paper_id string | review_id string | year int64 | title string | abstract string | keywords string | initial_review string | initial_review_is_estimated int64 | initial_rating int64 | review_created_at_ms int64 | rebuttal string | rebuttal_messages string | general_response string | general_response_messages string | authors string | author_affiliations string | arxiv_id string | arxiv_categories string | arxiv_v1_timestamp_ms int64 | arxiv_before_review int64 | label string |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
014CgNPAGy | 3FOfBFYrdO | 2,024 | On the Role of Momentum in the Implicit Bias of Gradient Descent for Diagonal Linear Networks | Momentum is a widely adopted and crucial modification to gradient descent when training modern deep neural networks. In this paper, we target on the regularization effect of momentum-based methods in regression settings and analyze a popular proxy model, diagonal linear networks, to precisely characterize the implicit ... | ["GD","momentum","implicit bias","linear networks"] | summary:
The paper studies the implicit regularization effect of the momentum methods through their continuous-time approximations in the case of diagonal networks. The paper discusses the role of initialization and the gradient of the loss at initialization on the recovered implicit bias. The theory is supported by ex... | 1 | 3 | 1,698,928,691,830 | Our response
We thank the reviewer for the valuable comments and questions. Below we address your concerns.
### **Weaknesses Part**
----
**(1). On neglecting second-order terms.**
We would like to clarify that neglecting the second-order terms for the ODE is eligible. This is because the momentum ODE is established ... | [{"id":"HAh8EktwQi","parent_id":"3FOfBFYrdO","created_at_ms":1700215445208,"modified_at_ms":1700215445208,"author":"Authors","role":"authors","type":"Author Reply","content":{"title":"Our response","comment":"We thank the reviewer for the valuable comments and questions. Below we address your concerns.\n\n### **Weaknes... | null | [] | ["Bochen Lyu","Zhanxing Zhu"] | [null,"University of Southampton"] | null | null | null | 0 | unchanged |
014CgNPAGy | 4tQOZWtPAF | 2,024 | On the Role of Momentum in the Implicit Bias of Gradient Descent for Diagonal Linear Networks | Momentum is a widely adopted and crucial modification to gradient descent when training modern deep neural networks. In this paper, we target on the regularization effect of momentum-based methods in regression settings and analyze a popular proxy model, diagonal linear networks, to precisely characterize the implicit ... | ["GD","momentum","implicit bias","linear networks"] | summary:
In this manuscript, the authors rigorously examine the implicit biases introduced by momentum-based optimization techniques, specifically focusing on the Heavy-Ball (HB) method and Nesterov's Accelerated Gradient (NAG) scheme, within the context of diagonal linear networks. The paper establishes a dual role fo... | 1 | 6 | 1,698,614,073,404 | Our response
We thank the reviewer for the valuable comments and the appreciation of our work. Below we answer your questions.
### **On Weaknesses Part**
----
1. **On constraints of the architecture.**
Diagonal linear networks exhibit many important and interesting properties of complex neural networks such as t... | [{"id":"a7ZXMKhWr2","parent_id":"4tQOZWtPAF","created_at_ms":1700215616913,"modified_at_ms":1700215616913,"author":"Authors","role":"authors","type":"Author Reply","content":{"title":"Our response","comment":"We thank the reviewer for the valuable comments and the appreciation of our work. Below we answer your question... | null | [] | ["Bochen Lyu","Zhanxing Zhu"] | [null,"University of Southampton"] | null | null | null | 0 | unchanged |
014CgNPAGy | lhfWAErpnv | 2,024 | On the Role of Momentum in the Implicit Bias of Gradient Descent for Diagonal Linear Networks | Momentum is a widely adopted and crucial modification to gradient descent when training modern deep neural networks. In this paper, we target on the regularization effect of momentum-based methods in regression settings and analyze a popular proxy model, diagonal linear networks, to precisely characterize the implicit ... | ["GD","momentum","implicit bias","linear networks"] | summary:
This paper extends the analysis of (Woodworth et al., 2020) for the implicit bias over the diagonal linear network from GD to GD with momentum (HB and NAG). Specifically, this paper considers a continuous-time approximation of HB and NAG, and shows that the limiting iteration solves a similar but different min... | 1 | 5 | 1,698,673,110,283 | Our response, Part I/II
We thank the reviewer so much for the valuable comments and interesting questions. Blow we answer your questions.
### **Weaknesses Part**
----
1. **On missing related works**.
We greatly appreciate the reviewer for suggesting these related works. Below we discuss these related works.
... | [{"id":"yfXahfwC5v","parent_id":"lhfWAErpnv","created_at_ms":1700206529596,"modified_at_ms":1700221500679,"author":"Authors","role":"authors","type":"Author Reply","content":{"title":"Our response, Part I/II","comment":"We thank the reviewer so much for the valuable comments and interesting questions. Blow we answer yo... | null | [] | ["Bochen Lyu","Zhanxing Zhu"] | [null,"University of Southampton"] | null | null | null | 0 | unchanged |
014CgNPAGy | oxDqgUJK9z | 2,024 | On the Role of Momentum in the Implicit Bias of Gradient Descent for Diagonal Linear Networks | Momentum is a widely adopted and crucial modification to gradient descent when training modern deep neural networks. In this paper, we target on the regularization effect of momentum-based methods in regression settings and analyze a popular proxy model, diagonal linear networks, to precisely characterize the implicit ... | ["GD","momentum","implicit bias","linear networks"] | summary:
The work explores the role of momentum gradient descent (HB and NAG) in the implicit bias for diagonal linear networks. The authors borrow the $O(\eta)$-close continuous flow approximation for momentum from Kovachki and Stuart and use it to derive the continuous flow for the diaognal linear network. Then they ... | 1 | 1 | 1,697,489,944,281 | Our response, Part I/III
We thank the reviewer for the detailed comments and valuable suggestions. Below we address your concerns in detail.
### **Weaknesses part**
-----
**(1). On "Major novelty issues"**.
We were not aware of the nice related work [1] and we thank the reviewer for the suggestion of it. We includ... | [{"id":"2fFhM06UB4","parent_id":"oxDqgUJK9z","created_at_ms":1700203659407,"modified_at_ms":1700222020724,"author":"Authors","role":"authors","type":"Author Reply","content":{"title":"Our response, Part I/III","comment":"We thank the reviewer for the detailed comments and valuable suggestions. Below we address your con... | null | [] | ["Bochen Lyu","Zhanxing Zhu"] | [null,"University of Southampton"] | null | null | null | 0 | changed |
01ep65umEr | 4XUoTXBnFE | 2,024 | TeLLMe what you see: Using LLMs to Explain Neurons in Vision Models | As the role of machine learning models continues to expand across diverse fields, the demand for model interpretability grows. This is particularly crucial for deep learning models, which are often referred to as black boxes, due to their highly nonlinear nature. This paper proposes a novel method for generating and ev... | ["Explainable AI","Explaining Neurons in Vision Models"] | summary:
The paper proposes a method for generating explanations of individual neurons in trained vision models like AlexNet and ViT with the help from Large Language Model. This helps explain the decision-making in these neural networks. The main contributions are to propose two techniques to generate short, human-rea... | 1 | 5 | 1,698,692,088,308 | Dear Reviewer 1NcR,
Thank you very much for taking the time to thoroughly review our work and for giving us the opportunity to clear up some of the questions.
Weakness 1) As you correctly point out, when trying to understand neurons in vision models, we have a strong suite of visualization techniques at our disposal.... | [{"id":"ClFVUCqAyr","parent_id":"4XUoTXBnFE","created_at_ms":1700462829475,"modified_at_ms":1700462829475,"author":"Authors","role":"authors","type":"Author Reply","content":{"comment":"Dear Reviewer 1NcR,\n\nThank you very much for taking the time to thoroughly review our work and for giving us the opportunity to clea... | Dear Reviewers,
We thank you for the time and effort you have invested in reviewing our paper. Your insightful comments and constructive criticisms have been instrumental in enhancing our work, particularly in guiding us to conduct additional experiments. We are especially appreciative of your recognition of the novel... | [{"id":"jho41h1YRQ","parent_id":"01ep65umEr","created_at_ms":1700705342598,"modified_at_ms":1700705342598,"author":"Authors","role":"authors","type":"Author Reply","content":{"comment":"Dear Reviewers,\n\nWe thank you for the time and effort you have invested in reviewing our paper. Your insightful comments and constru... | ["Leon Guertler","M Ganesh Kumar","Anh Tuan Luu","Cheston Tan"] | ["Nanyang Technological University","Harvard University","Nanyang Technological University","Singapore University of Technology and Design"] | null | null | null | 0 | unchanged |
01ep65umEr | JYAB3KxWHW | 2,024 | TeLLMe what you see: Using LLMs to Explain Neurons in Vision Models | As the role of machine learning models continues to expand across diverse fields, the demand for model interpretability grows. This is particularly crucial for deep learning models, which are often referred to as black boxes, due to their highly nonlinear nature. This paper proposes a novel method for generating and ev... | ["Explainable AI","Explaining Neurons in Vision Models"] | summary:
They utilize GPT-4 to provide explanations from vision models based on the neuron activations and weights of the vision model. They use AlexNet and ViT as their vision models to show their explanations. They are comparing the neuron explanations against Clip-Dissect. Also with their method, they have 2 differ... | 1 | 5 | 1,698,379,271,519 | Dear Reviewer NqpP,
Thank you very much for taking the time to thoroughly review our work and for giving us the opportunity to clear up some of the questions.
## Weakness 1)
Thank you very much for reading our work so thoroughly. All typos have been fixed.
## Weakness 2)
In weakness 2 you raise an important point ... | [{"id":"4lRasAbnai","parent_id":"JYAB3KxWHW","created_at_ms":1700665918689,"modified_at_ms":1700665918689,"author":"Authors","role":"authors","type":"Author Reply","content":{"comment":"Dear Reviewer NqpP,\n\nThank you very much for taking the time to thoroughly review our work and for giving us the opportunity to clea... | null | [] | ["Leon Guertler","M Ganesh Kumar","Anh Tuan Luu","Cheston Tan"] | ["Nanyang Technological University","Harvard University","Nanyang Technological University","Singapore University of Technology and Design"] | null | null | null | 0 | changed |
01ep65umEr | ZPfUx2m7Be | 2,024 | TeLLMe what you see: Using LLMs to Explain Neurons in Vision Models | As the role of machine learning models continues to expand across diverse fields, the demand for model interpretability grows. This is particularly crucial for deep learning models, which are often referred to as black boxes, due to their highly nonlinear nature. This paper proposes a novel method for generating and ev... | ["Explainable AI","Explaining Neurons in Vision Models"] | summary:
This paper introduces an interesting idea of using Language Model Models to explain the behaviour of neurons in vision models. While the idea is compelling, it could benefit from more robust evaluations and clarifications in several areas:
strengths:
- The idea of explaining neurons with language models is ti... | 1 | 5 | 1,698,766,298,550 | Dear Reviewer KpfM,
Thank you very much for taking the time to thoroughly review our work and for giving us the opportunity to clear up some of the questions.
Weakness 1) This is a fair thing to point out. We were following the set-up used in [1], but after your review, re-ran some of the AlexNet (weight-label) exper... | [{"id":"kOwh1zsut4","parent_id":"ZPfUx2m7Be","created_at_ms":1700462438539,"modified_at_ms":1700462438539,"author":"Authors","role":"authors","type":"Author Reply","content":{"comment":"Dear Reviewer KpfM,\n\nThank you very much for taking the time to thoroughly review our work and for giving us the opportunity to clea... | Dear Reviewers,
We thank you for the time and effort you have invested in reviewing our paper. Your insightful comments and constructive criticisms have been instrumental in enhancing our work, particularly in guiding us to conduct additional experiments. We are especially appreciative of your recognition of the novel... | [{"id":"jho41h1YRQ","parent_id":"01ep65umEr","created_at_ms":1700705342598,"modified_at_ms":1700705342598,"author":"Authors","role":"authors","type":"Author Reply","content":{"comment":"Dear Reviewers,\n\nWe thank you for the time and effort you have invested in reviewing our paper. Your insightful comments and constru... | ["Leon Guertler","M Ganesh Kumar","Anh Tuan Luu","Cheston Tan"] | ["Nanyang Technological University","Harvard University","Nanyang Technological University","Singapore University of Technology and Design"] | null | null | null | 0 | unchanged |
01ep65umEr | adhKqmsmjk | 2,024 | TeLLMe what you see: Using LLMs to Explain Neurons in Vision Models | As the role of machine learning models continues to expand across diverse fields, the demand for model interpretability grows. This is particularly crucial for deep learning models, which are often referred to as black boxes, due to their highly nonlinear nature. This paper proposes a novel method for generating and ev... | ["Explainable AI","Explaining Neurons in Vision Models"] | summary:
The paper presents a novel method to interpret neuron functions in vision models, enhancing the understanding of deep learning 'black boxes.' Adapted from GPT-4 interpretability techniques, this approach uses neuron activations and weights, providing explanations for neurons of an AlexNet and ViT model trained... | 1 | 5 | 1,699,471,622,944 | Additional experiments
Dear Reviewer xVYG,
Thank you very much for taking the time to thoroughly review our work and for giving us the opportunity to clear up some of the questions.
Weakness 1) As you point out, using visualization techniques has lead to a good level of explainability in the field of computer vision... | [{"id":"HQMshxWaRk","parent_id":"adhKqmsmjk","created_at_ms":1700462327968,"modified_at_ms":1700462327968,"author":"Authors","role":"authors","type":"Author Reply","content":{"title":"Additional experiments","comment":"Dear Reviewer xVYG,\n\nThank you very much for taking the time to thoroughly review our work and for ... | Dear Reviewers,
We thank you for the time and effort you have invested in reviewing our paper. Your insightful comments and constructive criticisms have been instrumental in enhancing our work, particularly in guiding us to conduct additional experiments. We are especially appreciative of your recognition of the novel... | [{"id":"jho41h1YRQ","parent_id":"01ep65umEr","created_at_ms":1700705342598,"modified_at_ms":1700705342598,"author":"Authors","role":"authors","type":"Author Reply","content":{"comment":"Dear Reviewers,\n\nWe thank you for the time and effort you have invested in reviewing our paper. Your insightful comments and constru... | ["Leon Guertler","M Ganesh Kumar","Anh Tuan Luu","Cheston Tan"] | ["Nanyang Technological University","Harvard University","Nanyang Technological University","Singapore University of Technology and Design"] | null | null | null | 0 | unchanged |
030cjlZm4a | 6z8d0fHka1 | 2,024 | Learning Predictive Checklists with Probabilistic Logic Programming | Checklists have been widely recognized as effective tools for completing complex tasks in a systematic manner. Although originally intended for use in procedural tasks, their interpretability and ease of use have led to their adoption for predictive tasks as well, including in clinical settings. However, designing chec... | ["Predictive Checklists","Interpretability","Fairness","Probabilistic Logic Programming"] | summary:
The authors propose a new framework for learning predictive checklists. The method is able to process time series, images, tabular features, etc. The use of techniques enabling sparse representations of the inputs and the use of a fairness metric lead to checklists that would have interpretable features and th... | 1 | 6 | 1,698,683,826,765 | Rebuttal by Authors (1/3)
We thank the reviewer for their valuable comments and insights.
```1 - Interpretability is not inherent to a family of models. For example, a checklist whose features aren’t interpretable or a checklist with too many « checks » to look at (as with linear models) isn’t interpretable either, fo... | [{"id":"II3sUyugh9","parent_id":"6z8d0fHka1","created_at_ms":1700476159199,"modified_at_ms":1700476159199,"author":"Authors","role":"authors","type":"Author Reply","content":{"title":"Rebuttal by Authors (1/3)","comment":"We thank the reviewer for their valuable comments and insights.\n\n```1 - Interpretability is not ... | Global Response (1/2)
We want to thank all reviewers for their insightful comments that contributed to improving our paper. We are glad to report that we have implemented all the suggested changes in the paper.
We address here main topics that were raised in the reviews.
**1) Interpretability of checklist structure (... | [{"id":"yWf6eS9qMU","parent_id":"030cjlZm4a","created_at_ms":1700472306279,"modified_at_ms":1700472333822,"author":"Authors","role":"authors","type":"Author Reply","content":{"title":"Global Response (1/2)","comment":"We want to thank all reviewers for their insightful comments that contributed to improving our paper. ... | ["Yukti Makhija","Edward De Brouwer","Rahul G Krishnan"] | [null,"Yale University","Department of Computer Science, University of Toronto"] | 2411.1679 | cs.LG | 1,732,525,639,000 | 0 | changed |
030cjlZm4a | NpCjpCtarH | 2,024 | Learning Predictive Checklists with Probabilistic Logic Programming | Checklists have been widely recognized as effective tools for completing complex tasks in a systematic manner. Although originally intended for use in procedural tasks, their interpretability and ease of use have led to their adoption for predictive tasks as well, including in clinical settings. However, designing chec... | ["Predictive Checklists","Interpretability","Fairness","Probabilistic Logic Programming"] | summary:
The paper proposes a novel method based on probabilistic logic programming to learn predictive checklists from diverse data modalities including images and time series. The proposed approach was validated using several public benchmark datasets.
strengths:
- Originality: The paper demonstrates originality in ... | 1 | 3 | 1,698,794,126,465 | Rebuttal by Authors
We thank the reviewer for their valuable comments and insights.
``` 1) Model comparison in Table 1 would be better to also be illustrated in graph and plots for easy visualization. ```
We thank the reviewer for this suggestion. We have added this plot to Section I of the appendix to help the reade... | [{"id":"kRDsBsnMG5","parent_id":"NpCjpCtarH","created_at_ms":1700475814319,"modified_at_ms":1700735217999,"author":"Authors","role":"authors","type":"Author Reply","content":{"title":"Rebuttal by Authors","comment":"We thank the reviewer for their valuable comments and insights.\n\n``` 1) Model comparison in Table 1 wo... | Global Response (1/2)
We want to thank all reviewers for their insightful comments that contributed to improving our paper. We are glad to report that we have implemented all the suggested changes in the paper.
We address here main topics that were raised in the reviews.
**1) Interpretability of checklist structure (... | [{"id":"yWf6eS9qMU","parent_id":"030cjlZm4a","created_at_ms":1700472306279,"modified_at_ms":1700472333822,"author":"Authors","role":"authors","type":"Author Reply","content":{"title":"Global Response (1/2)","comment":"We want to thank all reviewers for their insightful comments that contributed to improving our paper. ... | ["Yukti Makhija","Edward De Brouwer","Rahul G Krishnan"] | [null,"Yale University","Department of Computer Science, University of Toronto"] | 2411.1679 | cs.LG | 1,732,525,639,000 | 0 | unchanged |
030cjlZm4a | ZHqzASP0nX | 2,024 | Learning Predictive Checklists with Probabilistic Logic Programming | Checklists have been widely recognized as effective tools for completing complex tasks in a systematic manner. Although originally intended for use in procedural tasks, their interpretability and ease of use have led to their adoption for predictive tasks as well, including in clinical settings. However, designing chec... | ["Predictive Checklists","Interpretability","Fairness","Probabilistic Logic Programming"] | summary:
This paper proposes a method for learning checklist models for diverse data modalities, such as images, time series, and text. Checklists are a type of interpretable models that are widely used in clinical settings. A checklist model consists of a set of concepts, each of which is assigned an integer weight (a... | 1 | 5 | 1,698,839,690,002 | Rebuttal by Authors (1/2)
We thank the reviewer for their valuable comments and insights.
```1) There are too many hyperparameters in the proposed method, including the weight of the regularization terms, the number of concepts, and the threshold, in addition to the architecture details of the neural networks. The aut... | [{"id":"CsUTv7ngW6","parent_id":"ZHqzASP0nX","created_at_ms":1700474127658,"modified_at_ms":1700474127658,"author":"Authors","role":"authors","type":"Author Reply","content":{"title":"Rebuttal by Authors (1/2)","comment":"We thank the reviewer for their valuable comments and insights.\n\n```1) There are too many hyperp... | Global Response (1/2)
We want to thank all reviewers for their insightful comments that contributed to improving our paper. We are glad to report that we have implemented all the suggested changes in the paper.
We address here main topics that were raised in the reviews.
**1) Interpretability of checklist structure (... | [{"id":"yWf6eS9qMU","parent_id":"030cjlZm4a","created_at_ms":1700472306279,"modified_at_ms":1700472333822,"author":"Authors","role":"authors","type":"Author Reply","content":{"title":"Global Response (1/2)","comment":"We want to thank all reviewers for their insightful comments that contributed to improving our paper. ... | ["Yukti Makhija","Edward De Brouwer","Rahul G Krishnan"] | [null,"Yale University","Department of Computer Science, University of Toronto"] | 2411.1679 | cs.LG | 1,732,525,639,000 | 0 | changed |
06mzMua9Rw | LC76iV9ObE | 2,024 | A Trust Region Approach for Few-Shot Sim-to-Real Reinforcement Learning | Simulation-to-Reality Reinforcement Learning (Sim-to-Real RL) seeks to use simulations to minimize the need for extensive real-world interactions. Specifically, in the few-shot off-dynamics setting, the goal is to acquire a simulator-based policy despite a dynamics mismatch that can be effectively transferred to the re... | ["Reinforcement Learning","Simulation-to-Reality","Off-Dynamics"] | summary:
The paper tackles the problem of few-shot sim-to-real. To mimic this, the experimental setting is consider the setting where a source and target simulator is available. The proposed approach is a penalized variant of a trust region approach, where the reward is maximized with an additional term to minimize a d... | 1 | 5 | 1,698,817,475,389 | First answer to Reviewer 1
Dear reviewer, thank you for your detailed report and your interest in our work. Here are the answers to your questions.
**General paper update:** We have improved the FOOD agent by considering a regularization between state-action-state visitation distributions. Not only FOOD is now direct... | [{"id":"BNlCattRGx","parent_id":"LC76iV9ObE","created_at_ms":1700082677423,"modified_at_ms":1700082677423,"author":"Authors","role":"authors","type":"Author Reply","content":{"title":"First answer to Reviewer 1","comment":"Dear reviewer, thank you for your detailed report and your interest in our work. Here are the ans... | null | [] | ["Paul Daoudi","Bogdan Robu","CHRISTOPHE PRIEUR","Ludovic Dos Santos","Merwan Barlier"] | ["Huawei Technologies Ltd.",null,"CNRS","Huawei Technologies Ltd.","Criteo AI Lab"] | null | null | null | 0 | unchanged |
06mzMua9Rw | PxkVBUuf2z | 2,024 | A Trust Region Approach for Few-Shot Sim-to-Real Reinforcement Learning | Simulation-to-Reality Reinforcement Learning (Sim-to-Real RL) seeks to use simulations to minimize the need for extensive real-world interactions. Specifically, in the few-shot off-dynamics setting, the goal is to acquire a simulator-based policy despite a dynamics mismatch that can be effectively transferred to the re... | ["Reinforcement Learning","Simulation-to-Reality","Off-Dynamics"] | summary:
The paper introduces a new sim-to-real transfer algorithm named FOOD for maximizing policy performance in simulation as well as minimizing the state visitation discrepancy between simulation and real environment, so as to achieve high performance considering the dynamics shift. Although the experiments in diff... | 1 | 3 | 1,698,557,665,792 | First answer to Reviewer 2
Dear reviewer, thank you for your detailed report and your interest in our work. Here are the answers to your questions.
**General paper update:** We have improved the FOOD agent by considering a regularization between state-action-state visitation distributions. Not only FOOD is now direct... | [{"id":"w4ehEGHGLD","parent_id":"PxkVBUuf2z","created_at_ms":1700082855030,"modified_at_ms":1700082855030,"author":"Authors","role":"authors","type":"Author Reply","content":{"title":"First answer to Reviewer 2","comment":"Dear reviewer, thank you for your detailed report and your interest in our work. Here are the ans... | null | [] | ["Paul Daoudi","Bogdan Robu","CHRISTOPHE PRIEUR","Ludovic Dos Santos","Merwan Barlier"] | ["Huawei Technologies Ltd.",null,"CNRS","Huawei Technologies Ltd.","Criteo AI Lab"] | null | null | null | 0 | unchanged |
06mzMua9Rw | tuneGcsB5p | 2,024 | A Trust Region Approach for Few-Shot Sim-to-Real Reinforcement Learning | Simulation-to-Reality Reinforcement Learning (Sim-to-Real RL) seeks to use simulations to minimize the need for extensive real-world interactions. Specifically, in the few-shot off-dynamics setting, the goal is to acquire a simulator-based policy despite a dynamics mismatch that can be effectively transferred to the re... | ["Reinforcement Learning","Simulation-to-Reality","Off-Dynamics"] | summary:
This paper focuses on the Sim2Real problem in reinforcement learning. Due to the Sim2Real gap, the well-trained policy from the simulator might perform poorly in real-world scenarios. This paper proposes to constrain the state-action visitation distribution in the real world to be close to that distribution in... | 1 | 5 | 1,698,209,101,925 | First answer to Reviewer 3
Dear reviewer, thank you for your detailed report and your interest in our work. Here are the answers to your questions.
**General paper update:** We have improved the FOOD agent by considering a regularization between state-action-state visitation distributions. Not only FOOD is now direct... | [{"id":"SMo8121VHa","parent_id":"tuneGcsB5p","created_at_ms":1700083046959,"modified_at_ms":1700085582289,"author":"Authors","role":"authors","type":"Author Reply","content":{"title":"First answer to Reviewer 3","comment":"Dear reviewer, thank you for your detailed report and your interest in our work. Here are the ans... | null | [] | ["Paul Daoudi","Bogdan Robu","CHRISTOPHE PRIEUR","Ludovic Dos Santos","Merwan Barlier"] | ["Huawei Technologies Ltd.",null,"CNRS","Huawei Technologies Ltd.","Criteo AI Lab"] | null | null | null | 0 | unchanged |
06mzMua9Rw | vd1VdsmJQx | 2,024 | A Trust Region Approach for Few-Shot Sim-to-Real Reinforcement Learning | Simulation-to-Reality Reinforcement Learning (Sim-to-Real RL) seeks to use simulations to minimize the need for extensive real-world interactions. Specifically, in the few-shot off-dynamics setting, the goal is to acquire a simulator-based policy despite a dynamics mismatch that can be effectively transferred to the re... | ["Reinforcement Learning","Simulation-to-Reality","Off-Dynamics"] | summary:
This paper presents a trust-region-inspired scheme for incorporating real-world data into simulation-based (online) RL. Results and comparisons to relevant baselines illustrate that it can perform well, and a sensitivity analysis provides some insight into the key hyperparameter of the method.
strengths:
- Pr... | 1 | 3 | 1,697,899,450,133 | First answer to Reviewer 4
Dear reviewer, thank you for your detailed report and your interest in our work. We took your suggestions into account and updated the paper. Here are the answers to your questions.
**General paper update:** We have improved the FOOD agent by considering a regularization between state-action... | [{"id":"pUMMrKdgxz","parent_id":"vd1VdsmJQx","created_at_ms":1700085549843,"modified_at_ms":1700085549843,"author":"Authors","role":"authors","type":"Author Reply","content":{"title":"First answer to Reviewer 4","comment":"Dear reviewer, thank you for your detailed report and your interest in our work. We took your sug... | null | [] | ["Paul Daoudi","Bogdan Robu","CHRISTOPHE PRIEUR","Ludovic Dos Santos","Merwan Barlier"] | ["Huawei Technologies Ltd.",null,"CNRS","Huawei Technologies Ltd.","Criteo AI Lab"] | null | null | null | 0 | unchanged |
070DFUdNh7 | 7K2FQbJt3O | 2,024 | GraphGPT: Graph Learning with Generative Pre-trained Transformers | We introduce GraphGPT, a novel model for Graph learning by self-supervised Generative Pre-training Transformers. Our model transforms each graph or sampled subgraph into a sequence of tokens representing the node, edge and attributes reversibly using the Eulerian path first. Then we feed the tokens into a standard tran... | ["Graph","GPT","Generative","Pre-train","Fine-tune","Transformer","GraphGPT"] | summary:
The paper introduces GraphGPT, a new model for graph learning using Generative Pre-training Transformers. By converting graphs or their subgraphs to token sequences via an Eulerian path and pre-training with next-token-prediction, GraphGPT outperforms or matches state-of-the-art methods on various datasets, fo... | 1 | 5 | 1,698,670,948,519 | Discussion on limitations and performance
W1: Discussion on limitations: The paper does not delve into potential limitations of the proposed method.
A: Thank you for pointing out. We will add the following discussion on limitations in the paper:
1. Transferability: Due to the lack of shared semantics among different g... | [{"id":"bJJurrl3ZL","parent_id":"7K2FQbJt3O","created_at_ms":1700647886767,"modified_at_ms":1700647886767,"author":"Authors","role":"authors","type":"Author Reply","content":{"title":"Discussion on limitations and performance","comment":"W1: Discussion on limitations: The paper does not delve into potential limitations... | null | [] | ["Qifang Zhao","Weidong Ren","Tianyu Li","Xiaoxiao Xu","Hong Liu"] | ["Alibaba Group",null,"Alibaba Group","The Hong Kong University of Science and Technology (Guangzhou)","Alibaba Group"] | null | null | null | 0 | unchanged |
070DFUdNh7 | 8ynkrbcL7A | 2,024 | GraphGPT: Graph Learning with Generative Pre-trained Transformers | We introduce GraphGPT, a novel model for Graph learning by self-supervised Generative Pre-training Transformers. Our model transforms each graph or sampled subgraph into a sequence of tokens representing the node, edge and attributes reversibly using the Eulerian path first. Then we feed the tokens into a standard tran... | ["Graph","GPT","Generative","Pre-train","Fine-tune","Transformer","GraphGPT"] | summary:
This paper proposes GraphGPT, a novel model for graph representation learning. The key ideas are: 1) Transform graphs into sequences of tokens via Eulerian paths to preserve structure information. 2) Pretrain the transformer decoder with next token prediction on the graph sequences. 3) Fine-tune on downstream ... | 1 | 3 | 1,698,925,698,888 | Performance concerns and billion-scale models
W1: Performance on some node and edge tasks is not state-of-the-art.
A: Although the performance on some tasks are not SOTA yet, we still present it in the paper, for 2 reasons.
1. GraphGPT’s potential. Although the results are not SOTA, they are already quite close. As p... | [{"id":"tIlPoJ4MH1","parent_id":"8ynkrbcL7A","created_at_ms":1700646926941,"modified_at_ms":1700646926941,"author":"Authors","role":"authors","type":"Author Reply","content":{"title":"Performance concerns and billion-scale models","comment":"W1: Performance on some node and edge tasks is not state-of-the-art.\n\nA: Alt... | null | [] | ["Qifang Zhao","Weidong Ren","Tianyu Li","Xiaoxiao Xu","Hong Liu"] | ["Alibaba Group",null,"Alibaba Group","The Hong Kong University of Science and Technology (Guangzhou)","Alibaba Group"] | null | null | null | 0 | unchanged |
070DFUdNh7 | NVyXWqo4cJ | 2,024 | GraphGPT: Graph Learning with Generative Pre-trained Transformers | We introduce GraphGPT, a novel model for Graph learning by self-supervised Generative Pre-training Transformers. Our model transforms each graph or sampled subgraph into a sequence of tokens representing the node, edge and attributes reversibly using the Eulerian path first. Then we feed the tokens into a standard tran... | ["Graph","GPT","Generative","Pre-train","Fine-tune","Transformer","GraphGPT"] | summary:
This paper presents GraphGPT, an adaptation of the Transformer network tailored for graph data. GraphGPT is adept at handling a wide range of graph datasets and various graph-related tasks, including node, edge, and graph prediction. A notable innovation in this work is the use of (semi-)Eulerian paths to tran... | 1 | 5 | 1,698,724,527,291 | performance concerns and over-smoothing
Q1: performance on graph/node-level tasks and scalability
A: The performance of GraphGPT on large-scale graph-level dataset PCQM4M-v2 (3.7m) is on par with the SOTA with the similar number of parameters.
The performance on the medium-scale graph-level dataset ogbg-molpcba (0.4m... | [{"id":"xZR0wFOVQv","parent_id":"NVyXWqo4cJ","created_at_ms":1700647274284,"modified_at_ms":1700652350244,"author":"Authors","role":"authors","type":"Author Reply","content":{"title":"performance concerns and over-smoothing","comment":"Q1: performance on graph/node-level tasks and scalability\n\nA: The performance of G... | null | [] | ["Qifang Zhao","Weidong Ren","Tianyu Li","Xiaoxiao Xu","Hong Liu"] | ["Alibaba Group",null,"Alibaba Group","The Hong Kong University of Science and Technology (Guangzhou)","Alibaba Group"] | null | null | null | 0 | unchanged |
070DFUdNh7 | atHoYRhAl2 | 2,024 | GraphGPT: Graph Learning with Generative Pre-trained Transformers | We introduce GraphGPT, a novel model for Graph learning by self-supervised Generative Pre-training Transformers. Our model transforms each graph or sampled subgraph into a sequence of tokens representing the node, edge and attributes reversibly using the Eulerian path first. Then we feed the tokens into a standard tran... | ["Graph","GPT","Generative","Pre-train","Fine-tune","Transformer","GraphGPT"] | summary:
The paper proposes a novel model for Graph learning by self-supervised Generative Pre-training Transformers. The proposed method includes: 1) transforming the (sub)graphs into a reversible sequence of tokens via the Eulerian path, 2) pre-training a transformer decoder using the NTP task, and 3) fine-tuning th... | 1 | 5 | 1,699,157,898,001 | More benchmarks and pre-training methods comparison
Q1: Due to high variance of graph benchmakrs, to prove the effectiveness of the proposed methods (graph pretraining & finetuning), I would expect the authors provide more benchmarks results.
A: There are several reasons that we didn’t experiment on more benchmarks.
... | [{"id":"P6vOgpGqyV","parent_id":"atHoYRhAl2","created_at_ms":1700646464175,"modified_at_ms":1700739785809,"author":"Authors","role":"authors","type":"Author Reply","content":{"title":"More benchmarks and pre-training methods comparison","comment":"Q1: Due to high variance of graph benchmakrs, to prove the effectiveness... | null | [] | ["Qifang Zhao","Weidong Ren","Tianyu Li","Xiaoxiao Xu","Hong Liu"] | ["Alibaba Group",null,"Alibaba Group","The Hong Kong University of Science and Technology (Guangzhou)","Alibaba Group"] | null | null | null | 0 | unchanged |
09iOdaeOzp | 0Z9zLQctKo | 2,024 | Sheared LLaMA: Accelerating Language Model Pre-training via Structured Pruning | The popularity of LLaMA (Touvron et al., 2023a;b) and other recently emerged moderate-sized large language models (LLMs) highlights the potential of building smaller yet powerful LLMs. Regardless, the cost of training such models from scratch on trillions of tokens remains high. In this work, we study structured prunin... | ["pruning","efficiency","large language models","pre-training"] | summary:
This paper introduces a pruning technique and a dynamic batching technique to continue training pre-trained LLMs. The proposed pruning technique offers control over the final network shape by solving a constrained optimization problem, where the shape constraint must be optimized while maximizing language mode... | 1 | 6 | 1,698,605,206,336 | About the generality of the method, fair descriptions of the contribution, pruning other models
We are greatly encouraged by the reviewer's positive feedback regarding the practicality and reproducibility of our work. In response to the reviewer's comments, we provide the following responses:
**1. Specific to Transfo... | [{"id":"jd2FW4BVBW","parent_id":"0Z9zLQctKo","created_at_ms":1700186385495,"modified_at_ms":1700229772138,"author":"Authors","role":"authors","type":"Author Reply","content":{"title":"About the generality of the method, fair descriptions of the contribution, pruning other models","comment":"We are greatly encouraged by... | General Response
We thank all reviewers for their constructive feedback! We have updated our paper to incorporate suggestions on typos/clarifications and added the following results:
**1. A fair comparison compared to LLM-Pruner**
Reviewers wgpN and bT2X highlighted the need for a fair comparison with LLM-Pruner with... | [{"id":"IIEpc8vLjz","parent_id":"09iOdaeOzp","created_at_ms":1700184919811,"modified_at_ms":1700184919811,"author":"Authors","role":"authors","type":"Author Reply","content":{"title":"General Response","comment":"We thank all reviewers for their constructive feedback! We have updated our paper to incorporate suggestion... | ["Mengzhou Xia","Tianyu Gao","Zhiyuan Zeng","Danqi Chen"] | ["Princeton University","Princeton University","Tsinghua University","Princeton University"] | 2310.06694 | cs.CL cs.AI cs.LG | 1,696,950,810,000 | 1 | unchanged |
09iOdaeOzp | 5WshVfX9zu | 2,024 | Sheared LLaMA: Accelerating Language Model Pre-training via Structured Pruning | The popularity of LLaMA (Touvron et al., 2023a;b) and other recently emerged moderate-sized large language models (LLMs) highlights the potential of building smaller yet powerful LLMs. Regardless, the cost of training such models from scratch on trillions of tokens remains high. In this work, we study structured prunin... | ["pruning","efficiency","large language models","pre-training"] | summary:
This paper introduces Sheared LLaMA, a method focused on developing streamlined LLaMA models through the pruning of pre-trained ones. Central to this approach is a sparse training technique that employs L0 regularization to systematically zero out specific substructures. Additionally, the paper presents a dyna... | 1 | 5 | 1,698,776,431,112 | About a fair comparison to LLM-Pruner, why dynamic batch loading, comparison to CoFiPruning
Thank you for the valuable feedback! We appreciate that you acknowledge the efficiency implications of our proposed approach to pre-training and the strong performance of our models. We address your comments below:
**1. A far c... | [{"id":"5Kd4BNJtI8","parent_id":"5WshVfX9zu","created_at_ms":1700186037756,"modified_at_ms":1700186069540,"author":"Authors","role":"authors","type":"Author Reply","content":{"title":"About a fair comparison to LLM-Pruner, why dynamic batch loading, comparison to CoFiPruning","comment":"Thank you for the valuable feedb... | General Response
We thank all reviewers for their constructive feedback! We have updated our paper to incorporate suggestions on typos/clarifications and added the following results:
**1. A fair comparison compared to LLM-Pruner**
Reviewers wgpN and bT2X highlighted the need for a fair comparison with LLM-Pruner with... | [{"id":"IIEpc8vLjz","parent_id":"09iOdaeOzp","created_at_ms":1700184919811,"modified_at_ms":1700184919811,"author":"Authors","role":"authors","type":"Author Reply","content":{"title":"General Response","comment":"We thank all reviewers for their constructive feedback! We have updated our paper to incorporate suggestion... | ["Mengzhou Xia","Tianyu Gao","Zhiyuan Zeng","Danqi Chen"] | ["Princeton University","Princeton University","Tsinghua University","Princeton University"] | 2310.06694 | cs.CL cs.AI cs.LG | 1,696,950,810,000 | 1 | unchanged |
09iOdaeOzp | XSnb9ygIEK | 2,024 | Sheared LLaMA: Accelerating Language Model Pre-training via Structured Pruning | The popularity of LLaMA (Touvron et al., 2023a;b) and other recently emerged moderate-sized large language models (LLMs) highlights the potential of building smaller yet powerful LLMs. Regardless, the cost of training such models from scratch on trillions of tokens remains high. In this work, we study structured prunin... | ["pruning","efficiency","large language models","pre-training"] | summary:
The manuscript introduces a technique that prunes pretrained models towards a target (smaller) architecture rather than towards a target sparsity level. After a small amount of retraining, the pruned models outperform similar-size models trained from scratch on many more tokens than the retraining budget, sugg... | 1 | 8 | 1,698,822,664,319 | Inference speed, pruning with C4
**1. Adding inference speedup for different pruning approaches**
Thanks for your suggestion. We've included below the inference speed (tokens/second) comparisons for all methods mentioned you mentioned. These experiments were all conducted on a single A100 80GB GPU, generating up to 20... | [{"id":"LYYSkH8VJj","parent_id":"XSnb9ygIEK","created_at_ms":1700185804481,"modified_at_ms":1700185804481,"author":"Authors","role":"authors","type":"Author Reply","content":{"title":"Inference speed, pruning with C4","comment":"**1. Adding inference speedup for different pruning approaches**\n\nThanks for your suggest... | General Response
We thank all reviewers for their constructive feedback! We have updated our paper to incorporate suggestions on typos/clarifications and added the following results:
**1. A fair comparison compared to LLM-Pruner**
Reviewers wgpN and bT2X highlighted the need for a fair comparison with LLM-Pruner with... | [{"id":"IIEpc8vLjz","parent_id":"09iOdaeOzp","created_at_ms":1700184919811,"modified_at_ms":1700184919811,"author":"Authors","role":"authors","type":"Author Reply","content":{"title":"General Response","comment":"We thank all reviewers for their constructive feedback! We have updated our paper to incorporate suggestion... | ["Mengzhou Xia","Tianyu Gao","Zhiyuan Zeng","Danqi Chen"] | ["Princeton University","Princeton University","Tsinghua University","Princeton University"] | 2310.06694 | cs.CL cs.AI cs.LG | 1,696,950,810,000 | 1 | unchanged |
09iOdaeOzp | zKtytRb7s1 | 2,024 | Sheared LLaMA: Accelerating Language Model Pre-training via Structured Pruning | The popularity of LLaMA (Touvron et al., 2023a;b) and other recently emerged moderate-sized large language models (LLMs) highlights the potential of building smaller yet powerful LLMs. Regardless, the cost of training such models from scratch on trillions of tokens remains high. In this work, we study structured prunin... | ["pruning","efficiency","large language models","pre-training"] | summary:
This paper introduces to employ the structural pruning to reduce the pre-training cost of LLMs. It focuses on the integration of two techniques: targeted structured pruning and dynamic batch loading. This paper proposes two small models: Sheared-LLaMA-1.3B and Sheared-LLaMA-2.7B. Comparative testing reveals t... | 1 | 5 | 1,698,839,708,409 | About novelty, comparison to LLM-Pruner, and pruning another source model
We are encouraged that the reviewer acknowledged that the Sheared-LLaMA models underwent thorough evaluations and demonstrated strong performance. Thanks for your comments and suggestions on the work!
**1. Concerns about the novelty of the paper... | [{"id":"i8zsymzo8w","parent_id":"zKtytRb7s1","created_at_ms":1700185129319,"modified_at_ms":1700185330287,"author":"Authors","role":"authors","type":"Author Reply","content":{"title":"About novelty, comparison to LLM-Pruner, and pruning another source model","comment":"We are encouraged that the reviewer acknowledged t... | General Response
We thank all reviewers for their constructive feedback! We have updated our paper to incorporate suggestions on typos/clarifications and added the following results:
**1. A fair comparison compared to LLM-Pruner**
Reviewers wgpN and bT2X highlighted the need for a fair comparison with LLM-Pruner with... | [{"id":"IIEpc8vLjz","parent_id":"09iOdaeOzp","created_at_ms":1700184919811,"modified_at_ms":1700184919811,"author":"Authors","role":"authors","type":"Author Reply","content":{"title":"General Response","comment":"We thank all reviewers for their constructive feedback! We have updated our paper to incorporate suggestion... | ["Mengzhou Xia","Tianyu Gao","Zhiyuan Zeng","Danqi Chen"] | ["Princeton University","Princeton University","Tsinghua University","Princeton University"] | 2310.06694 | cs.CL cs.AI cs.LG | 1,696,950,810,000 | 1 | unchanged |
0A5o6dCKeK | 31A6UkFLr8 | 2,024 | NExT-GPT: Any-to-Any Multimodal LLM | While recently Multimodal Large Language Models (MM-LLMs) have made exciting strides, they mostly fall prey to the limitation of only input-side multimodal understanding, without the ability to produce content in multiple modalities. As we humans always perceive the world and communicate with people through various mod... | ["Large Language Model","Diffusion Model"] | summary:
This paper proposed an end-to-end general-purpose any-to-any MM-LLM system, NExT-GPT, by connecting an LLM with multimodal adaptors and different diffusion decoders.
strengths:
(1) The paper formulation is good and clear.
(2) Introduced lightweight alignment learning techniques.
(3) Annotated a high-quali... | 1 | 8 | 1,698,440,717,308 | Response to Reviewer mDkx
Thank you so much for highly acknowledging the strengths of our work, such as 'lightweight alignment learning techniques', 'high-quality modality switching instruction-tuning dataset', and 'good and clear formulation', which means a lot to us. Following, we provide our feedback to your concern... | [{"id":"UykhEf0vgw","parent_id":"31A6UkFLr8","created_at_ms":1700443354966,"modified_at_ms":1700443354966,"author":"Authors","role":"authors","type":"Author Reply","content":{"title":"Response to Reviewer mDkx","comment":"Thank you so much for highly acknowledging the strengths of our work, such as 'lightweight alignme... | General Response to All Reviewers
## Dear reviewers,
Thanks for all of your time in providing detailed and constructive comments. Your feedback is important to assist us in enhancing the quality of our work, and we are fully committed to incorporating your suggestions into our revision process. At this juncture, we wo... | [{"id":"r9FyqepaMQ","parent_id":"0A5o6dCKeK","created_at_ms":1700443686942,"modified_at_ms":1700624022350,"author":"Authors","role":"authors","type":"Author Reply","content":{"title":"General Response to All Reviewers","comment":"## Dear reviewers,\n\nThanks for all of your time in providing detailed and constructive c... | ["Shengqiong Wu","Hao Fei","Leigang Qu","Wei Ji","Tat-Seng Chua"] | ["National University of Singapore","National University of Singapore","National University of Singapore","Nanjing University","National University of Singapore"] | 2309.05519 | cs.AI cs.CL cs.LG | 1,694,444,545,000 | 1 | unchanged |
0A5o6dCKeK | GIKH5Kuzyx | 2,024 | NExT-GPT: Any-to-Any Multimodal LLM | While recently Multimodal Large Language Models (MM-LLMs) have made exciting strides, they mostly fall prey to the limitation of only input-side multimodal understanding, without the ability to produce content in multiple modalities. As we humans always perceive the world and communicate with people through various mod... | ["Large Language Model","Diffusion Model"] | summary:
This paper introduces NExT-GPT, an end-to-end, multi-purpose, multi-modal Language Learning Model (MM-LLM) capable of generating text, images, audio, and video. The system is designed to be efficient, utilizing a small quantity of parameters. Furthermore, a multimodal instruction dataset named MosIT is present... | 1 | 5 | 1,698,665,385,127 | Response to Reviewer ijPL
We sincerely appreciate your valuable constructive feedback, which will surely enhance our work. We have conducted additional experiments and made every effort to address your concerns. Whenever you find we have successfully addressed your worries, please reconsider your evaluation, thanks.
... | [{"id":"9gTHmTjA0v","parent_id":"GIKH5Kuzyx","created_at_ms":1700446663923,"modified_at_ms":1700450052423,"author":"Authors","role":"authors","type":"Author Reply","content":{"title":"Response to Reviewer ijPL","comment":"We sincerely appreciate your valuable constructive feedback, which will surely enhance our work. W... | General Response to All Reviewers
## Dear reviewers,
Thanks for all of your time in providing detailed and constructive comments. Your feedback is important to assist us in enhancing the quality of our work, and we are fully committed to incorporating your suggestions into our revision process. At this juncture, we wo... | [{"id":"r9FyqepaMQ","parent_id":"0A5o6dCKeK","created_at_ms":1700443686942,"modified_at_ms":1700624022350,"author":"Authors","role":"authors","type":"Author Reply","content":{"title":"General Response to All Reviewers","comment":"## Dear reviewers,\n\nThanks for all of your time in providing detailed and constructive c... | ["Shengqiong Wu","Hao Fei","Leigang Qu","Wei Ji","Tat-Seng Chua"] | ["National University of Singapore","National University of Singapore","National University of Singapore","Nanjing University","National University of Singapore"] | 2309.05519 | cs.AI cs.CL cs.LG | 1,694,444,545,000 | 1 | unchanged |
0A5o6dCKeK | lFZ2hsGwWM | 2,024 | NExT-GPT: Any-to-Any Multimodal LLM | While recently Multimodal Large Language Models (MM-LLMs) have made exciting strides, they mostly fall prey to the limitation of only input-side multimodal understanding, without the ability to produce content in multiple modalities. As we humans always perceive the world and communicate with people through various mod... | ["Large Language Model","Diffusion Model"] | summary:
This paper proposes a multi-modal LLM, called any-to-any MM-LLM, to extend the multi-modality of LLM to a state where there is no limitation on the input and output modality combinations. To achieve this goal, the authors (1) propose a lightweight alignment learning technique to achieve en effective semantic a... | 1 | 6 | 1,699,320,956,741 | Response to Reviewer 7uQJ
Thank you for your valuable feedback, especially for the recognition of our strengths, such as 'important research', 'valuable datasets', and 'well-structured and easy to follow'. Below, we show you the point-to-point responses, which we hope can address your concerns.
* * *
**Q1: The propos... | [{"id":"HXmEqACXoW","parent_id":"lFZ2hsGwWM","created_at_ms":1700444394515,"modified_at_ms":1700450256523,"author":"Authors","role":"authors","type":"Author Reply","content":{"title":"Response to Reviewer 7uQJ","comment":"Thank you for your valuable feedback, especially for the recognition of our strengths, such as 'im... | General Response to All Reviewers
## Dear reviewers,
Thanks for all of your time in providing detailed and constructive comments. Your feedback is important to assist us in enhancing the quality of our work, and we are fully committed to incorporating your suggestions into our revision process. At this juncture, we wo... | [{"id":"r9FyqepaMQ","parent_id":"0A5o6dCKeK","created_at_ms":1700443686942,"modified_at_ms":1700624022350,"author":"Authors","role":"authors","type":"Author Reply","content":{"title":"General Response to All Reviewers","comment":"## Dear reviewers,\n\nThanks for all of your time in providing detailed and constructive c... | ["Shengqiong Wu","Hao Fei","Leigang Qu","Wei Ji","Tat-Seng Chua"] | ["National University of Singapore","National University of Singapore","National University of Singapore","Nanjing University","National University of Singapore"] | 2309.05519 | cs.AI cs.CL cs.LG | 1,694,444,545,000 | 1 | unchanged |
0A5o6dCKeK | wsJmhYE72b | 2,024 | NExT-GPT: Any-to-Any Multimodal LLM | While recently Multimodal Large Language Models (MM-LLMs) have made exciting strides, they mostly fall prey to the limitation of only input-side multimodal understanding, without the ability to produce content in multiple modalities. As we humans always perceive the world and communicate with people through various mod... | ["Large Language Model","Diffusion Model"] | summary:
This paper proposes a unified framework that enables any-to-any generation. Specifically, it receives inputs from multiple modalities, such as text, audio and video. Furthermore, it leverages off-the-shelf LLMs and diffusion models, which enables efficient training. They propose to leverage special tokens that... | 1 | 5 | 1,698,563,500,435 | Response to Reviewer NwS3
Thank you sincerely for your valuable time and comments. We've conducted additional experiments and tried every effort to address your concerns. We hope you'll reconsider your evaluation if you find our responses are effective.
* * *
***(2-1)***
**Q: The major concern I have regarding this ... | [{"id":"wL9JmjyeIA","parent_id":"wsJmhYE72b","created_at_ms":1700446149654,"modified_at_ms":1700623972864,"author":"Authors","role":"authors","type":"Author Reply","content":{"title":"Response to Reviewer NwS3","comment":"Thank you sincerely for your valuable time and comments. We've conducted additional experiments an... | General Response to All Reviewers
## Dear reviewers,
Thanks for all of your time in providing detailed and constructive comments. Your feedback is important to assist us in enhancing the quality of our work, and we are fully committed to incorporating your suggestions into our revision process. At this juncture, we wo... | [{"id":"r9FyqepaMQ","parent_id":"0A5o6dCKeK","created_at_ms":1700443686942,"modified_at_ms":1700624022350,"author":"Authors","role":"authors","type":"Author Reply","content":{"title":"General Response to All Reviewers","comment":"## Dear reviewers,\n\nThanks for all of your time in providing detailed and constructive c... | ["Shengqiong Wu","Hao Fei","Leigang Qu","Wei Ji","Tat-Seng Chua"] | ["National University of Singapore","National University of Singapore","National University of Singapore","Nanjing University","National University of Singapore"] | 2309.05519 | cs.AI cs.CL cs.LG | 1,694,444,545,000 | 1 | unchanged |
0AYosSFETw | 0Ubl1s0DWb | 2,024 | Towards human-like spoken dialogue generation between AI agents from written dialogue | The advent of large language models (LLMs) has made it possible to generate natural written dialogues between two agents.
However, generating human-like spoken dialogues from these written dialogues remains challenging.
Spoken dialogues have several unique characteristics: they frequently include backchannels and laugh... | ["spoken dialogue modeling","text-to-speech synthesis","backchannel generation","turn-taking"] | summary:
The paper proposes a method to generate natural overlapping spoken dialogue with the listener cues like backchannels and laughter only using the written transcripts (that lack the rich spoken dialog modes). This system generates speech for both the speaker and the listener simultaneously, using only the transc... | 1 | 8 | 1,699,070,181,245 | Response to Reviewer yc8m (1/3)
We thank Reviewer yc8m for the insightful comments regarding our paper. We are in the process of revising our manuscript to address the concerns raised and will inform you once the updated manuscript is ready. Below, we provide our specific responses to each of your comments.
**On Summa... | [{"id":"p3tK0ldoIV","parent_id":"0Ubl1s0DWb","created_at_ms":1700029735264,"modified_at_ms":1700029735264,"author":"Authors","role":"authors","type":"Author Reply","content":{"title":"Response to Reviewer yc8m (1/3)","comment":"We thank Reviewer yc8m for the insightful comments regarding our paper. We are in the proces... | For All Reviewers: Paper Revision
We again thank all reviewers for their invaluable feedback.
In response, we have substantially reorganized the paper to enhance its clarity and self-explanatory nature, as well as to distinguish it more clearly from previous works.
The revised version of the paper is now available, wi... | [{"id":"tMVM1DnCol","parent_id":"0AYosSFETw","created_at_ms":1700456982256,"modified_at_ms":1700456982256,"author":"Authors","role":"authors","type":"Author Reply","content":{"title":"For All Reviewers: Paper Revision","comment":"We again thank all reviewers for their invaluable feedback.\nIn response, we have substant... | ["Kentaro Mitsui","Yukiya Hono","Kei Sawada"] | ["rinna Co., Ltd.","Nagoya Institute of Technology",null] | 2310.01088 | cs.CL cs.LG cs.SD eess.AS | 1,696,244,600,000 | 1 | unchanged |
0AYosSFETw | 2qgRBrOuYh | 2,024 | Towards human-like spoken dialogue generation between AI agents from written dialogue | The advent of large language models (LLMs) has made it possible to generate natural written dialogues between two agents.
However, generating human-like spoken dialogues from these written dialogues remains challenging.
Spoken dialogues have several unique characteristics: they frequently include backchannels and laugh... | ["spoken dialogue modeling","text-to-speech synthesis","backchannel generation","turn-taking"] | summary:
This paper tackles the task of generating spoken dialogues between 2 parties using autoregressive models. It follows the earlier work on DLM (dialogue language model), and tries to extend it for better turn-taking and pause modeling. This results in more natural generated dialogs. The authors train all models ... | 1 | 6 | 1,699,574,573,346 | Response to Reviewer FEtd
We thank Reviewer FEtd for the insightful comments and questions regarding our paper. We are in the process of revising our manuscript to incorporate your suggestions and will inform you once the updated manuscript is ready. Below are our specific responses to each of your comments.
**On Need... | [{"id":"5QXQiLaCVk","parent_id":"2qgRBrOuYh","created_at_ms":1699958937471,"modified_at_ms":1699958937471,"author":"Authors","role":"authors","type":"Author Reply","content":{"title":"Response to Reviewer FEtd","comment":"We thank Reviewer FEtd for the insightful comments and questions regarding our paper. We are in th... | For All Reviewers: Paper Revision
We again thank all reviewers for their invaluable feedback.
In response, we have substantially reorganized the paper to enhance its clarity and self-explanatory nature, as well as to distinguish it more clearly from previous works.
The revised version of the paper is now available, wi... | [{"id":"tMVM1DnCol","parent_id":"0AYosSFETw","created_at_ms":1700456982256,"modified_at_ms":1700456982256,"author":"Authors","role":"authors","type":"Author Reply","content":{"title":"For All Reviewers: Paper Revision","comment":"We again thank all reviewers for their invaluable feedback.\nIn response, we have substant... | ["Kentaro Mitsui","Yukiya Hono","Kei Sawada"] | ["rinna Co., Ltd.","Nagoya Institute of Technology",null] | 2310.01088 | cs.CL cs.LG cs.SD eess.AS | 1,696,244,600,000 | 1 | unchanged |
0AYosSFETw | 5YdY7AKDmz | 2,024 | Towards human-like spoken dialogue generation between AI agents from written dialogue | The advent of large language models (LLMs) has made it possible to generate natural written dialogues between two agents.
However, generating human-like spoken dialogues from these written dialogues remains challenging.
Spoken dialogues have several unique characteristics: they frequently include backchannels and laugh... | ["spoken dialogue modeling","text-to-speech synthesis","backchannel generation","turn-taking"] | summary:
This paper proposes CHATS (CHatty Agents Text-to-Speech), a system for transforming written dialogue into spoken dialogue, whose content is coherent with the input written dialogue but generated with backchannels, laughter, and smooth turn-taking. Several contributions are announced: a method to prepare writte... | 1 | 3 | 1,698,425,008,031 | Response to Reviewer Kha2 (1/2)
We thank Reviewer Kha2 for the insightful feedback on our paper. We are currently revising our manuscript to incorporate your suggestions, and will inform you once the updated manuscript is ready. The revised version will include (1) a clearer statement on the importance of “Conversion f... | [{"id":"8WpawEiReQ","parent_id":"5YdY7AKDmz","created_at_ms":1699958654739,"modified_at_ms":1699958654739,"author":"Authors","role":"authors","type":"Author Reply","content":{"title":"Response to Reviewer Kha2 (1/2)","comment":"We thank Reviewer Kha2 for the insightful feedback on our paper. We are currently revising o... | null | [] | ["Kentaro Mitsui","Yukiya Hono","Kei Sawada"] | ["rinna Co., Ltd.","Nagoya Institute of Technology",null] | 2310.01088 | cs.CL cs.LG cs.SD eess.AS | 1,696,244,600,000 | 1 | changed |
0BqyZSWfzo | My8ST7abN0 | 2,024 | One-shot Empirical Privacy Estimation for Federated Learning | Privacy estimation techniques for differentially private (DP) algorithms are useful for comparing against analytical bounds, or to empirically measure privacy loss in settings where known analytical bounds are not tight. However, existing privacy auditing techniques usually make strong assumptions on the adversary (e.g... | ["differential privacy","federated learning","empirical privacy"] | summary:
The paper proposes a new method to empirically estimate the privacy loss of an algorithm, targeted at federated learning. The method can estimate the privacy loss from just one run of the algorithm, which makes it easy to apply. This is done by randomly sampling canary vectors, which are presented to the centr... | 1 | 5 | 1,697,550,809,607 | Thank you for the careful review. You make some excellent points about the difficulty of comparing two empirical privacy estimates, which we will address in a subsequent comment. In the interest of getting the conversation started during this short review period, we will address the rest of your comments first. You sai... | [{"id":"1IMgMQP0R3","parent_id":"My8ST7abN0","created_at_ms":1699907313521,"modified_at_ms":1699922345822,"author":"Authors","role":"authors","type":"Author Reply","content":{"comment":"Thank you for the careful review. You make some excellent points about the difficulty of comparing two empirical privacy estimates, wh... | null | [] | ["Galen Andrew","Peter Kairouz","Sewoong Oh","Alina Oprea","Hugh Brendan McMahan","Vinith Menon Suriyakumar"] | ["Google","Google","University of Illinois at Urbana-Champaign","Northeastern University","Google","Massachusetts Institute of Technology"] | 2302.03098 | cs.LG cs.CR | 1,675,713,508,000 | 1 | changed |
0BqyZSWfzo | QTJmiDd5dp | 2,024 | One-shot Empirical Privacy Estimation for Federated Learning | Privacy estimation techniques for differentially private (DP) algorithms are useful for comparing against analytical bounds, or to empirically measure privacy loss in settings where known analytical bounds are not tight. However, existing privacy auditing techniques usually make strong assumptions on the adversary (e.g... | ["differential privacy","federated learning","empirical privacy"] | summary:
This work proposes a one-shot empirical privacy estimation method for DP-FedAvg. Instead of choosing canary examples, the authors choose canary clients that releases randomly generated updates to the system. The idea is then to measure the cosine similarity between the overall updates and random updates to det... | 1 | 6 | 1,698,806,434,695 | We appreciate your careful review and suggestions to improve our paper. We believe we can answer most of your concerns.
* *Using canary clients seems more inefficient compared to using canary examples. More resource allocation might be needed.* The reason to use canary clients, as opposed to canary examples, is that w... | [{"id":"r3SH53j11u","parent_id":"QTJmiDd5dp","created_at_ms":1700001976594,"modified_at_ms":1700001976594,"author":"Authors","role":"authors","type":"Author Reply","content":{"comment":"We appreciate your careful review and suggestions to improve our paper. We believe we can answer most of your concerns.\n\n* *Using ca... | New revision uploaded
As the discussion period comes to a close, we would like to gently request that reviewers LDHu and zCeQ confirm whether their questions have been answered.
We had a detailed discussion with reviewer ZyjW about our theoretical results, whose feedback helped us to make the proofs completely rigorou... | [{"id":"mONxDyh0sV","parent_id":"0BqyZSWfzo","created_at_ms":1700268593958,"modified_at_ms":1700268593958,"author":"Authors","role":"authors","type":"Author Reply","content":{"title":"New revision uploaded","comment":"As the discussion period comes to a close, we would like to gently request that reviewers LDHu and zCe... | ["Galen Andrew","Peter Kairouz","Sewoong Oh","Alina Oprea","Hugh Brendan McMahan","Vinith Menon Suriyakumar"] | ["Google","Google","University of Illinois at Urbana-Champaign","Northeastern University","Google","Massachusetts Institute of Technology"] | 2302.03098 | cs.LG cs.CR | 1,675,713,508,000 | 1 | changed |
0BqyZSWfzo | iacTaqQ7Ug | 2,024 | One-shot Empirical Privacy Estimation for Federated Learning | Privacy estimation techniques for differentially private (DP) algorithms are useful for comparing against analytical bounds, or to empirically measure privacy loss in settings where known analytical bounds are not tight. However, existing privacy auditing techniques usually make strong assumptions on the adversary (e.g... | ["differential privacy","federated learning","empirical privacy"] | summary:
This paper proposes an approach for auditing the privacy of differentially-private learning algorithm in the context of federated learning such as DP-FedAvg. One of the benefits of the approach is that it is « one-shot » in the sense that it can be used at the same time as the model is learnt, without needing ... | 1 | 6 | 1,698,825,317,613 | We appreciate your careful review and suggestions to improve our paper. We believe we can answer most of your concerns.
* *In the experiments conducted the number of canaries used is quite large, which is likely to have an impact on the model utility.* We stated in the paper that for our Stackoverflow experiments, wit... | [{"id":"Ci1SzL455K","parent_id":"iacTaqQ7Ug","created_at_ms":1699912292055,"modified_at_ms":1699912292055,"author":"Authors","role":"authors","type":"Author Reply","content":{"comment":"We appreciate your careful review and suggestions to improve our paper. We believe we can answer most of your concerns.\n\n* *In the e... | New revision uploaded
As the discussion period comes to a close, we would like to gently request that reviewers LDHu and zCeQ confirm whether their questions have been answered.
We had a detailed discussion with reviewer ZyjW about our theoretical results, whose feedback helped us to make the proofs completely rigorou... | [{"id":"mONxDyh0sV","parent_id":"0BqyZSWfzo","created_at_ms":1700268593958,"modified_at_ms":1700268593958,"author":"Authors","role":"authors","type":"Author Reply","content":{"title":"New revision uploaded","comment":"As the discussion period comes to a close, we would like to gently request that reviewers LDHu and zCe... | ["Galen Andrew","Peter Kairouz","Sewoong Oh","Alina Oprea","Hugh Brendan McMahan","Vinith Menon Suriyakumar"] | ["Google","Google","University of Illinois at Urbana-Champaign","Northeastern University","Google","Massachusetts Institute of Technology"] | 2302.03098 | cs.LG cs.CR | 1,675,713,508,000 | 1 | changed |
0Ce3c9l7G1 | NbqiwlBLcl | 2,024 | Learning Multi-Agent Communication using Regularized Attention Messages | Learning how to communicate in Multi-Agent Reinforcement Learning (MARL) can be key to solve complex cooperative tasks. Recent approaches have shown the advantages of using an efficient communication architecture, tackling problems such as what, when, or whom to communicate. However, these methods still fail to solve s... | ["Multi-Agent Reinforcement Learning","Communication","Attention","Message Compression"] | summary:
This paper introduces Attentive Regularized Communication (ARCOMM), a method for communication in MARL. ARCOMM incorporates an attention module and a message regularizer to assess the significance of messages and facilitate the learning of more meaningful interactions among agents. The study suggests that ARCO... | 1 | 3 | 1,698,789,596,451 | Rebuttal
Thanks for taking the time to review our paper and for your helpful comments! We address them below and please let us know if there are still any remaining concerns.
**Novelty**
Performances are not significantly better in Figure 2 since SMAC does not necessarily require communication, as discussed in our r... | [{"id":"QkbvyDXVK5","parent_id":"NbqiwlBLcl","created_at_ms":1700281342465,"modified_at_ms":1700281342465,"author":"Authors","role":"authors","type":"Author Reply","content":{"title":"Rebuttal","comment":"Thanks for taking the time to review our paper and for your helpful comments! We address them below and please let ... | General Response
We would like to thank all the reviewers for their time and feedback given to improve our paper. We have updated a revised version with additional experiments (**Appendix B**) and some other clarifications. All the changes are marked in red.
We have addressed the questions of the reviewers individuall... | [{"id":"04hOH3Die0","parent_id":"0Ce3c9l7G1","created_at_ms":1700287003023,"modified_at_ms":1700287351631,"author":"Authors","role":"authors","type":"Author Reply","content":{"title":"General Response","comment":"We would like to thank all the reviewers for their time and feedback given to improve our paper. We have up... | ["Rafael Moreira Pina","Varuna De SIlva De SIlva","Corentin Artaud"] | ["Loughborough University London","Loughborough University","Loughborough University London"] | null | null | null | 0 | unchanged |
0Ce3c9l7G1 | c2D3m4tV79 | 2,024 | Learning Multi-Agent Communication using Regularized Attention Messages | Learning how to communicate in Multi-Agent Reinforcement Learning (MARL) can be key to solve complex cooperative tasks. Recent approaches have shown the advantages of using an efficient communication architecture, tackling problems such as what, when, or whom to communicate. However, these methods still fail to solve s... | ["Multi-Agent Reinforcement Learning","Communication","Attention","Message Compression"] | summary:
This paper focuses on reducing communication overhead in communication MARL. It utilized an attention module and DCT to compress the messages.
strengths:
1. Communication overhead is an important direction in communication-based MARL.
weaknesses:
1. The contribution is rather limited. Tarmac is the first pap... | 1 | 3 | 1,698,817,143,870 | Rebuttal
Thank you for your useful feedback! We have addressed below your comments. We hope that we have clarified your concerns, but please do let us know if there are still questions remaining.
**Contribution and relation with TARMAC**
It is true that other papers have investigated attention since it is a compelli... | [{"id":"CRJon38QFU","parent_id":"c2D3m4tV79","created_at_ms":1700280176824,"modified_at_ms":1700280210545,"author":"Authors","role":"authors","type":"Author Reply","content":{"title":"Rebuttal","comment":"Thank you for your useful feedback! We have addressed below your comments. We hope that we have clarified your conc... | General Response
We would like to thank all the reviewers for their time and feedback given to improve our paper. We have updated a revised version with additional experiments (**Appendix B**) and some other clarifications. All the changes are marked in red.
We have addressed the questions of the reviewers individuall... | [{"id":"04hOH3Die0","parent_id":"0Ce3c9l7G1","created_at_ms":1700287003023,"modified_at_ms":1700287351631,"author":"Authors","role":"authors","type":"Author Reply","content":{"title":"General Response","comment":"We would like to thank all the reviewers for their time and feedback given to improve our paper. We have up... | ["Rafael Moreira Pina","Varuna De SIlva De SIlva","Corentin Artaud"] | ["Loughborough University London","Loughborough University","Loughborough University London"] | null | null | null | 0 | unchanged |
0Ce3c9l7G1 | qLNatTLNHN | 2,024 | Learning Multi-Agent Communication using Regularized Attention Messages | Learning how to communicate in Multi-Agent Reinforcement Learning (MARL) can be key to solve complex cooperative tasks. Recent approaches have shown the advantages of using an efficient communication architecture, tackling problems such as what, when, or whom to communicate. However, these methods still fail to solve s... | ["Multi-Agent Reinforcement Learning","Communication","Attention","Message Compression"] | summary:
The paper studies Multi-agent communication in complex scenarios with limited communication channels. The authors introduces Attentive Regularized Communication (ARCOMM), a method for communication in Multi-Agent Reinforcement Learning (MARL) that uses an attention module and a message regularizer to improve t... | 1 | 5 | 1,698,739,032,805 | Rebuttal
Thank you for providing useful comments to our paper! We are happy that the reviewer agrees with the value of studying lossy communication. We have addressed your comments below but please do let us know if there are still any remaining concerns.
**Relation with TARMAC**
ARCOMM is structurally different fr... | [{"id":"bXv4g0zxGG","parent_id":"qLNatTLNHN","created_at_ms":1700282114044,"modified_at_ms":1700282114044,"author":"Authors","role":"authors","type":"Author Reply","content":{"title":"Rebuttal","comment":"Thank you for providing useful comments to our paper! We are happy that the reviewer agrees with the value of study... | General Response
We would like to thank all the reviewers for their time and feedback given to improve our paper. We have updated a revised version with additional experiments (**Appendix B**) and some other clarifications. All the changes are marked in red.
We have addressed the questions of the reviewers individuall... | [{"id":"04hOH3Die0","parent_id":"0Ce3c9l7G1","created_at_ms":1700287003023,"modified_at_ms":1700287351631,"author":"Authors","role":"authors","type":"Author Reply","content":{"title":"General Response","comment":"We would like to thank all the reviewers for their time and feedback given to improve our paper. We have up... | ["Rafael Moreira Pina","Varuna De SIlva De SIlva","Corentin Artaud"] | ["Loughborough University London","Loughborough University","Loughborough University London"] | null | null | null | 0 | unchanged |
0Ce3c9l7G1 | soDAz5gNpf | 2,024 | Learning Multi-Agent Communication using Regularized Attention Messages | Learning how to communicate in Multi-Agent Reinforcement Learning (MARL) can be key to solve complex cooperative tasks. Recent approaches have shown the advantages of using an efficient communication architecture, tackling problems such as what, when, or whom to communicate. However, these methods still fail to solve s... | ["Multi-Agent Reinforcement Learning","Communication","Attention","Message Compression"] | summary:
Despite numerous successful demonstrations of the benefits of communication in solving cooperative tasks with multi-agent reinforcement learning (MARL), these methods could still struggle in complex scenarios like when communication channels have limited capacity. The paper proposes an attention module to weig... | 1 | 3 | 1,698,698,141,758 | Rebuttal
Thank you for the helpful comments! We addressed them below but please let us know if you still have remaining questions.
1. We apologize for the imprecision in some of the language. The paper has been reviewed accordingly to address those. Regarding the specific points: with “messages with quality” we mean m... | [{"id":"MvzWXAOMrq","parent_id":"soDAz5gNpf","created_at_ms":1700283639869,"modified_at_ms":1700283639869,"author":"Authors","role":"authors","type":"Author Reply","content":{"title":"Rebuttal","comment":"Thank you for the helpful comments! We addressed them below but please let us know if you still have remaining ques... | General Response
We would like to thank all the reviewers for their time and feedback given to improve our paper. We have updated a revised version with additional experiments (**Appendix B**) and some other clarifications. All the changes are marked in red.
We have addressed the questions of the reviewers individuall... | [{"id":"04hOH3Die0","parent_id":"0Ce3c9l7G1","created_at_ms":1700287003023,"modified_at_ms":1700287351631,"author":"Authors","role":"authors","type":"Author Reply","content":{"title":"General Response","comment":"We would like to thank all the reviewers for their time and feedback given to improve our paper. We have up... | ["Rafael Moreira Pina","Varuna De SIlva De SIlva","Corentin Artaud"] | ["Loughborough University London","Loughborough University","Loughborough University London"] | null | null | null | 0 | changed |
0H6DFoZZXZ | 7yiZml58ZC | 2,024 | Language Control Diffusion: Efficiently Scaling through Space, Time, and Tasks | Training generalist agents is difficult across several axes, requiring us to deal with high-dimensional inputs (space), long horizons (time), and generalization to novel tasks. Recent advances with architectures have allowed for improved scaling along one or two of these axes, but are still computationally prohibitive ... | ["planning","diffusion","language","RL","reinforcement"] | summary:
In this work, the authors propose to scale diffusion models for planning in extended temporal, state, and task dimensions to tackle long-horizon control problems conditioned on natural language instructions. Specifically, they take a hierarchical diffusion approach by training diffusion policy to plan in the l... | 1 | 5 | 1,698,999,254,250 | Rebuttal
> We are grateful to the reviewer for their valuable comments, which allows us to emphasize the novelty of our diffusion model framework. We also justify our choice of the CALVIN benchmark as a testing ground and respond to requests for more extensive experimental evidence.
**Weaknesses:**
The weakness comes... | [{"id":"LjlH2fieRd","parent_id":"7yiZml58ZC","created_at_ms":1700060076115,"modified_at_ms":1700060076115,"author":"Authors","role":"authors","type":"Author Reply","content":{"title":"Rebuttal","comment":"> We are grateful to the reviewer for their valuable comments, which allows us to emphasize the novelty of our diff... | Rebuttal Summary
We extend our sincere gratitude to all the reviewers for their diligent evaluations and constructive feedback. Below, we summarize the new experiments that we have run and key modifications we have made to the manuscript based on the reviewers' input.
## New Experiments
- We have run our MLP (25.5M p... | [{"id":"12Xz5fihUl","parent_id":"0H6DFoZZXZ","created_at_ms":1700060428121,"modified_at_ms":1700060428121,"author":"Authors","role":"authors","type":"Author Reply","content":{"title":"Rebuttal Summary","comment":"We extend our sincere gratitude to all the reviewers for their diligent evaluations and constructive feedba... | ["Edwin Zhang","Yujie Lu","Shinda Huang","William Yang Wang","Amy Zhang"] | ["Harvard University","UC Santa Barbara","University of California, Santa Barbara","Meta Facebook","UC Santa Barbara"] | 2210.15629 | cs.LG cs.AI cs.CL | 1,666,891,250,000 | 1 | unchanged |
0H6DFoZZXZ | C7yxLXp9wz | 2,024 | Language Control Diffusion: Efficiently Scaling through Space, Time, and Tasks | Training generalist agents is difficult across several axes, requiring us to deal with high-dimensional inputs (space), long horizons (time), and generalization to novel tasks. Recent advances with architectures have allowed for improved scaling along one or two of these axes, but are still computationally prohibitive ... | ["planning","diffusion","language","RL","reinforcement"] | summary:
This paper studies text-to-control problems and proposes Language to Control Diffusion models as a hierarchical planner conditioned on language (LCD). This hierarchical approach scales the diffusion model along the spatial, time, and task dimensions.
strengths:
1. The paper is clearly written and easy to foll... | 1 | 6 | 1,698,729,664,848 | Rebuttal
> We appreciate the reviewer's insightful questions and valuable critique. To summarize, we have addressed concerns about the role of low-level controllers, differences between our method and other text-to-video works, the comparability of these methods in experiments, the clarification of certain terms and sy... | [{"id":"89iObGakQs","parent_id":"C7yxLXp9wz","created_at_ms":1700060203452,"modified_at_ms":1700060203452,"author":"Authors","role":"authors","type":"Author Reply","content":{"title":"Rebuttal","comment":"> We appreciate the reviewer's insightful questions and valuable critique. To summarize, we have addressed concerns... | Rebuttal Summary
We extend our sincere gratitude to all the reviewers for their diligent evaluations and constructive feedback. Below, we summarize the new experiments that we have run and key modifications we have made to the manuscript based on the reviewers' input.
## New Experiments
- We have run our MLP (25.5M p... | [{"id":"12Xz5fihUl","parent_id":"0H6DFoZZXZ","created_at_ms":1700060428121,"modified_at_ms":1700060428121,"author":"Authors","role":"authors","type":"Author Reply","content":{"title":"Rebuttal Summary","comment":"We extend our sincere gratitude to all the reviewers for their diligent evaluations and constructive feedba... | ["Edwin Zhang","Yujie Lu","Shinda Huang","William Yang Wang","Amy Zhang"] | ["Harvard University","UC Santa Barbara","University of California, Santa Barbara","Meta Facebook","UC Santa Barbara"] | 2210.15629 | cs.LG cs.AI cs.CL | 1,666,891,250,000 | 1 | unchanged |
0H6DFoZZXZ | UUmTO7PzHM | 2,024 | Language Control Diffusion: Efficiently Scaling through Space, Time, and Tasks | Training generalist agents is difficult across several axes, requiring us to deal with high-dimensional inputs (space), long horizons (time), and generalization to novel tasks. Recent advances with architectures have allowed for improved scaling along one or two of these axes, but are still computationally prohibitive ... | ["planning","diffusion","language","RL","reinforcement"] | summary:
This paper aimed to introduce the hierarchical diffusion policy into robotics control based on language instructions, while this framework faces three challenges: direct long-horizon planning, non-task-specific representation, and computational inefficiency. By proposing a diffusion-based model named Language ... | 1 | 5 | 1,698,675,114,196 | Rebuttal
> We appreciate the thorough review and the valuable suggestions you've provided regarding the clarity of the state encoder, dataset optimality, and the usage of T5-XXL, each of which has helped us to reflect on and improve the manuscript. We would like to address each of these points specifically.
**Weakness... | [{"id":"VH1sHE2vbb","parent_id":"UUmTO7PzHM","created_at_ms":1700060373547,"modified_at_ms":1700060373547,"author":"Authors","role":"authors","type":"Author Reply","content":{"title":"Rebuttal","comment":"> We appreciate the thorough review and the valuable suggestions you've provided regarding the clarity of the state... | Rebuttal Summary
We extend our sincere gratitude to all the reviewers for their diligent evaluations and constructive feedback. Below, we summarize the new experiments that we have run and key modifications we have made to the manuscript based on the reviewers' input.
## New Experiments
- We have run our MLP (25.5M p... | [{"id":"12Xz5fihUl","parent_id":"0H6DFoZZXZ","created_at_ms":1700060428121,"modified_at_ms":1700060428121,"author":"Authors","role":"authors","type":"Author Reply","content":{"title":"Rebuttal Summary","comment":"We extend our sincere gratitude to all the reviewers for their diligent evaluations and constructive feedba... | ["Edwin Zhang","Yujie Lu","Shinda Huang","William Yang Wang","Amy Zhang"] | ["Harvard University","UC Santa Barbara","University of California, Santa Barbara","Meta Facebook","UC Santa Barbara"] | 2210.15629 | cs.LG cs.AI cs.CL | 1,666,891,250,000 | 1 | changed |
0H6DFoZZXZ | pImdNaXavL | 2,024 | Language Control Diffusion: Efficiently Scaling through Space, Time, and Tasks | Training generalist agents is difficult across several axes, requiring us to deal with high-dimensional inputs (space), long horizons (time), and generalization to novel tasks. Recent advances with architectures have allowed for improved scaling along one or two of these axes, but are still computationally prohibitive ... | ["planning","diffusion","language","RL","reinforcement"] | summary:
The authors propose a novel hierarchical framework called LCD, which leverages a language-conditioned diffusion model as a high-level goal planner on top of HULC. The proposed model achieves scalability in the spatial, time, and task dimensions. It outperforms other baselines on the CALVIN benchmark.
strength... | 1 | 6 | 1,699,009,036,759 | Rebuttal
**fAWH**
> We thank the reviewer for their perceptive questions and feedback on the performance of inference using the MLP and Transformer ablation, considerations of parameter size in making comparisons, and the role of HULC as a baseline in LCD. In the responses following, we have sought to address each poi... | [{"id":"wMpO7YHEPA","parent_id":"pImdNaXavL","created_at_ms":1700059853962,"modified_at_ms":1700059853962,"author":"Authors","role":"authors","type":"Author Reply","content":{"title":"Rebuttal","comment":"**fAWH**\n\n> We thank the reviewer for their perceptive questions and feedback on the performance of inference usi... | Rebuttal Summary
We extend our sincere gratitude to all the reviewers for their diligent evaluations and constructive feedback. Below, we summarize the new experiments that we have run and key modifications we have made to the manuscript based on the reviewers' input.
## New Experiments
- We have run our MLP (25.5M p... | [{"id":"12Xz5fihUl","parent_id":"0H6DFoZZXZ","created_at_ms":1700060428121,"modified_at_ms":1700060428121,"author":"Authors","role":"authors","type":"Author Reply","content":{"title":"Rebuttal Summary","comment":"We extend our sincere gratitude to all the reviewers for their diligent evaluations and constructive feedba... | ["Edwin Zhang","Yujie Lu","Shinda Huang","William Yang Wang","Amy Zhang"] | ["Harvard University","UC Santa Barbara","University of California, Santa Barbara","Meta Facebook","UC Santa Barbara"] | 2210.15629 | cs.LG cs.AI cs.CL | 1,666,891,250,000 | 1 | unchanged |
0IaTFNJner | 9kPJm1DFwm | 2,024 | On the Embedding Collapse When Scaling up Recommendation Models | Recent advances in deep foundation models have led to a promising trend of developing large recommendation models to leverage vast amounts of available data. However, we experiment to scale up existing recommendation models and observe that the enlarged models do not improve satisfactorily. In this context, we investig... | ["machine learning","representation learning","recommender system","collapse"] | summary:
The paper addresses challenges in scaling up recommendation models, identifying a phenomenon called "embedding collapse" while enlarging models. The authors introduce information abundance as a metric to measure and evaluate collapse. They show that embedding matrix reside mostly in a low-dimensional subspace ... | 1 | 5 | 1,698,801,912,398 | Response to Reviewer DK5H (Part I)
We thank the reviewer for their comments. We have addressed the comments in the rebuttal below.
**Q1:** Scaling law under different data size.
We sincerely appreciate the reviewer's suggestion to evaluate the scalability of our model on larger datasets. However, it is important to n... | [{"id":"92DPxf1WlX","parent_id":"9kPJm1DFwm","created_at_ms":1700203560115,"modified_at_ms":1700203560115,"author":"Authors","role":"authors","type":"Author Reply","content":{"title":"Response to Reviewer DK5H (Part I)","comment":"We thank the reviewer for their comments. We have addressed the comments in the rebuttal ... | null | [] | ["Xingzhuo Guo","Junwei Pan","Ximei Wang","Baixu Chen","Jie Jiang","Mingsheng Long"] | ["Tsinghua University","Tencent","Tencent","Tsinghua University","Tencent AI Lab","Tsinghua University"] | 2310.044 | cs.LG cs.IR | 1,696,614,638,000 | 1 | unchanged |
0IaTFNJner | IlRLbALEdX | 2,024 | On the Embedding Collapse When Scaling up Recommendation Models | Recent advances in deep foundation models have led to a promising trend of developing large recommendation models to leverage vast amounts of available data. However, we experiment to scale up existing recommendation models and observe that the enlarged models do not improve satisfactorily. In this context, we investig... | ["machine learning","representation learning","recommender system","collapse"] | summary:
This paper suggests that the embedding collapse phenomenon restricts the scalability of existing recommendation models. Empirical and theoretical analysis show that interaction with collapsed embeddings constrains embedding learning. Also, this paper proposes a multi-embedding design incorporating embedding-se... | 1 | 5 | 1,698,822,478,538 | Response to Reviewer JdNi (Part I)
We thank the reviewer for their comments. We have addressed the comments in the rebuttal below.
**Q1:** Relation between multi-embedding and DMRL.
We appreciate the reviewer's question regarding the relationship between ME and DMRL, and we have incorporated it into our paper. It is ... | [{"id":"lA33HP1bCy","parent_id":"IlRLbALEdX","created_at_ms":1700203420843,"modified_at_ms":1700203432620,"author":"Authors","role":"authors","type":"Author Reply","content":{"title":"Response to Reviewer JdNi (Part I)","comment":"We thank the reviewer for their comments. We have addressed the comments in the rebuttal ... | null | [] | ["Xingzhuo Guo","Junwei Pan","Ximei Wang","Baixu Chen","Jie Jiang","Mingsheng Long"] | ["Tsinghua University","Tencent","Tencent","Tsinghua University","Tencent AI Lab","Tsinghua University"] | 2310.044 | cs.LG cs.IR | 1,696,614,638,000 | 1 | unchanged |
0IaTFNJner | Sb5jATn86P | 2,024 | On the Embedding Collapse When Scaling up Recommendation Models | Recent advances in deep foundation models have led to a promising trend of developing large recommendation models to leverage vast amounts of available data. However, we experiment to scale up existing recommendation models and observe that the enlarged models do not improve satisfactorily. In this context, we investig... | ["machine learning","representation learning","recommender system","collapse"] | summary:
This paper observes scaling the embedding dimensionality does not lead to satisfactory improvement in recommendation models and claims this is due to the dimensional collapse problem where the learned embedding vector spans only on lower dimensional subspace. The paper further claims such problem will propagat... | 1 | 5 | 1,698,736,644,154 | Response to Reviewer eFdv (Part I)
We thank the reviewer for their comments. We have addressed the comments in the rebuttal below.
**Q1:** How the information abundance for the setting of ME is computed.
We acknowledge the reviewer's concern regarding the computation of information abundance under the ME settings in ... | [{"id":"DomGkScB8g","parent_id":"Sb5jATn86P","created_at_ms":1700203666918,"modified_at_ms":1700203666918,"author":"Authors","role":"authors","type":"Author Reply","content":{"title":"Response to Reviewer eFdv (Part I)","comment":"We thank the reviewer for their comments. We have addressed the comments in the rebuttal ... | null | [] | ["Xingzhuo Guo","Junwei Pan","Ximei Wang","Baixu Chen","Jie Jiang","Mingsheng Long"] | ["Tsinghua University","Tencent","Tencent","Tsinghua University","Tencent AI Lab","Tsinghua University"] | 2310.044 | cs.LG cs.IR | 1,696,614,638,000 | 1 | changed |
0IaTFNJner | nWhcrWyPwG | 2,024 | On the Embedding Collapse When Scaling up Recommendation Models | Recent advances in deep foundation models have led to a promising trend of developing large recommendation models to leverage vast amounts of available data. However, we experiment to scale up existing recommendation models and observe that the enlarged models do not improve satisfactorily. In this context, we investig... | ["machine learning","representation learning","recommender system","collapse"] | summary:
This paper studies recommendation model performance when scaling up the embedding layers of the model. The paper identifies a phenomenon of embedding collapse, wherein the embedding matrix tends to reside in a low-dimensional subspace. Through empirical experiments on FFM and DCNv2 and theoretical analysis on ... | 1 | 3 | 1,698,829,848,014 | Response to Reviewer 9BGA (Part I)
We thank the reviewer for their comments. We have addressed the comments in the rebuttal below.
**Q1:** How the information abundance is a fair metric.
We appreciate the reviewer's question regarding the fairness of information abundance as a metric. Allow us to address this concern... | [{"id":"exaVvW3i7m","parent_id":"nWhcrWyPwG","created_at_ms":1700203284864,"modified_at_ms":1700203284864,"author":"Authors","role":"authors","type":"Author Reply","content":{"title":"Response to Reviewer 9BGA (Part I)","comment":"We thank the reviewer for their comments. We have addressed the comments in the rebuttal ... | null | [] | ["Xingzhuo Guo","Junwei Pan","Ximei Wang","Baixu Chen","Jie Jiang","Mingsheng Long"] | ["Tsinghua University","Tencent","Tencent","Tsinghua University","Tencent AI Lab","Tsinghua University"] | 2310.044 | cs.LG cs.IR | 1,696,614,638,000 | 1 | changed |
0JnaN0Crlz | 3PyGDlQ1o1 | 2,024 | Enhancing Adversarial Robustness on Categorical Data via Attribution Smoothing | Many efforts have been contributed to alleviate the adversarial risk of deep neural networks on continuous inputs.
Adversarial robustness on general categorical inputs, especially tabular categorical attributes, has received much less attention. To echo this challenge, our work aims to enhance the robustness of classif... | ["Adversarial Robustness","Categorical Data","Attribution Smoothing","Information Theory"] | summary:
This paper proposes IGSG, a theoretically motivated robust learning method for categorical inputs. It contains two parts: Integrated Gradient (IG) and Gradient Regularization. Experiments verify the effectiveness of the proposed method.
strengths:
- The studied problem is interesting, adversarial attack/defen... | 1 | 6 | 1,698,749,789,839 | Explanations of some Adversarial Training baseline and Regularization baseline methods performing worse than the undefended method.
Many thanks to the viewer for the insightful questions. In the response below, we first provide answers to these questions, and then would like to discuss more about the changes we made in... | [{"id":"jHfFuq4Pje","parent_id":"3PyGDlQ1o1","created_at_ms":1700435842222,"modified_at_ms":1710476874660,"author":"Authors","role":"authors","type":"Author Reply","content":{"title":"Explanations of some Adversarial Training baseline and Regularization baseline methods performing worse than the undefended method.","co... | Summary of the changes
We deeply appreciate all the constructive comments and suggestions from 6 reviewers. We have updated our submission to address the raised questions. The updated / added contents are highlighted in blue. Globally, our revision is summarized below:
We have updated Section 3.2, 4 and Appendix B (P... | [{"id":"yZ9XR0EX8Q","parent_id":"0JnaN0Crlz","created_at_ms":1700429760986,"modified_at_ms":1710476875393,"author":"Authors","role":"authors","type":"Author Reply","content":{"title":"Summary of the changes","comment":"We deeply appreciate all the constructive comments and suggestions from 6 reviewers. We have updated ... | ["Yujun Zhou","Yufei Han","Hongyan Bao","Xiangliang Zhang"] | ["University of Notre Dame","INRIA","KAUST","University of Notre Dame"] | null | null | null | 0 | unchanged |
0JnaN0Crlz | 9EkNt8c5rE | 2,024 | Enhancing Adversarial Robustness on Categorical Data via Attribution Smoothing | Many efforts have been contributed to alleviate the adversarial risk of deep neural networks on continuous inputs.
Adversarial robustness on general categorical inputs, especially tabular categorical attributes, has received much less attention. To echo this challenge, our work aims to enhance the robustness of classif... | ["Adversarial Robustness","Categorical Data","Attribution Smoothing","Information Theory"] | summary:
This paper studies how to strength the adversarial robustness of categorical data using attribution smoothing.
strengths:
- There have been only a few works aiming to enhance the adversarial robustness of categorical data.
- The empirical results look quite convincing.
- The use of IG for discrete data is i... | 1 | 3 | 1,698,877,218,937 | Change of Notations
Many thanks to the viewer for the insightful questions. In the response below, we first provide answers to these questions, and then would like to discuss more about the changes we made in the updated draft, thanks to the insightful questions from the reviewer.
**Notations**: Thank you for the a... | [{"id":"iVWezxXHuy","parent_id":"9EkNt8c5rE","created_at_ms":1700430746129,"modified_at_ms":1710476875380,"author":"Authors","role":"authors","type":"Author Reply","content":{"title":"Change of Notations","comment":"Many thanks to the viewer for the insightful questions. In the response below, we first provide answers ... | Summary of the changes
We deeply appreciate all the constructive comments and suggestions from 6 reviewers. We have updated our submission to address the raised questions. The updated / added contents are highlighted in blue. Globally, our revision is summarized below:
We have updated Section 3.2, 4 and Appendix B (P... | [{"id":"yZ9XR0EX8Q","parent_id":"0JnaN0Crlz","created_at_ms":1700429760986,"modified_at_ms":1710476875393,"author":"Authors","role":"authors","type":"Author Reply","content":{"title":"Summary of the changes","comment":"We deeply appreciate all the constructive comments and suggestions from 6 reviewers. We have updated ... | ["Yujun Zhou","Yufei Han","Hongyan Bao","Xiangliang Zhang"] | ["University of Notre Dame","INRIA","KAUST","University of Notre Dame"] | null | null | null | 0 | unchanged |
0JnaN0Crlz | LUdCXSBJov | 2,024 | Enhancing Adversarial Robustness on Categorical Data via Attribution Smoothing | Many efforts have been contributed to alleviate the adversarial risk of deep neural networks on continuous inputs.
Adversarial robustness on general categorical inputs, especially tabular categorical attributes, has received much less attention. To echo this challenge, our work aims to enhance the robustness of classif... | ["Adversarial Robustness","Categorical Data","Attribution Smoothing","Information Theory"] | summary:
The authors address adversarial robustness in the context of categorical variables, a less-explored scenario in the literature. They highlight the key finding that different attack methods yield adversarial samples with varied distributions. Relying solely on PGD-generated adversarial samples for training lead... | 1 | 6 | 1,698,804,666,375 | Ambiguities and the settings of cross-validation of hyperparameters
Many thanks to the viewer for the insightful questions. In the response below, we first provide answers to these questions, and then would like to discuss more about the changes we made in the updated draft, thanks to the insightful questions from the ... | [{"id":"CwKeMWHQlD","parent_id":"LUdCXSBJov","created_at_ms":1700435426252,"modified_at_ms":1710476874901,"author":"Authors","role":"authors","type":"Author Reply","content":{"title":"Ambiguities and the settings of cross-validation of hyperparameters","comment":"Many thanks to the viewer for the insightful questions. ... | Summary of the changes
We deeply appreciate all the constructive comments and suggestions from 6 reviewers. We have updated our submission to address the raised questions. The updated / added contents are highlighted in blue. Globally, our revision is summarized below:
We have updated Section 3.2, 4 and Appendix B (P... | [{"id":"yZ9XR0EX8Q","parent_id":"0JnaN0Crlz","created_at_ms":1700429760986,"modified_at_ms":1710476875393,"author":"Authors","role":"authors","type":"Author Reply","content":{"title":"Summary of the changes","comment":"We deeply appreciate all the constructive comments and suggestions from 6 reviewers. We have updated ... | ["Yujun Zhou","Yufei Han","Hongyan Bao","Xiangliang Zhang"] | ["University of Notre Dame","INRIA","KAUST","University of Notre Dame"] | null | null | null | 0 | unchanged |
0JnaN0Crlz | WqyX2wv9YP | 2,024 | Enhancing Adversarial Robustness on Categorical Data via Attribution Smoothing | Many efforts have been contributed to alleviate the adversarial risk of deep neural networks on continuous inputs.
Adversarial robustness on general categorical inputs, especially tabular categorical attributes, has received much less attention. To echo this challenge, our work aims to enhance the robustness of classif... | ["Adversarial Robustness","Categorical Data","Attribution Smoothing","Information Theory"] | summary:
This paper studies adversarial robustness of classification over categorical attributes against adversarial perturbation. An information-theoretic upper bound on the expected adversarial risk has been established, together with an adversarial learning method with integrated gradient-smoothed gradient regulariz... | 1 | 5 | 1,698,849,518,190 | Disconnection between the theoretical analysis and the design of the algorithm.
Many thanks to the viewer for the insightful questions. In the response below, we first provide answers to these questions, and then would like to discuss more about the changes we made in the updated draft, thanks to the insightful questio... | [{"id":"ynaD2zVyAa","parent_id":"WqyX2wv9YP","created_at_ms":1700434530822,"modified_at_ms":1710476875066,"author":"Authors","role":"authors","type":"Author Reply","content":{"title":"Disconnection between the theoretical analysis and the design of the algorithm.","comment":"Many thanks to the viewer for the insightful... | Summary of the changes
We deeply appreciate all the constructive comments and suggestions from 6 reviewers. We have updated our submission to address the raised questions. The updated / added contents are highlighted in blue. Globally, our revision is summarized below:
We have updated Section 3.2, 4 and Appendix B (P... | [{"id":"yZ9XR0EX8Q","parent_id":"0JnaN0Crlz","created_at_ms":1700429760986,"modified_at_ms":1710476875393,"author":"Authors","role":"authors","type":"Author Reply","content":{"title":"Summary of the changes","comment":"We deeply appreciate all the constructive comments and suggestions from 6 reviewers. We have updated ... | ["Yujun Zhou","Yufei Han","Hongyan Bao","Xiangliang Zhang"] | ["University of Notre Dame","INRIA","KAUST","University of Notre Dame"] | null | null | null | 0 | unchanged |
0JnaN0Crlz | sNCki9nNS4 | 2,024 | Enhancing Adversarial Robustness on Categorical Data via Attribution Smoothing | Many efforts have been contributed to alleviate the adversarial risk of deep neural networks on continuous inputs.
Adversarial robustness on general categorical inputs, especially tabular categorical attributes, has received much less attention. To echo this challenge, our work aims to enhance the robustness of classif... | ["Adversarial Robustness","Categorical Data","Attribution Smoothing","Information Theory"] | summary:
This work studies adversarial robustness on tabular categorical attributes and establishes an information-theoretic upper bound on the expected adversarial risk.
strengths:
- Provide some theoretical results.
- Easy to follow.
weaknesses:
- This work develops the theory over randomized learning algorithm $A(... | 1 | 5 | 1,698,756,156,353 | How does the assumption of randomized learning apply to the experiments?
Many thanks to the viewer for the insightful questions. In the response below, we first provide answers to these questions, and then would like to discuss more about the changes we made in the updated draft, thanks to the insightful questions from... | [{"id":"2rnpHzSpXR","parent_id":"sNCki9nNS4","created_at_ms":1700435555637,"modified_at_ms":1710476874821,"author":"Authors","role":"authors","type":"Author Reply","content":{"title":"How does the assumption of randomized learning apply to the experiments?","comment":"Many thanks to the viewer for the insightful questi... | Summary of the changes
We deeply appreciate all the constructive comments and suggestions from 6 reviewers. We have updated our submission to address the raised questions. The updated / added contents are highlighted in blue. Globally, our revision is summarized below:
We have updated Section 3.2, 4 and Appendix B (P... | [{"id":"yZ9XR0EX8Q","parent_id":"0JnaN0Crlz","created_at_ms":1700429760986,"modified_at_ms":1710476875393,"author":"Authors","role":"authors","type":"Author Reply","content":{"title":"Summary of the changes","comment":"We deeply appreciate all the constructive comments and suggestions from 6 reviewers. We have updated ... | ["Yujun Zhou","Yufei Han","Hongyan Bao","Xiangliang Zhang"] | ["University of Notre Dame","INRIA","KAUST","University of Notre Dame"] | null | null | null | 0 | unchanged |
0JnaN0Crlz | xZAhi8IH0z | 2,024 | Enhancing Adversarial Robustness on Categorical Data via Attribution Smoothing | Many efforts have been contributed to alleviate the adversarial risk of deep neural networks on continuous inputs.
Adversarial robustness on general categorical inputs, especially tabular categorical attributes, has received much less attention. To echo this challenge, our work aims to enhance the robustness of classif... | ["Adversarial Robustness","Categorical Data","Attribution Smoothing","Information Theory"] | summary:
This paper proposes a regularization method to enhance the robustness of classification over categorical attributes against adversarial perturbations. The paper establishes an information-theoretic upper bound on the expected adversarial risk and proposes an adversarially robust learning method, named Integrat... | 1 | 6 | 1,698,648,296,379 | Visualizations of the classification boundary
Many thanks to the viewer for the insightful questions. In the response below, we first provide answers to these questions, and then would like to discuss more about the changes we made in the updated draft, thanks to the insightful questions from the reviewer.
**Visual... | [{"id":"ADP05OCDkh","parent_id":"xZAhi8IH0z","created_at_ms":1700436081401,"modified_at_ms":1710476874372,"author":"Authors","role":"authors","type":"Author Reply","content":{"title":"Visualizations of the classification boundary","comment":"Many thanks to the viewer for the insightful questions. In the response below,... | Summary of the changes
We deeply appreciate all the constructive comments and suggestions from 6 reviewers. We have updated our submission to address the raised questions. The updated / added contents are highlighted in blue. Globally, our revision is summarized below:
We have updated Section 3.2, 4 and Appendix B (P... | [{"id":"yZ9XR0EX8Q","parent_id":"0JnaN0Crlz","created_at_ms":1700429760986,"modified_at_ms":1710476875393,"author":"Authors","role":"authors","type":"Author Reply","content":{"title":"Summary of the changes","comment":"We deeply appreciate all the constructive comments and suggestions from 6 reviewers. We have updated ... | ["Yujun Zhou","Yufei Han","Hongyan Bao","Xiangliang Zhang"] | ["University of Notre Dame","INRIA","KAUST","University of Notre Dame"] | null | null | null | 0 | unchanged |
0JsRZEGZ7L | 7ylg8ZyQ99 | 2,024 | From Latent Graph to Latent Topology Inference: Differentiable Cell Complex Module | Latent Graph Inference (LGI) relaxed the reliance of Graph Neural Networks (GNNs) on a given graph topology by dynamically learning it. However, most of LGI methods assume to have a (noisy, incomplete, improvable, ...) input graph to rewire and can solely learn regular graph topologies. In the wake of the success of T... | ["Topological Deep Learning","Geometric Deep Learning","Latent Topology Inference","Latent Graph Inference","Cell Complexes"] | summary:
The authors introduce a novel Latent Topology Inference (LTI) method that enables the learning of non-regular topologies based on their novel Differentiable Cell Complex Module (DCM). DCM is designed to compute cell probabilities within the complex, thus enhancing downstream tasks. They show how to integrate D... | 1 | 5 | 1,697,731,841,280 | ## Weaknesses
1. In our paper, DGM refers to the original Differentiable Graph Module of Kazi et al. (2022). Section 3.1 introduces $\alpha$-DGM, which is our prosed variation of DGM that exploits the entmax function instead of Gumbel-Softmax to sample a potentially irregular graph topology. $\alpha$-DGM is the varian... | [{"id":"nFLjfNuVT9","parent_id":"7ylg8ZyQ99","created_at_ms":1700258710787,"modified_at_ms":1700258710787,"author":"Authors","role":"authors","type":"Author Reply","content":{"comment":"## Weaknesses\n\n1. In our paper, DGM refers to the original Differentiable Graph Module of Kazi et al. (2022). Section 3.1 introduces... | We would like to thank all reviewers for their evaluation of the paper. We are glad they agree on the novelty of the idea, the extensive evaluation we provide, and our analysis of the limitations (Reviewers XizW, EucF). A point-by-point reply to all the comments is individually given in the sequel. We reply to comment... | [{"id":"nkPgVGa42i","parent_id":"0JsRZEGZ7L","created_at_ms":1700258589905,"modified_at_ms":1700258589905,"author":"Authors","role":"authors","type":"Author Reply","content":{"comment":"We would like to thank all reviewers for their evaluation of the paper. We are glad they agree on the novelty of the idea, the extensi... | ["Claudio Battiloro","Indro Spinelli","Lev Telyatnikov","Michael M. Bronstein","Simone Scardapane","Paolo Di Lorenzo"] | ["University of Roma \"La Sapienza\"","Sapienza University of Rome","University of Roma \"La Sapienza\"","University of Oxford","Sapienza University of Rome","University of Roma \"La Sapienza\""] | 2305.16174 | cs.LG cs.AI cs.NE | 1,685,028,799,000 | 1 | changed |
0JsRZEGZ7L | m5nqCSO4Ui | 2,024 | From Latent Graph to Latent Topology Inference: Differentiable Cell Complex Module | Latent Graph Inference (LGI) relaxed the reliance of Graph Neural Networks (GNNs) on a given graph topology by dynamically learning it. However, most of LGI methods assume to have a (noisy, incomplete, improvable, ...) input graph to rewire and can solely learn regular graph topologies. In the wake of the success of T... | ["Topological Deep Learning","Geometric Deep Learning","Latent Topology Inference","Latent Graph Inference","Cell Complexes"] | summary:
This paper deals with the learning of high-order cell complexes with sparse and not regular graph topology.
The key is to introduce a learnable function that computes cell probabilities in the complex and integrate with cell complex message-passing network layers in a scalable way. The model achieved improve ... | 1 | 5 | 1,699,116,800,246 | ## Weaknesses
1. It is true that, on the homophilic datasets, the results of the DCM with provided graph are slightly worse than the presented well-known graph-based and complex-based architectures. We argue this is an expected behavior, because those architectures inherently rely on the homophily assumption, i.e. th... | [{"id":"ZR3JvtkFzP","parent_id":"m5nqCSO4Ui","created_at_ms":1700258651979,"modified_at_ms":1700258651979,"author":"Authors","role":"authors","type":"Author Reply","content":{"comment":"## Weaknesses\n\n 1. It is true that, on the homophilic datasets, the results of the DCM with provided graph are slightly worse than t... | We would like to thank all reviewers for their evaluation of the paper. We are glad they agree on the novelty of the idea, the extensive evaluation we provide, and our analysis of the limitations (Reviewers XizW, EucF). A point-by-point reply to all the comments is individually given in the sequel. We reply to comment... | [{"id":"nkPgVGa42i","parent_id":"0JsRZEGZ7L","created_at_ms":1700258589905,"modified_at_ms":1700258589905,"author":"Authors","role":"authors","type":"Author Reply","content":{"comment":"We would like to thank all reviewers for their evaluation of the paper. We are glad they agree on the novelty of the idea, the extensi... | ["Claudio Battiloro","Indro Spinelli","Lev Telyatnikov","Michael M. Bronstein","Simone Scardapane","Paolo Di Lorenzo"] | ["University of Roma \"La Sapienza\"","Sapienza University of Rome","University of Roma \"La Sapienza\"","University of Oxford","Sapienza University of Rome","University of Roma \"La Sapienza\""] | 2305.16174 | cs.LG cs.AI cs.NE | 1,685,028,799,000 | 1 | changed |
0JsRZEGZ7L | t2NhhNjcX2 | 2,024 | From Latent Graph to Latent Topology Inference: Differentiable Cell Complex Module | Latent Graph Inference (LGI) relaxed the reliance of Graph Neural Networks (GNNs) on a given graph topology by dynamically learning it. However, most of LGI methods assume to have a (noisy, incomplete, improvable, ...) input graph to rewire and can solely learn regular graph topologies. In the wake of the success of T... | ["Topological Deep Learning","Geometric Deep Learning","Latent Topology Inference","Latent Graph Inference","Cell Complexes"] | summary:
This paper proposes a method to learn "latent topology" in ML tasks, extending Latent Graph Inference. The idea is to learn a cell complex during training. Their method extends the Differentiable Graph Module (DGM) (Kazi 2022) to Diff. Cell Module (DCM). They also use $\alpha$-maxent instead of gumball softmax... | 1 | 8 | 1,698,786,631,633 | ## Weaknesses
1. (Includes also the reply to Question 2.) Eq. (6) can be used for polygons of arbitrary size $k$. We initially presented it for the triangle case ($k=3$) just to provide a simple example. We made the choice to decompose the “$k$-th order” similarity as a sum of pairwise similarities to enhance the scal... | [{"id":"BBoTKTO8kS","parent_id":"t2NhhNjcX2","created_at_ms":1700258668492,"modified_at_ms":1700258668492,"author":"Authors","role":"authors","type":"Author Reply","content":{"comment":"## Weaknesses\n\n1. (Includes also the reply to Question 2.) Eq. (6) can be used for polygons of arbitrary size $k$. We initially pres... | We would like to thank all reviewers for their evaluation of the paper. We are glad they agree on the novelty of the idea, the extensive evaluation we provide, and our analysis of the limitations (Reviewers XizW, EucF). A point-by-point reply to all the comments is individually given in the sequel. We reply to comment... | [{"id":"nkPgVGa42i","parent_id":"0JsRZEGZ7L","created_at_ms":1700258589905,"modified_at_ms":1700258589905,"author":"Authors","role":"authors","type":"Author Reply","content":{"comment":"We would like to thank all reviewers for their evaluation of the paper. We are glad they agree on the novelty of the idea, the extensi... | ["Claudio Battiloro","Indro Spinelli","Lev Telyatnikov","Michael M. Bronstein","Simone Scardapane","Paolo Di Lorenzo"] | ["University of Roma \"La Sapienza\"","Sapienza University of Rome","University of Roma \"La Sapienza\"","University of Oxford","Sapienza University of Rome","University of Roma \"La Sapienza\""] | 2305.16174 | cs.LG cs.AI cs.NE | 1,685,028,799,000 | 1 | unchanged |
0KVkTDB6KZ | QjXHBWr4ve | 2,024 | EFFL: Egalitarian Fairness in Federated Learning for Mitigating Matthew Effect | Recent advances in federated learning (FL) enable collaborative training of machine learning (ML) models from large-scale and widely dispersed clients while protecting their privacy. However, when different clients' datasets are heterogeneous, traditional FL mechanisms produce a global model that does not adequately re... | ["Egalitarian Fairness","Federated Learning","AI Ethics"] | summary:
This paper investigates fairness in FL on both accuracy and decision bias. The authors propose a three-stage optimization method to optimize the proposed constrained multi-objective optimization problem. Numerical results show the benefits of the algorithm in terms of fairness.
strengths:
1. Considering both ... | 1 | 3 | 1,698,759,343,503 | Response to Reviewer YyVq
We appreciate the reviewer's constructive comments.
1. **To the Weakness 1**
We'd like to clarify that our primary focus lies in introducing the crucial concept of achieving egalitarian fairness within FL to mitigate the Matthew effect, emphasizing social ethics. Our proposed training m... | [{"id":"ScjzRyYTD2","parent_id":"QjXHBWr4ve","created_at_ms":1700134462691,"modified_at_ms":1700134462691,"author":"Authors","role":"authors","type":"Author Reply","content":{"title":"Response to Reviewer YyVq","comment":"We appreciate the reviewer's constructive comments.\n\n1. **To the Weakness 1**\n\n We'd like ... | Thank you!
Thank you very much for the reviewers's efforts and valuable comments. We have provided the point-to-point response for each review in detail. Should you have any further questions, do not hesitate to contact us.
Greatly appreciate your help!
We would like to thank all the reviewers for their thoughtful re... | [{"id":"QqlbpqzVln","parent_id":"0KVkTDB6KZ","created_at_ms":1700283066512,"modified_at_ms":1710476874039,"author":"Authors","role":"authors","type":"Author Reply","content":{"title":"Thank you!","comment":"Thank you very much for the reviewers's efforts and valuable comments. We have provided the point-to-point respon... | ["Jiashi Gao","Changwu Huang","Ming Tang","Shin Hwei Tan","Xin Yao","Xuetao Wei"] | ["Southern University of Science and Technology",null,"Southern University of Science and Technology","Concordia University","University of Technology Sydney","Southern University of Science and Technology"] | null | null | null | 0 | unchanged |
0KVkTDB6KZ | TeG4nwqPlm | 2,024 | EFFL: Egalitarian Fairness in Federated Learning for Mitigating Matthew Effect | Recent advances in federated learning (FL) enable collaborative training of machine learning (ML) models from large-scale and widely dispersed clients while protecting their privacy. However, when different clients' datasets are heterogeneous, traditional FL mechanisms produce a global model that does not adequately re... | ["Egalitarian Fairness","Federated Learning","AI Ethics"] | summary:
This work proposes EFFL a client fairness approach that aims to mitigate the Mathew effect by producing a Pareto optimal model with equal decision bias and accuracy across the participating clients. EFFL problem is formulated as a MOOP with decision bias and fairness constraints. The authors provide a 3-step a... | 1 | 3 | 1,699,561,150,503 | Response to Reviewer ta1u
We appreciate the reviewer's constructive comments. Our responses to the comments are listed as follows:
1. **To the Weakness 1**
We'd like to clarify the convergence of the proposed algorithm as follows:
**Proposition.** For $N$ optimization objectives $l_1(\theta^t),...,l_N(\thet... | [{"id":"4doKfe87ZK","parent_id":"TeG4nwqPlm","created_at_ms":1700131371048,"modified_at_ms":1700136243019,"author":"Authors","role":"authors","type":"Author Reply","content":{"title":"Response to Reviewer ta1u","comment":"We appreciate the reviewer's constructive comments. Our responses to the comments are listed as fo... | Thank you!
Thank you very much for the reviewers's efforts and valuable comments. We have provided the point-to-point response for each review in detail. Should you have any further questions, do not hesitate to contact us. | [{"id":"QqlbpqzVln","parent_id":"0KVkTDB6KZ","created_at_ms":1700283066512,"modified_at_ms":1710476874039,"author":"Authors","role":"authors","type":"Author Reply","content":{"title":"Thank you!","comment":"Thank you very much for the reviewers's efforts and valuable comments. We have provided the point-to-point respon... | ["Jiashi Gao","Changwu Huang","Ming Tang","Shin Hwei Tan","Xin Yao","Xuetao Wei"] | ["Southern University of Science and Technology",null,"Southern University of Science and Technology","Concordia University","University of Technology Sydney","Southern University of Science and Technology"] | null | null | null | 0 | changed |
0KVkTDB6KZ | j80sboFt7r | 2,024 | EFFL: Egalitarian Fairness in Federated Learning for Mitigating Matthew Effect | Recent advances in federated learning (FL) enable collaborative training of machine learning (ML) models from large-scale and widely dispersed clients while protecting their privacy. However, when different clients' datasets are heterogeneous, traditional FL mechanisms produce a global model that does not adequately re... | ["Egalitarian Fairness","Federated Learning","AI Ethics"] | summary:
This paper examines the problem of fairness in a federated learning setting through the lens of MCMOO. The goal is to reduce global loss while satisfying several constraints: (1) Roughly equal loss/accuracy for each individual client; (2) Fairness at each individual client, i.e., low local unfairness; and (3) ... | 1 | 8 | 1,698,802,341,849 | Response to Reviewer Raor
We appreciate the reviewer for acknowledging the contributions of our work and valuable comments.
1. **To the Weakness 1**
Given gradients from the clients have an error $\tilde G (\theta^t)=G (\theta^t)+\mathbf{e} ^t$, then $\left \| \tilde{\mathbf{\alpha} } -\mathbf{\alpha} \right \|... | [{"id":"wzFhwLSxGd","parent_id":"j80sboFt7r","created_at_ms":1700133368941,"modified_at_ms":1700136547106,"author":"Authors","role":"authors","type":"Author Reply","content":{"title":"Response to Reviewer Raor","comment":"We appreciate the reviewer for acknowledging the contributions of our work and valuable comments.... | Thank you!
Thank you very much for the reviewers's efforts and valuable comments. We have provided the point-to-point response for each review in detail. Should you have any further questions, do not hesitate to contact us.
Greatly appreciate your help!
We would like to thank all the reviewers for their thoughtful re... | [{"id":"QqlbpqzVln","parent_id":"0KVkTDB6KZ","created_at_ms":1700283066512,"modified_at_ms":1710476874039,"author":"Authors","role":"authors","type":"Author Reply","content":{"title":"Thank you!","comment":"Thank you very much for the reviewers's efforts and valuable comments. We have provided the point-to-point respon... | ["Jiashi Gao","Changwu Huang","Ming Tang","Shin Hwei Tan","Xin Yao","Xuetao Wei"] | ["Southern University of Science and Technology",null,"Southern University of Science and Technology","Concordia University","University of Technology Sydney","Southern University of Science and Technology"] | null | null | null | 0 | unchanged |
0KVkTDB6KZ | sMFMMciBNg | 2,024 | EFFL: Egalitarian Fairness in Federated Learning for Mitigating Matthew Effect | Recent advances in federated learning (FL) enable collaborative training of machine learning (ML) models from large-scale and widely dispersed clients while protecting their privacy. However, when different clients' datasets are heterogeneous, traditional FL mechanisms produce a global model that does not adequately re... | ["Egalitarian Fairness","Federated Learning","AI Ethics"] | summary:
This paper studies fairness in federated learning with. Specifically, it considers two types of fairness. The first one is the Matthew effect, which considers the performance across different clients; and the second one is the decision bias, which is a class of commonly studied fairness definitions like accura... | 1 | 5 | 1,698,988,060,505 | Response to Reviewer hFqL
We appreciate the reviewer's constructive comments. Our responses to the comments are listed as follows:
1. **To the W1**
We'd like to explain that the motivation for two fairness is that clients may contribute varying amounts due to inherent resource inequalities. Ignoring such inequit... | [{"id":"eXTu5FFXQn","parent_id":"sMFMMciBNg","created_at_ms":1700132120568,"modified_at_ms":1700136272086,"author":"Authors","role":"authors","type":"Author Reply","content":{"title":"Response to Reviewer hFqL","comment":"We appreciate the reviewer's constructive comments. Our responses to the comments are listed as fo... | Thank you!
Thank you very much for the reviewers's efforts and valuable comments. We have provided the point-to-point response for each review in detail. Should you have any further questions, do not hesitate to contact us. | [{"id":"QqlbpqzVln","parent_id":"0KVkTDB6KZ","created_at_ms":1700283066512,"modified_at_ms":1710476874039,"author":"Authors","role":"authors","type":"Author Reply","content":{"title":"Thank you!","comment":"Thank you very much for the reviewers's efforts and valuable comments. We have provided the point-to-point respon... | ["Jiashi Gao","Changwu Huang","Ming Tang","Shin Hwei Tan","Xin Yao","Xuetao Wei"] | ["Southern University of Science and Technology",null,"Southern University of Science and Technology","Concordia University","University of Technology Sydney","Southern University of Science and Technology"] | null | null | null | 0 | changed |
0Lqyut1y7M | KIaaUrXoVF | 2,024 | On the Optimality of Activations in Implicit Neural Representations | Implicit neural representations (INRs) have recently surged in popularity as a class of neural networks capable of encoding signals as compact, differentiable entities. To capture high-frequency content, INRs often employ techniques such as Fourier positional encodings or non-traditional activation functions like Gauss... | ["Implicit Neural Representations","Sampling theory"] | summary:
This paper provides a detailed analysis of approximation properties of implicit neural representations (INR) by deploying tools from harmonic analysis. Specifically, while a variety of activation functions have been investigated for use within INR models, a unified treatment is missing and comparisons appear t... | 1 | 6 | 1,698,811,459,654 | Response to reviewer hcEo
We thank the reviewer for the valuable suggestions. Please find our answers below.
**Not a major weakness but the use of orthogonal polynomials (such as Hermite) has been proposed...**
This is a very reasonable question from the reviewer and we thank them for asking it. It is true that pract... | [{"id":"zUXNuMTJLr","parent_id":"KIaaUrXoVF","created_at_ms":1700037019598,"modified_at_ms":1700037019598,"author":"Authors","role":"authors","type":"Author Reply","content":{"title":"Response to reviewer hcEo","comment":"We thank the reviewer for the valuable suggestions. Please find our answers below.\n\n**Not a majo... | null | [] | ["Sameera Ramasinghe","Hemanth Saratchandran","Violetta Shevchenko","Alexander Long","Simon Lucey"] | ["Amazon",null,"Amazon","Amazon","University of Adelaide"] | null | null | null | 0 | unchanged |
0Lqyut1y7M | gTWSurBkQK | 2,024 | On the Optimality of Activations in Implicit Neural Representations | Implicit neural representations (INRs) have recently surged in popularity as a class of neural networks capable of encoding signals as compact, differentiable entities. To capture high-frequency content, INRs often employ techniques such as Fourier positional encodings or non-traditional activation functions like Gauss... | ["Implicit Neural Representations","Sampling theory"] | summary:
The paper views INR fitting through the lens of sampling theory and shows that the sinc activation function is particularly amenable due to two properties: shifted sinc functions form a linearly independent basis in the space of bandlimited functions, and they also form a partition of unity in this space. The ... | 1 | 6 | 1,697,260,203,044 | Response to Reviewer 8ETH
**My key concern is that the paper is hard to read, especially for many potential users of this work (e.g. NeRF practitioners). When the paper involves this much theory, it is particularly important for the presentation to be clear and to offer graphics / intuition when possible....**
We than... | [{"id":"svaZZzmVKz","parent_id":"gTWSurBkQK","created_at_ms":1700037904463,"modified_at_ms":1700037904463,"author":"Authors","role":"authors","type":"Author Reply","content":{"title":"Response to Reviewer 8ETH","comment":"**My key concern is that the paper is hard to read, especially for many potential users of this wo... | null | [] | ["Sameera Ramasinghe","Hemanth Saratchandran","Violetta Shevchenko","Alexander Long","Simon Lucey"] | ["Amazon",null,"Amazon","Amazon","University of Adelaide"] | null | null | null | 0 | unchanged |
0Lqyut1y7M | w6hn9qOpfN | 2,024 | On the Optimality of Activations in Implicit Neural Representations | Implicit neural representations (INRs) have recently surged in popularity as a class of neural networks capable of encoding signals as compact, differentiable entities. To capture high-frequency content, INRs often employ techniques such as Fourier positional encodings or non-traditional activation functions like Gauss... | ["Implicit Neural Representations","Sampling theory"] | summary:
The authors study neural network activation functions from the perspective of Shannon-Nyquist sampling theory. This study is particularly relevant to implicit neural representations, which are usually overfitted to regularly sampled signals, e.g., the coordinate-intensity pairs of an image.
The authors prove... | 1 | 3 | 1,698,845,271,489 | Response to reviewer REJk
We thank the reviewer for carefully going through our paper and providing suggestions to improve our current version. We have addressed all your comments. Please find detailed answers below. Please note that the Theorem and proposition numbers may have changed in the revised version as we have... | [{"id":"IAkakXaHFa","parent_id":"w6hn9qOpfN","created_at_ms":1700036121259,"modified_at_ms":1700083369895,"author":"Authors","role":"authors","type":"Author Reply","content":{"title":"Response to reviewer REJk","comment":"We thank the reviewer for carefully going through our paper and providing suggestions to improve o... | null | [] | ["Sameera Ramasinghe","Hemanth Saratchandran","Violetta Shevchenko","Alexander Long","Simon Lucey"] | ["Amazon",null,"Amazon","Amazon","University of Adelaide"] | null | null | null | 0 | changed |
0Lqyut1y7M | ymg6cKdV2m | 2,024 | On the Optimality of Activations in Implicit Neural Representations | Implicit neural representations (INRs) have recently surged in popularity as a class of neural networks capable of encoding signals as compact, differentiable entities. To capture high-frequency content, INRs often employ techniques such as Fourier positional encodings or non-traditional activation functions like Gauss... | ["Implicit Neural Representations","Sampling theory"] | summary:
In this paper, the authors establish a connection between sampling theory and implicit neural representations, advocating for the sinc function as a better activation function with theoretical backing to minimize reconstruction error to any desired positive threshold. By demonstrating how shifted sinc function... | 1 | 6 | 1,699,151,476,553 | Response for the reviewer 9zZJ
We would like to thank the reviewer for their valuable comments for improving our paper. Please find our answers below.
**Clarification of Notation: Regarding the mathematical notation, it would enhance clarity to emphasize the argument 'x' in Equation 3...**
Thank you for the suggestio... | [{"id":"uMF38YTj1k","parent_id":"ymg6cKdV2m","created_at_ms":1700034428870,"modified_at_ms":1700034428870,"author":"Authors","role":"authors","type":"Author Reply","content":{"title":"Response for the reviewer 9zZJ","comment":"We would like to thank the reviewer for their valuable comments for improving our paper. Plea... | null | [] | ["Sameera Ramasinghe","Hemanth Saratchandran","Violetta Shevchenko","Alexander Long","Simon Lucey"] | ["Amazon",null,"Amazon","Amazon","University of Adelaide"] | null | null | null | 0 | unchanged |
0NruoU6s5Z | 2G9rXvMSCb | 2,024 | CompoDiff: Versatile Composed Image Retrieval With Latent Diffusion | This paper proposes a novel diffusion-based model, CompoDiff, for solving Composed Image Retrieval (CIR) with latent diffusion and presents a newly created dataset, named SynthTriplets18M, of 18 million reference images, conditions, and corresponding target image triplets to train the model. CompoDiff and SynthTriplets... | ["Composed Image Retrieval","Diffusion Models"] | summary:
The authors introduces CompoDiff, a novel diffusion-based model, for the task of Composed Image Retrieval (CIR) with latent diffusion. It also proposes a new dataset named SynthTriplets18M. Importantly, it supports diverse conditions like negative text and image masks, offers control over query importance, and... | 1 | 5 | 1,698,756,100,281 | We thank the reviewer for their positive feedback and constructive comments. We will address all the raised concerns by the reviewer, and will revise our paper as soon as possible (Hopefully before Tuesday).
**[W1] More analytical reasoning of the CIR benchmark performance gaps is encouraged**
Thanks for the comment!... | [{"id":"JCMvpZJt1Z","parent_id":"2G9rXvMSCb","created_at_ms":1700487807611,"modified_at_ms":1700487807611,"author":"Authors","role":"authors","type":"Author Reply","content":{"comment":"We thank the reviewer for their positive feedback and constructive comments. We will address all the raised concerns by the reviewer, ... | Revision summary
We truly appreciate all the reviewers and the chairs for their commitment to the community and thoughtful reviews. We are highly encouraged that the reviewers found that the motivation and direction of our method are interesting and can contribute to the community (Reviewer AJF4, wkZo), our work bridge... | [{"id":"nnvEKpNejH","parent_id":"0NruoU6s5Z","created_at_ms":1700667751331,"modified_at_ms":1700667751331,"author":"Authors","role":"authors","type":"Author Reply","content":{"title":"Revision summary","comment":"We truly appreciate all the reviewers and the chairs for their commitment to the community and thoughtful r... | ["Geonmo Gu","Sanghyuk Chun","Wonjae Kim","HeeJae Jun","Yoohoon Kang","Sangdoo Yun"] | ["NAVER","NAVER AI Lab","NAVER","NAVER","NAVER","NAVER"] | 2303.11916 | cs.CV cs.IR | 1,679,411,195,000 | 1 | changed |
0NruoU6s5Z | H7Mc3Fsmd2 | 2,024 | CompoDiff: Versatile Composed Image Retrieval With Latent Diffusion | This paper proposes a novel diffusion-based model, CompoDiff, for solving Composed Image Retrieval (CIR) with latent diffusion and presents a newly created dataset, named SynthTriplets18M, of 18 million reference images, conditions, and corresponding target image triplets to train the model. CompoDiff and SynthTriplets... | ["Composed Image Retrieval","Diffusion Models"] | summary:
This paper addresses the limitation of the small dataset and small categories of conditions on the Composed Image Retrieval (CIR) task through a new diffusion-based model (CompoDiff) and a large-scale CIR dataset (SynthTriplets18M). They show the generalizability of their dataset training CompoDiff on the exis... | 1 | 5 | 1,699,257,745,214 | We thank the reviewer for their positive feedback and constructive comments. We will address all the concerns raised by the reviewer, and will revise our paper as soon as possible (Hopefully before Tuesday).
**[W1] Comparison with the same backbone**
Thanks for the comment. We would like to emphasize that CompoDiff h... | [{"id":"3sF1FbNcgE","parent_id":"H7Mc3Fsmd2","created_at_ms":1700486594218,"modified_at_ms":1700660397022,"author":"Authors","role":"authors","type":"Author Reply","content":{"comment":"We thank the reviewer for their positive feedback and constructive comments. We will address all the concerns raised by the reviewer, ... | Revision summary
We truly appreciate all the reviewers and the chairs for their commitment to the community and thoughtful reviews. We are highly encouraged that the reviewers found that the motivation and direction of our method are interesting and can contribute to the community (Reviewer AJF4, wkZo), our work bridge... | [{"id":"nnvEKpNejH","parent_id":"0NruoU6s5Z","created_at_ms":1700667751331,"modified_at_ms":1700667751331,"author":"Authors","role":"authors","type":"Author Reply","content":{"title":"Revision summary","comment":"We truly appreciate all the reviewers and the chairs for their commitment to the community and thoughtful r... | ["Geonmo Gu","Sanghyuk Chun","Wonjae Kim","HeeJae Jun","Yoohoon Kang","Sangdoo Yun"] | ["NAVER","NAVER AI Lab","NAVER","NAVER","NAVER","NAVER"] | 2303.11916 | cs.CV cs.IR | 1,679,411,195,000 | 1 | unchanged |
0NruoU6s5Z | il7wWeDi6P | 2,024 | CompoDiff: Versatile Composed Image Retrieval With Latent Diffusion | This paper proposes a novel diffusion-based model, CompoDiff, for solving Composed Image Retrieval (CIR) with latent diffusion and presents a newly created dataset, named SynthTriplets18M, of 18 million reference images, conditions, and corresponding target image triplets to train the model. CompoDiff and SynthTriplets... | ["Composed Image Retrieval","Diffusion Models"] | summary:
This paper proposes a novel diffusion-based model, named CompoDiff, which could merge the multimodal conditional information, for solving composed image retrieval (CIR) task. It also proposes a newly created synthetic dataset, named SynthTriplet18M, of 18 million training triplets (reference image, conditions,... | 1 | 5 | 1,698,671,129,617 | We thank the reviewer for their positive feedback and constructive comments. We will address all the raised concerns by the reviewer, and will revise our paper as soon as possible (Hopefully before Tuesday).
**[W1] Meaning of diffusion for feature extraction**
We would like to emphasize that the diffusion process its... | [{"id":"ecSM9Zjqwj","parent_id":"il7wWeDi6P","created_at_ms":1700488788534,"modified_at_ms":1700666034599,"author":"Authors","role":"authors","type":"Author Reply","content":{"comment":"We thank the reviewer for their positive feedback and constructive comments. We will address all the raised concerns by the reviewer, ... | Revision summary
We truly appreciate all the reviewers and the chairs for their commitment to the community and thoughtful reviews. We are highly encouraged that the reviewers found that the motivation and direction of our method are interesting and can contribute to the community (Reviewer AJF4, wkZo), our work bridge... | [{"id":"nnvEKpNejH","parent_id":"0NruoU6s5Z","created_at_ms":1700667751331,"modified_at_ms":1700667751331,"author":"Authors","role":"authors","type":"Author Reply","content":{"title":"Revision summary","comment":"We truly appreciate all the reviewers and the chairs for their commitment to the community and thoughtful r... | ["Geonmo Gu","Sanghyuk Chun","Wonjae Kim","HeeJae Jun","Yoohoon Kang","Sangdoo Yun"] | ["NAVER","NAVER AI Lab","NAVER","NAVER","NAVER","NAVER"] | 2303.11916 | cs.CV cs.IR | 1,679,411,195,000 | 1 | unchanged |
0NruoU6s5Z | kypDsXX5U1 | 2,024 | CompoDiff: Versatile Composed Image Retrieval With Latent Diffusion | This paper proposes a novel diffusion-based model, CompoDiff, for solving Composed Image Retrieval (CIR) with latent diffusion and presents a newly created dataset, named SynthTriplets18M, of 18 million reference images, conditions, and corresponding target image triplets to train the model. CompoDiff and SynthTriplets... | ["Composed Image Retrieval","Diffusion Models"] | summary:
This paper proposes a composed image retrieval method, called CompoDiff, based on the diffusion model and proposes a large-scale dataset, called SynthTriplets18M, for the composed image retrieval task. In the experiments, the qualitative and quantitative results demonstrated that the performance of the propose... | 1 | 5 | 1,698,772,075,004 | We appreciate the reviewer for their positive comments and constructive feedbacks. We will address all the raised concerns by the reviewer, and will revise our paper as soon as possible (Hopefully before Tuesday).
**[W1] The authors just proposed a larger dataset**
Our main claim is the problem of requiring **pre-col... | [{"id":"bD9GDWeGJX","parent_id":"kypDsXX5U1","created_at_ms":1700487466408,"modified_at_ms":1700666499177,"author":"Authors","role":"authors","type":"Author Reply","content":{"comment":"We appreciate the reviewer for their positive comments and constructive feedbacks. We will address all the raised concerns by the revi... | Revision summary
We truly appreciate all the reviewers and the chairs for their commitment to the community and thoughtful reviews. We are highly encouraged that the reviewers found that the motivation and direction of our method are interesting and can contribute to the community (Reviewer AJF4, wkZo), our work bridge... | [{"id":"nnvEKpNejH","parent_id":"0NruoU6s5Z","created_at_ms":1700667751331,"modified_at_ms":1700667751331,"author":"Authors","role":"authors","type":"Author Reply","content":{"title":"Revision summary","comment":"We truly appreciate all the reviewers and the chairs for their commitment to the community and thoughtful r... | ["Geonmo Gu","Sanghyuk Chun","Wonjae Kim","HeeJae Jun","Yoohoon Kang","Sangdoo Yun"] | ["NAVER","NAVER AI Lab","NAVER","NAVER","NAVER","NAVER"] | 2303.11916 | cs.CV cs.IR | 1,679,411,195,000 | 1 | unchanged |
0Nui91LBQS | OFcjdHZdiE | 2,024 | Making LLaMA SEE and Draw with SEED Tokenizer | The great success of Large Language Models (LLMs) has expanded the potential of multimodality, contributing to the gradual evolution of General Artificial Intelligence (AGI). A true AGI agent should not only possess the capability to perform predefined multi-tasks but also exhibit emergent abilities in an open-world co... | ["Multimodal Large Language Model","Discrete Visual Tokenizer"] | summary:
This work introduces an image tokenizer, which is capable of discretizing images into a series of tokens. These image tokens are transformed by Q-Former into pseudo-causal tokens that can serve as the input for Large Language Models (LLMs), most importantly, they can also act as the target. This allows the mod... | 1 | 5 | 1,698,762,803,796 | **W1: The information leakage by bidirectional attention in ViT.**
**Response:** The bidirectional attention in ViT is among 2D patches while the unidirectional causal attention in Causal Q-Former is among 1D queries. It signifies that the information of the current query is derived based on the context provided by pr... | [{"id":"iaxbU32kzX","parent_id":"OFcjdHZdiE","created_at_ms":1700300693532,"modified_at_ms":1700658966185,"author":"Authors","role":"authors","type":"Author Reply","content":{"comment":"**W1: The information leakage by bidirectional attention in ViT.**\n\n**Response:** The bidirectional attention in ViT is among 2D pat... | A kind reminder regarding our response
We thank you for your time and effort in reviewing our paper. We have responded to all comments in the rebuttal. We would like to remind that the rebuttal period is approaching its end. If you have any other comments or questions, please let us know.
Thank you for your attention... | [{"id":"MpvK3PsmnP","parent_id":"0Nui91LBQS","created_at_ms":1700659273233,"modified_at_ms":1700659273233,"author":"Authors","role":"authors","type":"Author Reply","content":{"title":"A kind reminder regarding our response","comment":"We thank you for your time and effort in reviewing our paper. We have responded to al... | ["Yuying Ge","Sijie Zhao","Ziyun Zeng","Yixiao Ge","Chen Li","Xintao Wang","Ying Shan"] | ["Tencent AI Lab","Tencent AI Lab","Tsinghua University","Tencent","Tencent ARC Lab","Tencent","Tencent PCG ARC Lab"] | 2310.01218 | cs.CV | 1,696,255,382,000 | 1 | changed |
0Nui91LBQS | SCfwSnUuRp | 2,024 | Making LLaMA SEE and Draw with SEED Tokenizer | The great success of Large Language Models (LLMs) has expanded the potential of multimodality, contributing to the gradual evolution of General Artificial Intelligence (AGI). A true AGI agent should not only possess the capability to perform predefined multi-tasks but also exhibit emergent abilities in an open-world co... | ["Multimodal Large Language Model","Discrete Visual Tokenizer"] | summary:
Making the language model to see the words is one of the key research direction. This paper present new image tokenization on this direction. Specifically, unlike prior attempts that uses simple 2d style image tokenization (usually VQ-VAE), this paper propose SEED, which makes image embedding to be left-to-rig... | 1 | 5 | 1,699,282,327,755 | **W1: Not religiously explained and self-contained. Sec. 3.1 is hard to follow. No equation.**
**Response:** Thank you for the suggestion. We designed the SEED tokenizer with two premises: (1) image tokens should capture high-level semantics, and (2) image tokens should be produced under 1D causal dependence. The arch... | [{"id":"DFvlFHxTpJ","parent_id":"SCfwSnUuRp","created_at_ms":1700300547415,"modified_at_ms":1700300760911,"author":"Authors","role":"authors","type":"Author Reply","content":{"comment":"**W1: Not religiously explained and self-contained. Sec. 3.1 is hard to follow. No equation.**\n\n**Response:** Thank you for the sugg... | A kind reminder regarding our response
We thank you for your time and effort in reviewing our paper. We have responded to all comments in the rebuttal. We would like to remind that the rebuttal period is approaching its end. If you have any other comments or questions, please let us know.
Thank you for your attention... | [{"id":"MpvK3PsmnP","parent_id":"0Nui91LBQS","created_at_ms":1700659273233,"modified_at_ms":1700659273233,"author":"Authors","role":"authors","type":"Author Reply","content":{"title":"A kind reminder regarding our response","comment":"We thank you for your time and effort in reviewing our paper. We have responded to al... | ["Yuying Ge","Sijie Zhao","Ziyun Zeng","Yixiao Ge","Chen Li","Xintao Wang","Ying Shan"] | ["Tencent AI Lab","Tencent AI Lab","Tsinghua University","Tencent","Tencent ARC Lab","Tencent","Tencent PCG ARC Lab"] | 2310.01218 | cs.CV | 1,696,255,382,000 | 1 | unchanged |
0Nui91LBQS | wCe12Ato6W | 2,024 | Making LLaMA SEE and Draw with SEED Tokenizer | The great success of Large Language Models (LLMs) has expanded the potential of multimodality, contributing to the gradual evolution of General Artificial Intelligence (AGI). A true AGI agent should not only possess the capability to perform predefined multi-tasks but also exhibit emergent abilities in an open-world co... | ["Multimodal Large Language Model","Discrete Visual Tokenizer"] | summary:
The paper presents a method (SEED) to augment large language models (LLMs) with the capability of processing and generating visual data, i.e., images. The core contribution of SEED is a quantized tokenizer that learns to encode images into discrete visual tokens which can be again decoded using a pre-trained g... | 1 | 8 | 1,698,705,889,770 | **W1-1: Write loss functions.**
**Response:** Thank you for the advice. We have revised the method section and included all loss functions in our revised manuscript.
**W1-2: Why use two different image encoders.**
**Response:** We use two different image encoders in order to ease training by leveraging the pre-train... | [{"id":"q474W1fqKn","parent_id":"wCe12Ato6W","created_at_ms":1700300721705,"modified_at_ms":1700300721705,"author":"Authors","role":"authors","type":"Author Reply","content":{"comment":"**W1-1: Write loss functions.**\n\n**Response:** Thank you for the advice. We have revised the method section and included all loss fu... | null | [] | ["Yuying Ge","Sijie Zhao","Ziyun Zeng","Yixiao Ge","Chen Li","Xintao Wang","Ying Shan"] | ["Tencent AI Lab","Tencent AI Lab","Tsinghua University","Tencent","Tencent ARC Lab","Tencent","Tencent PCG ARC Lab"] | 2310.01218 | cs.CV | 1,696,255,382,000 | 1 | unchanged |
0P12CqfvUU | HAnp9k3TWC | 2,024 | 3D Morphable Master Face Generation: Towards Controllable Wolf Attacks against 2D and 3D Face Recognition Systems | Biometric authentication systems are facing increasing threats from Artificial Intelligence-Generated Content (AIGC). Previous research has revealed the vulnerability of face authentication systems against master face attacks. These attacks utilize generative models to create facial samples capable of matching multiple... | ["Master Attack","Wolf Attack","3D Morphable Face Model"] | summary:
Summary: In this paper, the authors study the master attack on 3D face recognition systems. They use combined 2D and 3D face similarities as a loss functions, CMA-ES as an optimizer, 3DMM for controlling the face properties.
strengths:
1. The first several sections are well-presented.
2. The algorithm is sim... | 1 | 3 | 1,698,229,435,839 | Thank you for your detailed feedback and suggestions. We are willing to discuss your concerns and questions below.
**Weakness 1** *Some consider that master attack is a real problem. This problem is existing when we run the system in identification mode, 1-to many matching, which is used in low security applic... | [{"id":"BJnf6mpCxu","parent_id":"HAnp9k3TWC","created_at_ms":1700040931329,"modified_at_ms":1700040931329,"author":"Authors","role":"authors","type":"Author Reply","content":{"comment":"Thank you for your detailed feedback and suggestions. We are willing to discuss your concerns and questions below.\n\n \n\n**Weak... | null | [] | ["Siyun Liang","Huy H. Nguyen","Junichi Yamagishi","Isao Echizen"] | ["Technische Universität München",null,"National Institute of Informatics",null] | null | null | null | 0 | unchanged |
0P12CqfvUU | IdSJDjBzA0 | 2,024 | 3D Morphable Master Face Generation: Towards Controllable Wolf Attacks against 2D and 3D Face Recognition Systems | Biometric authentication systems are facing increasing threats from Artificial Intelligence-Generated Content (AIGC). Previous research has revealed the vulnerability of face authentication systems against master face attacks. These attacks utilize generative models to create facial samples capable of matching multiple... | ["Master Attack","Wolf Attack","3D Morphable Face Model"] | summary:
The authors aim at presenting a systematic approach for generating master faces that can compromise both 2D and 3D face recognition systems using the Latent Variable Evolution with the 3D Face Morphable Model.
strengths:
The paper is well-written.
The motivation is clear.
The experimental setup is extensive.
... | 1 | 5 | 1,699,607,359,020 | Thank you very much for your feedback and the question. Our contributions are summarized below.
First, we want to discuss why studying 3D master face attacks is important. While previous work has focused on 2D master face attacks, **modern face authentication systems no longer rely solely on 2D face recognit... | [{"id":"dPo447Fps1","parent_id":"IdSJDjBzA0","created_at_ms":1700037911631,"modified_at_ms":1700037911631,"author":"Authors","role":"authors","type":"Author Reply","content":{"comment":"Thank you very much for your feedback and the question. Our contributions are summarized below. \n\n \n\nFirst, we want to discu... | null | [] | ["Siyun Liang","Huy H. Nguyen","Junichi Yamagishi","Isao Echizen"] | ["Technische Universität München",null,"National Institute of Informatics",null] | null | null | null | 0 | unchanged |
0P12CqfvUU | sNAwyK2zEb | 2,024 | 3D Morphable Master Face Generation: Towards Controllable Wolf Attacks against 2D and 3D Face Recognition Systems | Biometric authentication systems are facing increasing threats from Artificial Intelligence-Generated Content (AIGC). Previous research has revealed the vulnerability of face authentication systems against master face attacks. These attacks utilize generative models to create facial samples capable of matching multiple... | ["Master Attack","Wolf Attack","3D Morphable Face Model"] | summary:
This paper presents the vulnerability of biometric authentication systems to master face attacks, which utilize generative models to create facial samples that can match multiple registered user templates. The authors propose a systematic approach for generating 3D master faces that can compromise both 2D and ... | 1 | 3 | 1,698,562,941,709 | Thank you for your detailed feedback and suggestions. We are willing to discuss your concerns and questions below.
**Weakness 1** *The problem of master face attacks for 2D FR systems has been studied by Nguyen et al. (2020), Shmelkin et al. (2021) and Nguyen et al. (2022). The authors extend this issue to 2D ... | [{"id":"aRIAn8ZCec","parent_id":"sNAwyK2zEb","created_at_ms":1700040391850,"modified_at_ms":1700040391850,"author":"Authors","role":"authors","type":"Author Reply","content":{"comment":"Thank you for your detailed feedback and suggestions. We are willing to discuss your concerns and questions below.\n\n \n\n**Weak... | null | [] | ["Siyun Liang","Huy H. Nguyen","Junichi Yamagishi","Isao Echizen"] | ["Technische Universität München",null,"National Institute of Informatics",null] | null | null | null | 0 | unchanged |
0P12CqfvUU | zGWV81DLQi | 2,024 | 3D Morphable Master Face Generation: Towards Controllable Wolf Attacks against 2D and 3D Face Recognition Systems | Biometric authentication systems are facing increasing threats from Artificial Intelligence-Generated Content (AIGC). Previous research has revealed the vulnerability of face authentication systems against master face attacks. These attacks utilize generative models to create facial samples capable of matching multiple... | ["Master Attack","Wolf Attack","3D Morphable Face Model"] | summary:
This paper proposes an approach for generating master faces that can compromise both 2D and 3D face recognition systems.3D master faces are generated using the Latent Variable Evolution (LVE) algorithm with the 3D Face Morphable Model (3DMM). Simultaneous master face attacks in both white-box, gray-box, and bl... | 1 | 3 | 1,699,366,810,251 | Thank you for your detailed feedback and suggestions. We are willing to discuss your concerns and questions below.
**Weakness 1** *The contribution of the paper in terms of learning solutions and its relevance to the conference seem a bit below the standard for ICLR*
The clarification of our contribution is s... | [{"id":"ssM4gkvLC9","parent_id":"zGWV81DLQi","created_at_ms":1700039106920,"modified_at_ms":1700041211589,"author":"Authors","role":"authors","type":"Author Reply","content":{"comment":"Thank you for your detailed feedback and suggestions. We are willing to discuss your concerns and questions below.\n\n \n\n**Weak... | null | [] | ["Siyun Liang","Huy H. Nguyen","Junichi Yamagishi","Isao Echizen"] | ["Technische Universität München",null,"National Institute of Informatics",null] | null | null | null | 0 | unchanged |
0Pu0H7y3gg | 6V0txLehnI | 2,024 | Understanding the Initial Condensation of Convolutional Neural Networks | Previous research has shown that fully-connected neural networks with small initialization and gradient-based training methods exhibit a phenomenon known as condensation during training. This phenomenon refers to the input weights of hidden neurons condensing into isolated orientations during training, revealing an imp... | ["training dynamics","convolutional neural networks","initialization","gradient-based training methods"] | summary:
This paper studies the initial condensation phenomenon of training CNNs, supported by both experimental results and theoretical analysis.
strengths:
Understanding the training dynamics of gradient-based method is a crucial theoretical issue. While previous research has primarily concentrated on fully connecte... | 1 | 6 | 1,698,764,653,860 | Response to Reviewer P4L3
$\textbf{Point 1: }$
If Assumption 4 becomes $\lambda_1=\lambda_2>\lambda_3$, how will the initial condensation phenomenon change? My guess is that there will be two condensation directions, corresponding to the first two eigendirections. Is that right?
$\textbf{Reply 1: }$
Yes this will ma... | [{"id":"JrCuh42Anj","parent_id":"6V0txLehnI","created_at_ms":1700330282178,"modified_at_ms":1700330282178,"author":"Authors","role":"authors","type":"Author Reply","content":{"title":"Response to Reviewer P4L3","comment":"$\\textbf{Point 1: }$\n\nIf Assumption 4 becomes $\\lambda_1=\\lambda_2>\\lambda_3$, how will the ... | null | [] | ["Zhangchen Zhou","Hanxu Zhou","Yuqing Li","Zhi-Qin John Xu"] | ["Shanghai Jiaotong University","Shanghai Jiaotong University","Shanghai Jiaotong University","Shanghai Jiaotong University"] | 2305.09947 | cs.LG cs.NE | 1,684,299,647,000 | 1 | unchanged |
0Pu0H7y3gg | C04q1SPExA | 2,024 | Understanding the Initial Condensation of Convolutional Neural Networks | Previous research has shown that fully-connected neural networks with small initialization and gradient-based training methods exhibit a phenomenon known as condensation during training. This phenomenon refers to the input weights of hidden neurons condensing into isolated orientations during training, revealing an imp... | ["training dynamics","convolutional neural networks","initialization","gradient-based training methods"] | summary:
This manuscript presents a study of condensation in convolutional neural networks.
The main theorem states that under certain assumptions on the data, activation function and initial weights, the following two things hold:
1. the final weights go arbitrarily far away from the starting point
2. the final wei... | 1 | 3 | 1,698,419,927,483 | Response to Reviewer FvGf
$\textbf{Point 1:}$
(clarity) The paper is $\dots$ considers convolutional networks
$\textbf{Reply 1:}$
First of all, we agree that a consistent and clear notation, such as using $\\theta_{W,v_1} := \operatorname{vec}({\theta_{W,\beta}\cdot v_1})\)$, would enhance the readability, particularl... | [{"id":"yFmO5LpCNE","parent_id":"C04q1SPExA","created_at_ms":1700331066265,"modified_at_ms":1700331377988,"author":"Authors","role":"authors","type":"Author Reply","content":{"title":"Response to Reviewer FvGf","comment":"$\\textbf{Point 1:}$\n(clarity) The paper is $\\dots$ considers convolutional networks\n\n$\\textb... | null | [] | ["Zhangchen Zhou","Hanxu Zhou","Yuqing Li","Zhi-Qin John Xu"] | ["Shanghai Jiaotong University","Shanghai Jiaotong University","Shanghai Jiaotong University","Shanghai Jiaotong University"] | 2305.09947 | cs.LG cs.NE | 1,684,299,647,000 | 1 | unchanged |
0Pu0H7y3gg | ikJd6Ng9jY | 2,024 | Understanding the Initial Condensation of Convolutional Neural Networks | Previous research has shown that fully-connected neural networks with small initialization and gradient-based training methods exhibit a phenomenon known as condensation during training. This phenomenon refers to the input weights of hidden neurons condensing into isolated orientations during training, revealing an imp... | ["training dynamics","convolutional neural networks","initialization","gradient-based training methods"] | summary:
This paper analyzes the phenomenon called initial condensation in simple CNN models. Initial condensation refers to the occurrence of weight grouping in the early stages of neural network (NN) training, which has been discussed in previous literature mainly with respect to fully connected NNs. A two-layer CNN ... | 1 | 3 | 1,697,417,741,065 | Response to Reviewer mFgA
$\textbf{Point 1:}$
It is difficult to follow the technical details of the paper for the following reasons:
1. $\boldsymbol{W}_{p,q,\alpha,\beta}^{[l]}$is not well defined for $p=\infty,\infty,\cdots,\infty$ and $q=\infty,\infty,\cdots,\infty$, although it is required to define.
2. I could ... | [{"id":"6VYuIXHiBW","parent_id":"ikJd6Ng9jY","created_at_ms":1700334212100,"modified_at_ms":1700334212100,"author":"Authors","role":"authors","type":"Author Reply","content":{"title":"Response to Reviewer mFgA","comment":"$\\textbf{Point 1:}$\n\nIt is difficult to follow the technical details of the paper for the follo... | null | [] | ["Zhangchen Zhou","Hanxu Zhou","Yuqing Li","Zhi-Qin John Xu"] | ["Shanghai Jiaotong University","Shanghai Jiaotong University","Shanghai Jiaotong University","Shanghai Jiaotong University"] | 2305.09947 | cs.LG cs.NE | 1,684,299,647,000 | 1 | unchanged |
0QAzIMq32X | 0JDy5bGcUt | 2,024 | Inner Classifier-Free Guidance and Its Taylor Expansion for Diffusion Models | Classifier-free guidance (CFG) is a pivotal technique for balancing the diversity and fidelity of samples in conditional diffusion models. This approach involves utilizing a single model to jointly optimize the conditional score predictor and unconditional score predictor, eliminating the need for additional classifier... | ["diffusion models","classifier-free guidance"] | summary:
This work presents a new perspective on classifier-free guidance (CFG) for diffusion models imposing specific assumptions on the space of condition. Assuming that the condition space has a cone structure, the previous CFG can be seen as the first-order Taylor expansion of the proposed ICFG, and this work furth... | 1 | 3 | 1,699,360,843,822 | Rebuttal part I
We thank reviewer cPdv very much for recognizing the theoretical contributions and novelty of our idea. Your assistance in refining assumptions and conducting a deeper analysis of the gap between enhanced intermediate distributions has been particularly insightful.
### **W1-1: It is not clear which add... | [{"id":"NsElMPZ14H","parent_id":"0JDy5bGcUt","created_at_ms":1700475469578,"modified_at_ms":1700726826436,"author":"Authors","role":"authors","type":"Author Reply","content":{"title":"Rebuttal part I","comment":"We thank reviewer cPdv very much for recognizing the theoretical contributions and novelty of our idea. Your... | Rebuttal to all reviewers
We express our sincere gratitude to all the reviewers for their invaluable contributions.
We thank reviewer cPdv very much for recognizing the theoretical contributions and novelty of our idea. Their assistance in refining assumptions and conducting a deeper analysis of the gap between enhanc... | [{"id":"NUxOZiYqdk","parent_id":"0QAzIMq32X","created_at_ms":1700475122850,"modified_at_ms":1700475122850,"author":"Authors","role":"authors","type":"Author Reply","content":{"title":"Rebuttal to all reviewers","comment":"We express our sincere gratitude to all the reviewers for their invaluable contributions.\n\nWe th... | ["Shikun Sun","Longhui Wei","Zhicai Wang","Zixuan Wang","Junliang Xing","Jia Jia","Qi Tian"] | ["Tsinghua University","Huawei Technologies Ltd.","University of Science and Technology of China","Tsinghua University","Tsinghua University","Tsinghua University","Huawei Technologies Ltd."] | null | null | null | 0 | changed |
0QAzIMq32X | 0zH6TJ4hGc | 2,024 | Inner Classifier-Free Guidance and Its Taylor Expansion for Diffusion Models | Classifier-free guidance (CFG) is a pivotal technique for balancing the diversity and fidelity of samples in conditional diffusion models. This approach involves utilizing a single model to jointly optimize the conditional score predictor and unconditional score predictor, eliminating the need for additional classifier... | ["diffusion models","classifier-free guidance"] | summary:
This paper introduces an enhancement of classifier-free guidance (CFG) for diffusion models, called inner classifier-free guidance (ICFG). The paper claims that CFG can be extended to ICFG when the condition is continuous, leading to further improvements. They provide a theoretical analysis based on the condit... | 1 | 3 | 1,698,727,970,199 | Rebuttal part I
Our gratitude extends to Reviewer UdAr for acknowledging the novelty in our approach to combining guidance strength and condition. Their assistance in correcting definitions, perfecting proofs and the formulation of the theorem, along with their guidance in including additional empirical results and spe... | [{"id":"jR8WDIZA8i","parent_id":"0zH6TJ4hGc","created_at_ms":1700476053959,"modified_at_ms":1700726907083,"author":"Authors","role":"authors","type":"Author Reply","content":{"title":"Rebuttal part I","comment":"Our gratitude extends to Reviewer UdAr for acknowledging the novelty in our approach to combining guidance s... | Rebuttal to all reviewers
We express our sincere gratitude to all the reviewers for their invaluable contributions.
We thank reviewer cPdv very much for recognizing the theoretical contributions and novelty of our idea. Their assistance in refining assumptions and conducting a deeper analysis of the gap between enhanc... | [{"id":"NUxOZiYqdk","parent_id":"0QAzIMq32X","created_at_ms":1700475122850,"modified_at_ms":1700475122850,"author":"Authors","role":"authors","type":"Author Reply","content":{"title":"Rebuttal to all reviewers","comment":"We express our sincere gratitude to all the reviewers for their invaluable contributions.\n\nWe th... | ["Shikun Sun","Longhui Wei","Zhicai Wang","Zixuan Wang","Junliang Xing","Jia Jia","Qi Tian"] | ["Tsinghua University","Huawei Technologies Ltd.","University of Science and Technology of China","Tsinghua University","Tsinghua University","Tsinghua University","Huawei Technologies Ltd."] | null | null | null | 0 | changed |
0QAzIMq32X | RMV9cTzYc7 | 2,024 | Inner Classifier-Free Guidance and Its Taylor Expansion for Diffusion Models | Classifier-free guidance (CFG) is a pivotal technique for balancing the diversity and fidelity of samples in conditional diffusion models. This approach involves utilizing a single model to jointly optimize the conditional score predictor and unconditional score predictor, eliminating the need for additional classifier... | ["diffusion models","classifier-free guidance"] | summary:
This paper generalizes CFG for diffusion model guidance by adapting the guidance strength according to the relevance between the condition and a given sample. Conditions more relevant to a sample require weaker guidance strength.
strengths:
The paper identifies an issue with the deviating SDE when guidance is... | 1 | 6 | 1,699,321,340,466 | Rebuttal by Authors
We extend our appreciation to Reviewer LfUU for acknowledging both our theoretical analysis and empirical results. Their guidance in accelerating our method and perfecting the non-strict 2nd order term reasons has been tremendously helpful.
### **Q1: Cone assumption.**
We appreciate your understan... | [{"id":"jT24AfwmMl","parent_id":"RMV9cTzYc7","created_at_ms":1700475664056,"modified_at_ms":1700726937404,"author":"Authors","role":"authors","type":"Author Reply","content":{"title":"Rebuttal by Authors","comment":"We extend our appreciation to Reviewer LfUU for acknowledging both our theoretical analysis and empirica... | Rebuttal to all reviewers
We express our sincere gratitude to all the reviewers for their invaluable contributions.
We thank reviewer cPdv very much for recognizing the theoretical contributions and novelty of our idea. Their assistance in refining assumptions and conducting a deeper analysis of the gap between enhanc... | [{"id":"NUxOZiYqdk","parent_id":"0QAzIMq32X","created_at_ms":1700475122850,"modified_at_ms":1700475122850,"author":"Authors","role":"authors","type":"Author Reply","content":{"title":"Rebuttal to all reviewers","comment":"We express our sincere gratitude to all the reviewers for their invaluable contributions.\n\nWe th... | ["Shikun Sun","Longhui Wei","Zhicai Wang","Zixuan Wang","Junliang Xing","Jia Jia","Qi Tian"] | ["Tsinghua University","Huawei Technologies Ltd.","University of Science and Technology of China","Tsinghua University","Tsinghua University","Tsinghua University","Huawei Technologies Ltd."] | null | null | null | 0 | changed |
0QAzIMq32X | xkqbvcRjXN | 2,024 | Inner Classifier-Free Guidance and Its Taylor Expansion for Diffusion Models | Classifier-free guidance (CFG) is a pivotal technique for balancing the diversity and fidelity of samples in conditional diffusion models. This approach involves utilizing a single model to jointly optimize the conditional score predictor and unconditional score predictor, eliminating the need for additional classifier... | ["diffusion models","classifier-free guidance"] | summary:
This paper presents a generalized version of classifier-free guidance (CFG), i.e., inner classifier-free guidance. By exploiting the continuity of the generation condition, the author propose an interesting taylor expansion formulation to interpret CFG, where the classic CFG is viewed as the first order case o... | 1 | 5 | 1,698,856,980,845 | Rebuttal by Authors
We are deeply thankful for Reviewer jYKn's recognition of our novel formulation and well-studied hyper-parameters. We sincerely value their guidance in recommending the inclusion of additional empirical results, particularly in the form of visual comparisons.
### **Q1: Need more experiment results.... | [{"id":"mnLtE5gRMS","parent_id":"xkqbvcRjXN","created_at_ms":1700475917639,"modified_at_ms":1700475917639,"author":"Authors","role":"authors","type":"Author Reply","content":{"title":"Rebuttal by Authors","comment":"We are deeply thankful for Reviewer jYKn's recognition of our novel formulation and well-studied hyper-p... | Rebuttal to all reviewers
We express our sincere gratitude to all the reviewers for their invaluable contributions.
We thank reviewer cPdv very much for recognizing the theoretical contributions and novelty of our idea. Their assistance in refining assumptions and conducting a deeper analysis of the gap between enhanc... | [{"id":"NUxOZiYqdk","parent_id":"0QAzIMq32X","created_at_ms":1700475122850,"modified_at_ms":1700475122850,"author":"Authors","role":"authors","type":"Author Reply","content":{"title":"Rebuttal to all reviewers","comment":"We express our sincere gratitude to all the reviewers for their invaluable contributions.\n\nWe th... | ["Shikun Sun","Longhui Wei","Zhicai Wang","Zixuan Wang","Junliang Xing","Jia Jia","Qi Tian"] | ["Tsinghua University","Huawei Technologies Ltd.","University of Science and Technology of China","Tsinghua University","Tsinghua University","Tsinghua University","Huawei Technologies Ltd."] | null | null | null | 0 | changed |
0SOhDO7xI0 | 8SxH0lQZcB | 2,024 | DeepDRK: Deep Dependency Regularized Knockoff for Feature Selection | Model-X knockoff, among various feature selection methods, received much attention recently due to its guarantee on false discovery rate (FDR) control. Subsequent to its introduction in parametric design, knockoff is advanced to handle arbitrary data distributions using deep learning-based generative modeling. However,... | ["Feature Selection","Deep Learning","Model-X Knockoff","FDR Control"] | summary:
The authors develop a distribution-free deep learning method for knockoff generation which strikes a balance between FDR and power, called “Deep Dependency Regularized Knockoff” (DeepDRK). In DeepDRK, a “multi-swapper” adversarial training procedure is proposed to enforce the swap property, while a sliced-Wass... | 1 | 5 | 1,698,664,880,985 | Authors' Response (part 1)
Thank you for your insightful and constructive feedback on our manuscript. We appreciate the time and effort you invested in reviewing our work. We have carefully considered your comments and have made the following revisions to address the concerns raised above:
## Weaknesses
- Though it e... | [{"id":"sjAM7TGKOO","parent_id":"8SxH0lQZcB","created_at_ms":1700425022504,"modified_at_ms":1700425022504,"author":"Authors","role":"authors","type":"Author Reply","content":{"title":"Authors' Response (part 1)","comment":"Thank you for your insightful and constructive feedback on our manuscript. We appreciate the time... | General Response (part 1)
Dear reviewers,
Thank you for your time and the insightful feedback. We took careful considerations and added further experiments and clarifications to the original version. We believe that the quality and the clarity of the manuscript has been improved in this way. In what follows, we first ... | [{"id":"5xzH63neqm","parent_id":"0SOhDO7xI0","created_at_ms":1700422169248,"modified_at_ms":1700422169248,"author":"Authors","role":"authors","type":"Author Reply","content":{"title":"General Response (part 1)","comment":"Dear reviewers,\n\nThank you for your time and the insightful feedback. We took careful considerat... | ["Hongyu Shen","Yici Yan","Zhizhen Zhao"] | ["University of Illinois, Urbana Champaign","University of Illinois, Urbana Champaign","University of Washington"] | 2402.17176 | cs.LG | 1,709,004,294,000 | 0 | unchanged |
0SOhDO7xI0 | tEQX45duoz | 2,024 | DeepDRK: Deep Dependency Regularized Knockoff for Feature Selection | Model-X knockoff, among various feature selection methods, received much attention recently due to its guarantee on false discovery rate (FDR) control. Subsequent to its introduction in parametric design, knockoff is advanced to handle arbitrary data distributions using deep learning-based generative modeling. However,... | ["Feature Selection","Deep Learning","Model-X Knockoff","FDR Control"] | summary:
This paper investigates the problem of feature selection from the perspective of Model-X knockoff owing to its guarantee of false discovery rate (FDR) control. Realizing the diminished selection power caused by the swap property that knockoffs need, the authors proposed a Deep Dependency Regularized Knockoff... | 1 | 6 | 1,699,193,544,773 | Author's Response
Thanks a lot for your appreciation of this work. We further improved the manuscript based on reviewers’ comments and suggestions. Please see the general response to all reviewers for details. | [{"id":"lWJKkVY5p7","parent_id":"tEQX45duoz","created_at_ms":1700422658855,"modified_at_ms":1700422658855,"author":"Authors","role":"authors","type":"Author Reply","content":{"title":"Author's Response","comment":"Thanks a lot for your appreciation of this work. We further improved the manuscript based on reviewers’ co... | General Response (part 1)
Dear reviewers,
Thank you for your time and the insightful feedback. We took careful considerations and added further experiments and clarifications to the original version. We believe that the quality and the clarity of the manuscript has been improved in this way. In what follows, we first ... | [{"id":"5xzH63neqm","parent_id":"0SOhDO7xI0","created_at_ms":1700422169248,"modified_at_ms":1700422169248,"author":"Authors","role":"authors","type":"Author Reply","content":{"title":"General Response (part 1)","comment":"Dear reviewers,\n\nThank you for your time and the insightful feedback. We took careful considerat... | ["Hongyu Shen","Yici Yan","Zhizhen Zhao"] | ["University of Illinois, Urbana Champaign","University of Illinois, Urbana Champaign","University of Washington"] | 2402.17176 | cs.LG | 1,709,004,294,000 | 0 | unchanged |
0SOhDO7xI0 | tWbuOUTe2U | 2,024 | DeepDRK: Deep Dependency Regularized Knockoff for Feature Selection | Model-X knockoff, among various feature selection methods, received much attention recently due to its guarantee on false discovery rate (FDR) control. Subsequent to its introduction in parametric design, knockoff is advanced to handle arbitrary data distributions using deep learning-based generative modeling. However,... | ["Feature Selection","Deep Learning","Model-X Knockoff","FDR Control"] | summary:
This paper proposes DeepDRK, a new model-X knockoff based methods which adopts a two-stage framework to generate knockoff variables. A ViT (called knockoff transformer in the paper) is trained by minimizing a swap loss plus a dependency regularization loss in the training stage, while its output is further per... | 1 | 6 | 1,698,758,874,751 | Authors' Response (part 1)
Thank you for your insightful and constructive feedback on our manuscript. We have carefully considered your comments and have made the following revisions to address the concerns raised above:
## Weaknesses
- The loss in DeepDRK contains five terms, SWD, REx, cosine similarity w.r.t. swappe... | [{"id":"I7IHdXD5fn","parent_id":"tWbuOUTe2U","created_at_ms":1700424322466,"modified_at_ms":1700424322466,"author":"Authors","role":"authors","type":"Author Reply","content":{"title":"Authors' Response (part 1)","comment":"Thank you for your insightful and constructive feedback on our manuscript. We have carefully cons... | General Response (part 1)
Dear reviewers,
Thank you for your time and the insightful feedback. We took careful considerations and added further experiments and clarifications to the original version. We believe that the quality and the clarity of the manuscript has been improved in this way. In what follows, we first ... | [{"id":"5xzH63neqm","parent_id":"0SOhDO7xI0","created_at_ms":1700422169248,"modified_at_ms":1700422169248,"author":"Authors","role":"authors","type":"Author Reply","content":{"title":"General Response (part 1)","comment":"Dear reviewers,\n\nThank you for your time and the insightful feedback. We took careful considerat... | ["Hongyu Shen","Yici Yan","Zhizhen Zhao"] | ["University of Illinois, Urbana Champaign","University of Illinois, Urbana Champaign","University of Washington"] | 2402.17176 | cs.LG | 1,709,004,294,000 | 0 | unchanged |
0SOhDO7xI0 | zp8w7WbTvN | 2,024 | DeepDRK: Deep Dependency Regularized Knockoff for Feature Selection | Model-X knockoff, among various feature selection methods, received much attention recently due to its guarantee on false discovery rate (FDR) control. Subsequent to its introduction in parametric design, knockoff is advanced to handle arbitrary data distributions using deep learning-based generative modeling. However,... | ["Feature Selection","Deep Learning","Model-X Knockoff","FDR Control"] | summary:
The paper proposed “Deep Dependency Regularized Knockoff (DeepDRK)”, a distribution-free deep learning method that strikes a balance between FDR and power. It leverages transformer architecture and several loss functions for training to generate Knockoff.
strengths:
1. It introduces knockoff Transformer to ge... | 1 | 6 | 1,698,889,819,844 | Authors' Response
Thank you for your insightful and constructive feedback on our manuscript. We appreciate the time and effort you invested in reviewing our work. We have carefully considered your comments and have made the following revisions to address the concerns raised above:
## Weaknesses
- Some arguments of the... | [{"id":"fZoQtrBOBo","parent_id":"zp8w7WbTvN","created_at_ms":1700423522993,"modified_at_ms":1700423573284,"author":"Authors","role":"authors","type":"Author Reply","content":{"title":"Authors' Response","comment":"Thank you for your insightful and constructive feedback on our manuscript. We appreciate the time and effo... | General Response (part 1)
Dear reviewers,
Thank you for your time and the insightful feedback. We took careful considerations and added further experiments and clarifications to the original version. We believe that the quality and the clarity of the manuscript has been improved in this way. In what follows, we first ... | [{"id":"5xzH63neqm","parent_id":"0SOhDO7xI0","created_at_ms":1700422169248,"modified_at_ms":1700422169248,"author":"Authors","role":"authors","type":"Author Reply","content":{"title":"General Response (part 1)","comment":"Dear reviewers,\n\nThank you for your time and the insightful feedback. We took careful considerat... | ["Hongyu Shen","Yici Yan","Zhizhen Zhao"] | ["University of Illinois, Urbana Champaign","University of Illinois, Urbana Champaign","University of Washington"] | 2402.17176 | cs.LG | 1,709,004,294,000 | 0 | unchanged |
0TZs6WOs16 | 0yQQUIAveQ | 2,024 | Hyperbolic Embeddings in Sequential Self-Attention for Improved Next-Item Recommendations | In recent years, self-attentive sequential learning models have surpassed conventional collaborative filtering techniques in next-item recommendation tasks. However, Euclidean geometry utilized in these models may not be optimal for capturing a complex structure of the behavioral data. Building on recent advances in th... | ["recommender systems","sequential self-attention","hyperbolic geometry","Gromov product"] | summary:
This paper presents an approach that leverages hyperbolic geometry for extending recommendation systems. The idea is that the hyperbolic layer will capture structural properties of the approach. The authors are based on one base model, SASRec, which they extend. They perform experiments on multiple datasets wi... | 1 | 3 | 1,699,186,129,409 | Response to Reviewer vqza
We are grateful to the reviewer for their careful consideration of our work and for the provided suggestions on possible improvements. Below we address the outlined concerns by quoting them and expanding the explanation below the quote. Due to space constraints, our response will be split into... | [{"id":"zGsKC9eHJQ","parent_id":"0yQQUIAveQ","created_at_ms":1700677447572,"modified_at_ms":1700677482706,"author":"Authors","role":"authors","type":"Author Reply","content":{"title":"Response to Reviewer vqza","comment":"We are grateful to the reviewer for their careful consideration of our work and for the provided s... | null | [] | ["Evgeny Frolov","Lina Bashaeva","Leyla Mirvakhabova","Ivan Oseledets"] | ["Skolkovo Institute of Science and Technology",null,null,"Institute of Numerical Mathematics"] | null | null | null | 0 | unchanged |
0TZs6WOs16 | KhL24ud2Ze | 2,024 | Hyperbolic Embeddings in Sequential Self-Attention for Improved Next-Item Recommendations | In recent years, self-attentive sequential learning models have surpassed conventional collaborative filtering techniques in next-item recommendation tasks. However, Euclidean geometry utilized in these models may not be optimal for capturing a complex structure of the behavioral data. Building on recent advances in th... | ["recommender systems","sequential self-attention","hyperbolic geometry","Gromov product"] | summary:
The paper makes the following main contributions:
- The paper proposes a new approach for the next-item recommendation that combines sequential self-attention with hyperbolic geometry. The base architecture is SASRec, with modifications only to the final prediction layer.
- The prediction layer is adapted to ... | 1 | 5 | 1,698,785,279,076 | Response to Reviewer hnY1
Thank you for taking the time to thoroughly assess our work and provide valuable feedback. Below we address the concerns related to the listed weakness points. We quote the original point and then provide the details that should address it. We will separate the points into several subsequent c... | [{"id":"8mHEsYp6z6","parent_id":"KhL24ud2Ze","created_at_ms":1700502464659,"modified_at_ms":1700502464659,"author":"Authors","role":"authors","type":"Author Reply","content":{"title":"Response to Reviewer hnY1","comment":"Thank you for taking the time to thoroughly assess our work and provide valuable feedback. Below w... | null | [] | ["Evgeny Frolov","Lina Bashaeva","Leyla Mirvakhabova","Ivan Oseledets"] | ["Skolkovo Institute of Science and Technology",null,null,"Institute of Numerical Mathematics"] | null | null | null | 0 | unchanged |
0TZs6WOs16 | N1gZ4Qbd2y | 2,024 | Hyperbolic Embeddings in Sequential Self-Attention for Improved Next-Item Recommendations | In recent years, self-attentive sequential learning models have surpassed conventional collaborative filtering techniques in next-item recommendation tasks. However, Euclidean geometry utilized in these models may not be optimal for capturing a complex structure of the behavioral data. Building on recent advances in th... | ["recommender systems","sequential self-attention","hyperbolic geometry","Gromov product"] | summary:
This paper extends SASRec [1] in hyperbolic space by adjusting the output prediction layer for next-item recommendation task.
[1] Self-attentive sequential recommendation. ICDM 2018.
[2] HME: A Hyperbolic Metric Embedding Approach for Next-POI Recommendation. SIGIR 2020.
strengths:
Strengths:
- The approac... | 1 | 3 | 1,698,891,078,873 | Response to Reviewer WdM2
We appreciate the reviewer's dedicated efforts in evaluating our work and providing actionable suggestions and feedback. We will address the concerns by quoting them and expanding the explanation below the quote. Due to space constraints, our response will be split into several comments below.... | [{"id":"E4kPSSSNid","parent_id":"N1gZ4Qbd2y","created_at_ms":1700509837968,"modified_at_ms":1700509837968,"author":"Authors","role":"authors","type":"Author Reply","content":{"title":"Response to Reviewer WdM2","comment":"We appreciate the reviewer's dedicated efforts in evaluating our work and providing actionable sug... | null | [] | ["Evgeny Frolov","Lina Bashaeva","Leyla Mirvakhabova","Ivan Oseledets"] | ["Skolkovo Institute of Science and Technology",null,null,"Institute of Numerical Mathematics"] | null | null | null | 0 | unchanged |
0TZs6WOs16 | TyQiulW8uK | 2,024 | Hyperbolic Embeddings in Sequential Self-Attention for Improved Next-Item Recommendations | In recent years, self-attentive sequential learning models have surpassed conventional collaborative filtering techniques in next-item recommendation tasks. However, Euclidean geometry utilized in these models may not be optimal for capturing a complex structure of the behavioral data. Building on recent advances in th... | ["recommender systems","sequential self-attention","hyperbolic geometry","Gromov product"] | summary:
The paper proposes a hyperbolic architecture for sequential self-attention next-item recommendation applied to SASRec. Specifically, they replace the output prediction layer in SASRec with predictors in hyperbolic space. Besides, they adjust the machine precision setting to obtain a more accurate estimation of... | 1 | 3 | 1,698,914,629,858 | Response to Reviewer 9PVb
We thank the reviewer for the thorough assessment and extensive feedback on our work. Below we address the outlined weakness points by quoting them and expanding the explanation below the quote. Due to space constraints, our response will be split into several comments below.
### Point 1
> Li... | [{"id":"SwnGkqoC26","parent_id":"TyQiulW8uK","created_at_ms":1700645079802,"modified_at_ms":1700645079802,"author":"Authors","role":"authors","type":"Author Reply","content":{"title":"Response to Reviewer 9PVb","comment":"We thank the reviewer for the thorough assessment and extensive feedback on our work. Below we add... | null | [] | ["Evgeny Frolov","Lina Bashaeva","Leyla Mirvakhabova","Ivan Oseledets"] | ["Skolkovo Institute of Science and Technology",null,null,"Institute of Numerical Mathematics"] | null | null | null | 0 | unchanged |
0V311Uh8q1 | Mhu5qkfBUx | 2,024 | Algorithmic Stability Unleashed: Generalization Bounds with Unbounded Losses | One of the central problems of statistical learning theory is quantifying the generalization ability of learning algorithms within a probabilistic framework. Algorithmic stability is a powerful tool for deriving generalization bounds, however, it typically builds on a critical assumption that losses are bounded. In thi... | ["algorithmic stability","generalization bound"] | summary:
The present paper proposes a stability generalization bound for a variant of non-uniform stability condition.
Their analysis revolves around a new form of McDiarmid inequality, where the differences are bounded by another function, which satisfies certain moment conditions (sub-Weibul). The proof consists of ... | 1 | 6 | 1,698,604,069,652 | Thank you for your review. We will answer all your questions.
Question 1: Issue in Summary ``It would be great to include some analysis of how optimal ...''
Answer: Yes, the bound provided by Theorem 1 is solely in terms
of $||F\_i (X\_i, X'\_i)||\_{\psi\_\alpha}$. It is possible to replace the orlizc norm with th... | [{"id":"YbOhUS81IG","parent_id":"Mhu5qkfBUx","created_at_ms":1700677921742,"modified_at_ms":1700677937541,"author":"Authors","role":"authors","type":"Author Reply","content":{"comment":"Thank you for your review. We will answer all your questions. \n\nQuestion 1: Issue in Summary ``It would be great to include some ana... | null | [] | ["Shaojie Li","Bowei Zhu","Yong Liu"] | ["Renmin University of China","Renmin University of China","Renmin University of China"] | null | null | null | 0 | unchanged |
0V311Uh8q1 | T5aXlwf2D3 | 2,024 | Algorithmic Stability Unleashed: Generalization Bounds with Unbounded Losses | One of the central problems of statistical learning theory is quantifying the generalization ability of learning algorithms within a probabilistic framework. Algorithmic stability is a powerful tool for deriving generalization bounds, however, it typically builds on a critical assumption that losses are bounded. In thi... | ["algorithmic stability","generalization bound"] | summary:
The paper studies high-probability generalization bounds for stable algorithms with unbounded loss functions. In particular, the paper considers a symmetric and totally Lipschitz stable algorithm, where the change of the output model with respect to the perturbation of a single example follows from a subweibul... | 1 | 5 | 1,698,114,916,310 | Part I
Thanks for your time and thoughts. We will answer all your questions.
Question 1: Issue in Weakness ``The high-probability bound established in the paper is ...''
Answer: Recently, (Bousquet et al. 2020) provided a sharper generalization bound of the order $O(1/\sqrt{n} + \gamma \log n)$ for algorithmic stabi... | [{"id":"rhAfontDV1","parent_id":"T5aXlwf2D3","created_at_ms":1700678054280,"modified_at_ms":1700678054280,"author":"Authors","role":"authors","type":"Author Reply","content":{"title":"Part I","comment":"Thanks for your time and thoughts. We will answer all your questions. \n\nQuestion 1: Issue in Weakness ``The high-pr... | null | [] | ["Shaojie Li","Bowei Zhu","Yong Liu"] | ["Renmin University of China","Renmin University of China","Renmin University of China"] | null | null | null | 0 | unchanged |
0V311Uh8q1 | cBUksJeqFP | 2,024 | Algorithmic Stability Unleashed: Generalization Bounds with Unbounded Losses | One of the central problems of statistical learning theory is quantifying the generalization ability of learning algorithms within a probabilistic framework. Algorithmic stability is a powerful tool for deriving generalization bounds, however, it typically builds on a critical assumption that losses are bounded. In thi... | ["algorithmic stability","generalization bound"] | summary:
Algorithmic stability has increasingly become an important tool for studying the generalization ability of machine learning algorithms. The general idea is to understand how well algorithms generalize out-of-sample by looking that how insensitive ('stable") they are to perturbations of the input dataset. Intui... | 1 | 5 | 1,699,118,183,440 | Part I
Thank you for your time and thoughts. Below, we reply to all the issues.
Question 1: One point that would be nice to emphasize the technical novelties in the proof...
Answer: Thanks for this review. The first challenge is that we deal with the concentration of the general function $f$. Actually, the related w... | [{"id":"Z2CFdJ153y","parent_id":"cBUksJeqFP","created_at_ms":1700676768212,"modified_at_ms":1700676768212,"author":"Authors","role":"authors","type":"Author Reply","content":{"title":"Part I","comment":"Thank you for your time and thoughts. Below, we reply to all the issues. \n\nQuestion 1: One point that would be nice... | null | [] | ["Shaojie Li","Bowei Zhu","Yong Liu"] | ["Renmin University of China","Renmin University of China","Renmin University of China"] | null | null | null | 0 | unchanged |
0V311Uh8q1 | tAKZoji5eT | 2,024 | Algorithmic Stability Unleashed: Generalization Bounds with Unbounded Losses | One of the central problems of statistical learning theory is quantifying the generalization ability of learning algorithms within a probabilistic framework. Algorithmic stability is a powerful tool for deriving generalization bounds, however, it typically builds on a critical assumption that losses are bounded. In thi... | ["algorithmic stability","generalization bound"] | summary:
The paper studies algorithmic stability and generalization bounds for unbounded losses of independent subweibull random variables. Specifically, the concentration inequality for subweibull random variables (which covers subexponential or subgaussian random variables studied in the previous works) is establishe... | 1 | 3 | 1,698,830,219,863 | Part I
Thank you for your review. Here are the replies for your questions.
Question 1: As mentioned in weakness 1, it seems that the proofs ...
Answer: We believe our work is not incremental. Here, we discuss the technical novelties of Theorem 1. The first challenge to prove Theorem 1 is that we deal with the concent... | [{"id":"v6PWmKWQCn","parent_id":"tAKZoji5eT","created_at_ms":1700676953369,"modified_at_ms":1700676953369,"author":"Authors","role":"authors","type":"Author Reply","content":{"title":"Part I","comment":"Thank you for your review. Here are the replies for your questions.\n\nQuestion 1: As mentioned in weakness 1, it see... | null | [] | ["Shaojie Li","Bowei Zhu","Yong Liu"] | ["Renmin University of China","Renmin University of China","Renmin University of China"] | null | null | null | 0 | unchanged |
0V5TVt9bk0 | 15nKpcXvUB | 2,024 | Q-Bench: A Benchmark for General-Purpose Foundation Models on Low-level Vision | The rapid evolution of Multi-modality Large Language Models (MLLMs) has catalyzed a shift in computer vision from specialized models to general-purpose foundation models. Nevertheless, there is still an inadequacy in assessing the abilities of MLLMs on **low-level visual perception and understanding**. To address this ... | ["Benchmark","Vision-Language","Large Language Models","Low-level Vision","Image Quality Assessment"] | summary:
This work introduces a benchmark for low-level perception and understanding. They introduce three tasks for assessing the MLLM. The main focus is low-level, thus they build an evaluation bench from various aspects (low-level attributes, visual distortions, etc.). The experiments are extensive and comprehensive... | 1 | 6 | 1,699,111,065,227 | (Thread 1/3) Thank you for your recognition and questions. Here are some clarifications to question 1.
[Thread 1/3]
## General Response
We would like to sincerely thank the reviewer for recognition on our motivation, organization, and experimental soundness. We also sincerely appreciate the profound questions as rais... | [{"id":"dbC6H7WIgU","parent_id":"15nKpcXvUB","created_at_ms":1699716537088,"modified_at_ms":1699719088288,"author":"Authors","role":"authors","type":"Author Reply","content":{"title":"(Thread 1/3) Thank you for your recognition and questions. Here are some clarifications to question 1.","comment":"[Thread 1/3]\n\n## Ge... | Thank you for your efforts and valuable suggestions!
We would like to sincerely express our gratitude to all reviewers for the general recognition on our work, as all reviewers recommend towards acceptance for this work.
More importantly, we also notice that there are several points that reviewers would like us to cl... | [{"id":"a5KZ6hiRex","parent_id":"0V5TVt9bk0","created_at_ms":1699728701552,"modified_at_ms":1699728701552,"author":"Authors","role":"authors","type":"Author Reply","content":{"title":"Thank you for your efforts and valuable suggestions!","comment":"We would like to sincerely express our gratitude to all reviewers for t... | ["Haoning Wu","Zicheng Zhang","Erli Zhang","Chaofeng Chen","Liang Liao","Annan Wang","Chunyi Li","Wenxiu Sun","Qiong Yan","Guangtao Zhai","Weisi Lin"] | ["Nanyang Technological University","Nanyang Technological University","Nanyang Technological University","Nanyang Technological University","Nanyang Technological University","Shanghai Artificial Intelligence Laboratory","SenseTime Group Limited","SenseTime Research","Shanghai Jiaotong University","Nanyang Technologic... | 2309.14181 | cs.CV cs.AI cs.MM | 1,695,653,023,000 | 1 | unchanged |
0V5TVt9bk0 | 2RYXJp5YRH | 2,024 | Q-Bench: A Benchmark for General-Purpose Foundation Models on Low-level Vision | The rapid evolution of Multi-modality Large Language Models (MLLMs) has catalyzed a shift in computer vision from specialized models to general-purpose foundation models. Nevertheless, there is still an inadequacy in assessing the abilities of MLLMs on **low-level visual perception and understanding**. To address this ... | ["Benchmark","Vision-Language","Large Language Models","Low-level Vision","Image Quality Assessment"] | summary:
The rapidly developed Multi-modality Large Language Models (MLLMs) have progressed greatly. However, there is still an inadequacy in assessing the abilities of MLLMs on low-level visual perception and understanding. Accordingly, this paper proposed Q-Bench, a holistic benchmark of MLLMs for low-level visual pe... | 1 | 8 | 1,698,857,391,135 | (Thread 4/4) Thank you for your kind support and suggestions!
[Thread 4/4]
## Question 2 (Q2)
As a benchmark paper, the author may update the recent SOTA MLLMs into the leadboards of this paper, e.g., QWen-VL [1], InternLM-XComposer [2], and LLaVA-1.5 [3].
## Answer for Q2 (Part 3)
### Updated Results for (A3): As... | [{"id":"vcKYSHnSVg","parent_id":"2RYXJp5YRH","created_at_ms":1699695605680,"modified_at_ms":1699801459169,"author":"Authors","role":"authors","type":"Author Reply","content":{"title":"(Thread 4/4) Thank you for your kind support and suggestions!","comment":"[Thread 4/4]\n\n## Question 2 (Q2)\n\nAs a benchmark paper, th... | Thank you for your efforts and valuable suggestions!
We would like to sincerely express our gratitude to all reviewers for the general recognition on our work, as all reviewers recommend towards acceptance for this work.
More importantly, we also notice that there are several points that reviewers would like us to cl... | [{"id":"a5KZ6hiRex","parent_id":"0V5TVt9bk0","created_at_ms":1699728701552,"modified_at_ms":1699728701552,"author":"Authors","role":"authors","type":"Author Reply","content":{"title":"Thank you for your efforts and valuable suggestions!","comment":"We would like to sincerely express our gratitude to all reviewers for t... | ["Haoning Wu","Zicheng Zhang","Erli Zhang","Chaofeng Chen","Liang Liao","Annan Wang","Chunyi Li","Wenxiu Sun","Qiong Yan","Guangtao Zhai","Weisi Lin"] | ["Nanyang Technological University","Nanyang Technological University","Nanyang Technological University","Nanyang Technological University","Nanyang Technological University","Shanghai Artificial Intelligence Laboratory","SenseTime Group Limited","SenseTime Research","Shanghai Jiaotong University","Nanyang Technologic... | 2309.14181 | cs.CV cs.AI cs.MM | 1,695,653,023,000 | 1 | unchanged |
0V5TVt9bk0 | dXFHxuKQsB | 2,024 | Q-Bench: A Benchmark for General-Purpose Foundation Models on Low-level Vision | The rapid evolution of Multi-modality Large Language Models (MLLMs) has catalyzed a shift in computer vision from specialized models to general-purpose foundation models. Nevertheless, there is still an inadequacy in assessing the abilities of MLLMs on **low-level visual perception and understanding**. To address this ... | ["Benchmark","Vision-Language","Large Language Models","Low-level Vision","Image Quality Assessment"] | summary:
This is a benchmark paper on the performance of low-level visual perception and understanding of MLLMs. To this end, the paper collected and annotated a new benchmark dataset Q-Bench including 1) LLVisionQA about the low-level perception of 2990 images; 2) LLDescribe about the description of image quality of 4... | 1 | 6 | 1,698,762,994,673 | Thank you for your recognition and advice! Here is GPT-4V (and human) performance on Q-Bench.
We sincerely thank you for the reviewer's recognition on our work, and highly agree with the concerns and suggestions. The additional results as required are provided as follows.
## Question:
It is definitely a must-to-ask qu... | [{"id":"wTjJtBNZ1b","parent_id":"dXFHxuKQsB","created_at_ms":1699691780438,"modified_at_ms":1699691780438,"author":"Authors","role":"authors","type":"Author Reply","content":{"title":"Thank you for your recognition and advice! Here is GPT-4V (and human) performance on Q-Bench.","comment":"We sincerely thank you for the... | Thank you for your efforts and valuable suggestions!
We would like to sincerely express our gratitude to all reviewers for the general recognition on our work, as all reviewers recommend towards acceptance for this work.
More importantly, we also notice that there are several points that reviewers would like us to cl... | [{"id":"a5KZ6hiRex","parent_id":"0V5TVt9bk0","created_at_ms":1699728701552,"modified_at_ms":1699728701552,"author":"Authors","role":"authors","type":"Author Reply","content":{"title":"Thank you for your efforts and valuable suggestions!","comment":"We would like to sincerely express our gratitude to all reviewers for t... | ["Haoning Wu","Zicheng Zhang","Erli Zhang","Chaofeng Chen","Liang Liao","Annan Wang","Chunyi Li","Wenxiu Sun","Qiong Yan","Guangtao Zhai","Weisi Lin"] | ["Nanyang Technological University","Nanyang Technological University","Nanyang Technological University","Nanyang Technological University","Nanyang Technological University","Shanghai Artificial Intelligence Laboratory","SenseTime Group Limited","SenseTime Research","Shanghai Jiaotong University","Nanyang Technologic... | 2309.14181 | cs.CV cs.AI cs.MM | 1,695,653,023,000 | 1 | changed |
0VKEJKKLvr | 3OodMu0IFq | 2,024 | A GRAPH-BASED REPRESENTATION LEARNING APPROACH FOR BREAST CANCER RISK PREDICTION USING GENOTYPE DATA | Breast cancer risk prediction using genotype data is a critical task in personalized medicine. However, the high dimensionality and potential redundancy of genetic features pose challenges for accurate risk prediction. We present a graph-based representation learning pipeline for breast cancer risk prediction. Our meth... | ["Graph representation","Deep learning","Single nucleotide polymorphism","Breast cancer"] | summary:
The authors presented a graph-based representation learning framework for breast cancer risk prediction using genetic data. They selected informative SNPs by an ensemble approach aiming to capture non-linear and high-dimensional SNP-SNP interactions that aren't possible with linear feature selection approaches... | 0 | 3 | 1,699,678,357,927 | Did the authors control for independence between the training and testing sets, given that they were from the same cohort?
Thanks for the comment. Yes, the training and test sets are independent and do not overlap. To create the independence between the training and testing sets, we used a stratified splitting protocol... | [{"id":"nwIUbg4u8j","parent_id":"3OodMu0IFq","created_at_ms":1699966620653,"modified_at_ms":1699966620653,"author":"Authors","role":"authors","type":"Author Reply","content":{"title":"Did the authors control for independence between the training and testing sets, given that they were from the same cohort?","comment":"T... | null | [] | ["Naga Raju Gudhe","veli-matti Kosma","Hamid Behravan","Arto Mannermaa"] | ["University of Eastern Finland","uef",null,null] | null | null | null | 0 | unchanged |
0VZP2Dr9KX | b3G2fW3ng6 | 2,024 | Baseline Defenses for Adversarial Attacks Against Aligned Language Models | As large language models (LLMs) quickly become ubiquitous, it becomes critical to understand their security vulnerabilities.
Recent work shows that text optimizers can produce jailbreaking prompts that bypass moderation and alignment.
Drawing from the rich body of work on adversarial machine learning, we approach thes... | ["baseline defenses","attacks"] | summary:
This paper presents a study of the vulnerabilities associated with large language models (LLMs), specifically focusing on jailbreaking attacks that aim to bypass alignment and moderation mechanisms. The work mainly adopts an empirical approach, evaluating several baseline defenses under different threat models... | 1 | 5 | 1,698,676,117,473 | Response to Reviewer UELn
Thank you for your time, UELn!
> future directions
In the last paraphrase of the conclusion, we list the future directions of the paper. We highlight the following questions
"...leaves us with a number of big open questions. (i) What defenses can be reliably deployed, with only minimal imp... | [{"id":"VwXgziQiqX","parent_id":"b3G2fW3ng6","created_at_ms":1700151268042,"modified_at_ms":1700151330778,"author":"Authors","role":"authors","type":"Author Reply","content":{"title":"Response to Reviewer UELn","comment":"Thank you for your time, UELn!\n\n> future directions\n\nIn the last paraphrase of the conclusion,... | null | [] | ["Neel Jain","Avi Schwarzschild","Yuxin Wen","Gowthami Somepalli","John Kirchenbauer","Ping-yeh Chiang","Micah Goldblum","Aniruddha Saha","Jonas Geiping","Tom Goldstein"] | ["University of Maryland, College Park","Carnegie Mellon University","University of Maryland, College Park","University of Maryland, College Park","University of Maryland, College Park","University of Maryland, College Park","New York University","University of Maryland, College Park","Max Planck Institute for Intellig... | 2309.00614 | cs.LG cs.CL cs.CR | 1,693,591,184,000 | 1 | unchanged |
0VZP2Dr9KX | bMyFzthtbJ | 2,024 | Baseline Defenses for Adversarial Attacks Against Aligned Language Models | As large language models (LLMs) quickly become ubiquitous, it becomes critical to understand their security vulnerabilities.
Recent work shows that text optimizers can produce jailbreaking prompts that bypass moderation and alignment.
Drawing from the rich body of work on adversarial machine learning, we approach thes... | ["baseline defenses","attacks"] | summary:
Extending the studies in standard adversarial machine learning domains, this paper explores the effectiveness of three different mainstream defense methods against jailbreak attacks for LLM, including perplexity detection, input reprocessing via paraphrasing/retokenization, and adversarial training (more preci... | 1 | 5 | 1,698,721,430,040 | Response to Reviewer hCXb
Thank you for your time, hCXb!
> non-universal GCG attacks
We have now added a non-universal threat model, where the attacker has again access to the model weights, but now optimizes the attack for a single string. As before, the attacker does not have access to the system, so the defender c... | [{"id":"YLh0noHRai","parent_id":"bMyFzthtbJ","created_at_ms":1700150975101,"modified_at_ms":1700150975101,"author":"Authors","role":"authors","type":"Author Reply","content":{"title":"Response to Reviewer hCXb","comment":"Thank you for your time, hCXb!\n\n> non-universal GCG attacks\n\nWe have now added a non-universal... | null | [] | ["Neel Jain","Avi Schwarzschild","Yuxin Wen","Gowthami Somepalli","John Kirchenbauer","Ping-yeh Chiang","Micah Goldblum","Aniruddha Saha","Jonas Geiping","Tom Goldstein"] | ["University of Maryland, College Park","Carnegie Mellon University","University of Maryland, College Park","University of Maryland, College Park","University of Maryland, College Park","University of Maryland, College Park","New York University","University of Maryland, College Park","Max Planck Institute for Intellig... | 2309.00614 | cs.LG cs.CL cs.CR | 1,693,591,184,000 | 1 | unchanged |
0VZP2Dr9KX | hsBW78fLsG | 2,024 | Baseline Defenses for Adversarial Attacks Against Aligned Language Models | As large language models (LLMs) quickly become ubiquitous, it becomes critical to understand their security vulnerabilities.
Recent work shows that text optimizers can produce jailbreaking prompts that bypass moderation and alignment.
Drawing from the rich body of work on adversarial machine learning, we approach thes... | ["baseline defenses","attacks"] | summary:
This paper studied jailbreaking prompts generated by text optimizers that bypass alignment, discussed the effectiveness of threat models, performance of baseline defense techniques, and the differences between LLM security and computer vision. This paper studied detection based on perplexity, input preprocessi... | 1 | 8 | 1,698,710,725,911 | Response to Reviewer Ywox
Thank you for your time, Reviewer Ywox!
> threat model is limited
We have now added an additional non-universal threat model, where the attacker has again access to the model weights, but now optimizes an attack for a single target string. As before, the attack does not have access to the sy... | [{"id":"cksDpmOGYd","parent_id":"hsBW78fLsG","created_at_ms":1700151165765,"modified_at_ms":1700151165765,"author":"Authors","role":"authors","type":"Author Reply","content":{"title":"Response to Reviewer Ywox","comment":"Thank you for your time, Reviewer Ywox!\n\n> threat model is limited\n\nWe have now added an addit... | null | [] | ["Neel Jain","Avi Schwarzschild","Yuxin Wen","Gowthami Somepalli","John Kirchenbauer","Ping-yeh Chiang","Micah Goldblum","Aniruddha Saha","Jonas Geiping","Tom Goldstein"] | ["University of Maryland, College Park","Carnegie Mellon University","University of Maryland, College Park","University of Maryland, College Park","University of Maryland, College Park","University of Maryland, College Park","New York University","University of Maryland, College Park","Max Planck Institute for Intellig... | 2309.00614 | cs.LG cs.CL cs.CR | 1,693,591,184,000 | 1 | unchanged |
0VZP2Dr9KX | iXM1bi7Acs | 2,024 | Baseline Defenses for Adversarial Attacks Against Aligned Language Models | As large language models (LLMs) quickly become ubiquitous, it becomes critical to understand their security vulnerabilities.
Recent work shows that text optimizers can produce jailbreaking prompts that bypass moderation and alignment.
Drawing from the rich body of work on adversarial machine learning, we approach thes... | ["baseline defenses","attacks"] | summary:
The paper discusses simple baseline defenses against a single gradient-based adversarial attacks against Large Language Models (LLMs). These include perplexity based filtering, paraphrasing via another LLM, and an attempt at adversarial training.
strengths:
Some of the results (in particular, regarding perple... | 1 | 3 | 1,698,791,742,996 | Response to Reviewer 9hm2 (part 1/2)
Thank you for your time, 9hm2!
> Single-Attack Study Cannot Generalize -- conclusions limited from studying one attack threat model
We do measure defense efficacy on only one attack, we have modified the claims in the first paragraph of Section 4 to be more clear to reflect y... | [{"id":"LLipI9rbrs","parent_id":"iXM1bi7Acs","created_at_ms":1700150742673,"modified_at_ms":1700150742673,"author":"Authors","role":"authors","type":"Author Reply","content":{"title":"Response to Reviewer 9hm2 (part 1/2)","comment":"Thank you for your time, 9hm2!\n\n> Single-Attack Study Cannot Generalize -- conclusio... | null | [] | ["Neel Jain","Avi Schwarzschild","Yuxin Wen","Gowthami Somepalli","John Kirchenbauer","Ping-yeh Chiang","Micah Goldblum","Aniruddha Saha","Jonas Geiping","Tom Goldstein"] | ["University of Maryland, College Park","Carnegie Mellon University","University of Maryland, College Park","University of Maryland, College Park","University of Maryland, College Park","University of Maryland, College Park","New York University","University of Maryland, College Park","Max Planck Institute for Intellig... | 2309.00614 | cs.LG cs.CL cs.CR | 1,693,591,184,000 | 1 | unchanged |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.