problem in metric space with low doubling dimension. The approximation ratio is 3+$\epsilon$. The meaning of "fully dynamic" is that the algorithm should be able to efficiently adjust i... | Rebuttal 1:
Rebuttal: **Comment:** Time complexity. Table 2 in this paper presents a comparison of the time complexity of the proposed algorithm with some existing algorithms for solving the dynamic $k$ center with outliers problem. The table reveals that the previous works (such as [3, 16, 7] in the references) have u... | Summary: This paper presents a new algorithm for fully dynamic $k$-center clustering with outliers in spaces with bounded doubling dimension $dim$ and aspect ratio $\Delta$, where the number of outliers $z$, can be specified at query time. While the approximation ratio of $3+\epsilon$ is on par with previous results, i... | Rebuttal 1:
Rebuttal: We greatly appreciate your time in reviewing our paper and your feedback. | Summary: In k-center with z outliers problem, the input is a data set X that lies in metric space and the goal is to find a center set C of size k such that the z+1 largest distance from points in X to C is minimized. This paper stuided a dynamic version of this paper where data point can be inserted and deleted but we... | Rebuttal 1:
Rebuttal:
**Comment:** This result does not significantly improve previous results. For example, [16] already gave a dynamic algorithm with an $\epsilon^{-O(\text{ddim})} \log{\Delta}$ update time, an $\epsilon^{-O(\text{ddim})} (k+z)^2 \log{\Delta}$ query time, and the same approximation ratio. The new al... | Rebuttal 1:
Rebuttal: We would like to thank the reviewers. We appreciate your time in reviewing our paper and your feedback. Here are our responses to the questions. We will address any minor errors like typos and proposed references promptly. | NeurIPS_2023_submissions_huggingface | 2,023 | Summary: This paper consider the $k$-center problem with outliers in the fully dynamic setting (insertions and deletions of points).
The assumption is that the metric space has bounded doubling dimension (as a side note: some assumption is necessary to get better bounds due to known lower bounds).
The main result of ... | Rebuttal 1:
Rebuttal: **Question:** What factors in your running times are necessary? Is the exponential dependency on doubling dimension necessary etc?
**Response:** These are interesting questions. The exponential dependency on the doubling dimension seems unavoidable. We will pose this question as an open problem f... | null | null | null | null | null | null |
The Impact of Positional Encoding on Length Generalization in Transformers | Accept (poster) | Summary: Beyond model training, sequence length is one of the crucial aspect impacting the fluency, accuracy and completeness of LM outputs. For transformer based model's positional embeddings (PE) are used to explicitly represent the location of tokens. Over the past few years several works have attempted to propose a... | Rebuttal 1:
Rebuttal: We appreciate the reviewer's insightful feedback and questions, and we will now address them:
> Inclusion of other tasks such as summarization
We acknowledge the reviewer's concern about including more challenging tasks, such as summarization. However, we'd like to clarify why we focused on reas... | Summary: This work follows a long line of work analysing the effectiveness of various types of positional encodings and, in particular, a line of work exploring the effectiveness of causal masking without any positional encodings [1 - 8]. To that end, they evaluate the length generalisation performance of Transformers ... | Rebuttal 1:
Rebuttal: Thank you, we are delighted that you like our work!
> Novelty of results
Thank you for your insightful comment on the novelty of our results and the framing of our introduction. We agree that the field has seen previous work on Transformers without positional encodings, as you have detailed.
Wh... | Summary: Length generalization is one of hot topics for transformers. In this paper authors investigate how positional embedding influences this generalization for **decoder-only** models on **reasoning tasks**. Authors compare for variety of tasks (e.g. copy, summation, sorting, etc.) sinusoidal positional embedding, ... | Rebuttal 1:
Rebuttal: We would like to thank you for your positive and detailed feedback! We now address the key questions and concerns in detail below:
> Clarifying the Scope and Generalization Setup
Thank you for bringing attention to the perceived overstatement concerning generalization capability of our evaluated... | Summary: The paper presents a comprehensive study on the role of positional encodings in Transformer models, focusing on their ability to generalize over sequence lengths. The authors introduce a novel positional encoding method, which does not use explicit positional information.
Strengths: 1. The paper introduces a ... | Rebuttal 1:
Rebuttal: We thank the reviewer for their time in providing useful feedback and questions which we now address:
> Effect of large-scale pretraining on PEs
Thank you for raising the concern regarding the exploration of large-scale pretraining effects on different PEs. It's worth noting that the pretraining... | Rebuttal 1:
Rebuttal: ## General Response
We thank the reviewers for their positive feedback and constructive criticism.
We are glad that Reviewers 1 (VbRS) and 3 (rWd9) found our paper to be **“well-written and easy to follow,”** and that they believe it provides a “thorough and in-depth analysis,” as highlighted by ... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Gaussian Differential Privacy on Riemannian Manifolds | Accept (poster) | Summary: The paper extends definition of Gaussian DP to General Riemannian Manifolds and proposed Riemannian Gaussian mechanism that shows to achieve GaussianDP. Paper also develops MCMC procedure for Riemannian Manifolds with constant curvature.
Strengths: * Extending GaussianDP to Riemannian Manifolds
* MCMC proce... | Rebuttal 1:
Rebuttal: Dear Reviewer,
Thank you for your thorough review and constructive feedback. We appreciate your insights, and we have addressed your concerns and questions in our response below:
## Weakness
> The experiment compares the $\varepsilon$-differential privacy mechanism (Riemannian Laplace mechanism... | Summary: The authors consider the problem of achieving Gaussian Differential Privacy (GDP) on Riemannian manifolds. After defining what this concept means, they provide a mechanism capable of achieving GDP via a generalization of the Gaussian to manifolds. The major challenge in implementing the approach is in findin... | Rebuttal 1:
Rebuttal: Dear Reviewer,
Thank you for your thorough review and constructive feedback. We appreciate your insights, and we have addressed your concerns and questions in our response below:
## Weakness
> Implementing the mechanism seems quite challenging. Just to find the scaling parameter is difficult in ... | Summary: To obtain manifold-value statistic with the Gaussian differential privacy (GDP), the authors extend the GDP from Euclidean spaces to general Riemannian manifolds. Then they proved that their proposed Riemannian Gaussian distribution make GDP tractable. Furthermore, they proposed an efficient numerical algorith... | Rebuttal 1:
Rebuttal: Dear Reviewer,
Thank you for your thorough review and constructive feedback. We appreciate your insights, and we have addressed your concerns and questions in our response below:
## Weakness
> For the case of manifolds of dimension $d>1$, they make a too strong assumption that the manifold wit... | Summary: This work studies an extension of a Gaussian privacy mechanism on Riemannian manifolds, so as to extend the notion of mu-GDP for estimators with given global sensitivity. The idea is to sample a random point on the manifold with a Gibbs density (wrt to the Riemannian volume) whose potential is proportional to ... | Rebuttal 1:
Rebuttal:
Dear Reviewer,
Thank you for your thorough review and constructive feedback. We appreciate your insights, and we have addressed your concerns and questions in our response below:
## Weakness
> Laplace and Gaussian mechanisms can be very easily extended to Riemannian manifolds, and the computati... | null | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Navigating the Pitfalls of Active Learning Evaluation: A Systematic Framework for Meaningful Performance Assessment | Accept (poster) | Summary: The authors identify five main pitfalls that have impacted previous Active Learning (AL) empirical evaluations. They propose a novel evaluation framework that avoids these pitfalls, and they study both Active Learning and its potential synergies with Semi-Supervised and Self-supervised learning. The paper prov... | Rebuttal 1:
Rebuttal: We sincerely thank you for taking the time to carefully review our paper as well as for your insightful comments and questions. Note that the major changes made in our manuscript are detailed in the general reply above.
-> Presentational issues: “[...] the paper has "too much content in too litt... | Summary: The submission presents a large collection of experimental results for four active learning methods (entropy, core-set, BALD, BATCH) applied to image classification using ResNet-18 models, investigating the effect of differences in the data distribution (e.g., "roll-out data" from a different domain), size of ... | Rebuttal 1:
Rebuttal: We sincerely thank you for taking the time to carefully review our paper as well as for your insightful comments and questions. Note that the major changes made in our manuscript are detailed in the general reply above.
-> “Two of the main findings do not appear to be novel: the usefulness of ca... | Summary: Active Learning is a crucial problem that focuses on selecting a subset of examples from an unlabeled dataset to be labeled. Primarily community focus on two pillars for defining acquisition functions -- uncertainty and diversity based acquisition methods. With emergence of the highly powerful pretrained model... | Rebuttal 1:
Rebuttal: We sincerely thank you for taking the time to carefully review our paper as well as for your insightful comments and questions. Note that the major changes made in our manuscript are detailed in the general reply above.
-> “Weakness are more of some additional experimental suggestions. Given th... | Summary: The authors in this work try to address the common issues as to why active learning has not been in use compared to other methods like semi-supervised or self-supervised learning. The authors argue that the active learning methods are not evaluated in robust manner and in a cross-task generalisation setting w... | Rebuttal 1:
Rebuttal: We sincerely thank you for taking the time to carefully review our paper as well as for your insightful comments and questions. Note that the major changes made in our manuscript are detailed in the general reply above.
-> “Since it was mentioned in the limitation that the training may not have ... | Rebuttal 1:
Rebuttal: We sincerely thank all reviewers for their valuable comments. While the reviewers generally agreed on the added value of our work (“timely problem”,“extensive evaluation”, “useful insights”), concerns were raised regarding the clarity of the manuscript and the toning regarding Semi-SL. We have th... | NeurIPS_2023_submissions_huggingface | 2,023 | Summary: The paper identifies the pitfalls in current evaluation frameworks for AL, and proposed solutions that lead to a new evaluation framework (codebase) to tackle these problems. Empirical results reveal insights on the current state of AL methods.
Strengths: - The paper aims to tackle a very timely and important... | Rebuttal 1:
Rebuttal: We sincerely thank you for taking the time to carefully review our paper as well as for your insightful comments and questions. Note that the major changes made in our manuscript are detailed in the general reply above.
-> “Line 71 AL validation paradox: It is not very clear to me why AL paramet... | null | null | null | null | null | null |
Decentralized Randomly Distributed Multi-agent Multi-armed Bandit with Heterogeneous Rewards | Accept (spotlight) | Summary: This paper studies the problem of multi-agent multi-armed bandits where the arm rewards generated for each agent is heterogeneous across agents, and the goal is to find the global optimal arm, where the reward of the global optimal arm is the mean of the mean reward observed by agents. In addition, this work c... | Rebuttal 1:
Rebuttal: Thank you very much for the review and for providing your suggestions and comments. We would like to present the responses to the questions as follows.
Q1. This work, however, does not touch at all the communication complexity in their study. Can authors comment on whether or not it makes sense ... | Summary: Update: I've read the authors' rebuttal as well as all the reviews from other reviewers. The authors' clarification are helpful and mitigate my concerns. I've raised the score from 6 to 7 and I think this manuscript is acceptable for publication.
This paper aims at making contributions to decentralized multi-... | Rebuttal 1:
Rebuttal: Thank you so much for the review and for sharing your comments and perspectives. We would like to include the following perspectives.
Q1. the motivation/practical backgrounds of graphs
A1. We would like to refer to A5 in response to the comment Q5 from reviewer a3u4 where we provide detailed el... | Summary: This paper presents a novel algorithm for solving decentralized multi-agent multi-armed bandit problems for two graph dynamics under both sub-exponential and sub-gaussian reward distributions. Their whole algorithm framework can be divided into a Burn-in period and a Learning period, ensuring a process warm-up... | Rebuttal 1:
Rebuttal: Thank you very much for the review and providing the suggestions and insights. We would like to add responses as follows.
Q1. The paper is full of details but without a conclusion
A1. We will incorporate a concluding section into the paper, as shown in the official comment.
Q2. In line 368, is... | Summary: The paper studies stochastic multi-armed bandit over a network of agents communicating locally. Specifically, for each of the K arms, each agent has an associated underlying arm-mean. The agents can only communicate with their neighbors in each round to exchange information. The goal of the agents is to pull a... | Rebuttal 1:
Rebuttal: Thank you so much for reviewing the paper and for sharing the suggestions and perspectives. We would like to add some responses as follows.
Q1. Lack of experimental evidence.
A1. We will add an experimental illustration of the proposed algorithms using synthetic simulations to the paper, and c... | Rebuttal 1:
Rebuttal: We would like to thank all the reviewers for their valuable suggestions and comments. We have carefully considered the comments and provided additional perspectives accordingly. Thank you so much for your time and effort in reading the responses.
Additionally, in the uploaded file, we incorporate... | NeurIPS_2023_submissions_huggingface | 2,023 | Summary: This paper studies a multiplayer graph bandit setting. In particular, they consider heterogenous rewards in which the mean rewards for each agent are different, and are allowed to be both sub-gaussian and sub-exponential. They also consider random graphs between agents, which dictates which agents are allowed... | Rebuttal 1:
Rebuttal: Thank you very much for reviewing the paper and providing the detailed comments and suggestions. For the questions and limitations, we would like to include some responses as follows.
Q1. I think the organization of the settings as well as the key results can be better (Maybe each setting gets i... | null | null | null | null | null | null |
Selectively Sharing Experiences Improves Multi-Agent Reinforcement Learning | Accept (poster) | Summary: The paper proposes a modified prioritized experience replay algorithm where agents share a minimal amount of high-quality samples to improve learning.
Strengths: - The proposed method is very simple (therefore easy to implement and reproduce), and widely applicable for many cooperative RL environments, yet th... | Rebuttal 1:
Rebuttal: Dear Reviewer 3mbW,
We would like to thank you for your very helpful comments and suggestions. We would like to respond to some of your comments:
* Starcraft: While we agree that Starcraft is an interesting domain, it was designed specifically as a challenge domain for credit assignment in share... | Summary: The paper proposes a heuristic for sharing the experience between agents in Multi-Agent Reinforcement Learning (MARL) setting. It is a modification of Prioritised Experience Replay (PER) method, where the agents only share experiences, which td-error exceeds a certain threshold (paper proposes several ways to ... | Rebuttal 1:
Rebuttal: Dear Reviewer mQNR,
We would like to thank you for your insightful comments and suggestions. We would like to clarify a few details:
* Division of labor:
* Regarding the Overcooked domain in particular, while we agree that this is an exciting domain, the sparsity of reward make this an extre... | Summary: The paper proposes SUPER, a communication-based cooperative multi-agent reinforcement learning method with semi-decentralized training. In SUPER, each agent relays its highest-td-error experiences to the other agents, who insert them directly into their replay buffer, which they use for learning. The author be... | Rebuttal 1:
Rebuttal: Dear Reviewer 8wH6,
We would like to thank you for your constructive feedback and suggestions. Regarding your questions:
1. Could you clarify what algorithms you have in mind here? We are not aware of any other work that uses communication as part of its learning algorithm in a similar way; the ... | Summary: This paper introduces a semi-centralized multi-agent reinforcement learning algorithm named SUPER, by incorporating limited communication in the training phase.
With minor modifications, agents are able to achieve better performance than the baseline algorithms by sharing a limited amount of experiences.
St... | Rebuttal 1:
Rebuttal: Dear Reviewer aJ1o,
We would like to thank you for your helpful comments and suggestions, and would like to offer a few clarifications:
* Difference to previous work: To clarify, while SEAC uses importance sampling, this is only used to “translate” action probabilities for off-policy experiences.... | Rebuttal 1:
Rebuttal: Dear Reviewers,
In addition to individual responses, we are uploading a PDF with figures from testing SUPER in additional domains, as this was a common point raised in the reviews. Following the suggestion from reviewer mQNR, we used the Melting Pot package for this, and compared SUPER-DDQN again... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Statistically Valid Variable Importance Assessment through Conditional Permutations | Accept (poster) | Summary: In this work, the authors proposed a novel conditional permutation importance approach (CPI) that can well estimate the variable importance with the p-value provided under the condition that correlation exists across features. The authors theoretically justify the inflation of the feature permutation-based app... | Rebuttal 1:
Rebuttal: W1: “The authors should also provide the qqplot with a reasonable amount of tests (>=100) to demonstrate the calibration.”
R_W1: In $\textbf{PDF Fig. R1}$, we provided the qq-plot for the distribution of the p-values provided by Permfit-DNN and CPI-DNN for one randomly picked non-significant vari... | Summary: The authors argue for the importance of using conditional sampling in permutation importance (CPI) methods.
They show that when features are correlated, using non-conditional sampling for permutation importance (on a finite training set) leads to a bias in the estimated feature importance. (Specifically, the ... | Rebuttal 1:
Rebuttal: S1: “The first is a detailed critical discussion of the assumptions of Mi et al (...) ”
R_S1: Mi et al. 2021 demonstrated the intuitive estimation of variable importance in the biomedical context using an effective brute-force approach. This can work well thanks to their high-capacity function ap... | Summary: The authors present an approach that aims to tackle the problem with the default permutation importance technique of performing variable importance. The default setting is prone to detecting correlated variables even as important even when they are not (Type I error). The authors present Conditional permutatio... | Rebuttal 1:
Rebuttal: W1: “The clarity of the experiments can be improved. Firstly the selections of models, datasets and settings can be motivated better. Secondly, the first paragraph of section 5 mentioned that “default behavior” was implemented. It’s not very clear what this default behavior is.”
R_W1: First of al... | Summary: The paper proposes a statistical testing framework for more effectively determining variable importance in generic learning settings. Using a conditional permutation test, they show that correlated variables are better handled theoretically and empirically compared to existing unconditional setups.
Strengths... | Rebuttal 1:
Rebuttal: W1: “It's unclear to me how the performance of the method would stack up against aggregated Shapley or LIME-like methods. (...).”
R_W1: In Results (section 6) L.278-280, we mention that “Additional benchmarks on popular methods that do not provide p-values, e.g. BART [Chipman et al., 2010] or Sh... | Rebuttal 1:
Rebuttal: We thank all four reviewers for the time they spent studying our proposed work and for their thoughtful comments. We would also like to thank the reviewers for meaningfully recognizing the strengths of our work while giving us an opportunity to improve it based on their constructive criticism info... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
On the Implicit Bias of Linear Equivariant Steerable Networks | Accept (poster) | Summary: This paper studies an implicit bias of gradient descent (GD) on linear equivariant neural networks. The authors show that the asymptotic GD solution of a linear steerable neural network, an instance of the equivariant nets, agrees with that of an FC net trained with data augmentation. They also show that the f... | Rebuttal 1:
Rebuttal: **Comment 1:** The paper does not contain empirical evaluations. The authors admit that the analysis relies on the asymptotic approximation, which might have a gap toward a real solution. I feel the experiments make the paper significantly more convincing.
**Answer:** We agree with the reviewer ... | Summary: The paper present several results on linear equivariant steerable networks. 1) linear equivariant NN converge in the same direction as linear networks trained with data-augmentation, 2) equivariant networks always attain a larger margin then group-augmented fully connected networks, 3) also the generalization ... | Rebuttal 1:
Rebuttal: **Comment 1:** The paper is essentially a mathematics paper and is the question to what extend it fits the scope of NeurIPS. I find it hard to assess the significance of the work.
**Answer:** Our research is indeed theoretical in nature, and we think it fits within the scope of NeurIPS, as reflec... | Summary: The paper presents a collection of theoretical findings that highlight the equivalence between using a linear steerable network for group-invariant binary classification and employing data augmentation for non-invariant models within the specified context. Additionally, the paper demonstrates that linear steer... | Rebuttal 1:
Rebuttal: **Comment 1:** This paper is theoretical in nature. It does not present any new empirical results or experiments to validate its claims. Unfortunately, the reviewer has little knowledge of the topic.
**Answer:** We agree with the reviewer that empirical verification of the theory established in t... | Summary: This paper studies the implicit bias on the use gradient flow for the purpose of binary (group-invariant) classification while using linear equivariant steerable networks. There is now a fair amount of work on implicit bias in similar settings in the non-equivariant case. There is also some work for linear equ... | Rebuttal 1:
Rebuttal: **Minor Comment**: On page 3, line 122: "Equivalently, we want the following diagram to be commutative.." --> reword this since that diagram is already commutative, but that you want your network to adhere to it.
**Answer**: We thank the reviewer for pointing this out. Yes, the diagram has indica... | null | NeurIPS_2023_submissions_huggingface | 2,023 | Summary: The paper considers the implicit bias of gradient flow optimization for linear equivariant steerable networks. The paper consists of the following steps and contributions:
* The equivariant network is assumed to have regular representation at the input (Assumption 3.1). The architecture is characterized in Pr... | Rebuttal 1:
Rebuttal: **Comment 1**: When considering linear group invariant classifiers, the steerable network, based on Schur’s lemma, uses only the trivial representation. I.e., only a single basis in equation (4) matters for optimization and the output. The definition in (4) is misleading in this sense. Proposition... | Summary: This paper analyzes the implicit bias of gradient flow on linear group-equivariant steerable networks for group-invariant binary classification. It shows that the parameterized predictor converges in a direction that aligns with the unique group-invariant classifier with a maximum margin that is defined by the... | Rebuttal 1:
Rebuttal: **Comment 1**: The authors claim to fully characterize the implicit bias of the training algorithm. However, there is no experiment on empirical dataset other than theoretical results throughout the paper. It would be more promising if the conclusion can be supported by empirical results.
**Answe... | null | null | null | null |
Approximation-Generalization Trade-offs under (Approximate) Group Equivariance | Accept (poster) | Summary: The paper presents a theoretical investigation into the approximation and generalization errors for a model exhibiting approximate equivariance. The paper introduces the concept of "error to equivariance," a measure of the deviation of the error function from perfect equivariance. It then defines the $\epsilon... | Rebuttal 1:
Rebuttal: We thank the reviewer for the very encouraging feedback and recognizing the importance of our work. We try to address each of the points raised one by one below:
### Section 4
Thanks for pointing out that accessibility could be improved! Since section 4 mostly describes background, we decided to... | Summary: In this paper, the authors spell out explicitly the benefit of capturing task specific symmetries into machine learning models. This goal is achieved by deriving rigorously generalization error bounds that takes into account the cardinality of the set of transformations that a model is (approxaimately) equiva... | Rebuttal 1:
Rebuttal: Thank you for your appreciative comments about our work and recognizing its contribution to the literature. We hope to address some of the questions raised point-by-point below.
### Applicability and experiments evaluation
As we mentioned to reviewer CLQc, we have considered including experiments... | Summary: This paper provides theoretical investigation on the intuition that, machine learning models (e.g. neural networks) whose symmetry aligns with the data symmetry achieve best performance and generalization. The paper formulate the relationship between model equivariance error and data equivariance error by prov... | Rebuttal 1:
Rebuttal: Thank you for your very encouraging comments on our work. We are grateful that you appreciate its significance and novelty and that you also liked our treatment/survey of the historical results in the area. :) We also found your questions instructive, and acknowledge their role in improving our pr... | Summary: In their theoretic work, the authors presents quantitative bounds on the generalization error and approximation error. They extend the results to approximate symmetries, which refers to functions that are close, in the supremum norm, to a group action.
Their first results shows that when working with only th... | Rebuttal 1:
Rebuttal: We are grateful to the reviewer for a careful and critical reading of our paper and also for recognizing its significance in the literature on equivariance.
### Discussion
We acknowledge that our presentation might appear a bit terse. As we mentioned to reviewer CLQc and W1Je, if accepted, we wi... | null | NeurIPS_2023_submissions_huggingface | 2,023 | Summary: The submission discusses the role of task-specific symmetries in enhancing machine learning models. It presents a quantitative analysis substantiating the common intuition that task-specific symmetries lead to improved generalization. The authors delve into the interplay between model and data symmetry, partic... | Rebuttal 1:
Rebuttal: We appreciate the time that you have taken to read our paper and for providing feedback. Below we will try to address some specific points that you have raised.
### presentation
As we mentioned to reviewer CLQc, if accepted, we will use the additional page to add more context and discussion aro... | Summary: There are two main results in this paper. The first one (Theorem 4.4) obtains a PAC-generalization bound for a hypothesis space which is approximately equivariant to transformations by elements g in G. This bound shows that in addition to the classical complexity term that depends upon the size of the hypothes... | Rebuttal 1:
Rebuttal: We are grateful to the reviewer for the very encouraging feedback. In particular, we are glad that our theoretical perspective on equivariance is appreciated and recognized to not have been commonly addressed. Thanks for the Baxter reference, which we were not aware of.
Below we will address som... | null | null | null | null |
Data-Centric Learning from Unlabeled Graphs with Diffusion Model | Accept (poster) | Summary: The paper discusses the problem of graph property prediction tasks, where there is a limited amount of labeled data and a large set of unlabeled graphs available. The conventional approach is to train a model using self-supervised tasks on the unlabeled graphs and then fine-tune it for prediction tasks. Howeve... | Rebuttal 1:
Rebuttal: ### To weakness 1 about the literature review on graph diffusion model
Thanks for your comments. Here are discussions on recent advances for diffusion models on graphs.
Recent works have improved the generative modeling of diffusion models on graphs [1,2,3,4,5]. EDP-GNN [1] employed score matchi... | Summary: This paper proposes a new data augmentation method called DCT based on unlabeled graphs. The proposed DCT generates new augmented graphs using a diffusion model from unlabelled graphs and also with labeled graphs from down-stream tasks as guidance. Two specific objectives are proposed to guide the model's deno... | Rebuttal 1:
Rebuttal: Thank you for your feedback model details and literature review. We offer detailed responses to each of your points.
### To weakness 1 and question 2 about the diffusion model on graphs
We follow [1,2] to define graph diffusion models. [2] unified score-based generative models and denoising diff... | Summary: The paper leverages diffusion SDE to approach the hidden chemical space and help the prediction tasks of molecular properties.
Strengths: 1. It is a good idea to leverage mutual information for the alignment of the hidden chemical space within both labeled and unlabeled graphs.
2. The results presented in thi... | Rebuttal 1:
Rebuttal: We sincerely value your feedback. We respectfully and firmly defend the originality, the literature review, and a fair evaluation of the performance of our model. We begin by offering a concise overview of our rebuttal, followed by a comprehensive point-by-point elaboration.
### Summary
($\text{T... | Summary: This paper studies leveraging unlabeled graphs to aid the supervised graph learning tasks with limited labeled graphs. This paper starts with a motivation that previous self-supervised methods may conflict with the downstream task and lead to poor performance. This paper proposes to model the graph distributio... | Rebuttal 1:
Rebuttal: Thank you for your feedback on the model details. Below, we offer detailed responses to each of your points.
### To question 1 and 2 about the training of the diffusion model and dataset
We train one diffusion model for 14 molecule and polymer tasks on the unlabeled QM9 datasets.
We consider a... | null | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Sharpness Minimization Algorithms Do Not Only Minimize Sharpness To Achieve Better Generalization | Accept (oral) | Summary: This paper sheds light on the relationship between three concepts: (a) generalization (b) flatness and (c) explicitly inducing flatness via SAM. The paper identifies that the relationship is nuanced and is based on certain architectural properties of the model.
Specifically, depending on the architecture, th... | Rebuttal 1:
Rebuttal: We appreciate the detailed review and comment! We would like to answer the reviewers’ questions here.
1. **Q1a.** Our bound is independent of the width of the model. Under the setting of Thm 3.1 and Lemma 3.1, we show our result by using sharpness as an upper bound for the parameter norm square u... | Summary: This paper studies the connections between sharpness of the training loss and generalization performance under a simplified MLP architecture. In particular, authors show scenarios of when flat solution do/do not generalize, as well as cases when Sharpness-Aware Minimization (SAM) does/does not generalize.
Str... | Rebuttal 1:
Rebuttal: We appreciate the review and comment! Regarding the reviewer’s concern, there seem to be some misunderstandings that we wish to clarify.
1. **W1: The generality of the results**
While our positive result theorem 3.1 only holds for a 2-layer ReLU network without bias, our main results, which are ... | Summary: This work aims to examine different cases for which flatness implies generalization. The conclusion heavily depends on the underlying model, while there are three different scenarios that contradict with each other. Additionally, the theoretical guarantees study the generalization error at the solution, howeve... | Rebuttal 1:
Rebuttal: We thank the reviewer for the effort and time invested in the review. However, there are a lot of misunderstandings in the review and we clarify them below one by one.
1. To our best knowledge, there is no reported correlation between progressive sharpening and better generalization. Arora et al. ... | Summary: This paper explores the relationship between sharpness and generalization in overparameterized neural networks. The authors challenge the popular belief that flatness of the loss function implies better generalization and show that sharpness minimization algorithms do not always lead to better generalization. ... | Rebuttal 1:
Rebuttal: We appreciate the reviewer’s positive review! Below are our responses to the reviewer’s concerns:
* **W1.** We would like to clarify that we only perform experiments on synthetic tasks instead of CIFAR10 or CIFAR100 in the paper. However, we added an experiment on MNIST in the rebuttal phase. See... | Rebuttal 1:
Rebuttal: We would like to thank the reviewers for the detailed reviews. Below we address a common question (by reviewers fkTA and ex27) on data distribution, whether the phenomena observed on synthetic datasets in this paper extend to vision datasets, by providing additional MNIST experiments.
**MNIST ex... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
How to Scale Your EMA | Accept (spotlight) | Summary: In duality to previous work on learning rate scalings, the authors present a novel scaling strategy for the momentum parameter in teacher-student settings, i.e. how to adapt the momentum in the learning dynamics when the batch-size is changed. The work puts an emphasis on the large batch-size setting, which is... | Rebuttal 1:
Rebuttal: Thank you for taking the time to read our work. We are happy you found the paper easy to read, appreciate the empirical and theoretical validation of the EMA Scaling Rule, and recognize the broad applicability of the result to many modern ML settings.
### Weaknesses
#### *1. No EMA Scaling Rule*... | Summary: This work proposes a scaling rule for EMA in optimization and empirically evaluates its effectiveness across a range of tasks and optimizers. Authors find that proposed method enables large batch training using 24k batch sizes with a 6x wall-clock time reduction.
Strengths: - Idea of formulating a scaling rul... | Rebuttal 1:
Rebuttal: Thank you for taking the time to read our manuscript. We are are happy you find our formulation novel, the theoretical motivations clear, and the study design strong.
We begin by clarifying a potential misunderstanding. Our work is devoted to the behaviour of exponential moving averages (EMAs) of... | Summary: The exponential moving average (EMA) of model parameters is used in many contexts, including to improve robustness in supervised learning, and as an important component of self-supervised learning. Given model parameters $\theta_t$ at iteration $t$, the exponential moving average $\zeta_t$ is updated as $\zeta... | Rebuttal 1:
Rebuttal: Thank you for taking the time to read our work. We are happy you found the plots clear, appreciate the mixture of controlled toy and diverse real world investigations, and find utility in enabling large batch training of BYOL with Vision Transformers.
### Weaknesses
#### *Obfuscation*
The purpos... | Summary: This paper talks about the scaling rule for model EMA, and proposes that the momentum coefficient of EMA update should be scaled exponentially with the batch size instead of linearly. The authors first informally derive such rule (Equation 4) under the assumption that the gradients change slowly for a few step... | Rebuttal 1:
Rebuttal: Thank you for taking the time to read our work. We are happy you found the paper easy to read, acknowledge the contributions made, and see the broad applicability of an EMA Scaling Rule.
### Weaknesses
#### *Figure 6 - ViT scaling*
We also note that the use of an EMA Scaling Rule is more importa... | Rebuttal 1:
Rebuttal: We thank all our reviewers for taking the time to review our work. We will adapt the paper based on their insightful comments, feedback, and questions.
There were a number of questions prompting additional investigations. We provide here a pdf containing the results of those investigations. Its c... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Optimal Rates for Bandit Nonstochastic Control | Accept (poster) | Summary: The authors present an algorithm fro the LQG problem, in which a linear system has to be controlled in the presence of semi-adversarial perturbations in order to minimise an adversarially-chosen cost, with the controller only having access to a projection of the real dynamics onto a lower-dimensionality subspa... | Rebuttal 1:
Rebuttal: Thank you for taking the time to read and review our paper. We tried to address your concerns in the following paragraph:
Weakness:
**FiLK**: The paper is extremely theoretical and deals with a very specialised problem. This problem has connections with more mainstream AI problems (Optimal Contr... | Summary: This paper studies LQG with semi-adversarial perturbations and time varying adversarial bandit loss functions.
The best known bound for such problems has a T^{3/4} time horizon dependence.
They provide an algorithm for LQR/LQG problems that attains $\tilde{\mathcal{O}}(\sqrt{T})$ regret.
Their algorithm goes ... | Rebuttal 1:
Rebuttal: Thank you very much for your review and valuable feedback. Here are our responses to the points you raised:
Weaknesses:
**iuZo**: I must express my concerns about its presentation and structure. The writing is not as clear as it needs to be, making it quite challenging to follow the paper. The a... | Summary: This paper presents an algorithm to perform control in the semi-adversarial setting. After presenting the gap in the literature they want to fill, the authors present an algorithm that shows optimal regret ($O(\sqrt{T})$) for this setting, considerably improving the previous bound known from the literature ($O... | Rebuttal 1:
Rebuttal: Thank you very much for your review and feedback. Here is the response to your concern, with which we hope to further boost your confidence in our work:
**2R7J**: Algorithms 1, 2, and 3 should be explained instead of only presenting the pseudo-code. Some steps may not be clear to all the readers ... | Summary: This paper studies LQG with semi-adversarial disturbances and bandit feedback. The authors proposed a novel algorithm with $\sqrt T$ regret bound, which improves the $T^{3/4}$ regret bound in the literature. The algorithm utilizes the quadratic cost function structure and the ellipsoidal gradient estimator.
S... | Rebuttal 1:
Rebuttal: Thank you for taking the time to read and review our work. We believe that your concerns are valuable for us to improve our presentation. We hope that the following responses can explain our novelty and contribution better and boost your confidence in our work.
Weakness comments:
**7wFb**: Sim... | null | NeurIPS_2023_submissions_huggingface | 2,023 | Summary: This paper studies the online control problem with a quadratic cost function, adversarially chosen dynamics, and semi-adversarially chosen losses (they are the sum of a zero-mean term and an obliviously chosen adversarial term). The algorithm proposed lowers the previous $T^{3/4}$ algorithm to a $\sqrt{T}$ alg... | Rebuttal 1:
Rebuttal: Thank you for your thorough review. We respond to your points here:
1. **6aiM**: Assumption 3.3 seems unnatural. Can you provide a bit more context on its necessity? More generally, how are ideas from learning with delay present in your algorithms? I did not find this point very clear. **Response... | null | null | null | null | null | null |
NAR-Former V2: Rethinking Transformer for Universal Neural Network Representation Learning | Accept (poster) | Summary: This paper studies how to model and learn the representations of neural networks themselves. Inspired by Transformer and GNN, NAR-Former V2 is proposed which is a modified Transformer-based universal neural network representation learning model. Extensive experiments demonstrate its effectiveness.
Strengths: ... | Rebuttal 1:
Rebuttal: Thank you for your positive comments. The responses to your questions is as follows.
>1. The latency of the proposed NAR-Former V2.
The latency of the proposed NAR-Former V2 is presented in Table 3 of the supplementary. More latency comparison experiments are provided as follows:
| ... | Summary: This paper proposes to design graph-aided attention Transformer blocks that can be used to model and learn the representation of neural networks. Specifically, one can leverage this kind of representation learning framework to provide more accurate latency and accuracy predictions of given deep neural networks... | Rebuttal 1:
Rebuttal: Many thanks for your review and detailed sugession. Next, we will endeavor to address your concerns point by point.
> 1. Provide a context for what "NAR" refers to.
Thank you very much for your suggestions; this was an oversight on our part. We should consider more for the audiences who are new ... | Summary: This submission builds on previous work and proposes NAR-Former V2, an updated transformer architecture to learn representations of neural network architectures. NAR-Former V2 targets NAS tasks, predicting model properties like performance or latency, to improve NAS runtime and cost. NAR-Former V2 makes use of... | Rebuttal 1:
Rebuttal: We appreciate the many detailed and insightful questions you have raised. Below are our responses.
> 1. add standard deviations or other metrics in Table 3 and Table 4
Thanks for your suggestion. Indeed, this is a good way to evaluate two methods. Compared with Nar-Former, Nar-Former V2 exhibits... | Summary: The paper proposes a transformer based neural network representation learning model that incorporates the inductive representation capabilities of graph neural networks into the transformer architecture to improve it's generalizability on unseen architectures. The NAR-Former framework, which the proposed model... | Rebuttal 1:
Rebuttal: We appreciate the valuable questions and suggestions you've given, and here are our responses.
> 1. More explanation for the proposed Type-aware Enhancement module
The ``Enhancement" in the Type-Aware Enhancement module refers to the use of intrinsic information of the input data to modify the h... | Rebuttal 1:
Rebuttal: We thank all reviewers for their valuable comments. Reviewers 9nHL,L4c1 and sJnp comment this paper is well written and easy to read. Reviewers L4c1, kdkx and sJnp agree with our motivation and comment the idea of using network connectivity by conditioning on the adjacency matrix of the network's ... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Automatic Integration for Spatiotemporal Neural Point Processes | Accept (poster) | Summary: In this paper, the authors propose a new method for spatiotemporal neural point processes (STPP). A key innovation of their method is that they introduced AutoInt for efficient integration of the intensity function of STPP. Using the divergence theorem and Green’s theorem, they convert the original triple inte... | Rebuttal 1:
Rebuttal: Thank you for your review of our paper. We appreciate your positive feedback on the novelty and methodology contribution of our proposed method.
**Q: So the authors are encouraged to clarify contributions of this paper over the published paper above.**
**A:** This paper was scheduled to appear i... | Summary: The paper proposes a novel framework to extend the automatic integration of conditional intensity function to spatio-temporal point processes, aiming to overcome the computational challenges of triple integration. Specifically, the proposed model adopts an integral network and converts the triple integral into... | Rebuttal 1:
Rebuttal: Thank you for your review of our paper. We appreciate your positive feedback on the proposed framework and the comprehensive experimental results supporting our claims. We acknowledge the weaknesses and limitations pointed out and will address them accordingly.
**Q: I have to admit that the prese... | Summary: The paper introduces a new approach to address the challenge of integrating spatiotemporal point processes in a flexible and efficient manner. The existing methods either lack flexibility or introduce numerical errors. The authors propose AutoSTPP as a novel paradigm that overcomes this limitation and showcase... | Rebuttal 1:
Rebuttal: Thank you for providing feedback on our paper. We appreciate your positive remarks regarding the soundness and presentation of the paper. Below, we address the concerns you raised regarding the contribution.
**Q: An essential related work is missing: Automatic Integration for Fast and Interpretab... | Summary: In this paper is proposed an automatic integration scheme for efficient learning for spatiotemporal process models, where the corresponding function is approximated with a deep neural network. A priori, the method can be used to compute the likelihood of any intensity, even for sophisticated ones. The method i... | Rebuttal 1:
Rebuttal: Thank you very much for reviewing our paper. We greatly appreciate the reviewer's thorough feedback and the positive evaluation of our paper.
**Q: Some parameters in Eq.(1) are not properly defined.**
**A:** Sorry for not properly defining tau in Equation (1). Thank you for pointing this out! W... | Rebuttal 1:
Rebuttal: Once again, we sincerely thank you for your valuable feedback and constructive criticism. We will incorporate the suggested improvements and address the raised concerns in the final version of the manuscript. We will emphasize the differences between our approach and the work by Zhou et al. publis... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Sketchy: Memory-efficient Adaptive Regularization with Frequent Directions | Accept (poster) | Summary: This paper looks at using Frequent-Directions to do (non-diagonal) adagrad more efficiently. It suggests doing this for both full-matrix adagrad and Shampoo, a version that takes into account the matrix structure of the parameters.
Strengths: I think it's a pretty reasonable and natural idea, and both the th... | Rebuttal 1:
Rebuttal: Thank you for taking the time and effort to read and review our paper! Here are the responses to your concerns:
Weaknesses:
1. The experiments don't get any actual improvement in time or memory relative to existing implementations.
**Response**: Please see the top comment regarding performance ... | Summary: Motivated by empirical observation that Kronecker-factored gradient covariance matrix is (approximately), this work proposed a sketching scheme for memory-efficient preconditioner for gradient-based optimizer.
Strengths: - The observation of the low-rankedness of the KF convariance matrix in Figure 3 is quite... | Rebuttal 1:
Rebuttal: Thank you for taking the time and effort to review our paper. Here are our responses to your concerns:
Weaknesses:
1. From Figure 1, it seems that the main application of sketchy is S-Shampoo that achieves the sublinear memory. However, I wonder for shampoo (where the factorization has a certain... | Summary: This paper proposed a sketched online optimization method. The main idea is using frequent directions (FD) with adaptive regularization to maintain a preconditioner in the training process, which leads to $O(dk)$ memory and additive error associated with the small eigenvalues of the gradient covariance.
Stre... | Rebuttal 1:
Rebuttal: Thank you for taking the time and effort to review our paper. Here are our responses to your concerns:
Weaknesses:
RFD has the same regret bound (though different constants) as FD-SON, which as per Table 1, page 3 can be linear in T when there is no spectral decay or curvature (see our response ... | Summary: It is known that dense preconditioning can be highly beneficial for optimization, but this has high memory requirements and is often prohibitive. The classic solution to this problem is the use of a diagonal preconditioner. This work proposes the use of a low rank preconditioner by using a streaming approach t... | Rebuttal 1:
Rebuttal: Thank you for taking the time to read and review our work. Here are our responses to your questions and concerns, with which we hope to boost your confidence in our work:
Weaknesses:
**rVs5**: I am confused about the primary contribution of this work, and perhaps the writing needs to be improved... | Rebuttal 1:
Rebuttal: Thank you for all the reviewers for taking the time to read and review our work. Many reviewer comments focus on the differences between S-Adagrad and prior work which inspects FD for optimization.
We invite the reviewers to look at the following sections of our paper. As mentioned in L31, the Kr... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Uncovering Prototypical Knowledge for Weakly Open-Vocabulary Semantic Segmentation | Accept (poster) | Summary: This paper proposes that the traditional WOVSS approach lacks effective supervision of group tokens. In this paper, visual patch tokens are modeled through the GMM model to obtain prototype features that can be used to supervise each group token. Experimental results show that the proposed PGSeg achieves state... | Rebuttal 1:
Rebuttal: ### Q1
>As shown in Table 3, The performance of the baseline (GroupViT) drops a lot when using CC12M+RedCaps compared to CC12M+YFCC14M(Setting in the original GroupViT). To demonstrate the effectiveness of PGSeg, the performance of PGSeg trained with CC12M+YFCC14M should be reported.
Thanks for ... | Summary:
This paper presents a method for addressing the challenge of granularity inconsistency in weakly open-vocabulary semantic segmentation (WOVSS). To tackle this issue, the authors propose a technique called non-learnable prototypical regularization (NPR). NPR leverages explicit supervision from prototypical kno... | Rebuttal 1:
Rebuttal: ### Q1
>The term "granularity inconsistency" is initially unclear in the paper, specifically in lines 37-40 and the caption of Fig.2. Although Sec.3 provides clarification, it would be preferable to have a clear explanation in the abstract or introduction, as it serves as the main motivation for t... | Summary: This paper studies the problem of weakly open-vocabulary semantic segmentation, which aims to segment objects of arbitrary classes using only image-text pairs. The paper proposes a prototypical guidance segmentation network, which uses non-learnable prototypes estimated from image and text features to regulari... | Rebuttal 1:
Rebuttal: ### Q1
>Ablation studies are not enough. The number of prototypes should be more for ablation study. The affect of the position and number of PG-Unit should be explored in ablation studies.
Thanks for this suggestion! Here we first present the improved Table 5 by expanding the range of the number... | Summary: The paper addresses the problem of weakly open-vocabulary semantic segmentation, which is to segment images of novel classes without any segmentation annotation.
To segment unseen classes, a model is trained to align between image features and text caption to learn to associate between image patches and class ... | Rebuttal 1:
Rebuttal: ### Q1
>The reviewer finds that the emphasis on the non-learnable aspect of the algorithm is misleading. Specifically, the proposed method does use expectation-maximization to compute the prototype which is a form of learning algorithm, which contradicts with the claim about non-learnable prototyp... | Rebuttal 1:
Rebuttal: ## Summary
We thank reviewers for their valuable feedback, and appreciate the great efforts made by all reviewers, ACs, SACs and PCs.
We are invigorated by the **positive evaluation** from all reviewers. Specifically, they find the method **novel and reasonable** (nyAz, ae73), **well-motivated... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
How2comm: Communication-Efficient and Collaboration-Pragmatic Multi-Agent Perception | Accept (poster) | Summary: This article proposes a communication-efficient and collaboration-pragmatic framework that effectively reduces object detection errors caused by transmission delays in multi-agent collaborative perception tasks, while achieving a balance between bandwidth utilization and detection accuracy. This paper introduc... | Rebuttal 1:
Rebuttal: We thank the reviewer for recognizing our effective solution, complete experiments, and reliable results. Below are some specific responses. The common ablation experiments can be found in the global response.
**Q1**: Reference problem for the figure in line 150.
**A1**: Thank you for the cons... | Summary: This paper presents a collaborative perception system called How2comm, which aims to address three common challenges encountered in the collaborative perception setting:
1) Communication redundancy: To mitigate communication redundancy, the paper introduces a Mutual Information-aware communication approach. T... | Rebuttal 1:
Rebuttal: We thank the reviewer for recognizing our effective designs and comprehensive experiments. Below are some specific responses. The common questions regarding the definition of collaboration heterogeneity, ablation studies, and inference time can be found in the global response.
**Q1**: Design log... | Summary: This paper proposes How2comm, a collaborative perception framework, which seeks a trade-off between perception performance and communication bandwidth. This framework leverages a mutual information-aware communication to select the most informative messages to transmit and save bandwidth, where ego agents comp... | Rebuttal 1:
Rebuttal: Insightful comments! Below are some specific responses. The common questions can be found in the global response.
**Q1**: About technical novelties.
**A1**: Our three contributions fundamentally differ from previous works regarding design philosophy and structure modeling. We have highlighted co... | Summary: The paper introduces three key contributions in tackling the aforementioned challenges. Firstly, the authors devise a mutual information-aware communication mechanism that maximizes the sharing of informative features among collaborators. They employ spatial-channel filtering for efficient feature sparsificati... | Rebuttal 1:
Rebuttal: We thank the reviewer for recognizing our contribution of novel approach, comprehensive evaluation, and clear presentation. We present our responses below.
**Q1**: Discussion of any limitations or potential areas for improvement.
**A1**: 1. Thanks for your constructive comments! We have discuss... | Rebuttal 1:
Rebuttal: We thank all reviewers for their time and effort. Here we provide detailed responses to common questions.
**Q1**: Trade-off between training cost, inference time, and perception performance.
**A1**:
* We compare the evaluated models regarding space-time complexity, inference time, and detection... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
AiluRus: A Scalable ViT Framework for Dense Prediction | Accept (poster) | Summary: ViTs process input images as a sequence of tokens, and as the token size increases, the computational complexity escalates. The paper aims to accelerate the training and inference of ViTs by reducing the number of tokens. To achieve this, the authors introduce an adaptive resolution strategy, leveraging the ob... | Rebuttal 1:
Rebuttal:
**Q1: The novelty**
**A1:**
1. Please note that both STViT and PaCa-ViT rely on training and specific designs, which limits their ability to accelerate well-trained models. In contrast, Ailurus can be seamlessly integrated into pre-trained models for instant acceleration. To illustrate the ver... | Summary: This paper first points out that when handling long token sequences, ViTs particularly suffer in terms of computation. Motivated by an observation that dense prediction tasks focus more on the contours or shapes of object, the authors propose to leverage adaptive resolution strategy for different regions depen... | Rebuttal 1:
Rebuttal:
**Q1: Improve visualization**
**A1:** Thank you for your suggestion! We greatly appreciate it. Based on the concept of coloring a map, we have made improvements to our visualization. Specifically, we use highly distinguishable colors to differentiate adjacent patches belonging to different assig... | Summary: This paper proposed a spatial-aware density-based clustering algorithm for ViT-based dense prediction tasks. The proposed method can merge the original tokens into fewer representative tokens to accelerate the transformer inference without finetuning. The experiments on different datasets show that the propose... | Rebuttal 1:
Rebuttal: **Q1: The relation to [29]**
**A1:** We respectfully disagree. TCFormer and Ailurus exhibit significant differences in terms of motivation, implementation, and application scenarios.
1) The motivation behind TCFormer lies in the utilization of DPC to incorporate saliency information in human-ce... | Summary: This paper proposes a new approach to improving the efficiency of Vision Transformers (ViTs) for dense prediction tasks such as semantic segmentation. Recognizing that these tasks emphasize the contours or shapes of objects more than the textures within, the authors introduce a method to apply adaptive resolut... | Rebuttal 1:
Rebuttal: **Q1: Typographical errors**
**A1:** We are sorry for these typos and have corrected them in the revised manuscript. Many thanks for your patience.
**Q2: The placements of figures and tables**
**A2:** We sincerely appreciate your suggestion and will rearrange them to make it easier to refer to.... | Rebuttal 1:
Rebuttal: We sincerely appreciate all reviewers for their efforts, and first give general responses for some common issues.
**Why does Ailurus performs better than Expedite?**
As presented in the original paper, Ailurus significantly outperforms Expedite, especially in cases **when performing token reduct... | NeurIPS_2023_submissions_huggingface | 2,023 | Summary: This paper studies the problem of accelerating vision transformers for dense prediction without fine-tuning and proposes a spatial-aware density-based clustering algorithm to select anchors from the token sequence. The proposed method is evaluated across three different
datasets and demonstrates promising perf... | Rebuttal 1:
Rebuttal: **Q1: Other dense prediction tasks like instance segmentation and depth estimation**
**A1:** Thanks for your suggestion! We have presented the performance on object detection and instance segmentation in the appendix and summarize the results here for your reference. Specifically, we integrate Ai... | null | null | null | null | null | null |
RoboCLIP: One Demonstration is Enough to Learn Robot Policies | Accept (poster) | Summary: This study seeks to address the problem of data efficiency in imitation learning, aiming to train a robot to perform manipulation tasks after only one demonstration. In order to accomplish this, the research proposes using a pre-trained visual language model (VLM) to encode the demonstration video or task desc... | Rebuttal 1:
Rebuttal: Thanks for your detailed review! Addressing your comments has improved the paper significantly! :)
We address your concerns below:
1. **“Why use dot products between embeddings as the reward function?”**: We will add the following to the text: VLMs such as VideoCLIP and S3D are pretrained using a... | Summary: RoboCLIP is an innovative online imitation learning method that addresses the challenging problem of reward specification in reinforcement learning. Unlike traditional approaches that rely on expert supervision and extensive data, RoboCLIP utilizes a single demonstration, either in the form of a video or a tex... | Rebuttal 1:
Rebuttal: Thanks for your comments! We're glad that you would like to see our work published!
We address your concerns below:
1. **“Out-of-distribution results are underwhelming? (Sec 4.4)”**: Indeed, especially with rewards from OOD videos and language, the robot policy often approaches the target object,... | Summary: This paper presents a novel approach to leverage pretrained Vision-Language Models (VLMs) for training Reinforcement Learning (RL) agents by utilizing them to generate rewards. Specifically, the authors employ S3D, pretrained on HowTo100M, as the backbone and measure the dissimilarity between online agent expe... | Rebuttal 1:
Rebuttal: Thank you very much for your insightful comments! Addressing them has made this work much better! :)
We address each of your concerns below:
1. **“VLM confuses open and close, the domain alignment is not good?”**: The alignment metric for the confusion matrix was calculated by aligning a single de... | Summary: This paper proposes using the similarity between two embedded spaces, one from real interaction and the other from visual/textual interaction, to develop the rewards.
Strengths: It's a well-written paper. The idea is simple and clearly stated. The proposed method has been sufficiently evaluated on realistic s... | Rebuttal 1:
Rebuttal: Thank you for your thoughtful review! We are glad you would like to see our work published! :)
We address your questions below:
1. **“Why submit to NeurIPS over robot conferences?”**: Thanks for your suggestion! Sequential decision-making is a generally useful problem to the community and recent... | Rebuttal 1:
Rebuttal: We would like to thank the reviewers for their insightful comments and suggestions. They have helped improve the paper significantly and we are grateful for their time and effort :). We have run 4 additional experiments to address their concerns. We summarize the findings below. For details, pleas... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Chasing Fairness Under Distribution Shift: A Model Weight Perturbation Approach | Accept (poster) | Summary: This paper tackles the issue of fairness in situations where there are distribution shifts, viewing it from the angle of model weight perturbation. The authors establish a theoretical link between distribution shift, data perturbation, and model weight perturbation, leading to the conclusion that distribution ... | Rebuttal 1:
Rebuttal: Thank you for spending time and effort reviewing our paper. We appreciate your constructive comments and support to our paper. This feedback truly encouraged us to redouble our efforts in conducting quality and impactful research.
**[Q1] The hyperparameter tuning seems important. Do you record t... | Summary: The paper is in the context of fairness measures under distribution shift.
There is first a theoretical connection between distribution shift, data perturbation and model perturbations.
The proposed method Robust Fairness Regularizatin (RFR), proposes a regularization using 3 terms L_CLF-- classifier loss fun... | Rebuttal 1:
Rebuttal: We thank the reviewer for the constructive comments.
**[W1] The method is designed/tested in Neural Network architectures in tabular datasets. In tabular data algorithms like gradient boosting decision trees achieve SOTA[1] and are restricted to Linear Regression in regulated environments.**
**A... | Summary: theoretical proof: prove that distribution shifting can be written as data perturbation, and the model trained with the perturbed data can be written as the perturbed model trained with original source data.
methodology: motivated by the connection between distribution shifting and data weight perturbation, th... | Rebuttal 1:
Rebuttal: We thank the reviewer for the constructive comments.
**[W1] I think that the objective function which the author suggested did not reflect the true effect of $L_{RFR}$. Please add the experiments without $L_{DP}$.**
**Ans:** We thank the reviewer for this constructive comment. As required by the... | Summary: In this work, the authors seek to learn neural networks that are fair not just w.r.t. the observed distribution of individuals, but also under shifts to this distribution. In order to achieve this, they propose robust fairness regularization (RFR) which uses adversarial weight-space perturbations at train time... | Rebuttal 1:
Rebuttal: We thank the reviewer for the constructive comments.
**[Q1] The major weakness that I see with this paper is the underlying theoretical motivation. Theorem 2.3 appears to have a critical flaw. (1) There is no given relationship between the magnitude of the perturbations in either domain and thus... | Rebuttal 1:
Rebuttal: We thank all reviewers for the constructive comments and the positive evaluation for most reviewers. We provide the summary for the key revision.
* Add experiments on hyperparameter study on perturbation $\rho$.
* Add comparison results (Table. 1 and Figure 1 in one-page rebuttal) for the setting ... | NeurIPS_2023_submissions_huggingface | 2,023 | Summary: The paper studies the problem of training a fair ML model in the presence of distribution shift. The proposed method aims to bound the biases on the target distribution and use it as a regularizer.
Strengths: S1 The problem of distribution shift is quite fundamental.
S2 The idea of bounding the biases and us... | Rebuttal 1:
Rebuttal: We thank the reviewer for the constructive comments.
**[W1] The approach of using the worst-case bound as a regularizer has been explored before for selection bias [1]. Please discuss the differences with this approach.**
**Ans:** We will add the discussion on the difference with [1]. The diffe... | null | null | null | null | null | null |
R-divergence for Estimating Model-oriented Distribution Discrepancy | Accept (poster) | Summary: A new divergence metric, called R-divergence, is proposed in the paper. This metric states that two distributions are identical if the optimal hypothesis on the mixture distribution has the same expected risk on each individual distribution.
If two distributions are the same, then the optimal hypothesis on the... | Rebuttal 1:
Rebuttal: We are grateful for your comments and address them below.
**Q1. Different corruptions:**
Thank you for your insightful comments and suggestions. Following the construction methods of corrupted samples in [56] and the setup of H-divergence [63], we have performed experiments of R-divergence with ... | Summary: This work proposes R-divergence as a measure of discrepancy between two distribution. The R-divergence is the difference between the empirical risks computed on two datasets using a hypothesis trained on a mixture of both datasets. The authors use the R-divergence and the permutation test to identify if two da... | Rebuttal 1:
Rebuttal: We appreciate your comments, address them while also clarify potential misunderstandings below.
**Q1. Advantages of R-divergence:**
R-divergence is a good measure for both theoretical and practical advantages because R-divergence benefits from addressing the over-fitting issue and estimates the... | Summary: This paper proposes R-divergence to measure the distributional discrepancy of two datasets. The general idea is to learn a minimum hypothesis on the mixed data and then define R-divergence as the difference of the empirical risk of the learned hypothesis between two datasets. The proposed divergence leverages ... | Rebuttal 1:
Rebuttal: We are grateful for your comments, below we address all and also clarify potential misreading and misunderstanding.
**W1. Significance of R-divergence:**
R-divergence is significantly different from L- and H-divergence, discussed in the last paragraph of Introduction and Table 6. R- and L-dive... | Summary: A new distance for comparing two samples of two distributuions p and q is proposed that is based on learning a model on the mixtre distribution of p and q. More precisely, the distance is the difference between errors of the learned model evaluated on the individual distributions. Theoretical statements on the... | Rebuttal 1:
Rebuttal: We thank your constructive comments.
**W1. Theoretical upper bound:**
We agree that a large L-divergence $\mathcal{D}_{\textup{L}} (p || q)$ will affect the convergence of the empirical estimator. However, a large L-divergence contributes to identifying whether samples of two datasets are drawn... | Rebuttal 1:
Rebuttal: We are grateful for all the reviewers for their valuable feedback and insightful comments on our paper.
We are heartened to see that two of the reviewers have acknowledged the significance, contribution and quality of our work, where both assign a score of 7. We appreciate their positive remarks... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Causal Interpretation of Self-Attention in Pre-Trained Transformers | Accept (poster) | Summary: The paper establishes a formal link between a (pre-trained) self-attention layer and a causal graph underlying a sequence of symbols. Specifically, it shows that self-attention learns to represent pairwise associations for which there exists a linear-Gaussian Structural Causal Model that has the exact same pai... | Rebuttal 1:
Rebuttal: We sincerely thank you for your review, and the important questions. We hope that our response below makes this paper's contribution clearer, and mitigates the concerns you raised.
**Re: weaknesses.**
We wish to clarify that the main contribution of the paper is twofold. The first is a link betwe... | Summary: The ABCD paper starts off by motivating the need for understanding the complex dependencies that underly current Transformer architectures, pointing to causality as the missing puzzle piece. Following brief introductions of both SCMs and Self-attention, the authors then provide covariance matrix derivations fo... | Rebuttal 1:
Rebuttal: We wish to thank you for you detailed review, positive feedback, and identifying weakness that addressing them significantly improves this paper.
**Re: Weaknesses.**
1. **(a) Re: Causal Markov and faithful assumptions.** The causal Markov and faithfulness assumptions are required for constraint-b... | Summary: In this work, authors, interpret self attention in transformers as linear gaussian SCM over output embeddings.
SCM over input embeddings is same as SCM over output embeddings, since deepest attention layer is trained to predict input symbol.
Existing conditional independence based testing algorithm is applied... | Rebuttal 1:
Rebuttal: First, we wish to thank you for your review, positive feedback, and questions that allowed us to further improve this contribution.
**Re: Weaknesses.**
Thank you for spotting the cut-off sentence, and finding it complete in the supplementary material. We will correct that.
**Re: Questions.**
Eac... | Summary: The paper proposes Attention-Based Causal Discovery (ABCD), a causal interpretation for the attention matrix of transformers (e.g., LLMs like BERT). The authors suggest CLEANN, an algorithmic approach that first finds a causal structure of the input symbols (e.g., tokens (words) of a sentence as input to BERT)... | Rebuttal 1:
Rebuttal: We wish to thank you for the very detailed review and your positive feedback. Also, thank you for your comments for improvements that mitigate potential weaknesses. Addressing them as suggested strengthens the contribution and makes it clearer.
**Re: Your comments for improvements (weaknesses).**... | null | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
ZoomTrack: Target-aware Non-uniform Resizing for Efficient Visual Tracking | Accept (spotlight) | Summary: The paper studies a very interesting problem in local visual tracking with crop-then-resize operations: if it is possible to reduce the high computational demand of attending to a large visual field while retaining its potential benefit of achieving SOTA tracking accuracy. To this end, the paper takes inspirat... | Rebuttal 1:
Rebuttal: Thank you for your review and feedback! We are happy that you find our method quite interesting and solid, which studies a little-explored problem, and acknowledge that the performance gain on LaSOT is impressive and the experiments clearly show the superiority of our elegant solution. We provide ... | Summary: The authors propose a non-uniform resizing method for visual tracking, where they aim to boost the performance of transformer-based visual tracking methods at negligible additional computational cost. They achieve this by reducing the input size while attending to the relevant parts of the target, and this app... | Rebuttal 1:
Rebuttal: Thank you for your review and feedback! We are happy that you find our method technically plausible while bringing adequate enlightenment to the visual tracking field by studying an easily overlooked problem. We provide responses to the specific points below:
**About more experimental validations... | Summary: This paper propose a novel method to achieve high tracking speed with smaller input size. Non-uniformly resizing operation is introduced with respect to the object appearing probability. This resizing operation is then modeled as QP problem and can be effectively soved. This proposed method can be a flexible ... | Rebuttal 1:
Rebuttal: Thank you for your review and feedback! We are happy that you find our method is novel and can be a flexible plugin for other deep trackers. We provide responses to the specific points below:
**About more experimental validations on performance-oriented trackers to further show the generalization... | Summary: The paper presents a ZoomTrack, a target-aware non-uniform resizing technique for efficient visual tracking. It narrows the gap between speed-oriented and performance-oriented trackers by non-uniformly resizing the cropped image to have a smaller input size while retaining more raw information for the target. ... | Rebuttal 1:
Rebuttal: Thank you for your review and feedback! We are happy that you find our paper well written and acknowledge that we are the first time to apply non-uniform resizing to the field of object tracking. We provide responses to the specific points below:
**About the comparison with CVPR 2023 work "Learni... | Rebuttal 1:
Rebuttal: Thank you for the valuable reviews pointing out that our *simple but clever* (`Gq4E`) and *novel* (`Fe6p`) method, focusing on *a very interesting problem which is little-explored* (`2uk1`) and *easily overlooked* (`P4Ao`), *makes a large step towards closing the gap between fast and accurate trac... | NeurIPS_2023_submissions_huggingface | 2,023 | Summary: This paper proposes a simple but clever method to improve the accuracy-speed trade-off of state-of-the-art methods for single object tracking. Existing methods can achieve greater accuracy by using higher-resolution images, however this comes at the cost of inference speed. This paper proposes to adopt a non-u... | Rebuttal 1:
Rebuttal: Thank you for your review and feedback! *We will release our code upon acceptance and include the new analyses below in the revision*. We provide responses to the specific points below:
**About the edits**: We will revise the edits accordingly. The citation for 100MB/1MB in the human vision syste... | null | null | null | null | null | null |
Transfer Learning with Affine Model Transformation | Accept (poster) | Summary: The paper provides a comprehensive investigation of transfer learning and proposes an affine model transfer framework. The authors summarize various transfer learning approaches under the class of affine model transformation and establish theoretical properties such as generalization error and excess risk.
In ... | Rebuttal 1:
Rebuttal: We express our gratitude to the reviewer for providing us with valuable feedback. Below is our point-by-point response to each question.
> **[Q1]** I would like to know that why were the experiments not conducted on commonly adopted transfer learning datasets, such as MIT Indoors, Caltech 256, an... | Summary: The authors consider the transfer learning setting by affine models adapting source features to the target domain by separately estimating cross-domain shift and domain-specific factors. They motivate the choice of such models deriving the affine form as an optimal class based on the squared loss for the targe... | Rebuttal 1:
Rebuttal: Thank you very much for the valuable comments. Below we respond to each concern / question raised in your review.
> **[W1]** Considering only the squared loss is quite limiting. The authors comment about this, but it would be nice to understand if there is possibility to extend their framework to... | Summary: This paper presents a simple yet effective approach to transfer learning in regression problems by introducing an affine transform model. The proposed model incorporates the source feature function and the product of the source feature function and the function associated with the target input augmentation. Th... | Rebuttal 1:
Rebuttal: Thank you for your time in writing in-depth feedback. Please find our response to your concerns below.
> **[W1]** The existence of inverse functions appears to be of significance. However, it can be considered a potential limitation as it may not always be feasible to obtain inverse functions in ... | Summary: This paper focuses on the theoretical perspective on the affine model transfer learning. Specifically, this paper presents the optimal class of models on transfer learning with l2 loss takes the form of an affine decoupling. The decoupling consists of three functions, \$g_1 + g_2g_3\$, where \$g_1, g_2\$ encod... | Rebuttal 1:
Rebuttal: We are delighted to receive a positive evaluation from the reviewer. Below is our point-by-point response to each question.
> **[Q1]** All the discussions are limited to the $\mathbb{R}$. It remains elusive whether the analysis can be extended to higher dimensional spaces naturally.
When $y$ and... | null | NeurIPS_2023_submissions_huggingface | 2,023 | Summary: The authors focus on the hypothesis transfer learning (HTL) problem. They show that under some hypothesis the data and model transformation is given by an affine transformation. They consider an associated kernel regression problem to identify the transformation and then characterise the generalisation bound w... | Rebuttal 1:
Rebuttal: We would like to express our gratitude to the reviewer for providing insightful comments. Below is our point-by-point response to each of the comments and questions.
> **[W1]** The model is not clear (sec.2.1). Maybe adding a scheme or introducing the linear models as a first example before movin... | null | null | null | null | null | null |
A Theory of Unsupervised Translation Motivated by Understanding Animal Communication | Accept (poster) | Summary: This paper proposes a theoretical framework for understanding unsupervised translation and provides bounds for sample complexity. Specifically, the unsupervised machine translation setting they consider is that the translator $f_\theta$ is the one that maximizes the plausibility $\rho$ of the translation of ex... | Rebuttal 1:
Rebuttal: > **I doubt whether Theorem 3.4 is correct.** ...
You indeed found a typo in our proof overview intuition in section C.1.1. However, the theorem and proof are correct. We apologize for the mistake on line 717 and will correct it. Thank you for identifying it! Let us explain first the intuition an... | Summary: ### After rebuttal
Thanks the authors for the detailed rebuttal, especially with new experiments. Provided the condition that the authors will add the new experiments and adjust the presentation after reviewers' comments, I gladly change my scores and recommendation!
----
The paper presents a rather interest... | Rebuttal 1:
Rebuttal: > There is no experiments to validate the theory at all, even with a toy experiments.
Thank you for this great suggestion. To address this, we have added synthetic experiments to validate our theoretical results. See the pdf attached to our overall response.
> The paper is very math heavy with l... | Summary: This paper introduces a theoretical framework for unsupervised machine translation (UMT) when no parallel data available and source and target language do not share similar linguistic structures. To do this, they first define and clarify the three main challenges: understanding the goal, no linguistic structur... | Rebuttal 1:
Rebuttal: Thank you for your review and feedback.
> The paper does not provide any empirical validation of the proposed theoretical framework and models.
Thank you for this great suggestion. To address this, we have added synthetic experiments to validate our theoretical results. See the pdf attached to ... | Summary: The paper describes a theoretical analysis of error bounds for unsupervised machine translation, motivated by considering the unusual, and so far only speculative, case of translation between animal "languages" to human natural languages.
The paper consider the case where a number of examples of the source (a... | Rebuttal 1:
Rebuttal: Thank you for your review and feedback. We are glad you find the topic novel and interesting.
> Assumptions not well justified
We make the assumptions and justification more explicit in the paper, as follows. The two assumptions underlying all our models, are:
1. There exists a hypothetical tra... | Rebuttal 1:
Rebuttal: # Overall Response
**BREAKING NEWS**: We have added experiments as requested by all reviewers. These synthetic experiments validate the theoretical findings and help illustrate the main message, that more complex languages can be translated with lower error. The focus of these experiments, like o... | NeurIPS_2023_submissions_huggingface | 2,023 | Summary: ### Update
I'm satisfied with the author's rebuttal and decided to overall improve my score.
This paper proposes a theoretical framework for analyzing Unsupervised MT motivated by animal communication. The paper is only the first step in the theoretical analysis of UMT. The theoretical framework was analyze... | Rebuttal 1:
Rebuttal: Thank you for your review and detailed feedback.
Weaknesses:
> Purely descriptive: what is the key takeaway from the paper? How can we improve unsupervised MT based on these findings? Where can we use the theoretical framework?
We believe strongly that our theoretical findings are illuminating d... | Summary: The first contribution of this work is formalizing and analyzing a model of Unsupervised Machine Translation (UMT). The model applies even to low-resource source languages with massive domain gap and linguistic distance. Second, the paper exhibits two simple complementary models for which it is proved that:
(a... | Rebuttal 1:
Rebuttal: Thank you for your enthusiastic review and recognition of the importance of our findings, and more generally of introducing a theoretical UMT framework.
> There isn't any empirical evaluation aside from using text-davinci-02 to evaluate sentence likelihoods
Thank you for this great suggestion. ... | null | null | null | null |
Off-Policy Evaluation for Human Feedback | Accept (poster) | Summary: This paper proposes an off-policy evaluation method that utilizes sparse human feedback. Experiments are conducted on two real-world environments and one simulation environment.
Strengths: The idea of incorporating human feedback into policy evaluation is interesting and significant. The proposed framework is... | Rebuttal 1:
Rebuttal: We greatly appreciate the reviewer's thoughtful questions and constructive feedback. Please find our responses below. We hope they addressed the reviewer's outstanding concerns sufficiently, and we would be more than happy to further respond to any follow-ups the reviewer may have.
Q1: Are 4 pati... | Summary: The authors examine an off-policy evaluation scenario where we have an environment with an intrinsic environment reward $r$, and an extrinsic human reward $r^H$. A further assumption is made that the human reward is only observed at the end of the episode, and comes as an rating on the total return of the epis... | Rebuttal 1:
Rebuttal: Please find our responses below. We hope they addressed the reviewer's outstanding concerns sufficiently, and we would be more than happy to respond to any follow-ups the reviewer may have.
Q1:The bulk of the paper is really about learning a reward function. It could theoretically be done online ... | Summary: This paper discusses a framework that learns to decompose human feedback reward, which is usually sparse and only given at the end of the episode, to intermediate step reward. They formulate a reward decomposition objective that considers original environment trajectory information using a variational auto-enc... | Rebuttal 1:
Rebuttal: We greatly appreciate the reviewer's thoughtful questions and constructive feedback. Please find our responses below. We would be more than happy to further respond to any follow-ups the reviewer may have.
Q1: The methodology proposed in this paper is not very novel — the authors assembled a few ... | Summary: The paper introduces a framework for Off-Policy Evaluation for Human Feedback (OPEHF). Off-policy evaluation allows the evaluation of learned policies on offline recorded data, which is especially valuable in scenarios where online deployment to evaluate policies is dangerous or expensive. Especially in scenar... | Rebuttal 1:
Rebuttal: We greatly appreciate the reviewer's thoughtful questions and constructive feedback. Please find our responses below. We hope they addressed the reviewer's outstanding concerns sufficiently, and we would be more than happy to further respond to any follow-ups the reviewer may have.
Q1: While I ve... | Rebuttal 1:
Rebuttal: # General Response
We sincerely thank the reviewers for the time and efforts dedicated to the review process.
We feel that currently the main discrepancy centered around the evaluation of OPEHF. While reviewers KHGy and cuXN pointed that our real-world experiments are convincing and comprehensi... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Weakly Coupled Deep Q-Networks | Accept (poster) | Summary: The paper studies value-based control agents for Weakly Coupled MDPs (WCMDPs), which comprise independent subproblems linked by an action constraint. The authors generalize an existing Lagrangian relaxation approach to action-value methods learned by stochastic approximation to obtain Weakly Coupled Q-Learning... | Rebuttal 1:
Rebuttal: We thank the reviewer for their time and effort spent providing a thorough review of the paper and for the positive comment on the writing (“Self-contained descriptions of Weakly Coupled MDPs and Lagrangian relaxation, as well as a complete related works discussion, make the paper accessible... Th... | Summary: The authors present a new algorithm, "WCDQN" for solving structured problems know as weakly coupled MDPs. The method dynamically estimates upper bounds via Lagrangian relaxation essentially to improve the value of collected experience across the structured subproblems. The authors present a (re)formulation of ... | Rebuttal 1:
Rebuttal: We thank the reviewer for their kind comments (“Nicely written, enjoyable reading", "relatively simple and clear problem formulation”). It seems that the main concern that the reviewer had is due to the lack of an experiment showing the effect of the number of subproblems on the solution quality. ... | Summary: This paper considers the problem of learning optimal policies for a weakly coupled Markov decision process (WCMDP). The paper extends the classical DQN algorithm to this setting by introducing separate lagrangian relaxations. Experiments are conducted on 3 operational research testbeds.
Strengths: This paper ... | Rebuttal 1:
Rebuttal: We thank the reviewer for acknowledging the strengths of our paper: “This paper has a very complete structure, with a detailed problem introduction, an insightful algorithm derivation, a theoretical analysis as well as solid numerical studies. The presentation is particularly clear and easy to fol... | Summary: This paper introduces novel Q-learning based approaches (WCQL and WCDQN) to solving weakly-coupled Markov decision processes. The approaches work by running Q-learning for the sub-problems of the WCMDP to construct an upper bound on the optimal Q-values of the overall problem. The WCMDP constraint is incorpora... | Rebuttal 1:
Rebuttal: We thank the reviewer for recognizing the originality (“proposed approaches appear to be simple and relatively original”) and the importance of our work (“the problem setting is well-motivated”, “represents a significant advance for this problem setting”). We’d appreciate it if the reviewer would ... | Rebuttal 1:
Rebuttal: We thank all reviewers for their thoughtful comments and suggestions. Since receiving the reviews, we have completed a new experiment that studies the performance improvement from WCQL over vanilla QL as the number of subproblems increases, as suggested by Reviewer da8o. This is a great suggestion... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Crystal Structure Prediction by Joint Equivariant Diffusion | Accept (poster) | Summary: DiffCSP is proposed in this paper for the task of crystal structure prediction task, which generates the stable crystal structures of given compositions. The diffusion process of DIffCSP considers both lattice matrix and fractional coordinates of atoms in the unit cell. Experimental results demonstrate the eff... | Rebuttal 1:
Rebuttal: We are thankful to the reviewer for the constructive comments. We answer the reviewer’s questions as follows.
> **W1: Limited novelty.**
There could be some misunderstandings here. We admit that the way we diffuse lattice matrix is similar to previous works for molecule conformer generation, bu... | Summary: DiffCSP is a novel diffusion model for predicting crystal structures conditioned on the crystal's atom composition; it can also be extended to an initio crystal generation (sampling both atom composition and positions). Crystal modeling is challenging due to its periodic symmetry. DiffCSP proposes to solve per... | Rebuttal 1:
Rebuttal: We are thankful to the reviewer for the detailed comments. We answer the reviewer’s questions as follows.
For weaknesses:
> **W1: Why a generative viewpoint of CSP is beneficial.**
Thank you for your question. The generative viewpoint in CSP holds substantial benefits, particularly in addressin... | Summary: This paper addresses the crystal structure prediction problem, one of the important problems in this field.
The authors propose a diffusion model generating both lattices and fractional coordinates given chemical composition.
The proposed denoising model satisfies O(3) equivariance and periodic translation inv... | Rebuttal 1:
Rebuttal: We are thankful to the reviewer for the constructive comments. We answer the reviewer’s questions as follows.
> **Q1: The study does not report any new crystal structures. It will increase the impact of this paper if the proposed method can find a new crystal that achieves lower energies via DFT ... | Summary: The paper studies crystal structure prediction, an important topic in the field of AI for scientific discovery. Diffusion models are designed considering necessary symmetries for crystal study, with diffusing lattice vectors and fractional coordinates for generating lattices and atoms, respectively. Experiment... | Rebuttal 1:
Rebuttal: We are thankful to the reviewer for the constructive comments. We answer the reviewer’s questions as follows.
> **Q1: The technical contribution is not clear or strong enough. Diffusion models for 3D scientific data are not new. The authors may need to clearly state the significant and unique des... | Rebuttal 1:
Rebuttal: We sincerely thank all reviewers and ACs for their time and efforts on reviewing the paper. We are glad that the reviewers recognized the contributions of our paper, which we briefly summarize as follows.
- **Novelty.** "The paper presents novel and shining ideas."(1HBn)"The problem setting is ne... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
On the Stability-Plasticity Dilemma in Continual Meta-Learning: Theory and Algorithm | Accept (poster) | Summary: This paper aims to address the balance between stability (preserving past knowledge) and plasticity (rapid learning from new experiences) in Continual Meta-Learning (CML). The authors propose a novel theoretical framework for CML in both static and shifting task environments. They discuss the bi-level (task- a... | Rebuttal 1:
Rebuttal:
We sincerely thank the reviewer for appreciating the strength of our work and the constructive comments we can incorporate into the revision. Below are our responses to your comments.
> Weakness 1. “Real-world examples for practicality and necessity of CML.”
* A representative real-world applic... | Summary: The paper considers a continual meta-learning setting with both static and dynamically changing environments and seeks to study the trade-off between learning and forgetting. It derives an upper bound on the excess risk, i.e. the expected gap between the true risk of the predictor and the optimal true risk and... | Rebuttal 1:
Rebuttal: We thank the reviewer for the insightful comments on the experimental details. Below are our responses.
> Weakness 1. Novelty and citation
- We appreciate the suggestion for a more prominent citation of [1], which used a bound to derive the meta-objective for ***statistical meta-learning***. ... | Summary: This paper theoretically studies the stability and plasticity dilemma in Continual Meta-Learning (CML).
This paper formulates the CML objective as selecting u_t at each time t to ensure a small Average Excess Risk (AER) upper bound for the task sequence. Based on the AER objective, it proposes a continual me... | Rebuttal 1:
Rebuttal:
We sincerely thank the reviewer for appreciating the strength of our work and the constructive suggestions. Below are our responses to your comments.
> The paper lacks a clear figure that contextualizes the base model, learner, and how continual meta-learning fits into the setup. This visual re... | Summary: The authors present a framework for continual meta-learning, that is the training of meta-learning models through exposure to a sequence of tasks. The paper presents a formalism for the setting which allows for non-stationary task distributions.
Theoretical analysis is provided both for the framework and for t... | Rebuttal 1:
Rebuttal: We sincerely thank the reviewer for appreciating our work and the constructive comments. Below are our responses.
> "...if comparison with non-continual meta-learning method is fair."
* Thanks for mentioning this point. Due to the page limit, the detailed description for the baselines was not in... | Rebuttal 1:
Rebuttal:
We appreciate the diligent efforts of the reviewers in thoroughly reviewing our paper and providing insightful feedback. In response to the valuable comments, we have included supplementary figures and tables to better address the raised concerns.
Pdf: /pdf/34ac5c4451f0d81a7ac1fe290131de181e891a6... | NeurIPS_2023_submissions_huggingface | 2,023 | Summary: The paper presents a comprehensive theoretical framework for continuous meta-learning in both static and shifting task environments. It also conducts a theoretical analysis of the bi-level trade-off in shifting environments. Furthermore, the paper proposes a method to adapt meta-parameters and the correspondin... | Rebuttal 1:
Rebuttal: We sincerely appreciate the reviewer's efforts to provide constructive comments which we can incorporate in the revision. Below are our responses to your comments.
> Q1. "Both continual meta-learning and continual learning share the common objective," "What is the key difference between them?"
*... | null | null | null | null | null | null |
Optimal Convergence Rate for Exact Policy Mirror Descent in Discounted Markov Decision Processes | Accept (poster) | Summary: This paper proposes a new analysis of the policy mirror descent algorithm and achieves a linear convergence rate with exact $Q$-function that does not contain any distribution mismatch coefficient. Furthermore, it also proposes lower bounds showing that the linear convergence rate is optimal under specific set... | Rebuttal 1:
Rebuttal: We thank reviewer GGgc for their valuable comments and feedback. We have now performed numerical experiments to validate our theoretical findings (see “global” Author Rebuttal and attached pdf), which we will include in the final version.
### D.1 What are the difficulties of extending t... | Summary: This paper studies the convergence rate of a general family of policy gradient methods called Policy Mirror Descent (PMD) in the setting of discounted MDP, unregularized objective, and tabular policy model. In prior work ((Xiao, 2022), [7] in the paper), it has been known that with a geometrically increasing s... | Rebuttal 1:
Rebuttal: We thank reviewer TTxd for their valuable comments and feedback. We have now performed numerical experiments to validate our theoretical findings (see “global” Author Rebuttal and attached pdf), which we will include in the final version.
### C.1 I am not very sure about the significanc... | Summary: This work considers the class of Policy Mirror Descent (PMD) methods and establishes a linear $\gamma$-rate which is dimension-free using adaptive step sizes. In particular, the convergence rate does not depend on visitation distribution mismatch coefficients or the state/action space sizes. It is further show... | Rebuttal 1:
Rebuttal: We thank reviewer YE79 for their valuable comments and feedback and for pointing our typos. We have now performed numerical experiments to validate our theoretical findings (see “global” Author Rebuttal and attached pdf), which we will include in the final version.
### B.1 Weaknesses
-... | Summary: The paper studies the convergence rate of Policy Mirror Descent. The authors prove a $\gamma$-rate for exact PMD update and a matching lower bound for this class of algorithm. Then they prove a sample complexity bound of $O(\epsilon^{-2}(1-\gamma)^{-8})$.
Strengths: The paper analyses an interesting problem... | Rebuttal 1:
Rebuttal: We thank reviewer wj3u for their valuable comments and feedback. We have now performed numerical experiments to validate our theoretical findings (see “global” Author Rebuttal and attached pdf), which we will include in the final version.
### A.1 In the reviewer's opinion the paper does... | Rebuttal 1:
Rebuttal: We have now performed numerical experiment to validate our theoretical findings (see attached pdf), which we will include in the final version.
The plots in the attached PDF are made by running NPG (PMD with negative-entropy mirror map, see [9]) on the hard MDP used to prove Theorem 4.2. We use $... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Don’t just prune by magnitude! Your mask topology is a secret weapon | Accept (poster) | Summary: The paper suggests two pruning strategies: pruning at initialization as graph sampling (PAGS) and pruning early as graph sampling (PEGS). The PAGS approach is one of those that doesn't call for any pre-training dietary intake. The mask is oversampled each time, and the masks with the greatest l2 movement dista... | Rebuttal 1:
Rebuttal: 1) The actual method suggested is relatively naive, and too much length in the paper is devoted to analysis and introduction.
> We thank the reviewer for their comment. We understand that both PAGS and PEGS can be considered as simple greedy algorithms. However, we consider this to be an advantag... | Summary: This paper proposes a new method for pruning deep neural networks that optimizes both the network's structure and weights. The experimental results show that it can be naturally applied as a “zero-shot" PaI method and outperforms existing PaI methods with large margins. Alternatively, it can be applied after l... | Rebuttal 1:
Rebuttal: 1) Despite the appeal of theory, I don’t think the current experiments are sound enough. In general, all pruning methods in-use tend to perform similarly on CIFAR-10/100. As seen in Tables 1 and 2, different methods’ numbers are often within each other’s error bar, making it hard to judge whether ... | Summary: This paper proposes a new approach to deep network pruning based on graph theory, specifically the use of Ramanujan graphs. The authors argue that their approach, which focuses on optimizing both the graph's structure and weights, can achieve notable performance improvements across various pruning methods whil... | Rebuttal 1:
Rebuttal: 1) I have two major critiques: (1) poor readability, and (2) unimpressive, sometimes confusing experiments.
* Poor readability
* Experiments: there are several presented results confusing me. Why VGG shows much poorer linear correlation in Figures 1-4, compared to ResNets? Also the authors didn’t... | Summary: This work studies network pruning in a way that takes into account both the graph topology and values of weights. This connects purely magnitude-based and purely graph-based pruning methods. In particular, the authors find correlations between performance of masks and 4 properties derived per layer of the neur... | Rebuttal 1:
Rebuttal: 1) Page 5's "physics intuition" is hard to understand and unrigorous, and does not seem to do much to alleviate the mentioned "lack of theory" from the Pal et al. approach.
> We appreciate the reviewer's feedback. In our paper, we describe $\Delta r$ in graph theory as the degree of connectivity ... | Rebuttal 1:
Rebuttal: We thank all our reviewers for their feedback and will try to address each concern individually. While the reviewers tend to have consensus on the novelty and contribution of our work, there are two common criticisms that we wish to respond globally, namely:
* Insufficient experimentation
* Reada... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
First Order Stochastic Optimization with Oblivious Noise | Accept (poster) | Summary: This paper initiates the study of stochastic optimization with oblivious noise that might be biased and have unbounded variance, which is a generalization of the heavy-tailed noise setup. The key assumption regarding the oblivious noise is that it assumes a value of $0$ with a probability within the range of $... | Rebuttal 1:
Rebuttal: We first give a common response, and then addresses the specific concerns of the reviewer.
**Common Response**
To summarize, our paper investigates a generalization of the heavy-tailed stochastic optimization problem. As highlighted by reviewers M73n, NDsr, and 2f9g, our work addresses a challen... | Summary: The paper presents an algorithm for first-order stochastic optimization where the algorithm has access to an oracle that returns a noisy version of the gradient of the objective function. The considered noise model includes two components: A bounded-variance observation noise (which is the typical well-studied... | Rebuttal 1:
Rebuttal: We first give a common response, and then addresses the specific concerns of the reviewer.
**Common Response**
To summarize, our paper investigates a generalization of the heavy-tailed stochastic optimization problem. As highlighted by reviewers M73n, NDsr, and 2f9g, our work addresses a challen... | Summary: The paper considers the problem of stochastic optimization with oblivious noise. Here, one receives noisy gradients of a non-convex function $f(x) = \mathbb{E} [f(x, \gamma)]$ (where $\gamma$ is bounded variance observational noise) and the noisy gradient samples are generated as follows $\nabla_x f(x, \gamma)... | Rebuttal 1:
Rebuttal: We first give a common response, and then addresses the specific concerns of the reviewer.
**Common Response**
To summarize, our paper investigates a generalization of the heavy-tailed stochastic optimization problem.
As highlighted by reviewers M73n, NDsr, and 2f9g, our work addresses a chall... | Summary: This paper studies robust first order optimization in a challenging setting where noise may be unbounded, a setting that arises often for real world optimization problems. Because this problem is intractable in general, one needs to make plausible assumptions on the noise, that are realistic on one hand but al... | Rebuttal 1:
Rebuttal: We first give a common response, and then addresses the specific concerns of the reviewer.
**Common Response**
To summarize, our paper investigates a generalization of the heavy-tailed stochastic optimization problem.
As highlighted by reviewers M73n, NDsr, and 2f9g, our work addresses a chall... | null | NeurIPS_2023_submissions_huggingface | 2,023 | Summary: This paper introduces a new setup for stochastic optimization, where in addition to random observation noise, the stochastic gradient may be subject to independent oblivious noise. This noise might not have bounded moments and isn't necessarily centered. The authors propose a Noisy Location Estimation approach... | Rebuttal 1:
Rebuttal: We first give a common response, and then addresses the specific concerns of the reviewer.
**Common Response**
To summarize, our paper investigates a generalization of the heavy-tailed stochastic optimization problem.
As highlighted by reviewers M73n, NDsr, and 2f9g, our work addresses a chall... | null | null | null | null | null | null |
Zero-One Laws of Graph Neural Networks | Accept (poster) | Summary: The authors analyze standard GNN binary classifiers on Erdos-Renyi graphs, and conclude that the output satisfies a zero-one law. To be more precise, this means that an Erdos-Renyi graph with the number of nodes tending to infinity and random node features will be classified as 1 with probability tending towar... | Rebuttal 1:
Rebuttal: Thank you for your comments on our paper, which we address below.
> Do the proofs extend easily to different random graph structures and/or wider families of distributions?
We haven’t found an easy extension to different distributions of graphs with node features. Once edges and features become ... | Summary: This paper is primarily a theory paper - proposing a proof that, as the number of nodes in a graph drawn from the ER graph distribution increases, a class of GNNs binary classifiers converges to either 1 or 0.
This observation is backed up with a small set of empirical experiments of GCNs and variants, but no... | Rebuttal 1:
Rebuttal: Thank you very much for your review, and your comments on the clarity of the exposition.
> How would an attention aggregation change the analysis? For me, that would hugely improve the impact of the paper because you're heading towards transformers at that point and so a far wider audience.
The ... | Summary: This paper introduces the zero-one laws pertaining to graph neural networks. More specifically, the authors utilize concentration analysis on Erdos-Renyi graphs to demonstrate that the likelihood of GNN classifiers producing either zero or one as output approaches 1. These findings are further expanded to enco... | Rebuttal 1:
Rebuttal: Thank you for your review and for your comments. We address your concerns below.
> I still lack access to the specific paper mentioned, so I can offer only general suggestions for rephrasing the paragraph based on the information provided: I find it challenging to fully comprehend the correct int... | Summary:
This paper proves that for Erdos-Renyi random graphs, the hidden representations of nodes for GNNs, as well as the graph-level representation, will converge to a constant. In Theorem 4.6, which is the main result for GCN and binary graph classification, they prove that the binary output is always zero or one ... | Rebuttal 1:
Rebuttal: Thank you for your review and your comments on our paper. We address each below.
> (0) One major problem regarding the writing of the paper is that many paragraphs try to deliver the simple concept of the zero-one law for GNNs and its importance/motivation, which was too long for me.
To the best... | Rebuttal 1:
Rebuttal: We wish to thank the reviewers for their comments. We have responded to each concern in detail in our individual responses. In addition, we include a PDF containing the results of additional experiments as suggested by reviewer comments.
Here is a summary of the changes to be made to the paper in... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Propagating Knowledge Updates to LMs Through Distillation | Accept (poster) | Summary: This paper tackles updating the knowledge in LMs, focusing on allowing LMs to make new inferences consistent with the updated facts. To do this, the authors propose using the LM itself (or a teacher) to generate natural continuations for the "updated/new entity" definition. These continuations are used to upda... | Rebuttal 1:
Rebuttal: Thank you very much for the review! We are glad that you found our contribution to be novel and well-motivated. Please let us know if you have any further questions or comments.
---
Rebuttal Comment 1.1:
Comment: Thank you.
Following up on your discussion with reviewer ARok on prompting with up... | Summary: This paper studies the problem of injecting new entity knowledge in LLMs, such that these knowledge can be propagated and utilized when LLMs make inference on related queries. The paper proposes a context distillation method that consists of two steps to inject entity knowledge in a definition sentence: 1) Use... | Rebuttal 1:
Rebuttal: Thank you very much for the review! We are happy you found our contribution and analysis to be novel and interesting.
>On Entity Inferences dataset, the conclusion that the proposed method improves the model ability to make inference using the injected knowledge is suspected. The reported perform... | Summary: This paper proposes a method to propagate knowledge update to LMs via training a student model through context distillation, such that the LM can make inference on an entity even though the relative context/knowledge of the entity is not given. The framework involves two steps: 1) create a transfer set that co... | Rebuttal 1:
Rebuttal: Thank you very much for the review!
>Another baseline is to try prompting the LLMs with new knowledge and see how it propagates. If the existing LLMs can handle such knowledge updates well, it may be hard to justify why we need to train a separate student model.
Our “prepending definition” base... | Summary: The paper propose a context distillation-based approach that can both impart knowledge about entities and propagate that knowledge to enable broader inferences. This approach consists of two stages: transfer set generation and distillation on the transfer set. In the first stage, a transfer set is generated by... | Rebuttal 1:
Rebuttal: Thank you very much for the review! We are glad you found our method to be well-motivated and effective.
>As the authors concede in Section 'Limitations', their proposed methodology has yet to be substantiated on models of a larger scale. For instance, LLaMA-65B may present a fitting candidate f... | Rebuttal 1:
Rebuttal: Thanks very much to each of you for taking the time to review our paper! We are happy you thought our work was insightful and novel. We include additional figures in the PDF attached below and refer to this PDF in our replies.
We will first address common questions, and then address reviewer-spec... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
ANQ: Approximate Nearest-Neighbor Q Learning | Reject | Summary: This paper introduces the ANQ (Approximate Nearest Neighbor Q-Learning) framework, which aims to provide explainability in reinforcement learning models. ANQ combines neural networks for high performance and memory-based structures for explainability, offering a promising solution for domains like autonomous d... | Rebuttal 1:
Rebuttal: Dear Reviewer,
Firstly, we are deeply grateful for your thoughtful review and the recognition of the significance of our work. We truly appreciate your insights on the novelty of our approach, the efficacy of the Sim-Encoder, our evaluation methodologies, and the potential real-world applications ... | Summary: The submission creates a framework called Approximate Nearest Neighbor Q-Learning (ANQ). ANQ uses a sim-encoder contrastive learning and approximate nearest neighbor search to find which states are similar. Utilizing this approach, they can use it to find similar states in aiding for the decision the framework... | Rebuttal 1:
Rebuttal: Dear Reviewer,
We appreciate the detailed feedback you provided and recognize the significance and originality of our approach, especially with respect to the sim-encoder and the ablation experiments across four environments. We would like to address the concerns and confusion you've pointed out.... | Summary: This work proposes a memory-based Q-learning algorithm aimed at better explainability. The method extends the prior work "Episodic Control" and enables learning with continuous action space.
This work provides 2 main contributions:
1. a one-step-away contrastive-learning objective to learn embeddings from sta... | Rebuttal 1:
Rebuttal: Dear Reviewer,
Thank you for your thorough review and insightful feedback. We appreciate your recognition of the potential ubiquity of our approach in real-world applications with continuous state and action spaces. We are pleased to know that the algorithm and experiment settings came across as ... | Summary: Instead of using deep neural networks to approximate Q functions as it's done in deep RL methods, the paper investigates the potential of using nearest neighbor methods to approximate Q functions. They used contrast learning with a Sim encoder. They argue that such a method is more explainable than the ones wi... | Rebuttal 1:
Rebuttal: Dear Reviewer,
We greatly appreciate your insightful comments and the time you took to review our manuscript. Your feedback is essential to improving our work, and here we address your concerns:
Q1: The proposed nearest neighbor is doubted its effectiveness under video games.
A1: We have not ye... | null | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
DreamSim: Learning New Dimensions of Human Visual Similarity using Synthetic Data | Accept (spotlight) | Summary: This paper introduces a new dataset consisting of images generated by prompting the Stable Diffusion models, aiming for studying the mid-level image similarities. Human perceptual judgments are collected for image triplets via both 2AFC and JND tests. A perceptual metric is built upon large pretrained vision m... | Rebuttal 1:
Rebuttal: We thank the reviewer for their comments. We are glad that the reviewer finds the contribution of our dataset to be significant, and the experiments well organized. Below we address specific questions and suggestions.
**Definition of distortion**
We use the term “distortions” to refer to varia... | Summary: In this paper, the authors develop a perceptual metric that assesses images similarity holistically. They first collect a new dataset of human similarity judgments over image pairs that are alike in diverse ways. During the dataset construction, the authors utilized Stable Diffusion to create synthetic data,... | Rebuttal 1:
Rebuttal: We thank the reviewer for their helpful comments, and address their questions and concerns below.
**Inclusion of human faces**
As the reviewer notes, humans are uniquely sensitive to faces. We believe that face similarity is important but merits its own specialized studies, and a general-purpose... | Summary: The paper expands the fundamental task of measuring perceptual image similarity to encompass factors
beyond low-level and high-level similarity. To achieve this, it leverages the recent advances in generative visual AI and generate synthetic images with text prompts. These synthetic images are expected to con... | Rebuttal 1:
Rebuttal: We thank the reviewer for the helpful suggestions, and are glad that the reviewer finds the analysis extensive and our contribution beneficial to the vision community. We address specific suggestions below.
**Image Quality Assessment**
The reviewer mentions that our model may be useful for vari... | Summary: This paper leverages recent advances in synthetic image generation (namely text-to-image) to explore the space of image similarity assessment. They generate a novel dataset based on human forced choice input on pairs of synthetic images generated with respect to a reference image as well as interleaved visual ... | Rebuttal 1:
Rebuttal: We thank the reviewer for their helpful comments. We address questions and concerns from the reviewer below.
**Reasoning for JND study**
To clarify our purpose for JND, we pair it with 2AFC following prior works like Zhang et al. [1] and Nazeri et al. [2] to verify if our 2AFC annotations genera... | Rebuttal 1:
Rebuttal: We thank all reviewers for their feedback. We are glad they found our research well-motivated [Mq9y, 71tx]. Additionally, we appreciate their acknowledgment of our synthetic dataset and metric as novel [23q7, 71tx, hsYj], and of our experiments and analyses as well-organized and rigorous [23q7, M... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Coupled Reconstruction of Cortical Surfaces by Diffeomorphic Mesh Deformation | Accept (poster) | Summary: This paper is solving a solid practical problem, handling the cortical thickness estimation problem that was historically ignored. Moreover, it improves the quality of surface reconstruction. Authors observed that deforming an over-smoothed template results in lower accuracy if the template is relatively furth... | Rebuttal 1:
Rebuttal: Thanks for your comments and suggestions. We will revise our paper accordingly.
**C1: Rather than integrating all the equations within the text, it will enhance clarity if you present them as numbered equations. Furthermore, providing explanations for each variable immediately after equations wo... | Summary: This paper proposes a method for reconstructing cortical surface (CSR) (including white matter (WM) and gray matter (GM)) from MRI images. The method first estimates the “midthickness” surface, which is the surface between WM and GM, and learn jointly diffeomorphic flows to deform it to WM and GM, respectively... | Rebuttal 1:
Rebuttal: Thanks for your comments and suggestions.
**C1: Novelty\
C1-1: Neural ODE is widely used for deformation in this task (CortexODE, CorticalFlow).**\
**A1-1**: While ODE based formation has been widely used for diffeomorphic (image/surface) deformation, our method is different from CortexODE and Co... | Summary: The paper focuses on the cortical surface-mesh reconstruction in Magnetic Resonance Imaging (MRI) data. The goal is to reconstruct both the inner and outer surfaces of the cortical region. The proposed method begins by estimating an intermediate midthickness layer, which is positioned halfway between the inner... | Rebuttal 1:
Rebuttal: Thanks for your comments and suggestions. We will revise our paper accordingly.
**C1: Surface granularity\
C1-1: The choice of spatial resolution for applying Marching Cubes on the Signed Distance Function (SDF) volume directly impacts resulting mesh granularity. More vertices can reduce the rec... | Summary: This paper presents a deep-learning-based approach for cortical surface reconstruction (CSR) with topology constraints to ensure the physical meaning of reconstructed surfaces. Experiments on public benchmarks show that the proposed approach outperforms current state of the art networks.
Strengths: The overal... | Rebuttal 1:
Rebuttal: **C1: Contributions over-stated, several related works ignored, and potential for reader misunderstanding.\
The statement "the cortical thickness estimation is neglected in all existing DL-based CSR methods" is incorrect. Related methods on cortical surface reconstruction/prediction/parcellation c... | null | NeurIPS_2023_submissions_huggingface | 2,023 | Summary: This paper proposed a method for cortical surface reconstruction by deforming a middle layer inwards and outwards to be the gray matter and white matter surface respectively. First, segmentation network and topology correction are used to extract the initial surface. Then, three velocity fields are predicted t... | Rebuttal 1:
Rebuttal: Thanks for your comments and suggestions.
**C1: This paper seems to overstate its contribution.\
C1-1: First, the initialization pipeline is following/the same as CortexODE (seg + WM surf.).**
**A1-1**: Our objectives and emphases are different from them.
+ First, our method jointly reconstruct... | Summary: This paper develops a new deep-learning-based framework to jointly reconstruct the inner, outer, and their in-between (midthickness) surfaces and estimate cortical thickness. The proposed method has been evaluated on two neuroimaging datasets and shown superior cortical surface reconstruction performance. This... | Rebuttal 1:
Rebuttal: Thanks for your comments and suggestions. We will revise our paper accordingly.
**C1: Lack demographics for the two testing datasets. And experiments can be performed on more populations for evaluation.**
**A1**: We will provide detailed demographic information for all datasets used in our expe... | null | null | null | null |
Discrete-Smoothness in Online Algorithms with Predictions | Accept (poster) | Summary: This paper studies online algorithms with predictions, in which a prediction is provided to the algorithm without any guarantee on its quality. On this topic, the general goal is to provide an algorithm which is (1) consistent (i.e. if the prediction is perfect the algorithm should be much better than a pure o... | Rebuttal 1:
Rebuttal: We thank the reviewer for their comments. Here is our response:
* Regarding the novelty of the combiner, existing combiners do not achieve our smoothness guarantee. The standard online combiner achieves a competitiveness bound of $O(min(ALG_1(Q), ALG_2(Q)))$, which could be arbitrarily worse than... | Summary: This paper studies online facility location and online set cover with predictions, and provides improved theoretical guarantees for these problems. The authors state that these improvements come from a new property that they call discrete-smoothness. Learning-augmented algorithms take as input a prediction on ... | Rebuttal 1:
Rebuttal: We thank the reviewer for their comments. Regarding the more semantic comments, we will be sure to implement the necessary changes. Here are our responses to the remaining questions:
* Regarding the question in the Weaknesses section, the discrete-smoothness guarantee makes sense for many proble... | Summary: Ideal learning-augment online algorithms should have the properties of \emph{consistency} (meaning that they perform comparable to the optimum when the predictions are perfect) and \emph{robustness} (meaning that they don't perform much worse than their classic counterparts even for arbitrary predictions) and ... | Rebuttal 1:
Rebuttal: We thank the reviewer for their comments. Here is our response:
l32: We refer the reviewer to the example in the general rebuttal, which exemplifies what we meant by combinatorial problems, and how this notion captures the (min-cost, bipartite) matching problem. Note that in our paper we consider... | Summary: This paper presents the concept of "discrete smoothness" in the context of online algorithms with predictions. This notion smoothly bridges the gap between consistency and robustness, terms originally introduced by Purohit et al. For a class of problems called "online covering problem", the authors first show ... | Rebuttal 1:
Rebuttal: We thank the reviewer for their comments. Here is our response:
* We are unaware of a general construction that can encapsulate any algorithm for a covering problem and add the subset-competitiveness or Lagrangian properties.
* The term “smoothness” has often been used in online algorithms with ... | Rebuttal 1:
Rebuttal: We would like to thank the reviewers for their thorough reviews and thoughtful comments. We would like to clarify some aspects of the discrete-smoothness notion and its broad applicability.
The structure of many combinatorial problems is the following: one must assign binary values in {0,1} to a s... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Brant: Foundation Model for Intracranial Neural Signal | Accept (poster) | Summary: This paper proposes a foundation model (called BPT) for modeling intracranial brain signals. BPT consists of a Transformer encoder for temporal encoding, and a separate Transformer encoder for spatial encoding. BPT is pre-trained for reconstructing masked input signals, and is fine-tuned for a variety of downs... | Rebuttal 1:
Rebuttal: Thank you for your constructive and detailed comments. Responses to specific comments are listed below.
**1. Q1: Clarification on notations**
Thank you for pointing out this potential ambiguity. Actually, $N_p$ denotes the number of all the patches from an entire channel, while $L$ denotes the n... | Summary: This paper presents Brain Pre-trained Transformer (BPT) for intracranial neural signal modeling. The model was trained on a large-scale dataset and achieved SOTA performance on multiple downstream tasks.
Strengths: - originality: the authors developed a transformer model for brain signal modeling, which can l... | Rebuttal 1:
Rebuttal: Thank you for your careful reading and comments. We are happy that you acknowledged that our model can learn rich representations from intracranial neural recordings. Responses to specific comments are listed below.
**1. Q1: Our planned steps for the dataset release**
Thank you for the valuabl... | Summary: The authors propose a foundation model named BPT for modeling intracranial neural signals. BPT is pre-trained on a large corpus (1.01 TB) of intracranial recordings collected by the authors. Animportant feature is that BPT is designed to capture long-term temporal dependency and spatial correlation from neural... | Rebuttal 1:
Rebuttal: Thank you for your constructive and detailed comments. We are excited that you felt our work has high potential impact and significance for the medical field and perhaps a step towards future neural interfaces. Responses to specific comments are listed below.
**1. Q1: Discussion on why not use o... | Summary: This manuscript trains a foundation model for intracranial signals based upon a very large electrophysiological dataset. The methods are based upon a transformer architecture with adaptations to account for the electrode structure and important frequency properties in electrophysiological systems. The manusc... | Rebuttal 1:
Rebuttal: Thank you for your constructive and detailed comments. We apologize for the imprecise claims in certain parts of our manuscript and misunderstandings caused by our writing. We will reconsider the claims to be more precise. We hope that the responses below could address your specific comments.
**1... | Rebuttal 1:
Rebuttal: # Global Response to AC and all reviewers
Thanks to all reviewers for the careful reading and thoughtful feedback. We are grateful for acknowledging "the model size is of unparalleled precedent" and our work "has high potential impact and significance for the medical field and perhaps a step tow... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
ParaFuzz: An Interpretability-Driven Technique for Detecting Poisoned Samples in NLP | Accept (poster) | Summary: This work proposes a novel framework for defending against poisoned samples. The authors argue that poisoning triggers, like infrequent words, shouldn't drastically change the samples' semantic meaning, and paraphrasing should keep the predictions of clean samples stable but cause poisoned samples to revert to... | Rebuttal 1:
Rebuttal: ## Response to reviewer 4qgi
1. Hinge on ChatGPT
Our technique is model-agnostic and can be generalized to any language models. For example, section 10 in the global response shows that alternative model davinci-003 can also outperform the baselines.
2. Ablation study
See the global response ... | Summary: This paper presents a test-time poisoned sample detection approach in NLP. The underlying intuition is that models' predictions for clean samples should stay the same under paraphrasing, while their predictions for the poisoned samples should change. The core of their approach is to find an appropriate prompt ... | Rebuttal 1:
Rebuttal: ## Response to reviewer 6kxG
1. Assumptions
We adopt the **same** assumptions as the baselines. It’s a standard practice for model users or owners to maintain a small set of clean samples to evaluate model performance. PARAFUZZ is designed to identify poisoned samples, implying that we are alre... | Summary: In this paper, the authors propose PARAFUZZ, a novel framework for test-time detection of poisoned samples in NLP models. Based on that backdoor triggers should not fundamentally alter the semantic meaning of poisoned samples, they adopt a timely large language model, i.e., ChatGPT to paraphrase the poisoned s... | Rebuttal 1:
Rebuttal: ## Response to reviewer 7yjk
1. Alternative LLMs
Please see the global response section 10.
2. Overhead
In experiments, we set the maximum iterations to 300 for each model and the fuzzing process takes 143.88 minutes on average.
3. ChatGPT as classifier
We do not directly apply ChatGPT to c... | Summary: This paper proposes ParaFuzz for discovering if a piece of input text for a poisoned NLP model has backdoor trigger. The idea is is that a paraphrased clean input maintains the same semantic meaning thus the same NLP decision, whereas a poisoned input after paraphrasing will result in a different NLP decision ... | Rebuttal 1:
Rebuttal: ### Response to reviewer oHTu
1. Relation between PICCOLO and PARAFUZZ’s contribution
Please refer to global response section 7.
2. Configuration of PARAFUZZ
The configuration of PARAFUZZ is the same for all NLP models, even with different input lengths.
3. Running time and iterations
**Iter... | Rebuttal 1:
Rebuttal: We sincerely thank all the reviewers for the constructive comments. Please refer to the PDF for details, thanks!
Pdf: /pdf/bc193b39063fd746a66e342fb7b5db832063038d.pdf | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Payoff-based Learning with Matrix Multiplicative Weights in Quantum Games | Accept (poster) | Summary: The paper studies the setting of learning in quantum games, in the case where agents only have access to a scalar (for instance, a payoff value) instead of their full individual gradient, and have to update their strategies in order to converge to some equilibrium. The primary contribution is a variant of the ... | Rebuttal 1:
Rebuttal: Thank you for your encouraging input and strongly positive evaluation. To streamline our response, we reply to your questions and comments in a point-by-point thread below, and we will revise our manuscript accordingly at the next revision opportunity:
1. "*In the experiments, 3MW is compared wit... | Summary: This paper studies online learning of quantum algorithms based on the matrix multiplicative weight (MMW) updates. A main technical contribution is to introduce a weaker version called the minimal-information matrix multiplicative weights (3MW), which only requires zeroth-order gradient sampler and can achieve ... | Rebuttal 1:
Rebuttal: Thank you for your input and positive evaluation. We reply to your questions and comments point-by-point below:
1. "*[The $O(\sqrt{T})$ regret of Jain & Watrous is not the tightest known bound.] A recent paper by Gao et al. gave a quantum algorithm with poly-log regret for zero-sum games.*"
T... | Summary: This paper studies Quantum Games (mainly in the two-player zero-sum setting) and develops payoff-based learning algorithms with provable convergence-rate guarantees. The learning algorithm combines matrix multiplicative weights (MMW) and zeroth order method for gradient estimation. The overall convergence rate... | Rebuttal 1:
Rebuttal: Thank your for your input and positive evaluation. To streamline our response, we reply to your questions and comments in a point-by-point thread below, and we will revise our manuscript accordingly at the next revision opportunity:
1. *Do the authors expect the bound to be tight? Are there any k... | Summary: The authors consider the quantum games in the bandit setting, which players observe only the payoff value of their actions. In order to facilitate this, the authors extend the online bandit estimation methodology into the quantum-state space. As a result, they are able to realize a duality gap of $\mathcal O(T... | Rebuttal 1:
Rebuttal: Thank you for your encouraging input and positive evaluation. To streamline our response, we reply to your questions and comments in a point-by-point thread below, and we will revise our manuscript accordingly at the next revision opportunity:
---
1. "*In classical games or online learning probl... | Rebuttal 1:
Rebuttal: Dear AC, dear reviewers,
We are sincerely grateful for your time, input and positive evaluation. To streamline the discussion phase, we reply to each reviewer's questions in a separate point-by-point thread below. We only include in this global rebuttal a pdf with some additional experiments to i... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Provably Fast Finite Particle Variants of SVGD via Virtual Particle Stochastic Approximation | Accept (spotlight) | Summary: The authors propose a implementable stochastic approximation algorithm based on Stein Variational Gradient Descent, which the idea is based on a stochastic approximation at the probability measure level. The efficiency and convergence analysis of the proposed algorithm, i.e. VP-SVGD and GB-SVGD, are presente... | Rebuttal 1:
Rebuttal: $\newcommand{\vxl}[1]{\vx^{(l)}\_{#1}} \newcommand{\s}[1]{\\{0, \dots, #1\\}} \newcommand{\muFt}{\mu\_t | \cF\_t} \newcommand{\xtt}{\vx^{(t)}\_t} \newcommand{\XS}[1]{\vx^{(0)}\_0, \dots, \vx^{(#1)}\_0}$Thanks for your helpful feedback! The reviewer mentions that some of our proofs are not rigorous... | Summary: This study analyzed the finite particle behavior of Stein Variational Gradient Descent (SVGD) in an asymptotic manner. For that purpose, the authors designed two computationally efficient variants of SVGD, namely VP-SVGD and GB-SVGD, and prove their convergence rates. In those methods, the authors introduced a... | Rebuttal 1:
Rebuttal: $\newcommand{\e}{\epsilon} \newcommand{\gd}{\nabla} \newcommand{\KSD}[3]{KSD\_{#1}(#2||#3)}$ Thank you for your constructive feedback. We hope to address your questions in our response below:
## Notation in Sec. 4
Thanks for the suggestion! This is incorporated in our updated draft where we consis... | Summary: The paper studies Stein Variational Gradient Descent (SVGD) which is a popular variational inference algorithm that simulates an interacting particle system to approximately sample from a target distribution. The authors propose two variants of SVGD, Virtual Particle SVGD (VP-SVGD) and Global Batch SVGD (GB-SV... | Rebuttal 1:
Rebuttal: $ \newcommand{\iidsim}{{\overset{\mathrm{iid}}{\sim}}} \newcommand{\cH}{\mathcal{H}} \newcommand{\hmu}{\hat{\mu}^{(n)}}$
Thank you for your thoughtful review and suggestions, which have significantly helped us improve our work. We have updated our draft to include a detailed discussion of the lim... | Summary: This paper propose two variants of stein variational gradient descent (SVD) named as Virtual Particle SVGD (VP-SVGD) and Global Batch SGVD (GB-SVGD) with provably fast finite-particle convergence rates to a target distribution. Introducing the notion of virtual particles, the authors show that VP-SVGD is an ex... | Rebuttal 1:
Rebuttal: $\newcommand{\cF}{\mathcal{F}} \newcommand{\vx}{\mathbf{x}} \newcommand{\mub}{\bar{\mu}} \newcommand{\KL}[2]{KL\left(#1||#2\right)} \newcommand{\bE}{\mathbb{E}} \newcommand{\KSDs}[3]{KSD^{2}\_{#1}(#2||#3)} \newcommand{\iidsim}{{\overset{\mathrm{iid}}{\sim}}}$ Thank you for your kind and thoughtful... | Rebuttal 1:
Rebuttal: $\newcommand{\cF}{\mathcal{F}} \newcommand{\vx}{\mathbf{x}} \newcommand{\mub}{\bar{\mu}} \newcommand{\KL}[2]{KL\left(#1||#2\right)} \newcommand{\bE}{\mathbb{E}} \newcommand{\KSDs}[3]{KSD^{2}\_{#1}(#2||#3)}$
We thank the reviewers for their thoughtful comments & feedback that have significantly he... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Brain-like Flexible Visual Inference by Harnessing Feedback Feedforward Alignment | Accept (poster) | Summary: The authors present a novel artificial neural network architecture which they relate to visual processing in the brain. In particular, the authors avoid the weight transport problem by proposing that feedback weights can be learned by optimising them to reconstruct the original input (autoencoder). The authors... | Rebuttal 1:
Rebuttal: We appreciated the detailed evaluation of the reviewer and their comments to improve the manuscript.
> "Weaknesses:”
We now added a pseudocode to the paper (see general response).
> "The authors ... “
In the general response, we outline the theoretical motivation behind the work and the limit... | Summary: This article proposes and exploits a bio-plausible mechanism inspired from the Feedback Alignement : the Feedfback-Feedforward Alignement (FFA). Through a series of experiments (classification, robustness, denoising, generation…) the authors demonstrate that the FFA could be used for various applications.
St... | Rebuttal 1:
Rebuttal: We would like to thank the reviewer for their constructive comments and suggestions to improve the manuscript.
> "Weaknesses ...”
We acknowledge that a model of the brain ideally should encompass the structural constraints of the neural circuits (bio-plausible) and exhibit hallmarks of human co... | Summary: The paper presents a new neural network architecture called Feedback Alignment (FFA) that demonstrates the effectiveness of co-optimizing classification and reconstruction tasks on MNIST and CIFAR10 datasets. FFA uses a goal-driven feedback pathway to alleviate weight transport problems encountered in traditio... | Rebuttal 1:
Rebuttal: > "Strengths:
S1. This paper proposes a novel strategy to train neural networks to co-optimize for two objective functions based on the role of feedback connections in the brain, and leverages the credit assignment computational graph as feedback connections during learning and inference S2. FFA f... | Summary: This paper explores the intriguing subject of integrating feedback connections into neural networks during both learning and inference stages, drawing inspiration from the brain. This study has the potential to significantly advance artificial intelligence (AI). Feedback connections are a topic of interest in ... | Rebuttal 1:
Rebuttal: > "Summary:
This paper explores the intriguing subject of integrating feedback connections into neural networks....
.”
We would like to thank the reviewer for recognizing the significance of the aim of this work.
We agree that the algorithm was not clarified in detail and we apologize for this o... | Rebuttal 1:
Rebuttal: We thank the reviewers for their helpful comments to improve the work. Below, we address the general theme of concerns which were shared between reviewers.
**FFA algorithm in pseudo code**
Weight updates in a three-layer linear network:
parameters: $W_{f_1}$, $W_{f_2}$, $W_{b_1}$, $W_{b_2}$ ... | NeurIPS_2023_submissions_huggingface | 2,023 | Summary: The paper proposes a bio-inspired computer vision algorithm for classification that includes feedback and aims to use alignment rather than backprop for training the models of feedforward and feedback.
Strengths: The paper provides a novel idea for including feedback (or recurrence) in visual classification t... | Rebuttal 1:
Rebuttal: > " Strengths:
The paper provides a novel idea for including feedback (or recurrence) in visual classification that would be interesting for neuroscientists to investigate further. The emerging noise robustness is a big strength for this model and confirms a lot of neuro theories about feedback.”
... | Summary: The authors propose a mechanism for biologically plausible learning in visual circuits that can endow the system with additional epiphenomenal capabilities that resemble phenomena humans are known to exhibit. The mechanism resembles feedback alignment, where the weight transport problem is resolved by substitu... | Rebuttal 1:
Rebuttal: > "
Strengths:
The concept for the paper is a very strong one: it is a major failing of the 'feedforward paradigm' of visual cortex modeling that it cannot account for capabilities and phenomena the visual system is known to exhibit ..."
We appreciate that the reviewer recognizes the significance... | null | null | null | null |
Optimal Transport for Treatment Effect Estimation | Accept (poster) | Summary: In this paper, authors have applied optimal transport to address selection-bias issue in binary treatment setting for individualized treatment effect estimation. They have proposed a relaxed mass-preserving regularizer to address the mis-alignment generated from outcome imbalance and outliers in non-ideal mini... | Rebuttal 1:
Rebuttal: **[W1] Comparison with recent techniques**
- Thanks for your kind recognition.
We agree that comparison with recent techniques could improve this work. Thus, we incorporated two more recently baselines founded on representation learning, DRCFR [1] and DRCFR-MIM [2], in appendix, with the aim of mo... | Summary: The authors propose to minimize a (relaxed) Sinkhorn distance between representations under the two treatment groups in order to balance the latent space and better predict the treatment effects. They highlight two problems with existing TARNet-based approaches that do not regularize with optimal-transport cos... | Rebuttal 1:
Rebuttal: **[W1] About the hidden-confounding regularizer**
- **Formal connection.** Please refer to our response to the Reviewer VwsR, section "Theoretical Guarantees for the UCE Regularizer".
- **Substituting Potential Outcomes with Model Predictions.** Pseudo label (using model predictions to substitue g... | Summary: The objective of this research paper is to tackle two primary concerns: (1) the effects of mini-batch sampling (MSE), which lead to misalignment in non-ideal mini-batches characterized by outcome imbalance and outliers, and (2) the impacts of unobserved confounders (UCE), resulting in inaccurate discrepancy ca... | Rebuttal 1:
Rebuttal: **[W1] The main contribution of the proposed relaxed mass-preserving regularizer may be seen as an incremental addition to the field of causal inference, raising concerns about the originality and novelty of the paper.**
- Thank you for your sincere comments. We acknowledge your concern regarding... | Summary: The paper introduces a new approach called Entire Space CounterFactual Regression (ESCFR) to estimate individual treatment effects from observational data, addressing two critical problems that existing methods fail to solve. The first problem is mini-batch sampling effects (MSE), which lead to misalignment in... | Rebuttal 1:
Rebuttal:
We extend our sincere gratitude for your insightful comments and constructive critique. We acknowledge your point concerning the theoretical support of the UCE regularizer. We are pleased to elaborate our thoughts and methodologies to clarify the theoretical foundation.
**Is there some theoretic... | null | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
A Smooth Binary Mechanism for Efficient Private Continual Observation | Accept (poster) | Summary: The binary mechanism is a simple, efficient and well-known method to answer linear prefix queries (like counting) when a database of size T is not given in advanced, but rather in each time point $t \in [T]$ we are given a new row $x_t$, and we are required to update our result based on the new row (while taki... | Rebuttal 1:
Rebuttal: We thank the reviewer for their reading and agree that the algorithm has the potential of being useful in practice.
*W1: The improvement is clever and simple, but perhaps too simple for a NeurIPS submission.*
While we agree that our algorithm is simple, we would argue that this is a strength of ... | Summary: This paper studies the continual counting with differential privacy problem. Here there are $T$ time steps; at time step $t \\in [T]$, we receive a $x_t \\in \\{0, 1\\}$ and we have to output an estimate to the sum so far $x_1 + \\cdots + x_t$. We would like to entire stream of output to be differentially priv... | Rebuttal 1:
Rebuttal: We thank the reviewer for their careful reading and insightful questions.
*W1: [...] a tradeoff between time-space efficiency and privacy-utility tradeoff. In my opinion, the latter is typically more important in practice.*
While often true, time-space efficiency is the reason why papers based o... | Summary: This paper studies the important problem of optimizing continual observation algorithms while also making them efficient. There has been some recent advances in the classical binary mechanism for releasing a streaming counter subject to differential privacy. One of the main issues with the binary mechanism i... | Rebuttal 1:
Rebuttal: We thank the reviewer for their reading and positive comments.
*Q1: Can you give a better presentation for the algorithm and how it works?*
We will add further details about how the algorithm works, including a graphical illustration of how the algorithm functions reminiscent of Figure 1(b) for ... | Summary: This paper studies the problem of releasing private prefix sums in the continual release model of differential privacy. A continual release mechanism for prefix sums receives a binary stream $x_1, x_2, x_3, ...$ one element at a time and continually outputs, at each time step, the number of ones received so fa... | Rebuttal 1:
Rebuttal: We thank the reviewer for their reading and questions.
*W1: [...] formal portions of the paper are (in several instances) not well specified.*
We disagree with the claim that the formal portions are not well-specified, see answer to Q1 below.
*W2: [...] analysis of the mechanism seems incomplet... | Rebuttal 1:
Rebuttal: In response to reviewer kh8m's request for a better explanation of how the final algorithm, Algorithm 2, works, see the attached PDF below containing two graphical examples of how the algorithm computes $\mathcal{M}(2)$ and $\mathcal{M}(5)$ when $T=5$.
We commit to including at least one of the e... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
MGDD: A Meta Generator for Fast Dataset Distillation | Accept (spotlight) | Summary: In this paper, the authors address the efficiency concerns associated with dataset distillation. Initially, the authors identify the challenges arising from the dependencies within the intensive loop of forward-backward propagation across numerous neural networks, as well as the repetitive training procedures ... | Rebuttal 1:
Rebuttal: We sincerely thank Reviewer Bfcg for the positive feedback on the manuscript and are very excited that the reviewer mentioned the strengths of effectiveness, theoretical analysis, algorithmic workflows, writing, and experiments. The questions are addressed below.
* > What is the physical adaptio... | Summary: The reviewed paper makes a clear and novel contribution to the field of dataset distillation, which is expected to have a beneficial impact. The paper is well written, demonstrating a structured and easily understandable presentation. Notably, the empirical evaluation is exceptional, covering various aspects o... | Rebuttal 1:
Rebuttal: We are more than grateful to Reviewer CHTg for the positive feedback, valuable comments, and frank discussions regarding the manuscript. We are very delighted that the reviewer found our contribution novel, paper well written, and evaluation exceptional. For the concerns and questions, we would li... | Summary: - This paper proposing a meta-learning version of dataset distillation (DD)
- The main idea is that dataset distillation is slow, and meta-learning can be used to amortize the dataset distillation procedure
- The goal a function $g_\omega (X_{init}) = X_S$, which changes an initialized distilled dataset to the... | Rebuttal 1:
Rebuttal: We would like to express our sincere gratitude for the pertinent comments and suggestions of Reviewer pHTr. We are glad that the reviewer finds our method novel and valuable, improvements in efficiency dramatic, and cross-IPC generalization impressive. The concerns, which withhold the reviewer's r... | Summary: This paper proposes a meta-learning conditional generator method that takes some initialization of synthetic images and quickly learns to render it for Dataset Distillation purposes. The authors extensively verified the performance and show that the proposed method is substantially faster and achieves good per... | Rebuttal 1:
Rebuttal: We sincerely appreciate Reviewer jcg2 for the constructive comments. We are happy that the reviewer finds our work well-motivated, the method interesting, and the experiments extensive. We would like to address the concerns and questions reflected in the review below.
* > Why do the updated image... | Rebuttal 1:
Rebuttal: We would like to express our heartfelt gratitude to all reviewers for their constructive reviews. The questions and concerns are addressed in the rebuttal under the Official Review block of each reviewer. We would like to upload a pdf file here, denoted as "rebuttal pdf" in the following parts. Th... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
A One-Size-Fits-All Approach to Improving Randomness in Paper Assignment | Accept (spotlight) | Summary: This paper focuses on randomized paper assignments in the reviewer-paper matching problem. First, the authors state the motivation and emphasize the importance of randomness in paper assignments. Then, they point out the weakness of the existing method, PLRA, and propose additional metrics for evaluating the p... | Rebuttal 1:
Rebuttal: Thank you for spending time understanding our work and providing insightful comments. We are happy to see that you find our work well-written, practical, and universal. Below we respond to the questions and weaknesses raised in your review, hoping to address your concerns about our paper.
**Regar... | Summary: In this paper, the authors propose a new algorithm to improve randomness in peer-review paper assignment without sacrificing assignment quality. The authors first identify a set of good randomness properties in paper assignment besides the max probability used in the PLRA. The authors propose a simple and prac... | Rebuttal 1:
Rebuttal: Thank you for spending time understanding our work and providing insightful comments. We are very glad to see your positive review and you recognize our work as superior in performance and very practical. Below we briefly respond to the weaknesses raised in your review.
**Regarding comparison to ... | Summary: The paper presents an improvement on PLRA, an important (used in major reviewing deployments like AAAI and OpenReview) algorithm for randomizing reviewer assignments while preserving a prescribed fraction of assignment quality, by finding randomized assignments that increase additional randomization metrics of... | Rebuttal 1:
Rebuttal: Thank you for spending time understanding our work and providing insightful comments. We are very glad to see your positive review and that you recognize our work as significant, substantial, and ready to deploy. We respond to the questions and limitations raised in your review below.
**Regarding... | Summary: This work is concerned with randomness in paper matching for peer review. It provides normative motivations for the value of randomness in this process and proposes a family of generalizations of the Probability Limited Randomized Assignment (PLRA) paper matching framework of Jecmen et al. This generalization,... | Rebuttal 1:
Rebuttal: Thank you for spending time understanding our work and providing insightful comments. We are happy you find our contributions relevant, interesting, and practical. Below we respond to the questions and limitations raised in your review, hoping to address your concerns about our paper.
**Regarding... | Rebuttal 1:
Rebuttal: We firstly want to thank all the reviewers for spending the time to understand our work and provide insightful comments.
We are happy to see that all the reviewers concurred that our proposed method “has better performance” (reviewer 625T) in randomness with “little to no apparent downside (iL2x... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Fairness-guided Few-shot Prompting for Large Language Models | Accept (poster) | Summary: The paper addresses the problem of constructing an appropriate prompt for improving the performance of in-context learning in large language models (LLMs). It introduces a metric based on predictive bias to evaluate prompt quality and empirically demonstrates that prompts with higher bias lead to unsatisfactor... | Rebuttal 1:
Rebuttal: **Thank you for carefully reviewing our work and valuable suggestions. We believe the following point-to-point response can address all the concerns:**
---
***Q1: The complexity of the search strategies.***
(1) We acknowledge the concern regarding the computational cost of G-fair Prompting as N ... | Summary: The paper examines the sensitivity of large language models to prompts and analyzes the issue from a predictive bias standpoint. They utilize a fairness metric called a content-free strategy to evaluate the predictive bias of a fixed prompt. Methodologically, the paper proposes two strategies, T-fair-Prompting... | Rebuttal 1:
Rebuttal: **Thank you for your time and valuable suggestions. Next, we will answer the questions in detail point by point to address any concerns.**
---
***Q1: Contribution and difference from [1].***
Our paper shares a similar metric with calibration-before-use [1] to access the predictive bias of a give... | Summary: The paper proposes two new procedures to select examples to be used for few-shot learning (in-context learning) for LLMs. The key idea for example selection is to select those examples that lead to a distribution over classes that is as close to uniform as possible . The two proposed variants (a) select the to... | Rebuttal 1:
Rebuttal: **Thank you for taking the time to provide valuable suggestions and have a positive attitude towards our work.**
---
***Q1: The proposed method is limited to multi-class classification tasks.***
(1) We acknowledge the limitation that the proposed method is currently applicable only to classific... | Summary: The paper shows that the accuracy of large language models on various text classification tasks for a given prompt tends to be correlated with the entropy of the output prediction when using that prompt for a context free input: in general, the more uniform the output probability distribution for a given promp... | Rebuttal 1:
Rebuttal: ***Q1: The time complexity of G-fair-prompting.***
(1) We acknowledge the concern regarding the computational cost of G-fair Prompting as N increases. However, we would like to emphasize that the worst-case computation cost of 1/2*N^2 is not a common scenario. In practice, only a few demonstratio... | null | NeurIPS_2023_submissions_huggingface | 2,023 | Summary: Finding an (nearly) optimal prompt is crucial for successful in-context learning (ICL) of large language models (LLMs). It is known that performance of downstream tasks is sensitive to prompts themselves, few-shot examples, and their orders, often requiring large degree of non-trivial tuning. This paper presen... | Rebuttal 1:
Rebuttal: **Thank you for the reviewers' valuable feedback. We have carefully reviewed the comments and offer the following responses to address each point:**
----
***Q1: Computational cost of using G-fair Prompting becomes non-trivial as N goes larger.***
(1) We acknowledge the concern regarding the co... | null | null | null | null | null | null |
On Transfer of Adversarial Robustness from Pretraining to Downstream Tasks | Accept (poster) | Summary: This paper studies the question of how the robustness of pretrained representations transfers to downstream tasks. Under the linear probing setting (i.e., the downstream classifier $f$ is linearly related to the pretrained representations $g_{\theta}$), they prove that the gap between the standard and robust r... | Rebuttal 1:
Rebuttal: We want to thank you for your feedback and in the following we address the weaknesses and questions raised in your review.
**Weaknesses:**
1. We are aware of results on robust risk bounds for linear classifiers or VC-class functions, which, roughly speaking, provide insights into the learnability... | Summary: This paper study how the robustness properties of a given representation might be transferred to a linear probe which was trained on this representation. Since the linear probe is a simple multiplication between the weights parameters and the representation vector, the authors were able to demonstrate that by... | Rebuttal 1:
Rebuttal: Thank you for your thorough feedback. In the following, we would like to comment on the weaknesses and questions raised in your review.
**Weaknesses:**
1. Thank you for highlighting this topic. We agree that the impact of regularizing the weight matrix of the linear probe on downstream robustness... | Summary: The paper studies the robust transfer of a pre-trained model, which demonstrate that the robustness of a linear predictor on downstream tasks can be constrained by the robustness of its underlying representation, and theoretically prove a bound on the difference between clean and adversarial loss in terms of t... | Rebuttal 1:
Rebuttal: Thank you for your review. In the following, we address all your questions and concerns.
**Questions:**
1. In Figure 1, we distinguish between robust and non-robust representations based on their response to adversarial perturbations. We consider two separate representation functions in green and... | Summary: The paper considers the problem of robustness transfer from pretraining to downstream finetuning. The authors focus on the setup of linear probing, where downstream tasks are learned with a linear model on top of frozen pretrained features. They however give some insights on how to go beyond linear probing and... | Rebuttal 1:
Rebuttal: Thanks for your review and feedback. We would like to comment on the weaknesses and questions raised in your review.
**Weaknesses:**
1. While we agree that empirical validation in additional modalities and with large-scale models would be valuable, we have limited computational resources. With th... | Rebuttal 1:
Rebuttal: We want to thank the reviewers for their thorough reviews and meaningful suggestions which will help us improve our paper.
We are also happy to hear that our article is considered “*well written and easy to follow*” (R-*2ud3*), that our contribution is “*significant and important to the community*... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Optimal Transport Model Distributional Robustness | Accept (poster) | Summary: This paper explores the relationship between OT-based distributional robustness and sharpness-aware minimization (SAM), and shows that SAM is a special case of the proposed framework when a Dirac delta distribution is used over a single model. To demonstrate the effectiveness of the framework, the authors perf... | Rebuttal 1:
Rebuttal: We sincerely appreciate your constructive comments. We are dedicated to addressing all the questions listed below to the best of our capabilities.
**Theorem 4.2 holds for $p=+\infty$, can the results be extended to more general settings?**
Our distance of interest is $d(\theta,\tilde{\theta})=\... | Summary: This paper proposes an optimal-transport-based distribution robustness framework on model space. The proposed theory considers the worst-case loss w.r.t a Wasserstein distance from the model distribution, which can be considered a probabilistic extension of sharpness-aware minimization (SAM).
Strengths: - The... | Rebuttal 1:
Rebuttal: We sincerely appreciate your constructive comments. We are dedicated to addressing all the questions listed below to the best of our capabilities.
**Do not conduct generalization analysis, which is important in learning theory**
Thanks for this comment. Generalization bound is an interesting the... | Summary: This paper explores the optimal transport distributional robustness framework in the model parameters spaces, and provides practical algorithms to solve the corresponding optimization problems. This flexible framework can be used to train a single model, ensemble models and Bayesian neural networks in a sharpn... | Rebuttal 1:
Rebuttal: We sincerely appreciate your constructive comments. We are dedicated to addressing all the questions listed below to the best of our capabilities.
**Q1. In SAM [1], their experiments show that using 2-Wasserstein distance is better. But based on theorem 4.2, the algorithm solves the minimization... | Summary: This paper proposes a novel distributional robustness optimization problem on the space of model parameters (OT-MDR). By formulating the uncertainty set with Wasserstein distance and introducing an entopy regularization, the explicit solution to the primal problem is derived. A practical algorithm is proposed ... | Rebuttal 1:
Rebuttal: We sincerely appreciate your constructive comments. We are dedicated to addressing all the questions listed below to the best of our capabilities.
**Advantage of OT-MDR in the case of a single model and ensembling**
In Theorem 4.3, we show a connection of our OT-MDR developed by applying distri... | Rebuttal 1:
Rebuttal: We appreciate the reviewers' valuable comments and would like to report additional experiments in the attached pdf.
Pdf: /pdf/332cb08fbabba5ba82145a7845e5e3197d8dce9d.pdf | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Flow Matching for Scalable Simulation-Based Inference | Accept (poster) | Summary: The paper introduces a new solution for Simulation Based Inference (SBI) by proposing Flow Matching Posterior Estimation (FMPE) as an alternative to Neural Posterior Estimation (NPE). The authors leverage the power of Flow Matching (FM) generative models as amortizors to enhance the accuracy and efficiency of ... | Rebuttal 1:
Rebuttal: We thank the reviewer for the positive report and helpful comments.
### Weaknesses
* *[Cost of likelihood evaluations.]* We agree that flow matching likelihood evaluations are expensive due to the many backward network passes required (taking ~6 minutes for the gravitational waves example). Howe... | Summary: This paper proposes solving Bayesian inference problems where we would like to estimate & sample from p(theta | x) when p(theta) and p(x | theta) are known. This is done by randomly sampling from the joint distribution and training conditional generative models as amortized approximate posteriors.
Unlike ELBO... | Rebuttal 1:
Rebuttal: We thank the reviewer for the positive review and helpful comments. We are happy that they liked our Theorem 1!
### Weaknesses
* *[Benchmarks not challenging.]* As we argue in the response to all reviewers, the gravitational wave (GW) inference problem (Sec. 5) is in fact extremely challenging. ... | Summary: The paper proposes a novel method for posterior estimation when samples of likelihood are given. Specifically, the authors propose the use of flow matching (FM) to train continuous normalizing flow-based (CNF-based) posterior distributions, a method they term "flow matching posterior estimation" (FMPE). The au... | Rebuttal 1:
Rebuttal:
We thank the reviewer for the careful reading of our manuscript and positive comments on our writing and perspective. We fully agree with the reviewer that the main challenge was not purely technical; rather, **we combine existing methods in a new way, and show that this can solve a very hard tas... | Summary: This a **very well written** paper presenting an extension to the flow matching approach for generative modelling to the simulation-based inference setting. It goes in line with preceding works that adapted normalizing flows to SBI as well.
The core idea of the paper is described in Equation (7): one can app... | Rebuttal 1:
Rebuttal: We thank the reviewer for the thorough report and are pleased to hear that they enjoyed reading the paper.
### Weaknesses
* *[“Results on real data seem rather preliminary”, "Why [...] use FMPE for it then?"]* Indeed, results are not comprehensive from a physics perspective, and we plan to carry... | Rebuttal 1:
Rebuttal: We thank the reviewers for the positive evaluation of our work and for providing detailed comments and suggestions for improvement. In particular, all reviewers recommend acceptance to varying degrees, with the paper “very well written” (1iUK), “easy to understand” (7UFh), “offering meaningful con... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Temporal Dynamic Quantization for Diffusion Models | Accept (poster) | Summary: This paper presents a novel quantization method for diffusion models that dynamically adjusts the quantization interval. To this end, the authors present a unique Temporal Dynamic Quantization (TDQ) module, which utilizes the time step as an input to determine the quantization interval, which significantly imp... | Rebuttal 1:
Rebuttal: Thank you for your positive comment and insightful feedback.
---
### Weaknesses
**W1: Comparsion with Q-Diffusion**
We have thoroughly prepared a response to address your concern. Kindly refer to the global response section 2 where we detail the comparison with Q-Diffusion.
### Questions:
**... | Summary: While diffusion models provide outstanding quality and diversity of generated data, they demand high storage and computation, limiting its applications on mobile devices. Further, the paper points out that current existing works on quantization are not suitable for diffusion models as diffusion models have uni... | Rebuttal 1:
Rebuttal:
Thank you for your positive comment and constructive feedbacks
---
### Weaknesses
**W1: Extra overhead of input-dependent quantization method**
We appreciate the feedback and understand the concerns raised. We would like to direct the reviewer's attention to the recently published paper, "Mob... | Summary: This paper proposes Temporal Dynamic Quantization (TDQ), a new quantization approach for diffusion models. This method utilizes a learned module to adjust the quantization interval with respect to timestep to deal with activation quantization errors. Unlike conventional dynamic quantization techniques, the TDQ... | Rebuttal 1:
Rebuttal:
Thank you for bringing up important points with productive feedback.
---
### Weaknesses
**W1: Quantization Configuration Details for Different Layers**
Thank you for raising the pivotal aspect of activation quantization for attention multiplicands. To clarify, our implementation does indeed a... | Summary: The authors present and empirically evaluate a novel quantization scheme for diffusion models. Static quantization of the evolving activation distribution in diffusion models leads to performance degradation. By using a quantization range which is a function of the timestep, the authors improve performance.
S... | Rebuttal 1:
Rebuttal: Thank you for your valuable feedback and constructive reviews.
---
### Weakness
**W1: TDQ Module Specifications**
In this manuscript, we present the TDQ module, a carefully designed 4-layer MLP with ReLU activation. The module takes 128 input elements and has a hidden layer with 64 elements. ... | Rebuttal 1:
Rebuttal: # Global response for reviewers
We appreciate all the reviewers for taking time to review our work and giving valuable comments and feedbacks that help to improve the completion of our paper.
---
## 1. Learning quantization interval directly for each time step
We highlighted that as the activat... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Selectivity Drives Productivity: Efficient Dataset Pruning for Enhanced Transfer Learning | Accept (poster) | Summary: The paper proposes two new dataset pruning (DP) methods, label mapping and feature mapping, for transfer learning. The aim is to improve pretraining efficiency and fine-tuning accuracy on downstream target tasks without sacrificing performance. The proposed methods are evaluated on numerous transfer learning t... | Rebuttal 1:
Rebuttal: Thank you for your constructive feedback. We provide response to **the eight points you listed in the weakness column (W1 to W8) and the limitations.** Additional results are indexed with "R" (e.g., Fig. R1) to distinguish them from the submission (e.g., Fig. 4), and are detailed in the one-page P... | Summary: This paper addresses the problem of dataset pruning, which serves to smaller the dataset for faster computation and learning. The authors claim that some existing methods require overly much computation for selecting the pruned dataset and thus propose a method with less computational requirements. However, in... | Rebuttal 1:
Rebuttal: We thank the reviewer for the constructive comments. **We use [W] and [Q] to refer to the specific points listed in the “Weakness” and “Question” columns in your official review respectively.** All the additional results use the index starting with “R” (e.g., Tab. R2, Fig. R1) to differentiate fro... | Summary: This paper presents the problem of dataset pruning for transfer learning. The idea is to remove redundant samples in the source dataset to improve the source pre-training efficiency while maintaining the downstream fine-tuning accuracy. The label and feature mapping method proposed for the problem utilizes a s... | Rebuttal 1:
Rebuttal: We thank the reviewer for the detailed feedback. We present a point-to-point response below, where **we use [W] and [Q] to refer to the specific points listed in the “Weakness” and “Question” columns in your official review.** All the additional results use the index starting with “R” (e.g., Tab. ... | Summary: The following paper tackles the problem of dataset pruning from the lens of transfer learning, where the main goal is to achieve lossless or improved transfer learning accuracy of the source model on a target task using less amount of source data. To achieve this, this paper proposes two dataset pruning algori... | Rebuttal 1:
Rebuttal: We sincerely appreciate your constructive comments and a precise understanding of our contributions. **We use [W] and [Q] to refer to the specific points listed in the “Weakness” and “Question” columns in your official review correspondingly.** Please see our response below.
**Q1: [Q] [W] More ex... | Rebuttal 1:
Rebuttal: ## Highlighted General Response
We thank all the reviewers for their valuable comments. First, we would like to clarify some possible misunderstandings about our work based on the comments.
### GR1 - Possible misunderstandings of the motivation and the setting of this work. (Reviewer BXTC, vmFL,... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Optimal cross-learning for contextual bandits with unknown context distributions | Accept (poster) | Summary: The paper studies the cross-learning problem in contextual bandits, and proposes a $\sqrt{KT}$ regret algorithm.
Strengths: The paper proposes an algorithm achieving $\sqrt{KT}$ regret using an epoch-based schedule. The algorithm achieves optimal regret without requiring knowledge of the context distribution.... | Rebuttal 1:
Rebuttal: ``In applications (section 1.1 and section 4), it might be better to point out the explicit connection between cross-learning and bidding in 1st price auction / sleeping bandits. E.g. what is the context here in 1st price auction?''
In 1st price auction, the context is the personal value $v_t$ fo... | Summary: This paper studies contextual bandits in the ``cross-learning' setting. They propose an algorithm that attains nearly tight regret of $\tilde{\mathcal{O}}(\sqrt{TK})$. They achieves this by carefully separating the horizon into blocks of fixed size and coordinate the estimation of unknown distribution and the ... | Rebuttal 1:
Rebuttal: Re: dropping the cross-learning condition, this paper is fundamentally focused on designing algorithms under this cross-learning assumption, so our algorithms only make sense in this setting. Without a cross-learning assumption, the regret bounds can be far worse (in particular, we must pay a fact... | Summary: This manuscript studies the following missing piece in [Balserio et al. 2019]: in contextual bandits with cross learning (i.e. with a complete feedback graph across contexts), suppose that the contexts are stochastic with an unknown distribution, is there an algorithm that achieves a \tilde{\Theta}(\sqrt{KT}) ... | Rebuttal 1:
Rebuttal: Thank you for the suggestions on improving the writing.
---
Rebuttal Comment 1.1:
Title: AC response to rebuttal
Comment: Hello authors,
Could you please give some details about how exactly you are planning to improve the write-up according to the reviewer's complaints?
Thank you,
the AC
--... | Summary: This paper studies the sequential learning problem with the presence if context, where the context is drawn iid from some unknown distribution and the losses are chosen by the adversary. The feedback, that learner observes, is the loss for any context for the chosen action. This paper improves the previous re... | Rebuttal 1:
Rebuttal: $p_{eL,c,k}$ is $p_{t,c,k}$ at time $t=eL$, i.e. the last timestep of epoch e. $s_{e,c,k}$ is defined after line 193, it determines the rate at which we use observations in epoch e via downsampling.
The loss estimate for $t_f$ is simply 0. We scale the loss estimate for $t_{\ell}$ by a factor two... | Rebuttal 1:
Rebuttal: Several reviewers questioned the motivation of the cross-context learning. This setting is of interest because it generalizes several distinct problems in the literature. Besides bidding in first-price auctions and sleeping bandits, this also includes dynamic pricing with variable costs, where eac... | NeurIPS_2023_submissions_huggingface | 2,023 | Summary: The paper proposes an efficient contextual bandit algorithm in the cross-learning setting, introduced in Balseiro et al., in the case of contexts sampled from an unknown distribution and where losses are chosen adversarially. In contextual bandits, the learner only utilizes the reward associated with an action... | Rebuttal 1:
Rebuttal: Thank you for spotting the typos, the correction in L.152/159 is correct.
We will add a comparison for the computational efficiency. The $T^{2/3}$ baseline of Balseiro et al. has the same computational complexity.
---
Rebuttal Comment 1.1:
Title: AC response to rebuttal
Comment: Hello authors,
... | null | null | null | null | null | null |
Coneheads: Hierarchy Aware Attention | Accept (poster) | Summary: This paper introduces a simple drop-and-replace operation for dot-product attention, based on cones. The new operation is designed to model hierarchical information better than standard attention. The proposed method outperforms standard attention on various tasks. I will say upfront that I am not very familia... | Rebuttal 1:
Rebuttal: - Hierarchical nature of model.
- Please see the general response. Beyond that, discovering hierarchical structures in an unsupervised or semi-supervised manner is an active research area, and it remains to be seen if cone attention can be applied there. Using cone attention for structure disc... | Summary: The paper explores a variant of hyperbolic attention to induce a hierarchical inductive bias by exploiting the corollaries of points in a hyperbolic space and nodes in a tree. The paper integrates shadow cone construction for entailment cones to define partial order relations between points (the partial order ... | Rebuttal 1:
Rebuttal: - “The limitation says that the model can be sensitive to initialization. While I am not going to take that itself as a negative point but this fact may put some concern given that the models are only trained once per data. However, since the improvement is quite consistent across multiple tasks a... | Summary: This paper introduces hyperbolic shadow cone similarity and induced Hierarchy Aware Attention. The main contribution lies in their proposal to utilize hyperbolic shadow cones (specifically, umbral cone and penumbral cone) as an entailment setting to induce pairwise similarity scores denoted as $K(\cdot,\cdot)$... | Rebuttal 1:
Rebuttal: - Why is the model hierarchical?
- We apologize if our use of “hierarchical” caused confusion and are open to using other terms in the final paper. Please see the general response for more information about hierarchies and the model.
- The use of "Attention" in the title may be a bit over-clai... | Summary: This paper proposes replacing the traditional softmax attention kernel with one that tries to captures the hierarchical relationship between two vectors by mapping them to hyperbolic space.
Strengths: The idea of using hyperbolic attention cones to define hierarchy-aware attention is (to my knowledge) novel ... | Rebuttal 1:
Rebuttal: - Hierarchical behavior of trained models
- Please see the general response.
- Comparison with hyperbolic distance attention, hyperbolic geometry, and entailment cones
- We suspect that most of the performance difference between the two methods comes from our use of entailment cones vs. Gu... | Rebuttal 1:
Rebuttal: Dear reviewers, thank you for your detailed and thorough reviews. As reviewers Rx82 and AaRb noted, our use of entailment cones to define an attention mapping is “novel” and “innovative.” All reviewers noted that throughout our “extensive” suite of experiments across graph prediction, NLP, and vis... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Exploring and Interacting with the Set of Good Sparse Generalized Additive Models | Accept (poster) | Summary: In contrast to the conventional machine learning paradigm, which typically yields a single model, Rashomon sets comprise a collection of near-optimal models. These sets allow users to select a model based on their specific preferences. This paper focuses on constructing Rashomon sets for sparse generalised add... | Rebuttal 1:
Rebuttal: Thank you so much for your review! We really appreciate it! See below for our response to your questions.
> Q: I have difficulty understanding the distinction between fixed support sets and different support sets in Section 3.2. Could you please provide further clarification or explanation on thi... | Summary: this paper designs a new editable gam model. They leverage the ellipsoid to approximate the optimal solution set. The model allows to edit the model according to different use-cases and requirements.
The edited parameter could be solved by constrained quadratic problem. They experiment the model and edited use... | Rebuttal 1:
Rebuttal: Thank you so much for your review! We really appreciate it! See below for our response to your questions.
> Q: the procedure of approximating the Rashomon set has high complexity. The experiments don't evaluate the overhead of this part.
We do show the running time of approximating the Rashomon ... | Summary: This paper presents a novel algorithm (and an additional variant) to learn the Rashomon set of sparse generalized additive models (GAMs) approximated by an ellipsoid in the parameter space. The paper also presents 4 interesting use cases of the proposed algorithm. It includes a rich experiments section. In the... | Rebuttal 1:
Rebuttal: Thank you so much for your review! We really appreciate it! See below for our response to your questions.
> Q: The problem this paper studies is a niche in the general ML problem space.
In fact, GAMs are a widely used form of interpretable predictive model. The vast majority of healthcare scorin... | null | null | 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.