title string | paper_decision string | review_1 string | rebuttals_1 string | review_2 string | rebuttals_2 string | review_3 string | rebuttals_3 string | review_4 string | rebuttals_4 string | global_rebuttals string | dataset_source string | conference_year int64 | review_5 string | rebuttals_5 string | review_6 string | rebuttals_6 string | review_7 string | rebuttals_7 string | review_8 string | rebuttals_8 string |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
On the Generalization Error of Stochastic Mirror Descent for Quadratically-Bounded Losses: an Improved Analysis | Accept (poster) | Summary: This paper analyzes the generalization bound for stochastic mirror descent applied to linear models and a general class of quadratically-bounded losses. In addition to the more common IID sampled data, this paper also considers where the data is sampled from a Markov chain. This paper follows up from the work ... | Rebuttal 1:
Rebuttal: We thank the reviewer for the valuable feedback.
**Regarding the comparison between our results and [T22]**: Both theorems 3 and 13 in our paper improve the guarantee of theorems 5 and 8 in [T22] by a $\log T$ factor. This is also the case of Theorem 7 compared with Theorem 10 part 1 for sub-gau... | Summary: The authors study Stochastic Mirror Descent, and analysis its generalization error under a quadratic loss assumptions, self boundedness of the loss (a relaxation of smoothness), iid or Markovian data, and realizable or non-realizable cases, which amount to assume that the error of the reference point is of dif... | Rebuttal 1:
Rebuttal: We thank the reviewer for the valuable feedback and for pointing out the references that we missed in our submission. We will include them in the revision.
**Regarding [A]**: We agree that this paper can be seen in part as giving an in-expectation generalization error bound for a setting that cor... | Summary: The authors propose a novel approach to analyze the generalization error of Stochastic Mirror Descent (SMD) for quadratically bounded losses. Compared to previous methods, their approach demonstrates improved accuracy by a logarithmic factor when the data is realizable. Additionally, in scenarios where the dat... | Rebuttal 1:
Rebuttal: We thank the reviewer for the feedback. We answer the concerns of the reviewer below.
1. The main contribution in the realizable case is not only an improvement in the log factors, but also a new approach to analyze generalization bounds in high probability of stochastic gradient methods. As note... | Summary: This paper revisits the generalization error of stochastic mirror descent for quadratically bounded losses studied in Telgarsky (2022). The authors depart significantly from the prior work of Telgarsky (2022), and introduce a novel approach for establishing high probability generalization guarantees. In contra... | Rebuttal 1:
Rebuttal: We thank the reviewer for the feedback and the endorsement of our novel approaches.
Regarding the question: We believe that line 391-line 392 are correct and consistent with our notations. We recall some notations to clarify the assumptions. We use $\ell_{t}(w)$ as the short form for $\ell(y_{t},... | null | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
The probability flow ODE is provably fast | Accept (poster) | Summary: In this work, authors propose improved solution scheme with the specially chosen corrector step for the the score-based generative models. Within the denoising diffusion probabilistic modeling (DDPM) and the underdamped Langevin diffusion approach authors obtain better dimension dependence within certain restr... | Rebuttal 1:
Rebuttal: Thanks for your positive comments.
**I am confused by such a weak study of the experimental part**
Please refer to our answer to all reviewers above.
As mentioned in the paper: “Although we provide preliminary numerical experiments in the Appendix, we leave it as a question for future work to ... | Summary: This paper provides polynomial-time convergence guarantees for the probability flow ODE implementation. This paper obtains a better better dimension dependence than prior works on DDPM. However, there is no experiments to prove the effectiveness of the proposed methods.
Strengths: 1. The paper is well written... | Rebuttal 1:
Rebuttal: Thank you for the review.
**No experiments are performed to prove the proposed method.**
We experimented our main algorithm DPUM in the appendix on a toy model, and **the Python code is included in the Supplementary material.** The other algorithm we analyze in the paper, DPOM, was already exp... | Summary: The authors proposed the convergence guarantees for the probability flow ODE implementation. Through the use of a specially chosen corrector step based on the underdamped Langevin diffusion, a better dimension dependence of $O(\sqrt{d})$ than DDPM $O(d)$ is achieved.
Strengths: 1. the theoretical comparison b... | Rebuttal 1:
Rebuttal: Thank you for your positive review.
**The term of fast looks a bit confusing to me. It is fast in inference or training?**
Our theory shows that sampling is fast, therefore the term fast refers to the inference time, not the training time. We will clarify this point in the introduction.
**Flow-... | Summary: The current study offers a polynomial-time total variation convergence guarantee for score-based generative modeling, given that the generative process is a probabilistic Ordinary Differential Equation (ODE). This differs from previous work, which typically involves a reverse Stochastic Differential Equation (... | Rebuttal 1:
Rebuttal: Thank you very much for the generous review.
**The inclusion of an additional stochastic corrector in the generative process is somewhat unsatisfying [...]**
Regarding the use of the corrector, we are not the first ones to propose this; in fact, it was shown to achieve superior performances in ... | Rebuttal 1:
Rebuttal: **TO ALL REVIEWERS and AC:**
We thank all of the reviewers for their hard work in reviewing our article. All reviewers are in agreement regarding the high quality and novelty of our theoretical results, and that our results are of interest to the NeurIPS community. In particular, we leverage rece... | NeurIPS_2023_submissions_huggingface | 2,023 | Summary: In this article the authors propose polynomial-time convergence guarantees for probability flow ODEs related to score based generative models.
Previous analyses on score based diffusion models have resorted to the use of SDEs and relied on properties of stochastic flows to alleviate error accumulation due to ... | Rebuttal 1:
Rebuttal: Thank you for the detailed review.
**They propose two variants of the algorithm, one relying on overdamped Langevin dynamics with dimension dependence that scales as $O(d)$, and one with underdamped Langevin dynamics that scales more favourably as $O(\sqrt{d})$ and thus has improved dimension dep... | null | null | null | null | null | null |
Towards Free Data Selection with General-Purpose Models | Accept (poster) | Summary: This paper presents a data selection methods without training, so saving lots of time than previous methods. By a publicly available pretrained model, FreeSel can select data from datasets with a single-pass inference, without need for additional training. This method uses comparison between semantic patterns ... | Rebuttal 1:
Rebuttal: Thank you for your efforts to improve the quality of our papers. We thank you for pointing out your concerns, which we believe have been carefully clarified and addressed, as follows.
**Q1: [Our method is plain and cannot outperform all active learning algorithms.]**
Thank you for the question.... | Summary: This work introduces a method named FreeSel, which is designed to efficiently select information-rich data points from an unlabeled pool. These selected data points can then be annotated and utilized for training deep learning models, thereby mitigating the costs and resources needed to annotate an entire data... | Rebuttal 1:
Rebuttal: Thank you for the positive comments and insightful suggestions. The concerns are answered as follows.
**Q1: [Our method is more like subset selection instead of active learning.]**
+ We totally agree with your points. Our method should be considered as subset selection more than active learning.... | Summary: This paper aims to introduce a unique method called FreeSel for data selection. This method is accompanied by a newly designed pipeline that incorporates existing general-purpose models, requiring minimal additional time. By leveraging a publicly available pre-trained model, FreeSel can select data from divers... | Rebuttal 1:
Rebuttal:
Thank you for the positive comments and insightful suggestions. The concerns are answered as follows.
**Q1: [Performance decreases as the sample number increases. Sensitivity to the dataset size.]**
- Existing active learning methods train their models on the selected samples after each batch s... | Summary: This paper proposed a diversity-based method for active learning. Given an image, features are extracted from a pretrained visual transformer model (DeiT-S), features with [CLS] token attention above a threshold are kept and clustered into K clusters. Each cluster is represented by the average of features belo... | Rebuttal 1:
Rebuttal:
Thanks for your efforts to improve the quality of our papers. We thank you for pointing out your concerns, which we believe have been carefully clarified and addressed, as follows.
**Q1: [Outperformed by some active learning methods in late stage.]**
Good question! This observation is reasonabl... | Rebuttal 1:
Rebuttal:
We sincerely appreciate all reviewers’ time and efforts in reviewing our paper. We are glad to find that reviewers generally recognized our contributions including the novelty of our proposed pipeline (Reviewers SbPr,7NQc), the efficiency of our method (Reviewers xhjx, SbPr, yLeE), detailed exper... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
LIMA: Less Is More for Alignment | Accept (poster) | Summary: In this paper, the authors present an LLM finetuned from 65 LLaMa with very limited 1000 samples. LIMA achieves very impressive results in the human evaluation, with competitive results against GPT4/Bard/davinci003. According to my understanding, the main contribution of this paper lies on two aspects: 1. veri... | Rebuttal 1:
Rebuttal: Question 1: While collecting prompts, Group B assigned each prompt in the test set to a category based on their own definitions. Here are the categories and distribution of the prompts in the test set:
- writing: 13.7%
- coding and stem: 9%
- knowledge: 18.7%
- advice: 18.7%
- commonsense: 2.7%
-... | Summary: This paper presents a superficial alignment hypothesis. It claims that most knowledge and abilities of large language models (LLMs) are learned in the pre-training stage, and a few datasets are just enough to transform the styles of LLMs for alignment.
To support the argument, the authors of the paper finetu... | Rebuttal 1:
Rebuttal: Weakness A. While collecting prompts, Group B assigned each prompt in the test set to a category based on their own definitions. This is a noisy process, and the categories are not entirely well-defined, which is why we did not share this information in the paper. However, per your request, we sha... | Summary: The authors investigate on instruction tuning an LLM on high-quality manually collected/created data (only 1K training examples). They provide detailed explanation on how to collect the data, tune a 65B LLaMA model and compare it with other commercial/open-source LLMs. The main experiment results show that the... | Rebuttal 1:
Rebuttal: Weakness 1: The test set does indeed include naturally occurring prompts from Reddit.
Weakness 2: We agree that there are limitations to evaluating models by human preferences. However, there are even more significant limitations to evaluating models by other existing benchmarks, e.g. MMLU, which... | Summary: This paper presents LIMA, a 65 billion-parameter large language model fine-tuned on a dataset of 1,000 carefully curated prompts and responses without the use of reinforcement learning or human preference modeling. The aim of this research was to measure the relative importance of pretraining and large-scale i... | Rebuttal 1:
Rebuttal: - Weakness1: We intentionally did not evaluate using MMLU and other multiple-choice or classification benchmarks since they are unrelated to the functionality that we tried to imbue in the model. The experiments in LIMA show that an assistant (chatbot) can be trained from a strong pretrained model... | Rebuttal 1:
Rebuttal: ## General Response
A general concern among reviewers is that the human evaluation in LIMA is limited, and that the paper should also include automatic benchmark evaluations. While we agree that there are limitations to evaluating models by human preferences, we argue that there are even more sig... | NeurIPS_2023_submissions_huggingface | 2,023 | Summary: This paper hypothesizes that alignment of pretrained LLMs to become user friendly and useful AI chatbots is mainly about learning the style or format of interacting with users, and that all/most knowledge and capabilities were already acquired by the LLM during pretraining. To test this hypothesis, this paper ... | Rebuttal 1:
Rebuttal: Thank you for pointing out this potential issue!
We acknowledge the reviewer’s concern regarding potential similarities between Group A and Group B data, and provide the test set results (LIMA win / tie / LIMA lose), split 230 / 70 by prompt source, demonstrating that the trends are not over-spec... | null | null | null | null | null | null |
PromptIR: Prompting for All-in-One Image Restoration | Accept (poster) | Summary: The paper proposes PromptIR, a prompt-based learning approach for all-in-one image restoration. The method utilizes prompts to encode degradation-specific information and dynamically guide the restoration network. It achieves state-of-the-art results on various image restoration tasks.
Strengths: (1) The prop... | Rebuttal 1:
Rebuttal: `R5.1`: *Ablations on number of prompt blocks and multi-degraded images*:
The number of prompt components is chosen empirically. Please refer to response 2.4 (Table R3). And for a detailed response on images corrupted with multiple degradations, please see R1.3.
`R5.2`: *Learned prompts*:
We pe... | Summary: This paper proposes an all-in-one image restoration method that can restore images from various types and levels of degradation. Specifically, it utilizes prompts to encode degradation-specific information, which is used to dynamically guide the restoration network. Experimental evaluations on image denoising,... | Rebuttal 1:
Rebuttal: `R4.1`: *Degradation embeddings*:
In Figure 2, we plot the prompt P ,generated through the Prompt Generation Module within the Prompt block. Here, the term ”degradation embedding” is used to refer to the generated prompt P, as it provides
additional contextual information regarding the type of de... | Summary: In this paper, authors study to restore images with unknown degradation types with a unified framework. To achieve it, a prompt-based block is proposed which generates suitable prompt first and then interact them with original features. Experiments show that combining the proposed block with restormer network ... | Rebuttal 1:
Rebuttal: `R3.1`: *Typos*:
We thank the reviewer for pointing out these issues. We will improve clarity and fix typos in the revised manuscript.
`R3.2`: *Additional Comparisons*:
As suggested, we include Restormer and Transweather in comparisons. Please refer to Table R1. Our PromptIR provides favorable ... | Summary: This paper restores images from various types and levels of degradation. In detail, degradation information is encoded as prompts and used to guidance the restoration. Based on the classical UNet architecture for restoration, there are two added blocks: prompt generation module and prompt interaction module. W... | Rebuttal 1:
Rebuttal: `R2.1`: *Ablation study (directly incorporating softmax outputs):*
To show the benefit and contribution coming from only prompts in our design, we perform an ablation experiment in which we adjust the baseline Restormer's capacity such that the params/FLOPs become similar to our PromptIR framewor... | null | NeurIPS_2023_submissions_huggingface | 2,023 | Summary: The paper proposes prompt based image restoration method that can handle different degradations. the proposed method uses prompts to encode degradation-specific information, which is then used to dynamically guide the restoration network. Given any instance, prompt generation module (PGM) and a prompt interact... | Rebuttal 1:
Rebuttal: `R1.1`:*Other Degradations*:
We opted for deraining, dehazing and denoising to demonstrate the effectiveness of our approach on a diverse range of low-level vision tasks. However, as per the reviewer suggestions, we retrained our PromptIR network on five distinct tasks, including two new tasks i.... | null | null | null | null | null | null |
DropCompute: simple and more robust distributed synchronous training via compute variance reduction | Accept (poster) | Summary: This paper considers a typical scenario in which workers are straggling due to variability in compute time, in which the authors find an analytical relation between compute time properties and scalability limitations. They propose a decentralized method to reduce the variation among workers and improve the r... | Rebuttal 1:
Rebuttal: ### Weakness 1 response
We prove convergence under the standard assumptions that are used in the analysis of distributed SGD. To be concrete, we make 2 assumptions: Smoothness, and bounded variance, both of them are standard in the analysis of distributed/parallel SGD, see e.g. the classical paper... | Summary: The authors proposed a method to drop computation (DropCompute) of **microbatches** of a **minibatch** on decentralized workers if their execution time exceeds a certain threshhold. The author provided theoretical convergence proof of this method as well as theoretical/experimental study of the performance gai... | Rebuttal 1:
Rebuttal: ### Weakness 1 response
It is true that some learning tasks do not use many microbatches, however, we argue that for LLMs the number of microbatches can be significantly larger. For example, the recent leading performance benchmark for LLMs, GPT-3 175B training with 3584 workers (including pipelin... | Summary: This paper introduces DropCompute, a method to mitigate load imbalance across workers when doing distributed training. It splits mini-batches into micro-batches, and uses a dynamically-calculated time budget, which, when exceeded, results in the remaining micro-batches being dropped and communication to aggreg... | Rebuttal 1:
Rebuttal: ### Weakness 1 response
We thank the reviewer for the suggestion to add more concrete use case examples. Heterogeneity in computation causes issues when training in parallel schemes, and so many such workloads rely on “tricks” (such as packing of sentences and padding) in order to reduce computati... | Summary: This paper proposes a strategy to improve the scalability of DNN training for highly distributed workloads by allowing nodes that are lagging behind, in terms of processing time, to stop computing updates and share partial results with the other nodes in the cluster. The proposed method, DropCompute, allows th... | Rebuttal 1:
Rebuttal: ### Weakness 1 response
We appreciate the reviewer's concern regarding the limited audience that may directly benefit from DropCompute. However, the growing investment in large-scale training environments and the potential impact on efficiency justifies our focus on optimizing such settings. Also,... | Rebuttal 1:
Rebuttal: **We thank the reviewers for the insightful and constructive feedback. The evaluations provided in the paper will be improved along with the required modifications and clarifications. We will be happy for further questions and feedback.**
We thank reviewer Zc33's recognition of the novelty and im... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Graph Mixture of Experts: Learning on Large-Scale Graphs with Explicit Diversity Modeling | Accept (poster) | Summary: This paper considers the problem of graph neural networks. After discussing the drawback of existing methods, this paper presents a new model, called Graph Mixture of Expert (GMoE), which can dynamically select its own optimal information aggregation experts. Extensive experiments conducted on several graph d... | Rebuttal 1:
Rebuttal: Thank you for your supportive comments and insightful questions. We addressed each point in your comment carefully.
Question: Although the authors have shown the benefit of the proposed method over GCN, the improvement seems somewhat limited. For example, in several tables, the improvement over ... | Summary: This paper introduces the Graph Mixture of Expert (GMoE) model, enhancing Graph Neural Networks' (GNNs) ability to handle diverse graph structures without additional computational costs. GMoE allows nodes to select their information aggregation experts dynamically, catering to different subgroups of graph stru... | Rebuttal 1:
Rebuttal: Thank you for your thorough review and insightful comments. We carefully addressed every point raised in your feedback. We hope this can address any concerns you may have had and foster a more positive perspective on our paper.
Question: What is the motivation of the paper? What issue does the ar... | Summary: The paper claims that increasing the diversity of graph structures plays an important role in GNN training, which makes GNNs more robust and generalizable. To make GNN models leverage the structural diversity better, the paper proposes the Graph Mixture of Experts (GMoE). In each layer of GMoE, the representat... | Rebuttal 1:
Rebuttal: Thank you for your careful review and constructive comments. We carefully addressed your comments below.
Question. In the abstract, the paper argues that GMoE enables each node to select its own optimal information aggregation experts. But we do not see evidence why GMoE will find optimal experts... | Summary: This paper proposes a novel Graph Mixture of Experts (GMoE) architecture that enhances the ability of Graph Neural Networks (GNNs) to leverage and benefit from the structural diversity present in training graphs. The GMoE model comprises multiple "experts" at each layer, with each expert being an independent a... | Rebuttal 1:
Rebuttal: We greatly appreciate your recognition of the strengths in our paper and your valuable insights shared in your comments. In response, we have carefully addressed each of your points.
Question. While the proposed GMoE model is designed to avoid additional computational overhead, the paper does no... | null | NeurIPS_2023_submissions_huggingface | 2,023 | Summary: This paper tracks the problem of training a good GNN. To enhance the generalization ability of GNNs, it has become common to increase the diversity of training graph structures by incorporating graph augmentations and/or performing large-scale pre-training on more graphs. Therefore, it is essential for a GNN t... | Rebuttal 1:
Rebuttal: Thank you for your supportive comments and insightful questions. We addressed each point in your comment carefully.
Question: This paper can be improved further in terms of writing. You introduce two interesting modules but without good motivations. You can verify your motivation via conducting ... | null | null | null | null | null | null |
A Recurrent Neural Circuit Mechanism of Temporal-scaling Equivariant Representation | Accept (poster) | Summary: Empirical data have shown the existence of neurons that represent time and the ability of neural circuits to generate temporal sequences at different speeds. The authors propose a model that takes into account these findings: a continuous attractor network that acts as a neural clock which can evolve with a sp... | Rebuttal 1:
Rebuttal: We are grateful to your positive review and honored to answer your questions.
> “What does the Lie group perspective bring? To the best of my understanding, the Lie group perspective is not used in the analysis and requires introducing additional jargon that makes the paper harder to read. Can th... | Summary: This paper introduces a continuous attractor neural network that has a simple way of generating a specific sequence of outputs at a speed tunable by changing the gain of a constant (but neuron-dependent) input. The authors show how this can be used in conjunction with a multi-layer perceptron to generate arbit... | Rebuttal 1:
Rebuttal: Thanks for your positive comments about the strengths, especially the analytical derivations, of our work. We will revise the manuscript accordingly based on your suggestions. Below please finds our detailed reply for your comments.
> "I find that the machinery of Lie group theory is not really n... | Summary: Using Lie group algebra, this paper derives the general conditions for a network that evolves through a series of states at a continuous set of possible rates. This would be useful for temporal rescaling, as is frequently observed in motor generation (imagine playing a melody at a different rate than it was l... | Rebuttal 1:
Rebuttal: We deeply appreciate your positive review and insightful comments.
> “The dependence of this approach on a common zero to the scale ($t_{start}$) ..."
Our analysis implicitly assumes that the sequences with different speeds start from the same $t_{start}$, and time reversal requires exchanging s... | Summary: The authors present a framework to analyze and realize temporal scaling in networks. Studies in neuroscience have thought about how the brain achieves outputs at different speeds. Here, the authors present a theoretical analysis of temporal scaling group operators, and then posit a disentangled neural circuit ... | Rebuttal 1:
Rebuttal: Thank you for the reviewer’s comments, especially the appreciation of our theoretical derivations in identifying how temporal scaling group operators are represented in the recurrent network dynamics.
> “Originality: While the authors … which is not cited.”
Thank you for directing us to this pap... | Rebuttal 1:
Rebuttal: We sincerely appreciate all reviewers' comments! Here we reply to some common questions raised by reviewers.
### Novelty and contribution
As some reviewers asked whether Lie group theory is necessary for this study, we emphasize the objectives and contributions of our work that include **three** ... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Corruption-Robust Offline Reinforcement Learning with General Function Approximation | Accept (poster) | Summary: The paper studies offline RL with general function approximation in the presence of adversarial corruption. Specifically, an adversary can corrupt the rewards and transitions in the offline dataset with a fixed cumulative budget. The corruption formulation encompasses prior corruption formulations. The propose... | Rebuttal 1:
Rebuttal: Thank you for your positive review and constructive evaluation!
**Q1** The main limitation of this work is that the proposed algorithm requires the corruption level to be known. I did not find any attempts at possible heuristics for estimating the corruption level, which limits the practicality o... | Summary: The research investigates the issue of corruption robustness in offline reinforcement learning with general function approximation, where an adversary can corrupt samples in the offline dataset. The objective is to find a policy that remains robust against such corruption and minimizes the suboptimality gap co... | Rebuttal 1:
Rebuttal: Thank you for your positive feedback and valuable comments!
**Q1** On Assumption 2.1. I don't see what is pessimism helping in this work if not to avoid completeness. I am okay with this assumption if the work is trying to address the corruption-robust problem which requires (?) completeness?
**... | Summary: This paper proposed a new robust offline RL algorithm called the Corruption Robust PEVI (CR-PEVI). The CR-PEVI achieves smaller suboptimality error compared to existing methods. The key idea behind CR-PEVI is to utilize the uncertainty-weighting technique when solving the least-squares problem in PEVI. The wei... | Rebuttal 1:
Rebuttal: Thank you for your positive feedback and suggestions!
**Q1** The definition of cumulative corruption is less intuitive and harder to quantize. Is the goal to favor theoretical analysis? It makes more sense to define corruption level as the real modification of the reward/state transition over the... | Summary: This work considers corrupted offline RL with general function approximation, and propose an algorithm based on PEVI and the uncertainty weighting technique. Under certain date coverage and well-explored dataset assumptions, the proposed algorithm achieves neal-optimal bound, where the corruption-independent t... | Rebuttal 1:
Rebuttal: Thank you for your positive feedback and valuable comments!
**Q1** The uncertainty-weighting technique and PEVI are not new. This work extends corrupted online setting to offline setting.
**A1** Our work is not a trivial extension from online RL to offline RL. We would like to highlight our tech... | Rebuttal 1:
Rebuttal: We express our gratitude to the reviewers for their positive feedback and constructive suggestions!
During the rebuttal period, we make some explanations and illustrations for our theory. For experiments, we complement two experiments and present the results in the following pdf. Firstly, we sho... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Regularization properties of adversarially-trained linear regression | Accept (spotlight) | Summary: This paper provides an in-depth analysis of adversarial training with linear models and its relationship to regularized regression methods under the overparametrized regime.
Depending on the value of the perturbation radius, it is revealed that there exist three modes.
When the radius is small, solutions to ad... | Rebuttal 1:
Rebuttal: We greatly appreciate the reviewer's comments and feedback. We are thankful for the overall positive assessment of our work and also believe the comments will greatly improve the paper.
**About Theorem 4 and its relation to Xu et al (2009).** Thank you for the comment. We updated the manuscript i... | Summary: This paper studies the connection between adversarial training and regularization methods in linear regression problem. Simulation studies are provided to justify the correctness of their theoretical observations.
Strengths: The authors conducts a comprehensive study on the relationship between adversarial tr... | Rebuttal 1:
Rebuttal:
We thank the reviewer for useful feedback and comments.
**Why restrict the analysis to linear models?** We understand your concern and believe you raise a very valid point: adversarial training is very often used in the context of neural networks and it is natural to be interested in the behavio... | Summary: The paper studies adversarial training (AT) for linear regression for which the inner maximization problems has a closed form solution. They then attempt at relating the solutions to solutions of other optimization problems:
- They show that the minimum norm interpolator also minimizes the adversarial loss (i... | Rebuttal 1:
Rebuttal: We thank the reviewer for the detailed feedback. We are grateful the reviewer pointed out aspects that were unclear. We answer your comments below and will update the paper accordingly.
- **Thm 1:** Indeed, we did not establish uniqueness. We will tone down our claims and rewrite so it is not mi... | Summary: This paper investigates adversarial training of linear regression. The authors compared the solution of adversarial training and other regularization frameworks (minimum-norm interpolating, ridge regression, Lasso and square-root Lasso), and established close relations between adversarial training and other me... | Rebuttal 1:
Rebuttal: We thank the reviewer for the careful reading of our paper (including even going into the code). We answer the comments below.
1. **On the equivalence with parameter shrinking methods.** Thank you for the comment. A possible interpretation is to view
$$\min\_{\beta}\frac{1}{n} \sum\_{i=1}^n (y\... | Rebuttal 1:
Rebuttal: We sincerely thank all reviewers for taking the time to read and reflect on our paper. We are pleased to have received a generally positive response and that all the reviewers highlight the clarity of the paper and that it provides insight into adversarial training. We also believe the feedback pr... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Neural Modulation for Flash Memory: An Unsupervised Learning Framework for Improved Reliability | Accept (poster) | Summary: A neural modulator is proposed and sufficient results are presented.
Strengths: Sufficient results are presented.
Weaknesses: No
Technical Quality: 2 fair
Clarity: 3 good
Questions for Authors: 1. What kind of noise added in the figure 1.
2. Elaborate briefly about the neural modulation comparison with co... | Rebuttal 1:
Rebuttal: Dear reviewer,
Thank you for the time and effort you have dedicated to reviewing our manuscript.
Below are detailed answers to your questions:
Q: What kind of noise added in the figure 1.
A: Figure 1 is only an illustration of a real Vth distribution. We used a gaussian noise to generate the l... | Summary: The paper introduces a programming method for 3D NAND flash memory. In 3D NAND, data is stored in the voltage level of cells. A cell can typically store a few bits of information (2,3,4 etc). As device density increases, the data is prone to multiple types of errors. The paper is targets two types of errors, t... | Rebuttal 1:
Rebuttal: Dear reviewer,
We appreciate your thoughtful feedback on our work. Below are detailed answers to your questions:
Q: It seems not clear whether any part of the experimental evaluation was performed on a physical device, or it was simulated. Could you please give some details?
A: We used data col... | Summary: This paper presents a novel approach for improving the reliability of NAND flash memory by utilizing unsupervised learning and neural modulation techniques. The authors constructed and trained a neural modulator that reduces programming errors and compensates for data degradation over time. The proposed framew... | Rebuttal 1:
Rebuttal: Dear reviewer,
Thank you for carefully reading our manuscript, and pointing out some unclear points.
Specifically, we plan to further elaborate on the required resources:
First, to ensure no additional latency is introduced due to the modulator, the set of targets for WL_n+1 is calculated while ... | Summary: * This paper introduces a unsupervised learning method using GAN for flash memory.
Strengths: * The paper is well written and organized, generally easy to follow.
* To the best of my knowledge, this paper is the first to introduce an unsupervised technique to this kind of problem.
* The core problem is well ... | Rebuttal 1:
Rebuttal: Dear reviewer,
We appreciate your thoughtful engagement with our work and the valuable feedback.
Below are detailed answers to your questions:
Q: On Figure 7, it does give a view of generated WL from the generated model, is there quantized way to measure the quality?
A: In the paper we measure... | Rebuttal 1:
Rebuttal: Dear reviewers,
Following your feedback, we would like to provide a detailed answer for two specific questions, both have relevant figures in the attached PDF:
Q: We customize the Vth target for a given cell, but assume the neighbors use default target values. Would it not be better for the enc... | NeurIPS_2023_submissions_huggingface | 2,023 | Summary: The authors propose a generative model architecture for compensating for errors in NAND flash memory, with the goal of extending flash memory block lifetime. The architecture consists of a GAN and an autoencoder, where the encoder and decoder represent the memory’s modulator and demodulator respectively. The e... | Rebuttal 1:
Rebuttal: Dear reviewer,
Thank you for carefully reading our manuscript. We plan to extend the discussion in the revised version to increase the clarity of the paper, incorporate the references and fix the labeling of Fig 9. Specifically, in figure 9a we plot the BER as a function of retention time for sev... | null | null | null | null | null | null |
Invariant Learning via Probability of Sufficient and Necessary Causes | Accept (spotlight) | Summary: This paper proposes the new concept of 'essential causal information' to achieve OOD generalization and proposes the probability of necessarity and sufficiency (PNS) risk to estimate the representation if satisfies the new concept. Then, it proposes a new algorithm to achieve loss PNS risk on the test domain.
... | Rebuttal 1:
Rebuttal: ## Response to Reviewer CQCE:
We appreciate the suggestions from reviewer CQCE and are grateful for the recognition of the novelty of our approach, the rigor of our theory, and our writing style. We will respond to each of the questions posed by the reviewer below.
**Q1 The fox also has long whi... | Summary: This paper presents a novel approach to enhance out-of-distribution (OOD) generalization by identifying both sufficient and necessary conditions, thereby extending the existing literature that primarily focuses on causal invariance. The proposed method employs the probability of sufficient and necessary causes... | Rebuttal 1:
Rebuttal: ## Response to Reviewer pPMP:
We thank reviewer pPMP for the constructive feedback and recognition of our novel and intuitive idea, as well as our experiments on various datasets. We respond to each of the reviewer's questions in the sections below.
**Q1 It is difficult to understand the definiti... | Summary: The paper proposes a new formulation of the out-of-distribution generalization task via finding necessary and sufficient features, similar to invariant feature learning but with some added constraints.
The paper uses the definition of necessary and sufficient from Pearl (2008) and proposes a related risk measu... | Rebuttal 1:
Rebuttal: # Response to Reviewer qVZ2:
We appreciate reviewer qVZ2's constructive feedback and positive comments on our idea, theoretical contributions, and experiments contributions. We respond to each of the reviewer's questions in the following.
**Q1.1** **Need to discussed more to understand PNS**
W... | Summary: A mostly theoretical paper that formulates the necessity and sufficiency of causal variables in order to optimize models to rely on necessary & sufficient variables. The paper provides rigorous theory connecting the empirical loss when models are optimized using the Probability of Necessity and Sufficiency Los... | Rebuttal 1:
Rebuttal: ## Response to Reviewer jeo8:
We thank reviewer jeo8 for their suggestions and appreciate their recognition that our method is intuitive and our theory is rigorous. We will address each of the questions posed by the reviewer in the sections below.
**Questions. Rigorous empirical evaluation in th... | Rebuttal 1:
Rebuttal: # General Response:
We sincerely appreciate all reviewers's great efforts on review and comments of our work. We thank for the positive comments:
- **Clear and well-explained intuition** (Reviewer #jeo8 #pPMP )
- **Interesting and reasonable idea** (Reviewer #qVZ2 #pPMP #CQCE)
- **Novel idea** (R... | NeurIPS_2023_submissions_huggingface | 2,023 | Summary: This paper addresses the crucial issue of out-of-distribution (OOD) generalization in machine learning models, where the testing data distribution differs from the training data. To tackle this limitation, the authors introduce the concept of the Probability of Sufficiency and Necessary Causes (PNS), represent... | Rebuttal 1:
Rebuttal: # Response to Reviewer aG69:
We would like to thank reviewer aG69's comments for our novelty as well as our experiments to validate the effectiveness of the proposed PNS risk. We respond to each of the reviewer's questions in the following.
**Q1.** **The differences between existing PNS risk and... | null | null | null | null | null | null |
MAG-GNN: Reinforcement Learning Boosted Graph Neural Network | Accept (poster) | Summary: The authors present MAG-GNN, a subgraph-based Graph Neural Network (GNN) method that utilizes reinforcement learning to effectively identify discriminative subgraphs. They empirically assess the performance of MAG-GNN using both synthetic and real-world datasets.
Strengths: - The main elements of the RL frame... | Rebuttal 1:
Rebuttal: We are grateful to reviewer 9Mgd’s effort in helping us improve our work and paper. We address the reviewer’s concerns and questions as follows.
- Q1: “The property” refers to many graphs’ properties that a discriminative small subset, rather than the whole set, of a graph’s subgraphs is sufficie... | Summary: The paper considers the problem of increased complexity of subgraph GNNs with respect to standard MPNNs, due to the enumeration of the entire bag of subgraphs for every given graph.
The authors first study the need of enumerating all subgraphs, and show that there exist graphs where using a small subset of su... | Rebuttal 1:
Rebuttal: We are very grateful for reviewer jrFK’s in-depth review and insightful comments. We address the reviewer’s concerns and questions as follows:
- W1: Full comparison with a random approach
As mentioned in Q1 in the paper, we also believe that outperforming random subgraphs is critical to the actu... | Summary: The paper aims to locate expressive subgraphs for graph representation learning. To do so, the author proposed to use deep Q Learning for training an efficient agent which optimizes the combinatorial problem of choosing the optimal subgraph. The proposed agent-aided GNN shows good performance and is able to tr... | Rebuttal 1:
Rebuttal: We thank Reviewer SjU6 for acknowledging our contribution and appreciate the reviewer's suggestions to improve our work. We address the reviewer's concerns and questions as follows,
- W1: How to train an effective agent seems to be a critical point. Despite that, the author has a brief discussion... | Summary: this paper proposes a strategy to select the optimal subset instead of extensive enumeration for the subgraph GNN. The proposed method can achieve comparative results and reduce the running time.
Strengths: 1. the paper is well written and easy to follow.
2. the experiment part is thoroughly detailed。
3. the ... | Rebuttal 1:
Rebuttal: We thank the reviewer for the positive feedback and refer the reviewer to our other responses in case the reviewer is interested in more details about our work. We are also happy to answer any further questions the reviewer might have. | Rebuttal 1:
Rebuttal: We thank all reviewers for their constructive comments that help us improve our work. Here we address some common questions and concerns that reviewers raise, and we also respond to their comments individually.
- GR1: MAG-GNN is a random method that uses an RL agent to refine the selected subgrap... | NeurIPS_2023_submissions_huggingface | 2,023 | Summary: This paper proposes a reinforcement learning boosted approach to subgraph graph neural networks (GNNs) that achieves high expressivity with constant complexity, instead of the exponential complexity of existing methods. They demonstrate the effectiveness and efficiency of their approach through experiments tha... | Rebuttal 1:
Rebuttal: We appreciate the feedback from reviewer FiDZ, and would like to introduce our research's background and intuition further. In traditional deep-learning fields (e.g. CV and MLP), expressivity usually is not an issue, and reducing overfitting is more critical. However, in GNN, the expressivity is l... | Summary: The paper introduces MAG-GNN, an RL approach for identifying important subgraphs. This is valuable since models using fewer subgraphs incur lower inference costs. Experimentally, the authors show that MAG-GNN retains the strong performance of sub-graph approaches while reducing the run time.
Strengths: The gu... | Rebuttal 1:
Rebuttal: We thank reviewer egpV for the insightful review, and we address reviewer egpV’s questions and concerns as follows,
- W1: **Training efficiency**. While MAG-GNN accelerates inference, the training process becomes more time-consuming. On smaller graphs, this training pipeline may dominate the over... | null | null | null | null |
Fast Optimal Locally Private Mean Estimation via Random Projections | Accept (poster) | Summary: This paper presents an efficient distributed mean estimation framework, which first reduces local data dimension using (different) random projections and then applies an existing private processor (Asi et al [7]) on the projected data before sending them to the server. Further speedups of the framework are als... | Rebuttal 1:
Rebuttal: We thank the reviewer for their feedback. Below we answer all of their questions.
1. Novelty: While the techniques we use are similar to techniques used in the *non-private* version of the problem, our paper is the first to use them with privacy and several challenges arise in doing so. One of th... | Summary: The paper introduces a fast algorithm for private mean estimation under communication constraints. The proposed scheme primarily relies on random projection, and the algorithm enhances speed through the utilization of Hadamard matrix and sampling techniques.
Strengths: - Theorem 1 offers a theoretical analysi... | Rebuttal 1:
Rebuttal: We thank the reviewer for their positive feedback and comments. Below we answer all of their questions.
1. Comparison to SQRK (communication):
The primary emphasis in this paper is on getting the optimal error bounds with practical communication: our work is the first to achieve optimal utility ... | Summary: The paper studies the problem of locally private mean estimation, and design algorithms with optimal (upto $1+o(1)$ factor) error with better communication complexity and runtime. The key idea behind improved communication complexity and runtime is (a certain) random projection based pre-processing step.
Stre... | Rebuttal 1:
Rebuttal: We thank the reviewer for their detailed and positive feedback. Below we answer all of their questions.
1. Technical challenges: While the techniques we use are similar to techniques used in the *non-private* version of the problem, our paper is the first to use them with privacy and several chal... | Summary: The paper considers the problem of performing mean estimation for worst-case data under local differential privacy. The precise setting considered involves $n$ users, each of whom has one data vector $v_i$ (which, without loss of generality, can be assumed to exist within the unit ball). The users must privati... | Rebuttal 1:
Rebuttal: We thank the reviewer for their positive feedback. | null | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Two-timescale Derivative Free Optimization for Performative Prediction with Markovian Data | Reject | Summary: The performative prediction problem targets real-world application using deployed models, and this method studies the optimiazation method for such a scenario. Since directly optimiazing the non-convex objective is challenging, this paper utilizes weighted Markovian samples of states to estimate the gradient ... | Rebuttal 1:
Rebuttal: Thank you for the review. We address your concerns as follows.
> The experiments are conducted on toy examples ... challenging task setting. Also, how does the method perform in comparison with other strategies that also fit with such tasks?
We acknowledge that the experiments are mostly on toy ... | Summary: This paper proposes a derivative-free two-time scale for performative prediction problem. The two-timescale facilitates a faster accumulation of samples to compute a gradient with smaller bias. For smooth nonconvex objective, the work proves a iteration complexity of $O(\epsilon^{-3})$ for the $\epsilon$-stai... | Rebuttal 1:
Rebuttal: Thank you for the comments which inspired us to explore our contributions more. Referring to the _General Responses_, we reiterate that the our DFO algo is new which has not been analyzed before, nor is it studied with Markov data, further it is motivated by **practical** performative prediction (... | Summary: The paper studies performative prediction when the data is stateful, in particular generated via a controlled Markov chain, and the learner's loss is possibly nonconvex. The paper develops a two-timescale derivative-free optimization algorithm for this setting and shows a O(1/eps^3) sample complexity for findi... | Rebuttal 1:
Rebuttal: We are glad that the reviewer liked our paper recongnizing the theoretical contributions of performative prediction problem under the nonconvex, stateful settings. Our point-to-point responses to concerns on weakness and questions are:
> Discussion about the differences between recent performativ... | Summary: This paper is on performative prediction i.e. a setting where the data distribution changes in response to the predictions of a learned model. The prototypical example of such a setting is where a learned model is used to make loan decisions, and then people or companies adjust their behavior based on knowledg... | Rebuttal 1:
Rebuttal: Thank you for the careful review. As mentioned in the general response, we emphasize that our contributions lie on proposing a practical algorithm for performative prediction (aiming at finding the performative optimal solution) with limited knowledge assumed on the learner's side. Below we shall ... | Rebuttal 1:
Rebuttal: We thank all the five reviewers for their careful reviews. We summarize our responses and highlight the main contributions of our paper.
The performative prediction (PP) problem constitutes a new class of stochastic optimization problems with decision dependent distribution. This is challenging a... | NeurIPS_2023_submissions_huggingface | 2,023 | Summary: The paper considers the performative prediction problem, where the goal of learner is to optimize the expectation of the known loss function over a decision-dependent unknown data distribution that evolves according to an underlying controlled Markov chain. Authors presents a stochastic derivative-free optimiz... | Rebuttal 1:
Rebuttal: Thank you for the review. We address your concerns as follows.
> lower bound for the problem
We are not aware of any such lower bound result for the **decision-dependent** stochastic optimization problem (1) without access to $\Pi_{\theta}$. In fact, to our best knowledge, even in the setting w... | null | null | null | null | null | null |
Controlling Text-to-Image Diffusion by Orthogonal Finetuning | Accept (poster) | Summary: This paper proposes a parameter-efficient fine-tuning algorithm that preserves the so-called "hyperspherical energy" to achieve a great trade-off between prevervation and generalization and high training efficiency. The authors conduct experiments on subject-driven generation and controllable generation, which... | Rebuttal 1:
Rebuttal: We would like to sincerely thank the reviewer for the positive comments on our work. We take every comment seriously and hope our response can address the reviewer’s concerns. If there are any remaining questions, we are more than happy to address them.
---
**Q1: The related work of subject-driv... | Summary: This work proposes Orthogonal Finetuning (OFT), a new finetuning method for t2i diffusion models, centered on maintaining the hyperspherical energy from the pretrained model during finetuning to preserve semantic structures. It also introduces Constrained Orthogonal Finetuning (COFT), a variant that limits ang... | Rebuttal 1:
Rebuttal: We would like to sincerely thank the reviewer for the positive comments on our work. We take every comment seriously and hope our response can address the reviewer’s concerns. If there are any remaining questions, we are more than happy to address them.
---
**Q1: DB does not add additional infer... | Summary: The paper proposes a method for efficiently fine-tuning pre-trained text-to-image generation models, good results are obtained in experiments.
Strengths: The paper proposes OFT and COFT, which can efficiently fine-tune the pre-trained text-to-image generation model as less parameters needed to be fine-tuned c... | Rebuttal 1:
Rebuttal: We would like to sincerely thank the reviewer for the positive comments on our work. We take every comment seriously and hope our response can address the reviewer’s concerns. If there are any remaining questions, we are more than happy to address them.
---
**Q1: Comparing supervised FID on a gi... | Summary: The paper introduces a principled finetuning method called Orthogonal Finetuning (OFT) for adapting text-to-image diffusion models to different downstream tasks. The key idea of the proposed approach is to preserve hyperspherical energy which characterizes the pairwise neuron relationship on the unit hypersphe... | Rebuttal 1:
Rebuttal: We would like to sincerely thank the reviewer for the positive comments on our work. We take every comment seriously and hope our response can address the reviewer’s concerns. If there are any remaining questions, we are more than happy to address them.
---
**Q1: On OFT in section 3.2 and COFT i... | Rebuttal 1:
Rebuttal: Dear Reviewers and ACs,
We are deeply appreciative of all the reviewers and ACs for spending time on our submission. **Rebuttal PDF** contains the figures that are referred in the response to reviewers.
Best,
Authors
Pdf: /pdf/ee3e625a5c33c4624ec17ac9af0ab57c4035224e.pdf | NeurIPS_2023_submissions_huggingface | 2,023 | Summary: This paper proposes a new method for finetuning text-to-image diffusion models for better guidance and control. The proposed method OFT is a low-cost finetuning method even in comparison with LoRA, and it can preserve semantic generation ability well. Furthermore, this work proposes COFT to improve the stabili... | Rebuttal 1:
Rebuttal: We sincerely thank the reviewer for the recognition of our contribution, motivation and novelty. We take every comment seriously and hope our response can address the reviewer’s concerns. If there are any remaining questions, we are more than happy to address them.
---
**Q1: Unfortunately, this ... | null | null | null | null | null | null |
Learning Mask-aware CLIP Representations for Zero-Shot Segmentation | Accept (poster) | Summary: This paper proposes a new topic and method for training a mask-aware CLIP, which could serve as a core component for open-vocabulary segmentation. The designed structure could be used as a flexible plug-in but brings significant improvements for existing methods on various benchmarks.
Strengths: 1. This topic... | Rebuttal 1:
Rebuttal: # Response to Reviewer XVP7
Thank you so much for acknowledging the strength of our method. We have carefully considered your constructive and insightful comments and here are the answers to your concerns.
**Q1. Extending the method into open-vocabulary panoptic settings**
First of all, we full... | Summary: This paper mainly discusses how to use pre-trained CLIP to solve zero-shot segmentation task, and proposes a new method called Mask-aware Fine-tuning (MAFT) to address the issue of significant false positives in CLIP's classification of mask proposals. Specifically, the paper introduces an Image-Proposals CLIP... | Rebuttal 1:
Rebuttal: # Response to Reviewer uMA5
Thank you so much for acknowledging the strength of our method. We have carefully considered your constructive and insightful comments and here are the answers to your concerns.
**Q1. The ability to handle any number of mask proposals is not unique.**
There may be so... | Summary: The paper proposes a mask-aware fine-tuning method to address challenges faced by frozen-CLIP-based zero-shot segmentation methods. It addresses the problem of CLIP being insensitive to different mask proposals and tending to produce similar predictions regardless of the variation in proposals. The proposed IP... | Rebuttal 1:
Rebuttal: # Response to Reviewer q7Qn
Thank you so much for acknowledging the strength of our method. We have carefully considered your constructive and insightful comments and here are the answers to your concerns.
**Q1. Experiments with other backbones and proposal generators**
To enhance the comprehen... | Summary: The general goal of the paper is to leverage CLIP for zero-shot segmentation. For this, in contrast to prior work, an CLIP-inspired IP-CLIP encoder is trained to enable mask-level encodings for segmentation. The core of the approach is the so-called IP-encoder that uses as input a frozen mask generator, as wel... | Rebuttal 1:
Rebuttal: # Response to Reviewer fJQP
Thank you so much for acknowledging the strength of our method. We have carefully considered your constructive and insightful comments and here are the answers to your concerns.
**Q1. Explanation of $A^c \in\mathbb{R}^{C \times N}$, $A^c_{IP}$$\in\mathbb{R}^{C \times ... | Rebuttal 1:
Rebuttal: # General Response
**Q1. Extending the method with SAM**
As some Reviewers suggest us use stronger proposal generators, *e.g.*, Segment Anything Model (SAM). We explore the performance with SAM-H using an original CLIP (dubbed `SAM+oriCLIP`) or a mask-aware fine-tuned CLIP (dubbed `SAM+MAFT`). I... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Expressive dynamics models with nonlinear injective readouts enable reliable recovery of latent features from neural activity | Reject | Summary: This paper develops a new latent variable model for neural recordings ('ODIN') that uses a flow-based readout to map from the latent space to neural activity. The authors motivate their method by the injectivity of such flow-based methods and compare it to baseline models with linear and MLP readouts. ODIN exh... | Rebuttal 1:
Rebuttal: We thank the reviewer for their detailed and insightful comments. While they were impressed by the breadth of our evaluation metrics, they were unconvinced by some of our arguments about the importance of readout injectivity for accurate latent inference.
- I am left unconvinced that injectivity ... | Summary: This manuscript looks at the challenge of capturing latent dynamics from neural data, a common modeling challenge in neuroscience that has attracted significant recent attention. One common approach is to use deep learning methods, and recent highly predictive methods have been proposed that rely on high-dim... | Rebuttal 1:
Rebuttal: Reviewer GbAv
We thank the reviewer for their great suggestions to improve the manuscript.
- The theoretical backing of this work is sparse, at best, and the proposed technique is very heuristic. When will this technique be good enough? When will it theoretically hold? The long-term impact of th... | Summary: The paper proposes Ordinary Differential equations autoencoder with Injective Nonlinear readout (ODIN) model, a model for inferring latent representations underlying high-dimensional neural spikings. Additionally, the ODIN model is designed to recover the latent dynamics and generative process. Through evaluat... | Rebuttal 1:
Rebuttal: Reviewer x81c:
We thank the reviewer for their suggestions. We believe that clarifying a few points will help to improve the reach and impact of the manuscript.
- The effect of injectiveness in the generative process is not evaluated
Unfortunately, we are not sure if this comment referred to t... | Summary: The paper presents ODIN, a new autoencoder model of population neural dynamics that uses a neural ODE (NODE) to model latent dynamics. The core innovation is the use of a simplified invertible ResNet to ensure that a nonlinear and approximately _injective_ readout from latent dynamics to population neural acti... | Rebuttal 1:
Rebuttal: We thank the reviewer for their insightful feedback and their suggestions for improvements!
- the improvement of ODIN over MLP+NODE at the true dimensionality ($\hat{D}=3$) is small and it's unclear whether the improvement is significant.
We apologize for the confusion: while the slight improv... | Rebuttal 1:
Rebuttal: *General Response:*
We thank the reviewers for their attention and feedback, which has allowed us to strengthen this body of work.
Overall, reviewers pointed out that our focus on readout injectivity was a novel approach to an important problem in neural data modeling (UThX, GbAv). They pointed ... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Block Low-Rank Preconditioner with Shared Basis for Stochastic Optimization | Accept (poster) | Summary: In this paper, the authors introduce a novel method to perform adaptive gradient updates. Similarly to other works in the literature, this method focuses on capturing curvature information by considering an approximation to the second-moment matrix $H_t = \sum_t g_t g_t^T$, where $g_t$ is the gradient at the t... | Rebuttal 1:
Rebuttal: Thank you for the detailed review!
> The approach might be rather limited in its applicability, as it mostly focuses on fully-connected layers.
We have experiments on the transformer, which also contains self-attention layers. In general, our method can be used to handle any tensor product opera... | Summary: Computing the full Gram matrix is practically infeasible for large models even if it's done layer-wise. This work first confirms the block structure of each layer's Gram matrix experimentally. Previous work considered approximating each of these blocks as a low-rank matrix to improve the memory and computation... | Rebuttal 1:
Rebuttal: Thank you for the valuable feedback.
> The evaluation of the method remains limited. The auto-encoding task is only tried with MNIST. it should be feasible to also run CIFAR 10.
In addition to the autoencoder benchmark, we also conduct experiments on transformers. We will further expand the auto... | Summary: Following up on the line of work stemming from AdaGrad, the paper presents a novel adaptive optimizer that uses a block-diagonal preconditioner. The block-diagonal elements of the second-moment matrix are approximated through a low-rank approximation that shares the same basis, computed via randomized SVD, for... | Rebuttal 1:
Rebuttal: Thank you for the detailed review!
> I believe the paper would benefit from a wider experimental section, perhaps including convolutional networks or vision transformers (on vision tasks).
We will expand the experimental section with our future experiments.
> As a very minor comment: I think t... | Summary: This paper proposes an adaptive optimizer similar to Shampoo and GGT, which considers non-diagonal components of the preconditioner matrix. They reduce the cost of the preconditioner by using a unit-wise block diagonalization on the output units and a further low-rank approximation with a shared basis between ... | Rebuttal 1:
Rebuttal: Thank you for the valuable response.
> The novelty with respect to the existing work [4] and [11] seems incremental
We are sorry that our writing might cause some confusion, but our method is very different from the work of Shampoo [4] and Sketchy [11]. Shampoo and Sketchy approximate the secon... | Rebuttal 1:
Rebuttal: We thank all the reviewers for the valuable feedback. We have added figures for the ablation study and the training curves for the transformer benchmark.
Pdf: /pdf/9b35d8e633ad85f7eaaa190931a58f5c4f70277b.pdf | NeurIPS_2023_submissions_huggingface | 2,023 | Summary: This paper studies how to improve the memory efficiency of adaptive methods with (non-)diagonal preconditioning. The authors propose approximating the diagonal blocks of the second moment matrix by low-rank matrices and enforcing the same basis for the blocks within each layer. The authors also provide theoret... | Rebuttal 1:
Rebuttal: Thank you for the valuable feedback!
> The document is not very easy to understand.
We are sorry for the inconvenience and will further polish the paper.
> The proposed methodology is not clearly superior to the baseline methodology.
Our proposed method outperforms Adam on both benchmarks. On ... | Summary: This paper implements an online block version of AdaGrad to capture off-diagonal information in the preconditioner at each block level. A low-rank approximation is used for each block, which is a common approach in the literature to decrease time and memory complexities. To further reduce the memory usage, a s... | Rebuttal 1:
Rebuttal: Thank you for the detailed review!
> the memory complexity scales quadratically with rank k.
As stated in the paper, the memory complexity scales quadratically with the rank k. But since each block has a different coefficient matrix, our approach requires a much smaller k compared with Sketchy.... | null | null | null | null |
Opening the Vocabulary of Egocentric Actions | Accept (poster) | Summary: In this paper, a new method is proposed for open vocabulary object understanding within egocentric videos for the task of action recognition. The method uses CLIP features to generalise to new objects which haven't been seen during training by focusing on objects using a HoI model, combining nouns via object m... | Rebuttal 1:
Rebuttal: We thank the reviewer for their valuable feedback and acknowledging our task of expanding egocentric action recognition beyond a closed vocabulary of objects.
**W1**
*"Figure 2 could be improved..."* - Thank you for the feedback, we will improve this figure in our revision and add a more detaile... | Summary: The paper tackles open vocabularly action recognition where novel objects may be seen at test time and will be composed with a known verb. Note that the paper does not investigate the open vocabularly of verbs. To address this problem the paper proposes 'object agnostic pretraining' to become less reliant on t... | Rebuttal 1:
Rebuttal: We thank the reviewer for their detailed review, great suggestions and acknowledging the additional experiment on Something-Else. We also thank the author for appreciating the motivation behind our benchmark design. Due to space constraints, kindly find the required results (tables) in the attache... | Summary: The paper proposes a novel open vocabulary action recognition task, and this task aims to generalize the ability of the model to an open vocabulary of actions with seen and novel objects. To do this, the authors propose a new method, which mainly contains two components: Object Agnostic Pretraining (OAP) and A... | Rebuttal 1:
Rebuttal: We thank the reviewer for their thoughtful comments and acknowledging the importance of our proposed open vocabulary action recognition task, the simplicity of the proposed approach and sufficient experimentation.
**W1**
- *"why interested in recognizing active objects"* - Our task is action rec... | Summary: Existing approaches for egocentric action recognition considers the problem as recognizing actions from a closed set. The manuscript instead proposes to consider this as an open set recognition problem where some objects are unseen during training. Towards, a novel benchmark for evaluating open set egocentric ... | Rebuttal 1:
Rebuttal: We thank the reviewer for their valuable feedback and acknowledging our novel task, benchmarks and the set of ablations.
**W1**
- In the main paper, we have provided the results for *“One important baseline would be...”* The 1st row of `Tab. 4` provides S3D (without OAP) results on EPIC100-OV. T... | Rebuttal 1:
Rebuttal: We thank all the reviewers for their valuable time and constructive feedback. We also thank them for appreciating our novel problem formulation of open vocabulary action recognition.
We appreciate the concern from R3 (DGnw) and R4 (FCUw) for not considering an open vocabulary of verbs; here we e... | NeurIPS_2023_submissions_huggingface | 2,023 | Summary: Learning representations that generalise across novel objects and actions is an important topic of research that is of great practical utility. This is especially true due to most datasets containing only a few combinations of object-action pairs or triplets, which often leads to model overfitting. A key quest... | Rebuttal 1:
Rebuttal: We thank the reviewer for the thoughtful comments and acknowledging our open vocabulary benchmark, our method’s potential and our ablations in Table 5.
**S1**
*“This paper mentions that the “pre” and “post” are….”* - Using learnable prefix and postfix context prompts [13, 25, 66] are used as a... | null | null | null | null | null | null |
Provably (More) Sample-Efficient Offline RL with Options | Accept (poster) | Summary: The paper investigates the sample complexity of offline reinforcement learning (RL) with options. An information-theoretic lower bound for the suboptimality of the expected return of the policy over options that is learned offline is derived (by constructing hard MDP instances). The main contribution of the pa... | Rebuttal 1:
Rebuttal: Thanks for your reviews, we would like to address your concerns as follows.
**1. (Paragraph 1 in Weakness Section) Data Splitting Technique.**
We point out that the data splitting procedure for dataset $\mathcal{D}\_1$ is only to facilitate theoretical analysis. Indeed, by Definition 4.1, a $\x... | Summary: In this paper, the authors analyze the sample complexity for offline reinforcement learning with options. In particular, the sample complexity is measured by the suboptimality (the shortfall in the value function) of a hierarchical policy compared to the optimal hierarchical policy. The authors propose the pes... | Rebuttal 1:
Rebuttal: Thanks for your reviews, we would like to address your concerns as follows.
**(Q1) Explanation of $\theta_h^\mu(s,o)$.**
By its definition in (L121), $\theta_h^\mu(s,o)$ is the probability of visiting the state-option pair $(s,o)$ at timestep $h$, when following the hierarchical policy $\mu$. F... | Summary: The authors provide a sample complexity analysis of offline RL with options. A sub-optimality analysis was performed considering two data collection schemes, (D1) state-option-utility collection and (D2) state-action-reward collection for the proposed pessimistic value iteration algorithm (PEVIO). The analysis... | Rebuttal 1:
Rebuttal: Thanks for your reviews, we would like to address your concerns as follows.
**(Q1) Scope of the Theory.**
The results in this paper are limited to tabular MDPs. We will clearly illustrate this assumption in the preliminaries section in the later version of the paper.
**(Q2) Finite Concentrabil... | Summary: This analysis is the study of sample complexity in offline reinforcement learning (RL) with options, highlighting potential risks in online environmental exploration. The paper proposes the innovative PEssimistic Value Iteration for Learning with Options (PEVIO) algorithm and provides near-optimal suboptimalit... | Rebuttal 1:
Rebuttal: Thanks for your reviews, we would like to address your concerns as follows.
**(Q1) Connecting to Hybrid RL.**
The reviewer asks whether our results can also facilitate hybrid RL. While this is out of the scope of the paper, we would like to provide some comments as follows. In hybrid RL, it is ... | null | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
CODA: Generalizing to Open and Unseen Domains with Compaction and Disambiguation | Accept (spotlight) | Summary: This paper proposes a two-stage framework for the task of open-class domain generalization. During training, it introduces virtual unknown classes for the learning objective to insert unknowns into the latent space by compacting the embedding space of known classes. During the test phase, it proposes a prototy... | Rebuttal 1:
Rebuttal: Thank you for your constructive feedback and for recognizing our novel contributions. We list our responses to your questions below.
> **Q1. The authors initially utilize the term 'adaptivity gap', but introduce a new terminology, 'optimality gap', in Definition 1. What are their distinction and ... | Summary: This research paper introduces a novel experimental configuration referred to as Open Test-time Domain Generalization (OTDG), which can be viewed as an amalgamation of domain generalization, test-time adaptation, and open-set recognition. In order to address this particular experimental configuration, the auth... | Rebuttal 1:
Rebuttal: Thank you for the constructive feedback and for acknowledging the innovative aspects of our study. We address the concerns as follows and will improve the manuscript accordingly.
> **Q1. List some practical scenarios of the OTDG setup.**
Great question! We would like to discuss three practical s... | Summary: This paper focuses on the problem of domain adaptation. In particular, in addition to the already-studied setting of domain shift at test time, this paper considers the practical scenario of unknown classes being presented at test time. Often current models will incorrectly predict a known class label for exam... | Rebuttal 1:
Rebuttal: Thanks very much for your insightful comments and suggestions! We have properly revised our paper based on your reviews and below are our detailed responses.
> **Q1. More background on test-time adaptation can be given before introducing the "disambiguation" protocol.**
We appreciate the reviewe... | Summary: This paper introduces a novel problem formulation termed Open Test-Time Domain Generalisation, which accounts for the presence of both domain shift and unseen classes during the testing phase. To tackle this problem, a two-stage methodology is proposed, comprising Compaction and Disambiguation. The primary obj... | Rebuttal 1:
Rebuttal: We sincerely appreciate the reviewer for the positive feedback and insightful comments! We extend our apologies for any inconvenience or misunderstanding that may have occurred when reading our manuscript, and we assure you that we will incorporate all of your valuable suggestions in our final ver... | Rebuttal 1:
Rebuttal: We thank all the reviewers for their time, insightful suggestions, and valuable comments. We are glad that **ALL** reviewers appreciate our work and find our method simple and effective, highlighting that a novel problem (bs2E, BmaN, nkce, shJ8) - Open Test-Time Domain Generalization (OTDG), a nov... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
DASpeech: Directed Acyclic Transformer for Fast and High-quality Speech-to-Speech Translation | Accept (poster) | Summary: This paper presents a non-autoregressive direct speech-to-speech translation model based on Directed Acyclic Transformer. Experiments are conducted on CVSS-C and CVSS-T datasets, demonstrating translation performance comparable to the SOTA (Translatotron 2) but more than 16x speed up. It also demonstrates capa... | Rebuttal 1:
Rebuttal: We are deeply grateful for your positive evaluation of our work. Your valuable comments have been carefully considered, and we have provided responses to each of them below.
1. **Re: It would be helpful to include more results on languages with significant differences in ordering, such as English... | Summary: DASpeech is a non-autoregressive direct speech-to-speech translation (S2ST) model. The model adopts a two-pass architecture, where a linguistic decoder first generates the target text, and an acoustic decoder then generates the target speech based on the hidden states of the linguistic decoder. The linguistic ... | Rebuttal 1:
Rebuttal: Thank you for providing your valuable and constructive feedback! Below, you will find our responses to each comment.
1. **Re: Can you evaluate the proposed method with more languages?**
Thank you for your suggestions. Based on your advice, we conducted additional experiments on the CVSS-C De-En ... | Summary: The paper describes a non-autoregressive direct Speech-to-Speech Translation (S2ST) model called DASpeech that achieves both high-quality translations and fast decoding speeds. DASpeech uses a two-pass architecture where a linguistic decoder generates the target text, and an acoustic decoder generates the targ... | Rebuttal 1:
Rebuttal: Thank you for your helpful feedback! Please find below the responses to each comment.
1. **Re: The novelty of the paper is not significant. It is not clear that the proposed model can solve some unique problems in speech-to-speech translation scenario.**
Thank you for pointing that out. In our o... | Summary: This paper proposes a method for fast and high quality speech to speech translation.
The paper adopts the two-pass decoding framework.
The first pass decoder adapts prior work on non autoregressive decoding for machine translation.
The second pass decoder reuses fastspeech 2 except that the input are the hidde... | Rebuttal 1:
Rebuttal: We truly appreciate your positive evaluation of our work. Below, we have provided the responses to each of your valuable comments.
1. **Re: 227: "We implement the following baseline systems." Would you be able to explain this in more details. Does reimplement mean reusing existing results, using ... | Rebuttal 1:
Rebuttal: We sincerely appreciate the valuable comments from all the reviewers, which have provided us with many insightful suggestions. **We are greatly encouraged by the fact that most reviewers have found our work to be sound and effective, with a novel application in the speech-to-speech translation (S2... | NeurIPS_2023_submissions_huggingface | 2,023 | Summary: The paper introduces DASpeech, a directed acyclic transformer model for fast and high-quality speech-to-speech translation (S2ST). DASpeech employs a two-pass architecture, using a linguistic decoder based on DA-Transformer and an acoustic decoder based on FastSpeech 2. The proposed model captures the complex ... | Rebuttal 1:
Rebuttal: Thank you for providing your valuable and constructive feedback! Please find below the responses to each comment.
1. **Re: The novelty is not enough, the method of this article seems to just learn from the DAT method of NMT and use it in ST (Specific differences between DASpeech and the DA-Transf... | null | null | null | null | null | null |
Train Hard, Fight Easy: Robust Meta Reinforcement Learning | Accept (poster) | Summary: This is an interesting paper that harnesses a risk-averse optimization from RL and applies the trick to Meta RL (called RoML). It has a theory section that proves that biased gradients in ordinary robust RL objectives are no longer biased when applied to meta RL, an interesting result. there is also a data ine... | Rebuttal 1:
Rebuttal: We thank the reviewer for reading our work, and are encouraged that the reviewer found it cool, intuitive and solid.
---
Rebuttal Comment 1.1:
Comment: Thanks for the reply, best of luck to the authors. | Summary: This paper focuses on the meta-reinforcement learning setting, where the agent learns a meta-policy that adapts to new tasks with few observations. However, traditional meta-RL methods optimize the average return over tasks but often suffer from poor results in tasks of high risk or difficulty. This may make t... | Rebuttal 1:
Rebuttal: We thank the reviewer for their detailed and helpful feedback. Please see our responses below, as well as the new experiments we added following the review.
**What does robustness mean in meta RL?**
We defined the robustness of an agent in meta RL as the CVaR return of the agent. High CvaR retur... | Summary: The paper presents a novel approach to robust Meta-Reinforcement Learning (MRL), aiming to improve the consistency of returns across tasks, crucial for many RL applications. The authors introduce Conditional Value-at-Risk (CVaR) optimization to MRL, and propose a new method called Robust Meta RL algorithm (RoM... | Rebuttal 1:
Rebuttal: We thank the reviewer for their helpful comments. Please see our responses below.
**Could you elaborate on the conditions under which RoML improves the average return?**
This result is empirical, and we specify it mostly to show that optimizing the mean and the CVaR are not necessarily contradic... | Summary: This paper aim to address the limitations in the standard meta reinforcement learning (MRL) methods, where they optimize the average return over tasks and often struggle with high-risk or difficult tasks, hence limiting system reliability. To this end, the authors define a robust MRL objective with a controll... | Rebuttal 1:
Rebuttal: We thank the reviewer for their helpful comments. Please see our responses below.
**RoML requires the proper set up of parameter alpha and beta.**
The parameter alpha is part of the problem definition and it is not a hyper-parameter of RoML. It expresses the desired robustness level, which is an... | Rebuttal 1:
Rebuttal: We are excited that all the reviewers recommended accepting the paper, and thank them for their helpful comments.
The reviewers found that our work **“contributes to advancing RL in real-world application”** (ib31) and is **“crucial for many RL applications”** (CM4Y). They consider our **theoreti... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Penalising the biases in norm regularisation enforces sparsity | Accept (poster) | Summary: This work considers a single hidden layer ReLU network $f_\theta: \mathbb{R} \rightarrow \mathbb{R}$ with biases in the hidden layer as well as skip connections. In this context the representational cost $R(f)$ required to interpolate a function $f: \mathbb{R} \rightarrow \mathbb{R}$ is studied, which is defin... | Rebuttal 1:
Rebuttal: We thank the reviewer for the insightful comments. We now answer the concerns.
*Incremental contribution*
===
In the theory of NNs, the considered settings are often incremental wrt previous works. This incremental approach stems from the difficulty of the considered problems. Even relaxing a sma... | Summary: This paper studies the representational cost of a toy ReLU MLP with a single hidden layer, and a scalar input. The main result of the paper is that, when regularizing both the weights and the biases, the minimum cost functions are min norm interpolators which are also sparse (defined in the paper as having the... | Rebuttal 1:
Rebuttal: We thank the reviewer for the insightful comments. We now answer the raised concerns.
We agree with the weaknesses pointed out by the reviewer. As discussed in the paper, we indeed consider a simplified setting (1d, skip connection, etc.) as it allows for a tractable analysis. Studying a more ... | Summary: The paper addresses two main questions: (1) Can the parameters' norm be understood in terms of a quantity that is more insightful from a functional analysis point of view? (2) What kind of functions are learned when minimizing the empirical loss, either with an explicit or implicit regularization of the parame... | Rebuttal 1:
Rebuttal: We thank the reviewer for the many insightful comments. We now answer the raised concerns.
*Restricting setting*
===
As you rightly highlighted, analyzing any of these potential extensions is undoubtedly much more challenging. When conducting theoretical studies on neural networks, it is often d... | Summary: This paper theoretically investigates the effect of regularizing bias parameters’ norm, mostly ignored or omitted in previous analyses, for univariate one hidden layer ReLU networks. It turns out that minimizing $L_2$ norms for both weights and biases corresponds to regularizing a functional norm defined as th... | Rebuttal 1:
Rebuttal: We thank the reviewer for the many insightful comments. We now answer the raised concerns.
*Sparsity*
===
The generalization benefit of the considered sparsity remains unclear in the literature. Indeed, generalization bounds rely on the parameters’ norm rather than the network width (ie sparsity... | null | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
T2T: From Distribution Learning in Training to Gradient Search in Testing for Combinatorial Optimization | Accept (poster) | Summary: This paper proposes a novel framework, T2TCO, for solving combinatorial optimization problems. Specifically, it trains a generative model (i.e., discrete diffusion solver from [1]) to estimate the distribution of high-quality solutions for each instance, and, during testing, conducts an effective local search ... | Rebuttal 1:
Rebuttal: Thanks for the thorough review and valuable comments. We are encouraged with your acknowledgment of our writing, motivation, methodology, empirical results, and contribution to the ML4CO community. Below we respond to your specific comments.
**Q1: The proposed method is only applicable to diffusi... | Summary: This paper proposed a diffusion model based solution distribution learning method for combinatorial optimization. The main novelty is that a search mechanism is proposed, which uses the learned distribution to conduct a gradient based search for each instance where the gradient is computed w.r.t the optimiztio... | Rebuttal 1:
Rebuttal: Thanks for the valuable comments, constructive advice, and recognition of our idea and empirical results. Below we respond to your specific comments.
**Q1: What kind of problems can the proposed approach solve? Specifically, what kind of variables and constraints, and is it easy to design a probl... | Summary: The paper proposes a training to testing (T2TCO) framework to better align with the goal of combinatorial optimization (CO) problems, instead of pursuing the average expected performance, Real-world CO problems care more about instance-wise optimal solution. And thus, the paper first models the solution distri... | Rebuttal 1:
Rebuttal: Thanks for the valuable comments and for acknowledging the motivation practicality, framework novelty, and methodology soundness of our paper. Below we respond to your specific comments. Since your concerns are representative and common, we have discussed them mainly in the general response.
**Q1... | Summary: The paper presents T2TCO, a new approach for neural combinatorial optimization based on using gradient-based search in the solution space during test time. The proposed approach leverages generative modeling (specifically diffusion) for graduated solution improvement but employs an additional energy-based mod... | Rebuttal 1:
Rebuttal: Thanks for the thorough review and valuable comments. We are encouraged with your acknowledgement to our idea, experiments and contribution to the ML4CO community. Below we respond to your specific comments.
**Q1: Some related works are missed.**
Thanks for sharing these related works. We will d... | Rebuttal 1:
Rebuttal: Dear Area Chairs and Reviewers,
We greatly appreciate the reviewers' time, valuable comments, and constructive suggestions. We are delighted that all the reviews have expressed a positive inclination towards accepting our submission. Overall, the reviewers deem our paper as "well written" (p83a, ... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Attentive Transfer Entropy to Exploit Transient Emergence of Coupling Effect | Accept (spotlight) | Summary: This article expands the concept of transfer entropy to include attentive transfer entropy. Implementing attentive transfer entropy on connectomes of realistic biological networks demonstrated that this extension encompasses the transient coupling effect within the connectivity constraints of realistic biologi... | Rebuttal 1:
Rebuttal: We thank the referee for highlighting how we provided convincing empirical evidence that the attentive transfer entropy is particularly useful in understanding coupling effects of realistic biological networks.
To Weakness:
1. Similarity of performance of different classifiers. Both the classifi... | Summary: This paper proposes "attentive transfer entropy" as a way to infer connectivity of real and synthetic coupling networks. The submission suggests extending a 'neural estimator of mutual information' to transfer entropy and adding "attention" to focus on moments in time where the influence is larger.
Strengths... | Rebuttal 1:
Rebuttal: We thank the referee for noting the strengths of our method: it provides higher performance and requires fewer data.
To weakness and issues:
1. Discrete and continuous dynamics: In the analytical example and in the writing, for convenience, we used subscripts $t$ and $t+1$ because both experimen... | Summary: The authors extend the MiNE method [23], of using a neural network to estimate mutual information, to estimating transfer entropy between two coupled time series. When the coupling is weak on average due to being intermittent, they propose a differentiable attention mechanism that learns to transiently gate th... | Rebuttal 1:
Rebuttal: We thank the referee for noting that our paper addresses an important issue.
To Weakness:
1. Consistency of ATEn. So far, we cannot theoretically demonstrate that ATEn is consistent in some way. At present, we have only experimentally verified that the attention model can converge to stability w... | Summary: This paper proposes a novel method to reconstruct coupled networks from observational time series data, where the coupling effects between variables are sparse and weak. The authors propose a new attention mechanism called coupling attention is introduced to identify critical regions in time series where coupl... | Rebuttal 1:
Rebuttal: We thank the referee for stating that ATEn is well motivated and for offering us constructive suggestions for improvement.
To Weaknesses:
1. Preventing overfitting.
The coupling attention model is trained to maximize ATEn on the train samples. When the ATEn reaches stability (or converge to t... | null | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
ALGO: Synthesizing Algorithmic Programs with Generated Oracle Verifiers | Accept (poster) | Summary: This paper presents ALGO, a new method for code generation for LLMs. Given a problem description, ALGO first generates a brute-force implementation and dozens of test cases. Then, ALGO generates efficient implementations and ranks them by the agreement with the brute-force implementation on the generated test ... | Rebuttal 1:
Rebuttal: Thank you for the thoughtful comments.
## 0. Terminology
We understand your concerns. We will replace the term *verifier* in the revision. To avoid confusion, we use the terms "verifier" and "oracle" with quotation marks in the rebuttal.
## 1. Motivation and discussion about existing works
We high... | Summary: The paper introduces a new approach that utilizes Language Models (LLMs) to generate an oracle program for a given task, which is a brute-force search algorithm that is likely correct but potentially suboptimal in terms of runtime. The oracle program is then used to rank candidate programs based on their equiv... | Rebuttal 1:
Rebuttal: Thank you for your valuable comments!
## 1 Inference cost estimation
We argue that ALGO is NOT significantly more expensive than the baselines. We compare ALGO's cost with the baselines in two aspects - generation and validation.
**ALGO has the same generation cost.** ALGO is a code generation fr... | Summary: This paper presents ALGO which is a technique for generating programs for algorithmic challenges like LeetCode problems. To be successful, a code solution must be both correct and efficient enough. The core idea in ALGO is to first ask a code LLM to generate a slow brute force solution that is likely to be cor... | Rebuttal 1:
Rebuttal: Thank you for your thoughtful comments and inspiring questions.
## 1 pass@k vs n@k
**We followed CodeT's exact setting.** The "setting" in Line 216 means that we used ALGO to rerank the solution candidates (**n=1000** per problem) provided by CodeT, so our results is comparable to those in CodeT's... | Summary: The paper tries to improve code generation ability. The main idea of the paper is as follows
1. Use an LLM to generate a brute force solution to contest coding problems -- the idea being that the hard part for such problems is not solving them but solving them given the time and memory constraints.
2. Have a... | Rebuttal 1:
Rebuttal: Thank you for the thoughtful comments.
## 1 ChatGPT baselines on CodeContests
We ran three experiments with ChatGPT as the coder on CodeContests: ChatGPT + sample tests, ChatGPT + ChatGPT-generated tests, ChatGPT + ALGO tests. The experiments were conducted on a 50-problem subset of CodeContests.... | Rebuttal 1:
Rebuttal: We thank the reviewers for their thoughtful reviews and comments. We would like to address some common issues mentioned in multiple reviews and clarify the main points of our paper. For reviewer-specific issues, we have responded in reviewer-specific rebuttals.
## 0. Reminders
**Terminology.** We ... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
On Single-Index Models beyond Gaussian Data | Accept (poster) | Summary: This paper studies the problem of learning a single-index model beyond Gaussian data. Prior work (Ben Arous et al. 2021) studies the optimization landscape of recovering the hidden direction $\theta^*$ given a known link function $\phi$ when the covariates are drawn from the standard Gaussian distribution; in ... | Rebuttal 1:
Rebuttal: We thank very much the reviewer for the sharp comments and discussion that will help improve the paper greatly.
- *Dimension dependent constants.* This is the major concern of the reviewer, and we would like to thank him for the important comment: indeed constants like $\sup_{\theta} E[\|\nabla M... | Summary: This paper extends the results of Ben Arous et al., 2021 for learning a single-index model with a known link function via online SGD beyond the Gaussian input setting. The authors introduce a condition called Local Polynomial Growth (LPG) and establish the convergence rate of online SGD under LPG for spherical... | Rebuttal 1:
Rebuttal: Thank you for your response and questions, hopefully they will help improve greatly the paper. We give a precise answer to your concerns below:
- *Necessity of LPG.* Thanks for raising this question that allows us to clarify an important point. In *the Gaussian case*, the definition in the infor... | Summary: The paper deals with the study of online SGD in single-index models of the type $f(\mathbf x)=\phi(\boldsymbol\theta^\intercal\mathbf x)$ for a generic non-linear scalar link function $\phi$. The work builds on previous results of Ben Arous et al and aims at going beyond the Gaussian-data assumptions therein. ... | Rebuttal 1:
Rebuttal: Thank you for your response, comments and paper suggestion that will help improve the paper. We answer precisely to the questions below:
We agree that the LPG condition requires an input distribution for which one can at least approximate the loss landscape in terms of the correlation. Surely, it... | Summary: The main contributions of the authors are as follows:
(a) Rather than assuming the features are i.i.d. Gaussian, the authors identify a condition called Local Polynomial Growth (LPG, see Def. 2) on the population loss, under which the analysis of Ben Arous et. al. still goes through without any change in the... | Rebuttal 1:
Rebuttal: Thank you for your response, we answer questions below:
**General answer**
- *Tempering the overclaim.* Regarding Yehudai and Ohad (2020) and Wu (2022)', both contain theorems specifically demonstrating strong recovery under specific family of functions and spherically rotationally invariant dis... | null | NeurIPS_2023_submissions_huggingface | 2,023 | Summary: This paper studied the online SGD of learning a single-index teacher model with the link function as the student model. This work generalized the results of [Ben Arous et al. 2021], where the information exponent of the link function governs the sample complexity, to the non-Gaussian dataset, where the author ... | Rebuttal 1:
Rebuttal: Thank you for your review, our response follows:
**Main response**
- *Simulations.* We've included simulations of learning single index models that verify our theory under rotationally invariant and non-Gaussian input distributions in a new draft of the paper. Specifically, we see that inputs u... | null | null | null | null | null | null |
Towards Symmetry-Aware Generation of Periodic Materials | Accept (spotlight) | Summary: The problem of generating periodic materials with deep models is considered, focusing on capturing their unique symmetries. SyMat, a novel material generation approach, is proposed, using a variational auto-encoder to generate atom types, lattice lengths, and angles. A score-based diffusion model with a symmet... | Rebuttal 1:
Rebuttal: Thank you for your positive feedback and comments! We hope your concerns and questions can be addressed by the following responses.
> **Q1: Contributions to ML method development are minor.**
We agree that our contributions to general generative models or other machine learning methods are not s... | Summary: this paper proposes a method to enforce required symmetric properties in output generated by diffusion models. this method is applied to the problem of generation of new materials with desired symmetric properties. the method comprise of VAE for the generation of atom numbers, A, and lattice structures, L, whi... | Rebuttal 1:
Rebuttal: Thanks for your valuable feedback and comments! We make detailed responses below to address your concerns and questions. We hope you will raise your score if we address your concerns.
> **Q1: Could you please explain why $P(M) = P(M')$ in Section 3.1? There are inter-atom, $A$, interactions, whic... | Summary: This paper proposes SyMat, a new approach to generate materials that capture physical properties of perodic structures. This method generates atom types, lattice matrix of the unit cell, and then atom positions, all while maintaining invariance to permutations, rotation, translation, and periodic transformatio... | Rebuttal 1:
Rebuttal: Thanks for your positive feedback and helpful suggestions! We believe your concerns and questions can be addressed by the following responses.
> **Q1: I don't fully agree with not needing to evaluate the ability of the model to reconstruct materials from their latent representations. 1: even if t... | Summary: This paper introduces a deep learning based method for the automatic generation of periodic molecules. At the core of their method lies a variational autoencoder and a score-based diffusion models, which generate lattice lengths and angles, and atom coordinates, respectively. Theoretical derivations show that ... | Rebuttal 1:
Rebuttal: Thank you for your positive feedback and insightful comments! We hope your concerns and questions can be addressed by the following responses.
> **Q1: The paper could do a better job at motivating the need for modeling symmetry.**
Thanks for your suggestion. The main motivation of modeling symme... | Rebuttal 1:
Rebuttal: We sincerely appreciate all reviewers' hard work in reviewing our paper and writing comments. We are glad that all reviewers give positive feedbacks to our work and make many insightful suggestions to help improve our paper.
In the rebuttal PDF file, we include an updated version of the original ... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Module-wise Adaptive Distillation for Multimodality Foundation Models | Accept (poster) | Summary: The paper proposes a method called OPTIMA for reducing the size of pre-trained multimodal models. It uses layerwise distillation to train smaller student models to match the hidden representations of larger teacher models. The authors track the contributions of individual modules and choose modules with greate... | Rebuttal 1:
Rebuttal: We appreciate the reviewer for the insightful comments. We hope our response addresses your concerns. Should you have additional questions, we are readily available for further discussion.
**The authors did not provide enough evidence/motivation for why OPTIMA is only applied to CoCa. The authors... | Summary: This paper proposes a novel knowledge distillation (KD) method for multimodal foundation models. Motivated by the observation that distilling different modules yields distinct performance, the authors propose to dynamically choose an optimal combination of the modules for KD. Specifically, the choice of the op... | Rebuttal 1:
Rebuttal: We appreciate the reviewer for the valuable suggestions. We hope our response addresses your concerns. Should you have additional questions, we are readily available for further discussion.
**The model architectures adopted in this paper are limited to CoCa-base models.**
The commonly adopted mu... | Summary: This work proposed a new knowledge distillation method for multimodule multimodality models. This work pointed out that different modality modules contribute differently to the model performance, and it proposed to sample different modality modules for distillation during training. The paper proposed a multi-a... | Rebuttal 1:
Rebuttal: **In Table 1, how is the CoCa-Tiny baselines finetuned? Are they finetuned by Image+Text+Multi distillation?**
Yes, CoCa-Tiny baselines are fine-tuned by “Image+Text+Multi” distillation, where all modules are simultaneously distilled.
**It is hard to find a fair comparison in Table 1. The method... | Summary: They use layer-wise distillation to reduce model size for multi-modal foundation models. They develop a modified-Thompson sampling algorithm named OPTIMA to address the non-stationarity of module contributions resulting from model updating.
Strengths: The problem is valuable to reduce size for large models. T... | Rebuttal 1:
Rebuttal: **The result is based on COCA, experiments on other foundation models are needed.**
The commonly adopted multimodal foundation models are often computationally expensive to train and test. Given our limited computing resources, we were constrained to allocate them solely to a specific family of m... | null | NeurIPS_2023_submissions_huggingface | 2,023 | Summary: This paper proposes a Module-wise Adaptive Distillation method, OPTIMA, for a multimodal foundation model based on a modified multi-arm bandit algorithm. Comparisons with baseline multimodal pre-trainings and multimodal distillation methods on four different multimodal datasets demonstrate the effectiveness of... | Rebuttal 1:
Rebuttal: We appreciate the reviewer for the detailed comments and valuable suggestions. We hope our response addresses your concerns. Should you have additional questions, we are readily available for further discussion.
**Is that reasonable to use the change in $L_{hidn}$ of "Img" to evaluate the reward ... | null | null | null | null | null | null |
Minimax-Optimal Location Estimation | Accept (poster) | Summary: This work revisits one of the most basic questions in parametric statistics: location estimation. Two location estimators are proposed and proved to be optimal under different criteria.
From the reviewer's point of view, overall this paper is well-written and easy to understand. The problem is well-motivated,... | Rebuttal 1:
Rebuttal: Thank you for your suggestions and comments.
Q1: We will provide more motivation before Algorithm 1, as discussed in the overall response.
Q2: There are prior works studying the finite-sample guarantees of MLE for point estimation (see [Spo11, Pin17, VdG00, Mia10]), but they tend to lose constan... | Summary: This paper studies the problem of location estimation. We obtain n i.i.d. samples from a distribution f(x-u*) where f is known and u* is fixed but unknown. The goal is to estimate u* as accurately as possible. This is a generalization of Gaussian mean estimation though it's not as difficult of the general mean... | Rebuttal 1:
Rebuttal: **Running times.** One challenge in expressing the running time is it depends on how the distribution is represented as an input. If it's just given as a density discretized to $N$ bins, then Algorithm 1 can be done in $O(N \log N)$ time by an FFT convolution to compute the likelihoods. Algorith... | Summary:
This paper considers the problem of estimating an unknown shift given samples from a known density. The authors provide two estimators for this problem” a point estimator and an output interval, and they demonstrate these estimators are minimax optimal in slightly different ways. They provide an empirical eva... | Rebuttal 1:
Rebuttal: Thank you for your questions and suggestions.
(1) The goal of our (long-term) project is to build a finite-sample minimax theory of parametric statistics. Classical theory is asymptotic, without finite-sample guarantees; location estimation is a simple setting to start building this theory, so w... | Summary: This paper is about location estimation from sampling. That is, given n samples from a known probability distribution f, which is shifted by an unknown value $\mu$, the goal is to estimate that shift. Gaussian mean estimation is a special case of this parametric estimation type of problem. The paper describes ... | Rebuttal 1:
Rebuttal: Thank you for your positive appreciation of our results. | Rebuttal 1:
Rebuttal: We thank the reviewers for the positive reviews of our work. In particular, it is encouraging that the reviewers find our work 1) tackling an important problem (reviewers cZWD, j3jn), with 2) strong motivations (reviewers ng2F, j3jn) as well as 3) using highly non-trivial techniques (reviewer cZWD... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Neural Functional Transformers | Accept (poster) | Summary: Designing architectures for processing the weights of neural networks is an exciting and relatively new area with great implications. This paper proposes a novel, attention-based architecture for processing NN weights. The proposed architecture is equivariant to the permutation symmetries of neural weight spac... | Rebuttal 1:
Rebuttal: Thank you for your careful review of our work. We will try to address your questions and feedback individually here:
> the technical novelty of the paper is somewhat limited with respect to prior works ([1], [2], [3])
In addition to extending the equivariant framework of [1,2], we highlight our ... | Summary: The paper studies an interesting problem of encoding weights of neural network using another network (Neural Functional Transformer or NFT). NFTs are built based on previous works, mainly DWS [31] and NFN [46], proposing permutation equivariant linear models to encode neural network weights. NFT's main contrib... | Rebuttal 1:
Rebuttal: Thank you for your review--we appreciate the detailed and thoughtful questions. We will work to disentangle Section 5.1 and improve the clarity of the writing. Here we aim to answer them individually:
> Intuitive explanation of the difference between Eq. 4 and 5 is missing (why the union of sets ... | Summary: This paper introduces a new architecture called Neural Functional Transformer (NFT), that uses an attention-based layer that is equivariant to neuron permutations. NFT is an architecture that is used for tasks where the model is taking an input another neural network's weights (in their case, either an MLP or ... | Rebuttal 1:
Rebuttal: Thank you for your thorough analysis and review. We have updated our paper to correct the minor typos and missing definitions you found.
> What about direct image classification with NFTs instead of first encoding the INR as an array and then performing classification with that array instead of t... | Summary: This paper aims to construct neural functional architectures that can handle high-dimensional weight-space objects. Specifically, this paper designs a novel class of weight-space models with neuron permutation symmetries using nonlinear self-attention and cross-attention layers. The proposed method is called n... | Rebuttal 1:
Rebuttal: Thank you for your suggestions and insightful questions.
> It seems straightforward to make self-attention permutation equivariant. What is the actual contribution of section 3.1? Do you compare the performance of original attention and your modified attention?
The core of our contribution is th... | Rebuttal 1:
Rebuttal: We thank the reviewers for their thoughtful feedback and questions. Here are some changes and new experiments we have done that will help strengthen this paper:
* Created a diagram illustrating the equivariant self attention (SA) layer (see PDF response)
* Performed ablation experiments on the in... | NeurIPS_2023_submissions_huggingface | 2,023 | Summary: The authors consider the principled design of neural functionals, i.e. neural
networks that take other neural networks as inputs. Example tasks include
predicting the generalisation error of a trained network for a given set of
hyperparameters, or trying to reconstruct the images or 3D object encoded in an
imp... | Rebuttal 1:
Rebuttal: Thank you for your feedback and thoughtful review.
> I believe the article could benefit from some discussion of key concepts regarding neural functionals, for example the type of tasks in standard benchmarks, to make the article easier to read on its own.
We will expand the Experiments (Section... | null | null | null | null | null | null |
Scissorhands: Exploiting the Persistence of Importance Hypothesis for LLM KV Cache Compression at Test Time | Accept (poster) | Summary: The paper proposes a method for reducing the memory requirement of inference for large autoregressive language models.
The proposed techniques is based on the empirical observation that the high attention score tokens in a transformer self-attention layer maintain a remarkable persistence for different token ... | Rebuttal 1:
Rebuttal: We appreciate the reviewer’s supportive comments on novelty and constructive suggestions on the expositions. We have updated our manuscript and provided further details below. We hope our detailed clarification will clear the doubt about the significance and position of our work.
**Q1: The exposi... | Summary: This work studies attention patterns of LLMs and find that high-valued token-token patterns repeat for particular pivotal tokens. Since these patterns repeat, they can be compressed by storing only the pivotal tokens thus reducing the KV-cache size during inference.
Recommendation:
I recommend acceptance. T... | Rebuttal 1:
Rebuttal: Thank the reviewer for the supportive comments, recognizing the novelty of our analysis, and thorough evaluations. We understand that another model family would further demonstrate the generality of our hypothesis and methods. However, we couldn't finish this within the time frame due to the time ... | Summary: This work aims to increase the throughput of inference system by reducing the size of the KV cache.
- The KV cache comes from saving forward pass computations and can take up up to 5x the model size.
The paper proposes to discard part of the KV cache, based on the observation that the set of tokens that get h... | Rebuttal 1:
Rebuttal:
We appreciate the reviewer’s thoughtful suggestions and supportive comments. We have updated our manuscript to provide further clarification and improve writing details.
**Q1: The paper has mostly discussed the improvement in terms of memory. What's the improvement on wall clock time?**
The mai... | Summary: The paper introduces a new way to compress KV cache for LLM serving. It showed the proposed approach can reduce the memory usage for inference up to 5 times. However, there is no evaluation on the performance on the generative datasets to justify the claim that there is no loss in the serving accuracy.
Stren... | Rebuttal 1:
Rebuttal: We are glad that the reviewer finds our hypothesis and method interesting. We have tried to address your questions carefully. We hope the reviewer can consider raising your score in light of our response.
**Q1: There are no benchmarks on the inference efficiency (latency and throughput) with ... | Rebuttal 1:
Rebuttal: **Response to All Reviewers:**
We thank reviewers [R1(h9xt), R2(UGwp), R3(g2hZ), R4(AWSm), R5(H8Ve)] for their thoughtful and highly supportive feedback! We were glad that the reviewers found the problem significant and interesting [R1, R3, R5], the observations and theoretical analysis insightfu... | NeurIPS_2023_submissions_huggingface | 2,023 | Summary: This paper presents a new KV cache mechanism that uses an attention importance score to drop out tokens from the cache in transformer models. This allows up to 5x compression of the cache without degradation on assessed tasks.
Strengths: This is likely to be useful to practictioners and researchers alike by m... | Rebuttal 1:
Rebuttal: Thank the reviewer for the supportive comments on the significance of our work. We have carefully considered the reviewer’s constructive suggestions and updated the manuscript accordingly.
**Q1: The evaluation appears to focus on 4 datasets: HellaSwag, MathQA, PIQA, and Winograd. But I wonder ... | null | null | null | null | null | null |
Joint Attribute and Model Generalization Learning for Privacy-Preserving Action Recognition | Accept (poster) | Summary: This paper deals with the privacy-preserving action recognition.
They propose a meta privacy-preserving action recognition (MPPAR) framework based on the concept of meta learning to improve both generalisation abilities w.r.t. novel privacy attributes & privacy attack models in a unified manner.
They simulat... | Rebuttal 1:
Rebuttal: >**Q1:** *"Formulas: the author eglects the main idea and intention behind the module design."*
**A1:** Thank you for your suggestion. Below we provide a brief description to clarify the role of each component and the intention behind each design. (We will expand the following descriptions and ad... | Summary: The paper proposes a novel privacy-preserving action recognition (PPAR) framework built around the MAML meta-learning framework by Finn et al.. The goal of PPAR is to train an anonymization model that is able to anonymize video data so that 1) an action recognition model can still predict the depicted action i... | Rebuttal 1:
Rebuttal: >**Q1:** *"The font size in the figures and tables is a bit small. To improve readability, the font size should be increased in these cases. Also, the space after Fig. 1 should be increased."*
**A1:** Thank you for the feedback. To improve readability, we will increase the font size in the figur... | Summary: Considering that current privacy-preserving action recognition are difficult to cope with novel privacy attribute and privacy attack model, this paper propose a Meta Privacy-Preserving Action Recognition (MPPAR) framework to improve generalization abilities. Specifically, inspred by meta learning, the authors... | Rebuttal 1:
Rebuttal: >**Q1:** *"The description of the task is not clear, making it difficult to understand the mission objectives. To protect the privacy of individuals in the video and identify the action behavior, why not directly convert the RGB video into a skeleton sequence? It is well known that actions can alr... | Summary: To deal with novel privacy attributes and novel privacy attack models that are unavailable during the training phase, this paper applies the meta learning to the privacy-preserving action recognition. As a result, this work improves both generalization abilities above (i.e., generalize to novel privacy attribu... | Rebuttal 1:
Rebuttal: >**Q1:** *"Are there any theoretical guarantees about the generalization of the proposed method?"*
**A1:** Given the training set $D_{train}$, the conventional training scheme, which directly training the model $\phi$ on the whole training set $D_{train}$, can be formulated as:
$$\min_{\phi} L(\... | Rebuttal 1:
Rebuttal: We thank all reviewers for recognition of our contributions (Reviewer **hKip**: "this paper proposes a solution to address a important and interesting problem"; Reviewer **4pCF**: "the proposed method effectively defends against the privacy attack while maintaining the accuracy of action recogniti... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Switching Temporary Teachers for Semi-Supervised Semantic Segmentation | Accept (poster) | Summary: The authors argue that the teacher-student coupling problem leads to the performance bottleneck in semi-supervised semantic segmentation. To this end, this paper proposes a novel Dual Teacher framework. Specifically, it introduces two teachers that are switched every epoch during training to supervise the stud... | Rebuttal 1:
Rebuttal: We appreciate the insightful comments and positive reviews. We address all the concerns below and will revise them in the final version.
>Q1. Why does the Triple Teacher framework perform worse than the Dual Teacher alternative in Table 5? As mentioned, to alleviate the teacher-student coupling p... | Summary: This paper addresses the task of semi-supervised semantic segmentation. It proposes dual teachers that are alternatively updated to supervise a single student. Improvements are obtained compared with previous methods.
Strengths: - Improvements over previous best methods on Pascal VOC and Cityscapes.
- The met... | Rebuttal 1:
Comment: This is an AC of #12117. First of all, thank you for your great contribution to NeurIPS 2023.
I have discussed with PCs that if the authors can share experimental logs in the rebuttal period. We have concluded that such logs can be regarded as code in some way so the authors share them in a comple... | Summary: This work proposes the dual teacher knowledge distillation algorithm. For teachers trained with EMA during KD, there exists the issue of conformation bias which has been tackled in previous works. The motivation of this work is to diversify the training strategy using the dual-teacher student algorithm, in a w... | Rebuttal 1:
Rebuttal: We appreciate your valuable comments and positive reviews. We address your concerns and questions below in detail.
>Q1. Did the authors apply similar augmentations to other methods?
A1. Yes, we utilize basic augmentations such as CutMix and ClassMix, which is a similar level of augmentation comp... | Summary: This work proposes a method for semi-supervised semantic segmentation of images.
To mitigate the coupling problem of the student-teacher framework, the author proposes Dual teacher framework and Implicit Consistency Learning.
Dual teacher framework consists of a pair of teacher-student models, where the two t... | Rebuttal 1:
Rebuttal: We appreciate your constructive feedback. We address all the concerns raised by the reviewer and provide some new experimental results to support our contributions.
>Q1. What are the main differences from PS-MT?
A1. We would like to emphasize the notable differences.
1. Our approach differs fr... | null | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Parameterizing Non-Parametric Meta-Reinforcement Learning Tasks via Subtask Decomposition | Accept (poster) | Summary: This paper proposes SDVT, which addresses the non-parametric task variability problem in Meta-RL. SDVT utilizes a bi-level task encoding structure, namely a categorical encoding to address non-parametric variability and a VariBad-style continuous task belief. SDVT also utilizes imagined rewards to improve gene... | Rebuttal 1:
Rebuttal: We appreciate the reviewer for acknowledging the importance of the problem and the methodology. We address the questions as follows.
### **Q1.** Why not Imagine State Transition Dynamics?
Our current approach focuses on tasks with varying reward dynamics and doesn't extend to imagining unseen st... | Summary: The paper introduces a novel context-based meta-reinforcement learning (meta-RL) methodology designed for tasks that can be decomposed into sub-tasks, such as the pick and place operation. To tackle this specific type of environment, the authors propose the adoption of a novel encoder variant capable of predic... | Rebuttal 1:
Rebuttal: We appreciate the reviewer for recognizing our contribution and posing many insightful questions. We are confident that by addressing these points, our manuscript will offer a more comprehensive and clearer understanding.
### **W1.** and **Q1.** "Non-parametric" Problem Formulation
We apologize ... | Summary: This paper proposes an encoder architecture for meta-RL for tasks with non-parametric variation. The motivation for proposing a new encoder is that in order to achieve better generalization to a test-set of tasks with previously unseen non-parametric task variations, some specialized inductive bias can help. T... | Rebuttal 1:
Rebuttal: We appreciate the reviewer for recognizing the many strengths of our work. In response to the concerns raised, we outline below how we plan to improve the presentation of our work. We are excited to incorporate your valuable feedback, thereby contributing to the RL community more effectively.
###... | Summary: Meta-RL has proven successful in generalizing RL across tasks. Current methods struggle to generalize across tasks with novel sub-task decompositions. The authors introduce SVDT, a GM-VAE method that decomposes tasks into sub-tasks that help with this form of generalization. Virtual training is employed to fur... | Rebuttal 1:
Rebuttal: We are grateful to the reviewer for recognizing the many strengths of our paper, and we have taken the opportunity to address the concerns raised regarding the perceived weaknesses in our work as detailed below.
### **W1.** Examination of Why SDVT-LW Outperforms SDVT
What our model aims for is t... | Rebuttal 1:
Rebuttal: We appreciate the reviewers for dedicating their time and effort to recognize the strengths of our paper, and for offering constructive feedback that will undoubtedly enhance the presentation of our work.
Below are global responses that address concerns raised by multiple reviewers. Additionally... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
ImageReward: Learning and Evaluating Human Preferences for Text-to-Image Generation | Accept (poster) | Summary: The paper presents a human preference model for text-to-image generation and a method to enhance text-to-image models using this preference model. To achieve this, the authors develop a human preference annotation pipeline and create a dataset consisting of generated images and human ratings. The proposed mode... | Rebuttal 1:
Rebuttal: Thank you for your valuable comments and suggestions.
### Weakness 1: Formatting issues
We sincerely apologize for the unintentional formatting issues and would appreciate your leniency towards these easily rectifiable issues. In the final version, we will promptly correct these formatting probl... | Summary: This study presents ImageReward, a general-purpose text-to-image human preference reward model. They have collected 137k expert preference dataset, which contain a lot of rating and ranking annotation. Additionally, the authors propose Reward Feedback Learning (ReFL), a direct tuning algorithm designed to opt... | Rebuttal 1:
Rebuttal: Thank you for your valuable comments, we will explain your concerns point by point.
### Weaknesses: More advanced backbone for ImageReward
> BLIP, being outdated, falls short in generating accurate and comprehensive image descriptions. An alternative approach is to employ more recent models like... | Summary: The paper explores human preferences and introduces an ImageReward mechanism, which can be employed for evaluating text-to-image generation. The study further enhances the performance of existing text-to-image generation models through a Reward Feedback Learning (ReFL) approach. The experiments primarily focus... | Rebuttal 1:
Rebuttal: Thank you for your valuable comments and suggestions!
### Question 1: On sufficiency of ImageReward experiments
> The number of generative models considered in the experiments seems insufficient. With only six text-to-image generation models used for the evaluation of the proposed ImageReward, ... | Summary: The paper introduces ImageReward, a new dataset of human preference over generated images given a text prompt. Human preference for images is rated across three dimensionalities: text alignment, image fidelity, and harmlessness. Using the dataset, they train a reward model to score the generated image and text... | Rebuttal 1:
Rebuttal: Thank you for your valuable comments and suggestions!
### Weakness 1: Usefulness of the ImageReward dataset
Thank you for your suggestion. However, there might be some misunderstandings about the function of the dataset and how we improve T2I models upon them:
1. RMs’ model-agnostic generalizati... | null | NeurIPS_2023_submissions_huggingface | 2,023 | Summary: This paper aims to improve text-to-image (T2I) from human preference feedback. They first collect a human rating dataset to train their human preference model, ImageReward. With the reward model, they further optimize a pre-trained T2I via the proposed Reward Feedback Learning (ReFL). The experimental results ... | Rebuttal 1:
Rebuttal: Thank you for your valuable comments, we will explain your concerns point by point.
### Weakness 1: Novelty of ImageReward & ReFL
Despite drawing inspirations from ChatGPT, our work introduces several crucial novelties for T2I scenarios:
1. We are pioneering in the introduction of RLHF to the Te... | null | null | null | null | null | null |
DIFUSCO: Graph-based Diffusion Solvers for Combinatorial Optimization | Accept (spotlight) | Summary: This paper introduces DIFUSCO, a novel graph-based diffusion model for solving NP-complete combinatorial optimization problems. The authors propose two variants of the diffusion model: one with continuous Gaussian noise and one with discrete Bernoulli noise. Through extensive experiments on the Traveling Sales... | Rebuttal 1:
Rebuttal: We appreciate the reviewers taking the time to evaluate our work. Please find our responses to your concerns below:
> About more Benchmark Problems
We would like to cite the comments by reviewer WFTx: “It is extremely solid study in all regards. .. The experiments are impressive in their breadt... | Summary: This paper proposes and evaluates the use of diffusion models for solving combinatorial optimization problems on graphs.
The goal is to construct a probability distribution on the solution space that puts weight on high-quality solutions.
First, a dataset of optimal or near-optimal solutions is generated offl... | Rebuttal 1:
Rebuttal: We appreciate your valuable insights and constructive feedback on our paper. Below we address your comments and suggestions:
> Distribution Multimodality and Non-autoregressive Model Limitations
The expressiveness of non-autoregressive models tends to be limited when confronted with multiple opt... | Summary: The paper, "DIFUSCO: Graph-based Diffusion Solvers for Combinatorial Optimization," introduces a novel methodology for solving NP-Complete combinatorial optimization problems using graph-based diffusion models, named DIFUSCO. By formulating each NP-Complete problem as a {0, 1}-valued vector with N variables th... | Rebuttal 1:
Rebuttal: We appreciate Reviewer NgPq for their time and feedback on our work. Let's address your concerns:
> The traditional heuristic method still performs better.
It's critical to highlight that our proposed method, in its simplicity, surpasses other neural solvers consistently. To date, there has been... | Summary: This paper proposes a new, non-autoregressive, approach to Neural Combinatorial Optimization. As in similar non-autoregressive NCO works, a heat map characterizing the likelihood of a node being part of a solution is generated. The novelty is in the use of a generative model based on diffusion, trained in a su... | Rebuttal 1:
Rebuttal: We express our sincere gratitude to the reviewer for providing us with valuable insights on our paper. We would like to address your concerns as follows:
> The performance on large TSP instances seems to significantly rely on the 2-opt heuristic
We value your suggestion on the application of the... | null | NeurIPS_2023_submissions_huggingface | 2,023 | Summary: This paper investigates using diffusion models to solve combinatorial optimization problems. The authors propose two alternative methods to adapt diffusion models to this task: (1) a discrete diffusion model with Bernoulli noise, and (2) a more standard continuous diffusion model, where post-processing is need... | Rebuttal 1:
Rebuttal: We appreciate the time and effort the reviewer put into examining our work and providing valuable feedback. We would like to address your concerns as follows:
> The feasibility of a solution is only treated as a Boolean.
In our current design, feasibility is treated as a binary-valued variable. ... | null | null | null | null | null | null |
Diffusion-Based Adversarial Sample Generation for Improved Stealthiness and Controllability | Accept (poster) | Summary: The paper proposes a framework called Diffusion-Based Projected Gradient Descent (Diff-PGD) for generating realistic and stealthy adversarial samples to deceive neural network models. Diff-PGD utilizes a diffusion model to guide the optimization process, ensuring that adversarial samples remain close to the or... | Rebuttal 1:
Rebuttal: We express our gratitude to the reviewer for conducting a careful review and providing valuable feedback. We highly appreciate the reviewer's recognition of the effectiveness of our method. In this response, we address the questions raised by the reviewer:
> The phenomenon of the vulnerability ... | Summary: This paper proposes to use diffusion models to improve the quality of adversarial examples. It improves the Projected Gradient Descent (PGD) attack by performing a few steps of Stochastic Diffusion Editing (SDEdit) before taking each projected gradient descent (i.e., adversarial) step. The proposed method, Dif... | Rebuttal 1:
Rebuttal: We thank the reviewer for the careful review and valuable comments, we provide answer below.
> The proposed method is much more computationally expensive (in terms of both time and GPU memory) than regular adversarial attacks, while its advantage over existing methods seems marginal (Fig 6(a)).
... | Summary: This paper propose a gradient-based adversarial attack method that use diffusion model to generate adversarial examples.The proposed method can generate adversarial examples with high stealthiness and controllability.
Strengths: 1.The proposed method can improve the stealthiness, anti-purification ability and... | Rebuttal 1:
Rebuttal: We express our gratitude to the reviewer for conducting a careful review and providing valuable feedback. We highly appreciate the reviewer's recognition of the effectiveness of our method. In this response, we aim to address the questions raised by the reviewer:
>The novelty of the paper is some... | Summary: This paper proposes a novel Diffusion-Based Projected Gradient Descent (Diff-PGD) framework which optimizes perturbations by using an off-the-shelf diffusion model. The proposed method works well on both digital world and physical world with high stealthiness.
Strengths: 1. The proposed method is novel and pe... | Rebuttal 1:
Rebuttal: We thank the reviewer for the careful review and valuable comments. Please see our response below.
> Lack of the efficiency report of the proposed method
We present the detailed efficiency (speed and memory occupation) of Diff-PGD in the section `The computational cost of Diff-PGD` in **General... | Rebuttal 1:
Rebuttal: # General Response
We thank the reviewers for their valuable comments on our paper. We are excited to see that all the reviewers (Reviewer PJE2, B84v, x4rZ, PvL7) identified the novelty of our technical contribution: utilizing DM to generate adversarial samples, appreciated the application to bot... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
MEGABYTE: Predicting Million-byte Sequences with Multiscale Transformers | Accept (poster) | Summary: The authors propose MegaByte, a decoder style model consisting of 3 parts: patching of bytes, processing of patches (with autoregressive model), and converting patches back to bytes (with autoregressive model). Their model allows for sub-quadratic attention, more expressive models for the same flop count, fast... | Rebuttal 1:
Rebuttal: Dear Reviewer W4cB,
We sincerely appreciate your thoughtful review and your kind words about the novelty and potential significance of our work.
"Extensions":
We report models with simple patch embedder, we use cross-attention for image and audio, but not for text. We used strided inference onl... | Summary: A new auto-regressive decoder-only Transformer variant for language modeling is presented which operates well on byte level. Similar to the Vision Transformer, the main idea is to operate on patches instead of single tokens (bytes). Those patches are composed of a fixed number of $P$ tokens. Then, a big global... | Rebuttal 1:
Rebuttal: Dear Reviewer RJHZ,
Thank you for your thoughtful review of our submission and for recognizing the novel aspects of our work. We appreciate your feedback and are eager to clarify the points you have raised:
Sub-quadratic computation runtime argument:
Thankyou for raising this issue, which is a b... | Summary: This paper proposes Megabyte, a Transformer architecture that combines global and local attention to scale sub-quadratically in sequence length. The authors evaluate Megabyte on byte-level language modeling, image modeling, and audio modeling.
Strengths: There is an extensive empirical validation, and Megabyt... | Rebuttal 1:
Rebuttal: Dear Reviewer bsD9,
We appreciate the time you have taken to review our paper and provide us with constructive feedback. We agree with your suggestions to improve the paper, and will address your concerns below:
Presentation and references:
We apologize for the presentation issues you've mentione... | Summary: This paper proposes a new architecture of sequential modeling called MEGABYTE. MEGABYTE splits the long data sequence into patches and encode the patches using the Global model. The output of Global model is forwarded into the Local model, which facilitates the within-patch encoding. The output of the Local mo... | Rebuttal 1:
Rebuttal: Dear Reviewer tnBu,
We greatly appreciate your positive and thoughtful review of our paper. We're glad you acknowledged our contributions and the effort we made in ensuring reproducibility. We would like to address your comments and questions:
Typo Correction:
We appreciate you pointing out the ... | null | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Randomized Sparse Neural Galerkin Schemes for Solving Evolution Equations with Deep Networks | Accept (spotlight) | Summary: The paper proposes a so-called randomized sparse neural galerkin (RSNG) algorithm to solve time-dependent PDEs. The basic idea is to
consider parameterized functions as ansatz spaces, with parameters that vary in time. Consequently, the method is of the sequential-in-time learning algorithms rather than global... | Rebuttal 1:
Rebuttal: We thank the reviewer for their comments and detailed reading of our paper.
> [...] where authors elaborate on the rather succinct notation, for instance in the appendix. [...]
The reviewer mentions several good points on how to further improve the clarity of the paper. We will address these in ... | Summary: The paper suggests a simple but effective way to train sequential-in-time models by randomly sampling weights to update. The method avoids overfitting and realizes speedup because only a small part of parameters is updated. The experiments show that the proposed model offers the best performance in terms of sp... | Rebuttal 1:
Rebuttal: We thank the reviewer for their comments and detailed reading of our paper.
> [...] reviewer recommends that authors show predicted solutions in figures to assess the smoothness of the prediction. Sometimes smoothness is key because some downstream tasks may perform additional differentiation on ... | Summary: The authors suggest an alternative optimization scheme for Neural Galerkin methods - a family of methods in which parameters of a neural network are optimized so that the gradients of the solution obey a given PDE with the notable quality being that rather than solving over all of space-time, the parameters of... | Rebuttal 1:
Rebuttal: We thank the reviewer for their comments and detailed reading of our paper.
> Were the PINN comparisons also performed with rational activation functions?
Tanh activations are usually a standard choice in the PINN literature, which is what we used too. However, to ensure the choice of the activa... | Summary: In this paper, the authors propose a modified Neural Galerkin scheme for the solution of time-dependent PDEs, aiming at reducing computational cost and improving accuracy by avoiding overfitting.
The target solution $u(x,t)$ is approximated with a nonlinear parameterisation $\theta(t)$, so that $u(x,t)\approx... | Rebuttal 1:
Rebuttal: We thank the reviewer for their comments and detailed reading of our paper.
> Some experiments might be expanded to further substantiate the claims (see questions below), but this should be straightforward to address
We provide additional figures/experiments that all further support the claims o... | Rebuttal 1:
Rebuttal: We thank the reviewers for their valuable comments and suggestions. Here is a summary of some of the key points in the response, including an uploaded PDF that includes additional figures:
- Additional numerical results (in response below and in figures in PDF) of ‘Methods of Line PINNs’ and othe... | NeurIPS_2023_submissions_huggingface | 2,023 | Summary: The paper introduces Neural Galerkin scheme to sparsely update neural network over the sequential-in-time training. In particular, the proposed method randomly update sparse subsets of network parameters at each time step. With randomized sparse updates, overfitting problem of sequential-in-time methods can be... | Rebuttal 1:
Rebuttal: We thank the reviewer for their comments and detailed reading of our paper.
> The proposed method is inspired by dropout in neural networks. Thus, I am curious about comparison results with PINN and CausalPINN, with different rates of dropouts applied.
To address the reviewer’s point, we perform... | null | null | null | null | null | null |
Latent exploration for Reinforcement Learning | Accept (poster) | Summary: This paper explores the exploration problem in RL algorithms by considering the action-space correlation. The approach is motivated by the fact that over-actuated systems with high-dof action space require correlated control actions. Such action-space correlation is ignored by action-space random perturbation ... | Rebuttal 1:
Rebuttal: We thank the reviewer for their comments and positive evaluation.
*Selection of log_std*
We have run additional ablation experiments on Hand Pose and Reorient, where we have tested initial values for log_std between -3 and 2. The following results show that values around 0 are optimal:
|std_ini... | Summary: This paper proposes an exploration method that adds noise to the last linear layer (latent) of the policy network to incorporate time-dependent noise. The method is tested on locomotion tasks and evaluated on the metrics of performance and energy consumption. Evaluation is done on PPO and SAC base algorithm an... | Rebuttal 1:
Rebuttal: We thank the reviewer for their comments and strong evaluation (soundess, presentation and contribution) and excellent questions. We hope to be able to provide sufficient motivation to reconsider the assigned score.
*It is unclear why the proposed method enabled energy-saving policies compared to... | Summary: This paper proposes a new method to conduct action exploration based on the latent space representation during the reinforcement learning process. The disturbance noise in this case is based on gSDE, and the further "interval-hold" mechanism makes the noise time-correlated. The authors conduct experiments on b... | Rebuttal 1:
Rebuttal: We thank the reviewer for their comments and score.
*Lack of exploration algorithms comparison: the authors mention many exploring techniques [20-37] in the related work section, but have not compared to any of those, nor add on top of those techniques to boost the performance.*
We compared Latt... | Summary: The paper proposes adding time- and actuator-correlated noise instead of simple gaussian noise to improve exploration in RL.
Strengths: 1. The paper is well written and easy to read.
2. The approach can be easily applied to any RL algorithm. And can be thought of as a generalization of earlier work in noise... | Rebuttal 1:
Rebuttal: We thank the reviewer for their comments and score.
*Is there any intuition as to which tasks are suitable for your method? It appears completely random that PPO outperforms everyone on handpose which is (I expect), an implicit subgoal in tasks like Baoding and Pen. Hence it is unclear why latti... | Rebuttal 1:
Rebuttal: This paper proposes a simple and easy-to-implement idea for exploration in reinforcement learning by exploiting the implicit knowledge in the weights of the policy network to drive exploration. We thank all reviewers for their assessment. We think the quality of our paper improved due to your com... | NeurIPS_2023_submissions_huggingface | 2,023 | Summary: The paper presents an extension to the previously introduced RL exploration method gSDE called Lattice. The main idea is to add noise to the full covariance matrix (instead of only onto the diagonal) of the multivariate Gaussian used for exploration. In particular, the authors multiply state dependent noise wi... | Rebuttal 1:
Rebuttal: We thank the reviewer for the precise summary and for praising the clarity of our exposition. We would like to address their concerns:
*no comparison to other exploration methods such as [11], [18] or [30]*
In our manuscript we have compared Lattice to SAC/PPO with two forms of action noise, sta... | null | null | null | null | null | null |
MMD-Fuse: Learning and Combining Kernels for Two-Sample Testing Without Data Splitting | Accept (spotlight) | Summary: [Update: After intensive discussion with the authors, I changed my score from 6 to 7. I further increase the score for "Contribution" from 2 to 3.]
The paper introduces two new test statistics for permutation-based two-sample tests that are related to the MMD -- called FUSE_N and FUSE_1. These are motivated b... | Rebuttal 1:
Rebuttal: **Reply to reviewer fFeS Part 1**
Thank you for your valuable feedback and questions which will help us improve the presentation of our ideas.
## Permutation tests
> The paper is very well written and clarifies important (although not really new) aspects of permutation-based tests in an accessi... | Summary: This paper looks into the problem of permutation based kernel two-sample testing. The paper is motivated by two main challenges:
1. The first one being the fact that selecting kernel hyper-parameters through standard data splitting often reduces the statistical power due to using less data than we actually h... | Rebuttal 1:
Rebuttal: *[Comment to all reviewers (Author Rebuttal) posted above]*
Thank you for your generous praise, and comments. We will incorporate answers to the following into our revised work.
> There is some recent work on not using permutation test at all to do two sample test, such as the work of Sh... | Summary: The paper studies the question of choice of critical values for the MMD test. The authors introduce a new MMD-based test that allows data-driven critical values based on permutation test, without the need to train a new kernel for every permutation. The main advantage of this new method is that it does not req... | Rebuttal 1:
Rebuttal: *[Comment to all reviewers (Author Rebuttal) posted above]*
Thank you for your thorough feedback, positive evaluation of our work, and for pointing out ways in which we can more clearly communicate our contribution. We also thank you for pointing out the small error in the proof of Th 2 due to in... | Summary: This paper proposes a new statistic that can incorporate with the permutation test where we can learn features of data in an unsupervised way. From the perspective of testing, this paper contributes a new idea and paves a new way to perform two-sample testing without data splitting. If we can safely use all da... | Rebuttal 1:
Rebuttal: *[Comment to all reviewers (Author Rebuttal) posted above]*
Thank you for your valuable feedback, for highlighting the strengths of our paper, as well as ways in which we can present and motivate our ideas more clearly. We will include the above discussions into the body of the paper.
> ... | Rebuttal 1:
Rebuttal: # Comment to all reviewers
We warmly thank all reviewers for their time and detailed commentary on our work. We are highly encouraged by the positive evaluations of all reviewers: **reviewer 1tja** points out that our FUSE method *'paves a new way to perform two-sample testing without data splitt... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Towards Combinatorial Generalization for Catalysts: A Kohn-Sham Charge-Density Approach | Accept (poster) | Summary: The paper proposes a new learning framework for using machine learning to approximate chemical simulations. While many papers focus on direct property prediction (e.g. energy, forces) when applying machine learning to chemical simulation, this paper motivates the prediction of charge-density which holds promis... | Rebuttal 1:
Rebuttal: Thanks for your review.
Regarding the performance of SCN on the entire dataset, please see the distributions of metrics in Figures 1 and 2. Most scores fall into this range of -0.5 and 3 in Figure 1, and between -0.2 and 0.8 in Figure 2.
Regarding other models, we did try several other models l... | Summary: In this work, the researchers created a novel dataset of bulk catalyst relaxations with charge densities using open-source software. They aimed to investigate the phenomenon of combinatorial generalization, which refers to the ability of a model to effectively perform on new combinations of elements that were ... | Rebuttal 1:
Rebuttal: Thanks for your review. We realize we didn't make a number of details clear in the main text. We will make sure to add these in the next revision.
(1)
The model is trained with an regression loss. Following other work in ML for catalysis, final prediction losses are reported as "mean absolute er... | Summary: - The authors investigate charge density modeling in catalyst systems. They generated a dataset using Quantum Espresso on Materials Project, defined a metric of SCF convergence savings, and tested models, focusing on generalization and comparing it to ACP initialization. They used GNNs with virtual atoms as qu... | Rebuttal 1:
Rebuttal: Thanks for your review.
(0)
We did try other model architectures such as Schnet, Dimenet++, and Gemnet, but we found the SCN to be the best performing. Our contribution is not specifically about finding the best architecture for density prediction, rather it is about investigating a specific ki... | Summary: The authors build and train approximately E(3) equivariant graph neural networks based on Spherical Channel Network to learn the Kohn-Sham Charge-Density scalar field as a means of accelerating density functional theory (DFT) calculations by shortening the number of self-consistent field calculations needed to... | Rebuttal 1:
Rebuttal: We thank the reviewer for the thoughtful review.
(1)
We are happy to add a discussion of Schütt et al. (2019) to our related work section. In short we think our results are complementary. The primary differences between our works is that (1) we study learning of the charge density for bulk cata... | Rebuttal 1:
Rebuttal: We thank the reviewers for their feedback. We address common concerns of reviewers below:
* (1) strength of performance and number of structures
* (2) assessing impact of results on ML+DFT community
* (3) comprehensiveness of evaluation
* (4) bias of learned densities
---
(1)
To address concer... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
On Slicing Optimality for Mutual Information | Accept (poster) | Summary: The paper estalishes a new quantity for measuring the dependence between random variables. The definition is similar to mutual information and sliced mutual information. As (for rather obvious reasons) it takes values between these two existing measures, it has some desireable properties. A variational represe... | Rebuttal 1:
Rebuttal: We thank you for the positive review and helpful comments. Below, we address your comments and questions.
**Why (2) is better than (1)?** In low rank settings, Sliced mutual information (1), $\mathcal{SI}$ [1] performs quite well despite the fact that our method (2) still surpasses $\mathcal{SI}$... | Summary: Recently, sliced mutual information ($\mathcal{SI}$) has been proposed as a statistical dependence measure that is scalable to high dimensions (Goldfeld and Greenewald, 2021). However, due to the use of uniform sampling for the slicing directions, $\mathcal{SI}$ generally requires many slices for accurate esti... | Rebuttal 1:
Rebuttal: Many thanks for the constructive feedback. We address all your concerns below.
1. Reviewer posed a good question. In lines 140-148, we discussed the reasoning behind the definition of eq 2 and continued to offer evidence that shows the effectiveness of our criteria. We further discuss the questio... | Summary: This work based on existing ideas of providing a tractable independence metric for high dimensional random variables, the proposed metric named "Optimal sliced mutual information" (SI*) improves upon existing method of uniform slicing to an adaptive optimal slicing with scalability design in mind. It has been ... | Rebuttal 1:
Rebuttal: We thank you for the positive review and helpful comments.
We conduct an experiment on the NN estimator at the sample size level while varying $d_x,d_y$. We have $X,Y\in\mathbb{R}^d$ with linear dependence. The results are in the attached pdf above. One can tell, similarly to the convergence anal... | Summary: This paper proposes a new dependence measure based on the sliced mutual information with both the theoretical analysis and the modern machine learning experiments. This measure is not strictly the mutual information but still could capture the correlation between two distributions.
The experiments are done in... | Rebuttal 1:
Rebuttal: We thank you for the positive review and helpful comments. Below, we address your concerns and questions in order.
1. Followed from Corollary 3, the Lagrangian dual form in Eq.6 is completely equivalent to the original definition of $\mathcal{SI}^*$. The last two terms in Eq.6 account for the $\ar... | Rebuttal 1:
Rebuttal: All reviews and feedback are appreciated, and many suggestions will be incorporated in the paper.
Attached you can find a file containing new experiments suggested by the reviewers.
* Reviewer 8cdY: (1) AUROC with high dimensions (>100); Figure 1. (2) An RL experiment on Half-Cheetah-v3; Figure ... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Generalised f-Mean Aggregation for Graph Neural Networks | Accept (poster) | Summary: The authors deal with the interesting problem of parametrising the space of aggregation functions. They propose a novel aggregation method that allows each MPNN to learn the most appropriate aggregator for each downstream task. In contrast to previous works, their proposed aggregator can interpret a broad pool... | Rebuttal 1:
Rebuttal: Thank you for your review! We will address the weaknesses that you cited:
- If I understand this correctly, this is about the explainability of GenAgg—is that correct? You are correct, some datasets (such as images, as you mentioned) are less likely to lead to an easily interpretable aggregator. ... | Summary: This paper presents GenAgg, a generalised aggregation operator, which parametrises a function space that includes all standard aggregators. They show using GenAgg as a drop-in replacement for an existing aggregator in a GNN usually leads to a boost in performance across various tasks.
Strengths: 1. The propos... | Rebuttal 1:
Rebuttal: Thank you for the review! To address the weaknesses:
- You’re right, GenAgg does introduce some computational overhead compared to an aggregator like ‘sum’. Unfortunately, an experiment where we tweak the architectures until all achieve the same latency is not possible—sum is a subcomponent of Ge... | Summary: The paper proposes a generalized aggregation operator, GenAgg, for GNNs that can significantly improve performance in various tasks. The authors argue that existing aggregation methods are limited in their expressiveness and propose GenAgg as a more flexible alternative. They demonstrate the effectiveness of G... | Rebuttal 1:
Rebuttal: Thank you for your review, you bring up some good points. We will first address the weaknesses:
- Yes, the method section is about 2 pages—the architecture itself is relatively simple (we regard this as a good thing). However, the rest of the paper also contains technical content, such as the exp... | Summary: EDIT: Due to author responses below, I have increased my score from a 6 to a 7.
This work proposes a generalized method for parameterizing the aggregation operation of a message passing neural network (i.e. a graph neural network, GNN). This relies on a modification of the generalized $f$-mean that adds two h... | Rebuttal 1:
Rebuttal: Thank you for your thoughtful review! We will address the weaknesses:
- The closest thing to a DeepSets-based GNN in prior literature is the original MPNN. Despite its representational capacity, subsequent research has focused on more constrained methods. A direct comparison can be seen in the PN... | Rebuttal 1:
Rebuttal: Two reviewers have requested that we add a table about the computational overhead of GenAgg to our Limitations section. We include our results in the attached PDF.
Pdf: /pdf/7289244ebc305a6c88fca272189c4e4e5adc94ab.pdf | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Vocabulary-free Image Classification | Accept (poster) | Summary: This paper proposes the task of vocabulary-free image classification - image classification without a predefined set of categories (which is required in zero-shot classification). The paper further proposes the first method for this task - Category Search from External Databases (CaSED). For each given image,... | Rebuttal 1:
Rebuttal: **[W2: How to select databases]** The choice of the database is indeed crucial for CaSED, as shown in Tab. 6 of the Supplementary. We also perform additional experiments and calculate the average semantic similarity between the closest caption retrieved and the ground-truth for all the images in t... | Summary: This work introduces Vocabulary-free Image Classification (VIC), a task that aims to assign an image to a class in a large, evolving semantic space without a known vocabulary. The proposed method, Category Search from External Databases (CaSED), utilizes a vision-language model and an external database to extr... | Rebuttal 1:
Rebuttal: **[W1: Relationship with [1]]** The cited paper was released on arXiv concurrently (April 5th) with the deadline of NeurIPS (May 17th), making it “contemporaneous” according to the NeurIPS call for papers. The provided OpenReview link leads us to a withdrawn paper from ICLR 2023, with unanimously ... | Summary: Current approaches to zero shot classification are limited by the assumption of an existing set of candidate classes. This paper presents a novel task, Vocabulary-free Image Classification, where semantic categories need to be automatically mined. This is important when generalising to a new domain, or the dom... | Rebuttal 1:
Rebuttal: **[W1: Text-to-text score]** When performing text-to-text scoring with a text-only model, Sentence-BERT [a], and considering only text scores (i.e., alpha=0), we get 39.8% cluster accuracy (-2.9%), 51.9 semantic similarity (+1.6), and 16.6% semantic IoU (-0.4%) w.r.t. CLIP. Thus, CLIP is better at... | Summary: The paper presents a method to perform image classification by using the similarity text and image embeddings of a large vision-language model (VLM). The method is called "vocabulary-free" because the large vision-language model doesn't have a particular set of classes (vocabulary) in mind, and (contrary to ze... | Rebuttal 1:
Rebuttal: **[W1: VLM built-in vocabulary]** Strictly speaking, VLMs, such as CLIP, process sequence of tokens. Thus, they have a possibly infinite vocabulary induced by all the possible combinations of tokens, i.e., the tokenizer can process any string/class name. While there is no performance guarantee for... | Rebuttal 1:
Rebuttal: We thank the reviewers for their valuable feedback. We are especially glad the reviewers appreciated i) the novel task, with a well-thought evaluation protocol and baselines (m2jc, xx1r, 3fhc), ii) the proposed training-free method, which consistently outperforms multiple baselines on several data... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Global Identifiability of $\ell_1$-based Dictionary Learning via Matrix Volume Optimization | Accept (poster) | Summary: This article proposes a new method for dictionary learning. The proposed optimisation problem minimises the determinant (volume) is derived from the l1 formulation of dictionary learning. Global identification results are proved under a general assumption of being 'sufficiently scattered'. The case of a Bernou... | Rebuttal 1:
Rebuttal: Thank you very much for the positive assessments. Let us address some of your concerns:
> There are no comparisons with previous methods in the experiments section of the main paper
We did compare with some existing methods that claim recovery guarantees in Appendix D.2. The advantage of our pro... | Summary: The paper studied the classic problem of dictionary learning. A new method is proposed based on optimizing the determinant of the dictionary matrix. Deterministic and probabilistic analysis are conducted for global recovery of the dictionary and coefficients. Experiments are provided to help demonstrate the e... | Rebuttal 1:
Rebuttal: Thank you for the positive assessments. Let us address some of your concerns:
> **Novelty**. A crucial related work is Tatli and Erdogan [2021] as the authors pointed out, which studied (1) except that they bound the norm of columns of S while this paper did so for rows of S. What differs in the ... | Summary: This paper presents a novel formulation for dictionary learning by minimizing the determinant of the dictionary matrix, subject to the constraint that each row of the sparse coefficient matrix has unit $\ell_1$ norm. The proposed method provides global identifiability guarantees for the ground truth dictionary... | Rebuttal 1:
Rebuttal: Thank you for the positive assessments. Regarding some of your concerns:
1. *Global optimality*. Unfortunately, dictionary learning is known to be a computationally hard problem, and most existing methods that provide guarantee to a global minimizer requires some kind of "good-enough" initializati... | null | null | null | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Simple, Scalable and Effective Clustering via One-Dimensional Projections | Accept (poster) | Summary: $k$-means clustering is a classical and fundamental method in machine learning. This paper proposes a very simple method that projects the $n$ data points onto a one-dimensional space by generating a random Gaussian vector. Moreover, the one-dimensional projections can be used for coreset construction and the ... | Rebuttal 1:
Rebuttal: Thank you for your helpful feedback! Here is our response to the individual comments and questions in your review:
**Dimension reduction techniques.** We would like to emphasize that our dimension reduction analysis is much more sophisticated than simply showing that the pairwise distances betwee... | Summary: The paper proposes a new clustering algorithm that seeks to have the accuracy of k-means++ while avoiding the time complexity of O(ndk). In particular, the paper focuses on settings where k is large and seeks a O(nd+nlogn) time complexity. This is a practically relevant problem.
The main algorithm proposed in... | Rebuttal 1:
Rebuttal: Thank you for the insightful feedback! Here is our response to the individual comments and questions in your review:
**Advantage over Lightweight coresets.** Compared to lightweight coresets, a crucial property of our algorithm is that it has a rigorous guarantee on its worst-case (multiplicative... | Summary: The submitted paper focuses on the problem of unsupervised clustering and aims at improving the $\Omega(ndk)$ runtime of the ```kmeans++``` algorithm in the clustering of $n$ $d$-dimensional datapoints in $k$ clusters. The proposed algorithm relies on the intuition of running ```kmeans++``` on a one-dimensiona... | Rebuttal 1:
Rebuttal: Thank you for your comprehensive review! Here are our answers to your questions:
**Sampling more than one vector.** Thanks for the great question! An obvious way to obtain improved performance by sampling more than one projection vector would be independently running our algorithm multiple times ... | Summary: The paper presents a new technique for clustering data such that the running time of the method is independent of the number of clusters, and the approximation error is polynomial only in the number of clusters and not the number of data points.
The paper experimentally tests the method and finds that it is ... | Rebuttal 1:
Rebuttal: Thank you for the thoughtful review! Here is our response to the individual comments and questions in your review:
**Experimental results.**
The experiments in the paper are mainly meant to demonstrate that we achieve significant speedups over the baseline algorithms (see Table 1 in the submissio... | Rebuttal 1:
Rebuttal: Here is a one-page pdf containing a figure and a table showing additional experimental results. More details can be found in our responses to the reviewers.
Pdf: /pdf/b958ad2d18773d82699172004ce721ee74f29ab8.pdf | NeurIPS_2023_submissions_huggingface | 2,023 | Summary: This paper introduces a randomized clustering algorithm that provably runs in expected time O(nnz(X) + n log n) for arbitrary k. In theoretical analysis, the proposed algorithm achieves approximation ratio O(k^4 ) on any input dataset for the k-means objective. In the experiments, the quality of the clusters f... | Rebuttal 1:
Rebuttal: Thank you for the helpful feedback! Here is our response to the individual comments in your review:
**Research on related work.** In our submission we extensively discussed prior results on fast algorithms for clustering with a focus on results with provable guarantees. We discussed algorithms in... | null | null | null | null | null | null |
DatasetDM: Synthesizing Data with Perception Annotations Using Diffusion Models | Accept (poster) | Summary: The paper proposes a new framework called DatasetDM that leverages the rich visual knowledge contained in diffusion models to generate synthetic data and corresponding annotations for various perception tasks, The core component of DatasetDM is the P-Decoder for generating perception annotations. The primary p... | Rebuttal 1:
Rebuttal: Thanks to the reviewers for the valuable comments.
**Weakness 1**: Comparison with VPD.
**R1**: We want to emphasize that our work and VPD (unpublished) are concurrent works. And we argue that our work is not a resemblance for the following reasons:
- **Extracting features.** Using the diffusi... | Summary: The paper introduces a novel dataset generation method producing diverse synthetic images along with perception annotations including semantic/instance segmentation, and depth/pose estimation. The technique leverages a pretrained text-to-image generation model, i.e., Stable Diffusion, to extract cross-attentio... | Rebuttal 1:
Rebuttal: Thanks to the reviewers for the valuable comments.
**Weaknesses 1**: Comparison with VPD (novelty). **R 1**: We want to emphasize that our work and VPD[1] (unpublished) are concurrent efforts. Besides, we believe that our work and VPD are entirely different for the following reasons:
- **Differ... | Summary: With the success of text-to-image generation models, the authors propose a novel approach to text-to-data that can produce synthetic data by adding a perception decoder (P-Decoder) on top of Stable Diffusion, which only requires a few labeled images for training. The authors show various experimental results t... | Rebuttal 1:
Rebuttal: Thanks to the reviewers for the valuable comments.
**Weaknesses 1**: In Section 4, the DatasetDM results are compared to the baselines trained with less than 1000 real images, which is unfair regarding the number of training samples (e.g. 400 vs. 80k). To claim the effectiveness of the proposed ... | Summary: The paper presents a framework based on pre-trained Diffusion Models to synthesize images along with their perception annotation for various downstream perception tasks. With limited real labeled images as a baseline, the extra synthetic training pairs can boost the performance by a large margin.
Strengths: -... | Rebuttal 1:
Rebuttal: Thanks to the reviewers for the valuable comments.
**Weakness 1**: Using the synthetic training pairs to boot an already well-established baseline with all available real training pairs, for example, using all the 10,000+ training images on Pascal VOC.
**Response 1**: Thank you for the valuabl... | Rebuttal 1:
Rebuttal: Thanks to all reviewers and ACs for the valuable comments and suggestions. We have summarized and addressed the main concerns raised by each reviewer as follows:
**To Reviewer 21oc.** Our work differs from Handsoff and DatasetDDPM primarily in terms of uniqueness and contribution (lines 69 - 85).... | NeurIPS_2023_submissions_huggingface | 2,023 | Summary: Large-scale datasets with quality annotations are often costly and time-consuming to collect and annotate. Since for most applications, large amounts of annotated data are required by deep neural networks, the authors propose a framework for text-guided synthetic data generation - called DatasetDM - which can ... | Rebuttal 1:
Rebuttal: Thanks to the reviewers for the valuable comments.
**Weakness 1-1**: Comparison against Handsoff at the **method** levels.
**R1**: We provide a comparison as follows:
- **Task.** HandsOff designs several CNN layers for the decoder, which cannot handle tasks such as instance segmentation and ope... | null | null | null | null | null | null |
Generalizing to Unseen Domains for Regression | Reject | Summary: The authors tackle domain generalization for regression tasks, or DGR. DGR is often different from classification tasks in DG as well as imbalanced regression tasks due to shift in labels.
Popular feature alignment type of algorithms using IPM may not be applicable to DGR because the labels of different domai... | Rebuttal 1:
Rebuttal: ## Dear Reviewer 2irV:
Thanks for your effort in giving detailed and constructive comments. Here are our point-to-point responses.
> Weakness-1: *The relationship between the margin and hardness that the authors demonstrated is not something new.*
**Response:** Thanks for your valuable comments... | Summary: This paper addressed the problem of domain generalization for regression, where the authors assume that different domains may have disjoint labeling space. The authors proposed a meta-learning-based method to learn the parameter initialization for DGR. In particular, the choosing frequency of different support... | Rebuttal 1:
Rebuttal: ## Dear Reviewer phkn:
Thanks for your effort in giving detailed and constructive comments. We hope your concerns can be successfully addressed by our point-to-point responses.
> Weakness-1: *It is unclear why using meta-learning to tackle the DGR problem in this paper. Although a range of works... | Summary: This work investigates and analyzes domain generalization for regression tasks and propose a margin-aware meta-learning method for DG regression, which can help learn long-range exploration and interpolation. Also, a benchmark for evaluating regression dg is built.
Strengths: - the paper is presented and deri... | Rebuttal 1:
Rebuttal: ## Dear Reviewer tR9s:
> Weakness-1: *Is there any possibility to apply the proposed meta-learning based DG to more general regression tasks like object detection?*
**Response:** Thank you for your support of our work. For more general regression tasks, e.g., object position regression and gaze ... | Summary: The paper describes an extension of domain generalization to the regression domain - traditionally, most of the research in Domain Generalization has occurred in image classification and to some degree in combined tasks such as object detection, depth estimation, and semantic segmentation and there has been a ... | Rebuttal 1:
Rebuttal: ## Dear Reviewer JfrS:
We appreciate your support of our work and your efforts in giving detailed and constructive comments. We hope your concerns can be successfully addressed by our point-to-point responses.
> Weakness-1: *The limitations section is currently limited to 3 bullets, none of whi... | Rebuttal 1:
Rebuttal: ## Global Response:
Thanks for the efforts of reviewers, we are encouraged by the detailed and constructive comments.
We have provided comprehensive responses to the comments from each reviewer. Additionally, we have prepared a **one-page PDF** containing the necessary justification experiments ... | NeurIPS_2023_submissions_huggingface | 2,023 | Summary: This paper focused on the generalization problem in the regression context and proposed an interesting setting, Domain Generalization in Regression (DGR). Compared to domain generalization for classification, DGR is a more challenging task due to its different domain shift forms between source and target domai... | Rebuttal 1:
Rebuttal: ## Dear Reviewer ajvr:
We appreciate your support of our work and your efforts in giving detailed and constructive comments. We hope your concerns can be successfully addressed by our point-to-point responses.
>Weakness-1: *Some assumptions in the theoretical analysis can be further discussed. F... | null | null | null | null | null | null |
Partial Multi-Label Learning with Probabilistic Graphical Disambiguation | Accept (poster) | Summary: The paper focuses on a specific problem called partial multi-label learning (PML), which is an emerging weakly-supervised learning problem. PML involves inaccurate annotations, where each training example is associated with a set of candidate labels, among which only some labels are valid. The paper proposes a... | Rebuttal 1:
Rebuttal: ------
Thank you for your reviews and insightful comments. The following are our responses to the **Questions** and **Weaknesses**:
------
1. *Any idea about how to explore alternative implementations for principled disambiguation?*
* **Response 1**: Thanks to the comments. We believe that new i... | Summary: In the manuscript, the authors adopt the probabilistic graphical model to solve the disambiguation problem in partial multi-label learning. Specifically, a directed graph is tailored to describe the generative process of partial multi-label data, and a partial multi-label learning approach named PARD is propos... | Rebuttal 1:
Rebuttal: ------
Thank you for your reviews and insightful comments. The following are our responses to the **Questions**:
------
1. *The motivation is not clear. Why do authors apply probabilistic graphical model for disambiguation in PML?*
* **Response 1**: Thanks to the comments. The probabilistic grap... | Summary: The paper introduces PARD , a new approach for addressing the problem of partial multilabel learning. PARD uses a probabilistic graphical model to recover ground-truth labeling information. By leveraging the framework of stochastic gradient variational Bayes, PARD simultaneously identifies labeling informatio... | Rebuttal 1:
Rebuttal: ------
Thank you for your reviews and insightful comments. The following are our responses to the **Questions** and **Weaknesses**:
------
1. *The derivation from Eq1 to Eq2 can be more detailedly explained.*
* **Response 1**: Thanks to the comments. Detailed derivation process has been provided... | Summary: This paper presents a novel, systematic approach to partial multi-label learning (PML) and label disambiguation using a probabilistic graphical model and the stochastic gradient variational Bayes framework. The authors offer a reliable method tested extensively on synthetic and real-world datasets. The results... | Rebuttal 1:
Rebuttal: ------
Thank you for your reviews and insightful comments. The following are our responses to the **Questions** and **Weaknesses**:
------
1. *Could the authors provide further clarification regarding the omission of the marginal distribution on $\mathbf{x}$?*
* **Response 1**: Thanks to the com... | null | NeurIPS_2023_submissions_huggingface | 2,023 | Summary: This paper attempts to tackle the Partial Multi-Label Learning task by employing the probabilistic graph method. The proposed method employs the directed graph to model the generative process of partial multi-label data, then derives a variational lower bound objective with a gumbel-softmax trick and closed fo... | Rebuttal 1:
Rebuttal: ------
Thank you for your reviews and insightful comments. The following are our responses to the **Questions**:
------
1. *The motivation of this paper is not clear. In this paper, the authors attempt to employ the probabilistic graphical model to perform a principled way for disambiguation and ... | null | null | null | null | null | null |
Adversarial Examples Might be Avoidable: The Role of Data Concentration in Adversarial Robustness | Accept (poster) | Summary: The paper studies the problem of adversarial robustness in multi-class classification. The authors show that a necessary condition for the existence of an adversarially robust classifier is the condition that at least one of the class conditional ditributions is "concentrated" in the sense that most of the pro... | Rebuttal 1:
Rebuttal: Thank you for finding our proof techniques and our polyhedral certificates novel, and the paper well written. We address your questions and concerns below:
> Q2: Is it possible to find the radius of the largest ball around x that fits inside C(x)?
Good question. While a closed form solution is n... | Summary: This paper gives a theoretical insight into the existence of robust classifiers for avoiding adversarial examples, which may be a potential direction for bridging the gap between human perception and machine intelligence. Specifically, this paper reveals the sufficiency and necessity of data concentration for ... | Rebuttal 1:
Rebuttal: Thank you for finding our paper well organized and our theory solid. We address your questions below:
> The empirical results are insufficient. The experiments are only conducted on a small dataset, whether the theoretical conclusion still holds on larger natural datasets is unknown.
**CIFAR-10 ... | Summary: The paper shows that adversarial examples can be avoided allowing for robust classification as long as the data distribution obeys certain properties. In particular, they show that the the input data distribution needs to be "well-concentrated" and not overlap too much in order to guarantee the existence of a ... | Rebuttal 1:
Rebuttal: Thank you for finding our paper well written, and our theory interesting. We address your comments below:
> My main concern is that Def 3.1 is possibly too strong. Is there some way to [relax it]?
Great question. In the [global response][9], we explain why Def 3.1 is not strong, is satisfied by ... | Summary: This paper first studies the existence problem of a robust classifier from a data concentration perspective. In particular, they prove that the (strong) concentration property of class-conditional data distributions characterizes a sufficient and necessary condition for determining whether a robust classifier ... | Rebuttal 1:
Rebuttal: Thank you for finding our paper well written and the results novel and interesting. Here are our responses to your comments:
> Q1. Detailed comparison to [1-4].
The results in these works support our claims. We detail several aspects below:
---
**Comparison of notions of concentration**
*Conc... | Rebuttal 1:
Rebuttal: **Comment to Everybody**
We thank the reviewers for their time and thoughtful reviews. We address broader questions here, and leave the rest to individual responses. At a high level, we have explained the importance of our concentration theorems, their applicability to real world datasets, and the... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Improving Few-Shot Generalization by Exploring and Exploiting Auxiliary Data | Accept (poster) | Summary: The paper presents a new approach to Few-shot Learning with Auxiliary Data (FLAD), a method that uses additional data to improve model generalization. Traditional methods for integrating auxiliary and target data have limitations, particularly in terms of computational scale with an increasing number of auxili... | Rebuttal 1:
Rebuttal: We thank the reviewer for their insightful questions and overall appreciation of our work, we share an excitement about the proposed methodology.
>Figure 2 presents an interesting picture of the model's performance, revealing an interesting dependency on the auxiliary data for some tasks, while s... | Summary: The paper addresses the problem of using auxiliary data for the few-shot learning. The main novelty of the proposed approach is to formulate the selection of the auxiliary dataset from a pool of datasets as a multi-arm bandit problem (balancing exploration and exploitation) and adapting EXP3 and UCB1 to that ... | Rebuttal 1:
Rebuttal: Thank you for your very insightful questions and for the opportunity to respond with the clarifications below:
>The novelty of the paper is incremental
We would like to clarify what we believe is an important misunderstanding. While the building blocks of our algorithms may be derived from exist... | Summary: In the context of Language Models, this paper studies the use of Auxiliary Data during training to gain in generalization. The Few Shot settings is also considered here where proposed tasks only have limited set of labeled data. Compared to previously proposed approaches using auxiliary data, here it is consid... | Rebuttal 1:
Rebuttal: Thank you for the overall supportive review of our work, and for the insightful question about complexity which led us to a very interesting analysis that we now include in the updated draft.
>As mentioned by the authors, the proposed training has to be performed from scratch for every domain.
I... | Summary: This paper focuses on the problem of few-shot learning with auxiliary data (FLAD) problem. The authors formulate FLAD as a multi-armed bandit problem, and employ exploration-exploitation algorithms to tackle it. Experiment results show that the proposed method outperforms exploration or exploitation-only appro... | Rebuttal 1:
Rebuttal: We thank the reviewer for their useful feedback, which points out some details of the paper that required clarification and for the suggested studies that can strengthen the paper. Our responses are below:
>The novelty is limited since the proposed multi-armed bandit method is a existing techniqu... | Rebuttal 1:
Rebuttal: We would like to express our gratitude to the reviewers for their valuable and detailed feedback. We appreciate the time and effort they took to review our draft and provide us with their insightful comments. We carefully consider all of the suggestions and incorporate them into our final version.... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Deep Non-line-of-sight Imaging from Under-scanning Measurements | Accept (poster) | Summary: This paper proposes a new deep learning based method for NLOS imaging reconstruction. Their method is an end-to-end deep network that directly acts on under scanning transient measurements to recover the NLOS image. The use of a single pass network enables fast, realtime inference. The key contribution is the ... | Rebuttal 1:
Rebuttal: We thank the reviewer for giving us valuable comments. The main concerns are addressed below.
**Q1-1:** *The ML contribution of this paper is limited. ... So I'm not sure of the suitability of the paper for a venue like Neurips.*\
**Reply:** First of all, we would like to kindly remind that our s... | Summary: 1, This paper proposed a new approach to NLOS reconstruction from USM based on deep learning, and achieved high-quality and rapid inference.
2, Two main reconstruction modules, dubbed transient recovery network (TRN) and volume reconstruction network (VRN), are designed to produce the hidden volume and the int... | Rebuttal 1:
Rebuttal: We thank the reviewer for giving us valuable comments. The main concerns are addressed below.
**Q1:** *The main modules in this work are heavily based on existing work. These points indicate that the contribution of this work is not convincing enough, although the performance is satisfying.*\
**R... | Summary: The paper proposes a deep learning based approach to non-line-of-sight (NLOS) imaging from under-scanning measurements (USM). This is achieved by first using a transient recovery network (TRN) that increases the spatial resolution of the input, and a volume reconstruction network (VRN) that uses linear physics... | Rebuttal 1:
Rebuttal: We are highly encouraged by the positive recommendation and comments from the reviewer. The main concerns are addressed below.
**Q1:** *What is the motivation behind using a Gaussian weighting instead of bilateral weight which preserves discontinuity?*\
**Reply:** Indeed, bilateral weights are co... | Summary: This paper introduces a deep-learning-based method for non-line-of-sight (NLOS) imaging from under-scanning measurements (USM). It comprises two main components: the transient recovery network (TRN) and the volume reconstruction network (VRN). The TRN extracts features and produces high-resolution measurements... | Rebuttal 1:
Rebuttal: We are highly encouraged by the positive recommendation and comments from the reviewer. The main concerns are addressed below.
**Q1:** *The paper does not mention how noise is modeled.*\
**Reply:** For the noise modeling of the NLOS simulation, the formulation is as follows:\
$$ \hat{\tau}(\math... | Rebuttal 1:
Rebuttal: # This is the global rebuttal for all the reviewers including a PDF containing only figures. Please find individual responses to each reviewer below.
The content is:
**Fig 1**: Qualitative results about the noise level experiments and our imaging system.
As suggested by Reviewer r4bJ, we build ... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Grounded Decoding: Guiding Text Generation with Grounded Models for Embodied Agents | Accept (poster) | Summary: The paper introduces a novel method to tackle a key challenge in utilizing Internet-scale knowledge from LLMs (Language Model Models) in robotics tasks. Specifically, the focus is on aligning the instructions generated by LLMs with the actual actions performed to manipulate objects. The paper endeavors to fram... | Rebuttal 1:
Rebuttal: We thank the reviewer for recognizing the novelty, experimental evaluations, and applicability of the proposed method to various domains. We also appreciate the constructive feedback. Please see below for our responses.
>*”The paper could benefit from improvements in certain sections of the writi... | Summary: The paper introduces Guiding Text Generation with Grounded Models for Robot Control. Previous works have shown that large pre-trained language models can generate long action plans for robots based on abstract instructions. However, these models struggle to handle unstructured environments with large action sp... | Rebuttal 1:
Rebuttal: We thank the reviewer for recognizing the number of experiments in diverse domains. We appreciate the constructive feedback, please see our responses below.
>*Lack of robot control*
- We recognize the comment that the title may not best suit the work and intend to amend the name to “Grounded Dec... | Summary: This paper studies grounding Large Language Models (LLMs) for robot control. The planning tokens are generated by the joint probability of an LLM and grounded models sequentially. The grounded models can be RL value functions, multi-modal perception models or human rules. As an extension of SayCan, this method... | Rebuttal 1:
Rebuttal: We thank the reviewer for the positive remarks on the comprehensiveness of our experimental results as well as the presentation quality. Please see below as we provide more clarification to the raised points.
>*”The contribution of this paper is mainly incremental to SayCan”*
- The main distincti... | Summary: This paper proposes Grounded Decoding, a methodology to ground the high level semantic plans output by an LLM into a realizable action sequence. Specifically, the authors construct a model which decodes a sequence that is likely to be realizable given the current state of the robot and environment, but also li... | Rebuttal 1:
Rebuttal: We thank the reviewer for the positive remarks on the comprehensiveness of our experimental results as well as the presentation quality. Please see below as we provide more clarification to the raised points.
>*”The authors include results on just two environments in simulation and one in the rea... | null | NeurIPS_2023_submissions_huggingface | 2,023 | Summary: This paper proposes Grounded Decoding (GD), a method that filters the decoded task descriptions by the probability of a grounded model, LLM-based task planning. This paper shows the framework can integrate different types of grounded models. Experiments across physical and simulated environments show that this... | Rebuttal 1:
Rebuttal: We thank the reviewer for the constructive feedback. Please see our responses below.
>*”the current form to include safety and preference is quite preliminary… these should be represented as statements”*
- The main contribution of GD is presenting the general framework of guiding LLM-based planni... | null | null | null | null | null | null |
Make Pre-trained Model Reversible: From Parameter to Memory Efficient Fine-Tuning | Accept (poster) | Summary: >**Rebuttal:** The provided details satisfy my concerns. I think this paper should be accepted after applying the agreed changes.
>**TL;DR:** The proposed MEFT reduces training activation memory, which is an important topic for the ML community. MEFT outperforms existing PEFT techniques while consuming up to ... | Rebuttal 1:
Rebuttal: Dear reviewer, we would like to thank you for your time, effort, and thorough reviews! We are really encouraged by the reviews that highlight our work:
(1) Tackles an important problem in the existing PEFT methods.
(2) Offers insightful observation for the success of existing PEFT methods.
(3) ... | Summary: This paper points out that the existing parameter-efficient fine-tuning (PEFT) methods are not memory-efficient, since they need to cache most of the intermediate activations for the gradient calculation. To resolve this problem, the authors introduce to use the reversible model to reduce the activation memory... | Rebuttal 1:
Rebuttal: Dear reviewer, we would like to thank you for your time, effort, and thorough reviews!
We have responded to your comments below and updated our draft accordingly. If you still have additional questions and suggestions, we would be happy to answer your questions and incorporate the suggestions in... | Summary: Authors propose a memory-efficient (MEFT) variant of parameter-efficient fine-tuning (PEFT), which utilizes reversible networks to save memory stored for activations during model forward propagation. The authors showed insights into the reasons for the effectiveness of PEFT. They evaluated several variants of ... | Rebuttal 1:
Rebuttal: Dear reviewer, we would like to thank you for your time, effort, and thorough reviews! We are really encouraged by the reviews that highlight our work:
(1) Offers competitive results;
(2) Conducts insightful experiments for LoRA's initialization;
(3) Is useful for practitioners.
We have respo... | null | null | Rebuttal 1:
Rebuttal: We would like to thank all reviewers for your time, effort, and thorough reviews! Regarding typos, presentation improvement, and missing baselines, we have incorporated suggestions into our updated draft.
Here we offer new results for some common concerns:
> 1. Compare MEFT to gradient checkpoin... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
LLM-Pruner: On the Structural Pruning of Large Language Models | Accept (poster) | Summary: This paper tackles the problem of reducing the parameter number of large language models with pruning.
It identifies structures within an LLM and uses them to compose a dependency graph-based.
The authors provide experiments where they prune the parameters of LLMS and evaluate them on zero-shoot tasks.
Namel... | Rebuttal 1:
Rebuttal: We truly appreciate the reviewer for the constructive comments.
---
> As the authors show, pruning effectively reduces the parameter of LLMs. But how does the pruned model compare to the next smaller model variation? Showing that with the proposed method, we can get a better-pruned model (7B to ... | Summary: The paper is trying to present an resource efficient task agnostic technique for pruning large language models. The authors tackle two main problems, 1) in a setup where there is no access to the original dataset that was used to train the LLM, and 2) the pruning should be fast.
The paper presents a three ste... | Rebuttal 1:
Rebuttal: We truly appreciate the reviewer for the constructive comments.
---
> W1: The paper starts with the problem that deployment \& inference are the key challenges faced by large language models, and presents a solution to it. However, no assessment of latency is provided in the paper. If we prune s... | Summary: The authors propose LLM-Pruner to compress large language models in a task-agnostic manner. Their method can achieve a fast recovery by tuning LoRA after pruning the model. Their experiment results show LLM-Pruner can achieve comparable performance with much smaller memory.
Strengths: 1. Their motivation is ... | Rebuttal 1:
Rebuttal: We truly appreciate the reviewer for the constructive comments.
---
> W1: What's the performance if no post-tuning is applied?
Thanks for the comment. In Table 1-3 in our submitted manuscript, we present the results without post-training (please refer to `Ratio=X% w/o tune`). Without post-train... | Summary: This paper present a method for LLM pruning, where some parameters are discarded by sacrificing the downstream accuracy.
The proposed method, LLM-Pruner, adopts structural pruning that selectively removes non-critical coupled structures based on gradient information.
By investigating different pruning settings... | Rebuttal 1:
Rebuttal: We truly appreciate the reviewer for the constructive comments.
---
> W1: Missing the comparison with the smaller models. For example, when pruning 50\% parameters for 7B model, how the results compared with a 3B model?
Thanks for the constructive suggestion. As advised, we show the experimenta... | null | NeurIPS_2023_submissions_huggingface | 2,023 | Summary: This work proposes an LLM compression framework, LLM-Pruner, that employs structural pruning to compress LLMs in a task agnostic manner so that their multi-task solving is preserved. They achieve this without using the original training data. In addition, they employ adaptors (LoRA) during the tuning phase for... | Rebuttal 1:
Rebuttal: We truly appreciate the reviewer for the constructive comments.
>W1: Although the authors presents the reduction in parameters, they do not mention the inference speedup with and without LLM-pruner. What would be specially interesting would be the inference speed under different types of group pr... | null | null | null | null | null | null |
Unlocking Feature Visualization for Deep Network with MAgnitude Constrained Optimization | Accept (poster) | Summary: The paper proposes MACO, a feature visualization technique that optimizes image phase but keeps the magnitude of the Fourier spectrum constant. This method is similar, in spirit, to the 'classic' feature visualizations by Olah et al. (2017), while at the same time enabling feature visualizations for architectu... | Rebuttal 1:
Rebuttal: We thank the reviewer for their time and for providing this detailed review with very relevant comments.
## Weaknesses:
**The strongest contribution is MACO itself. The approach of combining attribution and feature visualization, on the other hand, sounds like a great idea if at least one of tho... | Summary: ## Summary
This paper presents MACO, a simple and novel method to generate feature visualization for deep neural networks. MACO uses a constraint to optimize feature visualization in the space of natural images. It also presents an attribution-based mask to augment generated images to highlight spacial import... | Rebuttal 1:
Rebuttal: First of all, we thank the reviewer for taking the time to read our article.
## Major
**W1: The evaluation can be improved. The goal of feature visualization is to help human users interpret computer vision models. However, I don't see a strong connection between the proposed evaluation scores (... | Summary: This paper proposes a small but apparently effective modification to the common technique of feature visualizations to improve the visual quality of the produced explanations. Specifically, this approach uses a different image parameterization producing images with power spectra similar to that of natural/data... | Rebuttal 1:
Rebuttal: We thank the reviewer for taking the time to read our article, and we sincerely appreciate their review.
- **Section 4: The authors claim that they are the first to introduce quantitative methods for comparing feature visualizations. This is, however, incorrect as Borowski et al. [47] and Zimmer... | Summary: The paper proposes MACO, a feature visualization method which focuses the generation process of the visualizations on the image's phase spectrum while keeping the magnitude constant. This makes the generated explanation visualizations lie in the space of natural images, and hence, clearer for visual inspection... | Rebuttal 1:
Rebuttal: We thank the reviewer for taking the time to read our article and for the valuable feedback.
## Weaknesses
- **W1: The proposed method for generating the explanation visualizations raises from the observation that human recognition of objects in images is driven not by magnitude but by phase (l... | Rebuttal 1:
Rebuttal: We extend our gratitude to the reviewers for dedicating their time to thoroughly read and assess our paper. Their critiques have been insightful, and we believe all their concerns can be addressed. In particular, we appreciate that all five reviewers have converged on the following two issues (w... | NeurIPS_2023_submissions_huggingface | 2,023 | Summary: This paper proposes a simple yet effective approach to visualize deep learning features without using a parametric image prior. Improving upon the prior work that uses the Fourier domain formulation, the proposed method uses the polar form and optimizes the image phase to maximize the activation of specified n... | Rebuttal 1:
Rebuttal: We thank the reviewer for dedicating their time and for their kind words regarding our work.
## Major weaknesses
- **Lack of significant novelty. The other two contributions claimed in the paper are less convincing: the second contribution is proposed in the SmoothGrad paper as acknowledged by t... | null | null | null | null | null | null |
Glance and Focus: Memory Prompting for Multi-Event Video Question Answering | Accept (poster) | Summary: This paper presents a model for video question answering called "Glance-Focus". The idea is to first "Glance" at a video to obtain a set of N dynamic memories, each represented as vectors of dimension D.
The memories can either be learned in an unsupervised way or a supervised way. In the unsupervised case, ... | Rebuttal 1:
Rebuttal: We thank all the suggestions for a more comprehensive evaluation of our model.
**Ablation Study On the Loss Functions**
Due to space limitations, we provide a complete ablation study in the Supplementary Materials. According to the reviewer's suggestion, we try to give a more detailed analysis of... | Summary: The paper proposes a novel approach for multi-event video question answering (VideoQA) that mimics the human reasoning strategy of glancing and focusing. The proposed method consists of two stages: glancing and focusing. In the glancing stage, the model generates a set of episode memories that summarize the ke... | Rebuttal 1:
Rebuttal: *We will certainly improve our writing in the revised version.*
We thank for the reviewer motivating us to make a detailed analysis of how the event categories influence the model's performance.
**Comparisons of the overall Number of Event Categories Among Datasets**
We first compare the over... | Summary: This paper proposes a two-stage method of constructing memories of periodical summarization of the video, then using memory prompts to local key frames for reasoning.
Strengths: 1. The SOTA results on three Ego QA datasets.
2. The hierarchical construction of video abstraction might be a new way of tackling l... | Rebuttal 1:
Rebuttal: *We will certainly improve our writing in the revised version.*
**One-To-One Matching & Soft-Attention**
The one-to-one matching is only used under the supervised setting following the recently proposed series of detection works [DETR](https://arxiv.org/abs/2005.12872)/[MDETR](https://arxiv.org... | Summary: This paper aims to tackle the problem of multi-event VideoQA that involves multiple human-object interaction events. Inspired by the human's reasoning ability, this paper proposes the Glance-Focus model that can reason from coarse to fine. At the glancing stage, the Glance-Focus model can generate some dynamic... | Rebuttal 1:
Rebuttal: We are very glad to see all the positive reviews you given. Hope we can address your questions below.
**Relations with unsupervised learning**
We will add more relevant explanations in the revised version. Similar to individual discriminative learning, we aim to generate a set of event-level ind... | Rebuttal 1:
Rebuttal: **We thank all reviewers for the insightful feedback. We are encouraged that you find our motivation interesting and reasonable. We try to address the common issues below.**
P.S. For convenience, we directly link the paper of the related works. The best results in the tables are **bolded**, and ... | NeurIPS_2023_submissions_huggingface | 2,023 | Summary: The paper introduces a model called Glance-Focus for Video Question Answering (VideoQA) tasks, which involve understanding human behaviors in videos. While vision language models have shown success in multi-modal tasks, reasoning over long videos with multiple human-object interaction events remains challengin... | Rebuttal 1:
Rebuttal: **Corrections on STAR Dataset**
We are very grateful for the reviewer's reminder. The following are the updated results, with GF in the unsupervised setting achieving SOTA on both the Test and Val parts.
**Table 1. Corrections on STAR evaluations.**
| Model ... | Summary: This paper addresses the task of question-answering in videos which have multiple events. The authors draw inspiration from the human capability to create event memories as a way to anchor the reasoning process over multiple events, and propose the Glance-Focus model, which aggregates a set of event memories (... | Rebuttal 1:
Rebuttal: **Scale to longer video**
We thank the reviewer's suggestions. The evaluations on longer VideoQA benchmarks (NExT-QA/ActivityNet-QA) are shown in the [global rebuttal](https://openreview.net/forum?id=J6Niv3yrMq¬eId=LQPJMIxMFG). Note that the model is tested under the unsupervised setting since... | null | null | null | null |
Kernel Quadrature with Randomly Pivoted Cholesky | Accept (spotlight) | Summary: The kernel quadrature problem computes quadrature weights and nodes that minimize the maximum quadrature error over all functions in RKHS with the norm bounded by 1. This paper proposes new quadrature rules for functions in RKHS using nodes drawn by randomly pivoted Cholesky algorithm. The randomly pivoted Cho... | Rebuttal 1:
Rebuttal: We thank the reviewer for their careful reading and their many insightful questions and suggestions. Our responses are as follows:
## Applications
The new version of the paper discusses applications more and includes a new experiment demonstrating an application to analyzing large chemistry data... | Summary: This paper introduces a new method for kernel quadrature which relies on a randomly pivoted Cholesky decomposition.
The sampling of the quadrature points is done thanks to a rejection sampling algorithm. To improve performance, an optimized rejection sampling procedure is also introduced, so that the acceptanc... | Rebuttal 1:
Rebuttal: We thank the reviewer for their thoughtful feedback, and we are delighted that they appreciated the paper. Below, we address the reviewer’s comments and questions:
## Proof of Proposition 3
At the suggestion of yiWH, the proof of proposition 3 has been deferred to the supplementary material. We ... | Summary: The authors provide theoretical analysis and show that RPCholesky has near optimal error rates.
Meanwhile, the authors propose to combine RPCholesky with rejection sampling in order to use it in the continuous setting. Experiments show that the proposed method is compared favorably to continuous volume samplin... | Rebuttal 1:
Rebuttal: We are glad the reviewer found the paper well-written and found the experiments to “demonstrate the good performance of the proposed method”. We thank the reviewer for their suggestions for improvement, all of which we have accepted and incorporated into the main text:
- We now provide a derivati... | Summary: The recently proposed Randomly Pivoted Cholesky procedure is an efficient and accurate method for approximating symmetric PSD (e.g. kernel) matrices. It builds a Nystrom approximation via randomizing the choice of pivot in the classical Cholesky decomposition procedure.
The authors of the present submission tr... | Rebuttal 1:
Rebuttal: Thanks to the reviewer for their considered remarks. We are very glad to hear the reviewer found the theory and algorithms of the paper to be solid and practical. Below, we address the reviewer’s questions and suggestions:
## Application
At the suggestion of this review and others, we have incl... | Rebuttal 1:
Rebuttal: We thank the reviewers for their very thoughtful comments. We have revised the paper to address any questions and concerns raised and have accepted all specific recommended changes by all reviewers. We have responded to the questions of each reviewer in the individual responses, and we look forwar... | NeurIPS_2023_submissions_huggingface | 2,023 | Summary: This article studies a family of quadrature rules suited for functions that belongs to an RKHS. The proposed construction is based on an algorithm called randomly pivoted Cholesky (RPCholesky). The latter was studied for the task of approximating positive-semidefinite matrices. In particular, the authors exten... | Rebuttal 1:
Rebuttal: We appreciate the reviewer's very incisive comments. We are delighted to hear that they found the article well-written and that they enjoyed the further connections our article draws between kernel quadrature and numerical linear algebra. Below, we address your concerns and questions:
## Curse of... | null | null | null | null | null | null |
Stability-penalty-adaptive follow-the-regularized-leader: Sparsity, game-dependency, and best-of-both-worlds | Accept (poster) | Summary: The paper studies the problem of designing no-regret algorithm with three types of adaptivity to the setting: sparsity, game-dependency, and Best-of-Both-Worlds. To do so, the authors propose a new adaptive learning rate, called Stability-Penalty-Adaptive learning rate for Follow the regularized leader. They p... | Rebuttal 1:
Rebuttal: Thank you for your valuable time and for providing valuable suggestions.
> The paper presents many different results and hence it is hard to follow the paper in some points.
Indeed, multi-armed bandits and partial monitoring are very different, and the results obtained with each problem setting... | Summary: This paper proposes a new scheme called SPA to configure adaptive learning rates for FTRL in MAB. Developed from a stability-&penalty-dependent generic regret bound for FTRL, SPA is designed to select learning rates to optimize the bound. SPA is then applied to three scenarios with different types of adaptivit... | Rebuttal 1:
Rebuttal: Thank you for your valuable time and for carefully reading our paper.
> The design of SPA is rather straightforward, which seems a direct consequence of existing generic regret bound for FTRL. This somehow limits the novelty of the proposed method (though the subsequent extensions to scenarios wi... | Summary: This paper studies how to achieve BOBW regret bounds in sequential decision making problems such multi-armed bandits and partial monitoring games. To this end, this paper comes up with novel SPA learning rates for FTRL. For MAB, this paper shows FTRL with SPA could simultaneously be near-optimal for stochastic... | Rebuttal 1:
Rebuttal:
Thank you for your time and for providing a possible alternative procedure for exploiting sparsity in the adversarial regime.
> The improvement upon the previous literature seems not big enough. The main improvement for sparse MAB seems to be removing the sparsity requirement. It is possible I ... | Summary: This paper studies the Follow-the-Regularized-Leader (FTRL) framework for bandit problems. It proposes a learning rate dependent on both stability and penalty terms, which is inspired by the analysis. The proposal is helpful to achieve sparisty, game-dependent and Best of both worlds regret for bandit problems... | Rebuttal 1:
Rebuttal: Thank you for your time and helpful suggestions.
> It would be better to provide comparison on regret bounds with $\Delta_{\min}$ and $L_2$ with previous results (some of them don't have these parameters) in Table 1.
Thank you for the suggestion for improving Table 1.
We may not fully understand... | null | NeurIPS_2023_submissions_huggingface | 2,023 | Summary: This work considers the Follow-the-regularized-leader paradigm for multi-arm bandit and partial monitoring problems. It proposes a new framework for designing FTRL algorithms with a time-dependent learning rate to generate algorithms that is adaptive to the sparsity of the feedback, to the setting being stocha... | Rebuttal 1:
Rebuttal: Thank you for your time and insightful suggestions.
> While all the notations used in this paper are defined before being used, it would be great if some context/discussion can be provided before diving into the statements and analyses. For example, what is the intuitive way to see that $\mathrm{... | null | null | null | null | null | null |
DesCo: Learning Object Recognition with Rich Language Descriptions | Accept (poster) | Summary: The paper describes an approach to open world object detection that generates detailed object descriptions with GPT to improve performance. The main advance of the paper is showing that the removal of the entity name, and just using it's description, forces the network to learn the "context". This is turn lead... | Rebuttal 1:
Rebuttal: 1. More details on the training and testing procedure.
Thank you for raising the issue and we will include more details in the next version.
We have included the detailed data generation process in the attached PDF (Figure 1) and we briefly introduce the process in the following. DesCo is train... | Summary: This paper is about open-vocabulary object detection, in which a model is given an image and a sequence of object descriptions and has to predict bounding boxes matching the descriptions. While open-vocabulary detectors can in principle take arbitrarily detailed object descriptions as input, they tend to focus... | Rebuttal 1:
Rebuttal: Thank you for your detailed review.
1. “Insufficient Contributions”
We list our core contributions in the general response, and we would like to emphasize the following points.
As Reviewer Wg2t pointed out, DesCo is the first approach to support complex-language based detection, while “recent ... | Summary: This paper leverages rich descriptions generated by LLMs to encourage GLIP/FIBER to pay more attention to object attributes aside from object names when performing detection tasks. A more robust model is resulted from training on object-description alignment as the model is able to leverage attributes for reco... | Rebuttal 1:
Rebuttal: Thank you for the detailed review and constructive comments.
**1. False-positive rate / reject grounding attempt / error analysis**
A. The main metric AP penalizes false positives
While we did not explicitly include a false positive metric, the AP metric we used already penalizes false ... | Summary: The paper tackles the problem of open-world (language-based) object detection and proposes two main things. First, it uses an LLM to obtain detailed descriptions of the object classes of interests and, secondly, proposes a training methodology that forces the network to pay attention to such descriptions throu... | Rebuttal 1:
Rebuttal: Thank you for the review!
1. Clarification on the setting
Thank you for the suggestion! We used “object recognition” as an umbrella term because we believe the proposed technique can be applied to other vision tasks as well.
Indeed the specific experiment setting is object detection. We follo... | Rebuttal 1:
Rebuttal: We thank all the reviewers for the constructive reviews and suggestions! We are encouraged that the reviewers find our paper “**insightful and well-shaped for acceptance**” (Reviewer Wg2t); “demonstrates a viable path of distilling the knowledge present in LLMs into better object grounding abiliti... | NeurIPS_2023_submissions_huggingface | 2,023 | Summary: This work revisits a key aspect of open-vocabulary object detection: what makes for the class names? It shows that GLIP, the state-of-the-art vision-language model for open vocabulary object detection, behaves like bag of words and could not effectively understand text descriptions of the objects to detect. To... | Rebuttal 1:
Rebuttal: Thank you for the view!
1. Failure cases of the LLM-generated descriptions
Thank you for the suggestion! In Figure 4, we included failure cases of the smaller LM; but indeed even the largest LM can generate inaccurate descriptions. We will include examples in the next version.
2. More diverse d... | null | null | null | null | null | null |
Quantifying & Modeling Multimodal Interactions: An Information Decomposition Framework | Accept (poster) | Summary: This work is mainly concerned with the methodological contribution of estimators for partial information theory statistics, to quantify the levels of redundancy, uniqueness and synergy, which are expressed in information theoretic terms, through the use of mutual information.
Depending on dataset size and di... | Rebuttal 1:
Rebuttal: Thank you for your valuable feedback and insightful comments! We respond to some concerns below:
[multimodal VAEs] While multimodal VAEs are related, their goal is not to define and estimate multimodal interactions, but rather using intuition of interactions to guide model design (e.g., see figur... | Summary: This paper proposes an information-theoretic approach to quantify the degree of three PID(Partial Information Decomposition) statistics of a multimodal distribution: redundancy, uniqueness and synergy. To solve the difficulties in computing PID over high-dimensional continuous distributions, two scalable estim... | Rebuttal 1:
Rebuttal: Thank you for your valuable feedback and insightful comments! We respond to some concerns below:
[table 2] **The absolute value (i.e., actual bits of information) will never be fully 1 bit, because one can never extract the exact amount of task-relevant information**, which would imply 100% train... | Summary: This paper addresses a fundamental question in multimodal modeling: given multiple input modalities, how can we quantify the ways they must interact (as part of a model) to solve a particular task. Furthermore, can we use knowledge to learn how to design models that better leverage these interactions.
The au... | Rebuttal 1:
Rebuttal: Thank you for your valuable feedback and insightful comments! We respond to some concerns below:
[scale] **The absolute value (i.e., actual bits of information) will never be fully 1 bit, because one can never extract the exact amount of task-relevant information**, which would imply 100% train a... | Summary: This paper attempts to quantify the interactions between different input modalities using Principle Information Decomposition (PID) from Information Theory. Using this approach, one can quantify not only interactions within multimodal datasets but as well as interactions captured by multimodal models.
The pap... | Rebuttal 1:
Rebuttal: Thank you for your valuable feedback and insightful comments! We respond to some concerns below:
[visuals] Thank you for pointing this out - we do plan to add more visuals and have included a rebuttal pdf with some other figures we have included to the paper:
1. Figure 1: A venn diagram to give a... | Rebuttal 1:
Rebuttal: Dear all reviewers, we are extremely grateful for your valuable feedback and insightful comments. We are glad that you agree that our framework can be of fundamental importance for studying multimodal interactions, and that our proposed estimators are technically novel and comprehensively evaluate... | NeurIPS_2023_submissions_huggingface | 2,023 | Summary: This paper proposes an information-theoretic approach to quantify the degree of redundancy, uniqueness, and synergy relating input modalities with an output task. The authors introduce two new estimators for these PID statistics that scale to high-dimensional distributions. They validate the PID statistics on ... | Rebuttal 1:
Rebuttal: Thank you for your valuable feedback and insightful comments! We respond to some concerns below:
[other methods] Our framework allows for easy generalization to other definitions as well: **we implemented 3 alternative definitions for interactions, including the original called I-min, WMS, and CI... | null | null | null | null | null | null |
A graphon-signal analysis of graph neural networks | Accept (poster) | Summary: The authors introduce an extension of the graphon analysis to a graphon-signal analysis, in order to measure generalization bounds of and stability of MPNNs. Specifically, they propose a variant of cut distance as a new graph-signal similarity measure that transforms the space of graph-signals to a dense subse... | Rebuttal 1:
Rebuttal: We thank the reviewer for their constructive and supportive comments.
>
***Experiments:*** We ran new experiments that we will include in the camera ready version if accepted. The only non trivial assumption in our model is that it uses normalized sum aggregation. To verify that such a model gen... | Summary: This work focuses on analyzing message passing neural networks (MPNNs) by extending graphon analysis. It casts MPNNs as a function taking a graph-signal input (graph + signal on the graph) and proposing a similarity measure over graph signals, viewed as a dense subset of a compact metric space. This allows the... | Rebuttal 1:
Rebuttal:
We thank the reviewer for their important and insightful comments.
>
***Slow asymptotics and implications:*** The motivation may have been not clear enough in the first submission. We will extend the discussion in the introduction and generalization sections and stress the following point. All ... | Summary: This work proposes a new metric to measure the distance between graphon signals (graphon parameters and signals defined over graphon jointly). The metric generalizes the previous metric for graphon only. The work also uses the metric to characterize the approximation error of graphon signals by performing equi... | Rebuttal 1:
Rebuttal: We thank the reviewer for their constructive and in-depth comments.
>
***New insights from the theory and comparison to past results:*** Thank you for this comment. We see your point that the motivation is not explained clearly enough in the first submission. There is actually a clear reason to ... | Summary: This paper extends the graphon theory to the graphon-signal theory. Especially, the authors introduce a distance metric "cut distance" on graphon-signal and analyze their regularity, and compactness. Furthermore, the authors use Lipschitez properties of MPNN to bound the model generalization.
Strengths: I am ... | Rebuttal 1:
Rebuttal: We thank the reviewer for their constructive comments.
>
***Notations:*** We will add a notation table in the supplementary material of the camera ready version if accepted.
>
***References and discussion:*** We will add a discussion and references on benign overfitting and double descent to ... | Rebuttal 1:
Rebuttal: We thank the reviewers for their constructive and in-depth comments. Following the comments, we will edit the paper to clarify the motivation for our theory as explained below, and will add the experiments and figures described below and in the additional rebuttal PDF file.
>
***New insights fro... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
SUBP: Soft Uniform Block Pruning for 1$\times$N Sparse CNNs Multithreading Acceleration | Accept (poster) | Summary: This paper proposes a new pruning approach to train a uniform 1×N sparse structured network from scratch, dubbed as SUBP. SUBP pruned blocks according to block angular redundancy. The blocks are pruned in a uniform manner to make full use of multithread parallelism computation. During the training process, the... | Rebuttal 1:
Rebuttal: Thanks for your constructive and supportive comments.
**Q1:** The authors are encouraged to conduct experiments on other tasks.
**A1:** Following your advice, we further conduct experiments on two other tasks including object detection and instance segmentation. Below displays the experimental r... | Summary: SUBP trains a uniform 1×N sparse CNNs from scratch, with three characteristics: (1) a periodic block pruning and regrowing technique via importance sampling, (2) a pruning criterion based on angular redundancy across blocks, and (3) a uniform 1×N sparse pattern for multithreading acceleration. The authors als... | Rebuttal 1:
Rebuttal: Thanks for your in-depth review that will help us strengthen the manuscript. We hope our response can address your concern here.
**Q1:** Prune-and-regrow idea has been well-studied and acceleration of TVM is highly relies on TVM.
**A1:** With all due respect, though the concept of grow-and-prune... | Summary: This paper proposes a new fine-grained structured pruning method, which can train from scratch to reduce the training cost and prunes the weights in a uniform manner to handle the load imbalance problem. During pruning, the method uses block angular redundancy as the criterion to discard the redundant blocks a... | Rebuttal 1:
Rebuttal: Many thanks for your responsible reviews that will help us improve the manuscript. Please see the following answers to your questions.
**Q1:** Author should use the same hyperparameters for fair comparisons, as well as dense model baselines for other methods to better demonstrate the effectivenes... | Summary: Structural weight pruning methods have gain popularity due to the hardware friendly pattern. This work improves 1xN pruning with prune-and-grow and uniform block pruning. Evaluation shows outperforming accuracy results at similar levels of FLOPs than prior work.
Strengths: - This paper presents a new contribu... | Rebuttal 1:
Rebuttal: Thanks for your kind comments. We sincerely wish our response can well address your concerns so as to raise the rating score.
**Q1:** The author explain how it differs from previous block pruning. The author should add an algorithm to describe of the steps better.
**A1:** More specifically. Acc... | null | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Smoothed Online Learning for Prediction in Piecewise Affine Systems | Accept (spotlight) | Summary: The authors consider a smoothed nonlinear regression problem with time-series data in which the hypothesis class is decomposed into the product of (i) affine classes across polyhedral components and (ii) an indicator/classifier "choosing" the component. They first show that parameter recovery for the affine c... | Rebuttal 1:
Rebuttal: # Regret Bound
Thank you for your review. Regarding the regret bound, we agree that the resulting oracle complexity is likely suboptimal. Note that if one is happy with computationally inefficient algorithms, then running exponential weights on an exponentially large cover will attain optimal r... | Summary: This paper develops the theory to show that the online smoothed learning of piecewise affine model (PWA), under mild assumption, can achieve sublinear regret in dynamics prediction, and also the regret bound is polynomial in system parameters.
Strengths:
I think the results of this paper are of great impor... | Rebuttal 1:
Rebuttal: Thank you for your review. We will address each question below:
1. Directional smoothness means that for any direction, the random variable projected onto this direction is anti-concentrated in any interval. Because the boundaries are determined by affine sets, for the system to be near to the b... | Summary: This paper presents an algorithm for learning policies for piecewise affine systems, with both parameter-accuracy and regret bound guarantees, by using a kind of randomized smoothing of the boundaries of the affine regions. It identifies an anticoncentration property ("directional smoothness") that was previou... | Rebuttal 1:
Rebuttal: Thank you for your review. Regarding the regret bound, we agree that the resulting oracle complexity is likely suboptimal. Note that if one is happy with computationally inefficient algorithms, then running exponential weights on an exponentially large cover will attain optimal regret scaling li... | Summary: This paper studies the problem of prediction in piecewise affine systems over a finite horizon. The main results are a linear regret bound and a sublinear regret bound under the proposed metric of simulation regret. To obtain the regret bounds, the authors made many assumptions and assume the access to an empi... | Rebuttal 1:
Rebuttal: Thank you for your feedback. We will be sure to provide more concrete examples and motivation for our setting in our revision. While we deferred several explanations to the appendix for the sake of space, we will better clarify the intuition for the mathematical expressions.
We agree that there ... | null | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Swap Agnostic Learning, or Characterizing Omniprediction via Multicalibration | Accept (poster) | Summary: This paper introduces and studies 3 learning theoretic problems, namely "swap" versions of agnostic learning, omniprediction, and multicalibration. These swap versions are inspired by the notion of "swap regret" in online learning. They show that in some settings, all three of these learning-theoretic problems... | Rebuttal 1:
Rebuttal: We thank the reviewer for the questions. We hope the main response above helps to answer the motivation for our study. These swap variants are natural in a few ways. Perhaps the way to motivate these notions as natural is through multicalibration. Multicalibration was introduced as a notion of... | Summary: The authors propose a stronger variant of agnostic learning called *swap* agnostic learning, and show that this notion of learning is equivalent to “swap” variants of multicalibration and omniprediction. Multicalibration means that the predictor is simultaneously calibrated on a given collection of subpopulati... | Rebuttal 1:
Rebuttal: To reiterate from the main response, multicalibration and swap multicalibration are essentially the same, so there is little additional motivation. For swap omniprediction (and similarly agnostic learning), the guarantee seems to be much stronger. In some sense, one can view this as pushing the ... | Summary: The authors study swap agnostic learning, or a game between a predictor and an adversary where the predictor has to come up with a family of hypotheses to compete with a strong adversary's loss. They prove a surprising feasibility result for swap agnostic learning for any convex loss by proving that swap agnos... | Rebuttal 1:
Rebuttal: We appreciate you bringing up these points on presentation. If there is any jargon that you feel is particularly inaccessible to the general machine learning audience, we are happy to incorporate definitions and clarifications. We will be cognizant of aiming for accessibility in our revisions fo... | Summary: The paper introduces the notion of swap agnostic learning, and proves the feasibility of swap agnostic learning for any convex loss using only weak agnostic learners. To prove this result, the authors establish the equivalence between swap agnostic learning and the swap variants of multi-calibration and omnipr... | Rebuttal 1:
Rebuttal: Swap and standard multicalibration are equivalent in the max version you describe. Unfortunately, the max version is hard to achieve from sample access alone. This is why all known algorithms work with (some variant of) the expectation version, which allows us to ignore unlikely predictions. (Ev... | Rebuttal 1:
Rebuttal: We thank all of the reviewers for their thoughtful, and overall, positive feedback on the submission on Swap Agnostic Learning, Multicalibration, and Omniprediction. We hope to clarify our views on the distinction between the Swap and Standard learning tasks—a point that came up across many revie... | NeurIPS_2023_submissions_huggingface | 2,023 | Summary: This paper studies algorithmic fairness from a learning theoretic lens, specifically extensions of the recent notion omniprediction by Gopalan et al and its connections to to agnostic learning - multicalibration is sufficient to obtain omniprediction and guarantee agnostic learning for all convex loss function... | Rebuttal 1:
Rebuttal: To be very clear, the equivalences we show for the swap variants are NOT known to be true in the standard models: the implications are only known to go in one direction. In fact, our work suggests that they do not generally hold. This is a key contribution of our work. On a technical note, (Go... | Summary: This paper provides a new learning framework referred to as Swap Agnostic Learning, a variant of agnostic learning aiming to a stronger goal.
It is shown that this framework is equivalent ot swap variants of the notions of Omniprediction and Multicalibration.
An algorithm for Swap Agnostic Learning is provided... | Rebuttal 1:
Rebuttal: We apologize for some of the confusions. Multicalibration is a fairly young notion, but has had significant developments and contributions to the literature in the past few years. As such, the presentation and nomenclature has become a bit ambiguous. Let us clarify.
MCBoost—i.e., Mutlicalibrat... | null | null | null | null |
Conditional Adapters: Parameter-efficient Transfer Learning with Fast Inference | Accept (poster) | Summary: The paper proposes Conditional Adapter (CODA), which not only enables parameter-efficient fine-tuning but is also beneficial for the inference speedup. The main idea of CODA is to pass all tokens to a lightweight routing adapter path while selectively passing only certain tokens to the original path. The selec... | Rebuttal 1:
Rebuttal: Thank you for your review and suggestions!
**Differences between CoLT5 and CODA:**
We will add a better explanation in the next version of our paper. While both CODA and CoLT5 use conditional activation (token selection) to enhance inference efficiency, the focus of the two works are ve... | Summary: This paper proposes a unique modification over parameter-efficient transfer learning by employing a conditional adapter parallel to the pre-trained model. In doing so, a dynamic token selection mechanism is proposed for each block. The selected tokens undergo heavy pre-trained operations, while all the tokens ... | Rebuttal 1:
Rebuttal: Thank you for the review and suggestions!
**Presentation:**
We agree that the scale and location of the tables and plots need additional work. Due to the page limit, we have to shrink the size of tables and plots and move some results to the supplementary materials. We will improve the... | Summary: This paper proposed Conditional Adapter (CODA), which can achieve good trade-offs between speed and accuracy. CODA achieves 2x to 8x inference speed-up compared to the state-of-the-art Adapter approach while maintaining moderate to no accuracy loss and the same parameter efficiency across language, vision, and... | Rebuttal 1:
Rebuttal: Thank you for your review and comments! Please see our response to the weakness points.
**Contributions:**
One of our major contributions is the insight that the combination of adapters and token pruning is a natural fit. While token pruning is effective for faster inference, it is a d... | Summary: The paper introduces a novel technique called Conditional Adapters (CODA), which aims to enhance transfer learning and improve inference efficiency. CODA surpasses the state-of-the-art Adapter approach by achieving a remarkable 2x to 8x acceleration in inference speed while maintaining moderate accuracy loss a... | Rebuttal 1:
Rebuttal: Thank you for your review and comments!
**Open sourcing and reproducibility:**
We will request to make our code available. Open sourcing the code has to go through our employer’s approval process. In addition, we provide detailed descriptions of our implementation and experimental setup... | Rebuttal 1:
Rebuttal: We thank all reviewers for their insightful comments and feedback.
Most reviewers agree that there is an thorough evaluation and analysis in the paper (R1, R2, R3, R4), the tackled problem and proposed technique are important and have a broad practical value (R3, R4), and the paper is very-well p... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Block Coordinate Plug-and-Play Methods for Blind Inverse Problems | Accept (poster) | Summary: The paper proposes a block-coordinate optimization strategy for plug-and-play (PnP) methods to solve blind inverse problems in imaging. Empirical validation is performed with blind deblurring and parallel CS-MRI. Theoretical study on the convergence of the proposed BC-PnP was extended from the theoretical resu... | Rebuttal 1:
Rebuttal: Thank you for your valuable feedback and positive comments on our work. Please see below for our point-by-point responses to your comments. The PDF with additional simulations is available within the Author Rebuttal panel above.
**Novelty in the theory:**
We respectfully disagree with the reviewe... | Summary: This paper presents a new plug-and-play algorithm for solving blind inverse problems, i.e., where the forward sensing operator is not fully known. The method leverages pretrained denoisers for both the signals and the unknown parameters of the sensing model. A theoretical analysis of the proposed algorithm is ... | Rebuttal 1:
Rebuttal: Thank you for your valuable feedback and positive comments on our work. Please see below for our point-by-point responses to your comments. The PDF with additional simulations is available within the Author Rebuttal panel above.
**Commentary on deep unrolling:**
We fully agree with the reviewer t... | Summary: The paper generalizes the plug-and-play framework from single-block to a multi-block case by embedding unknown inverse operators. It provides convergence analysis for the proposed method and conducted various experiments on two imaging applications, i.e., parallel MRI and blind image deblurring. Numerical resu... | Rebuttal 1:
Rebuttal: Thank you for your valuable feedback and positive comments on our work. Please see below for our point-by-point responses to your comments. The PDF with additional simulations is available within the Author Rebuttal panel above.
**Computational cost:** Prompted by your comment, we computed runnin... | Summary: Develops a plug-and-play technique for blind (unknown forward model) imaging inverse problems. Key idea is to use two denoising algorithms: One as a prior on the image and one as a prior model on the measurement operator (e.g., blur kernel). Also analyzes the convergence of the algorithm. Successfully applies ... | Rebuttal 1:
Rebuttal: Thank you for your valuable feedback and positive comments on our work. Please see below for our point-by-point responses to your comments. The PDF with additional simulations is available within the Author Rebuttal panel above.
**Comparison with [60]:**
Prompted by your comment, we ran simulati... | Rebuttal 1:
Rebuttal: Thank you all for providing us with valuable feedback. We provide detailed answers to all the comments below. To better address some of them, we ran additional simulations, albeit within the time constraints of the limited rebuttal window. The PDF file attached to this response reports the results... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Assumption violations in causal discovery and the robustness of score matching | Accept (poster) | Summary: The authors focus on evaluating causal structure learning algorithms. These algorithms, in order to claim that their outputs are causal, require various assumptions, most of which can't be tested easily in observational data. This work aims to test how robust a set of causal structure learning algorithms are... | Rebuttal 1:
Rebuttal:
We thank the reviewer for considering our work and for their valuable comments and insights.
- > The diversity of method is lacking. In Table 1, FCI and DirectLiNGAM are included, but neither are present in the actual evaluation.
In the case of FCI, when confounders are not present in t... | Summary: The authors present the results of an extensive empirical study evaluating the performance of algorithms for learning the structure of causal graphical models from observational data. In particular, the study examines the robustness of these algorithms to violations of various untestable assumptions that are t... | Rebuttal 1:
Rebuttal: We thank the reviewer for their effort spent in providing valuable feedback on our work. In the response, we will follow the paragraph listing structure of the review.
- > The authors use false-positive and false-negative rates (and F1) as their measures of error. This ignores a great deal of r... | Summary: The paper considers the important problem of what happens to the reliability of the causal model output of different types of causal discovery algorithms when the underlying assumptions behind the method are violated. It finds that score matching-based methods tend to be more robust and provides a theoretical ... | Rebuttal 1:
Rebuttal: We thank the reviewer for the valuable insights provided with their comment. In the response, we will follow the structure of the bullet list in the review.
- The reviewer comments that our "comparison is too limited and biased to be meaningful”, supporting their claim with a series of points.
... | null | null | Rebuttal 1:
Rebuttal: We thank the reviewers for their effort in reading our paper also providing valuable insights. In general, there is agreement that this is _very important_ work (R QK5D , eJF4, JGO2), both in terms of the subject of the paper and in the scale of the experiments (_highly commendable attempt at prov... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
A Unified Approach to Count-Based Weakly Supervised Learning | Accept (poster) | Summary: The paper describes a unified loss for different weakly supervised paradigms: learning from label proportions (LLP), multiple instance learning (MIL) and learning from positive and unlabeled data (PU).
The main idea is to compute the probability of a count sum constraint by summing over all possible combinatio... | Rebuttal 1:
Rebuttal: We thank reviewer 5jbX for such detailed insights. We are pleased to hear that you found the paper easy to follow and the unification as a strong novelty.
**[Multi-class Extension Possibility]**
The multi-class extension is certainly possible though admittedly, it comes with an exponential time ... | Summary: This paper proposes a unified framework for weakly-supervised binary classificasion, where instances can be considered as bags and their weak supervision can be seen as count of positive instances in each bag.
The framework is based on the probability of the count and the authors derived loss functions based o... | Rebuttal 1:
Rebuttal: We thank reviewer YHmN for the review of our paper. We are delighted to hear that you found our solution well-written, easy to follow, and the empirical evaluation effective.
**[Originality]**
“It is better to discuss concrete descriptions of how the existing work trains models.”
We will make s... | Summary: The paper proposes a unified approach for three problems of weak supervision: LLP, MIL, and PU learning. The approach is based on estimation of a common computational block – count probability. For its estimation, the dynamic programming method is used. The method is compared in three problem scenarios with di... | Rebuttal 1:
Rebuttal: We thank reviewer rn5b for the comments. We are happy that you found the paper well-written and easy to follow.
**[Training Algorithm]**
“My main question is how the training algorithm actually works, and can the authors provide a pseudocode for that?”
Let us consider the example of LLP. As not... | Summary: The author proposes a differentiable count loss for three weak supervision settings: LLP, MIL, PU learning. The paper provides a clear explanation of the proposed loss, demonstrates strong empirical performance, and highlights the efficiency of its implementation using dynamic programming.
Strengths: 1. The ... | Rebuttal 1:
Rebuttal: We would like to thank reviewer Q9rM for the feedback. We are glad you found approach's efficiency, explanation, and applicability to be solid.
**[Articulating Contribution]**
“The paper could benefit from a clearer articulation of its contribution. Providing a more explicit comparison and empha... | Rebuttal 1:
Rebuttal: Thank you all for your reviews and insights!
**[Novelty]**
While we do acknowledge that [3] showcases the “exactly k” dynamic programming algorithm we employ, our novelty lies in the *unified view* of weakly supervised learning for *count-based* settings. We are the first to leverage the dynamic... | NeurIPS_2023_submissions_huggingface | 2,023 | Summary: This paper proposes a unification of three different kinds of weak supervision, viz., Learning from label proportions (LLP), Multi-instance learning (MIL), and learning from positive and unlabeled data (PU learning). They all involve supervision at the bag level instead of the instance level (except for PU lea... | Rebuttal 1:
Rebuttal: We would like to thank reviewer jVGG for the feedback. We are glad they found the paper well-written and enjoyed our approach.
**[Related Work - Partial Label Learning]**
Thank you for the comment, we have included this as part of our general rebuttal section.
**[Limited Novelty]**
We unify $3... | null | null | null | null | null | null |
Perceptual adjustment queries and an inverted measurement paradigm for low-rank metric learning | Accept (poster) | Summary: This paper introduces perceptual adjustment query (PAQ), which combines cardinal reporting and ordinal reasoning using a sliding bar for relational questions through identifying the transition point (similar vs. dissimilar) along a continuum to the reference item. Formulating the problem as Mahalanobis distanc... | Rebuttal 1:
Rebuttal: Thank you for your comments and feedback. We are grateful that you appreciate the advantages of our proposed query and the writing of the paper.
> What are the main contributions of this work?
Please see the global response for a detailed response. In summary, our paper has three main contributi... | Summary: People evaluate things in two ways: scoring and comparing, which is the process of generating cardinal data and ordinal data. However, since different people have different scales, the cardinal data are more influenced by people's subjective perception, while the ordinal data are fundamentally limited and lace... | Rebuttal 1:
Rebuttal: Thank you for your comments and feedback. We are grateful that you appreciate the novelty of our proposed query.
>While novel ideas and theories are important, it is sufficient experimentation that will prove the value of the proposed method. It is recommended that sufficient experiments be added... | Summary: this paper introduces the Perceptual Adjustment Query (PAQ) mechanism for collecting human feedback in a cognitively lightweight manner, which combines advantages from both ordinal and cardinal queries.
The authors apply PAQs to the problem of metric learning under a Mahalanobis metric.
Strengths: - The ... | Rebuttal 1:
Rebuttal: Thank you for your thoughtful comments and feedback. We are grateful that you appreciate the advantages of PAQs and our sample complexity guarantees.
> Limited Evaluation: While the paper mentions ”extensive numerical simulations” to demonstrate the performance and properties of the estimator, i... | Summary: This paper proposes to collect users' perceptual adjustment queries (PAQ) to capture change points in a semantic space. In contrast to ordinal and cardinal data, PAQ better aligns with human perception and is cheap to collect. PAQ can be used to learn the semantic space manifold by solving a row-rank optimizat... | Rebuttal 1:
Rebuttal: Thank you for your comments and questions. We are delighted that you find our proposed query conceptually novel, well-motivated, and easy to implement.
> The paper does not provide a downstream use case for PAQ.
We agree with the reviewer that a discussion of downstream use cases for PAQs would... | Rebuttal 1:
Rebuttal: We thank the reviewers for their time and thoughtful feedback. The reviews especially recognized the novelty of the parametric adjustment query (reviewers Ehde, sAC6, eVNV, H5Pr) in combining “the advantages of both cardinal and ordinal queries” to alleviate cognitive burden on respondents while m... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Conformal Meta-learners for Predictive Inference of Individual Treatment Effects | Accept (oral) | Summary: This paper develops a meta-learner-based approach for conformal inference of individual treatment effects (ITEs). Instead of predicting the missing outcome (the counterfactual), the proposed approach uses plug-in pseudo outcomes as the inference proxy. The validity of ITE inference is based on careful analysis... | Rebuttal 1:
Rebuttal: We would like to thank the reviewer for the thoughtful feedback and comments. Regarding condition (ii) in Theorem 1, the ordering is written correctly for the SOSD condition. Because these notations come from the decision theory literature, they are written to express risks associated with differe... | Summary: The paper introduces a framework for inferring treatment effects, merging concepts of conformal prediction and meta-learner. This framework is characterized by its distribution-free validity and coverage guarantees. The authors utilize stochastic ordering techniques to substantiate the framework's validity.
S... | Rebuttal 1:
Rebuttal: We would like to thank the reviewer for the thoughtful feedback and comments. In the final manuscript, we will fix the overlapping elements in Fig. 4(c).
Please see below our point-by-point responses to your comments.
- **Uncheckable assumptions involved in the stochastic ordering framework**
... | Summary: This paper adresses individual treatment effect (ITE) inference using conditional average treatment effect (CATE) meta learners.
The authors propose to wrap the meta-learners with a conformal prediction step so as to obtain valid confidence intervals for the estimated treatment effect. Since the CATE estimat... | Rebuttal 1:
Rebuttal: Thank you for your thoughtful comments and the feedback on our paper. Please see below some points that we would like to clarify in response to your comments.
- **Knowledge of the propensity scores**
Please note that this is not a unique limitation to our method. To date, all valid inference p... | Summary: The paper deals with confidence intervals for the ITE prediction task, where the authors propose to use the framework of conformal prediction for meta learners. The learning of nuisance models for generating pseudo outcomes is done one separate data split, and the final regression models trained to predict pse... | Rebuttal 1:
Rebuttal: We would like to thank the reviewer for the thoughtful and detailed feedback. Please see below our point-by-point responses to your comments.
- **Conformal meta-learners and inductive biases**
In the points below, we clarify our argument on how the proposed framework enables a more flexible set... | null | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
It begins with a boundary: A geometric view on probabilistically robust learning | Reject | Summary: This paper claims to identify a flaw in the formulation of probabilistically robust learning. To correct this supposed flaw, the authors propose a regularization scheme based on a notion of probabilistic perimeter. This leads to a set of theoretical results that generalize the losses used in probabilistic ro... | Rebuttal 1:
Rebuttal: We thank the reviewer for their detailed comments! In our response we follow the structure of the review and give point to point replies.
----
**Overall evaluation**
- *The mathematical insights $\ldots$ don't lead to $\ldots $ notable improvements on the algorithmic end.*
We note that the goal ... | Summary: Congratulations to authors for a timely and interesting paper. The paper advances the work of Robey et al. [2022] that established computational less expensive version of robust learning, dubbed Probabilistic Robust Learning (PRT). In particular, random perturbations and gradient based attacks had been used be... | Rebuttal 1:
Rebuttal: Thank you very much for your kind words and for seeing merit in our paper! Below we provide point to point replies to your questions and remarks.
- _Intriguing results are mentioned but not further explored or commented: "Note that ... or other worst-case optimization techniques."_
We will ma... | Summary: The probabilistically robust learning (PRL) framework aims to trade-off clean accuracy and adversarial robustness in model training by introducing a loss that penalizes the suceptibility towards adversarially attacks in the perimeter of the data points. The authors of the paper under review note that the loss ... | Rebuttal 1:
Rebuttal: Thank you very much for your valuable comments on our paper. Below we provide point to point replies to your questions and remarks.
- _Regarding the soundness score._
We realize the mathematics is a tad sophisticated. However, we stand by the correctness of our results and the validity of the... | Summary: This paper takes a fresh and geometric view on the method—Probabilistically Robust Learning (PRL) [Robey et al., 2022], an approach which interpolates between the robustness offered by adversarial training and the higher clean accuracy and faster training times of Empirical Risk Minimization. The paper propose... | Rebuttal 1:
Rebuttal: Thank you very much for your kind words and for seeing merit in our paper! Below we provide point to point replies to your questions.
- _Do the theoretical results hold for other attacks e.g. Carlini & Wagner, or for larger datasets?_
Yes, as our results are based on mathematical proofs and ... | Rebuttal 1:
Rebuttal: We would first like to thank all reviewers for their valuable comments on our paper.
We would like to clarify the reasons why we believe our slight modification to the probabilistically robust learning (PRL) framework is meaningful. Therefore, before moving on to the individual responses to each... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Accelerating Exploration with Unlabeled Prior Data | Accept (poster) | Summary: This work presents an off policy RL algorithm that use prior unlabeled data in a novel manner. The mechanism use an "optimistic" reward labeled based on a UCB estimate, that it is practically an iterative RND full update of the unlabeled dataset based on a stepwise exploration update of the RND functions. Thi... | Rebuttal 1:
Rebuttal: Thank you for your kind words and thoughtful feedback about our paper. We addressed your main concern regarding the reward resetting by providing more insights on the nature of the underlying issue (that is likely orthogonal to our main focus of the paper) as well as providing additional empirical... | Summary: The paper proposes a simple method to leverage reward-free data in order to accelerate exploration in sparse reward tasks. This is done by learning a reward model and an uncertainty model during online learning, labeling the unlabeled data with an upper confidence bound based on these models, and sampling from... | Rebuttal 1:
Rebuttal: Thank you for your kind words and thoughtful feedback about our paper. We addressed your main concern regarding the coverage assumption of the offline data by conducting additional ablation experiments with different offline data distributions (e.g., poor data coverage, orthogonal trajectories) an... | Summary: The paper propose a method exploiting unlabeled prior data (i.e. trajectories without the reward signal) for more efficient online learning. To this end, a reward model is trained (based on the online data) and the so-labeled offline data are utilised jointly with the offline data.
Strengths: The paper tack... | Rebuttal 1:
Rebuttal: Thank you for your kind words and thoughtful feedback about our paper. We addressed one of your main concerns on “how the effect of offline data distribution affects training” by conducting additional ablation experiments with different offline data distributions (e.g., poor data coverage, orthogo... | Summary: This paper proposes a method to leverage unlabelled offline data to expedite online exploration. The idea is to train a reward model and random network distillation models on online data and use them to construct a UCB estimate for offline data. By training an off-policy RL algorithm on both online and relabel... | Rebuttal 1:
Rebuttal: Thank you for your kind words and thoughtful feedback about our paper. We addressed your first concern on offline data quality by conducting additional ablation experiments with different offline data distributions (e.g., poor data coverage, orthogonal trajectories) and showed that our method is a... | Rebuttal 1:
Rebuttal: We appreciate the feedback from all reviewers. We added three additional sets of experimental results in the one-page PDF attached to address the three shared main concerns from the reviewers --
**1. How does offline data quality affect online performance? -- Figure 1**
We ran additional ablatio... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Hidden Poison: Machine Unlearning Enables Camouflaged Poisoning Attacks | Accept (poster) | Summary: The paper proposes a novel attack in machine unlearning. There are two step in the attack process. In the first step, the poisoned data and the camouflaged data are fed into the initial model training, together with the benign data. The machine unlearning process would take out the camouflaged data and the ret... | Rebuttal 1:
Rebuttal: Unfortunately, we were not fully able to parse your concerns. Please find the following response to the best of our interpretation. We would love to engage further with the reviewer if any of their concerns are not answered:
> Retraining from scratch vs approximate unlearning
In machine unlearn... | Summary: This paper studies poisoning attacks against machine unlearning systems, where the attacker injects a set of poisoning samples and a set of camouflage samples. When both sets of samples are present in the training set, the trained model performs similarly to the clean model but shows adversarial behaviors and ... | Rebuttal 1:
Rebuttal: We are heartened to hear that the reviewer finds our paper to be an interesting attack, which is effective, and that we performed a thorough investigation which shows its efficacy across a variety of settings. We believe such a contribution to be above the bar for NeurIPS. We hope to assuage any c... | Summary: This paper proposes a new attack for security risks in the field of machine unlearning, namely the camouflaged data poisoning attack. The attacker adds poisoned and camouflaged points to the training data, and then triggers a deletion request to remove the camouflaged points, causing the model to misclassify t... | Rebuttal 1:
Rebuttal: We appreciate that the reviewer likes the direction, and for their thoughtful questions to keep us honest! We have investigated all these questions, and found that the attack is fairly robust to the modifications suggested. Details follow.
> Extension of the attack to the black-box setting.
Th... | Summary: This paper introduces a very interesting novel attack vector for tampering with the behavior of a machine learning model by influencing its training dataset. It poses that the practice of "unlearning" examples that are requested to be taken down can be exploited. The attack proceeds in several steps: 1. The ad... | Rebuttal 1:
Rebuttal: Thank you for your positive support for the paper! We agree that these other settings are also interesting. However, rather than cover many disparate settings poorly, we chose to investigate computer vision datasets in extreme depth, evidenced by our paper being already quite long (27 pages), with... | null | NeurIPS_2023_submissions_huggingface | 2,023 | Summary: This paper introduces a clean-label targeted data poisoning attack for machine unlearning pipelines. Gradient matching [A] is used to generate the poisons, and experiments use the retraining from scratch method for machine unlearning.
The core idea is for an attacker to introduce poisons and camouflages to th... | Rebuttal 1:
Rebuttal: We thank the reviewer for their kind comments. They give strong support for the paper, pointing out novelty and interest of the attack, thoroughness of the investigation (covering several datasets and models) and the fact that we highlight new technical questions for the data poisoning community. ... | null | null | null | null | null | null |
Model-free Posterior Sampling via Learning Rate Randomization | Accept (poster) | Summary: The paper introduces Randomized Q-learning (RandQL), a novel model-free algorithm for regret minimization in episodic MDPs. Leveraging a new idea of learning rate randomization, two respect algorithms are proposed for tabular and non-tabular metric space settings, achieving near-optimal regret upper bound. Pro... | Rebuttal 1:
Rebuttal: We are very thankful to Reviewer XC7g for detailed reading and informative feedback. Additionally, we are very thankful for pointing out the misprints. In the sequel, we address the particular questions raised in the review.
- “Intuition of the value of J”.
This value of J is used in the proof ... | Summary: This work aims to develop computationally efficient, posterior sampling-style, model-free RL algorithms. They propose a family of algorithms based on the idea of learning rate randomization, which they show are able to provably learn and mimic to some degree the updates of posterior sampling. In the tabular se... | Rebuttal 1:
Rebuttal: We are very thankful to Reviewer 4Am8 for detailed reading and informative feedback. In the sequel we answer the questions raised in the review.
- "This work is motivated by noting that “a provably tractable model-free posterior sampling algorithm has remained a challenge”. However, I’m not entir... | Summary: The paper introduces a novel algorithm for regret minimization in episodic Markov Decision Processes. The proposed algorithm, RandQL, achieves optimistic exploration without using bonuses, relying on learning rate randomization. The authors provide theoretical analysis of the algorithm's regret bound and demon... | Rebuttal 1:
Rebuttal: We are very thankful to Reviewer xW74 for detailed reading and informative feedback. In the sequel we answer the particular questions raised in the review.
- “Lack of baseline comparison: using ensembles of Q values is not new, for example, bootstrapped DQN [1]. It would be an essential baseline ... | Summary: The paper titled Model-free Posterior Sampling via Learning Rate Randomization proposes a novel model-free algorithm called Randomized Q-learning (RandQL) for regret minimization in episodic Markov Decision Processes (MDPs). The algorithm introduces the concept of learning rate randomization to achieve optimis... | Rebuttal 1:
Rebuttal: We are very thankful to Reviewer axbs for detailed reading and informative feedback. In the sequel we answer the particular questions raised in the review.
- “Although the author claims the algorithm do not require bonus, the learning rate randomization still needs (s, a) counting. Counting is no... | Rebuttal 1:
Rebuttal: ## Model-based vs model-free
First, we would like to stress that in the tabular setting, according to the Definition 1 by Jin et al. 2018, RLSVI is a model-based algorithm since it needs a space complexity of order $O(S^2AH)$ independently of $T$, large enough to store an estimate of the MDP (emp... | NeurIPS_2023_submissions_huggingface | 2,023 | Summary: This paper introduces a new model-free Q-learning algorithm that utilizes a randomized learning rate. The effectiveness of this algorithm is theoretically validated by demonstrating a promising regret bound of $\widetilde{O}(\sqrt{H^{5} SA T})$ under the tabular setting. Furthermore, a theoretical bound is est... | Rebuttal 1:
Rebuttal: We are very thankful to Reviewer eHBt for detailed reading and informative feedback. We will add all missing references. In the sequel we answer the questions raised in the review.
- “What are the benefits of eliminating the exploration bonus? What are the promising parts compared to alternative... | Summary: The paper proposes a new algorithm, Randomized Q-learning (RandQL), for regret minimization in episodic Markov Decision Processes (MDPs). RandQL is the first manageable model-free posterior sampling-based algorithm. The authors study RandQL's performance in both tabular and non-tabular metric space settings. I... | Rebuttal 1:
Rebuttal: We are very thankful to Reviewer PYAU for detailed reading and informative feedback. In the following, we answer the concerns raised in the review.
- “In the experiments, the performance gap between model-free algorithms and model-based algorithms seems too large. It would be better to provide st... | null | null | null | null |
Why Did This Model Forecast This Future? Information-Theoretic Saliency for Counterfactual Explanations of Probabilistic Regression Models | Accept (poster) | Summary: This paper proposes a framework to provide counterfactual explanations for time-series forecasting models. This method attempts to detect the salient timestep which is critical to a future event. The method is built based on the framework of bottom-up preattentive saliency. Experiments are conducted on one syn... | Rebuttal 1:
Rebuttal: Thank you for the review and comments, f9yh!
### Clarifying Misinterpretation: Distinguishing from critical event detection and action localization
> (2) The current method functions more as a “critical event detector”...the tasks can be done by an action localization model.
[In L38-42 and App.... | Summary: The authors create a saliency method for time-series data, where they quantify the impact that a previous time-step has on the model's prediction of a future time-step. They deploy this method on synthetic and real-world video datasets, and show that this method is able to pick up on key points in the time-ser... | Rebuttal 1:
Rebuttal: Thanks for your review and comments, reviewer 82xZ! We are delighted that you found our formulation to be elegant (we especially worked hard on this aspect), and our experimental results to be strong.
### Answering Questions
1\.
> Is it possible to use such a formulation for sequential but non-t... | Summary: The paper proposes a computational XAI framework for counterfactual reasoning in probabilistic multivariate time-series forecasting. The key idea of the paper is to establish the conceptual foundation for linking saliency-based explanations with counterfactual reasoning. Specifically, the paper leverages a uni... | Rebuttal 1:
Rebuttal: Thank you for the review, N3J3. We are glad that you found our derivation of the conceptual link between saliency-based explanations and counterfactual reasoning interesting and well-explained.
### Addressing minor weakness comments
1\.
> … (and section 4.3 in particular). Improving the reada... | Summary: This paper proposes a method for counterfactual explanation using Information-Theoretic Saliency. The method trains a model using data collected up to the present moment and applies it to predict the outcomes caused by events occurring in the future time. This paper provides experiments on synthetic and real-w... | Rebuttal 1:
Rebuttal: ### Clarifying Central Misinterpretations
Thanks for the detailed review, L4MH! There seem to be some misinterpretations (perhaps expecting alignment with traditional causal inference rather than XAI literature):
1\.
> ..the model is designed as the causal path between nodes. Exogenous variables ... | Rebuttal 1:
Rebuttal: ## Updates: Additional Baselines, Real-World Examples (82xZ, VPQB, N3J3)
Following the suggestions, we have added the following (see attachment):
1\.
(82xZ). We added results from IntegratedGradients (IG) and SmoothGrad in Fig. R1. We also added GradientSHAP. The takeaway remains the same (the c... | NeurIPS_2023_submissions_huggingface | 2,023 | Summary: This paper presents a method for detecting salient points in time+feature space for models which make predictions from timeseries data. Salient points are defined as points that once observed produce a large change in the model’s distribution of predicted future outcomes (measured as differential entropy). The... | Rebuttal 1:
Rebuttal: Thank you for the review and your comments, VPQB! We are glad that you found the paper to be well-written and clearly explained.
1a\.
>I’m unclear why you can drop the prior (-log p_x(x)) between equations 3 and 4… is p_x uniform here... which is not always a good assumption for real-world image ... | null | null | null | null | null | null |
Debiasing Pretrained Generative Models by Uniformly Sampling Semantic Attributes | Accept (poster) | Summary: This paper studies how to learn a fair latent distribution so that a pretrained generator outputs samples that have equal likelihood to fall into different attributes. The paper assumes access to an imperfect attribute classifier. The proposed method learns the fair latent distribution by maximizing the entrop... | Rebuttal 1:
Rebuttal: Thank you for your feedback; and are happy to hear of your favorable view of our work.
**How far off are the learned weights of the mixture of class probabilities from optimal, and how close does the output of the Mapper match the target Minimally-Unfair Noise Distribution?**
We have added anot... | Summary: This paper introduces a method for mitigating biases in a pre-trained generator, without the need to retrain the generator or rely on real data. The authors propose the use of a classifier to create a fair distribution of noise vectors, which are then sampled to generate an equal number of instances for each g... | Rebuttal 1:
Rebuttal: Thank you for your valuable feedback.
**The paper lacks a discussion and comparison with baselines, such as [1] and [2], which address debiasing generative models without requiring retraining.**
As you have suggested, we have conducted new experiments to compare against the method proposed in [... | Summary: The paper proposes a method for training a Distribution Mapper that can generate a noise distribution capable of creating Fair Samples using an Imperfect Classifier. They conducted experiments on toy examples and several Face Generation models, demonstrating quantitatively superior performance compared to comp... | Rebuttal 1:
Rebuttal: Thank you for the time you put into reviewing our work, and for the constructive feedback you provided.
**It would be beneficial to report FID scores for CelebA as well.**
We computed the FID score for the Progressive GAN before and after applying our method, and found that the increase in FID ... | Summary: This paper proposes an approach to mitigate biases in pre-trained generative models. Specifically, the authors aim to generate a semantically fair distribution over the attributes. Authors propose to train a distribution mapper network that effectively learns to generate a fair noise distribution, which, in tu... | Rebuttal 1:
Rebuttal: We thank you for your thoughtful feedback.
**“How do you handle stochasticity that comes from the noise during training? Do you need the training process of the GAN to be deterministic?”**
We do not require the training of the generative model to be deterministic. We only require that, once tra... | Rebuttal 1:
Rebuttal: ## General Rebuttal
We thank the reviewers for the time they spent providing valuable feedback on our work. We have added additional experiments (see attached PDF) based on the reviewers’ suggestions.
In short, we:
- Performed a new experiment, where we apply our approach to a Latent Diffusio... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Emergent and Predictable Memorization in Large Language Models | Accept (poster) | Summary: This work presents an empirical study on predicting the memorization of examples in large language models. Specifically, using the Pythia model, they test two ideas for memorization prediction: (1) prediction via smaller models, and (2) prediction via partially trained models. Through extensive experiments on ... | Rebuttal 1:
Rebuttal: We thank the reviewer for their very thorough and positive comments on our work, and are excited that they find our work may be of immediate empirical usefulness!
> As the authors acknowledged in the limitations, the use of k-extractible and the memorization score simplifies the problem. In pract... | Summary: This paper discusses the issue of memorization in large language models, which can lead to the output of sensitive data. The authors propose to predict which sequences will be memorized by partially-trained checkpoints. They use the Pythia model in experiments and plot scaling laws for forecasting memorization... | Rebuttal 1:
Rebuttal: Thank you for your review, and we are glad you find our posed memorization behavior forecasting problem important.
> in practical scenarios, how can we know in advance which data should be memorized well (such as factual knowledge) and which data should not be remembered (such as private persona... | Summary: This paper introduces the difficulties of both forecasting a future model’s memorization of specific training data and forecasting the memorization behavior of fully-trained LLMs from partially-trained checkpoints. Scaling laws proposed for forecasting memorization, and recommendations for maximizing forecast re... | Rebuttal 1:
Rebuttal: Thank you for your feedback, and we are glad you feel the paper is well-written and effectively communicated the takeaways we intended.
> The essence of scaling laws is just the method that the author borrowed and tried instead of his own innovation.
We believe that our application of scaling l... | Summary: The paper focuses on predicting memorization of specific data point prior to training the model. The main approach of the paper relies on scaling law, where the memorization property of a large model is predicted using a smaller model, or a partially trained model.
The findings are (1) small models are not go... | Rebuttal 1:
Rebuttal: We thank the reviewer for their feedback on our work! We are grateful that you find the target of our work well-motivated and novel, and agree that it is an important target for future work to improve beyond our performance on the task we have defined.
> If the developer know which data is sensit... | Rebuttal 1:
Rebuttal: We thank all reviewers for their valuable feedback on our work. We have worked to revise the presentation of the paper where recommended, and hope to provide this updated version in the final camera-ready submission. We would like to clarify some details regarding the application of our prediction... | NeurIPS_2023_submissions_huggingface | 2,023 | Summary: This work empirically predict the memorization behaviors as measured by k-extractible and memorization scores. They tested on two settings 1) use small model to predict the memorization behavior or larger model, 2) use partially trained model to predict the fully trained model. The main finding is that 2) is h... | Rebuttal 1:
Rebuttal: Thank you for your review. We are grateful that you feel our problem setting is important and that the paper contains “clearly motivated and clean experiments with clear conclusions”.
> The heatmap is figure 3 is a wash, some variant of the Hinton diagram might show such correlation more clearly... | null | null | null | null | null | null |
Improving the Privacy and Practicality of Objective Perturbation for Differentially Private Linear Learners | Accept (poster) | Summary: This paper presents a new privacy analysis for the objective perturbation method for learning differentially private linear models. The new analysis makes use of gradient clipping (similar to DP-SGD) and is shown to have a competitive performance with respect to DP-SGD, while exhibiting better behavior for hyp... | Rebuttal 1:
Rebuttal: Thanks for your feedback! The notation went through a major overhaul during this paper’s lifespan and we obviously didn’t check for consistency as thoroughly as we should have. Our intention was to make this a fun read and we’re sorry to hear it missed the mark for you.
>The first item in the con... | Summary: The paper revisited a differentially private learning algorithm: objective perturbation. The authors gave an RDP analysis of objective perturbation and extended AMP to losses with unbounded gradients using clipped loss function.
Strengths: The main contributions of the paper are (1) RDP and HK analysis of obj... | Rebuttal 1:
Rebuttal:
Thank you for the review, we are happy to hear you enjoyed reading the paper!
> The 3rd bullet point in the contribution list emphasized that there is neglected subtlety in the analysis of objective perturbation which might lead to dimensional dependence of PLRV and the paper fixed the issue wi... | Summary: The paper revisits objective perturbation as a viable alternative to DP-SGD with tighter analysis and illustrating it competitiveness on a few problems.
Strengths: This paper explores the interesting direction of fairly comparing objective perturbation and DP-SGD specifically. The paper has multiple strengths... | Rebuttal 1:
Rebuttal: > DP-SGD is known to work best with large batch sizes, I would suggest increasing batch size for for the baselines -- even better is to have a plot with varying batch sizes. Similar argument can made for number of iterations as well.
The batch size and the number of iterations were simply reasona... | Summary: This paper derives new privacy analysis of the objective perturbation mechanism. Novel bounds are given for Rényi differential privacy and approximate differential privacy, that are tighter than existing bounds. A practical variant of objective perturbation (where the problem does not have to be solved exactly... | Rebuttal 1:
Rebuttal: > The proposed method is compared with DP-SGD for non-smooth functions [1], which is a different setting. In the smooth setting, DP-SGD and its variants converge faster, which makes the argument that the proposed Objective Perturbation method is more computationally efficient vacuous. (In general,... | null | NeurIPS_2023_submissions_huggingface | 2,023 | Summary: The paper considers a practical perspective on privacy for unconstrained convex generalized linear problems. While DP-SGD has been the workhorse for differential privacy in deep learning and prediction/classification problems, the authors suggest that objective perturbation might be more suitable for simpler ... | Rebuttal 1:
Rebuttal: > A gradient clipping parameter is used in the proposed algorithm - authors should explain why this does not lead a more 'dishonest' privacy mechanism.
The clipping parameter $C$ is indeed a parameter that affects the privacy guarantees, and tuning it would have an additional privacy cost. Howev... | null | null | null | null | null | null |
Skill-it! A data-driven skills framework for understanding and training language models | Accept (spotlight) | Summary: This paper presents a way to introduce a curriculum of tokens to train LLMs. To do that, the paper introduces the concept of skills, i.e. behaviors that can be acquired by training the LLM in a subset of data, and ordered skill sets: partial orderings of skills that can be learned faster when using the skill p... | Rebuttal 1:
Rebuttal: Thank you so much for your review! We are glad that you found the paper to be novel and address an important problem. We respond to your questions below.
**Robustness to errors or noise in dependency graph:** In our global response, we show that the performance of Skill-It across different random... | Summary: This paper addresses the problem of how to order the training data for a LM. The authors define the notion of a skill, show that skills can be (noisily) identified in training data, show that skills can be ordered, and show that sampling from the data defining a skill in an intelligent manner can improve pret... | Rebuttal 1:
Rebuttal: Thank you so much for your positive review! We are happy that you enjoyed the paper and found it to be well-written.
**How to identify skills:** We agree that identifying skills from data is a challenging problem and an exciting direction we are studying next. We will add this problem to a future... | Summary: The authors introduce a new way of curriculum learning that is thematically similar to active learning: subsets of the data are broken up into "skills" and in cases where there is a dependency structure of skills, learning precursor skills can help improve performance. When the dependency structure is either c... | Rebuttal 1:
Rebuttal: Thank you so much for your positive review. We are glad you liked the paper and think it could be of practical use. We address your concerns below.
**Learning the skills graph:** We hope your questions around the approximate approach and the random baseline with larger training budget to adjust f... | Summary: Review Summary:
The paper explores the concept of clustering data into skills to establish a curriculum, aiming to increase the efficiency of language model (LM) training. The concept of a 'skill' is introduced as a unit of behavior with a corresponding data distribution. A skill, in this context, signifies t... | Rebuttal 1:
Rebuttal: Thank you for your review! We are glad you appreciated the theoretical soundness and clarity of our paper. We address your concerns around the paper’s weaknesses below.
**Computational costs and 1.3B experiments:** In our global response, we demonstrate how computational costs of the skills graph... | Rebuttal 1:
Rebuttal: We thank the reviewers for their valuable feedback, which we have incorporated into our draft. We are glad that reviewers found the paper to address an important and timely problem (6a64, XvRR) and our method to be novel (XvRR), clean (paFF, 6a64), and sound (j55A). Reviewers thought the approach ... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
SLM: A Smoothed First-Order Lagrangian Method for Structured Constrained Nonconvex Optimization | Accept (poster) | Summary: The paper proposes a smoothed first-order Lagrange method(SLM) with the help of a primal-dual optimization framework for solving a class of structured nonconvex functional constrained optimization (FCO) problems. It can be used to deal with lower-level (LL) nonconvex bilevel optimization(BO) problems. Theoret... | Rebuttal 1:
Rebuttal: We thank reviewer VNHZ for your helpful comments and questions.
**Response to Weaknesses 1 (RW1)**
$\epsilon$-KKT points and $\epsilon$-stationary points essentially refer to a point $x^*$ that satisfies the condition:
\begin{equation}
\max \\{|| \mathcal{G}(x^*,y^*)||,|g(x^*, y^*) - g^*(x^*)-\d... | Summary: In this paper, a smooth first-order Lagrangian method (SLM) is proposed to solve a class of structured nonconvex functional constrained optimization problems, where the two optimization variable are nonlinearly coupled. The convergence rates are established by the dual error bounds. The proposed method is used... | Rebuttal 1:
Rebuttal: We thank reviewer Ggqj for your helpful comments and questions.
*Q1: This paper is able to handle nonconvexity and non-singleton at the lower level. However, the assumption that the PL condition is quite strong. If there exist multiple isolated minimizers, then such a condition does not hold. Tho... | Summary: This work studies a class of structured nonconvex functional constrained optimization problems, where the functional constraint satisfies the PL condition w.r.t. one block of variables. The authors propose a smoothed first-order Lagrangian method (SLM) and establish the theoretical convergence rates of SLM to ... | Rebuttal 1:
Rebuttal: We thank reviewer aS1n for your helpful comments and questions.
*W1: Major comment 1: BOME in Table 1 also covers the nonconvexity and non-singleton case. So SLM is not the first work.*
We have thoroughly reviewed reference [14] (i.e., BOME paper). We acknowledge that SLM is not the first work c... | Summary: This paper studies structured nonconvex functional constrained problems and proposed a smoothed first-order Lagrangian method to solve it. The proposed method decomposes the complexity of the problem by introducing a proximal term. Convergence of the proposed method is also studied. In general this is a well-w... | Rebuttal 1:
Rebuttal: We thank reviewer 7siX for your helpful comments and questions.
*Weaknesses: More numerical experiments will make this paper stronger.*
We have incorporated new numerical results, including new loss functions and data set partitions, as well as two additional baselines. Please kindly refer to th... | Rebuttal 1:
Rebuttal: We sincerely thank all the reviewers for their time spent reviewing our work and greatly appreciate the comments provided by each of you. In this general response, we would like to present the F1 score and test accuracy achieved by SLM and V-PBGD, as requested by reviewer aS1n, along with the newl... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Utilitarian Algorithm Configuration | Accept (poster) | Summary: This paper studies the algorithm configuration problem with a monotone (invertible) utility function. We are given a set of $n$ algorithms, and instances are drawn from a distribution, $\mathcal{D}$. We can observe samples of running times of $i\in [n]$ on $j\sim\mathcal{D}$ capped by $\kappa>0$, and we aim to... | Rebuttal 1:
Rebuttal: We thank the reviewer for their time and address each of their concerns below.
Weaknesses:
1. We see the strength of our procedure (UP) as being in its anytime-ness rather than in its use of procrastination. In our context, "procrastination" really just means capping runs at small captimes befo... | Summary: The paper presents the first nontrivial procedure for configuring heuristic algorithms to maximize the utility provided to their end users while also offering theoretical guarantees about performance. The paper shows how to avoid extremely long runs with low probability by putting limits on the runtime, assum... | Rebuttal 1:
Rebuttal: We thank the reviewer for their time. We note only that we are looking at much more extensive experiments in future work.
---
Rebuttal Comment 1.1:
Title: Rebuttal
Comment: I read the rebuttal and I am fine with this paper. | Summary: Theoretical guarantees for algorithm configuration (AC) when using a utilitarian objective function are proposed. This work differs from previous work in several ways, the first being that it focuses on the "utilitarian" objective function, which differs from the objective functions used in essentially all pre... | Rebuttal 1:
Rebuttal: We thank the reviewer for their time and address each of their concerns below.
Weaknesses:
1. We plead ignorance not arrogance. Being so familiar with the work it becomes hard to see it from the perspective of someone who is not, and easy to forget that what seems obvious now was not always so... | Summary: The paper introduces a setting and an algorithm for algorithm configuration based on utility of the algorithm's running time. Theoretical guarantees are provided and upper and lower bounds are compared against. The algorithm is evaluated on case studies.
Strengths: Utility-based decision making approach is ap... | Rebuttal 1:
Rebuttal: We thank the reviewer for their time, unfortunately we have had a hard time understanding their objections. We have tried our best to address each of their concerns below.
"The approach is based on a notion of utility that is not consistent neither with the common practice of utility functions, ... | null | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Improved Bayes Risk Can Yield Reduced Social Welfare Under Competition | Accept (poster) | Summary: This study highlights the inconsistencies of scaling laws in machine learning when model providers compete with each other in the market. The researchers conducted experiments with varying data representations and found that improving data representation quality may actually decrease the overall predictive acc... | Rebuttal 1:
Rebuttal: We thank the reviewer for their feedback. We present **new results** that address some concerns and also respond to other concerns.
**“The experiment only covers CIFAR-10 and reduces it to a binary classification task, which warrants further exploration with datasets of higher cardinality.”**
... | Summary: The paper explores the impact of competition on machine learning models and social welfare. It investigates how scaling trends can be altered by competition and how improving data representation quality may not always lead to better predictive accuracy. The authors provide a theoretical framework for analyzing... | Rebuttal 1:
Rebuttal: We thank the reviewer for their feedback. We present **new results** that address some concerns and also respond to other concerns.
**“The paper would benefit from a more thorough evaluation of the proposed framework, including experiments on real-world datasets.**
Beyond the experiments in the ... | Summary: This paper looks and trends that arise in predictive utility across a population as model providers adapt to improving representations. At the core of the paper is an analysis that characterizes the population utility of equilibria as the representation varies in quality, noise, and dimension, which shows that... | Rebuttal 1:
Rebuttal: We thank the reviewer for their feedback. We present **new results** that address some concerns and also respond to other concerns.
**“There are still some simplifications that weren't explicitly motivated (binarized CIFAR).”**
We provide new results showing that our empirical results in Section... | Summary: This paper shows that "social welfare" might (under some settings) be worse-off as the representation quality of machine learning models increase when there are multiple strategic machine-learning-model providers.
The paper focuses the analysis on binary classification tasks. The paper assumes that users cho... | Rebuttal 1:
Rebuttal: We thank the reviewer for their feedback. We present **new results** that address some of their concerns and also respond to other concerns below.
**“I am not sure whether binary classification is the right problem to analyze the concerned question.”**
We provide new results showing that our ... | Rebuttal 1:
Rebuttal: Thanks to the reviewers for their detailed feedback. In response to reviewer feedback, we provide the following new results: (1) we provide **additional experiments on CIFAR-10 and theoretical results for multi-class classification** and (2) we provide **additional experiments for non-global repre... | NeurIPS_2023_submissions_huggingface | 2,023 | Summary: The authors consider a setting where there exist multiple provides of a machine learning model, and all together choose to increase their representation quality. The authors show, via both theory and experiment, that this improvement in data representation does not necessarily translate to improved social welf... | Rebuttal 1:
Rebuttal: We thank the reviewer for their feedback. We present **new results** that address some concerns and also respond to other concerns.
**“Additional experiments / analyzing settings without these assumptions.”**
We provide new CIFAR-10 experiments on 10 classes (Figure N1, General Response) and ... | null | null | null | null | null | null |
Computational Complexity of Detecting Proximity to Losslessly Compressible Neural Network Parameters | Reject | Summary: This paper studies the replication of the same function using a smaller network. It provides a procedure for achieving optimal lossless compression in the context of single-hidden-layer hyperbolic tangent networks.
This paper introduces the idea of 'rank' of a parameter, defined as the least number of hidden ... | Rebuttal 1:
Rebuttal: We thank reviewer aLjx for their thorough review of our work. We thank the reviewer for pointing out strengths in our paper including our interesting results and clear presentation. We also thank the reviewer for pointing out several legitimate limitations in the scope of our analysis including so... | Summary: The paper studies the idealized phenomenon of lossless compressibility, whereby an identical
function can be implemented with a smaller network, in the setting of single-hidden-layer hyperbolic tangent networks. It introduces the notion of rank as the minimal number of hidden units (used to measure the "size"... | Rebuttal 1:
Rebuttal: We thank reviewer Coti for their review of our work. We thank the reviewer for acknowledging various strengths of our work including its clear presentation, and we graciously thank them for the recommendation that our paper be accepted.
**Weakness:** The main weakness listed by the reviewer is an... | Summary: The authors propose two notions for studying neural network complexity with, on single hidden layer neural networks. The first is _rank_, which corresponds to the smallest number of parameters that can produce a network that is functionally equivalent to the original, and the second is _proximate rank_, which ... | Rebuttal 1:
Rebuttal: We thank reviewer t3vj for their thorough review of our submission. We thank the reviewer for pointing out several strengths in our paper including the potential of our results and our clear structure and exposition. We especially thank the reviewer for their recommendation that the paper be (bord... | Summary: The paper deep-dives into the concept of lossless network compressibility. It presents an algorithm for optimal lossless compression in single-hidden-layer hyperbolic tangent networks, which can in part generalize to other more relevant feedforward architectures. The authors introduce the novel concept of "pro... | Rebuttal 1:
Rebuttal: We thank reviewer ubUt for taking the time to review our paper, especially given it is outside of their specialisation.
The main weakness identified by the reviewer (and some other reviewers) is the limitation in the technical scope of the study (to single-hidden-layer hyperbolic tangent networks... | Rebuttal 1:
Rebuttal: We thank all of our reviewers for their time spent thoroughly reviewing our submission. As a supplement to our individual review responses, we want to take the opportunity here in the top-level author rebuttal to expand on a common thread raised in three of our four reviews, namely the limitation ... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Perturbation Towards Easy Samples Improves Targeted Adversarial Transferability | Accept (poster) | Summary: The paper theoretically and experimentally demonstrated that adding perturbations towards High-Sample-Density-Regions (HSDR) of the target class can further enhance the transferability of targeted attacks. To avoided the impracticality of density estimation in high-dimensional data, authors propose a method fo... | Rebuttal 1:
Rebuttal: Thank you very much for your valuable feedback. Below is our response to your suggestions:
**Concern1**：The recent transfer-based attack [1,2,3] is missing. I suggest authors discuss or compare propsed method with these paper.
**Response to Concern1**: We sincerely appreciate your suggestio... | Summary: This paper presents a hypothesis that pushing adversarial samples towards high-density regions can enhance the transferability of target attacks. The author provides theoretical analysis and experimental validation to support the rationale behind this hypothesis. Based on this hypothesis, the author designs a ... | Rebuttal 1:
Rebuttal: Sincerely appreciate your valuable suggestions. Here is our sincere response to each of your suggestions (due to word limitations, a portion of the questions you raised has been indicated with ellipsis) :
**Concern1**：The paper introduces multiple modules in the proposed method ...... when pret... | Summary: The paper presents both theoretical and experimental evidence showing that incorporating perturbations towards the High-Sample-Density-Regions (HSDR) of the target class can significantly enhance the transferability of targeted attacks. To address the challenges of density estimation in high-dimensional data, ... | Rebuttal 1:
Rebuttal: Thank you very much for highlighting these issues, and we greatly appreciate your thorough examination of our work. Below are our responses to each of these questions and suggestions:
**Concern1**：There are some grammar mistakes (in lines 231-238, the sentences in the whole paragraph are all co... | Summary: This paper finds that optimizing perturbations towards easy examples located in HSDR can improve targeted adversarial transferability. The authors provide toy experiments and theoretical proof to illustrate these conclusions. They further utilize easy examples to guide the targeted attack in the GAN-like frame... | Rebuttal 1:
Rebuttal: We sincerely appreciate your thorough review of our paper and the valuable feedback you have provided. Allow us to address your inquiries in the following response:
**Concern1**： They miss previous works, hierarchical generative networks for multi-target attack [A], and iterative attacks [B, C ... | Rebuttal 1:
Rebuttal: We sincerely appreciate the valuable insights and suggestions provided by the esteemed reviewers during the review process. We extend our gratitude for the time and effort you dedicated to scrutinize our paper. Your reviews are of paramount importance to our research endeavor, and your insightful ... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
StableFDG: Style and Attention Based Learning for Federated Domain Generalization | Accept (poster) | Summary: This paper proposes a style and attention based learning strategy for federated domain generalization which consists of two stages. The first stage shares style information among clients and then explorates and augments styles, and the second stage highlights the essential characteristics of images with attent... | Rebuttal 1:
Rebuttal: We thank the reviewer for the positive comments and feedback. In particular, we appreciate the reviewer’s acknowledgment on the novelty of our work. Below, we reply to the comments raised by the reviewer.
### **[Question 1] Insights into equation (7)**
We appreciate this comment. The sim... | Summary: The authors observe that existing domain generalization (DG) algorithms face the issue in the setting of federated learning (FL) because of the lack of samples in each client's local datasets/domains. They propose StableFDG, which utilizes both style-based learning which allows the clients to explore styles di... | Rebuttal 1:
Rebuttal: We appreciate the reviewer for the positive comments and feedback. We are also thankful for the suggestions to discuss connections between our style-based learning and existing style transfer. Our responses to the comments raised by the reviewer are given below.
### **Intuition behind equati... | Summary: In this paper, the authors present StableFDG, where a style-based strategy enables the model to reach a variety of new styles beyond each client's source domain, and an attention-based approach captures and emphasizes the important/common features of each category.
Strengths: The StableFDG method proposed in ... | Rebuttal 1:
Rebuttal: We appreciate the reviewer’s constructive suggestions. We make all of the raised points clearer below, with additional experiments as needed.
### **Differentiation from [1] and novelty of Sec. 3.1**
Our style-based learning methodology in Sec 3.1 has three key components: (i) style sharing, (ii) s... | Summary: This paper presents a novel domain generation method in the FL framework. The authors make two key contributions to enhance domain generalization in FL. Firstly, they introduce style-based learning, which empowers individual clients to explore new styles by leveraging both their own local statistics and those ... | Rebuttal 1:
Rebuttal: We appreciate the reviewer for the time and efforts, and providing helpful comments. Our responses are given below.
### **Style-based learning**
We would like to first clearly differentiate the concept of existing style-augmentation [ICLR'21], [ICLR'22] and our style-based learning, and show that ... | Rebuttal 1:
Rebuttal: We appreciate all reviewers for providing constructive comments, which have greatly helped us to improve the paper.
Due to the limited content we can provide in each response, we would like to share additional experimental results that **Reviewer i1eF** and **Reviewer fAoj** suggested here. For ... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
On the Adversarial Robustness of Out-of-distribution Generalization Models | Accept (poster) | Summary: This paper studies the adversarial robustness of models on the unseen target domain. Firstly, exhaustive experiments show that existing OOD generalization methods are vulnerable to adversarial attacks. Then, a theoretical analysis for the OOD adversarial robustness is presented. Finally, this paper proposes tw... | Rebuttal 1:
Rebuttal: Thank you for your meaningful suggestions.
**For the weaknesses:** Thank you for your suggestion, we have extended our experimental results to the PACS dataset and OfficeHome dataset in DomainBed, please refer to **Table 1** in the reply to all reviewers part. From the table we can see that: on t... | Summary: The paper first evaluates the adversarial robustness of current OOD generalization algorithms and shows that existing OOD generalization algorithms are vulnerable to adversarial attacks. Then, the paper presents a theoretical analysis of the adversarial robustness (of a model) on the unseen test domain. After ... | Rebuttal 1:
Rebuttal: Thank you for your meaningful suggestions.
**For the weakness (1)**: Thank you for your kind suggestion. We are sorry for not noticing such a related work. Although the title of [1] is about the OOD adversarial robustness, the topic that [1] studies is quite different from ours. [1] investigates ... | Summary: This paper studies the adversarial robustness of OOD generalization algorithms. Authors claimed that existing OOD generalization algorithms fail in adversarial attacks (FGSM and PGD). To this end, authors first theoretically analyze the problem definition and generalization bound of OOD algorithms in adversari... | Rebuttal 1:
Rebuttal: Thank you for your suggestions.
**Comparison with related works.** We are sorry for ignoring the related works, we now provide comparisons and will incorporate them in the revision.
- [Yi et al'2021, ICML] studies the relationship between adversarial robustness and OOD generalization, it shows th... | Summary: The authors use statistical learning theory tools and the HdH divergence to motivate methods to improve the performance of a domain generalizing model dealing with adversarially perturbed, out-of-distribution domains. The authors evaluate two methodologies in that setting, and show how this can indeed defend m... | Rebuttal 1:
Rebuttal: Thank you for your valuable suggestions.
**For weakness 1:** We are sorry for missing the references and comparisons.
- "Adversarial Feature Desensitization" focuses on improving the adversarial robustness of the models by regarding the adversarial distribution as the target domain and then using... | Rebuttal 1:
Rebuttal: # The reply to all reviewers
We would like to thank all the reviewers for all your valuable suggestions, which help us improve the quality of our paper.
According to the suggestion of reviewers, we conduct extra experiments on the datasets PACS and OfficeHome. We present the new results in Table... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Calibration by Distribution Matching: Trainable Kernel Calibration Metrics | Accept (poster) | Summary: Accurate probability prediction is a crucial aspect of a trustworthy model, especially for deep neural networks. This paper focuses on the model calibration problem and proposes a unified framework via distribution matching. To leverage the distribution matching, they used
Maximum Mean Discrepancy to estimate... | Rebuttal 1:
Rebuttal: Thank you for your time and feedback! We have focused on improving the related works section to include more direct comparisons to [Kumar et al., 2018] and other kernel-based calibration metrics. We have included updated comparisons in our general author response. We respond to your questions and ... | Summary: This paper presents a general framework for train-time calibration based on kernel MMD. The authors demonstrate empirically that choosing good kernels (based on the task at hand) in their framework can lead to improvements in calibration while maintaining sharpness of predictions.
Update: I have increased my ... | Rebuttal 1:
Rebuttal: Thank you for your thoughtful feedback on our work! We are grateful that you appreciated the quality and originality of our work. We believe we have thoroughly responded to each of your comments, and that these changes have improved our work. If you find our responses and the corresponding updates... | Summary: This paper introduces differentiable kernel-based calibration metrics that unify and generalize popular forms of calibration.
The notions of calibration are presented as distribution matching constraints by using various kernels in the Maximum Mean Discrepancy (MMD) metric.
Through empirical evaluation, it is ... | Rebuttal 1:
Rebuttal: Thank you for your helpful comments! We are glad you appreciated the trainable nature of our metrics, the decision making connections, and the experimental results. We agreed with your suggestions to expand on the applications to classification and to add more detailed descriptions of related work... | Summary: Calibration as a training regularizer:
This paper focuses on investigating how to incorporate calibration metrics as training regularizers. It starts with summarizing different calibration metrics using a unified view from distribution matching. The paper then proposed a kernel-based general metric that maps ... | Rebuttal 1:
Rebuttal: Thank you for your time and the helpful feedback! Below, we clarify our contributions surrounding decision calibration, the role of post-hoc calibration, and respond to additional questions. Given our response to the main issues you raised in the review, we kindly request a re-evaluation of your s... | Rebuttal 1:
Rebuttal: We thank the reviewers for engaging thoughtfully and providing constructive feedback. We are glad that all reviewers appreciate that this work is “novel” [hWT9], “original” [1Qie], “a very interesting view for uncertainty calibration” [UW9V], and “offers valuable insights for practical problems” [... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Bayes beats Cross Validation: Efficient and Accurate Ridge Regression via Expectation Maximization | Accept (poster) | Summary: This paper tackles the problem of choosing the regularization parameter in ridge regression. The authors introduce a Bayesian model that has the ridge estimate as its posterior mode with a prior distribution over the regularization parameter; they then propose an EM algorithm to estimate the optimal regulariza... | Rebuttal 1:
Rebuttal: We thank you for the positive comments and the assessment that the paper is likely impactful.
**Regarding your concern about novelty:** We show for the first time how the specific structure of ridge regression can be exploited to derive substantially faster EM updates in this important special c... | Summary: This paper proposes a simplified implementation of the EM algorithm of the Bayesian Ridge regression for the hyperparameter estimation. The advantage of the proposed method is in its computational cost when searching the hyperparameter. The numerical experiments support their proposal in terms of computational... | Rebuttal 1:
Rebuttal: We appreciate the positive comments on the simplicity and effectiveness of the proposed method. However, we disagree with the assessment of our work’s significance. In particular, our technical contributions extend beyond merely proposing a singular-value decomposition (SVD) as pre-processing.
U... | Summary: This article proposes a new method based on an Expectation-Maximisation framework to train a Bayesian formulation of ridge regression for penalised regression. It is compared with the standard cross-validation approach for training the regularisation parameter lambda, instead iteratively updating the value alg... | Rebuttal 1:
Rebuttal: We appreciate the reviewers' positive comments on our method's novelty, high usability, and principled presentation, and we are delighted to learn that the reviewer enjoyed reading the paper.
Importantly, we want to emphasize that both the EM method and LOOCV share the same underlying model, whic... | Summary: This paper studies the comparison of cross validation (CV) and expectation maximization (EM) on tuning the regularization hyper-parameter $\lambda$.
The authors explained the running time of these two methods and showed that EM is faster than CV.
Then, the authors also provided some experimental results to su... | Rebuttal 1:
Rebuttal: We thank the reviewer for the positive evaluation. Regarding some further theoretical insights we refer to the [global rebuttal](https://openreview.net/forum?id=Ih2yL7o2Gq¬eId=OV1Ci3vVUB).
While we do not currently have a specific explanation for the typically similar predictive performance, ... | Rebuttal 1:
Rebuttal: We thank all reviewers for their thoughtful comments and mostly positive evaluation of our work. Some reviewers encouraged us to provide further theoretical insights into the behaviour of the proposed EM method. Here, we summarise some new considerations that we intend to include in the final vers... | NeurIPS_2023_submissions_huggingface | 2,023 | Summary: This paper proposes a new method for estimating the regularization parameter in ridge regression models that is faster, than the widely used leave one out cross validation (LOOCV) technique. The new method is derived from a Bayesian view for estimating this regularization parameter which, through standard Expe... | Rebuttal 1:
Rebuttal: Thank you for your positive feedback and especially your assessment that the paper will be a good addition to the literature. Regarding the questions you identified:
**When is the proposed method worse than the fast LOOCV?**
While this is a natural question, it is from our perspective not easy t... | null | null | null | null | null | null |
Efficient Model-Free Exploration in Low-Rank MDPs | Accept (poster) | Summary: This paper studied model-free exploration in low-rank MDP. The authors proposed an algorithm based on the idea of barycentric spanner, which is provably efficient given a reachability assumption weaker than previous literatures.
Strengths: The paper writing is clear. The comparsion with previous related work ... | Rebuttal 1:
Rebuttal: Thank you for the review and useful suggestions!
**“I have doubt about whether the algorithm can be called "efficient"...”**
Regarding the computational efficiency of our algorithm, as noted, many other works use the same minimax oracle (in fact, our oracle is slightly weaker because the maximum ... | Summary: The paper studies the reward-free exploration problem for low-rank MDPs. The approach is based on the combination of the barycentric spanner and a model-free representation module. The authors provide a theoretical analysis of the algorithm for both the reward-free exploration task and the downstream reward-ba... | Rebuttal 1:
Rebuttal: Thank you for the positive comments and helpful suggestions!
**“While the paper studies the exploration problem of low-rank MDPs, its contribution mainly lies in the planning phase…”**
Thank you for your review. We agree that our primary algorithmic contribution is not in the representation learn... | Summary: This paper studies model-free RL under low-rank MDPs. This paper first generalizes previous fromulation of low-rank MDPs to a uncountable state case. Then this paper proposes a non-optimistic, model-free RL algorithm called SpanRL, which can provide a good policy cover with polynomial sample complexity. With p... | Rebuttal 1:
Rebuttal: Thank you for the positive comments and helpful suggestions!
**“One advantage of the algorithm emphasized in this paper is its computational efficiency. However, there is no experiments to validate this claim.”**
Regarding the lack of experiments, the contribution of the paper is primarily theore... | Summary: The paper proposes a new algorithm for the low-rank MDPs setting. With a new reachability assumption, the paper shows that one can learn a policy cover in a statistically and computationally efficient manner. The proposed algorithm works in a modular manner, using barycentric spanner for exploration and the re... | Rebuttal 1:
Rebuttal: Thank you for the positive comments.
**“Unlike previous papers such as BRIEE or MUSIK, the proposed computational efficient algorithm is not empirically tested…”**
Regarding the lack of experiments, the contribution of the paper is primarily theoretical, and we believe that our theoretical contr... | null | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Conditional independence testing under misspecified inductive biases | Accept (spotlight) | Summary: The paper proposes a new Rao-Blackwellized Predictor Test which is a regression-based conditional independence test. It was shown to be robust in misspecification settings.
Strengths: The paper is in general well-written and considers an important scenario which is not covered by standard conditional indepen... | Rebuttal 1:
Rebuttal: ## Thank you
We want to thank you for your time and work on our paper.
We believe we have addressed your questions in our comments. Please let us know if you have further questions or comments.
Best regards,
Authors
--------------
## Rebuttal
**Question about universal approximators.** We i... | Summary: The authors study model-based tests of conditional independence under conditions where the models may be biased. They propose new approximations and bounds on testing errors resulting from bias in model-based CI tests. They also introduce RBPT, a new model-based CI test and evaluate it against several alternat... | Rebuttal 1:
Rebuttal: ## Thank you
We greatly appreciate your work on our paper.
Thank you for the raised points. We have addressed them below. Please let us know if you have any other comments or questions.
Best regards,
Authors
---------------------
## Rebuttal
**Superiority of RBPT/RBPT2 over STFR in the real... | Summary: This paper presents new robustness results for three regression-based conditional independence (CI) tests, namely, Significance Test of Feature Relevance (STFR), Generalized Covariance Measure (GCM) test, and REgression with Subsequent Independence Test (RESIT). It introduces a new CI test, the Rao-Blackwelliz... | Rebuttal 1:
Rebuttal: ## Thank you
We greatly appreciate your feedback on our paper.
Thank you for your questions. We have addressed them below. Please let us know if you have any other comments or questions.
Best regards,
Authors
------------------
## Rebuttal
**Misspecification discussion.** You're correct tha... | Summary: This paper discusses the challenges of conditional independence (CI) testing in modern statistics and machine learning. The authors note that many modern methods for CI testing rely on supervised learning methods to learn regression functions or Bayes predictors, but these methods can be unreliable when the mo... | Rebuttal 1:
Rebuttal: ## Thank you
Thank you for your time and work on our paper.
We believe we have addressed your concerns and questions in our comments. Please let us know if you have any other questions.
Best regards,
Authors
______________
## Rebuttal
**Time complexity of RBPT and RBPT2.** The running times ... | Rebuttal 1:
Rebuttal: We extend our sincere thanks to all reviewers and chairs. We have responded to all the points raised by reviewers individually and have attached a PDF containing additional experiments. The first of these experiments demonstrate that the robustness of RBPT2 in higher dimensions can be improved thr... | NeurIPS_2023_submissions_huggingface | 2,023 | Summary: The paper under review offers a thoughtful exploration of how model misspecification influences conditional independence testing. The significance and practical applications of such testing are aptly underscored in the introduction. The use of a toy example to demonstrate the impact of model misspecification o... | Rebuttal 1:
Rebuttal: ## Thank you
Thank you for your time and constructive feedback. Your suggestions made our paper more complete.
We believe we have addressed your requests and questions in our comments. Please let us know if you have further concerns.
Best regards,
Authors
--------------
## Rebuttal
**Semi-s... | null | null | null | null | null | null |
GOAt: Explaining Graph Neural Networks via Graph Output Attribution | Reject | Summary: This paper proposes a novel local-level post-hoc GNN explainer (GOAt) that takes the issue of discriminability and consistency into consideration. The discriminability and consistency of graph explanation are significant and meaningful for the community’s understanding of the GNN decision-process. The GOAt ex... | Rebuttal 1:
Rebuttal: * **(W1) The motivation behind GOAt in addressing the discriminability and inconsistency.**
* The existing approaches fail because most of them are trained or searched towards some human-defined metrics (such as explanation accuracy), which are not guaranteed to be faithful to GNNs. Other metho... | Summary: In this paper, the local explanation algorithm for Graph Neural Networks (GNN), Graph Output Attribution (GOAt), is proposed. The key idea of the algorithm is decomposing the GNN's forward propagation as sum of product form. After decomposing, authors compute the contributions of nodes and edges by proposed me... | Rebuttal 1:
Rebuttal: * (**W1, Q1**) **Why is zero-vector used in the _Equal Contribution_ algorithm?**
* The elements in the adjacency matrix can be {0,1}, indicating the absence or presence of edges between pairs of nodes. Any values between {0,1} are actually not relevant to the problem, as they do not carry any... | Summary: It proposes a local GNN explanation method called GOAt by analysing the output attribution of graphs. Specifically, it proposes a novel explanation framework by the notion of equation contribution, which is transparent compared to the existing black-box type explanation models. Experiments on real and simulati... | Rebuttal 1:
Rebuttal: * **(W1, W2) Should explain the differences between global-level explanations and local-level explanations and show the superiority of global-level explainability in the related work and problem formulation.**
Global-level GNN explainability is a parallel sub-domain to the local-level explain... | Summary: Unlike existing methods that rely on training auxiliary models, the paper introduces a computationally efficient instance-level GNN explanation technique (called GOAt), which enables attribution of GNN output to input features by direct algebraically computation. More specifically, a GNN is represented in an e... | Rebuttal 1:
Rebuttal: * **(W1) The explainer can become very complex as the network goes deeper.**
* Our paper only focuses on the Graph Neural Network explainability instead of the general neural network explainability. As outlined in the "Limitations and Broader Impacts" section, our proposed approach can effective... | Rebuttal 1:
Rebuttal: ### **Response to all the reviewers**
Thanks to all reviewers for the constructive feedback. First, we would like to clarify our contribution as compared to the existing literature of GNN explanation:
1. Existing GNN explaining methods suffer from issues of **insufficient discriminability** betw... | NeurIPS_2023_submissions_huggingface | 2,023 | Summary: This work studies the problem of GNN local explanation. It points out that existing GNN explainers are suffering from a few limitations including insufficient discriminability, inconsistency on same-class data samples, and overfitting to noise, and the aim to address these limitations by proposing the GOAt me... | Rebuttal 1:
Rebuttal: * (**W2 & Q1**) **How is the contribution of an edge determined?**
* _Equal Contribution_:
* Consider a product term $z = 10A_{11}A_{12}A_{23}$, where the variables $A_{11} = A_{12} = A_{23} = 1$ indicate that all three edges exist, resulting in $z = 10$. If any of the edges is missing (... | null | null | null | null | null | null |
Online Performative Gradient Descent for Learning Nash Equilibria in Decision-Dependent Games | Accept (poster) | Summary: This paper studies the problem of multi-agent performative game. The authors generalize the existing assumption in (Narang et al., 2022) from linear cases to generalized linear cases, and proposed algorithms consisting online stochastic approximation of the link function and individual gradient updates. They p... | Rebuttal 1:
Rebuttal: Thanks for your review and suggestions. We have explained our novelty over Narang et al. (2022) in the global response. We refer the reviewer to **Novel I** for the comparison of OPGD and Algorithm 1 and Narang et al. (2022). We refer to **Novel II** for the explanation that extending the linear a... | Summary: The authors study multi-agent decision-dependent games. Due to strategic interactions among agents, classical gradient-based methods cannot be directly applied in computing Nash equilibrium. Instead, the authors propose using parametric models to model the agents' interactions and the resulting OPGD algorithm.... | Rebuttal 1:
Rebuttal: Thanks for your insightful comments and kind suggestions! We clarify our novelty over Narang et al. (2022) in the global response. We refer the reviewer to **Novel I** for the comparison of OPGD and Algorithm 1 and Narang et al. (2022). We refer to **Novel II** for the explanation that extending t... | Summary: This work proposes stochastic approximation-based Online Performative Gradient Descent to compute the Nash equilibrium of multi-agent decision-dependent games with bandit feedback. To this end they extend the work of Narang et al. (2022) in two ways: 1. The work uses stochastic approximation method to estimate... | Rebuttal 1:
Rebuttal: Thanks for your detailed and valuable review. We have outlined the contribution of our work in the global comment. Next, we will address the questions not included in that particular response.
**"I am concerned about the use of $\cdots$":**
Thanks for this question. We regard the use of $u_i^t$ a... | Summary: This work studies the framework of decision dependent games, where the actions of the agents in the game do not only affect the utilities of the players but also the distribution of population data observed by them. The critical challenge of this type of games is that the data population sampling process is no... | Rebuttal 1:
Rebuttal: Thanks for your positive feedback and the time spent reviewing and understanding our paper! We have added more explanation on the contribution and innovation of our work in the global response. We address the other comments as follows:
**"I think the differentiation between Narang et al $\cdots$"... | Rebuttal 1:
Rebuttal: In the following, we present a detailed explanation of our novelty over Narang et al. (2022).
**Novelty I.** We extend the linear model (Assumption 6) in Narang et al. (2022) (which assumes agents share a common parametric matrix $A = [A_i~A_{-i}]$) into a general framework (parametric functions... | NeurIPS_2023_submissions_huggingface | 2,023 | Summary: The authors focus on the analysis of learning the Nash equilibrium (which is unique) in strongly monotone games that are decision-dependent. They provide a definition of decision-dependent games generalizing a recent work of Narang et al. Intuitively speaking, if an agent chooses an action, he gets an observat... | Rebuttal 1:
Rebuttal: Thanks for your time and effort in reviewing our paper. We have summarized the contribution and innovation of our paper over Narang et al. (2022) in the global response. In the following, we address other comments not covered in that response.
**"I think the paper would be more interesting $\cdot... | null | null | null | null | null | null |
Stochastic Approximation Approaches to Group Distributionally Robust Optimization | Accept (poster) | Summary: This paper provides near-optimal optimization algorithm to group distributionally robust optimization problem (GDRO). Since it is a special case of the stochastic convex-concave saddle point algorithm, the standard SMD could indeed achieve near-optimal sample complexity. Besides, the authors also apply online ... | Rebuttal 1:
Rebuttal: Many thanks for the constructive reviews!
---
Q1: No numerical study is presented to demonstrate the convergence rate of proposed algorithm
A1: We have carried out preliminary experiments, the results of which can be found in the global response to all reviewers. These empirical findings align ... | Summary: The paper considers group DRO and shows an SMD algorithm to optimize. This algorithm requires a sample from every distribution in each iteration and matches the sample complexity lower bound up to a log factor. The second algorithm is designed for the case when the online functions depend on past decisions. Fo... | Rebuttal 1:
Rebuttal: Many thanks for the constructive reviews! We believe there are some misunderstandings about the work of Soma et al. [2022], and provide detailed responses below. We hope the reviewer could examine them, and reassess the significance of our results. We are looking forward to addressing any further ... | Summary: The paper presents a comprehensive study on group distributionally robust optimization (GDRO) by formulating it as a convex-concave saddle point problem. The authors propose utilizing mirror descent algorithms and online convex optimization techniques to effectively address this problem. Furthermore, the autho... | Rebuttal 1:
Rebuttal: Many thanks for the constructive reviews! We provide detailed responses below, and hope the reviewer could reassess the significance of our results. We are looking forward to addressing any further question in the reviewer-author discussion period.
---
Q1: Theorem 1 seems to be directly derived f... | Summary: In this paper, the authors present a systematic investigation of group distributionally robust optimization (GDRO). First, they consider the standard standing where the algorithm can draw samples from all distributions freely. In this case, they demonstrate that stochastic mirror descent (SMD) attains an optim... | Rebuttal 1:
Rebuttal: Many thanks for the constructive reviews!
---
Q1: The authors have mentioned that the work of Haghtalab et al. [2022] suffers a dependency issue. Could you please elaborate on this point? In particular, which part of their analysis is problematic?
A1: In the work by Haghtalab et al. [2022], spe... | Rebuttal 1:
Rebuttal: We have conducted preliminary experiments to evaluate our algorithms. Below, we present the experimental setting and empirical results. All related figures can be found in the accompanying PDF.
## Datasets
We first use a **synthetic** data from Soma et al. [2022]. We set the group number $m=20$,... | NeurIPS_2023_submissions_huggingface | 2,023 | Summary: This paper provides algorithms for group distributionally robust optimization.
For the case in which the sample budget for each distribution is identical,
the proposed algorithm achieves nearly optimal sample complexity.
This paper also considers the sample budgets are different for each distribution and provi... | Rebuttal 1:
Rebuttal: Many thanks for the constructive reviews!
---
Q1: The obtained bounds in this work includes extra $\log m$ factors, unlike ones by [Soma et al., 2022].
A1: First, it is important to recognize that the original analysis of Soma et al. [2022] is *flawed*. For a detailed explanation, please refer t... | null | null | null | null | null | null |
Toward Causal-Aware RL: State-Wise Action-Refined Temporal Difference | Reject | Summary: In this paper, the authors introduce a method called State-Wise Action Refined to explore the causal relationship between actions and task rewards in reinforcement learning. They address the issue of action space redundancy and propose interventions on the primal action space as a means to discover causality. ... | Rebuttal 1:
Rebuttal: We sincerely appreciate the reviewer's efforts and constructive comments in improving our paper. Below please find our response:
---
Our **technical contribution** is to introduce the state-wise action refinery based on the causal relationship in temporal difference learning to improve learning e... | Summary: This paper presents a strategy for action space reduction, based on causality-inspired modeling. It employs a method of action space selection, using collected data to identify the specific dimensions of actions that are meaningful to the task, as well as to improve the efficiency of exploration in reinforceme... | Rebuttal 1:
Rebuttal: We deeply appreciate the reviewer's time and effort in evaluating our paper. We will now proceed to answer each of the posed questions:
---
### Q1: Intervention
### A1:
In general, an _intervention_ in causal discovery is an external manipulation of a variable in the system, and we observe the ... | Summary: The paper proposes an intervention-based method for understanding causality in Reinforcement Learning, aiming to enhance learning efficiency in tasks with redundant actions. The authors present two novel algorithms, Temporal Difference State-Wise Action Refined (TD-SWAR) and Dynamic State-Wise Action Refined (... | Rebuttal 1:
Rebuttal: We thank the reviewer for the insightful comments and constructive feedback. We will respond to each of the questions in turn:
---
### Q1: More Comparisons, Baselines
### A1:
In order to perform a more comprehensive evaluation and provide more empirical evidence, we include the **SAC and PPO a... | Summary: This paper shows a Q-learning-based RL algorithm that jointly performs feature selection to detect action variables that are not relevant to dynamics or rewards. The feature variable selection algorithm adopts INVASE, and the experiment confirms the idea of five artificially created extended environments.
---... | Rebuttal 1:
Rebuttal: We sincerely appreciate the reviewer's thoughtful insights and constructive feedback. We will respond to each of the comments and questions in turn.
---
### Point 1: Additional Experiment Results with Different Numbers of Nuisance Actions
### Reply 1:
**We use 100 redundant actions to stress-t... | Rebuttal 1:
Rebuttal: We extend our sincere gratitude to all reviewers for their insightful comments, valuable suggestions, time, and efforts in evaluating and improving our paper.
---
To address the concerns raised by reviewers, we would respond to each of their questions respectively. Below, as a general response, ... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Can Language Models Teach? Teacher Explanations Improve Student Performance via Personalization | Accept (poster) | Summary: This work centers around the research question: *can stronger LLMs teach weaker ones with their explanations? And if yes, when and how should the teacher do this*? It answers a few research questions, progressively building up from using a teacher LLM to generate explanations for a randomly sampled subset of t... | Rebuttal 1:
Rebuttal: We thank the reviewer for their thorough review, acknowledging our experimental setup and finding the result on Expected Utility compelling. Below we address your questions and comments.
> **Many details are left out of the paper**
Please refer to the common response and detailed answers here. W... | Summary: The authors investigate whether language models make good teachers for other, weaker language models. They do this through a systematic investigation of a teacher LM intervening at test time to improve a student LMs performance. They take seriously the idea that the teacher intervention imposes some cost (prim... | Rebuttal 1:
Rebuttal: We thank the reviewer for their encouraging comments and detailed review, along with providing useful references. We will add them in the final version. Below we address your questions and concerns.
> **Statistical analysis of results**
Thanks for the suggestion! Please refer to our common respo... | Summary: In this work, the authors tackle exciting research questions on how the teacher LLM intervention, which provides explanations in natural language, can enable the student LLM to make more accurate predictions. They specifically used StrategyQA, GSM8k, and CommonSenseQA for their experiments. The direction of th... | Rebuttal 1:
Rebuttal: We thank the reviewer for acknowledging our exciting research questions and our experimental setup with direct answers. Below we address your questions and concerns.
> **Missing appendix**
Please note that our paper **does have** an accompanying appendix. As per the rules, it is uploaded as a se... | Summary: The paper studies how an LLM can be used to teach another LLM how to perform reasoning tasks by providing explanations (chain-of-thoughts). The paper starts by establishing that when the teacher LLM intervenes at test-time by providing such explanations, the student LLM accuracy is indeed improved compared to ... | Rebuttal 1:
Rebuttal: We thank the reviewer for finding value in our research questions and their ability to motivate future research. We are also encouraged by the reviewer’s appreciation of our Expected Utility-based intervention function and generating personalized explanations. Below, we address your questions and ... | Rebuttal 1:
Rebuttal: > **Experiments with a stronger teacher and significance of results**
Some reviewers were curious about the statistical significance of our results. Since experiments with larger teachers are significantly more time and resource intensive, we had to choose a teacher model that while stronger than... | NeurIPS_2023_submissions_huggingface | 2,023 | null | null | null | null | null | null | null | null |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.